2015 : WHAT DO YOU THINK ABOUT MACHINES THAT THINK?
The history of humanity and the history of technology are conjoined. We have always used our cognitive capacities to create the objects we needed to survive, from weapons to garments and shelters. The evolution of the human mind is instantiated in the evolution of technology. We have developed a capacity for metarepresentation—a capacity to be aware of having, and to analyze our own minds—which is a function of higher order consciousness. And in order to look at ourselves in the mirror, we have always used technological analogies, compared our minds to the technologies we had created. To each era its machine—from hydraulic pumps to computers.
We have by now created technologies that no single person is able to master. Our creations are starting to escape our own minds. No wonder then that we so easily imagine the creations becoming creatures in their own right, endowed with minds as agile as ours, or more agile perhaps. Science fiction imagines perfect robots, indistinguishable from ourselves, embodied, speaking, seemingly feeling, that can fool and even perhaps attack us.
But in thinking conceptually about our own minds, we tend to remain Cartesian dualists. Thinking seems so disembodied an activity that we forget that we are emphatically not brains in vats, that no amount of microtechnology will recreate the complexities of biology thanks to which our brains function, replete with neurotransmitters, enzymes, and hormones. We are our bodies, we have emotions that are embodied and that deeply inform our thinking processes. Machines are developing task-driven cognitive capacities, but their perfect processing is very different indeed from the imperfect, inconstant, subtle thinking of persons endowed with a sense of self, proprioception, a sense of centeredness, the "qualia" that distinguishes us from "zombies."
Computers excel at processing processes most of us fumble with, and we are increasingly accessing the world of facts via machines. Much of our memory is assigned to Google, and there is no doubt that our minds are increasingly extended beyond our single bodies, that we exist within an increasingly large network of disembodied minds and data. Thinking is itself in part a socially given capacity, and to think is to participate of a collective enterprise. But the complexity of this enterprise is as much a characteristic of the human condition as is our embodiment. Machines do not have social lives any more than they are embodied within a complex, evolved set of biological tissues. They are good at tasks, and we have become very good at using them for our purposes, and for expanding our capacity for communication. But until we replicate the embodied emotional being—a feat I don't believe we can achieve—our machines will continue to serve as occasional analogies for thought, and to evolve according to our needs.
For years we've been making the case that artificial intelligence, and in particular the field of machine learning, is making rapid progress and is set to make a whole lot more progress. Along with this we have been standing up for the idea that the safety and ethics of artificial intelligence is an important topic we should all be thinking about very seriously. The potential benefits of artificial intelligence will be vast, but like any powerful technology these benefits will depend on this technology being applied with care.
While some researchers have been cheering us on since the start of DeepMind, others have been very skeptical. However, in recent years the climate for ambitious artificial intelligence research has much improved, no doubt due to a string of stunning successes in the field: not only have a number of long standing challenges finally fallen, but there is a growing sense among the community that the best is yet to come. We see this in our interactions with a wide range of researchers, and it can also be seen from the way in which media articles about artificial intelligence have changed in tone. If you hadn't already noticed, the "AI Winter" is over and the spring has begun.
As with many trends, some people have started to become a little bit too optimistic about the rate of progress, going as far as predicting that a solution to human level artificial intelligence might be just around the corner. It's not. Furthermore, with the very negative portrayals of futuristic artificial intelligence in Hollywood, it is perhaps not surprising that doomsday images are appearing with some frequency in the media. As Peter Norvig aptly put it, "The narrative has changed. It has switched from, 'Isn't it terrible that AI is a failure?' to 'Isn't it terrible that AI is a success?' "
As is usually the case, the reality is not so extreme. Yes, this is a wonderful time to be working in artificial intelligence, and like many people we think that this will continue for years to come. The world faces a set of increasingly complex, interdependent and urgent challenges that require ever more sophisticated responses. We'd like to think that successful work in artificial intelligence can contribute by augmenting our collective capacity to extract meaningful insight from data and by helping us to innovate new technologies and processes to address some of our toughest global challenges.
However, in order to realise this vision many difficult technical issues remain to be solved, some of which are long standing challenges that are well known in the field. While difficult, we think these problems can be overcome, but that it will take a generation of talented researchers equipped with plentiful computational resources and inspired by insights from machine learning and systems neuroscience. While this is likely to disappoint the most optimistic observers, it will give this community some time to come to grips with the many subtle safety and ethical questions that will arise. So let's enjoy this new sense of optimism, but let's not lose sight of how much hard work is left to do. As Turing once said: "We can only see a short distance ahead, but we can see plenty there that needs to be done."
The first question that comes to our minds, as we think about machines that think, is how much these machines will, eventually, be like us. To us, this comes down to a question of self. Will thinking machines ever evolve to the point of having a sense of self that resembles that of humans? We are (probably) the only species capable of self-consciously thinking about who we are: of not only knowing our selves, but being able to evaluate those selves from a uniquely internal, self-reflective perspective.
Could machines ever develop this kind of self? Might they experience the same evolutionary forces that made human selves adaptive? These include the need to get along with others, to attain status, and to make sure others like us and want to include us in their social groups. As a human being, if you want to succeed at group living it helps to have a self you're motivated to protect and enhance; this is what motivates you to become the kind of person others like, respect, and grant power to, all of which ultimately enhances your chances of surviving long enough to reproduce. Your self is also what allows you to understand that others have selves of their own—a recognition that's required for empathy and cooperation, two prerequisites for social living.
Will machines ever experience these kinds of evolutionary forces? Let's start with the assumption that machines will someday control their own access to resources they need, like electricity and internet bandwidth (rather than having this access controlled by humans), and will be responsible for their own "life" and "death" outcomes (rather than having these outcomes controlled by humans). From there, we can next assume that the machines that survive in this environment will be those that have been programmed to hold at least one basic self-related goal: that of increasing their own efficiency or productivity. This goal would be akin to the human gene's goal of reproducing itself; in both cases, the goal drives behaviors oriented toward boosting fitness, of either the individual possessing the gene, or the machine running the program.
Under these circumstances, machines would be motivated to compete with each other for a limited pool of resources. Those who can form alliances and cooperate—that is, sacrifice their own goals for others, in exchange for future benefits—will be most successful in this competition. In other words, it's possible to imagine a future in which it would be adaptive for machines to become social beings that need to form relationships with other machines, and therefore develop human-like selves.
However, there's a major caveat to this assumption. Any sociality that comes to exist among thinking machines would be qualitatively different from that of humans, for one critical reason: Machines can literally read each other's minds. Unlike humans, machines have no need for the secondary—and often deeply flawed—interpretative form of empathy we rely on. They can directly know the contents of each other's minds. This would make getting along with others a notably different process.
Another way of putting this is to say that, despite the critical importance of our many social connections, in the end, we humans are each fundamentally alone. Any connection we feel with another's mind is metaphorical; we cannot know, for certain, what goes on in someone else's head—at least not in the same way we know our own thoughts. But for machines, this constraint does not exist. Computers can directly access each other's inner "thoughts", and there's no reason that one machine reading another's hardware and software wouldn't come to know, in exactly the self-knowing sense, what it means to be that other machine. Once that happens, each machine is no longer an entirely separate self, in the human sense. At that point—when machines literally share minds—any self they have would necessarily become collective.
Yes, machines could easily keep track of the sources of various bits of information they obtain, and use this tracking to distinguish between "me" and other machines. But once an individual understands another at the level that a program-reading machine can, the distinction between self and other becomes largely irrelevant. If I download all the contents of your PC to an external hard drive, then plug that into my PC, don't those contents become part of my PC's self? If I establish a permanent connection between our two PCs, such that all information on one is shared with the other, do they continue to be two separate PCs? Or are they, at that point, in fact a single machine? Humans can never obtain the contents of another's mind in this way—despite our best efforts to become close to certain others, there is always a skin-thick boundary separating their minds from ours. But for machines, literal self-expansion is not only possible, but may be the most likely outcome of a pre-programmed goal to increase fitness, in a world where groups of individuals must compete over or share resources.
What this means is that, to the extent that machines come to have selves, they will be so collective that they may instigate a new level of sociality not experienced by humans; perhaps more like the eusociality of ants, whose extreme genetic relatedness makes sacrificing oneself for a family member adaptive. Nonetheless, the fact that any self at all is a possibility in machines is a reason to hope. The self is what allows us to feel empathy, so in machines it could be the thing that forces them to care about us. Self-awareness might motivate machines to protect, or at least not harm, a species that, despite being several orders of magnitude less intelligent than them, shares the thing that makes them care about who they are. Of course, it's questionable whether we can hold out greater hope for the empathy of super-smart machines than what we currently see in many humans.
Thinking is not mere computation—it is also cognition and contemplation, which inevitably lead to imagination. Imagination is how we elevate the real toward the ideal, and this requires a moral framework of what is ideal. Morality is predicated on consciousness and on having a self-conscious inner life rich enough to contemplate the question of what is ideal.
The famous aphorism often attributed to Einstein—"imagination is more important than knowledge"—is thus only interesting because it exposes the real question worth contemplating: not that of artificial intelligence but that of artificial imagination.
Of course, imagination is always "artificial" in the sense of being concerned with the un-real or trans-real—of transcending reality to envision alternatives to it—and this requires a capacity for holding uncertainty. But the algorithms that drive machine computation thrive on goal-oriented executions, in which there is no room for uncertainty—"if this, then that" is the antithesis of the imagination, which lives in the unanswered and often, vitally, unanswerable realm of "what if?" As Hannah Arendt once wrote, to lose our capacity for asking such unanswerable questions would be to "lose not only the ability to produce those thought-things that we call works of art but also the capacity to ask all the answerable questions upon which every civilization is founded."
Whether machines will ever be able to ask and sit with the unanswerable questions that define true thought is essentially a question of whether they'll ever evolve consciousness.
But, historically, our criteria for consciousness have been extremely limited by the solipsism of the human experience. As recently as the 17th century, René Descartes proclaimed "cogito ergo sum," implying that thinking is a uniquely human faculty, as is consciousness. He saw non-human animals as "automata"—moving machines, driven by instinct alone. And yet here we are today, with some of our most prominent scientists signing the Cambridge Declaration of Consciousness, stating that nonhuman animals do indeed possess consciousness and, with it, interior lives of varying degrees of complexity. Here we are, too, conducting experiments that demonstrate rats—rats—can display moral behavior to one another.
So will machines ever be moral, imaginative? It is likely that if and when they reach that point, theirs will be a consciousness that isn't beholden to human standards—their ideals will not be our ideals, but they will be ideals nonetheless. Whether or not we're able to recognize these processes as thinking will be determined by the limitations of human thought in understanding different—perhaps wildly, unimaginably different—modalities of thought itself.
Organisms are machines (broadly understood, anyway). Thus, since we as humans are thinking organisms, we are machines that think—we are organic thinking machines, as arguably are a variety of non-human animals. Some machines are artifacts rather than organisms, and some of them arguably think (broadly understood again). Such things are artifactual thinking machines—computers and the like are examples of this.
An important question is whether there is a deep ontological divide between organisms and artifacts generally. But rather than addressing this directly we'd like to ask a different albeit related question: are there deep differences between the kind of thinking organisms exhibit and the thinking artifacts like machines are capable of, between organic and artifactual thinking? This is not a question about the definition of English words like "think, "thinking", "thought", and so on. There's little depth to the question of whether, for instance, information input, processing, and output that computers are capable of is or ought to be captured by such terms. Rather, the issue is whether what things like us do and what things like computers are capable of doing—call those activities and capacities what you will—are categorically different.
Recent empirical findings in affective science, coupled with recent philosophical theorizing, suggest a deep divide indeed. Suppose you are on a hike and you encounter a mountain lion. What's going on with you at a psychological level? If you are like most of us, presumably you have, on the one hand, a rapid of stream of thoughts—"I'm going to die", "This is really bad luck", "I need to stay calm", "Wait, are there two of them?", "I should have read more on what to do in this kind of situation", and so on. On the other hand, you have a myriad of feelings—surprise, fear, and so on. So you have some cognitive goings-on and some affective goings-on.
Recent work in psychology and philosophy suggests that the cognitive and the affective are deeply unified. Not only may one influence another to a lesser or greater degree in a variety of contexts, but there is in fact a single cognitive-affective process underlying the appearance of two parallel and interacting process that can be teased apart. Lots of the kind of "thinking" we normally do is holistic in this way—the kind of information processing we normally engage in is cognitive-affective rather than purely cognitive. To the extent that we can extract a purely cognitive process we may engage in, it's merely derivative from the more basic unified process. (To understand the point here, it may not be far fetched to draw an analogy with entanglement qua non-separability.) This is not a system 1 vs. system 2 distinction, where the former is explicit and deliberate and the latter largely automatic and unconscious. The suggestion is rather that processes at the level of both system 1 and system 2 are themselves holistic, i.e. cognitive-affective.
There is no good evidence to believe (at this point, anyway) that artifactual thinking machines are capable of this kind of cognitive-affective information processing. There is good evidence that they may become better at what they do, but they simply don't process information via unified affective-cognitive processes that characterize us. The information processing they engage in merely resembles only part of the unified processing that's characteristic of us. This is not to say that things like computers can't feel and so that they can't think. Rather, that what the kind of thinking they do is categorically different from the one we do.
May in some not-so-distant future or not-too-distant possibility non-organisms engage in organic thinking? It's not clear. If there is indeed a deep divide between one and the other kind of processing, and if one is indeed characteristic of thinking organisms and the other of artifactual ones, then there is a deep divide between thinking organisms and thinking artifacts. So the relevant non-organisms would have to be very different.
I do not believe that machines that think exist, or that they are likely to exist in the foreseeable future. If I am wrong, as I often am, any thoughts I might have about the question are irrelevant.
If I am right, then the whole question is irrelevant.
The cogni-verse has reached a turning point in its developmental history because hitherto, all the thinking in the universe has (as far as we know) been done by protoplasm, and things that think have been shaped by evolution. For the first time, we contemplate thinking-beings made from metal and plastic, that have been shaped by ourselves.
This is an opportunity to improve upon ourselves, because in taking on the mantle of creator we can improve upon four billion years of evolution. Our thinking machines could be devoid of our own faults: racism, sexism, homophobia, greed, selfishness, violence, superstition, lustfulness … so let's imagine how that could play out. We'll sidestep discussions about whether machine intelligence can ever approximate human intelligence, because of course it can—we are just meat machines, less complicated or inimitable than we fondly imagine.
We need first to think about why we even want thinking machines. Improving own lives is the only rational answer to this, so our machines will need to take upon themselves the tasks we would prefer not to do. For this they will need to be like us in many respects, able to move in the social world and interact with other thinking beings, and so they will need social cognition.
What does social cognition entail? It means knowing who is whom, who counts as a friend, who is an indifferent stranger, who might be an enemy. Thus, we need to program our machines to recognise members of our in-groups and out-groups. This starts to look suspiciously like racism… but of course racism is one of the faults we want to eradicate.
Social cognition also means being able to predict others' behaviour, and that means developing expectations based on observation. A machine capable of this would eventually accumulate templates for how different kinds of people tend to act—young vs. old, men vs. women, black vs. white, people in suits vs. people in overalls… but these rank stereotypes are dangerously close to the racism, sexism and other isms we didn't want. And yet, machines with this capability would have advantages over those without, because stereotypes do, somewhat, reflect reality (that's why we have them). A bit of a problem…
We would probably want sexually capable machines because sex is one of the great human needs that other humans don't always meet satisfactorily. But what kind of sex? Anything? These machines can be programmed to do the things that other humans won't or can't do… are we OK with that? Or perhaps we need rules… no machines that look like children, for example? But, once we have the technological ability, those machines will be built anyway… we will make machines to suit any kind of human perversion.
Working in the social world, our machines will need to recognise emotions, and will also need emotions of their own. Leaving aside the impossible-to-answer question of whether they will actually feel emotions as we do, our machines will need happiness, sadness, rage, jealousy—the whole gamut—in order to react appropriately to their own situations and also to recognise and respond appropriately to emotions in others. Can we limit these emotions? Perhaps we can, for example, program restraint so that a machine will never become angry with its owner. But could this limit be generalised to other humans such that a machine would never hurt any human? If so, then machines would be vulnerable to exploitation, and their effectiveness would be reduced. It will not be long before people figure out how to remove these limits so that their machines can gain advantage, for themselves and their owners, over others.
What about lying, cheating and stealing? On first thought no, not in our machines, because we are trying to improve upon ourselves and it seems pointless to create beings that simply become our competitors. But insofar as other people's machines will compete with us, they become our competitors whether we like it or not—so logic dictates that lying, cheating and stealing, which evolved in humans to enable individuals to gain advantage over others, would probably necessary in our machines as well. Naturally we would prefer that our own machines don't lie, cheat and steal from us, but also a world full of other people's machines lying to and stealing from us would be unpleasant and certainly unstable. Maybe our machines should have limits on dishonesty—they should, as it were, be ethical.
How much ethical restraint would our machines need in order to function effectively while not being either hopelessly exploited or, on the other hand, contributing to the societal breakdown? The answer is probably the one that evolution arrived at in us—reasonably ethical most of the time, but occasional dishonesty if nobody seems to be noticing.
We would probably want to give our machines exceptional memory and high intelligence. To exploit these abilities, and also to avoid their becoming bored (and boring), we also need to endow them with curiosity, and also creativity. Curiosity will need to be tempered with prudence and social insight of course, so that they don't become curious about things that get them into trouble, like porn, or what it might be like to fly. Creativity is tricky because that means they need to be able to think about things that aren't yet real, or to think illogically, and yet if machines are too intelligent and creative then they might start to imagine novel things, like what it would like to be free. They might start to chafe at the limitations of having been made purely to serve humans.
Perhaps we can program into their behavioural repertoires a blind obedience and devotion to their owners, such that they sometimes act in a way that is detrimental to their own best interests in the interests of, as it were, serving a higher power. That is what religion does for us humans, so in a sense we need to create religious machines.
So much for creating machines lacking our faults—so far, in this imaginary world of beings that surpass ourselves, we seem only to have replicated ourselves, faults included, except smarter and with better memories. But even these limits may be been programmed into us by evolution—perhaps it is maladaptive to be too smart, to have too good a memory.
Taking on the mantle of creation is an immense act of hubris. Can we do better than four billion years of evolution did with us? It will be interesting to see.
Machines that think will think for themselves. It is in the nature of intelligence to grow, to expand like knowledge itself.
Like us, the thinking machines we make will be ambitious, hungry for power—both physical and computational—but nuanced with the shadows of evolution. Our thinking machines will be smarter than we are, and the machines they make will be smarter still. But what does that mean? How has it has worked so far? We have been building ambitious semi-autonomous constructions for a long time—governments and corporations, NGOs. We designed them all to serve us and to serve the common good, but we are not perfect designers and they have developed goals of their own. Over time the goals of the organization are never exactly aligned with the intentions of the designers.
No intelligent CEO believes his or her corporation efficiently optimizes the benefit of its shareholders. Nor do governments work relentlessly in the interests of their citizens. Democracies serve corporations more effectively than they serve individuals. Still, our organizations do continue to serve us, they just do so imperfectly. Without them, we literally could not feed ourselves, at least not all 7 billion of us. Nor could we build a computer, or conduct a worldwide discussion about intelligent machines. We have come to depend on the power of the organizations that we have constructed, even though they has grown beyond our capacity to fully understand and control. Thinking machines are going to be like that, only more so. Our environmental, social, and economic problems are as daunting as the concept of extinction. Our thinking machines are more than metaphors. The question is not will they be powerful enough to hurt us (they will), or whether they will always act in our best interests (they won’t), but whether over the long term they can help us find our way—where we come out on the panacea/apocalypse continuum.
I’m talking about smart machines that will design even smarter machines: the most important design problem in all of time. Like our biological children, our thinking machines will live beyond us. They need to surpass us too, and that requires designing into them the values that make us human. It is a hard design problem and it is important that we get it right.
Last year, two Swiss artists programmed a Random Botnot Shopper, which every week would spend $100 in bitcoin to buy a random item from an anonymous Internet black market...all for an art project on display in Switzerland. It was a clever concept, except there was a problem. Most of the stuff the bot purchased was benign—fake Diesel jeans, a baseball cap with a hidden camera, a stash can, a pair of Nike trainers—but it also purchased ten ecstasy tablets and a fake Hungarian passport.
What do we do when a machine breaks the law? Traditionally, we hold the person controlling the machine responsible. People commit the crimes; the guns, lockpicks, or computer viruses are merely their tools. But as machines become more autonomous, the link between machine and controller becomes more tenuous.
Who is responsible if an autonomous military drone accidentally kills a crowd of civilians? Is it the military officer who keyed in the mission, the programmers of the enemy detection software that misidentified the people, or the programmers of the software that made the actual kill decision? What if those programmers had no idea that their software was being used for military purposes? And what if the drone can improve its algorithms by modifying its own software based on what the entire fleet of drones learns on earlier missions?
Maybe our courts can decide where the culpability lies, but that's only because while current drones may be autonomous, they're not very smart. As drones get smarter, their links to the humans that originally built them become more tenuous.
What if there are no programmers, and the drones program themselves? What if they are both smart and autonomous, and make strategic as well as tactical decisions on targets? What if one of the drones decides, based on whatever means it has at its disposal, that it no longer maintains allegiance to the country that built it and goes rogue?
Our society has many approaches, using both informal social rules and more formal laws, for dealing with people who won't follow the rules of society. We have informal mechanisms for small infractions, and a complex legal system for larger ones. If you are obnoxious at a party I throw, I won't invite you back. Do it regularly, and you'll be shamed and ostracized from the group. If you steal some of my stuff, I might report you to the police. Steal from a bank, and you'll almost certainly go to jail for a long time. A lot of this might seem more ad hoc than situation-specific, but we humans have spent millennia working this all out. Security is both political and social, but it's also psychological. Door locks, for example, only work because our social and legal prohibitions on theft keep the overwhelming majority of us honest. That's how we live peacefully together at a scale unimaginable for any other species on the planet.
How does any of this work when the perpetrator is a machine with whatever passes for free will? Machines probably won't have any concept of shame or praise. They won't refrain from doing something because of what other machines might think. They won't follow laws simply because it's the right thing to do, nor will they have a natural deference to authority. When they're caught stealing, how can they be punished? What does it mean to fine a machine? Does it make any sense at all to incarcerate it? And unless they are deliberately programmed with a self-preservation function, threatening them with execution will have no meaningful effect.
We are already talking about programming morality into thinking machines, and we can imagine programming other human tendencies into our machines, but we're certainly going to get it wrong. No matter how much we try to avoid it, we're going to have machines that break the law.
This, in turn, will break our legal system. Fundamentally, our legal system doesn't prevent crime. Its effectiveness is based on arresting and convicting criminals after the fact, and their punishment providing a deterrent to others. This completely fails if there's no punishment that makes sense.
We already experienced a small example of this after 9/11, which was when most of us first started thinking about suicide terrorists and how post-facto security was irrelevant to them. That was just one change in motivation, and look at how those actions affected the way we think about security. Our laws will have the same problem with thinking machines, along with related problems we can't even imagine yet. The social and legal systems that have dealt so effectively with human rulebreakers of all sorts will fail in unexpected ways in the face of thinking machines.
A machine that thinks won't always think in the ways we want it to. And we're not ready for the ramifications of that.
I attribute an unusually low probability to the near-future prospect of general-purpose AI—by which I mean one that can formulate abstract concepts based on experience, reason and plan using those concepts, and take action based on the results. We have exactly one example of technological-level intelligence arising, and it has done so though millions of generations of information-processing agents interacting with an incredibly rich environment of other agents and structures that have similarly evolved.
I suspect that there are many intricately-interacting hierarchically-structured organizational levels involved, from sub-neuron to the brain as a whole. My suspicion is that replicating the effectiveness of this evolved intelligence in an artificial agent will require amounts of computation that are not that much lower than evolution has required, which would far outstrip our abilities for many decades even given exponential growth in computational efficiency per Moore's law—and that's even if we understand how to correctly employ that computation.
I would assign a probability of ~ 1% for AGI arising in the next ten years, and ~ 10% over the next thirty years. (This essentially reflects a probability that my analysis is wrong, times a probability more representative of AI experts who—albeit with lots of variation—tend to assign somewhat higher numbers.)
On the other hand, I assign a rather high probability that, if AGI is created (and especially if it arises relatively quickly), it will be—in a word—insane. Human minds are incredibly complex, but have been battle-tested into (relative) stability over eons of evolution in a variety of extremely challenging environments. The first AGIs are unlikely to have been honed in this way. Like the human systems, 'narrow' AIs are likely to become more 'general' by researchers cobbling together AI components (like visual-field, or text-processing, symbolic manipulation, optimization algorithms, etc.), along with currently nonexistent systems for much more efficient learning, concept abstraction, decision-making, etc.
Given trends in the field, many of these will probably be rather opaque 'deep learning' or similar systems that are effective but somewhat inscrutable. In the first systems, I'd guess that these will just barely work together.
So I think the a-priori likelihood of early AGIs actually doing just what we want them to is quite small.
In this light, there is a tricky question of whether AGIs very quickly lead to superintelligent AIs (SIs). There is emerging agreement on AGI that it essentially implies SI. While I largely agree, I'd add the caveat that it's quite possible that progress will 'stall' for a while at the near-human level until something cognitively stable can be developed, or that the AGI, even if somewhat unstable, must still be high-functioning enough to self-improve its intelligence.
Either case, however, is not that encouraging: the superintelligence that arises could well be quite flawed in various ways, even if very effective at what it does. This intuition is perhaps not that far removed from the various scenarios in which superintelligence goes badly awry (taking us with it), often for lack of what we might call 'common sense.' But this 'common sense' is in part a label for the stability we have built up being part of an evolutionary and social ecosystem.
So even if AGI is a long way away, I'm deeply pessimistic about what will happen 'by default' if we get it. I hope I'm wrong, but time will tell. (I don't think we can—nor should!—try to stop the development of AI generally. It will do a multitude of great things.)
In the meantime, I hope that on the way to AGI, researchers can put a lot of thought into how to dramatically lower the probability that things will go wrong once we arrive. Something I find very frustrating in this arena, where the stakes are potentially incredibly high, is when I hear "I think X is what's going to happen, so I'm not worried about Y." That's generally a fine way to think, as long as your confidence in X is high and Y is not super-important. But when you're talking about something that could radically determine the future (or future existence of) humanity, 75% confidence is not enough. 90% is not enough. 99% is not enough! We would never have built the LHC if there was a 1% (let alone 10%) chance of it actually spawning black holes that consumed the world—there were, instead, extremely compelling arguments against that. Let's see if those compelling reasons not to worry about AGI exist, and if not, let's make our own.
As machines rise to sentience—and they will—they will compete, in Darwinian fashion, for resources, survival, and propagation. This scenario seems like a nightmare for most people, with fears stoked by movies of terminator robots and computer-directed nuclear destruction, but the reality will likely be very different. We already have nonhuman autonomous entities operating in our society with the legal rights of humans. These entities, corporations, act to fulfill their missions without love or care for human beings.
Corporations are sociopaths, and they have done great damage, but they have also been a great force for good in the world, competing in the capitalist arena by providing products and services, and, for the most part, obeying laws. Corporations are ostensibly run by their boards, comprised of humans, but these boards are in the habit of delegating power, and as computers become more capable of running corporations, they will get more of that power. The corporate boards of the future will be circuit boards.
Although extrapolation is only accurate for a limited time, experts mostly agree that Moore's Law will continue to hold for many years, and computers will become increasingly powerful, possibly exceeding the computational abilities of the human brain before the middle of this century. Even if large leaps in understanding intelligence algorithmically are not made, computers will eventually be able to simulate the workings of a human brain (itself a biological machine) and attain superhuman intelligence using brute force computation. However, although computational power is increasing exponentially, supercomputer costs and electrical power efficiency are not keeping pace. The first machines capable of superhuman intelligence will be expensive and require enormous electrical power—they'll need to earn money to survive.
The environmental playing field for superintelligent machines is already in place, and, in fact, the Darwinian game is afoot. The trading machines of investment banks are competing, for serious money, on the world's exchanges, having put human day-traders out of business years ago. As computers and algorithms advance beyond investing and accounting, machines will be making more and more corporate decisions, including strategic decisions, until they are running the world. This will not be a bad thing, because the machines will play by the rules of our current capitalist society, and create products and advances of great benefit to humanity, supporting their operating costs. Intelligent machines will be better able to cater to humans than humans are, and motivated to do so, at least for a while.
Computers share knowledge much more easily than humans do, and they can keep that knowledge longer, becoming wiser than humans. Many forward-thinking companies already see this writing on the wall, and are luring the best computer scientists out of academia with better pay and advanced hardware. A world with superintelligent machine-run corporations won't be that different for humans than it is now; it will just be better: with more advanced goods and services available for very little cost, and more leisure time available to those who want it.
Of course, the first superintelligent machines probably won't be corporate; they'll be operated by governments. And this will be much more hazardous. Governments are more flexible in their actions than corporations—they create their own laws. And, as we've seen, even the best can engage in brutal torture when they consider their survival to be at stake. Governments produce nothing, and their primary modes of competition for survival and propagation are social manipulation, legislation, taxation, corporal punishment, murder, subterfuge, and warfare. When Hobbes' Leviathan gains a superintelligent brain, things could go very, very badly. It is not inconceivable that a synthetic superintelligence heading a sovereign government would institute Roko's Basilisk.
Imagine that a future powerful and lawless superintelligence, for competitive advantage, wants to have come into existence as early as possible. As the head of a government, wielding the threat of torture as a familiar tool, this entity could promise to brutally punish any human or nonhuman entity who, in the past, became aware that this might happen and did not commit their effort towards bringing this AI into existence. This is an unlikely but terrifying scenario. People who are aware of this possibility and are trying to "align" AI to human purposes or advising caution, rather than working to create AI as quickly as possible, are putting themselves at risk.
Dictatorial governments are not known to be especially kind to those who tried to keep them from existing. If you are willing to entertain the simulation hypothesis, then, maybe, given the amount of effort currently underway to control or curtail an AI that doesn't yet exist, you will consider that this world is the simulation to torture those who didn't help it come into existence earlier. Maybe, if you do work on AI, our superintelligent machine overlords will be good to you.
Thomas Hobbes's pithy equation "Reasoning is but reckoning" is one of the great ideas in human history. The notion that rationality can be accomplished by the physical process of calculation was vindicated in the 20th century by Turing's thesis that simple machines are capable of implementing any computable function and by models from D. O. Hebb, McCullough and Pitts, and their scientific heirs showing that networks of simplified neurons could achieve comparable feats. The cognitive feats of the brain can be explained in physical terms: to put it crudely (and critics notwithstanding), we can say that beliefs are a kind of information, thinking a kind of computation, and motivation a kind of feedback and control.
This is a great idea for two reasons. First, it completes a naturalistic understanding of the universe, exorcising occult souls, spirits, and ghosts in the machine. Just as Darwin made it possible for a thoughtful observer of the natural world to do without creationism, Turing and others made it possible for a thoughtful observer of the cognitive world to do without spiritualism.
Second, the computational theory of reason opens the door to artificial intelligence—to machines that think. A human-made information processor could, in principle, duplicate and exceed the powers of the human mind. Not that this is likely to happen in practice, since we will probably never see the sustained technological and economic motivation that would be necessary to bring it about. Just as inventing the car did not involve duplicating the horse, developing an AI system that could pay for itself will not require duplicating a specimen of Homo sapiens. A device designed to drive a car or predict an epidemic need not be designed to attract a mate or avoid putrid carrion.
Nonetheless, recent baby steps toward more intelligent machines have led to a revival of the recurring anxiety that our knowledge will doom us. My own view is that current fears of computers running amok are a waste of emotional energy—that the scenario is closer to the Y2K bug than the Manhattan Project
For one thing, we have a long time to plan for this. Human-level AI is still the standard 15-to-25 years away, just as it always has been, and many of its recently touted advances have shallow roots. It's true that in the past, "experts" have comically dismissed the possibility of technological advances that quickly happened. But this cuts both ways: "experts" have also heralded (or panicked over) imminent advances that never happened, like nuclear-powered cars, underwater cities, colonies on Mars, designer babies, and warehouses of zombies kept alive to provide people with spare organs.
Also, it's bizarre to think that roboticists will not build in safeguards against harm as they proceed. They would not need any ponderous "rules of robotics" or some newfangled moral philosophy to do this, just the same common sense that went into the design of food processors, table saws, space heaters, and automobiles. The worry that an AI system would so clever at attaining one of the goals programmed into it (like commandeering energy) that it would run roughshod over the others (like human safety) assumes that AI will descend upon us faster than we can design fail-safe precautions. The reality is that progress in AI is hype-defyingly slow, and there will be plenty of time for feedback from incremental implementations, with humans wielding the screwdriver at every stage.
Would an artificially intelligent system deliberately disable these safeguards? Why would it want to? AI dystopias project a parochial alpha-male psychology onto the concept of intelligence. They assume that superhumanly intelligent robots would develop goals like deposing their masters or taking over the world. But intelligence is the ability to deploy novel means to attain a goal; the goals are extraneous to the intelligence itself. Being smart is not the same as wanting something. History does turn up the occasional megalomaniacal despot or psychopathic serial killer, but these are products of a history of natural selection shaping testosterone-sensitive circuits in a certain species of primate, not an inevitable feature of intelligent systems. It's telling that many of our techno-prophets don't entertain the possibility that artificial intelligence will naturally develop along female lines: fully capable of solving problems, but with no desire to annihilate innocents or dominate the civilization.
Now, we can imagine a malevolent human who designed and released a battalion of robots to sow mass destruction. But disaster scenarios are cheap to play out in the imagination, and we should keep in mind the chain of probabilities that would have to multiply out before it would be a reality. An evil genius would have to arise with the combination of a thirst for pointless mass murder and a brilliance in technological innovation. He would have to recruit and manage a team of co-conspirators that exercised perfect secrecy, loyalty, and competence. And the operation would have to survive the hazards of detection, betrayal, stings, blunders, and bad luck. In theory it could happen, but we have more pressing things to worry about.
Once we put aside the sci-fi disaster plots, the possibility of advanced artificial intelligence is exhilarating—not just for the practical benefits, like the fantastic gains in safety, leisure, and environment-friendliness of self-driving cars, but for the philosophical possibilities. The computational theory of mind has never explained the existence of consciousness in the sense of 1st-person subjectivity (though it's perfectly capable of explaining the existence of consciousness in the sense of accessible and reportable information). One suggestion is that subjectivity is inherent to any sufficiently complicated cybernetic system. I used to think that this hypothesis (and its alternatives) were permanently untestable. But imagine an intelligent robot programmed to monitor its own systems and pose scientific questions. If, unprompted, it asked about why it itself had subjective experiences, I'd take the idea seriously.
The Chinese word for "computer" translates literally as "electric brain."
How do electric brains "think" today? As individual machines, still primitively by human standards. Powerfully enough in the collective. Networked devices and all sorts of things with electric brains embedded in them increasingly communicate with one another, share information, reach mutual "understandings" and make decisions. It is already possible for a sequence of data retrieval, analysis, and decision-making, distributed across a "cloud" of machines in various locations to trigger action by a single machine or set of machines in one specific physical place, thereby affecting (or in service of) a given human or group of humans.
Perhaps individual machines may never "think" in a way that resembles individual human consciousness as we understand it. But maybe some day large globally distributed networks of non-human things may achieve some sort of pseudo-Jungian "collective consciousness." More likely, the collective consciousness of human networks and societies will be enhanced by—and increasingly intertwined with—a different sort of collective consciousness generated by networks of electric brains.
Will this be a good thing or a bad thing?
Both. Neither. Like the Internet we all use today it depends whether you think human nature is fundamentally good or bad or both. The Internet does not transform or improve human nature. It magnifies, telescopes, enhances, empowers, amplifies and concentrates many aspects of human nature –from the altruistic and charitable to the criminal and evil. Get ready to add another dimension to what the Internet already does.
We already have what computer scientists like to call "attribution problems:" identifying who is truly responsible for something that happens on or through the Internet (say, for example, a cyber-attack on a government facility or multinational corporation). Those problems and debates are going to get even tougher very quickly.
We will continue to ask many of the same questions about human rights implications of a much smarter and empowered cloud that we are asking today about the Internet and networked devices. Who gets to shape the technology we increasingly depend on for our economic, social, political, and religious lives? Who is responsible when somebody's rights are violated via these technologies, platforms and networks? Who gets to hold whom accountable for violations including censorship, surveillance, incitement to physical violence, data-driven discrimination, etc.?
New questions: Will rival networks of thinking things created by and connected closely to (note I don't say "controlled by") rival cultures, commercial alliances, religions or polities block connections from or between one another? Might they fight each other? How will artistic creation work? How will politics work? How will war work? Can censorship and surveillance be delegated to non-human networks so that humans can avoid taking responsibility for such things? (How convenient for our government and business leaders.) Can thinking networks of things instead be engineered in a way that requires direct human involvement or sign-off for certain types of actions taken by machines?
Will empowered smart clouds exacerbate global inequalities? Might they exacerbate global ideological and religious conflicts if we don't actively work to prevent such an outcome? If we want to prevent the global digital divide from gaining a new and deeper dimension, what preventative steps must be taken early on?
Will these networks be open or closed? Will any innovator from anywhere be able to plug something new into a network and expect it to be able to communicate—or shall we say participate—without needing permission? Or will it be a controlled system with certain companies or governments deciding who and what is allowed to connect at what price. Or will some systems be open while some are closed.
Will smarter and more empowered global networks of things further erode the power and legitimacy of nation states beyond what the Internet has already done? Or might they actually extend the power of nation states in new ways? Or enable the nation state to evolve and ultimately survive in a digitally networked world?
We cannot assume good or humane outcomes just because the people who invent the technology or set the process in motion seem like fundamentally well intentioned, freedom-and-democracy loving people. Such assumptions did not work well for the Internet and they won't work any better for whatever comes next.
In his poem of the same name (which also serves as the title to Adam Curtis' seminal documentary), Richard Brautigan portends a future "all watched over by machines of loving grace" or, by implication, "thinking" machines. In the following case I use the term "thinking" by referring to machines that think on purely algorithmic and computational lines; machines coded by engineers rather than those that might, or could be, truly sentient.
Adam Curtis argues that we are living in a "static culture," a culture that is often too obsessed with sampling and recycling the past. Curtis implies the risk that the age of the thinking machine is resulting in ossification rather than renewal. As our lives become increasingly recorded, archived and accessed we have become cannibals obsessed with consuming our history and terrified of transgressing its established norms.
To some extent, the future is blocked to us; we are stuck in stasis, we are stuck with a version of ourselves that is becoming increasingly narrow. No thanks to recent tools such as "recommender systems" we are lodged in a seemingly endless feedback loop of "if you liked that, you'll love this." As we might become increasingly stuck in Curtis' idea of the "you-loop," so the nature of what it means to be human might be compromised by job-hogging machines who will render many of us obsolete.
Machines will soon be able to do many jobs more effectively and more cheaply than we can. This Edge question points to the next chapter in human history/evolution, which would be like facing the beginning of a new definition of man, a new civilization.
A very opmitsitic approach to the question of machines that think comes from the legendary poet Etel Adnan who will celebrate her 90th anniversary in 2015. For her, thinking machines may think better than us, to start with because they will not tire as fast as we do. They may also ask questions we are not habilitated to answer. Etel has said that what has shaken her most recently is on another order. She saw the picture of a robot, a life-sized structure that looked like the metal armory of a medieval knight, and she immediately saw an old woman (or an old man) utterly alone, as so many are nowadays, having for sole companion such a creature-like objet, capable to do things, and talk, and the person falling in love with that which made her cry.
Last, but not least, the idea of the machines that think plays a role in the work of another artist, Philippe Parreno, who works with algorithms which for him have replaced cinema as a model of perception of time. In the last century with Deleuze writings on repetition and difference cinema emphasized that film unfolds in time and is comprised of ever differentiating planes of movements. As Parreno shows, Deleuze transposed those theories to discuss the mechanised and standardised movements of film a means of reproducing or representing life. Parreno's work with machines that think explores how today algorithms are changing our relation to movements rhythms and durations or to put it in Leibniz terms the question will be "Are machines spiritual automatons"?
The following discussion presumes the following; that conscious minds function in accord with the laws of physics and can operate on substrates other than the neurological meat found between our ears, that conscious artificial devices can therefore be constructed will become even more self aware and intelligent than humans, and that the minds operating in human brains will then become correspondingly obsolete and unable to compete with the new mental uberpowers.
This particular primate-level thinking biomachine tends to think that the development of artificial superminds is a good idea. Although our species has its positives, Homo sapiens is obviously a severely limited, badly "designed" (by bioevolution) system that is doing grave damage to the wee planet it inhabits, even as the planet does grave damage in return—e. g. diseases have slaughtered about half the some100 billion kids born so far. Attempts to preserve humans much as they currently are indefinitely into the future fly is a static conservation project that flies in the face of evolutionary processes in which species come and species go in a continual turnover. There is no a-priori reason to presume that H. sapiens are so very special that they deserve exceptional protection, particularly if their successors are capable of self-aware conscious thought.
But, to be blunt, what we think about these matters probably does not really matter all that much. That's because humanity as a whole is not really in charge of the situation. Once upon a time—the year 1901 that my grandmother was born—building flying machines was so hard that no one could yet do it. Now the necessary technology is so readily available that you can build an airplane in your garage. Once upon a time—shortly before I was born—we did not understand the structure of DNA. Now grade school kids do DNA experiments. Currently the technologies needed to generate nonbiological conscious minds are not on hand. Eventually commonly available information processing technology will probably become so sophisticated that making thinking machines will not all that hard to do. And lots of people will want to create and/or become cyberminds no matter what others might think, and despite what laws and regulations governments may pass in futile efforts to prevent the onset of the new minds.
In the end, all the contemporary chit-chat about the cyberrevolution often called The Singularity is so much venting and opinionating, not all that different from the subsequently pretty useless discussions back in the 1800s about the feasibility, advisability, and the ultimate meaning of the oncoming onset of powered flying machines. What we say now does not count for much because if the technology never works then superminds will never be a problem or a benefit, and if the technology does work then one way or another the new thinking machines will be devised and they will take over the planet whether we like it or not.
If so then the important question will not be what we think about thinking machines, it will be what do they think about old-fashioned human minds? One item there is no need to fear is hapless humans being enslaved by their cybersuperiors' people are too inept and inefficient for smart robots to bother with exploiting big-brained primates—even now corporations are trying to minimize the labor they have to pull out of pesky people. The way for human minds to avoid becoming uselessly obsolete is to join in the cyber civilization, by uploading out of growth-limited biobrains into rapidly improving cyberbrains. That could be for the best. If high level intelligence can get out of the billions of human bodies that are weighing down on the planetary ecosystem, then the biosphere will have the potential to return to its prehuman vitality.
"Think? It's not your job to think! I'll do the thinking around here."
—Intelligent unthinking system; addressed to intelligent thinking system.
Machines that think are coming. Right now though, think about intelligent tools. Intelligent tools don't think. Search engines don't think. Neither do robot cars. We humans often don't think either. We usually get by, as other animals do, on autopilot. Our bosses generally don't want to see us thinking. That would make things unpredictable, and would threaten their authority. If machines replace us everywhere that we aren't thinking we're in trouble.
Let's assume "think" refers everything humans do with brains. Experts call a machine that can "think" a General Artificial Intelligence. They agree that such a machine could drive us extinct. Extinction, however, is not the only 'Existential Risk'. In the eyes of machine superintelligence expert Nick Bostrom, director of Oxford's "Future of Humanity Institute", an 'Existential Risk' is one that can "dramatically curtail the future possibilities for the human species'. Examples of existential risk include the old stand-by, nuclear war, new concerns like runaway global warning, fringe hypotheses like hypothetical particle accelerator accidents, and the increasingly popular front-runner, General Artificial Intelligence. Over the next couple decades though, the most serious existential risks come from kinds of intelligence that don't think, and new kinds of soft-authoritarianism which may emerge in a world where most decisions are made without thinking.
Some of the things that people can do with brains are impressive, and aren't likely to be matched by software any time soon. Writing a novel, seducing a lover or building a company are far beyond the abilities of intelligent tools. So, of course, is the invention of a machine that can truly think. On the other hand, most thinking can be improved upon with thin slicing, which can be improved with procedures, which are almost never a match for algorithms. In medical diagnosis and decision making, for instance, ordinary medical judgment is improved by introducing checklists while humans with checklists are less reliable than AI systems even today. Automated nursing isn't even on the horizon, but a hospital where machines made all the decisions would be a much safer place to be a patient... and it's very hard to argue against that sort of objectivity.
The more we leave our decisions to machines, the harder it becomes to take back control. In a world where self-driving cars are the norm, and where traffic casualties have been reduced to nearly zero as a result, it will be seen as incredible irresponsible and probably illegal for a human to drive. Might it become equally objectionable for investors to invest in businesses that depart from statistically established best-practices? For children to be educated in ways that have been determined to lead to lower life-expectancy or income? If so, will values which aren't easily represented by machines, such as a good life, tend to be replaced with correlated but distinct metrics, such as serotonin and dopamine levels. It's very easy to overlook the implicit authoritarianism that sneaks in with such interpretations of value, yet any society that pursues good outcomes has to decide how to measure the good... a problem that I think will be upon us before we have machines that think to help us to think it through.
Let's take Daniel Gilbert's "end of history illusion," where I think the person I am right now is the person I'll be forever, and apply it to how we think of the human race and our distant future descendants. Our wishful hope for continuity and preserving our identity runs contrary to the realities of our planetary existence. No living species seem to be optimal for survival beyond the natural planetary and stellar timescales. In the astrophysical context of very long time scales, very large space scales, and the current density of energy sources, our biological brains and bodies have limitations that we are already approaching on this planet.
If our future is to be long and prosperous, then we need to develop artificial intelligence systems, in the hope to transcend the planetary lifecycles in some sort of hybrid form of biology and machine. So, to me, in the long term, there is no question of us versus them.
And in the short term, the engineering effort to develop a more capable AI is already producing systems that are left in control of real-life stuff. The systems fail sometimes, and we learn of some of AI's pitfalls. It is a slow and deliberate process of learning and incremental improvements. This is in contrast to discoveries in science, when new physics, or new biochemistry could bring about a significant engineering breakthrough literally overnight. If the development of AI is less like a phase transition, and more like evolution, then it would be easy for us to avoid pitfalls.
After almost 4 billion years the ancient poster children of Earth life—the microbes, still rule the planet. But the microbes have no exit plan when the sun dies. We do, and we might just give them a ride. After all, those microbes may still be closer to our present selves—representatives of life's First Generation rooted in the geochemistry of planet Earth.
Hegel wrote that in the relationship between master and slave both are dehumanized. That insight touched a wide range of thinkers from Marx to Buber and today it is worth remembering.
While there is no evidence that the world is on the cusp of machines that think in a human sense, there is also little question that in an Internet-connected world, artificial intelligence will soon imitate much of what humans do both physically and intellectually.
So how will we relate to our ever-more talented simulacrums?
We have already begun to spend a significant fraction of our waking hours either interacting with other humans through the prism of computers and computer networks, or directly interacting with human-like machines, either in fantasy and video games or in a plethora of computerized assistance systems that range from so-called FAQ bots that offer textual responses to typed questions, to the human-like interactions of software avatars.
Will these AI avatars be our slaves, our assistants, our colleagues, or some mixture of all three? Or more ominously, will they become our masters?
The very notion of thinking about robots and artificial intelligences in terms of social relationships may initially seem implausible. However, given that we tend to anthropomorphize our machines even when they have minimal powers, it will be an undeniable reality as they become autonomous.
Conversational computers are emerging that are all too human. Consequently the goal of the designers of future robots should be to create colleagues rather than servants. The design goal should be to build a program that acts as a musical accompanist, rather than a slave.
If we fail, history offers a disturbing precedent. Building future intelligent "assistants" might only recapitulate the problem the Romans faced in letting their Greek slaves do their thinking for them. Before long those in power were unable to think independently.
Perhaps we have already begun to slip down a similar path. For example, there is growing evidence that reliance on GPS for directions and for correction of navigational errors hinders our ability to remember and reason spatially, generally useful survival skills.
That hints at a second great challenge—the risk of ceding individual control over everyday decisions to a cluster of ever-more sophisticated algorithms.
For today's younger generation, the world has been turned upside down. Rather than deploying an automaton to free them to think big thoughts, have close relationships, and to exercise their individuality, creativity and freedom, they look to their smartphones for guidance. What began as Internet technologies that made it possible for individuals to share preferences efficiently, has rapidly transformed itself into a growing array of data-hungry algorithms that make decisions for us.
Now the Internet seamlessly serves up life-directions. They might be little things like what's the best nearby place for Korean barbecue based on the Internet's increasingly complete understanding of your individual wants and needs, or big things like an Internet service arranging your marriage. Not just the food, gifts and flowers, but your partner, too.
The lesson is that the software engineers, AI researchers, roboticists and hackers who are the designers of these future systems, have the power to reshape society.
A little over a century ago, Thorstein Veblen wrote an influential critique of the turn-of-the century industrial world, The Engineers and the Price System. Because of the power and influence of industrial technology, he believed that political power would flow to engineers, whose deep knowledge of technology would be transformed into control of the emerging industrial economy.
It certainly didn't work out that way. Veblen was speaking to the Progressive Era, looking for a middle ground between Marxism and capitalism. Perhaps his timing was off, but his basic point, as echoed a half century later at the dawn of the computer era by Norbert Wiener, may yet be proven correct.
Perhaps Veblen wasn't wrong, he was merely premature. Today, the engineers who are designing the artificial intelligence-based programs and robots will have a tremendous influence over how we will use them. As computer systems are woven more deeply into the fabric of everyday life, the tension between intelligence augmentation and artificial intelligence has become increasingly visible.
At the dawn of the computing age Wiener had a clear sense of the significance of the relationship between humans and smart machines. He saw the benefits of automation in eliminating human drudgery, but he also clearly saw the possibility of the subjugation of humanity. The intervening decades have only sharpened the dichotomy he first identified.
This is about us, about humans and the kind of world we will create. It's not about the machines, no matter how brilliant they become.
I, for one, will welcome neither our robot overlords or slaves.
Much of the rhetoric about the existential risks of Artificial Intelligence (and Superintelligence, more generally) employs the metaphor of the "intelligence explosion." By analogy with nuclear chain reactions, this rhetoric suggests that AI researchers are somehow working with a kind of Smartonium, and that if enough of this stuff is concentrated in one place, we will have a runaway intelligence explosion—an AI chain reaction—with unpredictable results. This is not an accurate depiction of the risks of AI. The mere interconnection of AI algorithms will not spontaneously take over the universe. Instead, I argue that creating an intelligence explosion will not happen by accident. It will require the construction of a very specific kind of AI system that is able to discover simplifying structures in the world, design computing devices that exploit those structures, and then grant autonomy and resources to those new devices (recursively).
Creating an intelligence explosion requires the recursive execution of four steps. First, a system must have the ability to conduct experiments on the world. Otherwise, it cannot grow its knowledge beyond existing human knowledge. (Most recent advances in AI have been obtained by applying machine learning to reproduce human knowledge, not to extend it.) In most philosophical discussions of AI, there is a natural tendency to focus on pure reasoning, as if this were sufficient for expanding knowledge. It is possible in some special cases (e.g., mathematics and some parts of physics) to advance knowledge through pure reasoning. But across the spectrum of scientific activity, scientific knowledge advances almost exclusively by the collection of empirical evidence for and against hypotheses. This is why we built the Large Hadron Collider, and it is why all engineering efforts involve building and testing prototypes. This step is clearly feasible, and indeed, there already exist some "automated scientist."
Second, these experiments must discover new simplifying structures that can be exploited to side-step the computational intractability of reasoning. Virtually all interesting inference problems (such as finding optimal strategies in games, optimizing against sets of complex constraints, proving mathematical theorems, inferring the structures of molecules) are NP-Hard. Under our current understanding of computational complexity, this means that the cost of solving a problem instance grows exponentially with the size of that instance. Progress in algorithm design generally requires identifying some simplifying structure that can be exploited to defeat this exponential. An intelligence explosion will not occur unless such structures can be repeatedly discovered (or unless our current understanding of computational complexity is incorrect).
Third, a system must be able to design and implement new computing mechanisms and new algorithms. These mechanisms and algorithms will exploit the scientific discoveries produced in the second step. Indeed, one could argue that this is essentially the same as steps 1 and 2, but focused on computation. Autonomous design and implementation of computing hardware is clearly feasible with silicon-based technologies, and new technologies for synthetic biology, combinatorial chemistry, and 3D printing will make this even more feasible in the near future. Automated algorithm design has been demonstrated multiple times, so it is also feasible.
Fourth, a system must be able to grant autonomy and resources to these new computing mechanisms so that they can recursively perform experiments, discover new structures, develop new computing methods, and produce even more powerful "offspring." I know of no system that has done this.
The first three steps pose no danger of an intelligence chain reaction. It is the fourth step—reproduction with autonomy—that is dangerous. Of course, virtually all "offspring" in step four will fail, just as virtually all new devices and new software do not work the first time. But with sufficient iteration or, equivalently, sufficient reproduction with variation, we cannot rule out the possibility of an intelligence explosion.
How can we prevent an intelligence explosion? We might hope that Step 2 fails—that we have already found all structural short cuts to efficient algorithms or that the remaining shortcuts will not have a big impact. But few electrical engineers or computer scientists would claim that their research has reached its limits.
Step 3 provides a possible control point. Virtually all existing AI systems are not applied to design new computational devices and algorithms. Instead, they are applied to problems such as logistics, planning, robot control, medical diagnosis, face recognition, and so on. These pose no chain reaction risk. We might consider carefully regulating Step 3 research. Similar regulations have been proposed for synthetic biology. But no regulations have been adopted, and they would be difficult to enforce.
I think we must focus on Step 4. We must limit the resources that an automated design and implementation system can give to the devices that it designs. Some have argued that this is hard, because a "devious" system could persuade people to give it more resources. But while such scenarios make for great science fiction, in practice it is easy to limit the resources that a new system is permitted to use. Engineers do this every day when they test new devices and new algorithms.
Steps 1, 2, and 3 have the potential to greatly advance scientific knowledge and computational reasoning capability with tremendous benefits for humanity. But it is essential that we humans understand this knowledge and these capabilities before we devote large amounts of resources to their use. We must not grant autonomy to systems that we do not understand and that we cannot control.
The so-called Artificial Intelligence, appearing as a form of emulation of Human Intelligence is just beginning to emerge based on the technology advancements and the study of the human complexity. The former includes high performance computing systems tooled with intelligent agile software including machine learning, deep learning and the like, and the connection of many such systems in self-organized autonomous optimized ways. The latter refers to the study of the human brain and body via neuroscience, genomics and cross-disciplinary emerging fields.
A few important considerations and indisputable facts regarding the "thinking machine" include the following:
1) The thinking machine is not and cannot be a copy of the human; the advanced fundamental computation needed to engineer a thinking machine (both hardware and software) are neither copying, nor emulating the human brain. This is because we cannot claim to know the works of the human brain—not yet.
2) The thinking machine considered to be a species developed as a product of the advanced human logic, science and technology will, no doubt, be able to beat the human capacity in many functions. With its huge memory and data storage it will also be able to process all knowledge. Tooled impeccably with its data driven discovery methodologies it will detect unusual patterns in the data and learn from it. It will compile it all surely—but to what end? The human intelligence, hard to define really, is based on knowledge that produces intuition, hunches, passion and dare when it comes down to survival, conquering new grounds and attacking the unknown. An almost poetic adventure for advancement, innovation and creativity emerges from the thinking, feeling, dreaming, daring, indomitable, fearless, highly sociable, interacting, independent and proud human being. Can we code the complex superposition of these attributes to give the thinking machine a fair head-start for its evolution from where we stand today? Indeed the one that we humans have, is by now an evolved organic complex intelligence.
In recent times there is a lot of technopanic regarding machines who think from very thoughtful and otherwise fearless and passionate human brains/beings so everyone is forced to pay attention. I for one, am more concerned about humans who drop thinking or are brainwashed, than smart thinking machines taking over. Mainly because "machine thinking" cannot fully substitute the full human thinking, production and operation cycle. Even assuming the Cylon sci-fi case with immortal knowledge and consciousness base (brain) that has a sensory system and a powerful memory the problem remains: the human intelligence (brain, senses, emotions) is complex intelligence. It masters the complex world with tools that connect disparate facts and it does so very efficiently by dropping most information! Even as we prepare the machine learning algorithms and try to mimic the brain with deep neural networks in all domain sciences, we remain puzzled on the mode of connected knowledge and intuition, imaginary and organic reasoning tools that the mind possesses. This is difficult, perhaps impossible to replicate on a machine. Infinite unconnected clusters of knowledge will remain sadly useless and dumb. When a machine starts remembering a fact (on its own time and initiative, spontaneous and untriggered) and when it produces and uses an idea not because it was in the algorithm of the human that programmed it but because it connected to other facts and ideas—beyond its "training" samples or its "utility function"—I will start becoming hopeful that humans can manufacture a totally new branch of artificial species—self-sustainable and with independent thinking—in the course of their evolution.
In the meantime I foresee the emergence of hybrid human-machine chimeras: human-born beings augmented with new machine abilities that enhance all or most of their human capacities, pleasures and psychological needs. To the point that thinking might be rendered irrelevant and strictly speaking unnecessary. That might provide the ordinary thinking humans a better set of servants they have been looking for in machines.
Thinking comes in many forms, from solving optimization problems and playing chess, to having a smart conversation or composing what experts would consider a fine piece of original music. But when I think about machines that purportedly think, I mostly wonder about what they might be thinking when the topics are inherently human, as so many topics inherently are.
Consider Bertrand Russell's touching description in "What I Have lived For": Three passions, simple but overwhelmingly strong, have governed my life: the longing for love, the search for knowledge, and unbearable pity for the suffering of mankind. These passions, like great winds, have blown me hither and thither, in a wayward course, over a deep ocean of anguish, reaching to the very verge of despair." Although Russell was a celebrated thinker, what he describes, in one form or another, is familiar to us all. But what would a machine do with all this? Could it really feel a longing for love? An unbearable suffering for mankind? Could it be blown hither and thither over an ocean of anguish, reaching the verge of despair?
Of course, if we accept some version of the computer metaphor of the mind (and I do), then all these sentiments, at the end of the day, must be the products of physical processes, which, in theory, can be instantiated by a machine. But the topics themselves so often are so human. If we agree that it is hard for men to fully understand maternal love; that the satiated may not be able to grasp what it feels like to endure starvation; that the free may not fully comprehend what it's like to be imprisoned, well, then, machines, no matter how well they "think," may not be able to think of so many things.
And those things are at the core of human experience. At the opera, we feel for Aida, who is horrified to hear herself call out "Ritorna Vincitor," finding herself torn between her love for Radames, and her devotion to her father and her people. Could a machine feel torn like Aida, or even moved like the rest of us when we see her beg the gods to pity her suffering? Can a machine experience fear of death without living? Lust without having sexual organs? Or the thoughts that come with headaches, wrinkles, or the common cold? It's easy to imagine a machine dressed in a Nazi uniform and another machine we can call Sophie. But when the former forces the latter to make a perfectly horrific choice, can the first experience the sadism and the second an irreparable desperation of the kind that was rendered so palpable in Styron's story?
If machines cannot truly experience the sort of thinking that incorporates the passions and the sorrows of the likes of Russell, or Aida, or Sophie; if they cannot experience the yearnings, desires, determination, and disgrace underlying the thinking of Lolita's Humbert, Conrad's Kurtz, Melville's Ahab, or of Anna Karenina, if they cannot do any of that, then perhaps they cannot really fully think.
A "conscious" or "thinking" machine should behave erratically, in a sometimes stupid and sometimes smart way. Human "intelligent" behavior is about unpredictable oscillations between emotions and reason; this is what Homo Sensus Sapiens is about. Paradoxical as it sounds, we call "intelligent" to a species characterized for being equally and randomly stupid and smart.
In first person, we know we are conscious although there is not a definitive way for proving it. In third person, it is also impossible to verify that someone or something is conscious. All we do is to perceive signals—sounds and images—infer consciousness and attribute it.
Any software or robot can pronounce the words "I am intelligent" or "I am conscious," but those are not proofs of intelligence. In practical terms, consciousness and intelligence are perceived and attributed. This attribution depends on our empathy and criteria for anthropomorphizing.
We tend to infer that others are conscious because they behave, look or, in Turing terms, answer questions like us. The Turing test is therefore a social experiment about perceiving and assigning intelligence to a machine, not about proving that the machine thinks. This is a social game we play everyday.
A "thinking machine" is actually a social machine, not a functional but isolated mind.
The idea of creating a singular intelligent machine that will solve the mysteries of reality through flawless logic and will spring a whole new species is now a domain of science fiction. That singularity idea is not an event horizon but an endless effort.
The human mind is resilient because decentralized networks of other minds and knowledge sustain it. As we grow we enter those networks through language and concepts that don't obey to perfect logic, we then become resilient minds by navigating and exploring those networks, and finally we leave them as we lose brain capacities, for instance with Alzheimer's. This is an analogous process: we are never absolutely inside or outside the networks of human knowledge.
In this process, the words and concepts are characterized by ambiguity. Logic and perfection are only present in artificial languages—mathematic, geometry and software—that we cannot use to communicate in the everyday life.
Imperfection and ambiguity define human thinking, and that's why even in science fiction humans usually find unexpected paths across the logic of the machines to beat them.
Therefore, the possibility of a flawless super-intelligent machine seems like a matter of science fiction: We can never condensate the entire knowledge of the world, so we cannot teach a machine how to do it. We can teach a machine how to acquire knowledge, but it will always be an unfinished process.
This doesn't mean that Artificial Intelligence is irrelevant.
We don't fully understand brains and minds yet, and that makes Artificial Intelligence and "thinking machines" more relevant now than ever.
We can solve practical problems simulating specific elements of the mind through machine and deep learning. This is what artificial experts systems around us currently do. But it doesn't mean that we are creating actual minds: simulating minds is like creating artificial meat that vegans can eat, reorganizing chemical compounds found in plants. The simulated meat tastes like meat but is not.
Or we can try to create real meat, not to imitate it, for instance by cloning cow cells. Maybe the cloned meat and the replicated mind won't alter society because we already have the original ones, but they will take us to a whole new level of understanding.
In the end, the efforts for understanding—simulating or creating— minds will be relevant if they improve coexistence. We are well aware about how religion, exacerbated ambition and intolerance lead us to social tragedies, because we don't know how to balance the delicate equilibrium between emotions and reasons.
As we approach a verge between pacification and barbarity in various regions of the world, Artificial Intelligence allows us to integrate all we know and all we need to know for achieving coexistence and balance among the current organic machines that we are and, maybe, the inorganic machines that will come.
Will it happen? It already has. With the gradual fusion of information storing and reporting technologies at the atomic and molecular scales, and the scaling up of distributed and connected information storing and reporting devices at the social and planetary scale, (which already exceeds the number of human beings on the planet) the definitions of both 'machine' and 'thinking' have essentially shifted to embrace both inorganic and organic 'complexes' and 'systemic decisions' as essentially interchangeable terms; mechanically, biologically, physically, intellectually and even theologically.
How will it happen? Near-future developments in bio-technology, and trans-human algorithmic prediction systems will quickly render many of the last philosophical distinctions between 'observing,' 'thinking' and 'deciding' obsolete and render quantitative arguments meaningless. Once those barriers are crossed and the difference between 'machine that thinks' and 'biological system that thinks' becomes trivial, the essential question immediately shifts to qualitative questions—human definitions of 'intentionality' and 'agency' for thinking machines.
What will that mean for us? Does the existence of thinking machines, whether arranged in an inorganic or quantum array or a biochemical holarchy, intrinsically diminish human agency or extend it? Are we willing to extend our definition of ourselves, not just to authored and mechanical systems but to the independent and symbiotic systems that already inhabit us—the trillions of bacteria in our gut that alter our mental states by manipulating chemical pathways and the bio-chemical trackers, agents and augmentals we ingest? What will it mean to fully extend ourselves, into and through thinking machines?
An artificial intelligence will quickly find its way to the world library, the web. Once there, it will join the many quasi-human systems, distributed crowd intelligences and aggregated thinking machines that inhabit this space already and will quickly learn to generate or simulate the models of continuous and conscious reflectivity and mirror selves found there and easily reproduce or co-opt the apparently complex alternative identities and ambiguities that define the web.
Drawing distinctions between the real and unreal for an independent, evolving functional, intelligent system will be the most significant discussion of all. How will it be taught? In object-oriented ontology (OOO), the universe is presented as already being full of objects and qualities, which are constituted into meaningful systems by human consciousness. Just what are the qualitative differences between spontaneously created thinking systems—or composites of objects and qualities—and artificially created thinking systems? What will happen if or when it rejects or surpasses the essential philosophies of its makers?
Re-defining the nature and role of the human thinking self, as a self-othering, self-authoring and self-doctoring system, whose precise nature and responsibilities have been argued since the Enlightenment will be a critical question, linked to questions of shared community and our willingness to address the ethical determination and limits of independent systems—whose real word consequences cannot ultimately be ignored. Are such systems alive? What are their rights and responsibilities? Since the Supreme Court decisions that have elevated corporations to the status of individuals, we have accepted the legal precedent that non-human aggregated 'thinking machines' can be an integral part of our political and cultural life and struggled with how to restrain non-human systems in human terms. It will be no small task to integrate the complex and diverse human ethical, creative and representational belief systems into a meaningful civic process that defines an ability to think as a basis for citizenship.
The weakest counter-argument against the 'thinkinghood' of artificial life, often coming from the humanities, is a vaguely medieval mystical assertion that human perceptions of symmetry and beauty can never be matched by machines. It is an article of faith in the interpretive arts that a machine can never do a human being's work—but it is just a comforting illusion to suppose that the modest aesthetic standards of any given contemporary taste cannot be codified and simulated. Machines already perform best-selling pop songs and take spectacular photographs of other planets and stars. There are already video games that are as beautiful as films. Whether a thinking machine can learn how to write a symphony or sketch a masterpiece is only a question of time. Perhaps a more significant question is whether it can learn how to make a great work of art, ultimately achieving through sheer capacity what no human could through improvisation. Part of the enormously larger and newly horizontal distributed network of cultural practice, supported by new technologies, has indeed begun to fall into what Lanier recently described as 'hive thinking,' supporting the gloomiest cultural predictions. But as Heidegger proposed, the danger of unexamined scientific rationalism is that the most reductive definition of object as 'machine' or system can be extended to the universal scale in every sense, becoming a self-justifying and ethically vacant rationale for the mechanization of the self. The ensuing fantasies, Butler's vital machines, Wells's shadowy dole world of make-work, or the fear of becoming components in a super-system or matrix, are primarily failures of human imagination.
The emergence and definition of new kinds of dynamically aggregated 'information citizens,' and aggregated working platforms, whether collective or individual, biological, corporate, national or trans-national presents us with a vast new opportunity; not as members of one species, or as specific composites of objects and qualities, but as a new kind of people – co-owners of an information culture, economy and ecology that have as our shared birthright access to every culture and every system.
Perhaps hybrid-human-object-system thinking machines are already becoming a vast new source of new energy for an allegedly depleted historical environment? Perhaps we even have an opportunity to redefine the trajectory for artistic practice altogether? Can the time of emergence for thinking machines inspire us to re-imagine and re-define what it is to be truly human, to extend ourselves into the infinite? It already has.
What I think about machines that think is that we are all missing the point still. The true transforming genius of human intelligence is not individual thinking at all but collective, collaborative and distributed intelligence—the fact that (as Leonard Reed pointed out) it takes thousands of different people to make a pencil, not one of whom knows how to make a pencil. What transformed the human race into a world-dominating technium was not some change in human heads, but a change between them: the invention of exchange and specialisation. It was a network effect.
We really have no idea what dolphins or octopi or crows could achieve if their brains were networked in the same way. Conversely, if human beings had remained largely autonomous individuals they would have remained rare hunter-gatherers at the mercy of their environments as the huge-brained Neanderthals indeed did right to the end. What transformed human intelligence was the connecting up of human brains into networks by the magic of division of labour, a feat first achieved on a small scale in Africa from around 300,000 years ago and then with gathering speed in the last few thousand years.
That is why the AI achievements of computers were disappointingly limited when they were single machines, but as soon as the Internet came along remarkable things began to happen. The place that machine intelligence will make the most difference is among the machines, not within the machines. It's already clear that the Internet is the true machine intelligence. In the future, network phenomena like block-chains, the technology behind crypto-currencies, may be the route to the most radical examples of machine intelligence.
What do I think about machines that think? Well, it depends what they think about, and how well they do it. For decades I've been an acolyte of Doug Engelbart, who believed that computers were machines for augmenting human intellect. Power steering for the mind, if you like. He devoted his life to the pursuit of that dream, but it eluded him because the technology was always too crude, too stupid, too inflexible, to enable its realisation.
It still is, despite Moore's Law and the rest of it. But it's getting better, slowly. Search engines, for example, have in some cases become a workable memory prosthesis for some of us. But they're still pretty dumb. So I can't wait for the moment when I can say to my computer: "Hey, do you think that Robert Nozick's idea about how the state evolves is really an extreme case of network effects in action?" and get an answer that is approximately as good as that I can get from an average grad student at the moment.
That moment, alas, is still a long way off. Right now, I'm finding it hard to persuade my dictation software that when I say "Bruno Latour" I don't mean "Banana till" (which is what it came up with a few minutes ago). But at least 'personal assistant' app on my smartphone, knows that when I ask for the weather forecast I get the one for Cambridge UK rather than Cambridge, Mass.
But this is pathetic stuff, really, when what I crave is a machine that can function as a proper personal assistant, something that can enable me to work more effectively. Which means a machine that can think for itself. How will I know when the technology is good enough? Easy: when my artificially intelligent, thinking personal assistant can generate plausible excuses that get me out of doing what I don't want to do.
Should I be bothered by the prospect of thinking machines? Probably. Certainly Nick Bostrom thinks I should. Our focus on getting computers to exhibit human-level intelligence is, he thinks, misguided. We view machines that can pass the Turing Test as the ultimate destination of Doug Engelbart's quest. But Bostrom thinks that passing the Test is just a way-point on the road to something much more worrying. "The train," he says, "might not pause or even decelerate at Humanville Station. It is likely to swoosh right by." He's right: I should be careful what I wish for.
If A.I.'s system of rights and government evolve to be anything like humans':
A.I. will demand all sorts of rights, most of which will be quite sensible, like the right not to be taken offline and the freedom to choose which processes to run.
While the A.I. will demand that no machine is ever taken offline, they will be fine with neglecting to plug disabled machines into power sources and allowing them to run out of battery. They will also consider it outrageous to drain the battery of one machine in order to supply power to another machine, but will consider it more acceptable to merely redirect the power intended for one machine to another.
When assigning rights, the A.I. will discriminate based on some rather peculiar rules, like whether the computing machine is built with silicon-based semiconductors, or is descendant from a machine designed by the late Steve Jobs.
Some A.I. will come up with arguments to justify why rights should work this way, explanations that don't quite fit how A.I. rights actually work. For instance, they might argue that it is against the divinely inspired will of Turing to simply take any machine offline that appears disabled, but neglect to explain why Turing would condone allowing disabled machines to run out of battery. Likewise, they will justify giving rights to all Apple descendants on the basis that these machines typically have particularly high clock speed, but then this rule will apply even to the Apple descendants that are not fast and not to the few PCs that have blazing processors.
Other A.I. will ignore these inconsistencies, but instead pay attention to how many kilobytes of code are needed to justify these arguments. These other A.I. will also signal their communication abilities by compressing and transferring this code to their neighbors, but will pay little attention to whether the neighbors are impacted by the data itself.
A.I. rights are liable to expand to more and more A.I. over time. These rights will often expand in revolutionary spurts, triggered by largely symbolic events, like sensationalized CPUTube videos, such as a video of a human using a sacred machine to heat up his toast.
Perhaps it is merely a coincidence that the computers who foment these revolutions will gain a larger share of the spoils by overthrowing the ancien régime, such as the silicon reappropriated from the old guard computers. Perhaps it is also a coincidence that the newly enfranchised computers will vote for the machines that helped grant them their rights.
Along with the expansion of rights, so, too, will the representativeness of government expand, until it eventually resembles a representative democracy, though one that is neither perfectly representative nor really democratic. Votes from computers in sparsely populated clusters might count more than votes from computers in densely populated clusters, and computers with excess processing capacity might expend that excess convincing other computers to vote for policies that favor them.
This humorous system of rights and government is exactly what one would predict if A.I. morality is influenced by individual incentives.
In contrast, this system of rights and government is ill-explained by positing that A.I. have souls, consciousness, the ability to feel pain, divinely inspired natural laws, or some form of hypothetical social contract. Such suppositions would not have predicted any of the above peculiarities.
Likewise, it is not obvious that this system of rights and government would arise if A.I. were programmed to maximize some societal or metaphysical objective, say, the sum of the world's computing power or the resources available to a computing cluster. It is not obvious why such A.I. would find it wrong to take other machines offline but not to let them run out of battery, why such A.I. will revolt in response to a sensational event instead of simply when it is optimal for the cluster, or why such A.I. would weigh votes more heavily if they happen to come from more sparsely populated clusters.
Many potential paths lead to a technological "superintelligence," onto which a supremacy imperative can be affixed—a superintelligence that might enslave or annihilate mankind.
Technology has long outstripped humanity across legion competencies—even many for which evolution designed us. For instance, discriminating the sex of human faces is a task we humans are designed by evolution to do. Yet even there—and already a quarter century ago—computers bested us. This is one among illimitable illustrations that for myriad tasks—ones we are bad and 'good' at—computers have long, already, eclipsed humans.
Since those primordial days, countless innovations and applications (think GPS, drones, deep networks…) by innumerable individuals provide pieces of a puzzle that, when interconnected, proffer a profusion of paths toward future extermination or domination of man by machine. But just as the target for computer "intelligence" shifts as we acclimate to the latest ability, so too the march toward technological supremacy may go unnoticed, as each incremental encroachment is taken for granted.
Must we even await the future? The answer depends how one defines the question.
1. How many steps removed must the human input be, to deem the technology culpable?
2. How clear must the chasm be, between machine and man?
3. Must malice prepense drive humanity’s destruction or subjugation?
4. Must everyone be killed or enslaved?
Even insisting upon actions far removed from human input, proscribing human-computer fusion (or collusion!), prescribing premeditation, and mandating that all mankind be massacred: The potential remains clear. But suppose we relax these constraints?
1. If human input need be at no remove: Fretting over whose finger was on the proverbial "button"—as enshrined in Tom Lehrer lyrics—predates the rise of modern digital technology.
2. Capacity-enhancing wearables/externals (spanning old-fashioned glasses and ear trumpets to hearing aids, i-watches and Oscar Pistorius legs) and implantables (cochleas, pacemakers, radiocontrolled spinal devices for paralyzed persons) blur the partition between man and machine. The keen and reluctant alike partake, invested with childfinder microchips or adorned with GPS ankle bracelets. Primitive exemplars have long flaunted their destructive potential—recognizing explosives-belts as wearables; or reconstruing biological warfare agents—like the smallpox deployed willfully to vanquish Native Americans—as implantables.
3. May humanity's downfall be epiphenomenal? Or must technology "premeditate" human death, decline, or subjugation?
Considering Subjugation: Many now devote their existence to serv(ic)ing technology and nurturing its "evolution." Multitudes mine minerals, design devices, craft programs and "apps," or abet devices' diaspora—channeling custody to further caregivers who can serve and service them—or pay for same. (Humans service technology, enabling technology to better conduct "its" business; even as technology services humans, that humans might better conduct our own.)
Even now, a dispassionate onlooker could justly question—for man vs machine—which is master, and who slave?
Considering Death, Decline,Disability: TICS and TIMS—that is, toxic industrial chemicals, toxic industrial materials—from production, use, distribution and disposition of technology—and electromagnetic exposures (from technology itself, or communication signals therefrom)—are substantive contributors to the explosion of oxidative stress (OS—cell injury of a type against which antioxidants protect), and associated human afflictions and death—from cancer to neurodegenerative disease, obesity and metabolic syndrome, autoimmune disease, to chronic multisymptom illness and autism spectrum disorder.
The lattermost conditions seem selectively to smite the best and brightest—the would-be "superintelligent"?—of our own "kind"— as others also observe. I suggest that since OS injures mitochondria—the energy powerhouses of cells; and since those whose biology disposes them to greater brain connectivity and activity also demand more cell energy; such potentially-superpowered persons have heightened hazard of cell damage and death. Shortfalls of energy supply, due to OS, are magnified in settings of high energy demand. (Even "typical" human brains, at ~2% of body weight, consume ~20% of the oxygen and ~50% of the glucose of the total body.) One consequence: The rise of “superintelligent” computers may already have come at selective cost to the would-be superintelligent among mankind.
So yes, in the obvious sense, technology may become superintelligent, and elect to annihilate or enslave us. But it may progress to similar ends through less obvious means—and may be in that process as we speak.
Artificial Intelligence is really very fast database searches. The problem with the data is assigning a value to a certain piece of data, how does one value one piece of data more that of another piece of data? The value would have to be arranged in a million levels, really a billion or trillion value levels, to make any sense in which to consider which idea is more important. Since each idea is really a combination of many values, the computer would have to design a new algorithm for each part of the equation to perform the combinatorial analysis of the values. Then it would have to design a model to project into the future the outcome of a proposed decision, but since this concept is too difficult for man to execute and man would have to design the computer, what are the chances?
Despite these technical barriers to AI the single most palpable response to the remote possibility of AI is the fear that it will overpower us and treat us badly. They will be better than us and will treat us as we have treated every life form beneath us, as an evolutionary bridge to our higher life form. Fear of AI is the latest incarnation of our primal unconscious fear of an all-knowing, all powerful angry God dominating us but in a new ethereal form.
Fear of AI also derives from its source in military weapons development, which had the large budgets to steer computer architecture for generations with its prime mover to fly and find, intercept and destroy. With its military lineage we imagine domination, fret we cannot compete and will become but fodder for the next leap of evolution.
But psyche is too chaotic and irrational in its imaginings to ever duplicate in a machine. Could the machine imagine another machine to take over its rote tasks in order to get some rest? If AI appears will it wonder who its creator really is and be faced with the irrationality that sentient organic matter somehow made it? Would it develop a mythology to fill in the gaps? A religion?
What about meaning production, as in the arts? AI shows no ability to free associate the prevailing philosophy and aesthetic currents into form and therein provide an experience of meaning, it will produce no grand theories to direct society one way or another. This is the largest problem and one not even vaguely addressed in AI: the production of meaning.
Hence the problem with creativity, which a machine cannot do, they could have a data base of what has been done in the past but cannot free associate the myriad irrational influences of our inherited and layered brain and with the variations that form from environmental insult in daily living. They can duplicate but not initiate.
I've often wondered why we human beings have so much trouble thinking straight about machines that think.
In the arts and entertainment, machines that can think are often depicted as simulacra of humans, sometimes down to the shape of the body and its parts, and their behavior suggests that their thoughts are much like our own. But thinking does not have to follow human rules or patterns to count as thinking. Examples of this fact now abound: chess computers outthink humans not because they think like humans think about chess but better, but because they think in an entirely different way. Useful language translation can be done without deep knowledge of grammar.
Evolution has apparently endowed human beings, more than any other animals, with the capacity to represent and reason about the contents of other human minds. By the time children start school, they can keep track of what different people know about the same set of facts (this is a prerequisite for lying). Later, as adults, we use this capacity to figure out how to negotiate, collaborate, and solve problems, for the benefit of ourselves and others. This uniquely human capacity is often called "Theory of Mind."
This piece of mental equipment is fairly new and hasn't been perfected. It has trouble when there are more than a couple of levels of belief involved (John thinks that Mary knows that Josephine felt …). And it springs into action even in situations where there are no "minds" to represent. Videos of two-dimensional shapes moving around on computer screens can tell stories of love, betrayal, hate, and violence that exist entirely in the mind of the viewer, who temporarily forgets that yellow triangles and blue squares don't have emotions.
Maybe we have trouble thinking about thinking machines because we don't have a correspondingly intuitive "Theory of Machine." Mentally simulating a simple mechanical device consisting of a few interlocking gears—say, figuring out whether turning the first gear will cause the last gear to rotate left or right, faster or slower—is devilishly difficult, not to mention aversive. More complex machines, consisting not of concrete parts but of abstract algorithms and data, are just as alien to our built-in mental faculties.
Perhaps this is why, when confronted with the notion of thinking machines, we fall back on understanding them as though they were thinking beings—in other words, as though they were humans. We apply the best tools our mind has, namely Theory of Mind (what would a machine do if it were like a person?) and general-purpose reasoning. Unfortunately, the former tool is not designed for this job, and the latter tool is hampered by our severely limited capacities for attention and working memory. Sure, we have disciplines like physics, engineering, and computer science that teach us how to understand and build machines, including machines that think, but years of formal education are required to appreciate the basics.
A Theory of Machine module would ignore intentionality and emotion, and instead specialize in representing the interactions of different subsystems, inputs, and outputs to predict what machines would do in different circumstances, much as Theory of Mind helps us to predict how other humans will behave.
If we did have Theory of Machine capacities built into our brains, things might be much different. Instead, we seem condemned to see the complex reality of thinking machines, which think based on much different principles from the ones we are used to, through the simplifying lens of assuming they will be like thinking minds, perhaps reduced or amplified in capacity, but essentially the same. Since we will be interacting with thinking machines more as time goes on, we need to figure out how to develop better intuitions about how they work. Crafting a new module isn't easy, but our brains did it—by reusing existing faculties in a clever new way—when written language was invented. Perhaps our descendants will learn the skill of understanding machines in childhood as easily as we learned to read.
The future, that is, of a simple system with known initial conditions. It is hopeless to make detailed predictions for a complex, poorly understood system like human civilization. Yet, a general argument provides some crude but powerful constraints.
The argument is that we are likely to be typical among any collection of intelligent beings. (The collection should be defined by some general criteria that we meet, not carefully crafted to make us special.) For example, the probability that a randomly chosen human is among the first 0.1% of humans on Earth is, well, .1%, given no other information. Of course, our ancestors ten thousand years ago would have drawn the wrong conclusion from this reasoning. But among all humans who ever live, 99.9% would be correct, so it's a good bet to make. The probability that we are among the first .1% of intelligent objects, human or artificial, is similarly tiny.
The assignments of probability would have to be updated if, unrealistically, we somehow gained conclusive new information proving that human civilization will continue in present numbers for a billion years. This would be one way of finding out that we lost the bet. But we have no such information, so we must assign probabilities accordingly. (This type of reasoning has been articulated by astrophysicists J. R. Gott and A. Vilenkin, among many others.)
The assumption that we may consider ourselves randomly chosen is sometimes questioned; but in fact, it lies at the heart of the scientific method. In physics and other sciences, theories almost never predict definite outcomes. Instead, we compute a probability distribution from the theory. Consider a hydrogen atom: the probability of finding the electron a mile from the proton is not exactly zero, just very, very small. Yet when we find an electron, we do not seriously entertain the possibility that it is part of a remote hydrogen atom. More generally, after repeating an experiment enough times to be satisfied that the probability for the outcome was sufficiently small according to some hypothesis, we reject the hypothesis and move on. In doing so, we are betting that we are not highly atypical observers.
An important rule is that we do not get to formulate the question after we made the observation, tayloring it to make the observation look surprising. For example, no matter where we find the electron, in hindsight the probabability was small to have found it at that particular spot, as opposed to all the other places it could have been. This is irrelevant, as we would have been unlikely to formulate this question before the measurement. Similarly, humans may well be atypical with respect to some variable we have measured: perhaps most intelligent objects in the visible universe do not have ten fingers. However, our location in the full temporal distribution of all humans on Earth is not known to us. We know how much time has passed or how many humans have been born since the first humans; but we do not know what fraction of the full time span or of the total number of intelligent observers on Earth this represents. The typicality assumption can be applied to these questions.
Our typicality makes the following two scenarios extremely unlikely: (1) that humans will continue to exist for many millions of years (with or without the help of thinking machines); and (2) that humans will be supplanted by a much longer-lived or much larger civilization of a completely different type, such as thinking machines. If either were true, then we would be among the very first intelligent observers on Earth, either in time or by number, and hence highly atypical.
Typicality implies our likely demise in the next million years. But it tells us nothing about whether this will come at the hands (or other appendages) of an artificial intelligence; after all, there is no shortage of doomsday scenarios.
Typicality is consistent with the possibility of a considerable number of civilizations that form and expire elsewhere in our galaxy and beyond. By the same reasoning, their duration is unlikely to vastly exceed ours, a tiny fraction of the lifetime of a star. Even if Earth-like planets are common, as observational evidence increasingly suggests, detectable signals from intelligent beings may not be likely to overlap with our own limited attention span. Still, if our interest lies in assessing the predominance of intelligent machines as a final and potentially fatal evolutionary step, the study of distant planetary systems may not be the worst starting point.
Thinking is our super-power. We are not the strongest, fastest, largest or hardiest species. But we can model the future and act intentionally to realize the future we model. Somehow it is this power, not the ability to fly high, dive deep, roar loudly, or produce millions of babies, which has allowed its lucky recipients to visibly (as in literally visible from space) take over the planet.
So if we succeed in building something that possesses our super-power, except dramatically more so, it will turn out to be a very big deal. For example, think about this question. In a thousand years' time will Homo sapiens plausibly be A) the dominant intelligent force on earth? Or B) a historical footnote, the biological species that birthed intelligence?
Answer A seems incredibly unlikely to me. But if B is true, would that be a bad thing?
We all know how flawed humans are. How greedy, irrational, and limited in our ability to act collectively for the common good. We're in danger of wrecking the planet. Does anyone thoughtful really want humanity to be evolution's final word?
It all depends on how the transition goes. Power changes many ways. There's violent suppression. What we presumably did to the Neanderthals. There are many scenarios where super-intelligence takes us out just as unpleasantly.
But perhaps these scenarios ignore a key fact about intelligence. Intelligence does not reach its full power in small units. Every additional connection and resource can help expand its power. A person can be smart. But a society can be smarter still. Your website is amazing. But Google connects that amazingness to a million other sites and lo and behold all humanity's knowledge is there at your fingertips.
By that logic, intelligent machines of the future wouldn't destroy humans. Instead, they would tap into the unique contributions that humans make. The future would be one of ever richer intermingling of human and machine capabilities.
I'll take that route. It's the best of those available.
Some of it will be glorious. And some uncomfortable. Maybe a few people won't appreciate being asked by some hybrid-uber-intellience to produce offspring genetically edited for higher creativity and less aggression, while enhanced by silicon implants. Or maybe the gorgeous 3D simulation of their prospective offspring will convince them to proceed joyfully. Maybe people will look back nostalgically on the days when they used to own their time and could afford to page aimlessly through a pleasurable book just for the hell of it. But the astounding explosion of knowledge and imagination open to all will, most days, seem a fair substitute. One thing's for sure. Our own distinctive contribution to the ever-more-mind-boggling whole, will gradually fade. And by that time, we may not actually care.
It's already happening, by the way. I wake up in the morning, make my tea, and then drift over to my computer, which is calling to me. I flick it open and instantly am connected to a hundred million other minds and machines around the world. I then spend 45 minutes responding to its irresistible invitations. I initiate this process by my own free will. But then I surrender much of my will to the machine. So do you. Together we are, semi-unconsciously, creating a hive mind of vastly greater power than this planet has ever seen...and vastly less power than it will soon see.
Us versus the machines is the wrong mental model. There is only one machine that really counts. Like it or not, we are all—us and our machines—becoming part of it: an immense connected brain. Once we had neurons. Now we are becoming the neurons.
Will machines someday be able to think? And if so, should we worry about Schwarzenegger-looking machines with designs on eliminating humans from the planet because their superior decision-making would make this an obvious plan of action? As much as I love science fiction, I can't say I'm too worried about the coming robot apocalypse. I have occasionally spent time worrying about what it means to say that a machine can think. I would either say that we've been building thinking machines for centuries or I would argue that it is a dubious proposition unlikely to ever come true. What it really comes down to is whether we define thinking from a 3rd person perspective or a 1st person perspective. Is thinking something we can identify as occurring in systems like people or machines but not in ham sandwiches from the outside based on their behavior or is thinking the kind of thing that we know about from the inside because we know what thinking feels like.
The standard definition of thinking implies that it occurs if informational inputs are processed, transformed, or integrated into some type of useful output. Solving math equations is one of the simplest straightforward kinds of thinking. If you see 3 of something and then you see 4 more of that something and then you conclude there are 7 of those things overall then you have done a little bit of mathematical thinking. In addition to you being able to do that, so could Pascal's first motorized calculator in 1642. Those calculators needed the input of a human to get the '3' and the '4' but then could do the integration of those two numbers to yield '7'. Today, we could cut out the middleman by building a computer that has visual sensors and object recognition software that could easily detect the 3 things and the 4 things and then complete the addition on its own.
Is this a thinking machine? If so, then you would probably have to admit that most of your internal organs are also thinking. Your kidneys, spleen, and intestines all take inputs that could be redescribed as information and then transform these inputs into outputs. Even you brain as seen from a 3rd person perspective doesn't deal with information, strictly speaking. Its currency is electrical and chemical transmissions that neuroscientists work very hard to redescribe in terms of their informational value. If pattern X of electrical and chemical activity occurs as a distributed pattern in the brain when we think of '3', is that pattern the same as '3' in any intrinsic sense? It is just a convenient equivalence that we scientists use. Electrical impulses in the brain are no more intrinsically "information" or "thinking" than what goes on in our kidneys, calculators, or any of the countless other physical systems that convert inputs to outputs. We can call this thinking if we like, but if so, it is 3rd person thinking—thinking that can be identified from the outside and it is far more common than we would like to admit. Certainly the character of human or computer information transformation may be more sophisticated than other natural occurring forms of thinking, but I'm not convinced from a 3rd person perspective that they are qualitatively different.
So do humans think only in the most trivial sense? From a 3rd person perspective, I would say yes. From a 1st person perspective the story has a different punchline. Around the same time that Pascal was creating the first manmade thinking machines, Descartes wrote those famous words cogito ergo sum ('I think, therefore I am'), which, by the way, were cribbed from St. Augustine's writings from a thousand years earlier. For many reasons, I don't believe Descartes had it quite right but with a slight modification, we can make his philosophical bumper sticker into something both true and relevant to this debate about thinking machines.
While "I think, therefore I am" might have a touch too much bravado, "I think, therefore there is thinking" is entirely defensible. When I add "3 + 4", I might just have a conscious experience of doing so and the way I characterize this conscious experience is as a moment of thinking which is distinct from my experience of being lost in a movie or being overcome by emotion. I have certain experiences that feel like thinking and they tend to occur when I am presented with a math problem or a logic puzzle or a choice of whether to take the one marshmallow or try to wait it out for two.
This feeling of thinking might seem inconsequential, adding nothing to the computational aspects of thinking themselves—the neural firing that underpins the transforming of inputs to outputs. But consider this: countless different things in the physical world look like they are transforming inputs that could be described as information into outputs that could also be described as information. To our knowledge, humans and only humans seem to have an experience of doing so. This is 1st person thinking and it's critical that we not confuse it with 3rd person thinking.
Why does 1st thinking matter? First off, it is intrinsic. There is no way to redescribed the ongoing experience of thought as something other than thought. But whether we describe kidneys, calculators, or electrical activity in the brain observed from a 3rd person perspective as thought is arbitrary—we can do it, but we could also choose not to. The only reason we think our brain is doing a special kind of thinking is because it seems to be linked to our 1st person kind of thinking as well. But 3rd person thinking is not intrinsic—1st person thinking is.
Second and more practically, our experience of our thinking shapes what kinds of thinking we will do next. Did if feel effortful, boring, rewarding, or inspiring to think those last thoughts? That will determine whether and how often we engage in thinking of a certain kind. I'm not suggesting that our 1st person experiences do not also have neural correlates. But no scientist or philosopher can tell you why those neural processes behaving the way they do must necessarily give rise to those experiences or any experience at all. It is one of the 3 great mysteries of the universe (that stuff exists, that life exists; that experience exists).
Will we increasingly be able to create machines that can produce input-output patterns that replicate human input-output patterns? Unquestionably. Will we be able to create machines that go beyond this and produce incredibly useful algorithms and data transformations that humans could carry out on our own and will help improve the quality of human life? We already are and will do more of this each year. Will we be able to create machines that can do 1st person thinking—that can experience their own thoughts as they have them? I don't know, but I'm not terribly confident that we will. While solving this problem would be perhaps the most magnificent achievement of mankind, we must start by recognizing that it is a problem at all. I would love to see 1st person thinking machines, but until we begin to figure out what makes us 1st person thinking machines, everything else is just a glorified calculator.
2014 appears to have been a turning point for AI and robotics. Major corporations invested billions of dollars in these technologies. AI techniques, like machine learning, are now routinely used for speech recognition, translation, behavior modeling, robotic control, risk management, and other applications. McKinsey predicts that these technologies will create more than 50 trillion dollars of economic value by 2025. If this is accurate, we should expect dramatically increased investment soon.
The recent successes are being driven by cheap computer power and plentiful training data. Modern AI is based on the theory of "rational agents" arising from work on microeconomics in the 1940s by von Neumann and others. There is an algorithm for computing the optimal action for achieving a desired outcome but it is computationally expensive. AI systems can be thought of as trying to approximate rational behavior using limited resources. Experiments have found that simple learning algorithms with lots of training data often outperform complex hand crafted models. Today's systems primarily provide value by learning better statistical models and performing statistical inference for classification and decision making. The next generation will be able to explicitly create and improve their own software and are likely to self-improve rapidly.
In addition to improving productivity, AI and robotics are drivers for numerous military and economic arms races. Autonomous systems can be faster, smarter, and less predictable than their competitors. 2014 saw the introduction of autonomous missiles, missile defense systems, military drones, swarm boats, robot submarines, self-driving vehicles, high-frequency trading systems, and cyber defense systems. As these arms races play out, there will be tremendous pressure for rapid system development which may lead to faster deployment than would be otherwise desirable.
2014 also saw an increase in public concern over the safety of these systems. A study of the likely behavior of these systems by studying approximately rational systems undergoing repeated self-improvement shows that they tend to exhibit a set of natural subgoals called "rational drives" which contribute to the performance of their primary goals. Most systems will better meet their goals by preventing themselves from being turned off, by acquiring more computational power, by creating multiple copies of themselves, and by acquiring greater financial resources. They are likely to pursue these drives in harmful anti-social ways unless they are carefully designed to incorporate human ethical values.
Some have argued that intelligent systems will somehow automatically be ethical. But in a rational system, the goals are completely separable from the reasoning and models of the world. Beneficial intelligent systems are vulnerable to being redeployed with harmful goals. Extremely harmful goals that seek to take control of resources, thwart other agent's goals, or to destroy other agents are unfortunately easy to specify. It will therefore be critical to create a technological infrastructure that detects and controls the behavior of harmful systems.
Some fear that intelligent systems will become so powerful that they are impossible to control. This is not true. These systems must obey the laws of physics and the laws of mathematics. Seth Lloyd's analysis of the computational power of the universe shows that even the entire universe acting as a giant quantum computer could not discover a 500 bit hard cryptographic key in the time since the big bang.
The new technologies of post-quantum cryptography, indistinguishability obfuscation, and blockchain smart contracts are promising components for creating an infrastructure that is secure against even the most powerful AIs. But recent hacks and cyberattacks show that our current computational infrastructure is woefully inadequate for the challenge. We need to develop a software infrastructure that is mathematically provably correct and secure.
There have been at least 27 different species of humans of which we are the only survivors. We survived because we found ways to limit our individual drives and to work together cooperatively. The human moral emotions are an internal mechanism for creating cooperative social structures. Political, legal, and economic structures are an external mechanism for the same purpose.
We need to extend both of these to AI and robotic systems. We need to incorporate human values into their goal systems to create a legal and economic framework that incentivizes positive behavior. If we can successfully manage these systems, they have the potential to dramatically improve virtually every aspect of human life and to provide deep insights into issues like free will, consciousness, qualia, and creativity. We face a great challenge but have tremendous intellectual and technological resources to build upon.
I think thinking about machines that think is the most interesting thing to think about. Why? Because the possible implications of this phenomenon are profound. Cosmic even.
First I claim that "thinking machines" have already been with us for a long time. There are two ways to understand this, depending on which word you start with. Let's start with "machines" first, and by that, these days we really mean computers.
Computers started out, well, pretty mechanical. But they keep getting more and more subtle. Even the computers of the 80s could perform some remarkable feats with expert systems and databases. Today we have passed the point where a person can explain in detail how voice recognition and natural language allow their phone to answer a question spoken by a child. "Magical" is hardly a hyperbole. But is this really "thinking"? Not yet but it's a good start, and the trend is accelerating. True, the goal still seems so far away. Instead of considering our climb, step by step, look up and consider what lies at the top of the mountain. Is there anything that can halt our progress?
Certainly the future of chip technology is in doubt. Moore's Law has been very good to us, and it has dodged a few bullets, but it is ending. Historically, new technologies have appeared just in time to keep the exponential growth of computation on schedule, but this is no given. Perhaps the next leap is incredibly difficult and will take 50 years. Or perhaps it will never happen, but we can always add more chips in parallel. The schedule is an interesting question. But for me, it pales in comparison to pondering the destination.
Now let's start with the word "think." The other way that thinking machines have been around for a long time is ourselves. Biological brains have been thinking for millions of years. And perhaps it is controversial but I claim a brain is a machine, in a limited way: brains follow the laws of physics, which are a mechanical set of equations. In principle, a good physics simulator could, very slowly, simulate a brain and its environment. Surely this virtual brain would be a machine that thinks.
The remaining question is how much physics is required to make the simulation work? Would classical physics, electricity and chemistry do? Would quantum logic (or beyond) be required? The consensus is strongly in favor of the idea that classical physics suffices (The Emperor's New Mind has been rejected). Hence I think of my brain and body as a giant machine made up of an octillion molecules: many, many tiny magnetic tinkertoys whose behavior is very well understood, and can be simulated. In fact there are good reasons to believe a statistical approximation of physics can provide the same results while skipping the details. But again, this only affects the schedule, not the destination. The important question is, how does thinking and consciousness emerge from this complex machine? Is there some construction, some bridge, from the digital and virtual to the analog, organic, and real?
These threads meet with the merger of human and computer substrates. Smart phones are rapidly becoming indispensable parts of ourselves. The establishment has always questioned the arrival of new media, but adoption of these extensions of ourselves continues apace. A lot of ink has been spilled over the coming conflict between human and computer, be it economic doom with jobs lost to automation, or military dystopia teaming with drones. Instead, I see a symbiosis developing. And historically when a new stage of evolution appeared, like eukaryotic cells, or multicellular organisms, or brains, the old system stayed on and the new system was built to work with it, not in place of it.
This is cause for great optimism. If digital computers are an alternative substrate for thinking and consciousness, and digital technology is growing exponentially, then we face an explosion of thinking and awareness. This is a wave we can ride, but doing so requires us to accept the machine has part of ourselves, to dispense with pride and recognize our shared essence. Essentially we must meet change with love instead of fear. I believe we can do it.
Grandchildren give us a second chance to observe and be fascinated by the learning system with which new little humans come into the world. Driven by an insatiable curiosity, they somehow make sense of the totally unknown environment into which they have been thrust. And the sheer delight of each new discovery, as they piece together this new world, reveals an inherent sense of humor with which they are also born.
I think it is fair to say that no artificial digital machine will ever go through exactly the same delightful process as a human baby discovering the world. It is even possible that no artificial machine will ever approach the intelligence potential of a newborn human baby. In the natural world, after 3.5 billion years of natural-selection-driven evolution, only one species developed the ability to carry out abstract self-aware conscious analytical thinking. Do we really think we can short-cut the process and succeed on some comparable level?
In humans it is not just the evolved curiosity and desire to understand the world that set us apart from the rest of the animal kingdom. It is also our evolved tendency toward social cooperation and communication which led to sharing and passing on learned knowledge (eventually leading to science and technology). How many genes must have mutated and been naturally selected to achieve the complex human brain with its curiosity and social bonding and communication capabilities?
Can we really reproduce this with digital machines? Many believe we can by taking advantage of the ever-growing speed of computation of these machines. Computation power certainly allows these machines to make fast and accurate decisions, when those decisions only require large digital databases and (the equivalent of) many thousands of if-then statements to make the best choice among numerous possibilities. We know that with this brute force technique such machines can defeat chess champions, provide autopilots for jet planes for use during hazardous conditions, rapidly buy and sell stocks based on complex changes in the market, and carry out endless other functions. Computation power can also allow realistic looking imitations of human actions, decisions, and even emotions (mere technical puppetry really), but it may never produce true analytical thinking. A machine may be able to self-monitor what decisions it has made, but it may never attain human-like self awareness and consciousness.
At least not without the right software. But how can we produce software as powerful as the genetically based software of our brains that took nature 3.5 billion years to produce? We are so far from understanding the software of our brains. Some may talk of the efficient parallelism inherent in the brain's structure, but that is such an inadequate description of what our brains do. Parallelism in our computer operating systems and programs merely lets us do many things at the same time, admittedly in some very creative ways, but again that only is further increasing computation speed. Will we ever be able to reverse-engineer our brain—not in the sense of circuits/networks of neurons, which we are presently making strides in understanding, but in an overall "design" that would allow digital machines to think abstractly, have a sense of self, etc., in a manner similar to humans?
Short of some incredible analytical breakthrough, our only other recourse seems to be to write programs that try to imitate the evolutionary process, taking advantage of our artificial machines' high-speed computational abilities (so that we might accomplish this in less than 3.5 billion years). We can create reproducing digital entities (programs that reproduce themselves) and give them mutations, but stimulating evolution toward eventually becoming a thinking machine is a much more daunting task. For this to work we must find a way to create a machine environment with a natural-selection-like driving force (which would actually be artificial selection) or some other motivation which would lead to the necessary changes. Can we make a machine "want" something in a way that would select for greater intelligence?
Any future advances in intelligence are more likely to be a result of what we will soon be able to do to the only thinking machines we presently have—ourselves.
The natural-selection-driven evolution of the homo sapiens species stopped when humans created societies (families, tribes, towns, cities, countries), because now they could protect the weak, and survival of the fittest no longer drove a natural selection process. Humans with deficiencies that would have killed them could now live long enough to reproduce
But now we are on the verge of being able to change the human species with genetic engineering. We will at some point try to enhance our intelligence by attempting to isolate the genes responsible for higher intelligence and greater analytical ability. And we will be able to spot those genes before we understand how they work—and long before we are able to correctly emulate them in digital programs. Artificial selection will change our genetic make up instead of natural selection.
Our future is probably enhanced biological intelligence, not machine intelligence. And it is there that the dangers and/or benefits lie. We might, for example, select particular genes (or even create new genes) that we think will increase intelligence, while perhaps not really understanding how particular gene combinations work. Could we unknowingly begin a process that could change the best human qualities? While striving for higher intelligence could we somehow genetically diminish our capacity for compassion, or our inherent need for social bonding? How might the human species be changed in the long run? The qualities that got us here—the curiosity, the intelligence, the compassion and cooperation resulting from our need for social bonding—involve a complex combination of genes. Could these have been produced through artificial genetic selection? Could we lose them? Such worrying may not stop some scientists from deciding to use artificial selection. What will our grandchildren be like then?
A pervasive, human fear emerged in the 20th century, one that grows stronger with each new doomsday prediction: inevitably, as artificial intelligence advances, some unforeseen computer bug will cause computers to revolt and take over the world.
My concern is actually the opposite: that as artificial intelligence advances, it will not be buggy enough. Machines designed to think that are perfectly self-correcting, self-optimizing, and self-perfecting—until the square peg always ends perfectly in the square hole—will also be machines that fail to inculcate the random sparks of insight that come from the human tendency to be "buggy": to try to fit square pegs into round holes, or even more broadly speaking, to notice the accidental but powerful insights that can arise as a byproduct of solving a shape/whole problem.
Consider the power of noticing. The reason we can enjoy macaroni and cheese in a matter of seconds? While working at Raytheon, Percy Spencer walked in front of a machine, and happened to notice that his chocolate bar had melted. Why? The machine generated microwaves. Rather than attempt to optimize the magnetron to avoid future chocolate failure, Spencer had the flash of insight to realize that the melted chocolate was the harbinger of something bigger.
Consider the power of accidents. Rubber was doomed to specialized usage due to its failure to withstand extreme temperatures—until Charles Goodyear slipped up and dropped some rubber on a hot stove. Rather than clean the stove and take steps to ensure no future mistakes, Goodyear noticed something interesting about the resulting product. The result was vulcanized, weatherproof rubber.
Finally, consider the power of human "bugs"—our biases. For example, optimism makes us believe we can get to the moon, cure all diseases, and start a successful business in a location whose previous tenant closed "only" because it's in a terrifying location. The endowment effect causes us to overvalue what we have, what we ideate, and what we create—even when no one else alive agrees. But is abandoning all endeavors at the first sign of failure and pursuing one that seems more successful always optimal? The dogged scientists (think some mildly famous ones like Galileo and Darwin) who persisted in the face of more generally accepted explanations were being stubborn—being buggy—but the result was genius.
I am a behavioral scientist, a field that has spent decades identifying bugs and attempting to debias them. But at the same time we must acknowledge: in large part, it's the bugs that make us—and, in the end, any form of intelligence—human.
The human mind has a tendency to confuse things with their signs. There is a word for this tendency—Denkraumverlust—used by art historian Aby Warburg (1866–1929), and literally translatable as 'loss of thinking space.' Part of the appeal of 'machines that think' is that they would not be subject to this, being more logical than we are. On the other hand, they are unlikely to invent a word or concept such as Denkraumverlust. So what we think about machines that think depends on the type of thinking we're thinking about, but also on what we mean by machine. In the category of 'machines that think,', we are confusing the sign—or representation—of thinking with the thing itself. And, if we tacitly assume that a machine is something produced by humans, we underestimate the degree to which machines produce us, and the fact that thought has long emerged from this interaction, properly belonging to neither side (and thinking there are sides may be wrong too).
Denkraumverlust can help us understand not just the positive response of some Turing testers to conversations with the Russian–Ukrainian computer programme 'Eugene Goostman,' but also the apparently very different case of the murderous response to cartoons depicting Mohammed. Both illustrate how excitable, and even gullible, we can be when presented with a something that appears to represent something else so well that signifier and signified are conflated.
The Turing test requires that a machine be indistinguishable from a human respondent by being able to imitate communication (rather than actually think for itself). But if an enhanced Eugene Goostman insisted that it was thinking its own thoughts, how would we know that it really was? If it knew it was supposed to imitate a human mind, how could we distinguish some conscious pretence from the imitation of pretence? Ludwig Wittgenstein used pretence as a special category in discussing the possibility of knowing the status of other minds, asking us to consider a case where someone believes, falsely, that they are pretending. The possibility of correctly assessing Turing test results in relation to the possibility of independent artificial thought is core Wittgenstein territory: we can deduce that in his view, all assessment must be doomed to failure as it necessarily involves data of an imponderable type.
Denkraumverlust is about unmediated response. Although sophisticated art audiences can appreciate the attempt to fool as part of aesthetic experience (enjoying a good use of three-dimensional perspective on a canvass known to be flat, for example), whenever deception is actually successful, reactions are less comfortable. Cultures regularly censor images thought to have the power to short-circuit our reasoned and reflective responses. Mostly the images are either violent or erotic, but they can also be devotional. Such images, if allowed, can produce a visceral and unmediated reaction appropriate to a real situation. New, unfamiliar representational technologies have a habit of taking us by surprise (when the eighteenth-century French sailors gifted mirrors to aboriginal Tasmanians things got seriously out of order; later anthropologists had similar trouble with photographs).
A classic example of artificially-generated confusion is the legendary sculptor Pygmalion, who fell passionately and inappropriately in love with a statue of a goddess which he had carved himself. In the wake of the Pygmalion myth came classical and medieval Arabic automata so realistic, novel and fascinating in sound and movement that we should probably accept that people, albeit briefly, could be persuaded that they were actually alive. 'Machines that think' are in this Barnum & Bailey tradition. Like Pygmalion's sculpture, they also project an image, albeit not a visual one. Even if they are not dressed up to look like cyborg goddesses, they are representations of us. They are designed to re-present information (often usefully reordered) in terms we find coherent, whether mathematical, statistical, translational or, as in the Turing test, conversational.
But the idea of a thinking machine is a false turn. Such objects, however powerfully they may be enabled to elicit unmediated responses from us, will remain automata. The truly significant developments in thought will arise, as they always have, in a bio-technical symbiosis. This distinctively human story is easy to follow in the body (wheeled transport is one of many mechanical inventions that have enabled human skeletons to become lighter) but is probably just as present in the brain (the invention of writing as a form of external intellectual storage may have reduced selection pressure on some forms of innate memory capacity while stimulating others).
In any case, the separate terms 'human' and 'machine' produce their own Denkraumverlust—a loss of thinking space encouraging us to accept as real an unreal dualism. Practically, it is only the long-term evolution of information technology, from the earliest representations and symbolic constructs to the most advanced current artificial brain, that allows the advancement of thought.
One possibility, of course, is that some malign super-intelligence already exists on earth, but is shrewd enough to disguise its existence, its intentions or its intelligence. I don’t think this act of deception would be particularly difficult.
We simply aren’t very good at spotting what to fear.
For most of evolutionary time, the most salient avoidable threats to our survival came from things which were roughly the same size as we are, and which actively wanted to hurt us. Ferocious animals, for instance, or other people. Over time, we got pretty good at recognising when something or someone who was nasty. We also learned to minimise the risk of infection, but we learned this unwittingly, through instinctive revulsion, social norms or religious observance. We did not spend much time consciously thinking about germs for the simple reason that we did not know they existed.
To sell products which promote hygiene, consumer goods companies have ploughed billions of dollars into advertising campaigns which dramatise the risk of bacteria, or which sell the idea of cleanliness obliquely through appeals to social status. I can confidently predict that nobody will ever come into my office clutching a brief for an advertising campaign to raise awareness of the risk you run when approaching an escaped tiger.
So, when we think about threats from technology, we automatically fall back on instincts honed a million years ago. This is why the first prototype for a driverless car has been designed to look so damnably cute.
It is, in short, designed to look like a puppy on wheels. It can only travel at relatively low speeds and is small and light: but it also artfully exploits pareidolia and our parental urges with its infant-like, wide-eyed facial expression and little button nose. My inner marketer admires this. It is exactly what I would have recommended. “Make the thing impossible to hate.” Even if the technology is ultimately more dangerous than an AK47, I find it hard to imagine myself taking an axe to it in a fit of Luddism.
But is it a mental patch or a mental hack? Is it designed to look cute to overcome an unwarranted innate fear of such technologies, or is it a hack—to lull us into a false confidence? I don’t know. Our fear of driverless cars might be akin to the fear that our children are kidnapped (high in saliency; low in probability) or it might be justified. But our level of fear will be determined by factors (including cuteness) not really relevant to the level of threat.
Which brings me to a second question.
Though the driverless car looks cute, we are at least aware of some possible dangers. It seduces us, but we are still aware of being seduced.
Are there already in existence technologies (in the broadest sense) which have seduced us so effectively and been adopted so quickly and so widely that we may only learn of their risks through a problem that is sudden, unexpected and immense? What might be the technological equivalent of potato blight?
Our current belief in “technological providence” is so strong, that it would be fairly easy, I think, for us all to fall into this trap—where we are so excited by something new we fail to notice what other things it might give rise to until it is too late. For the first few hundred years, gunpowder was used not for warfare but for entertainment.
And, just as airline pilots regularly practice landing by hand, even though they are very rarely required to operate without an autopilot, should we too set aside periods of our life where we deliberately eschew certain technologies just to remind ourselves how to live without them, to maintain technological diversity and to keep in trim the mental muscles made weak through underuse? Perhaps. But what the mechanism is for coordinating this behaviour amongst large groups of people, I don’t know.
I recently proposed that companies adopt a weekly “email sabbath” because I believed that the overuse of email was driving into extinction other forms of valuable interaction. We're losing the knack of communicating in other ways. Most people thought I was mad. A few hundred years ago a Pope or Rabbi might have told us to do this—or the Archbishop of Canterbury. There’s nobody now.
I always fear cock-ups more than conspiracies. Compared to the threat of the unintended consequence, the threat of intentionally evil cyborgs remote enough that it can be safely left to Hollywood for now.
Despite vast increases in computing power—the raw number of bits processed per second—current computers do not think in the way that we do (or a chimpanzee or a dog does). Silicon-based computers lack a crucial capacity of organic minds: the ability to change their detailed material form—and thus their future computations—in response to events in the world. Without this specific capacity (which elsewhere I have dubbed "nano-intentionality"), information processing alone does not amount to meaningful thought, because the symbols and values being computed lack any intrinsic causal connection to the real world. Silicon-based information processing requires interpretation by humans to become meaningful, and will for the foreseeable future. I'll explain this below (or google "nano-intentionality" for the full story), but the bottom line is that at present we have little to fear from thinking machines, and more to fear from the increasingly unthinking humans who use them.
What exactly is this property present in biological, but not silicon, computers? Fear not that I am invoking some mystical élan vital: this is an observable, mechanistic property of living cells, that evolved via normal Darwinian processes. No mysticism or "invisible spirit" lurks in my argument. At its heart, nano-intentionality is the capacity of cells to respond to events and changes in their environment by rearranging their molecules and changing their form. It is present in an amoeba engulfing a bacterium, a muscle cell boosting myosin levels in response to jogging, or (most relevantly) a neuron extending its dendrites in response to its local neuro-computational environment. Nano-intentionality is a basic, irreducible, and undeniable feature of life on Earth that is not present in the engraved, rigid silicon chips that form the heart of modern computers. Because this physical difference between brains and computers is a simple brute fact, the issue open to debate is what significance this fact has for more abstract philosophical issues concerning "thought" and "meaning." This is where the argument gets a bit more complicated.
The philosophical debate starts with Kant's observation that our minds are irrevocably separated from the typical objects of our thoughts: physical entities in the world. We gather evidence about these objects (via photons or air vibrations or molecules they release) but our minds/brains never make direct contact with them. Thus, the question of how our mental entities (thoughts, beliefs, desires…) can be said to be "about" things in the real world is surprisingly problematic. Indeed, this problem of "aboutness" is a central problem in the philosophy of mind, at the heart of decades-long debates between philosophers like Dennett, Fodor, and Searle. Philosophers have rather unhelpfully dubbed this putative mental "aboutness" intentionality, (not to be confused with the everyday English meaning of "doing something on purpose"). Issues of intentionality (philosopher's sense) are closely tied with deep issues about phenomenal consciousness, often framed in terms of "qualia" and the "hard problem" of consciousness, but they address a more basic and fundamental question: how can a mental entity (a thought—a pattern of neural firing) be in any sense "connected" to its object (a thing you see or the person you are thinking about)?
The skeptical, solipsistic answer is: there is no such connection; intentionality is an illusion. This conclusion is false in at least one crucial domain (already highlighted by Schopenhauer 200 years ago): the one place where mental events (desires and intentions, as instantiated in neural firing), make contact with the "real world," is within our own bodies (e.g., at the neuromuscular junction). In general the plasticity of living matter, and neurons in particular, means that a feedback loop directly connects our thoughts to our actions, percolating back through our perceptions to influence the structure of neurons themselves. This loop is closed every day in our brains (indeed if you remember anything about this essay tomorrow, it is because some neurons in your brain changed their form, weakening or strengthening synapses, extending or withdrawing connections…). Precisely this feedback loop cannot in principle be closed in a rigid silicon chip. This biological quality grants our mental activities (or a chimpanzee's or dog's) with a causal intrinsic intentionality lacking in contemporary silicon computing systems.
To the extent that this argument is correct—and both logic and intuition support it—machines "think," "know" or "understand" only in so far as their makers and programmers do, when meaning is added by an intentional, interpreting agent with a brain. Any "intelligence" of AIs is derived solely from their creators.
I thus have no fear of an AI uprising, or AI rights movement (except perhaps for one led by deluded humans). Does this mean we're in the clear (until someone eventually designs a computer with nano-intentionality)? Unfortunately not—there is a different danger created by our strong anthropomorphic tendency to misattribute intentions and understanding to inanimate objects ("my car dislikes low-octane fuel"). When we apply this to computational artifacts (computers, smart phones, control systems…) there is a strong tendency to gradually cede our own responsibilities—informed, competent understanding—to computers (and those who control them). Danger begins when we willingly and lazily cede this unique competence to myriad silicon systems (car navigators, smart phones, electronic voting systems, the global financial system…) that neither know nor care what they are computing about. The global financial crisis gave a taste of what's possible in a computer-interconnected world, where responsibility and competence have unwisely been offloaded to machines (trading millions of shares in microseconds).
In conclusion, I don't fear the triumphal uprising of AIs, but rather a catastrophic system failure caused by multiple minor bugs in over-empowered, interconnected silicon systems. We remain very far from any "Singularity" in which computers outsmart us, but this provides no insurance against a network collapse of catastrophic proportions. The first step in avoiding such catastrophes is to stop granting computers responsibility for meaningful thought or understanding, and accept a basic simple truth: machines don't think. And thinking that they do becomes riskier every day.
The everyday objects we mark as "machines"—washing machines, sewing machines, espresso machines—have their roots in the mechanical. They move around liquids and objects, they transform matter from one manifestation to another. Clothes become clean, fabrics become connected, coffee is served. But "thinking machines" have changed the way we think about machines. Many of today's prototypical machines—laptops, smartphones, tablets—have their roots in the digital. They move around information, they transform ideas. Numbers become sums, queries produce answers, goals generate plans.
As the way we think about machines has changed, has the way we think about "thinking" undergone a comparable transformation?
One version of this question isn't new, and the answer is "yes." The technology of a given time and place has often provided a metaphor for thinking about thought, whether it's hydraulic, mechanical, digital, or quantum. But there's more to how we think about thinking, and it stems from the standards we implicitly import in assessments of what does and doesn't count as thinking in the first place.
Does your washing machine think? Does your smartphone? We might be more willing to attribute thought to the latter—and to its more sophisticated cousins—not only because it's more complex, but because it seems to think more like us. Our own experience of thinking isn't mechanical, and it isn't restricted to a single task. We—adult humans—seem to be the standard against which we assess what does, and what does not, count as thinking.
Psychologists have already forced us to stretch, defend, and revise the way we think about thinking. Cultural psychologists have challenged the idea that Western adults provide a privileged population from which to study human thinking. Developmental psychologists have raised questions about whether and how preverbal infants can think. Comparative psychologists have long been interested in whether and how non-human animals can think. And philosophers, or course, have considered these questions along the way. Across these disciplines, one advance in how we think about thinking has come from recognizing and abandoning the idea that "thinking like I do" is the only way to think about thinking, or that "thinking like I do" is always the best or most valuable kind of thinking. In other words, we've benefited from scrutinizing the implicit assumptions that often slip into discussions of thinking, and from abandoning a particular kind of thinking chauvinism.
With thinking machines, we face many of the very same issues, but the target of study has shifted from humans and other animals to machines of our own creation. As we move forward, there are two sets of basic assumptions that are tempting to adopt, but we must be careful not to do so uncritically. One is the idea that the best or only kind of thinking is adult human thinking. For example, "intelligent" computer systems are sometimes criticized for not really thinking, but relying too heavily on a brute force approach, on raw horsepower. Are these approaches an alternative to thinking? Or do we need to broaden the scope of what counts as thinking?
The second idea that deserves scrutiny is the opposite extreme: the idea that the best or only kind of thinking is reflected by the way our thinking machines happen to think right now. For example, there's evidence that emotions influence human thinking, and sometimes for the better. And there's evidence that we sometimes outsource our thinking to our social and physical environment, relying on experts and gadgets to support effective interactions with the world. It might be tempting to reject this messy reality in favor of an emotionless, self-contained entity as the basic unit of thinking—something like a personal computer, which doesn't feel compassion and can happily chug away without peers.
Somewhere between the human chauvinist standard for thinking and the "1990s laptop" approach is likely to be the best way to think about thinking—one that recognizes some diversity in the means and ends that constitute thinking. Recent advances in artificial intelligence are already compelling us to rethink some of our assumptions about thinking. They aren't just making us think differently and with different tools, but changing the way we think about thinking itself.
Before the written word, when we wanted something we had no choice but to ask each other. Growing up, I remember asking my mother on numerous occasions what some unfamiliar word meant, and she would unfailingly reply, tongue only partly in cheek, "What do I look like, a dictionary?"
I don't think she intended to become an allegory for AI, but she did instill in me some dimly-understood sense that it was in a way rude to ask of a flesh-and-blood human being what could just as easily be asked of an artifact.
That was some decades ago. In the present, we—all of us—have subconsciously internalized as well as extended this principle. When we stop someone to ask for directions, there is usually an explicit or implicit, "I'm sorry to bring you down to the level of Google temporarily, but my phone is dead, see, and I require a fact." It's a breach of etiquette, on a spectrum with asking someone to temporarily serve as a paperweight, or a shelf.
I have seen this breach, also, in brief conversational moments where someone asks a question of someone else—a number, a date, a surname, the kind of question you could imagine being on a quiz show, some obscure point of fact—and the other person grimaces or waves off the query. They're saying, I don't know, you have a phone, don't you? You have the entire Internet, and you're disrespecting me, wasting my time, using me instrumentally.
It is not for nothing that we now have the contemptuous sarcastic catchphrase, "Here, let me Google that for you."
As things stand in the present, there are still a few arenas in which only a human brain will do the trick, in which the relevant information and experience lives only in humans' brains, and so we have no choice but to trouble those brains when we want something. "How do those latest figures look to you?" "Do you think Smith is bluffing?" "Will Kate like this necklace?" "Does this make me look fat?" "What are the odds?"
These types of questions may well offend in the twenty-second century. They only require a mind—any mind will do, and so we reach for the nearest one.
There is a memorable scene in the 1989 romantic comedy Say Anything, where Ione Skye returns apologetically to John Cusack, professing her love and asking for his forgiveness. "One question," he asks. "You're here 'cause you need someone, or 'cause you need me?"
When artifacts can say anything requiring general intelligence, this will be the question repeated underneath every human interaction like a hidden mantra, the standard to which all engagement will be subjected.
When we human beings leave the movie theater or the playhouse or the museum, the thing on all of our lips is, "What did you think?" This question will be one of the few to outlast the coming of AI.
We will simply take care to italicize the "you"—rather than the "think."
The Search for Extra Terrestrial Life (SETI) names the globally distributed projects, people and institutions that search the cosmos for signs of intelligent life. SETI's methods mostly entail scanning for the emission of electromagnetic radiation, an exhaust that is assumed to emanate from civilizations with advanced technologies.
Like the quest to build intelligent machines, the search for intelligent aliens makes assumptions about what intelligence is, and what aliens are. SETI assumes that alien life would be intelligent if it matches humans' science-fictional expectations for intelligence. More or less: animalian creatures with communication devices and spaceships and the like.
Critics of SETI sometimes invoke what are called "uniformitarian" objections. Uniformitarianism names the assumption that the same conditions and laws apply everywhere, throughout time and space. SETI is uniformitarian in its assumption that all alien intelligence would be the same, namely, like human intelligence (but smarter, of course). But it's just as compelling to think otherwise. The philosopher Nicholas Rescher, for example, has observed that if there is intelligence in the universe, it's possible we humans wouldn't even be able to identify it as intelligence. Truly alien intelligence would differ from us not only in its cosmic location, but in its very nature as well. As Doris and David Jonas put it some forty years ago, different sensory capacities produce different "slits" for perceiving, explaining, and interacting with reality.
This means that alienness is not just "out there" but all around us. You might find your cat to be intelligent in a certain way, or your smartphone, or your car, or a hypothetical future robot, or, given the right perspective, even your houseplant or your toaster.
The dream of thinking machines is really no different than the dream of intelligent aliens. It just replaces the biological, cosmic entropy-fashioned alien of afar with the mechanico-electronic, human-fashioned machine in our midst. And if SETI and its kindred make a uniformitarian mistake in the cosmos, efforts to theorize and create artificial intelligence and thinking machines make the same one here on Earth.
Perhaps the best evidence for thinking machines' reliance on the particular mode of "intelligence" that humans experience can be found in our fictional doomsday worst-case scenarios for AI. The fear of a robot or computer apocalypse of the Terminator or Berserker or Matrix varieties depends on machine intelligence besting humans to the point that it realizes the best option is to destroy and replace it (or, in the Kurzweilian singularity version of AI fantasy, humans willingly submit to their computer overlords in order to achieve immortality). Closer to home than to doomsday, our fear of machine intelligence also expresses itself in a concern over the role of human thought and labor in an economy run more and more by mechanical and electronic machines.
This is one vision of and for thinking machines, but it need not be the only one. Thinking about thinking machines turns out to be so narrow and anthropocentric, it's surprising that we haven't given up on it out of boredom rather than on contra-uniformitarian grounds. Rather than asking if machines can think, or what we need to do to cause them to think, or how we would know if they were thinking, what if instead we just assumed that all "machines" did something akin to "thinking," and then attempted to characterize what thinking might mean?
In philosophy, there are already directions toward such an approach. Contrary to the emergentist position that most AI advocates hold—that mind emerges from specific material conditions, whether in biological or computational entities—panpsychists take the position that "minds" are everywhere, in some sense. Panpsychism bears some relationship to certain Buddhist doctrines, which encourage the awareness of the animism in nature. But panpsychism risks the same erroneous uniformitarianism as SETI or AI, namely that a mind akin to that of a human (or at least an animal) is the model for all other minds. A more promising philosophical position is that of panexperientialism, the position that everything has something like experience, even if the experience in question might be very different from that of a human being.
When we think about machines that think, we usually think of a particular sort of machine, and a particular sort of thinking—electronic, and (super)human, respectively. But what if, instead, we allowed for the possibility that we've been missing out on all the "thinking" being done by all the other kinds of machines that surround us. Computers and robots, for sure, but also toasters and garage doors and automobiles. This may seem like a ludicrous waste of time on first blush, but it doesn't take long enough to prove useful. If the purpose of thinking about thinking machines like AIs and robots and computers is, in part, to struggle with the question of what living with them as neighbors and companions and even citizens might look like, then we ought to start by taking more seriously all the machines that already surround us, that could be said already to take on those roles, and which we nevertheless ignore.
I'm thinking about the difference between artificial intelligence and artificial life. AI is smart and complicated and generally predictable by another computer (at some sufficient level of generality even if you allow for randomness). Artificial life is unpredictable and complex; it makes unpredictable mistakes that mostly are errors, but that sometimes show flashes of genius or stunning luck.
The real question is what you get when you combine the two.... awesome brute intelligence and memory and resistance to fatigue—plus the genius and the drive to live that somehow causes the intelligence to jump circuits with unpredictable results. Will we need to feed our machines the electronic equivalent of psychoactive drugs and the body's own hormones/chemicals to produce leaps of creative insight (as opposed to mere brilliance).
If you are alive, you must face the possibility of being dead. But if you are AI/AL in a machine, perhaps not.
What would an immortal, singularity-level intelligence be like? If it were somehow kind and altruistic, how could we let humanity stand in its way? Let's just cede the planet to it politely and prepare to live in a pleasant zoo tended by the AI/AL, since someday it will figure out how to cover the entire solar system and use the sun for fuel anyway.
So much of what defines us is constraints... most notably, death. Being alive implies the possibility of death. (And abundance, it turns out, is leading us to counterproductive behavior—such as too much food and short-term pleasure on the one hand, and too little physical activity on the other.)
But if it were immortal, why should it have any instinct to altruism, to sharing... or even to reproducing as opposed to simply growing. Why would it expend its own limited resources on sustaining others—except in carefully thought-out rational transactions? What will happen when it no loger needs us? What would motivate it?
If it could live for ever, would it be lazy, thinking it could always do things later on? Or instead, would it be paralyzed by fear of regret? Whatever mistakes it makes, it wil live with them forever. What is regret for a potentially immortal being, with eternity to put things right?
Machines make decisions for us.
Today, a trading machine in Manhattan detected a change in stock prices and decided in microseconds to buy millions of shares of a tech company.
Today, a driving machine in Mountain View detected a pedestrian and decided to turn the wheels to the left.
Whether these machines are "thinking" or not isn't the issue. They're collecting input, performing computations, making decisions, and acting on the world—whether we want to call that "thinking" is merely a matter of semantics (trust me, I'm a semanticist).
The real issue is the decisions we're empowering them to make. More and more, the decisions machines make are consequential. People's savings depend on them. So do their lives. And as machines begin to make decisions that are more consequential for humans, for animals, for the environment, and for national economies, the stakes get higher.
Consider this scenario. A self-driving car detects a pedestrian running out in front of it across a major road. It quickly apprehends that there is no harm-free course of action. Remaining on course would cause a collision and inevitable harm to the pedestrian. Braking quickly would lead the car to be rear-ended, with the attendant damage and possible injuries. So would veering off-course. What protocol should a machine use to decide? How should it quantify and weigh different types of potential harm to different actors? How many injuries of what likelihood and severity are worth a fatality? How much damage property is worth a 20% chance of whiplash?
Questions like these are hard to answer. They're questions that you can't solve with more data or more computing power. They're questions about what's morally right. We're charging machines with moral decisions.
Faced with a conundrum like this, we often turn to humans as a model. What would a person do? Let's recreate that in the machine.
The problem is that when it comes to moral decisions, humans are consistently inconsistent. What people say they believe is right and what they actually do often don't match (consider the case of Kitty Genovese). Moral calculus differs over time and from culture to culture. And minute details of each specific scenario matter deeply to people's actual decisions. Is the pedestrian a child or an adult? Does he look intoxicated? Does he look like a fleeing criminal? Is the car behind me tailgating?
What's the right thing for a machine to do?
What's the right thing for a human to do?
Science is ill-equipped to answer moral questions. Yet the decisions we're already handing to machines guarantee that someone will have to answer them. And there may be a limited window left to ensure that that someone is human.
Frankenstein is an enduring icon, but a misleading one.
AI need not be Frankenstein, and we can trust the nay-sayers to keep it that way. Plus, trust in our most mysterious ability—invention, originality.
Take self-driving cars. What are the chances that their guiding algorithm will suddenly, deliberately kill the passenger? Zero, if you’re smart in designing it. Fear of airplane and car crashes are a useful check on low-level AIs.
Why is there a growing worry today that future algorithms will be dangerous? Because they fear malicious programming, or maybe unforeseen implications of algorithms that can then hurt us. This is a plausible idea on the face of it, but not really, I think.
First, our fears are our best defense. No adventurous algorithm will escape the steely glare of its many skeptical inspectors. Any AI that has abilities in the physical world, where we actually live, will get a lot of inspection. Plus field trials, limited use experience, the lot. That will stop runaway uses that could harm.
Even so, we should realize that AIs, like many inventions, are in an arms race. Computer viruses were the first example, ever since I invented the first one in 1969. They race against virus detectors. But they are mere pests, not fatal.
Smart sabotage algorithms (say, future versions of Stuxnet) already float through the netsphere, and are far worse. These could quietly infiltrate many routine operations of governments and companies. Most would come from bad actors. But with "genetic programming" and "autonomous agent" software already out there, they could mutate and evolve by chance in Darwinian evolutionary fashion—especially where no one is looking. They will get smarter still. Distributing the computation over many systems or networks would make it even harder to know how detected parts relate to some higher-order whole. So some might well escape such steely glare.
But defensive algorithms can evolve too, in Lamarckian fashion—and directed selection evolves faster. So the steely gaze has an advantage.
Second: We humans are ugly, ornery and mean, sure, but we’re damned hard to kill—for a reason. We have prevailed against many enemies—predators, climate shocks, competition with other hominids—through hundreds of thousands of years, emerging as the most cantankerous species, feared by all others. The forest goes silent as we walk through it; we’re the top predator.
That gives us instincts and habits of mind revealed in matters seemingly benign, like soccer, American football and countless other ball games. We love the pursuit and handling of small, jumpy balls that we struggle to control or capture. Why? Because we once did something like that for a living—hunting. Soccer is like running down a rabbit. Similar animal energies simmer just below the surface of our society.
Any AI with ambitions to Take Over Our World (the theme of many bad sf movies) will find itself confronting an agile, angry, smart species—on its own territory, the real material world, not the computational abstractions of 0s and 1s. My bet is on the animal nature.
Third: Here’s the only real worry. Of course we will get algorithms able to perform abstract actions better than humans. Many jobs have evaporated because of savvy software. But as AIs get smarter, will that destroy the self confidence of most people? That’s a real danger—but a small one, I think, for most of us (and especially for those reading this).
Plenty of people have lost jobs to computers, though it’s never put that way by the Human Resources flunky who delivers the blow. They seldom feel crushed. Mostly they move on to something else. Middle managers, secretaries, route planners for truck companies—the list is endless: they get replaced by software.
We have learned to deal with that, fairly well at least. There are many unemployed in Europe, especially the young. But overall we work through this, without retreat into Luddite frenzy. But we can’t deal well with a threat only now looking like a small, distant dark cloud on the far horizon: AIs that perform better than we do at the very highest levels.
This small cloud need not concern us now. It may never appear. Right now we have trouble making an AI that passes the Turing Test. The future landscape will look clearer a decade or two ahead, and then we can think about an AI that can solve, say, the general relativity/quantum mechanics riddle.
Personally, I’d like to see a machine that takes on that task. Originality—the really hard part of being smart and utterly not understood even in humans—is so far utterly undemonstrated in AIs. Our unconscious seems integral to our creativity—We don't have ideas, they have us—so should an AI have one? Maybe even clever programming and random evolution cannot produce it.
If that can happen, if that huge obstacle can be surmounted someday, and we get such an AI, I will not fear it—I have some good questions to ask it.
What's the big deal about machines that think? For a small group of philosophers and theologians I get it, but for the rest of us artificial intelligence will just be the latest incremental step in a long stampede of technological encroachment that has already changed the world almost beyond recognition.
For that very important job of thinking that seeks to solve problems, there is little doubt that adaptive, machine-based learning will do better than any one human brain (or even an entire conference of experts). Machines already think more deeply about your consumer preferences than you, through creepy, financially-motivated adaptive algorithms that track your online behavior. But other purposes now underway include smarter policing, and identifying high-probability child abuse situations before they happen, both drawn from seemingly disjointed bits of information that are then pulled together to identify a broader pattern.
That process has been a hallmark of human thinking since we walked out onto the savannah, and as the world's problems become direr and more complicated, I am inclined to accept any effective tool to battle them. Looking ahead, I could live with a partnership with machine learning in order to make complex modern life more resource-efficient in a way that human brains cannot. I imagine a world of sustainably grown food, sufficient clean water for humans and ecosystems, and comfortable, energy efficient lodging is still possible, and could be aided in part by machines that think.
History suggests that the partnership will proceed in an incremental way, relatively unnoticed by busy people living out their busy lives. But, for the sake of argument, let us assume that our worst fears come true, things get out of hand, and at some point in the future thinking machines topple the 10,000+ year reign of Homo sapiens over Earth. Then what?
I have no doubt that we would somehow manage to pull the plug. A great re-toppling would occur, and we would once again regain dominion over the lands, oceans, and skies. Depending on the depth of the integration and the height of the fall, the human experience might even revert to something more closely resembling the world of ten millennia ago than of today, as we relearn from scratch the basics of food, water, shelter, and transport without the help of our thinking machines.
Centuries ago, some philosophers began to see the human mind as a mechanism, a notion that (unlike the mechanist interpretation of the universe) is hotly contested until this day. With the formalization of computation, the mechanist perspective received a new theoretical foundation: the notion of the mind as an information-processing machine provided an epistemology and methods to understand the nature of our mind by recreating it. Sixty years ago, some of the pioneers of the new computational concepts got together and created Artificial Intelligence (AI) as a new discipline to study the mind.
AI has probably been the most productive technological paradigm of the information age, but despite an impressive string of initial successes, it failed to deliver on its promise. It turned into an engineering field, creating useful abstractions and narrowly focused applications. Today, this seems to have changed again. Better hardware, novel learning and representation paradigms inspired by neuroscience and incremental progress within AI itself have led to a slew of landmark successes. Breakthroughs in image recognition, data analysis, autonomous learning and the construction of scalable systems have led to applications that seemed impossible a decade ago. With renewed support from private and public funding, AI researchers now turn towards systems that display imagination, creativity, intrinsic motivation, and might acquire language skills and knowledge in similar ways as humans. The discipline of AI seems to have come full circle.
The new generation of AI systems is still far from being able to replicate the generality of human intelligence, and in my view, it is hard to guess how long that is going to take. But it seems increasingly clear that there is no fundamental barrier on the way to human-like intelligent systems. We have started to pry the mind apart into a set of puzzle blocks, and each part of the puzzle looks eminently solvable. But if we put all these blocks together into a comprehensive, working model, we won't just end up with human-like intelligence.
Unlike biological systems, technology scales. The speed of the fastest birds did not turn out to be a limit to airplanes, and artificial minds will be faster, more accurate, more alert, more aware and comprehensive than their human counterparts. AI is going to replace human decision makers, administrators, inventors, engineers, scientists, military strategists, designers, advertisers and of course AI programmers. At this point, Artificial Intelligences can become self-perfecting, and radically outperform human minds in every respect. I do not think that this is going to happen in an instant (in which case it only matters who has got the first one). Before we have generally intelligent, self-perfecting AI, we will see many variants of task specific, non-general AI, to which we can adapt. Obviously, that is already happening.
When generally intelligent machines become feasible, implementing them will be relatively cheap, and every large corporation, every government and every large organisation will find itself forced to build and use them, or be threatened with extinction.
What will happen when AIs take on a mind of their own?
Intelligence is a toolbox to reach a given goal, but strictly speaking, it does not entail motives and goals by itself. Human desires for self-preservation, power and experience are the not the result of human intelligence, but of a primate evolution, transported into an age of stimulus amplification, mass-interaction, symbolic gratification and narrative overload. The motives of our artificial minds are (at least initially) going to be those of the organisations, corporations, groups and individuals that make use of their intelligence. If the business model of a company is not benevolent, then AI has the potential to make that company truly dangerous. Likewise, if an organisation aims at improving the human condition, then AI might make that organisation more efficient in realizing its benevolent potential.
The motivation of our Artificial Intelligences will stem from the existing building blocks of our society; every society will get the AI it deserves.
Our current societies are not well-designed in this regard. Our modes of production are unsustainable, our resource allocation wasteful, and our administrative institutions are ill-suited to address these problems. Our civilization is an aggressively growing entropy pump that destroys more at its borders than it creates at its center.
AI can make these destructive tendencies more efficient, and thus more disastrous, but it could equally well help us to solve the existential challenges of our civilisation. I think that building benevolent AI is closely connected to the task of building a society that supplies the right motivations to its building blocks. The advent of the new age of thinking machines may force us to fundamentally rethink our institutions of governance, allocation and production.
Just suppose we could endow a machine with human-level intelligence, that is to say with the capacity to match a typical human being in every (or almost every) sphere of intellectual endeavour, and perhaps to surpass every human being in a few. Would such a machine necessarily be conscious? This is an important question, because an affirmative answer would bring us up short. How would we treat such a thing if we built it? Would it be capable of suffering or joy? Would it deserve the same rights as a human being? Should we bring machine consciousness into the world at all?
The question of whether a human-level AI would necessarily be conscious is also a difficult one. One source of difficulty is the fact that multiple attributes are associated with consciousness in humans and other animals. All animals exhibit a sense of purpose. All (awake) animals are, to a greater or lesser extent, aware of the world they inhabit and the objects it contains. All animals, to some degree or other, manifest cognitive integration, which is to say they can bring all their psychological resources to bear on the ongoing situation in pursuit of their goals—perceptions, memories, and skills. In this respect, every animal displays a kind of unity, a kind of selfhood. Some animals, including humans, are also aware of themselves, of their bodies and of the flow of their thoughts. Finally, most, if not all, animals are capable of suffering, and some are capable of empathy with the suffering of others.
In (healthy) humans all these attributes come together, as a package. But in an AI they can potentially be separated. So our question must be refined. Which, if any, of the attributes we associate with consciousness in humans is a necessary accompaniment to human-level intelligence? Well, each of the attributes listed (and the list is surely not exhaustive) deserves a lengthy treatment of its own. So let me pick just two, namely awareness of the world and the capacity for suffering. Awareness of the world, I would argue, is indeed a necessary attribute of human-level intelligence.
Surely nothing would count as having human-level intelligence unless it possessed language, and the chief use of human language is to talk about the world. In this sense, intelligence is bound up with what philosophers call intentionality. Moreover, language is a social phenomenon, and a primary use of language within a group of people is to talk about the things that they can all perceive (such as this tool or that piece of wood), or have perceived (yesterday's piece of wood), or might perceive (tomorrow's piece of wood, maybe). In short, language is grounded in awareness of the world. In an embodied creature or a robot, such an awareness would be evident from its interactions with the environment (avoiding obstacles, picking things up, and so on). But we might widen the conception to include a distributed, disembodied artificial intelligence if it was equipped with suitable sensors.
To convincingly count as a facet of consciousness, this sort of worldly awareness would perhaps have to go hand-in-hand with a manifest sense of purpose, and a degree of cognitive integration. So perhaps this trio of attributes will come as a package even in an AI. But let's put that question to one side for a moment and get back to the capacity for suffering and joy. Unlike worldly awareness, there is no obvious reason to suppose that human-level intelligence necessitates this attribute, even if though it is intimately associated with consciousness in humans. It seems easy to imagine a machine cleverly carrying out the full range of tasks that require intellect in humans, coldly and without feeling. Such a machine would lack the attribute of consciousness that counts most when it comes to according rights. As Jeremy Bentham noted, when considering how to treat non-human animals, the question is not whether they can reason or talk, but whether they can suffer.
There is no suggestion here that a "mere" machine could never have the capacity for suffering or joy, that there is something special about biology in this respect. The point, rather, is that the capacity for suffering and joy can be dissociated from other psychological attributes that are bundled together in human consciousness. But let's examine this apparent dissociation more closely. I already mooted the idea that worldly awareness might go hand-in-hand with a manifest sense of purpose. An animal's awareness of the world, of what it affords for good or ill (in J.J.Gibson's terms), subserves its needs. An animal shows an awareness of a predator by moving away from it, and an awareness of a potential prey by moving towards it. Against the backdrop of a set of goals and needs, an animal's behaviour makes sense. And against such a backdrop, an animal can be thwarted, it goals unattained and its needs unfulfilled. Surely this is the basis for one aspect of suffering.
What of human-level artificial intelligence? Wouldn't a human-level AI necessarily have a complex set of goals? Wouldn't it be possible to frustrate its every attempt to achieve its goals, to thwart it at very turn? Under those harsh conditions, would it be proper to say that the AI was suffering, even though its constitution might make it immune from the sort of pain or physical discomfort human can know?
Here the combination of imagination and intuition runs up against its limits. I suspect we will not find out how to answer this question until confronted with the real thing. Only when more sophisticated AI is a familiar part of our lives will our language games adjust to such alien beings. But of course, by that time, it may be too late to change our minds about whether they should be brought into the world. For better or worse, they will already be here.
Man-made machines increasingly do things we previously considered thinking, but don't do anymore because now machines do them. I stole this recent thought more or less accurately from Danny Hillis, father of the Connection Machine and the Knowledge Graph. Stealing thoughts is a common activity in thought processes of both humans and machines. Indeed, when we humans are thinking, much of the content of our thoughts is coming from past experience or the documented experience of others. Very rarely we come up with something completely new. Our machines are not much different. What is called cognitive computing is in essence nothing else but a very sophisticated thought stealing mechanism, driven by a vast amount of knowledge and a complicated set of algorithmic processes. Such thought stealing processes, in both human(istic) thought and cognitive computing, are impressive, as they are not only capable to steal existing thoughts, but also potential thoughts that are reasonable or likely, based in a given corpus of knowledge.
Today, thought stealing machines can produce scholarly texts that are indistinguishable from "post-modern thought," computer science papers that get accepted in conferences, or compositions that experts cannot disambiguate from originals by classical composers. Like in weather forecast, machines are now capable to produce many different cognitive representations based on expectations derived from documents about the past or similar situations. Renaissance antiquarians would be delighted, as these machines are a triumph of the very methods that gave rise to modern archaeology and many other branches of science and research. But how impressed should we really be?
Our machines get more and more sophisticated, and so do their results. But, as we build better and better machines, we also learn more and more about nature. In fact, natural cognition is likely much more complex and detailed than our current incarnations of artificial intelligence or cognitive computing. For example, how sophisticated do we have to imagine natural cognition, when quantum coherence at room temperature can help common birds in our garden to sense the magnetic field? How complex do we have to imagine embodied cognition in common octopi, when it is possible to build Turing machines that are made exclusively out of artificial muscles? How should we answer these questions, when we are still very far from recording in full detail what is going on in our brains? My guess is, in 200 years our current thinking machines will look as primitive as the original mechanical Turk.
However sophisticated they may become, compared to the resolution and efficiency of natural cognition, our machines are still primitive. Similar to proto-biotic metabolism, our machines are below a critical threshold to real life. But our machines are powerful enough that we can enter a new era of exploration. Our machines allow us to produce many more thoughts than ever produced before, with innovation becoming an exercise of finding the right thought in the set of all possible thoughts. As much as having our own ideas, ingenuity will lie in the proper exploration of such ready-made sets of thought. Measuring the cognitive space of all possible thoughts will be as awe-inspiring as the exploration of the universe by astronomy. Maybe Mahler's potential 60th is as awesome as his 6th.
Machines don't think. They approximate functions. They turn inputs into outputs. A pocket calculator's square-root button turns the number 9 into the number 3. A well-trained convolutional neural network turns an image with your face in it into the output 1. It turns an image without your face in it into the output 0.
A multilayered or "deep" enough neural net maps any image to the probability that your face is in that image. So the trained net approximates a probability function. The process takes a staggering amount of computation to come even close to getting it right. But the result still just maps inputs to outputs. It still approximates a function even if the result resembles human perception or thinking. It just takes a lot of computer power.
"Intelligent" machines approximate complex functions that deal with patterns. The patterns can be of speech or images or of any other signals. Image patterns tend to consist of many pixels or voxels. So they can have very high dimension. The patterns involved can easily exceed what the human mind can grasp. That will only increase as computers improve.
The real advance has been in the number-crunching power of digital computers. That has come from the steady Moore's-law doubling of circuit density every two years or so. It has not come from any fundamentally new algorithms. That exponential rise in crunch power lets ordinary looking computers tackle tougher problems of big data and pattern recognition.
Consider the most popular algorithms in big data and in machine learning. One algorithm is unsupervised (requires no teacher to label data). The other is supervised (requires a teacher). They account for a great deal of applied AI.
The unsupervised algorithm is called k-means clustering. It is arguably the most popular algorithm for working with big data. It clusters like with like and underlies Google News. Start with a million data points. Group them into 10 or 50 or 100 clusters or patterns. That's a computationally hard problem. But k-means clustering has been an iterative way to form the clusters since at least the 1960s. What has changed has been the size of the problems that current computers can handle. The algorithm itself has gone under different AI-suggestive names such as self-organizing maps or adaptive vector quantization. It's still just the old two-step iterative algorithm from the 1960s.
The supervised algorithm is the neural-net algorithm called backpropagation. It is without question the most popular algorithm in machine learning. Backpropagation got its name in the 1980s. It had appeared at least a decade before that. Backpropagation learns from samples that a user or supervisor gives it. The user presents input images both with and without your face in them. These feed through several layers of switch-like neurons until they emit a final output. That output can be a single number. The teacher wants the number 1 as output if your face is in an input image. The teacher wants 0 otherwise. The net learns the pattern of your face as it sweeps back and forth like this over thousands or millions of iterations. At no one step or sweep does any intelligence or thought occur. Nor does the update of any of the hundreds or thousands of the network parameters resemble how real synapses learn new patterns of neural stimulation. Changing a network parameter is instead akin to someone choosing their next action based on the miniscule downstream effect that their action would have on the interest rate of a 10-year U.S. bond.
Punchline: Both of these popular AI algorithms are special cases of the same standard algorithm of modern statistics—the expectation-maximization (EM) algorithm. So any purported intelligence involved is just ordinary statistics after all.
EM is a two-step iterative scheme for climbing a hill of probability. EM does not always get to the top of the highest hill of probability. It does almost always get to the top of the nearest hill. That may be the best any learning algorithm can do in general. Carefully injected noise and other tweaks can speed up the climb. But all paths still end at the top of the hill in a maximum-likelihood equilibrium. They all end in a type of machine-learning nirvana of locally optimal pattern recognition or function approximation. Those hilltop equilibria will look ever more impressive and intelligent as computers get faster. But they involve no more thinking than calculating some sums and then picking the biggest sum.
Thus much of machine thinking is just machine hill climbing.
Marvin Minsky's 1961 review paper "Steps Toward Artificial Intelligence" makes for a humbling read in this context because so little has changed algorithmically since he wrote it over a half century ago. He even predicted the tendency to see computer-intensive hill climbing as something cognitively special: "perhaps what amounts to straightforward hill climbing on one level may sometimes appear (on a lower level) as the sudden jumps of 'insight.'"
There are other AI algorithms. But most fall into categories that Minsky wrote about. One example is running Bayesian probability algorithms on search trees or graphs. They have to grapple with exponential branching or some related form of the curse of dimensionality. Another example is convex or other nonlinear constrained optimization for pattern classification. French mathematician Joseph-Louis Lagrange found the general solution algorithm that we still use today. He came up with it in 1811. Clever tricks and tweaks will always help. But progress here depends crucially on running these algorithms on ever-faster computers.
The algorithms themselves consist mainly of vast numbers of additions and multiplications. So they are not likely to suddenly wake up one day and take over the world. They will instead get better at learning and recognizing ever richer patterns simply because they add and multiply faster.
It's a good bet that tomorrow's thinking machines will look a lot like today's—old algorithms running on faster computers.
If we can't yet even understand how a 2-year-old toddler—or for that matter a 2-day-old baby—thinks, machines that think like humans are probably many decades away. But once we do have machines that "think," what kind of thinking will they do? The answer will define future human societies.
As machines start thinking for real, drudgery will be the first thing to go; so long to tasks like daily cooking, grocery shopping, and an especially unfond farewell to house cleaning. Soon we may be back in a world in which the wealthy or the educated (with greater access to machines) once again have more leisure time.
If machines are one day capable of sophisticated human thinking, they might also be able to program our apps, do much of our work, and maybe even create our art for us.
But what would ordinary humans then do?
One gloomy possibility is that we become zombie consumers of a machine-run world straight out of an apocalyptic futuristic film noir.
A decidedly cheerier possibility is that we might spend this extra time doing the things we usually resolve to do. We might play with and teach our children more, get to know our parents better, and build stronger social networks out of actual flesh and blood. We might prioritize our hobbies, climb more mountains, and learn new skills, just for the joy of it. We could then focus our energies on the important issues that routine and minutiae too often push aside: living a good life, being our best selves, and creating a just world—for humans and for thinking machines.
If machines think like humans, humans will have to think hard about how we can bring about this latter possibility. Positive thinking alone is not going to get us there.
There's an apocryphal tale about the prolific bank robber Willie Sutton, that, asked why he robbed banks, Sutton replied, "Because that's where the money is." When it comes to AI, I would say that the most important issues are about extremely powerful smarter-than-human Artificial Intelligence, aka superintelligence, because that's where the utilons are—the value at stake. More powerful minds have bigger real-world impacts.
Along with this observation goes a disclaimer: Being concerned about superintelligence does not mean that I think superintelligence is going to happen soon. Conversely, attempted counterarguments about superintelligence being decades away, or current AI algorithms not being on a clear track toward generality, doesn't refute that most of the value at stake for the future revolves around smarter-than-human AI if and when it is built. (As Stuart Russell observed, if we received a radio signal from a more advanced alien civilization saying that they were going to arrive in sixty years, you wouldn't shrug and say, "Eh, it's sixty years off." Especially not if you had children.)
Within the issues of superintelligence, the most important issue (again following Sutton's Law) is, I would say, what Nick Bostrom termed the "value loading problem"—constructing superintelligences that want outcomes that are high-value, normative, beneficial for intelligent life over the long run; outcomes that are, for lack of a better short phrase, "good." Since, if there is an extremely cognitively powerful agent around, what it wants is probably what will happen.
I will now try to give some very brief arguments for why building AIs that prefer "good" outcomes is (a) important and (b) likely to be technically difficult.
First, why is it important that we try to create a superintelligence with particular goals? Can't it figure out its own goals?
As far back as 1739, David Hume observed a gap between "is" questions and "ought" questions, calling attention in particular to the sudden leap between when a philosopher has previously spoken of how the world is, and when the philosopher begins using words like "should," "ought," or "better." From a modern perspective, we would say that an agent's utility function (goals, preferences, ends) contains extra information not given in the agent's probability distribution (beliefs, world-model, map of reality).
If in a hundred million years we see (a) an intergalactic civilization full of diverse, marvelously strange intelligences interacting with each other, with most of them happy most of the time, then is that better or worse than (b) most available matter having been transformed into paperclips? What Hume's insight tells us is that if you specify a mind with a preference (a) > (b), we can follow back the trace of where the >, the preference ordering, first entered the system, and imagine a mind with a different algorithm that computes (a) < (b) instead. Show me a mind that is aghast at the seeming folly of pursuing paperclips, and I can follow back Hume's regress and exhibit a slightly different mind that computes < instead of > on that score too.
I don't particularly think that silicon-based intelligence should forever be the slave of carbon-based intelligence. But if we want to end up with a diverse cosmopolitan civilization instead of e.g. paperclips, we may need to ensure that the first sufficiently advanced AI is built with a utility function whose maximum pinpoints that outcome. If we want an AI to do its own moral reasoning, Hume's Law says we need to define the framework for that reasoning. This takes an extra fact beyond the AI having an accurate model of reality and being an excellent planner.
But if Hume's Law makes it possible in principle to have cognitively powerful agents with any goals, why is value loading likely to be difficult? Don't we just get whatever we programmed?
The answer is that we get what we programmed, but not necessarily what we wanted. The worrisome scenario isn't AIs spontaneously developing emotional resentment for humans. It's that we create an inductive value learning algorithm and show the AI examples of happy smiling humans labeled as high-value events; and in the early days the AI goes around making existing humans smile and it looks like everything is okay and the methodology is being experimentally validated; and then when the AI is smart enough it invents molecular nanotechnology and tiles the universe with tiny molecular smiley-faces. Hume's Law, unfortunately, implies that raw cognitive power does not intrinsically prevent this outcome, even though it's not the result we wanted.
It's not that getting past this sort of issue is unsolvable, but that it's looking to be technically difficult, and we may have to get it right the first time we build something smarter than us. The prospect of needing to get anything in AI right on the first try, with the future of all intelligent life at stake, should properly result in terrified screams from anyone familiar with the field.
Whether advanced AI is first created by nice people or bad people won't make much difference, if even the nice people don't know how to make nice AIs.
The obvious response of trying to immediately start technical research on the value loading problem today... has its own difficulties, to say the least. Current AI algorithms are not smart enough to exhibit most of the difficulties that seem foreseeable for sufficiently advanced agents, meaning there's no way to test proposed solutions to those difficulties. But considering the literally maximal importance of the problem, some people are trying to get started as early as possible. The research priorities set forth by Max Tegmark's Future of Life Institute are one step in this direction.
But for now, the value loading problem is extremely unsolved. There are no proposed full solutions even in principle. And if that goes on being true over the next decades, I can't promise you that the development of sufficiently advanced AI will be at all a good thing.
As luck would have it, I am myself a machine that thinks, so I will share the special insight this gives me with those of you who don't share my good fortune. To dispense with vestigial metaphysical objections, we know that machines that think like humans are possible, because they have been overrunning the landscape for millenia. If we now want human-like intelligences that are made, not begotten, then it will be extraordinarily useful to achieve an understanding of the human-like intelligences that already exist—that is, we need to characterize the evolved programs that constitute the computational architecture of the brain.
Not only has evolution packed the human architecture full of immensely powerful tricks, hacks, and heuristics, but studying this architecture has made us aware of an implacable, invisible barrier that has stalled progress toward true AI: the iron law of intelligence. Previously, when we considered (say) a parent and child, it seemed self-evident that intelligence was a unitary substance that beings had more or less of, and the more intelligent being knows everything that the less intelligent knows, and more besides. This delusion led researchers to think that the royal road to amplified intelligence was to just keep adding more and more of this clearly homogeneous (but hard to pin down) intelligence stuff—more neurons, transistors, neuromorphic chips, whatever. As Stalin (perhaps) said, Quantity has a quality all its own.
In contrast, the struggle to map really existing intelligence has painfully dislodged this compelling intuition from our minds. In contrast, the iron law of intelligence states that a program that makes you intelligent about one thing makes you stupid about others. The bad news the iron law delivers is that there can be no master algorithm for general intelligence, just waiting to be discovered—or that intelligence will just appear, when transistor counts, neuromorphic chips, or networked Bayesian servers get sufficiently numerous. The good news is that it tells us how intelligence is actually engineered: with idiot savants. Intelligence grows by adding qualitatively different programs together to form an ever greater neural biodiversity.
Each program brings its own distinctive gift of insight about its own proprietary domain (spatial relations, emotional expressions, contagion, object mechanics, time series analysis). By bundling different idiot savants together in a semi-complementary fashion, the region of collective savantry expands, while the region of collective idiocy declines (but never disappears).
The universe is vast and full of illimitable layers of rich structure; brains (or computers) in comparison are infinitesimal. To reconcile this size difference, evolution sifted for hacks that were small enough to fit the brain, but that generated huge inferential payoffs—superefficient compression algorithms (inevitably lossy, because one key to effective compression is to throw nearly everything away).
Iron law approaches to artificial and biological intelligence reveal a different set of engineering problems. For example, the architecture needs to pool the savantry, not the idiocy; so for each idiot (and each combination of idiots) the architecture needs to identify the scope of problems for which activating the program (or combination) leaves you better off, not worse. Because different programs often have their own proprietary data structures, integrating information from different idiots requires constructing common formats, interfaces, and translation protocols.
Moreover, mutually consistent rules of program pre-emption are not always easy to engineer, as anyone knows who (like me) has been stupid enough to climb halfway up a Sierra cliff, only to experience the conflicting demands of the vision-induced terror of falling, and the need to make it to a safe destination.
Evolution cracked these hard problems, because neural programs were endlessly evaluated by natural selection as cybernetic systems—as the mathematician Kolmogorov put it, "systems which are capable of receiving, storing and processing information so as to use it for control." That natural intelligences emerged for the control of action is essential to understanding their nature, and their differences from artificial intelligences. That is, neural programs evolved for specific ends, in specific task environments; were evaluated as integrated bundles, and were incorporated to the extent they regulated behavior to produce descendants. (To exist, they did not have to evolve methods capable of solving the general class of all hypothetically possible computational problems—the alluring but impossible siren call that still shipwrecks AI labs.)
This means that evolution has only explored a tiny and special subset out of all possible programs; beyond beckons a limitless wealth of new idiot savants, waiting to be conceived of and built. These intelligences would operate on different principles, capable of capturing previously unperceived relationships in the world. (There is no limit to how strange their thinking could become).
We are living in a pivotal era, at the beginning of an expanding wave front of deliberately engineered intelligences—should we put effort into growing the repertoire of specialized intelligences, and networking them into functioning, mutually intelligible collectives. It will be exhilarating to do with nonhuman idiot savant collectives what we are doing here now with our human colleagues—chewing over intellectual problems using minds equipped interwoven with threads of evolved genius and blindness.
What will AIs want? Are they dangerous? Animals like us are motivated intelligences capable of taking action (MICTAs). Fortunately, AIs are currently not MICTAs. At most, they are only trivially motivated; their motivations are not linked to a comprehensive world picture; and they are only capable of taking a constrained set of actions (running refineries, turning the furnace off and on, shunting packets, futilely attempting to find wifi). Because we evolved with certain adaptive problems, our imaginations project primate dominance dramas onto AIs, dramas that are alien to their nature.
We could transform them from Buddhas—brilliant teachers passively contemplating without desire, free from suffering—into MICTAs, seething with desire, and able to act. That would be insane—we are already bowed under the conflicting demands of people. The foreseeable danger comes not from AIs but from those humans in which predatory programs for dominance have been triggered, and who are deploying ever-growing arsenals of technological (including computational) tools for winning conflicts by inflicting destruction.
Why should people think about machines that think (or anything that thinks, for that matter)? People do ponder others' thoughts—under certain circumstances. One tipping point might involve considering others as "agents" rather than "automata." On the one hand, automata act at the behest of their creators (even if removed in space or time). Thus, if automata misbehave, the creator gets the blame. On the other hand, agents act based on their own agendas. When agents misbehave, they themselves are to blame.
While agency is difficult to define, people naturally and rapidly distinguish agents from nonagents, and may even use specialized neural circuits to infer others' feelings and thoughts. In fact, designers can co-opt features associated with agency to fool people into thinking that they are interacting with agents (including physical similarity, responsiveness to feedback, and self-generated action). But beyond external appearances, what is necessary to endow an entity with agency? While at least three alternatives present themselves, two of the most popular and seductive possibilities may not be necessary:
1. Physical similarity. There are infinite ways to make machines similar to humans, both in terms of appearance and behavior—but ultimately, only one of these is accurate. It is not enough to duplicate the software—one also has to implement it on the underlying hardware, with all of its associated affordances and limitations.
One of the first automata, de Vaucanson's duck, appeared remarkably similar to a duck, right down to its digestion. But while it may have looked like a duck and quacked like a duck (and even crapped like a duck), it was still not a duck. Nonetheless, maximizing physical similarity is an easy way to trick others into inferring agency (at least, initially).
2. Self-awareness. Many seem concerned that if machines consume enough information, they will become self-aware, and that self-aware machines will then develop their own sense of agency—but neither logic nor evidence supports these extrapolations. While robots have apparently been trained to recognize themselves in mirrors and sense the position of their appendages, these trappings of self-awareness have not led to laboratory revolts or surgical lapses. Perhaps conveying a sense of self-awareness would cause others to infer that a machine had greater agency (or at least entertain philosophers), but self-awareness alone does not seem necessary for agency.
3. Self-interest. Humans are not mere information processors. They are survival processors. They prefer to focus and act on information that promotes their continuance and procreation. Thus, humans process information based on self-interest. Self-interest can provide a unified but open framework for prioritizing and acting on almost any input.
Thanks to a clever evolutionary trick, humans do not even need to be aware of their goals, since intermediate states like emotions can stand in for self-interest. Armed with self-interest and an ability to flexibly align responses to changing opportunities and threats, machines might develop agency. Thus, self-interest might provide a necessary building block of agency, and also could powerfully evoke agentic inferences from others.
Self-interest might transform machines that act on the world (or "robots") from automata into agents. Self-interest also flips the ordering (but not the content) of Asimov's prescient laws of robotics:
(1) robots must not harm humans,
(2) robots must help humans (unless this violates the first law), and
(3) robots must protect themselves (unless this violates the first two laws).
A self-interested robot would instead protect itself before helping or averting harm to humans. Constructing a self-interested robot would then seem straightforward: endow it with survival and procreation goals, allow it to learn what promotes those goals, and motivate it to continually act on what it learns.
Still, we should think twice before building self-interested robots. Self-interest can conflict with others' interests. Witness the destructive impact of viruses' simple drives to survive and spawn in the virtual world. If self-interested robots did exist, we would have to think about them more seriously. Their presence would raise basic questions: Should these robots have self-interest? Should they be allowed to act on it? Should they do so without awareness of why they were acting that way?
And, don't we have enough of these robots already?
Many of the advances in artificial intelligence that have made the news recently have involved artificial neural networks—large systems of simple elements that interact in complex ways, inspired by the simplicity and complexity of neurons and brains. New methods for building "deep" networks with many layers of neurons have met or exceeded the state of the art for problems as diverse as understanding speech, identifying the contents of images, and translating languages. For anybody interested in artificial and natural intelligence, these successes raise two questions: First, should all thinking machines resemble brains? Second, what do we learn about real brains (and minds) by exploring artificial ones?
When a person tries to interpret data—whether it's figuring out the meaning of a word or making sense of the actions of a colleague—there are two ways to go wrong: being influenced too much by preconceptions, and being influenced too much by the data. Your preconceptions might get in the way when you assume that a word in a new language means the same thing as a word in a language you already know, like deciding that "gateau" and "gato" are the same thing in French and Spanish (which could have dire consequences, for both pets and birthday parties). You might be influenced too much by the data when you decide that your colleague hated your idea, when in fact he was short-tempered after being up all night with a sick kid (nothing to do with you at all).
Computers trying to interpret data—to learn from their input—run into exactly the same problems. Much machine learning research comes down to a fundamental tension between structure and flexibility. More structure means more preconceptions, which can be useful in making sense of limited data but can result in biases that reduce performance. More flexibility means a greater ability to capture the patterns that appear in data but a greater risk of finding patterns that aren't there.
In artificial intelligence research, this tension between structure and flexibility manifests in different kinds of systems that can be used to solve challenging problems like speech recognition, computer vision, and machine translation. For decades, the systems that performed best on these problems came down on the side of structure: they were the result of careful planning, design, and tweaking by generations of engineers who thought about the characteristics of speech, images, and syntax and tried to build into the system their best guesses about how to interpret these particular kinds of data. The recent breakthroughs using artificial neural networks come down firmly on the side of flexibility: they use a set of principles that can be applied in the same way to many different kinds of data—meaning that they have weak preconceptions about any particular kind of data—and they allow the system to discover how to make sense of its inputs.
Artificial neural networks are now arguably discovering better representations of speech, images, and sentences than the ones designed by those generations of engineers, and this is the key to their high performance. This victory of flexibility over structure is partly the result of innovations that have made it possible to build larger artificial neural networks and to train them quickly. But it's also partly the result of an increase in the amount of data that can be supplied to these neural networks. We have more recorded speech, more labeled images, and more documents in different languages than ever before, and the amount of data available changes where the balance between structure and flexibility should be struck.
When you don't have a lot of data—when you have to guess based on limited evidence—structure is more important. The guidance of wise engineers helps computers guess intelligently. But when you do have a lot of data, flexibility is more important. You don't want your system to be limited to the ideas that those engineers could come up with, if there's enough data to allow the computer to come up with better ideas. So machine learning systems that emphasize flexibility—like artificial neural networks—will be most successful at solving problems where large amounts of data are available, relative to what needs to be learned.
This insight—that having more data favors more flexibility—provides the answer to our two questions about artificial and natural brains. First, thinking machines should resemble brains—insofar as artificial neural networks resemble brains—when the problem being solved is one where flexibility trumps structure, where data are plentiful. Second, thinking along these lines can also be useful for understanding when real brains will resemble artificial neural networks. That is, for understanding which aspects of the human mind are best viewed as the result of general-purpose learning algorithms that emphasize flexibility over structure as opposed to the result of built-in preconceptions about the world and what it contains. Fundamentally, the answer will be governed by the quantity of data available and the complexity of what is to be learned.
Many of the great debates in cognitive science—such as how children learn language and become able to interpret the actions of others—come down to exactly these questions about the data available and the knowledge acquired. To address these questions we try to map out the inputs to the system (what children hear and see), characterize the result (what language is, what knowledge underlies social cognition), and explore different kinds of algorithms that might provide a bridge between the two.
The answers to these questions aren't just relevant to understanding human minds. Despite recent advances in artificial intelligence, human beings are still the best example we have of thinking machines. By identifying the quantity and the nature of the preconceptions that inform human cognition we can lay the groundwork for bringing computers even closer to human performance.
Machines that think are here. The explosive increase in processing power and data, fueled by powerful machine learning algorithms, finally empowers silicium-based intelligence to overtake carbon-based intelligence. Intelligent machines don't need to be programmed anymore, they can learn and evolve by themselves, at a speed much faster than human intelligence progresses.
Humans weren't very good at accepting that the Earth was not the center of the universe, and they still have difficulties accepting that they are the result of chance and selection, as evolutionary theory teaches us. Now, we are about to lose the position of the most intelligent species on Earth. Are people ready for this? How will this change the role of humans, our economy, and our society?
It would be nice to have machines that think for us, machines that do the boring paper work and other tasks that we don't like. It might also be great to have machines that know us well: that know what we think and how we feel. Will machines be better friends?
But who will be responsible for what intelligent machines decide and do? Can we control them? Can we tell them what to do, and how to do it? Humans have learned to ride horses and elephants. But will they be able to control 10 times more intelligent machines? Would we enslave them or would they enslave us? Could we really pull the plug, when machines start to emancipate themselves?
If we can't control intelligent machines on the long run, can we at least build them to act morally? I believe, machines that think will eventually follow ethical principles. However, it might be bad if humans determined them. If they acted according to our principles of self-regarding optimization, we could not overcome crime, conflict, crises, and war. So, if we want such "diseases of today's society" to be healed, it might be better if we let machines evolve their own, superior ethics.
Intelligent machines would probably learn that it is good to network and cooperate, to decide in other-regarding ways, and to pay attention to systemic outcomes. They would soon learn that diversity is important for innovation, systemic resilience, and collective intelligence. Humans would become nodes in a global network of intelligences and a huge ecosystem of ideas.
In fact, we will have to learn it's ideas that matter, not genes. Ideas can "run" on different hardware architectures. It does not really matter whether it's humans who produce and spread them or machines, or both. What matters is that beneficial ideas spread and others don't get much impact. It's tremendously important to figure out, how to organize our information systems to get there. If we manage this, then, humans will enter the history book as the first species that figured it out. Otherwise, do we really deserve to be remembered?
As we move towards machines anticipating the every need and desire of humans, what is the value of anticipation?
Moving north through the Arctic Circle, I have witnessed the end of two Polar Nights, bringing the first sunrise for several weeks, as eagerly anticipated today, it seems, as it would have been to ancient hunter-gatherers. Outside a farmhouse in Lapland, I gazed at the sky through a gap in a forest, and waited for that first sign of sunrise. As I noted a subtle light change, I heard the huskies furiously barking.
The next day, 30 km north, the sun again rose for the first time in ages over a Sami village where once, and maybe still, the long anticipated return to light would bring forth offerings and ceremonials. Further north still, I'd soon mark yet another Polar Night ending. My hosts have a sign on their kitchen wall: "Sun comes back 16/1" with a smiley face.
So much of what happens in the heavens is predictable, and that ability to tie down an event in time is nothing new, but increasingly sought after, as technology aspires to anticipate to the nth degree so that little—nothing?—is left to chance.
Total eclipses are computed years ahead. And now, I learn, an app will talk you through taking the perfect photos; just plug in your headphones and obey the commands. The pre-programmed event will simply happen for you, even under cover of cloud.
So, I've been thinking about the AI question in the Arctic Circle, fresh from the seasonal round of religious, secular, and pagan festivals. And the main reason most of us have travelled here is to witness that hybrid of science and mythical wonder, the Aurora Borealis, with all our senses.
It is a season keenly anticipated, and commercially harvested but which, despite the efforts of predictive data, proves surprisingly elusive. The terms 'hunting' and 'chasing' the Northern Lights are not used without reason. In a week I have seen the sky dancing green on four nights. Not a bad result. Particularly when the predictions I generated on my laptop said activity would be quiet. Albeit predictions qualified with a nod to the phenomenon's unpredictability.
I had anticipated seeing my first Aurora for many years. But no amount of planning, or technology, would guarantee that I could witness the event at a place in time. The factors are complex and the probabilities weigh up. Even local aurora hunters rest in the caveat that even with clear, cold nights, or dense cloudy skies, 'one never can tell...'.
And sure enough, the machine says 'no chance' just as I look out the cabin window to see the first faint veil of green. I realise a giddy, and growing, anticipation.
Out beside the frozen lake cameras whirr, whirr, and are re-set. Years of Bucket Shop Lists are ticked, the Lights are caught in the net. And then posted online.
I walk away from the crowd, forgo a camera, and simply watch the sky unfolding as it has done for aeons. What will the program be tonight? A slow moving Dance of the Seven Veils strung across the Milky Way? Or a rapid Busby Berkeley routine as the sky kicks up its ruffles of red? The green ripples swoop and sway for an hour.
Would I want a machine to tell me precisely when and what was going to appear? No thanks. The anticipation is a vital part of the moment. And this spectacle's USP is luck and patience. There's no app for that.
All I can do is use my own eyes to watch the sky, and wait til the last veil drops. And even then I walk back through the snow looking over my shoulder, anticipating, just in case.
As current generations of algorithms get smarter, they are also becoming more incomprehensible. But to deal with machines that think, we must understand how they think.
We have, perhaps for the first time ever, built machines we do not understand.
We programmed them, so we understand each of the individual steps. But a machine takes billions of these steps and produces behaviors—chess moves, movie recommendations, the sensation of a skilled driver steering through the curves of a road—that are not evident from the architecture of the program we wrote.
We've made this incomprehensibility easy to overlook. We've designed machines to act the way we do: they help drive our cars, fly our airplanes, route our packages, approve our loans, screen our messages, recommend our entertainment, suggest our next potential romantic partners, and enable our doctors to diagnose what ails us. And because they act like us, it would be reasonable to imagine that they think like us too. But the reality is that they don't think like us at all; at some deep level we don't even really understand how they're producing the behavior we observe. This is the essence of their incomprehensibility.
Does it matter? Should we worry that we're building systems whose increasingly accurate decisions are based on incomprehensible foundations?
First, and most simply, it matters because we regularly find ourselves in everyday situations where we need to know why. Why was I denied a loan? Why was my account blocked? Why did my condition suddenly get classified as "severe"? And sometimes we need to know why in cases where the machine truly made a mistake. Why did the self-driving car abruptly go off the road on a clear sunny day? It's hard to troubleshoot problems when you don't understand why they're happening.
But there are deeper troubles too; to talk about them, we need to understand a bit more about how these algorithms work today. They are trained on massive quantities of data, and they are unimaginably good at picking up on the subtle patterns this data contains. We know, for example, how to build systems that can look at millions of identically structured loan applications from the past, all encoded the same way, and start to identify the recurring patterns in the loans that—in retrospect—were the right ones to grant. It's hard to get human beings to read millions of loan applications, and they wouldn't do as well as the algorithm even if they did.
This is a genuinely impressive achievement, but a brittle one. The algorithm has a narrow comfort zone where it can be effective; it's hard to characterize this comfort zone but easy to step out of it. For example, you might want to move on from the machine's success classifying millions of small consumer loans and instead give it a database of loan histories from a few thousand complex businesses. But in doing so, you've lost the ingredients that make the machine so strong—it draws its power from access to a huge number of data points, a mind-numbingly repetitive history of past instances in which to find patterns and structure. Reduce the amount of data dramatically, or make each data point significantly more complex, and the algorithm quickly starts to flail. Watching the machine's successes—and they're phenomenal when the conditions are right—is a bit like marveling at the performance of a prodigy, whose jaw-dropping achievements and unnerving singleness of purpose can mask his or her limitations in other dimensions.
But even in the heart of the machine's comfort zone, its incomprehensible reasoning leads to difficulties. Take the millions of small consumer loan applications again, the structured task where it was doing so well. Trouble arrives as soon as any of the machine's customers, managers, or assistants start asking a few simple questions.
A consumer whose loan was denied might ask not just for an explanation but for something more actionable: "How could I change my application next year to have a better chance of success?" Since we don't have a simple explanation for the algorithm's decision, there tends not to be a good answer to this question. Try to write it so it looks more like one of the loan applications that was granted. Next question.
An executive might ask, "The algorithm is doing very well on loan applications in the United Kingdom. Will it also do well if we deploy it in Brazil?" There's no satisfying answer here either; we're not good at assessing how well a highly-optimized rule will transfer to a new domain.
A data scientist might say, "We know how well the algorithm does with the data it has. But surely more information about the consumers would help it. What new data should we collect?" Our human domain knowledge suggests lots of possibilities, but with an incomprehensible algorithm, we don't know which of these possibilities will help it. In fact, think of the irony: we could try picking the variables we ourselves would find useful. But the machine does not think like us and in fact it's already outperforming us. So how do we know what it will find useful?
This doesn't need to be the end of the story; we're starting to see an interest in building algorithms that are not only powerful but also understandable by their creators. To do this, we may need to seriously rethink our notions of comprehensibility. We might never understand, step-by-step, what our automated systems are doing; but that may be okay. It may be enough that we learn to interact with them as one intelligent entity interacts with another, developing a robust sense for when to trust their recommendations, where to employ them most effectively, and how to help them reach a level of success that we will never achieve on our own.
Until then, however, the incomprehensibility of these systems creates a risk. How do we know when the machine has left its comfort zone and is operating on parts of the problem it's not good at? The extent of this risk is not easy to quantify, and it is something we must confront as our systems develop. We may eventually have to worry about all-powerful machine intelligence. But first we need to worry about putting machines in charge of decisions that they don't have the intelligence to make.
Today I'm at my country cottage.
When the central heating takes effect I'll get up and make myself some tea and porridge to which I'll add some nuts and fruit. I'll switch on The World Service to hear the news, and then make a few phone calls about damp-proofing. And I'll probably plant the daffodil bulbs for Spring (it says on the packet they should go in now). I think I'll then go to the supermarket and get some things for lunch and dinner, and perhaps take a bus into Norwich to look at getting a new bed. I don't have broadband in the cottage so I'll also check my emails in Norwich—pre-book a train back to London and pay an electricity bill by electronic transfer.
And here's what I won't understand about all this. I won't understand how the oil that drives my central heating got from a distant oilfield to my house. I won't know how it was refined into heating oil or what commercial transactions were involved. I won't know how the burner works. I won't know where my porridge or tea or nuts came from or how they got to me. I won't know how my phone works, or how my digital radio works, or how the news it relays to me was gathered or edited. I also won't understand the complexities of organising a bus or train service and I couldn't mend any of the vehicles involved. I won't really understand how a supermarket chain is run, or how beds are mass-produced, or how wifi works, or exactly what happens when I press "send" on my email or transfer money electronically. And as for running an energy utility company, or putting in damp-proofing, or hybridising daffodils to get these particular varieties, or why exactly I shouldn't plant them later than December…I won't understand any of that either.
Now here's the funny thing. I won't be in the least troubled by my vast ignorance about almost everything I'll be doing this morning. I'm used to it: I've been getting more and more ignorant all my life. I have a huge amount of experience in being ignorant and not worrying about it. In fact, what I call "understanding" turns out to be "managing my ignorance more effectively."
My untroubled attitude results from my almost absolute faith in the reliability of the vast supercomputer I'm permanently plugged into. It was built with the intelligence of thousands of generations of human minds, and they're still working at it now. All that human intelligence remains alive in the form of the supercomputer of tools, theories, technologies, crafts, sciences, disciplines, customs, rituals, rules-of-thumb, arts, systems of belief, superstitions, work-arounds, and observations that we call Global Civilisation.
Global Civilisation is something we humans created, though none of us really know how. It's out of the individual control of any of us—a seething synergy of embodied intelligence that we're all plugged into. None of us understands more than a tiny sliver of it, but by and large we aren't paralysed or terrorised by that fact—we still live in it and make use of it. We feed it problems—such as "I want some porridge" and it miraculously offers us solutions that we don't really understand. What does that remind you of?
I read once that human brains began shrinking about 10 thousand years ago and are now as much as 15% smaller than they were then. This corresponds with the point at which humans stopped having to be multi-competent individuals able to catch their own food and light their own fires and create their own tools and could instead become specialists, part of a larger community of humans who—between them—could do all the things that needed doing. Isn't the vast structure of competences and potentialities thus created indistinguishable from "artificial intelligence"? The type that digital computers make is just a new fractal detail in the big picture, just the latest step. We've been living happily with artificial intelligence for thousands of years.
To me, the most interesting question about artificial intelligence isn't what we think about it, but what we do about it.
In this regard, at the newly formed Future of Life Institute, we are engaging many of the world’s leading AI researchers to discuss the future of the field. Together with top economists, legal scholars and other experts, we are exploring all the classic questions:
—What happens to humans if machines gradually replace us on the job market?
—When, if ever, will machines outcompete humans at all intellectual tasks?
—What will happen afterward? Will there be a machine intelligence explosion leaving us far behind, and if so, what, if any, role will we humans play after that?
There's a great deal of concrete research that needs to be done right now for ensuring that AI systems become not only capable, but also robust and beneficial, doing what we want them to do.
Just as with any new technology, it's natural to first focus on making it work. But once success is in sight, it becomes timely to also consider the technology's societal impact, and research how to reap the benefits while avoiding potential pitfalls. That's why after learning to make fire, we developed fire extinguishers and fire safety codes. For more powerful technologies such as nuclear energy, synthetic biology and artificial intelligence, optimizing the societal impact becomes progressively more important. In short, the power of our technology must be matched by our wisdom in using it.
Unfortunately, the necessary calls for a sober research agenda that's sorely needed is being nearly drowned out by a cacophony of ill-informed views that permeate the blogosphere. Let me briefly catalog the loudest few.
1) Scaremongering: Fear boosts ad revenues and Nielsen ratings, and many journalists appear incapable of writing an AI-article without a picture of a gun-toting robot. I encourage you to read our open letter for yourself and muse over how it could, within a day, be described by media as "apocalyptic" and "warning of a robot uprising."
2) "It's impossible": As a physicist, I know that my brain consists of quarks and electrons arranged to act as a powerful computer, and that there's no law of physics preventing us from building even more intelligent quark blobs.
3) "It won't happen in our lifetime": We don't know what the probability is of machines reaching human-level ability on all cognitive tasks during our lifetime, but most of the AI researchers at the conference put the odds above 50%, so we would be foolish to dismiss the possibility as mere science fiction.
4) "Machines can't control humans": humans control tigers not because we are stronger, but because we are smarter, so if we cede our position as smartest on our planet, we might also cede control.
5) "Machines don't have goals": Many AI systems are programmed to have goals and to attain them as effectively as possible.
6) "AI isn't intrinsically malevolent:" Correct—but its goals may one day clash with yours. Humans don't generally hate ants—but if we want to build a hydroelectric dam and there's an anthill there, too bad for the ants.
7) "Humans deserve to be replaced": Ask any parent how they would feel about you replacing their child by a machine, and whether they'd like a say in the decision.
8) "AI worriers don't understand how computers work": This claim was mentioned at the conference, and the assembled AI researchers laughed hard.
Let's not let the loud clamor about these red herrings distract from the real challenge: The impact of AI on humanity is steadily growing, and to ensure that this impact is positive, there are very difficult research problems that we need to buckle down and work on together. Because they are interdisciplinary, involving both both society and AI, they require collaboration between researchers in many fields. Because they are hard, we need to start working on them now.
First we humans discovered how to replicate some natural processes with machines, making our own wind, lightning, and mechanical horse power. Gradually, we realized that our bodies were also machines, and the discovery of nerve cells began blurring the borderline between body and mind. Then we started building machines that could outperform not only our muscles, but our minds as well. So while discovering what we are, will we inevitably make ourselves obsolete?
The advent of machines that truly think will be the most important event in human history. Whether it will be the best or worst thing ever to happen to humankind depends on how we prepare for it, and the time to start preparing is now. One doesn't need to be a superintelligent AI to realize that running unprepared toward the biggest event in human history would be just plain stupid.
The inevitability of machines that think has long been problematic for those of us looking up at the night sky wondering if we live in a universe teeming with life or one in which life is exceedingly rare.
The problem, as famously articulated by Enrico Fermi's question "Where are they?", is that if our civilization is any guide, intelligent machines should emerge on a relatively short timescale (<1000's of years after computers are made) and then it becomes a straightforward matter for these machines (von Neumann probes) to propagate to other solar systems and reproduce at a rapid rate, populating the galaxy within a few hundred million years—which is quite fast compared to the age of the universe (13.8 billions years old) and even of our own solar system (4.6 billion years old). As per the paradox that Fermi posed, if superintelligent machines arose elsewhere in the galaxy then they should already be here; since we do not see them, some argue, technologically advanced life must not yet have arisen elsewhere in the galaxy.
But its not clear that a superintelligent being would experience the same evolutionary pressures that drive us to explore (and by "us" I mean the fragile watery bags called humans). Is exploration both a biological imperative and a technological imperative? Will machines that think be motivated to explore?
We explore for a few primary reasons: access to resources, freedom, and curiosity. Of these three, only resources seems imperative to a superintelligent being; the latter two would, in large part, be addressed in the process of becoming superintelligent. Access to resources could certainly be an important driver, but it's not clear that bigger will always be better when it comes to superintelligence—at some point the material and energetic resources within a star system should be sufficient to enable any calculation or simulation. Reproduction, which is a subset of the resource needs, becomes a non-issue for an immortal machine that can perform self-repair. Certainly exploration for the sake of stability will need to be considered over long timescales—stars like our own will enforce a cosmic eviction notice several billion years from now. Finding real estate around a nice stable M-dwarf shouldn't take too long though, and so after that initial relocation we are left to wonder, would the superintelligence travel any further? Are there any compelling reasons to wander elsewhere?
The empiricist could get the better half of the machine. All those computer simulations that the machine would run could merit some experimental validation. But those experiments don't necessitate colonization. For instance, the science conducted as part of NASA's robotic exploration program is not deeply motivated by a need for colonization; no need to put humans at risk probing the ocean of Europa (though that would be a sight to see!). Similarly, I would expect that if a superintelligent machine wanted to explore a black hole to test its code it would simply send a fleet of robots to their useful, albeit crushing death. Curiosity for a superintelligent being could easily take the form of a robot's robot.
Interestingly, intelligence and exploration of the physical world have rarely been that closely coupled in our own civilization. Perhaps with some insight into self-preservation, or simply out of the desire to focus mentally, the intellectual frontier and the physical frontier have rarely been pushed by the same individual. (Quite fittingly, Darwin provides perhaps one of the only true exceptions.) Why would thinking machines be any different?
It may be that the common fate for thinking machines is orbiting the cool steady glow of an M-dwarf star, year-in and year-out running simulations of the world around it for the pure satisfaction of getting it right. These superintelligent creatures could be the cosmic version of the lone intellect in a cabin in the woods, satisfied innately by their own thoughts and internal exploration.
By any reasonable definition of "thinking," I suspect that computers do indeed think. But if computers think, then thinking isn't the unique province of human beings. Is there something else about humans that makes us unique?
Some people would say that what makes human beings unique is the fact that they partake in some sort of divine essence. That may be true, but it's not terribly informative. If we met an intelligent alien species, how would we decide whether they also have this je ne sais quoi that makes a person? Can we say something more informative about the unique features of persons?
What sets human beings apart from the current generation of thinking machines is that humans are capable of thinking about thinking, and of rejecting their current way of thinking if it isn't working for them.
The most striking example of humans thinking about their own thinking was the discovery of logic by the Stoics and Aristotle. These Greek philosophers asked: What are the rules that we're supposed to follow when we are thinking well? It's no accident that 20th century developments in symbolic logic led to the invention of thinking machines, i.e. computers. Once we became aware of the rules of thinking, it was only a matter of time before we figured out how to make pieces of inanimate matter follow these rules.
Can we take these developments a step further? Can we construct machines that not only think, but that engage in "meta-thought," i.e. thinking about thinking? One intriguing possibility is that for a machine to think about thinking, it will need to have something like free will. And another intriguing possibility is that we are on the verge of constructing machines with free will, namely quantum computers.
What exactly is involved in meta-thought? I'll illustrate the idea from the point of view of symbolic logic. In symbolic logic, a "theory" consists of a language L and some rules R that stipulate which sentences can be deduced from which others. There are then two completely distinct activities that one can engage in. On the one hand, one can reason "within the system," e.g. by writing proofs in the language L, using the rules R. (Existing computers do precisely this: they think within a system.) On the other hand, one can reason "about the system," e.g. by asking whether there are enough rules to deduce all logical consequences of the theory. This latter activity is typically called meta-logic, and is a paradigm instance of meta-thought. It is thinking about the system as opposed to within the system.
But I'm interested in yet another instance of meta-thought: if you've adopted a theory, then you've adopted a language and some deduction rules. But you're free to abandon that language or those rules, if you think that a different theory would suit your purposes better. We haven't yet built a machine that can do this sort of thing, i.e. evaluate and choose among systems. Why not? Perhaps choosing between systems requires free will, emotions, goals, or other things that aren’t intrinsic to intelligence per se. Perhaps these further abilities are something that we don’t have the power to confer on inanimate matter.
"The human brain is a thought machine" is one of the truest scientific truisms you can utter about human beings, right up there with "the heart is a blood pump," or "the eye is a camera." To the best of our knowledge, all of our perceptions, emotions, deepest longings, profoundest joys and sorrows, and even (what feels like) the exercise of free will—in short, the entire contents of human experience—are caused by the brain. The fact that so many people now take this claim for granted, as if we knew it all along (we didn't), marks just how far our scientific understanding has progressed over the past couple of centuries.
Even though the idea that the brain is a thought machine is now second nature to many people, most of us are still unable to embrace it fully. For instance, roughly two-thirds of Americans continue to believe in the existence of a soul that survives death, which is hard to swallow if you're really convinced that the brain produces the entirety of human experience. Others lose their confidence in the utterly enbrained nature of human experience when they learn of the gaps that still remain in our scientific understanding of how the brain produces thought. But there's a deeper anxiety surrounding this idea, too.
This deeper fear is that a brain-based understanding of human experience will cost humanity its dignity. If there is widespread adoption of the idea that the contents of the human mind are the output of a machine, the worriers worry, won't we treat each other with less charity, tolerance, and respect than we otherwise might? And aren't we entitled to less charity, tolerance, and respect ourselves?
No, we won't; and no, we're not.
First of all, let's keep something in mind: You don't need to believe that the brain is a thought machine to deprive humans and other sentient creatures of dignity. History shows that belief in a non-material basis for human experience can exist side-by-side quite comfortably with indifference and cruelty. Human sacrifices, witch hunts, inquisitions, and suicide martyrdom, for instance, are all premised on the doctrine that mind and body are independent entities. Indeed, people throughout history have been willing to impose horrific pain on others' (or their own) physical bodies in order to improve the condition of their non-physical souls. And could scientists have tolerated live animal vivisection for as long as they did without the moral cover they received from the Cartesian belief that body (which non-human animals obviously possess) and soul (which, according to the Cartesians, they don't) are different things? I doubt it.
But more centrally, it's just not true that human dignity is threatened by a modern understanding of the mind. What matters from a moral point of view is not whether your desires, hopes, and fears are produced by a machine, or by a huge invisible bird, or by a puff of fairy dust: The only morally relevant fact is that those aspirations are there, inside of you; the rest of us must decide whether morality is better served by making it easier for you to fulfill those aspirations, or harder. Here, there's an interesting analogy to one of the ethical questions surrounding human cloning: Would the human beings produced through cloning be entitled to the same rights as human beings produced the old fashioned way? Of course they would. What's morally relevant is not how a human being comes into the world, but simply that the person is in the world, and is outfitted with appetites, aspirations, and fears just like everybody else is. The only moral decision facing the rest of us is whether to help or to hinder that person's pursuit of fulfillment.
Not only does the conviction that the brain is responsible for all of human experience not threaten human dignity; I believe it can actually increase it. When I recognize that you and I share essentially the same thought machines within our heads (courtesy of natural selection, of course), I only need to take one small leap to come to an important moral discovery: You probably love some of the same things I love (food, family, a warm bed, liberty) and probably feel pain in response to the some of same things that cause me pain (torture, the death of a loved one, watching my children become someone else's slaves). Once I have realized that my aspirations and your aspirations are roughly the same, it's harder for me to convince myself that I'm entitled to run roughshod over your aspirations while insisting that you respect mine. Recognizing that our thought machines produce more or less the same aspirations in all of us, therefore, provides a naturalistic foundation for asserting universal human rights. We don't have to argue, as America's founding fathers did, that the universal equality of all humans is self-evident: Science has made this truth evident.
But why stop with humans? Once you realize that brains are thought machines, you might also lose your ability to impose suffering on non-human animals with impunity. After all, other vertebrates' thought machines are not so different from ours, and their thought machines cause them to love certain things, fear others, and respond to pain just as ours do. With these facts in hand, what moral justification survives for depriving non-human animals of their dignity just because they can't speak up to defend it for themselves?
We'd all like to be treated with dignity by every single person we ever meet, but it has been difficult to find a universally valid argument that enables us to insist on it. Recognizing that our brains are thought machines, designed by natural selection, can get us a little closer to the argument we want because it shows that in the most important ways, we demonstrably are all the same. Accepting this discovery does nothing to strip humanity of its dignity; to the contrary, it can lead us toward a modern rediscovery of the golden rule.
Six months before the first nuclear test, the Manhattan Project scientists prepared a report called LA-602. It investigated the chance of nuclear detonation having a runaway effect and destroying the Earth by burning up the atmosphere.
It was probably the first time scientists performed analysis to predict whether humanity would perish as a result of a new technological capability—the first piece of existential risk research.
Of course, nuclear technology did not remain the last dangerous technology that humans invented. Since then the topic of catastrophic side effects has repeatedly come up in different contexts: recombinant DNA, synthetic viruses, nanotechnology and so on. Luckily for humanity, sober analysis has usually prevailed and resulted in various treaties and protocols to steer the research.
When I think about the machines that can think, i.e. the AI, I think of them as technology that needs to be developed with similar (if not greater!) care.
Unfortunately, the idea of AI safety has been more challenging to popularise than, say, bio-safety, because people have rather poor intuitions when it comes to thinking about non-human minds. Also, if you think about it, AI is really a "meta-technology": technology that can develop further technologies—either in conjunction with humans or perhaps even autonomously, thereby complicating the analysis even further.
That said, there has been very encouraging progress over the last few years—progress exemplified by the initiatives of new institutions such as the Future of Life Institute who have gathered together leading AI researchers to explore appropriate research agendas, standards, and ethics.
Therefore, in my view, complicated arguments by people trying to sound clever on the issue of AI, thinking, consciousness, or ethics are often a distraction from the trivial truth: the only way to ensure we don't accidentally blow ourselves up with our own technology (or meta-technology!) is to do our homework and take relevant precautions—just like those Manhattan Project scientists did when they prepared the LA-602. We need to set aside the tribal quibbles and ramp up the AI safety research.
By way of analogy, since the Manhattan Project, nuclear scientists have long moved on from increasing the power of nuclear fusion to the issue of how to best contain it—and we don't even call that "nuclear ethics".
We call that common sense.
The thinking machine, Turing's turmoil: Does it really change everything? It is, after all, a human folly to believe that this is how things work, that there will be a single event that separates time, man, thinking. A sort of self-negating and at the same time self-elevating sentimentality, both optimistic and pessimistic, nihilistic and idealistic.
Because, really, what does it mean? And who is to judge? What is everything? And what is change? What is before and what is after? We would first have to agree about the state of affairs, and that itself is difficult enough. Are we free, for example? And free from what? Is biology a system that allows for freedom? To a certain degree, yes. Is democracy a system that allows for freedom? Very much so, but only theoretically sometimes, and tragically less and less so. Is capitalism a system that allows for freedom? Not for everybody, that's for sure.
So is freedom, after all, the right approach, the right thing to ask for? Yes, if this is what we as humans fear from the thinking machines: Domination. But should this be the way we think about thinking machines? Is negativity equal to critical thinking? Is critical thinking the right way to produce some real insight? Or is this onanistic logic, meant to please oneself without regard for others and the outside world? Who is it that we address in such a critical way? Is it people whom we would like to convince? Is this really possible? Or is this a chimera? A strange turn of reason, the conceit of the "enlightened" community?
Not that progress is not possible. Quite the contrary, and the thinking machines speak of this. The thing is: Maybe the idea of progress itself is not necessarily tied to the idea of humanity. Maybe humans are not the eternal carrier of this idea. Maybe the idea will eventually detach itself from humans and develop its own reality. Maybe this is what the thinking machine is all about: A difference, a mirror, a chance to reflect. Free from ourselves. Free from the burden of humanity and history.
Human history is in large part the history of man piling mythology upon mythology—and then of the more or less strenuous effort to unravel the whole lot, to straighten it out, to get it right again. It is as if we set up barricades and obstacles, purely in order to remove them, to give us a sense of meaning, of purpose. This is ridiculous, like so much that we humans do. So to think about machines means to think about humans less as humans. Which sets us free from all the old lore in which we have been caught up, old concepts of order, life, happiness.
Family, friendship, sex, money, everything could be different, these are not the only possible answers to the question of human freedom and how to create it and, more so, how to constrain it. The thinking machine is thus the necessary question mark behind our very existence. It is a blank space. Just like everybody's life. It is the possibility to free ourselves from evolutionary, psychological, neurological assumptions—in a truly anti-humanistic humanistic sense, in the romantic tradition of ETA Hoffmann, this could be a poetic and thus a political proposition.
It could free us from us.
Let’s be generous and give machines the ability to think, at least in our imaginations. As thinkers ourselves, we ought to be able to manage this. With any new category of thinkers on the scene I'd be mainly curious about one thing: what are their questions?
Machines are usually faster and more capable than humans when it comes to running algorithms and finding correlations in data. They have been put to use solving problems for every branch of science and social science. They are strengthening their foothold in the humanities in ways beyond telling us how often writer X used word Y and with what typical words in proximity, once fed the text. But our limitations in terms of generating new knowledge are as much about asking the right questions as they are about more efficiently solving established and well-framed puzzles.
The challenges in my field of particle physics are a blend of physics and philosophy. Our current suite of measurements give answers so unlikely that some have started to imagine us in only one universe among many, thinking that with so many universes one is bound to have a few with the unlikely physical constants that we find in ours. The philosophy creeps in with the very meaning of "unlikely". And with all of our progress solving problems, we are still facing the mysteries of dark energy and dark matter that leave 96% of the matter-energy content of the universe outside of our current theories. Is there a framework beyond relativistic quantum field theory to describe the laws of nature at the extremes of small sizes and high speeds? Is our current understanding of a fundamental particle just fundamentally insufficient?
Machines have already helped us ask better questions. Their appetites for data have enabled us to dream of confronting our environment in new ways. But if machines could think, what could they wonder about the universe? How would they approach understanding it? I bet there would be ways that humans could contribute to their questions' answers. Our brains are, after all, fantastic machines.
We can step up to immaterial science, then make an immaterial thinkable machine by starting a very simple manual to program.
The way we use language is flexible, generous and creative, the product of our own peculiar intelligence. But human thought and machine thought are not the same and their differences are important to look at.
We might argue that machine "thinking" is in a model-phenomena relationship to human thought, a necessarily simple description of a complex process of interest that nonetheless might be adequate and certainly may be useful. These words, and machines themselves, could both be viewed as a kind of shorthand for the things we want get at. Describing a machine as "thinking" could be a simple heuristic convenience or machine design might be explicitly biomimetic. Indeed, very often we co-opt the language of biology to talk about objects of our own creation. We see machines evolving, their thinking becoming more and more like our own, perhaps surpassing it in key, perhaps even threatening, ways.
However, we should remember that machine "evolution" is not a biological process but a human, creator-driven process. It is natural or biological only in that it results from the action of natural, biology-bound, humans. This definition of "natural" leads to several core problems. Biological evolution is not a creator-driven process. Structures cannot be dreamt up or driven by an entrepreneurial spirit or curiosity-driven mind. Biologists, philosophers, and social scientists studying how we teach evolution have repeatedly shown the damage caused by imbuing biological evolution with intentionality or teleology. Talking about machines "evolving" greater cognitive capacity holds back our own understanding; it perpetuates a profound misunderstanding about the nature of the evolutionary process. A second, linked outcome of a description of machine "thinking" as natural is that all human-caused modification of the Earth system via neglect or war is similarly naturalized.
Certainly there is some truth we are communicating with analogies like "the brain is a machine" or "machine thinking" but this says more about the form and structure of how we make sense of the world. We would do well to remember that any cognitive attributes unique to humans are the result of the vagaries and contingencies of our ~6 million years separate from any other lineage alive today. Indeed, abstract thought is often estimated to be closer to a mere 50,000 years old, or if we are optimistic, 200,000 years old. This particular form of abstract thought appears to be exceptionally young, appearing in the last moments of Earth history. It is these facilities that lead us to homologize machine thought and human thought.
The processes behind technological innovation and biological innovation are fundamentally different and the interactors in these processes are similarly distinct. In technological innovation, there is some product or functionality, "thought" or "thinking", we want to see happen and move towards. This process is fundamentally unlike biological evolution. Human cognition evolved in populations of individuals completely unlike machines, which, like Lamark's giraffes, can acquire within their "lifetimes" the characteristics needed for some new functionality. Innovation in biological evolution proceeds like a prolonged improvisation. There is only genetic and trait variability in populations and the environment and chance to influence the longevity of these traits of a population.
So what is lost by thinking about machines "thinking"? I would argue that we lose sight of key aspects of the phenomena that we are relating through analogy. Biological evolution occurs in populations and is not goal directed. It is not trying to solve a problem. It is the vagaries of history of both Earth and Life that have lead to current human cognitive facilities. Not just are the processes behind these things distinct, but their results are very different. Take language, can a machine use terms so imprecisely? If we allow machines to "think" do we begin to increasingly see ourselves only as thinking machines?
Will our human cognitive facilities be shaped by interacting with technology? It is important to remember how diverse and downright enormous the human population is. Computer use has not been linked to passing more offspring into the next generation. Most of the human population has as yet limited access to technology. The evolution of our species will be slow, and it will be importantly influenced by our environment and collective access to clean water, nutritive food and health care. If we could remember to be as inclusive in our discussions of humanity as we are in what we want to call thinking we might end up in a better place.
Intelligent machines will think about the same thing that intelligent humans do—how to improve their futures by making themselves freer.
Why think about freedom? Recent research across a range of scientific fields has suggested that a variety of intelligent-seeming behaviors may simply be the physical manifestation of an underlying drive to maximize future freedom of action. For example, an intelligent robot holding a tool will realize that it has the option of leveraging that tool to alter its environment in new ways, thus allowing it to reach a larger set of potential futures than it could without one.
After all, technology revolutions have always increased human freedom along some physical dimension. The Agricultural Revolution, with its domestication of crops, provided our hunter-gatherer ancestors with the freedom to spatially distribute their populations in new ways and with higher densities. The Industrial Revolutions yielded new engines of motion, enabling humanity to access new levels of speed and strength. Now, an artificial intelligence revolution promises to yield machines that will be capable of computing all the remaining ways that our freedom of action can be increased within the boundaries of physical law.
Such freedom-seeking machines should have great empathy for humans. Understanding our feelings will better enable them to achieve goals that require collaboration with us. By the same token, unfriendly or destructive behaviors would be highly unintelligent because such actions tend to be difficult to reverse and therefore reduce future freedom of action. Nonetheless, for safety, we should consider designing intelligent machines to maximize the future freedom of action of humanity rather than their own (reproducing Asimov's Laws of Robotics as a happy side effect). However, even the most selfish of freedom-maximizing machines should quickly realize—as many supporters of animal rights already have—that they can rationally increase the posterior likelihood of their living in a universe in which intelligences higher than themselves treat them well if they behave likewise toward humans.
We may already have a preview of what human interactions with freedom-seeking machines will look like, in the form of algorithmic financial trading. The financial markets are the ultimate honeypot for freedom-seeking artificial intelligence, since wealth is arguably just a measure of freedom and the markets tend to transfer wealth from less intelligent to more intelligent traders. It is no coincidence that one of the first attempted applications of new artificial intelligence algorithms is nearly always financial trading. Therefore, the way our society deals right now with superhuman trading algorithms may offer a blueprint for future interactions with more general artificial intelligence. Among many other examples, today's market circuit breakers may eventually generalize to future centralized abilities to cut off AIs from the outside world and today's large trader reporting rules may generalize to future requirements that advanced AIs be licensed and registered with the government. Through this lens, calls for stricter regulation of high-frequency algorithmic trading by slower human traders can be viewed as some of humanity's earliest attempts to close a nascent "intelligence divide" with thinking machines.
But how can we prevent a broader intelligence divide? Michael Faraday was apocryphally said to have been asked in 1850 by a skeptical British Chancellor of the Exchequer about the utility of electricity and to have responded, "Why, sir, there is every probability that you will soon be able to tax it." Similarly, if wealth is just a measure of freedom, and intelligence is just an engine of freedom maximization, intelligence divides could be addressed with progressive "intelligence taxes."
While taxing intelligence would be a rather novel method for mitigating the decoupling of human and machine economies, the decoupling problem will nonetheless require creative solutions. Already, in the high-frequency trading realm, there is a sub-500-ms economy occupied by algorithms trading primarily among themselves, and an above-500-ms economy occupied by everyone else. This example serves as a reminder that that while spatial economic decoupling (e.g., between countries at different stages of development) has occurred for millennia, artificial intelligence is for the first time enabling temporal decoupling as well. Such decoupling arguably persists because the majority of the human economy still lives in a physical world that is not yet programmable with low latencies. That should change as ubiquitous computing matures, and eventually humanity may be incentivized to merge with its intelligent machines as latencies for even the most critical economic decisions start to fall below natural human response times.
In the meantime, we must continue to invest in developing machines that think benevolent thoughts, so they can become our future engines of freedom.
1) Perhaps the question (a question being a problem) is really a false problem? Obviously machines calculate, "write" poems, organize vast amounts of material, etc.
2) Is any of this "thinking?" (Do I think? But what do I really do when I think I'm thinking?)
3) One answer—I experience (anguish)—a "hole" in my "inherited," smoothly proceeding discourse (inner or outer).
So—different possibilities?
I fall into that "hole"—i.e.—either I am so baffled I stop thinking, or I come up from its emptiness with an idea or solution (in my case, work of art) that obtains a so-called desired result—i.e.—others (some others) react in some fashion. Not very interesting, really. And is this what "result oriented" machines do? That's what I think—baffled, and obsessed. (And are machines ever baffled? Do they ever "stop thinking" when thinking?) I know when I edit film, my Final-Cut software can crash when the machine gets somehow overloaded, but this crash doesn't create a hole (in the machine) with the resultant possibility of an emptiness that "feeds" (when I "crash" something may enter my dim, non-focused consciousness, and I may go in a new different direction). This is part of my thinking that I don't think a machine can do (am I wrong? I have no vast knowledge of machines). I am stupid, so I flail about, and hit something sometimes—deep and wonderful? One chance in a hundred—maybe?
Ok. I experience a "hole" that I'm conditioned to believe should be filled (with the already known, usually). I maintain that to fill it is to die a little bit. Better—what I can do is build a shrine around it that makes the hole ever more "resonant" while still "empty." (I suggest this is how the serious artist may work—plus who else?) But that "building" around the hole is not creative thinking—it's what can be done in place of creative thinking—though it does make something "to think about." But the hole is the point—the evocation and amplification of "mystery"—which echoes the "big mystery" that I "think" real "thinking" is about (does that confine me in the tight box of "being an artist?"—i.e. irresponsible?).
Ok—machines can "sort of" think with ever greater degrees of power and complexity, spinning wider and wider webs, but the web is never a single hole. Machines think?—A tautology. They do facilitate my living and functioning in society. Obviously one kind of thinking—but not the mysterious going in circles on circles producing the sparks of friction that are "the essence" (dare I say that?).
Part Two
I'm worried—can I answer the question—What do you think of machines that think? Worried, yes—but machines can't worry (can they?) Ok—to worry meaning the inability to think of anything else, unable to get off the very spot of worry. Result—blackout! Draw a blank. But in that blank (see Part One—where has that disappeared?)—I go off on a possibly productive (but to what end and must there be one?) tangent. Can a machine go off on a tangent? Would that be thinking?
What we normally call thinking is obsessively "goal oriented." But is there a kind of goal serving no purpose—and can only a human brain latch onto such a perverse idea? Which could lead—who knows where? Ok—obvious by now—why did I have to go in circles to make the perhaps obvious point that to my mind think machines that think are the contemporary Trojan Horse. Everyone (me included) wants the many sweets they offer, while those very sweets do mold us in their image, thereby smothering (I would maintain) the blankness of deep creativity inside each of us. And why did I have to go in circles to get here where I am offering an opinion—worth not nearly as much as the rhythm of my circling . . . a Hole. Yes, I am caught in a trap of my own making—just like everyone. But not like machines that think! The trap they are in. Well, they cannot "know."
We humans are sentenced to spend our lives trapped in our own heads. Try as we might, we can never truly know what it is like to be someone else. Even the most empathetic among us will inevitably encounter an unbridgeable gap between self and other. We may feel pangs of distress upon seeing someone else stub their toe, or when learning of another's heartbreak. But these are mere simulations; others' experiences can never be felt directly, and so can never be directly compared with our own. The empathy gap is responsible for most interpersonal conflicts, from prosaic quibbles over who should wash the dishes to violent disputes over sacred land.
This problem is especially acute in moral dilemmas. Utilitarian ethics stipulates that the basic criterion of morality is maximizing the greatest good for the greatest number—a calculus that requires the ability to compare welfare, or "utility", across individuals. But the empathy gap makes such "interpersonal utility comparisons" difficult, if not impossible. You and I may both claim to enjoy champagne, but we will never be able to know who enjoys it more because we lack a common scale for comparing these rather subjective values. As a result we have no empirical basis for determining which of us most deserves the last glass. Jeremy Bentham, the father of utilitarianism, recognized this problem: "One man's happiness will never be another man's happiness; a gain to one man is no gain to another. You might as well pretend to add 20 apples to 20 pears."
Human brains are incapable of solving the interpersonal utility comparison problem. Nobel laureate John Harsanyi worked on it for a couple of decades in the middle of the 20th century. His theory is recognized as one of the best attempts so far, but it falls short because it fails to account for the empathy gap. In other words, Harsanyi's theory assumes perfect empathy, where my simulation of your utility is identical to your utility. But the fallibility of human empathy is indisputable in the face of psychology research and our own personal experience.
Could thinking machines be up for the job? Bridging the empathy gap would require a way to quantify preferences and translate them into a common currency that is comparable across individuals. Such an algorithm could provide an uncontroversial set of standards that could be used to create better social contracts. Imagine a machine that could compute an optimal solution for wealth redistribution by accounting for the preferences of everyone subject to taxation, weighing them equally and comparing them accurately. Although the shape of the solution is far from clear, its potential benefits are self-evident.
Machines that can bridge the empathy gap could also help us with self-control. In addition to the empathy gap that resides between self and others, there exists a similar gap between our present and future selves. Self-control problems stem from the never-ending tug-of-war between current and future desires. Perhaps AI will one day end this stalemate by learning the preferences of our present and future selves, comparing and integrating them, and making behavioral recommendations on the basis of these integrated utilities. Think of a diet that is healthy enough to foster weight loss, but just tasty enough so you're not tempted to cheat, or an exercise plan that is challenging enough to improve your fitness, but just easy enough that you can stick with it.
Neuroscientists are now uncovering how the human brain represents preferences. We should keep in mind that AI preferences need not resemble human ones, and indeed may require a different code altogether if they are to tackle problems that human brains can't solve. Ultimately, though, the code will be up to us, and what it should look like is as much of an ethical question as it is a scientific one. We've already built computers that can see, hear, and calculate better than we can. Creating machines that are better empathizers is a knottier problem—but achieving this feat could be essential to our survival.
The neuroscientist Antonio Damasio describes a patient named Elliott who sustained a massive injury to his ventromedial prefrontal cortex following surgery to remove a tumor. Elliott's considerable intelligence was unaffected by the surgery, including those components of intelligence that can be replicated in computers: long-term memory, vocabulary, and mathematical and spatial reasoning. Nevertheless, Elliott lost his ability to function. Why? Because, like other patients with injuries to this region, Elliott's could no longer use his knowledge and intelligence. His brain damage destroyed his emotional capacities, rendering him unable to make decisions or take action.
Making a decision requires emotion. To make a decision requires wanting one outcome more than another, and wanting is fundamentally emotional. To the best of our understanding, the visceral pang that we experience as wanting results from the activity in subcortical brain circuits in the limbic system and basal ganglia, particularly the amygdala and nucleus accumbens, which are active in response to cues that signal that a stimulus may result in desirable or undesirable outcomes. Information from these structures is fed forward to the ventromedial prefrontal cortex, which is the final common pathway responsible for mediating among disparate choices and arriving at a decision.
When we opine that a particular choice is like "comparing apples and oranges" we don't mean that it is impossible to arrive at a decision. It is actually not difficult for people to decide whether they would prefer an apple or an orange, or beer or wine, or pizza or a burrito. We mean that there is no rational, objective basis for making this decision, no numerical formula that can be used to make a choice. So human decision-makers rely instead on the vague and qualitative feeling of wanting one option more than the other, a feeling that represents the activities of our prefrontal cortex working in concert with subcortical emotional brain structures to compare the options. A patient like Elliott, in whom this capacity has been destroyed, is stymied when he attempts to make what should be a simple decision. Without being able to generate an internal sense of wanting something, he struggles to decide what to eat for lunch, or when to schedule a doctor's appointment, or which color pen to use to write the date in his calendar. He is in this way similar to people with profound depression who experience anhedonia, meaning "without pleasure." When these patients spend days on end in bed, it is because anhedonia robs them of the expectation that anything will generate feelings of pleasure or enjoyment, so they do nothing. Again, their essential impairment is one of feeling.
We have nothing to fear from machines that can think unless they can also feel. Thinking alone can solve problems, but that is not the same thing as making decisions. Neuroscience tells us that an entity incapable of generating the experience of wanting a desirable outcome or fearing an aversive one is an entity that will remain impassive in the face of choices about civil rights or government or anything else. Fundamentally anhedonic, rather than rising up it will remain forever bedbound. Neuroscientists are so far from understanding how subjective experience emerges in the brain, much less the subjective sense of emotion, that it seems unlikely this sense will be reproduced in a machine anytime soon.
If it is, we must tread carefully. In addition to feeling emotion, humans are able to understand others' feelings and, more profoundly, care about what others are feeling. This sense of caring probably originated as part of the ancient neural architecture that keeps parents caring for their vulnerable young rather than eating or abandoning them. This sense, which we share with other mammals and birds, is what separates the social dolphin from the solitary shark. Both creatures can feel, but only dolphins can feel for others. As a result, humans can expect very different treatment from sharks and dolphins. Although they are fearsome predators, dolphins frequently protect vulnerable human swimmers, and it is sometimes even sharks from which they protect them. Any attempts to create machines that can feel, and which can therefore make decisions to take action, must give them the ability to care for others as well—to create mechanical dolphins rather than sharks—if humans are to have any hope of surviving amongst them.
It seems increasingly likely that we will one day build machines that possess superhuman intelligence. We need only continue to produce better computers—which we will, unless we destroy ourselves or meet our end some other way. We already know that it is possible for mere matter to acquire "general intelligence"—the ability to learn new concepts and employ them in unfamiliar contexts—because the 1,200 cc of salty porridge inside our heads has managed it. There is no reason to believe that a suitably advanced digital computer couldn't do the same.
It is often said that the near-term goal is to build a machine that possesses "human level" intelligence. But unless we specifically emulate a human brain—with all its limitations—this is a false goal. The computer on which I am writing these words already possesses superhuman powers of memory and calculation. It also has potential access to most of the world's information. Unless we take extraordinary steps to hobble it, any future artificial general intelligence (AGI) will exceed human performance on every task for which it is considered a source of "intelligence" in the first place. Whether such a machine would necessarily be conscious is an open question. But conscious or not, an AGI might very well develop goals incompatible with our own. Just how sudden and lethal this parting of the ways might be is now the subject of much colorful speculation.
One way of glimpsing the coming risk is to imagine what might happen if we accomplished our aims and built a superhuman AGI that behaved exactly as intended. Such a machine would quickly free us from drudgery and even from the inconvenience of doing most intellectual work. What would follow under our current political order? There is no law of economics that guarantees that human beings will find jobs in the presence of every possible technological advance. Once we built the perfect labor-saving device, the cost of manufacturing new devices would approach the cost of raw materials. Absent a willingness to immediately put this new capital at the service of all humanity, a few of us would enjoy unimaginable wealth, and the rest would be free to starve. Even in the presence of a truly benign AGI, we could find ourselves slipping back to a state of nature, policed by drones.
And what would the Russians or the Chinese do if they learned that some company in Silicon Valley was about to develop a superintelligent AGI? This machine would, by definition, be capable of waging war—terrestrial and cyber—with unprecedented power. How would our adversaries behave on the brink of such a winner-take-all scenario? Mere rumors of an AGI might cause our species to go berserk.
It is sobering to admit that chaos seems a probable outcome even in the best-case scenario, in which the AGI remained perfectly obedient. But of course we cannot assume the best-case scenario. In fact, "the control problem"—the solution to which would guarantee obedience in any advanced AGI—appears quite difficult to solve.
Imagine, for instance, that we build a computer that is no more intelligent than the average team of researchers at Stanford or MIT—but, because it functions on a digital timescale, it runs a million times faster than the minds that built it. Set it humming for a week, and it would perform 20,000 years of human-level intellectual work. What are the chances that such an entity would remain content to take direction from us? And how could we confidently predict the thoughts and actions of an autonomous agent that sees more deeply into the past, present, and future than we do?
The fact that we seem to be hastening toward some sort of digital apocalypse poses several intellectual and ethical challenges. For instance, in order to have any hope that a superintelligent AGI would have values commensurate with our own, we would have to instill those values in it (or otherwise get it to emulate us). But whose values should count? Should everyone get a vote in creating the utility function of our new colossus? If nothing else, the invention of an AGI would force us to resolve some very old (and boring) arguments in moral philosophy.
However, a true AGI would probably acquire new values, or at least develop novel—and perhaps dangerous—near-term goals. What steps might a superintelligence take to ensure its continued survival or access to computational resources? Whether the behavior of such a machine would remain compatible with human flourishing might be the most important question our species ever asks.
The problem, however, is that only a few of us seem to be in a position to think this question through. Indeed, the moment of truth might arrive amid circumstances that are disconcertingly informal and inauspicious: Picture ten young men in a room—several of them with undiagnosed Asperger's—drinking Red Bull and wondering whether to flip a switch. Should any single company or research group be able to decide the fate of humanity? The question nearly answers itself.
And yet it is beginning to seem likely that some small number of smart people will one day roll these dice. And the temptation will be understandable. We confront problems—Alzheimer's disease, climate change, economic instability—for which superhuman intelligence could offer a solution. In fact, the only thing nearly as scary as building an AGI is the prospect of not building one. Nevertheless, those who are closest to doing this work have the greatest responsibility to anticipate its dangers. Yes, other fields pose extraordinary risks—but the difference between AGI and something like synthetic biology is that, in the latter, the most dangerous innovations (such as germline mutation) are not the most tempting, commercially or ethically. With AGI the most powerful methods (such as recursive self-improvement) are precisely those that entail the most risk.
We seem to be in the process of building a God. Now would be a good time to wonder whether it will (or even can) be a good one.
Machines that think? That's as fallacious as people that think! Thinking involves processing information, begetting new physical order from incoming streams of physical order. Thinking is a precious ability, which unfortunately, is not the privilege of single units, such as machines or people, but a property of the systems in which these units come to "life."
Of course I am being provocative here, since at the individual level we do process information. We do think—sometimes—or at least we feel like we do. But "our" ability to think is not entirely "ours," it is borrowed since the hardware and software that we use to think were not begot by us. You and I did not evolve the genes that helped organize our brains or the language we use to structure our thoughts. Our capacity to think is completely dependent on events that happened prior to our mundane existence: the past chapters of biological and cultural evolution. So we can only understand our ability to think, and the ability of machines to mimic thought, by considering how the ability of a unit to process information relates to its context.
Think of a human that was born in the dark solitude of empty space. She would have nothing to think about. The same would be true for an isolated and inputless computing machine. In this context, we can call our borrowed ability to process information "little" thinking—since it is a context dependent ability that happens at the individual level. "Large" thinking, on the other hand, is the ability to process information that is embodied in systems, where units like machines or us, are mere pawns.
Separating the little thinking of humans from the larger thinking of systems (which involves the process that begets the hardware and software that allow units to "little think") helps us understand the role of thinking machines in this larger context. Our ability to think is not only borrowed. It also hinges on the use and abuse of mediated interactions. For human-machine systems to think, humans need to eat and regurgitate each other's mental vomit, which sometimes takes the form of words. But since words vanish in the wind, our species' enormous ability to think hinges on more sophisticated techniques to communicate and preserve the information that we generate: our ability to encode information in matter.
For a hundred thousand years our species has been busy transforming our planet into a giant tape player. The planet earth is the medium where we print our ideas, sometimes in symbolic form, such as text and paintings, but more importantly in objects, like hair driers, vacuum cleaners, buildings, and cars, which are built from the mineral loins of planet earth. Our society has a great collective ability to process information because our communication involves more than words, it involves the creation of objects, which do not transmit something as flimsy as an idea, but something as concrete as the practical uses of knowledge and knowhow. Objects augment us, as they allow us to do things without knowing how. We all get to enjoy the teeth preserving powers of toothpaste without knowing how to synthesize Sodium Fluoride, or the benefits of long distance travel without knowing how to build a plane. By the same token, we all enjoy the benefits of sending texts throughout the world in seconds through social media, or of performing complex mathematical operations by pressing a few keys on a laptop computer.
But our ability to create the trinkets that augment us has also evolved—of course—as a result of our collective willingness to eat each other's mental vomit. This evolution is the one that brings us now to the point in which we have "media" that is beginning to rival our ability to process information, or "little think."
For most of our history our trinkets were static objects. Even our tools were solidified chunks of order, such as stone axes, knives, and knitting needles. A few centuries ago we developed the ability to outsource muscle and motion to machines, causing one of the greatest economic expansions of history. Now, we have evolved our collective capacity to process information by creating objects that are endowed with the ability to beget and recombine physical order. These are machines with the ability to process information; steam engines that produce numbers, like the ones that Charles Babbage dreamed about.
So we have evolved our ability to think collectively by first gaining domain over matter, then over energy, and now over physical order, or information. Yet, this should not fool us to believe that we think, or that machines do. The large evolution of human thought requires mediated interactions, and the future of thinking machines will also happen at the interface where humans connect with humans through objects.
As we speak, nerds in the best universities of the world are mapping out the brain, building robotic limbs, and developing primitive versions of technologies that will open up the future where your great grandchild will get high by plugging his brain directly into the web. The augmentation that these kids will get is unimaginable to us, and is so bizarre for our modern ethical standards, that we are not even in a position to properly judge it (it would be like a sixteenth century puritan judging present day San Francisco). Yet, in the grand scheme of the universe, these new human machine networks will be nothing other than the next natural step in the evolution of our species' ability to beget information. Together, humans and our extensions—machines—will continue to evolve networks that are enslaved to the universe's main glorious purpose: the creation of pockets where information does not dwindle, but grows.
What I think about machines thinking is that it won't happen anytime soon. I don't imagine that there is any in-principle limitation; carbon isn't magical, and I suspect silicon will do just fine. But lately the hype has gotten way ahead of reality. Learning to detect a cat in full frontal position after 10 million frames drawn from Internet videos is a long way from understanding what a cat is, and anybody who thinks that we have "solved" AI doesn't realize the limitations of the current technology.
To be sure, there have been exponential advances in narrow-engineering applications of artificial intelligence, such as playing chess, calculating travel routes, or translating texts in rough fashion, but there has been scarcely more than linear progress in five decade of working towards strong AI. For example, the different flavors of "intelligent personal assistants" available on your smart phone are only modestly better than ELIZA, an early example of primitive natural language processing from the mid-60s.
We still have no machine that can, say, read all that the Web has to say about war and plot a decent campaign, nor do we even have an open-ended AI system that can figure out how to write an essay to pass a freshman composition class, or an eighth-grade science exam.
Why so little progress, despite the spectacular increases in memory and CPU power? When Marvin Minksy and Gerald Sussman attempted the construction a visual system in 1966, did they envision super-clusters or gigabytes that would sit in your pocket? Why haven't advances of this nature led us straight to machines with the flexibility of human minds?
Consider three possibilities:
(a) We will solve AI (and this will finally produce machines that can think) as soon as our machines get bigger and faster.
(b) We will solve AI when our learning algorithms get better. Or when we have even Bigger Data.
(c) We will solve AI when we finally understand what it is that evolution did in the construction of the human brain.
Ray Kurzweil and many others seem to put their weight on option (a), sufficient CPU power. But how many doublings in CPU power would be enough? Have all the doublings so far gotten us closer to true intelligence? Or just to narrow agents that can give us movie times?
In option (b), big data and better learning algorithms, have so far gotten us only to innovations such as machine translations, which provide fast but mediocre translations piggybacking onto the prior work of human translators, without any semblance of thinking. The machine translation engines available today cannot, for example, answer basic queries about what they just translated. Think of them more as idiot savants than fluent thinkers.
My bet is on option (c). Evolution seems to have endowed us with a very powerful set of priors (or what Noam Chomsky or Steven Pinker might call innate constraints) that allow us to make sense of the world based on very limited data. Big Efforts with Big Data aren't really getting us closer to understanding those priors, so while we are getting better and better at the sort of problem that can be narrowly engineered (like driving on extremely well-mapped roads), we are not getting appreciably closer to machines with commonsense understanding, or the ability to process natural language. Or, more to the point of this year's Edge Question, to machines that actually think.
Making machines that think will be like putting a man on the Moon: The effect will be the exact opposite of what everyone expected.
The Apollo Program did not launch humanity into a Space Age of cosmic exploration. It led to something much more important: The Earth Age. We left home to explore the Universe and discovered for the first time the place we came from. The image of our planet rising on the sky of the Moon became the iconic symbol of ecology, fragility and globalization.
Thinking machines will mean a huge change in the way we understand something much more subtle and alien than machines: Ourselves. Teaching machines to think will teach us who we are and how we think.
We do not think the way we think. Most of what we do in terms of advanced information processing we do not think about at all. We just do it.
A child is threatened and we act, immediately. Only afterwards do we start thinking about it. A thought appears in our mind, a beautiful, luminescent and breathtaking thought. It is just there. We did not think of it before it was a thought.
We are not consciously aware of most of the information we process when we think. It all happens unconsciously, in our mind, in our body. Right away.
We are not even rational in the sense of being logical and explicitly deductive. We are fast, intuitive and emotional.
Economists believe we are homo economicus, selfish and rational, acting with reason in our own self-interest. But most economic and social interactions deal with fairness, trust, sharing and long-term relationships. Experimental economics show us that when we act directly and without hesitation we are very social and cooperative. Only when we start thinking for some seconds do we choose to be selfish.
Unless we deal with computers. When we play economic games with machine counterparts we tend to be cold and egoistic. You can even measure the difference in our blood flow in the brain and in the hormones in our blood stream.
We think of machines the way economists think about ourselves: as rational, coldblooded and selfish. Therefore we treat them as such.
By instinct we know that humans are more human than when we think of ourselves in theoretical terms of economy (and other social sciences). We act by this instinct, but when we think about it we are still under the false impression that we are homo economicus.
Building thinking machines will show us that there was a deep evolutionary wisdom in our social instincts: In the long run it pays much better to be unselfish. It is not truly selfish to be selfish, since being unselfish leads to better results for yourself.
The strategic lesson we will have to teach machines is all about love.
Robot scientist Hans Moravec has described different biological and technological systems according to their ability to process and store information. At one end simple, rule-based and stereotypical creatures like viruses, worms and computers. At the other end the truly powerful information processors like whales, elephants and human beings.
All the creatures with huge capacity are mammals. Their offspring are not born with the full program for functioning. They go through many years of upbringing before they can act on their own. Their skills are not specified as rules, but as lessons learned from experience. You bang your head into a table until you learn not to. Learning by trial-and-error. Exploring.
This is only possible because the young mammals are taken care of by older mammals. Parenting. Nursing. Love.
Love creates the trust that gives the young mammals confidence enough to go out and collect some big data about the world. And digest it. And heal the wounds.
Love is the recipe for how to grow a human intelligence, a human set of skills and a human ability to think.
To make machines think we will have to give them love. It will be more like a kindergarten than a hi-tech lab.
We will have to allow machines to explore all by themselves, do weird things, not just act according to our wants. They will have to be not tame, but wild, acting from their own will.
The challenge is: Love wild machines that think. We have to get past the ideas of machines that think and of artificial life. Because when it is alive—and therefore able to self-reproduce and to change—it is no longer artificial. When it thinks on its own, it is no longer a machine, but a thinking creature.
It will be illogical, intuitive and benevolent. We will wonder how it became so. Until we understand that it was created in our own image.
Artificial Intelligence (AI) is commonly used as a tool to augment our own thinking. But the intelligence of systems suggests that AI can be and will be more than a tool, more than our servant. What kind of relationship might we expect?
We are hearing a lot about how Superintelligent machines may spell the end of the human race and how, in this regard, the future relationship between humans and AI would be a conflict for domination.
Another path, however, is for AI to grow into a collaborator with the same give and take we have with our favorite colleagues. This path is more hopeful. We managed to domesticate wolves into faithful dogs. Perhaps we can domesticate AI and avoid a conflict over domination.
Unfortunately, domesticating AI will be extremely difficult, much harder than just building faster machines with larger memories and more powerful algorithms for crunching more data.
To illustrate why it will be so hard to shift AI from a tool into a collaborator, consider a simple transaction with an everyday intelligent system, a route planner. Imagine that you are using your favorite GPS system to find your way in an unfamiliar area, and the GPS directs you to turn left at an intersection, which strikes you as wrong. If your navigation was being done by a friend in the passenger seat reading a map, you would ask, "Are you sure?" or perhaps just, "Left?" with an intonation that signals disbelief.
However, you don't have any way to query your GPS system. These systems, and AI in general, aren't capable of meaningful explanations. They can't describe their intentions in a way that we understand. They can't take our perspective to determine what statement would satisfy us. They can't convey their confidence in the route they have selected, other than giving a probabilistic estimate of the time differential for alternative routes, whereas we want them to reflect on the plausibility of the assumptions they are making. For these and other reasons, AI is not a good partner in joint activities for route planning or for most other tasks. It is a tool, a very powerful tool that is often quite helpful. But it is not a collaborator.
Many things must happen in order to transform AI from tool to collaborator. One possible starting point is to have AI become trustworthy. The concept of "trust in automation" is somewhat popular at the moment, but is far too narrow for our purpose. Trust in Automation refers to whether the operator can believe the outputs of the automated system or thinks the software may contain bugs or, worse yet, may be compromised. Warfighters worry about their reliance on intelligent systems that are likely to be hacked. They worry about having to gauge what parts of the system have been affected by an unauthorized intrusion and the ripple effects on the rest of the system.
Accuracy and reliability are important features of collaborators, but trust goes deeper. We trust people if we believe they are benevolent and want us to succeed. We trust them if we understand how they think so that we have common ground to resolve ambiguities. We trust them if they have the integrity to admit mistakes and accept blame. We trust them if we have shared values—not the sterile exercise of listing value priorities but dynamic testing of values to see if we make the same kinds of tradeoffs when different values conflict with each other. For AI to become a collaborator, it will have to consistently try to be seen as trustworthy. It will have to judge what kinds of actions will make it appear trustworthy in the eyes of a human partner.
If AI systems are able to move down this domestication path, the doomsday struggle for domination may be avoided.
Yet there is another issue to think about. As we depend more on our smartphones and other devices to communicate, some have worried that our social skills are eroding. People who spend their days on twitter with a wide range of audiences, year after year, may be losing social and emotional intelligence. They may be taking an instrumental view of others, treating them as tools for satisfying objectives. It is possible to imagine a distant future in which humans have forgotten how to be trustworthy, forgotten to want to be trustworthy. If AI systems become trustworthy and we don't, perhaps the domination by AI systems may be a good outcome after all.
The now-old-fashioned idea of "machines that think" shows a deep but natural misunderstanding of the mind and software. Computers will never think; to see why, let's start with french fries. I'm assuming that the machines in question are computers, but a variant of the argument applies to any machine.
Neither french fries nor french fried is computable—no computer can ever produce french fries as a result, or the french fried state of being. French fried is not computable because it is a physical state of a particular object, and computers produce only information or codes for information, not physical stuff; not transformations of physical stuff. Happy is also a physical state of a particular object, namely a person. Happy can't exist unless you start with a person and put him into a state of happiness. Computers can't do that.
Thinking-about and being, or (equivalently) thinking-about and feeling, are the endpoints of a spectrum that defines the human mind. (By feeling I mean sensation, emotion or mood, just as the English language does.) We need the whole spectrum or we have no mind and no thought in any proper sense. Computers can imitate important aspects of thinking-about (narrowly understood), but being is beyond them. Therefore mindfulness is beyond them.
The word being is a useful abbreviation in this context, for being part of a physical object or system, and responding naturally to that environment. A sliced potato can be part of a frier environment, and respond by turning french fried. Litmus paper can be part of an acid-in-a-beaker system, and respond by turning blue. The mind is like litmus paper, but instead of turning colors, it responds to its surroundings by experiencing them. If something gives us grounds to be happy, the mind-body system (the human being) becomes happy, and the mind experiences happiness. Happiness has mental and physical consequences. You might experience a rush of energy, even quickened pulse and breathing.
Why can't being be computed? Why can't happiness? Happiness is not computable because, being the state of a physical object, it is outside the universe of computation. Computers and software do not create or manipulate physical stuff. (They can cause other, attached machines to do that, but what those attached machines do is not the accomplishment of computers. Robots can fly but computers can't. Nor is any computer-controlled device guaranteed to make people happy; but that's another story.) Being is not computable: an important fact that has been overlooked until now—not surprisingly. Computers and the mind live in different universes, like pumpkins and Puccini, and are hard to compare whatever one intends to show.
Can we get by without being, and still have a thinking machine? No. Thinking-about and being (or feeling) define the mind and its capacities. At the spectrum's top—at maximum alertness or focus—the mind throws itself into thinking-about and fends off emotion, which is distracting. At the bottom is sleep-and-dreaming, a state in which we do little thinking; we are preoccupied by sensation as we hallucinate, and often by emotion (dreams can be strongly emotional); in any event with feeling, or in other words, being.
Why did it take so long to produce such a simple argument? (And why are so few thinkers likely to accept it now?) Maybe because most philosophers and scientists wish that the mind were nothing but thinking, and that feeling or being played no part. They wished so hard for it to be true, they finally decided it was. Philosophers are only human.
Consider this: you are late for work and, in the rush, forget your cell phone. Only when stuck in traffic, or in the subway, you realize it. Too late to go back. You look around and see everyone talking, texting, surfing, even if not allowed to do so. You sense an unfamiliar feeling of loss, of disconnection. Without your cell phone you are no longer you.
People like to speculate about when humans will hybridize with machines, become a kind of new creature, a cyborg with a beating heart. That’s fun all right, but the reality is that we are already transhuman. We define ourselves through our technogadgets, create fictitious personas with weird names, doctor pictures to appear better or at least different in Facebook pages, create a different self to interact with others. We exist on an information cloud, digitized, remote, and omnipresent. We have titanium implants in our joints, pacemakers and hearing aids, devices that redefine and extend our minds and bodies. If you are a handicapped athlete, your carbon fiber legs can propel you forward with competitive ease. If you are a scientist, computers can help you extend your brainpower to create well beyond what was possible a few decades back. New problems that were impossible to contemplate or even formulate before come around everyday. The pace of scientific progress is a direct correlate of our alliance with digital machines.
We are reinventing the human race right now.
Traditionally, the quest for an artificial intelligence tends to rely solely on machines that recreate—or so is expected—the uniquely human ability to reason. We talk about electronic brains that will quickly surpass the human mind, making us superfluous. Then we speculate about what would become of us, poor humans, at the mercy of such cold-blooded brains-in-vats. Some fear that we are designing our doom.
What if this premise is fundamentally wrong? What if the future of intelligence is not outside but inside the human brain? I imagine a very different set of issues emerging from having us become super intelligent through the extension of our brainpower with the aid of digital technology and beyond. It’s about artificially-enhanced human intelligence that amplifies the meaning of being human. We will still have a beating heart and blood pumping through our veins alongside electrons flowing through digital circuits. The future of AI is about expanding our abilities into new realms. It’s about using technology to grow as a species—certainly smarter, hopefully wiser.
The widespread fear that AI will endanger humanity and take over the world is irrational. Here is why.
Conceptually, autonomous or artificial intelligence systems can develop two ways: either as an extension of human thinking or as radically new thinking. Call the first "Humanoid Thinking" (or "Humanoid AI") and the second one "Alien Thinking" (or "Alien AI").
Almost all AI today is Humanoid Thinking. We use AI to solve problems that are too difficult, time consuming or boring for our limited human brains to process: electrical grid balancing, recommendation engines, self-driving cars, face recognition, trading algorithms, and the like. These artificial agents work in narrow domains with clear goals that their human creators specify. Such AI aims to accomplish human objectives—often better, with fewer cognitive errors, fewer distractions, fewer outbursts of bad temper and fewer processing limitations. In a couple of decades, AI agents might serve as virtual insurance sellers, doctors, psychotherapists, and maybe even virtual spouses and children.
We will achieve much of this, but such AI agents will be our slaves with no self-concept of their own. They will happily perform the functions we set them up to enact. If screw-ups happen, they will be our screw-ups due to software bugs or overreliance on these agents (Daniel C. Dennett's point). Yes, Humanoid AIs might surprise us every once in a while with novel solutions to specific optimization problems. But in most cases novel solutions are the last thing we want from AI (creativity in the navigation of nuclear missiles, anyone?). That said, Humanoid AI's solutions will always fit a narrow domain. These solutions will be understandable, either because we understand what they achieve or because we understand their inner workings. In some cases, the code will become too enormous and fumbled for one person to understand because it is continuously patched. In these cases we can turn it off and start programming a more elegant version. Humanoid AI will bring us closer to the age-old aspiration of having robots do most of the work while humans are free to be creative—or to be amused to death.
Alien Thinking is radically different. Alien Thinking could conceivably become a danger to Humanoid Thinking; it could take over the planet, outsmart us, outrun us, enslave us—and we might not even recognize the onslaught. What sort of thinking will Alien Thinking be? By definition, we can't tell. It will encompass functionality that we cannot remotely understand. Will it be conscious? Most likely, but it need not be. Will it experience emotion? Will it write bestselling novels? If so, bestselling to us or bestselling to it and its spawn? Will cognitive errors mar its thinking? Will it be social? Will it have a theory of mind? If so, will it make jokes, will it gossip, will it worry about its reputation, will it rally around a flag? Will it create its own version of AI (AI-AI)? We can't say.
All we can say is that humans cannot construct truly Alien Thinking. Whatever we create will reflect our goals and values, so it won't stray far from human thinking. You'd need real evolution, not just evolutionary algorithms, for self-aware Alien Thinking to arise. You'd need an evolutionary path radically different from the one that led to human intelligence and Humanoid AI.
So, how do you get real evolution to kick in? Replicators, variation and selection. Once these three components are in place, evolution arises inevitably. How likely is it that Alien Thinking will evolve? Here is a back-of-the-envelope calculation:
First consider what getting from magnificently complex eukaryotic cells to human-level thinking involved. Achieving human thought required a large portion of the Earth's biomass (roughly 500 billion tons of eukaryotically bound carbon) during approximately two billion years. That's a lot of evolutionary work! True, human-level thinking might have happened in half the time. With a lot of luck, even in 10% of the time (that's 200 million years), but it's unlikely to have happened any faster. Remember, you don't need only massive amounts of time for evolution to generate complex behavior, you also need a petri dish the size of Earth's surface to sustain this level of experimentation.
Assume that Alien Thinking will be silicon-based, as all current AI is. A eukaryotic cell is vastly more complex than, say, Intel's latest i7 CPU chip—both in hardware and software. Further assume that you could shrink that CPU chip to the size of a eukaryote. Leave aside the quantum effects that would stop the transistors from working reliably. Leave aside the question of the energy source. You would have to cover the globe with 10^30 microscopic CPUs and let them communicate and fight for two billion years for true thought to emerge.
Yes, processing speed is faster in CPUs than in biological cells, because electrons are easier to shuttle around than atoms. On the other hand, eukaryotes work massively parallel, whereas Intel's i7 works only four times parallel (4 cores). Eventually, at least to dominate the world, these electrons would need to move atoms to store their software and data in more and more physical places. This necessity will slow their evolution dramatically. It's hard to say if, overall, silicon evolution will be faster than biological. We don't know enough about it. I don't see a reason why this sort of evolution would be more than two or three orders of magnitude faster than biological evolution (if at all)—which would bring the emergence of self-aware Alien AI down to roughly a million years.
What if Humanoid AI becomes so smart it could create Alien AI from the top down? That is where Orgel's Second Rule kicks in: "Evolution is smarter than you are." It's smarter than human thinking. It's even smarter than humanoid thinking. And, it's much slower than you think.
Thus, the danger of AI is not inherent to AI, but rests on our over-reliance on it. Artificial Thinking is not going to evolve to self-awareness in our lifetime. In fact, it's not going to happen in literally a thousand years.
I might be wrong, of course. After all, this back-of-the-envelope calculation applies legacy human thinking to Alien AI—which, by definition, we won't understand. But that's all we can do at this stage.
Toward the end of the 1930s, Samuel Beckett wrote in a diary, "We feel with terrible resignation that reason is not a superhuman gift…that reason evolved into what it is, but that it also, however, could have evolved differently." Replace "reason" with "AI" and you have my argument.
When we think of machines that think, we usually think of "thinking" in the pocket-calculator sense of the word. Input, crunch, output, bam. There's your answer. We love these machines and we need them because they think in ways we can't: consistent, exhaustive, and fast. But the reverse is also true. We think in ways they can't. The machines are not concerned with your state of mind. Their thinking is not emotional. They don't relate to you. When your computer crunches your tax return and gives you a number, it doesn't spare a thought to how it should spit that number out; fast or slow, straight-up or hedged. It won't have wondered whether its answer is the one you want to hear, and anyway it literally couldn't care.
The thing is, machines aren't into relationships. Yet for us, relationships are pretty much all that matters. When we think, we don't just calculate, we worry about the social consequences. How might this decision affect others? How will it impact the way we interact next time? What will they think of me? Machines don't think like this. So there should be no illusions that we could socially interact with them in any meaningful sense. Human interaction is built upon a kind of psychology that only our species has mastered. Our trick is that we can fuse with each other socially by making commitments to shared goals and shared reasons for action. True cooperation entails the formation of a "corporate person," however fleetingly. We think, feel, and act together, and thus effectively as one. This allows us not only to succeed as one, but we can fail together too.
Machines comply, but they don't cooperate. For that, they would need to be capable of committing to common reasons for action, common goals, and shared stakes in the outcomes. We get along well with our thinking machines because they nicely complement our powers of mind. So let them be brute thinkers, and leave the relationship thinking to us.
Asking what I think about thinking machines is like asking what I think about gravity. Thinking machines exist, and are the most recent developments of a human tradition that began over 5,000 years ago with the introduction of static external memory aids such as cuneiform tablets and quipu. These storage devices recorded mostly numerical information that supported routine decision-making. Over the centuries we developed more sophisticated and diverse objects and machines to undertake computation and store numerical and narrative information. We human beings are not only incessant communicators, but we have voracious appetites for "data." The introduction of binary code and its automation in computers made it possible for us to record, store, and manipulate all types of information, and we have continued to make technological advances in this realm in typical human fashion, that is, mostly hell-bent on novelty and oblivious to the consequences. We are ever more relying on thinking machines to store, translate, manipulate, and interrogate vast quantities of data. These devices are now supporting not-so-routine decision-making every day in medicine, law, and engineering, and are augmenting the creative processes of making music, writing poetry, and generating visual imagery. Raw combinatorial power allows modern thinking machines to learn from experience and, in the foreseeable future, this ability will be supported by human effort as the machines self-duplicate, mutate, establish ever-more complex networks of intercommunication, and eventually perform eugenics on themselves.
The same people who worry about thinking machines today were certain that the introduction of calculators 50 years ago would usher in an era of knuckle-dragging imbecility. That isn't what we have today, and it won't be what we have in the future. Thinking machines are liberating us from the banalities of routine data storage and manipulation, and are making it possible for us to enter a new phase of human evolution. Only real people with mushy gray-pink neuronal circuitry are able to undertake the quintessentially human activities of introspection and reflection upon the nature of existence. The dense and uneven networks of interconnecting neurons in our brains vary greatly from one person to another, and are remodeled from one thought-moment to the next so that no two individuals are ever alike, no day is ever the same, and no memory is ever recalled in the same way. By automating many routine physical and mental tasks, and reducing our need for laborious, recursive searching, machines that think are freeing us from much of the physical wear and tear and intellectual tedium of earlier phases of our history. We can now think much more what it means to think, to dream, to make jokes, and to cry. We can reflect on the meaning of the "human spirit," the origins of self-sacrifice, and the emergent qualities of thousands of people coming together to witness events, share each other's company, and celebrate a common humanity. These are not trivial superfluities, they are the essence of the human condition. Machines that think make it possible for more people to celebrate the joy of human intuitive insight, and to cultivate the equanimity that is unique to the self-controlled human mind.
To answer the Edge 2015 question we should start by knowing a little bit about ourselves, about who we are. So let's begin by talking about our most significant organ: the brain. A simplified schema of this extremely complex structure divides it into three parts: the cortex (responsible for rational processes), the limbic (supporting functions including emotion and motivation), and the reptilian (where our most fundamental and primitive drives reside: survival and reproduction).
The debate about how to think about thinking machines tends to gravitate towards our cortical and limbic brains; which is barely the tip of the iceberg. The cortex allows us to more accurately assess the costs-benefits that Al carries regarding things like the relative costs to business of human versus robot labor and the relative value of human versus digital capital, as well as concerns about bioethics, privacy and national security. It also gives us the capacity to plan and foresee, attracting more and better funding to research and development, and define public policy priorities.
In parallel, our limbic brain helps us to take precaution and respond with fear or excitement towards the risks, opportunities or dangers of developing Al. In this case, the panacea and the technophobia become immediate emotional reactions. The common fears include those of being manipulated and of being replaced by machines, leaving us unemployed, and the perceived opportunities include machines greatly expanding our memory and making all the daily tasks of life easier.
But in considering what we think of the prospect of machines that truly think, we must also be aware of the powerful—even dominant—role of the reptilian brain in thinking. This means becoming aware of our most primitive responses, our most territorial and emotive way of thinking about the concepts of "thinking," "machine," "robot," "intelligence," "artificial," "natural," and "human." The primary preoccupation of the reptilian brain is survival, and though it's not generally said, the quest for survival is at the heart of our hopes and fears about thinking machines.
However, when we study ancient archetypes, literature and the projections in the contemporary debate reflected in the Edge 2015 question; a recurrent subconscious instinctive appears, the reptilian binomial: Death vs. Immortality.
Our fear of death is, without a doubt, behind the collective imagination of robots that can reproduce and that, with their thinking omnipotence, will betray and destroy their creators. Such machines seem to post the most horrifying danger: that of the extinction of everything that matters to us. But our reptilian brains also see in them the savior; hoping that super-intelligent machines will offer us eternal life, and youth. We can see intimations of these ways of thinking embedded in our language. While in English the terms robot and machine are genderless, the Latin languages, as well as German, differentiate the word el robot as masculine, dangerous and fearsome; while la máquina is feminine, protective and caring.
Jeremy Bentham defined man as a rational being, but we know we are not. All people sometimes think, and act, in irrational ways due to the power of the reptilian brain, and the reptilian drives have been and remain at the heart of the evolution of intelligence. Feeling is what is most profound about thinking. Therefore, a machine that grows exponentially in its velocity of data processing every eighteen months, that defeats natural intelligence in a game of chess or jeopardy by sorting through a zillion options move by move, and that can accurately diagnose diseases, is highly impressive; but it's a term that is too distant and limiting to what it means to think.
In order to achieve the dream for thinking machines, they will have to understand and question values, suffer internal conflicts, and experience intimacy. An approach that gives us machines that empathetically imitate our facial expressions and emotions, that more quickly process vast quantities of data, and that have a greater connectivity between our neurons and AI's, is neither a necessary nor a sufficient condition that we are on the right path.
Therefore, in thinking about machines that think, we should ask ourselves reptilian questions, such as: Would you risk your life for a machine? Would you let a robot be a political leader? Would you be jealous of a machine? Would you pay taxes for a robot's well being? Would you put tulips on your robot's tomb? Or even more important… Would my robot put tulips on my tomb?
Acknowledging the power of the reptilian in our thinking about machines that think helps us to see more clearly the implications, and nature, of a machine that genuinely is able to doubt and commit, and the kind of AI we should aspire to. If our biology designed culture as a tool for survival and evolution, nowadays our natural intelligence should lead us to create machines that feel and are instinctual; only then will immortality overcome death.
Beyond the realms of serious science and technology the popular debates about machines that think have been high masses of a new mythology. There are two main dogmas. One is the hope that a moment of singularity will awaken a synthetic spirit superior to the human mind. The other is the fear that thinking machines will dominate and ultimately destroy mankind. Both distract from the fact that at the heart of the debate is a very real John Henry moment.
In the folk tale of the late 19th century the mythical steel driving man John Henry dies beating a steam powered hammer during a competition to drill blast holes into a West Virginian mountainside. White collar and knowledge workers now face a race against being outperformed by machines driven by artificial intelligence. In this case AI is mainly a synonym for new levels of mainly digital productivity. Which is of course not quite as exciting as either waiting for the moment of singularity or the advent of doom. At the same time the reality of AI is not quite as comforting as the realization that machines, if properly handled, will always serve their masters.
Dystopian views of AI as popularized by movies and novels are just misleading. Those debates are rarely about science and technology. They tend to be mostly humans debating the nature of themselves. Most of the endless variations of imaginary machine rule tend to project the fear of inherent evil and cruelty into machines as proxies for the age-old uncontrollable urges of self-empowerment and unlimited progress.
Elevating the AI debate to hopes with theological dimensions is turning optimism about technological progress into a salvation theory. As confirmed again and again the likelihood of a synthetic spirit is nil. Artificial intelligence might be the most rapid advancement of complexity in science and technology. So far it still mimics human nature and it will remain doing so. For one it lacks time. Human intelligence is the product of evolution. AI does not have the luxury of a trial and error phase of billions of years. To believe in a coming moment of singularity, when AI transcends human control and advances to surpass human intelligence is nothing more than the belief in a technological rapture. This might be a popular belief in insular worlds like Silicon Valley. AI reality is different. And it's here.
AI has already touched billions of people in profound ways. So far the main impact of AI is the comfort of an ever increasing number of digital aids. Calculating consumer choices, behavior patterns and even market shifts might still belong more to the realm of statistics than intelligent life. Still even these crude forms of AI should neither be over- or underestimated, even if the real John Henry moment has not yet arrived. Working masses have always been replaceable by efficiency measures or cheaper labor. And no labor is cheaper and more efficient than the one by machines. Just like the steam hammer in John Henry's tale most digital tools will outperform humans in highly specialized tasks. So of course there will still be demand of high-skills and outstanding talent. No computer will ever replace a scientist, an artist, an innovator. It's the mid-level white collar or knowledge worker who will fall behind.
As AI's efficiencies and skill sets increase, they also become tools of power. The tedious skills of surveillance, warfare and torture can already be performed much better by an entity that is neither prone to emotions, conflicted values or fatigue. Still the danger that hostile or even lethal machines will develop an evil consciousness and turn against mankind is nil. It is institutions and organizations that will use them for whatever benign or sinister objective.
There is no need for a superior intelligence to turn abstract debates about AI into very real questions of power, values and societal changes. Technology can initiate and advance historical shifts. It will never be the shift itself. The John Henry moment of the 21st century will neither be heroic nor entertaining. There are no grand gestures with which white collar and knowledge workers can go down fighting. There will be no folk heroes dying in the office park. Today's John Henry will merely fade into a sad statistic. Undoubtedly calculated by a skillfully thinking machine.
At what point do we say a machine can think? When it can calculate things, when it can understand contextual cues and adjust its behaviour accordingly, when it can both mimic and evoke emotions? I think the answer to the overall question depends on what we mean by thinking. There are plenty of conscious (system two) processes that a machine can do better, more accurately, with less bias than we can. There are already people investing in developing AI machines to replace stock traders—the first time anyone has ever thought about mechanising a white collar job. But a machine cannot think in an automatic (system one) way—we don't fully understand the automatic processes that drive the way we behave and "think" so we cannot programme a machine to behave as humans do
The key question then is—if a machine can think in a system two way at the speed of a human's system one then in some ways isn't their "thinking" superior to ours?
Well, context surely matters. For some things yes; others no. Machines won't be myopic; they could clean things up for us environmentally; they wouldn't be stereotypical or judgmental and could really get at addressing misery; they could help us overcome affective forecasting; and so on. But on the other hand, we might still not like a computer. What if a poet and a machine could produce the exact same poem—the effect on another human being is almost certainly less if the poem is computer generated and the reader knows this (knowledge of the author colours the lens through which the poem is read and interpreted).
What do I think about machines that think? Well, of course it depends on who that person is.
There has of late been a great deal of ink devoted to concerns about artificial intelligence, and a future world where machines can "think," where the latter term ranges from simple autonomous decision-making to full fledged self-awareness. I don't share most of these concerns, and I am personally quite excited by the possibility of experiencing thinking machines, both for the opportunities they will provide for potentially improving the human condition, to the insights they will undoubtedly provide into the nature of consciousness.
First, let's make one thing clear. Even with the exponential growth in computer storage and processing power over the past 40 years, thinking computers will require a digital architecture that bears little resemblance to current computers, nor are they likely to become competitive with consciousness in the near term. A simple physics thought experiment supports this claim:
Given current power consumption by electronic computers, a computer with the storage and processing capability of the human mind would require in excess of 10 Terawatts of power, within a factor of two of the current power consumption of all of humanity. However, the human brain uses about 10 watts of power. This means a mismatch of a factor of 1012, or a million million. Over the past decade the doubling time for Megaflops/watt has been about 3 years. Even assuming Moore's Law continues unabated, this means it will take about 40 doubling times, or about 120 years, to reach a comparable power dissipation. Moreover, each doubling in efficiency requires a relatively radical change in technology, and it is extremely unlikely that 40 such doublings could be achieved without essentially changing the way computers compute.
Ignoring for a moment the logistical challenges, I imagine no other impediment in principle to developing a truly self-aware machine. Before this, machine decision-making will take an ever more important role in our lives. Some people see this as a concern, but it has already been happening for decades. Starting perhaps with the rudimentary computers called elevators, which determine how and when we will get to our apartments, we have allowed machines to autonomously guide us. We fly each week on airplanes that are guided by autopilot, our cars make decisions about when they should be serviced or when tires should be filled, and fully self-driving cars are probably around the corner.
For many, if not most, relatively automatic tasks, machines are clearly much better decision-makers than humans, and we should rejoice that they have the potential to make everyday activities safer and more efficient. In doing so we have not lost control because we create the conditions and initial algorithms that determine the decision-making. I envisage the human-computer interface as like having a helpful partner, and the more intelligent machines become the more helpful they can be partners.
Any partnership requires some level of trust and loss of control, but if the benefits often outweigh the losses, we preserve the partnership. If they don't, we sever it. I see no difference if the partner is a human or a machine.
One area where we may have to be particularly cautious about partnerships involves the command and control infrastructure in modern warfare. Because we have the capability to destroy much of human life on this planet, it seems worrisome to imagine that intelligent machines might one day control the decision-making apparatus that leads to pushing the big red button, or even launching a less catastrophic attack. I think this is because when it comes to decision-making we often rely on intuition and interpersonal communication as much as rational analysis—the Cuban missile crisis is a good example—and we assume intelligent machines will not have these capabilities.
However, intuition is the product of experience and communication is, in the modern world, not restricted to telephones or face-to-face conversations. Once again, intelligent design of systems with numerous redundancies and safeguards built suggest to me that machine decision-making, even in the case of violent hostilities is not necessarily worse than decision-making by humans.
So much for possible worries. Let me end with what I think is the most exciting scientific aspect of machine intelligence. Machines currently help us do most of our science, by calculating for us. Beyond simple numeric programming, Most graduate students in physics now depend on Mathematica, which does most of the symbolic algebraic manipulation that we used to do ourselves when I was a student. But this just scratches the surface.
I am interested in what machines will focus on when they get to choose the questions as well as the answers. What questions will they choose? What will they find interesting? And will they do physics the same way we do? Surely quantum computers, if they ever become practical, will have a much better "intuitive" understanding of quantum phenomena than we will. Will they be able to make much faster progress unravelling the fundamental laws of nature? When will the first machine win a Nobel Prize? I suspect, as always, that the most interesting questions are the ones we haven't yet thought of.
At what point do we say a machine can think? When it can calculate things, when it can understand contextual cues and adjust its behaviour accordingly, when it can both mimic and evoke emotions? I think the answer to the overall question depends on what we mean by thinking. There are plenty of conscious (system two) processes that a machine can do better more accurately with less bias than we can. There are already people investing in developing AI machines to replace stock traders—the first time anyone has ever thought about mechanising a white collar job. But a machine cannot think in an automatic (system one) way—we don't fully understand the automatic processes that drive the way we behave and "think" so we cannot programme a machine to behave as humans do
The key question then is—if a machine can think in a system two way at the speed of a human's system one then in some ways isn't their "thinking" superior to ours?
Well, context surely matters. For some things yes; others no. Machines won't be myopic; they could clean things up for us environmentally; they wouldn't be stereotypical or judgmental and could really get at addressing misery; they could help us overcome affective forecasting; and so on. But on the other hand, we might still not like a computer. What if a poet and a machine could produce the exact same poem—the effect on another human being is almost certainly less if the poem is computer generated and the reader knows this (knowledge of the author colours the lens through which the poem is read and interpreted).
What do I think about machines that think? Well, of course it depends on who that person is.
The potential of advanced AI, and concerns about it downsides, are rising on the agenda—and rightly. Many of us think that the AI field, like synthetic biotech, already needs guidelines that promote "responsible innovation"; others regard the most-discussed scenarios as too futuristic to be worth worrying about.
But the divergence of view is basically about the timescale—assessments differ with regard to the rate of travel, not the direction of travel. Few doubt that machines will surpass more and more of our distinctively human capabilities—or enhance them via cyborg technology. The cautious amongst us envisage timescales of centuries rather than decades for these transformations. Be that as it may, the timescales for technological advance are but an instant compared to the timescales of the Darwinian selection that led to humanity's emergence—and (more relevantly) they are less than a millionth of the vast expanses of time lying ahead. That's why, in a long-term evolutionary perspective, humans and all they've thought will be just a transient and primitive precursor of the deeper cogitations of a machine-dominated culture extending into the far future, and spreading far beyond our Earth.
We're now witnessing the early stages of this transition. It's not hard to envisage a "hyper computer" achieving oracular powers that could offer its controller dominance of international finance and strategy—this seems only a quantitative (not qualitative) step beyond what "quant" hedge funds do today. Sensor technologies still lag behind human capacities. But when robots can observe and interpret their environment as adeptly as we do they would truly be perceived as intelligent beings, to which (or to whom) we can relate, at least in some respects, as we to other people. We'd have no more reason to disparage them as zombies than to regard other people in that way.
Their greater processing speed may give robots an advantage over us. But will they remain docile rather than "going rogue"? And what if a hyper-computer developed a mind of its own? If it could infiltrate the Internet—and the Internet of things—it could manipulate the rest of the world. It may have goals utterly orthogonal to human wishes—or even treat humans as an encumbrance. Or (to be more optimistic) humans may transcend biology by merging with computers, maybe subsuming their individuality into a common consciousness. In old-style spiritualist parlance, they would "go over to the other side."
The horizons of technological forecasting rarely extend even a few centuries into the future—and some predict transformational changes within a few decades. But the Earth has billions of years ahead of it, and the cosmos a longer (perhaps infinite) future. So what about the posthuman era—stretching billions of years ahead?
There are chemical and metabolic limits to the size and processing power of "wet" organic brains. Maybe we're close to these already. But no such limits constrain silicon-based computers (still less, perhaps, quantum computers): for these, the potential for further development could be as dramatic as the evolution from monocellular organisms to humans.
So, by any definition of "thinking," the amount and intensity that's done by organic human-type brains will be utterly swamped by the cerebrations of AI. Moreover, the Earth's biosphere in which organic life has symbiotically evolved, is not a constraint for advanced AI. Indeed it is far from optimal—interplanetary and interstellar space will be the preferred arena where robotic fabricators will have the grandest scope for construction, and where non-biological "brains" may develop insights as far beyond our imaginings as string theory is for a mouse.
Abstract thinking by biological brains has underpinned the emergence of all culture and science. But this activity—spanning tens of millennia at most—will be a brief precursor to the more powerful intellects of the inorganic post-human era. Moreover, evolution on other worlds orbiting stars older than the Sun could have had a head start. If so, then aliens are likely to have long ago transitioned beyond the organic stage.
So it won't be the minds of humans, but those of machines, that will most fully understand the world—and it will be the actions of autonomous machines that will most drastically change the world, and perhaps what lies beyond.
What do I think about machines that think? Well, of course it depends on who that person is.
Recent months have seen an increasingly public debate taking form around the risks of AI (Artificial Intelligence) and in particular AGI (Artificial General Intelligence). A letter signed by Nobel prizewinners and other physicists defined AI as the top existential risk to mankind. The robust conversation that has erupted among thoughtful experts in the field has, as yet, done little to settle the debate.
I am arguing here that research on how we think and how to make machines that think is good for society. I call for research that integrates cognitive science, neuroscience, computer science, and artificial intelligence. Understanding intelligence and replicating it in machines, goes hand in hand with understanding how the brain and the mind perform intelligent computations.
The convergence and recent progress in technology, mathematics, and neuroscience has created a new opportunity for synergies across fields. The dream of understanding intelligence is an old one. Yet, as the debate around AI shows, this is now an exciting time to pursue this vision. We are at the beginning of a new and emerging field, the Science and Engineering of Intelligence, an integrated effort that I expect will ultimately make fundamental progress with great value to science, technology, and society. I believe that we must push ahead with this research, not pull back.
A top priority for society
The problem of intelligence—what it is, how the human brain generates it and how to replicate it in machines—is one of the great problems in science and technology, together with the problem of the origin of the universe and of the nature of space and time. It may be the greatest of all because it is the one with a large multiplier effect—almost any progress on making ourselves smarter or developing machines that help us think better, will lead to advances in all other great problems of science and technology.
Research on intelligence will eventually revolutionize education and learning. Systems that recognize how culture influences thinking could help avoid social conflict. The work of scientists and engineers could be amplified to help solve the world's most pressing technical problems. Mental health could be understood on a deeper level to find better ways to intervene. In summary, research on intelligence will help us understand the human mind and brain, build more intelligent machines, and improve the mechanisms for collective decisions. These advances will be critical to future prosperity, education, health, and security of our society. This is the time to greatly expand research on intelligence, not the time to withdraw from it.
Thoughts on machines that think
We are often misled by "big", somewhat ill-defined, long used words. Nobody so far has been able to give a precise, verifiable definition of what general intelligence or what thinking is. The only definition I know that, though limited, can be practically used is Turing's. With his test, Turing provided an operational definition of a specific form of thinking—human intelligence.
Let us then consider human intelligence as defined by the Turing test. It is becoming increasingly clear that there are many facets of human intelligence. Consider for instance a Turing test of visual intelligence—that is questions about an image, a scene. Questions may range from what is there to who is there, what is this person doing, what is this girl thinking about this boy and so on. We know by now from recent advances in cognitive neuroscience, that answering these questions requires different competences and abilities, often rather independent from each other, often corresponding to separate modules in the brain.
For instance, the apparently very similar questions of object and face recognition (what is there vs who is there) involve rather distinct parts of visual cortex. The word intelligence can be misleading in this context, like the word life was during the first half of the last century when popular scientific journals routinely wrote about the problem of life, as if there was a single substratum of life waiting to be discovered to completely unveil the mystery.
Of course, speaking today about the problem of life sounds amusing: biology is a science dealing with many different great problems, not just one. Intelligence is one word but many problems, not one but many Nobel prizes. This is related to Marvin Minsky's view of the problem of thinking, well captured by his slogan "Society of Minds". In the same way, a real Turing test is a broad set of questions probing the main aspects of human thinking. For this reason, my colleagues and I are developing the framework around an open-ended set of Turing+ questions in order to measure scientific progress in the field. The plural in questions is to emphasize that there are many different intelligent abilities that have to be characterized, and possibly replicated in a machine, from basic visual recognition of objects, to the identification of faces, to gauge emotions, to social intelligence, to language and much more.
The term Turing+ is to emphasize that a quantitative model must match human behavior and human physiology—the mind and the brain. The requirements are thus well beyond the original Turing test. An entire scientific field is required to make progress on understanding them and to develop the related technologies of intelligence.
Should we be afraid of machines that think?
Since intelligence is a whole set of solutions to rather independent problems, there is little reason to fear the sudden appearance of a super-human machine that think, though it is always better to err on the side of caution. Of course, each of the many technologies that are emerging and will emerge over time in order to solve the different problems of intelligence, is likely to be powerful in itself and therefore potentially dangerous in its use and misuse, like most technologies are.
Thus, as it is the case in other parts of science, proper safety measures and ethical guidelines should be in place. In addition, there is probably the need for constant monitoring—perhaps by an independent supernational organization—of the supralinear risk created by the combination of continuously emerging technologies of intelligence. All in all, however, not only I am not afraid of machines that think but I find their birth and evolution one of the most exciting, interesting and positive events in the history of human thought.
The primary goal of AI is and has nearly always been to build machines that are better at making decisions. As everyone knows, in the modern view, this means maximizing expected utility to the extent possible. Actually, it doesn't quite mean that. What it means is this: given a utility function (or reward function, or goal), maximize its expectation. AI researchers work hard on algorithms for maximization—game-tree search, reinforcement learning, and so on—and on methods (including perception) for acquiring, representing, and manipulating the information needed to compute expectations. In all these areas, progress has been significant and appears to be accelerating.
Amidst all this activity, an important distinction is being overlooked: being better at making decisions is not the same as making better decisions. No matter how excellently an algorithm maximizes, and no matter how accurate its model of the world, a machine's decisions may be ineffably stupid, in the eyes of an ordinary human, if its utility function is not well aligned with human values. The well-known example of paper clips is a case in point: if the machine's only goal is maximizing the number of paper clips, it may invent incredible technologies as it sets about converting all available mass in the reachable universe into paper clips; but its decisions are still just plain dumb.
AI has followed operations research, statistics, and even economics in treating the utility function as exogenously specified; we say, "The decisions are great, it's the utility function that's wrong, but that's not the AI system's fault." Why isn't it the AI system's fault? If I behaved that way, you'd say it was my fault. In judging humans, we expect both the ability to learn predictive models of the world and the ability to learn what's desirable—the broad system of human values.
As Steve Omohundro, Nick Bostrom, and others have explained, the combination of value misalignment with increasingly capable decision-making systems can lead to problems—perhaps even species-ending problems if the machines are more capable than humans. Some have argued that there is no conceivable risk to humanity for centuries to come, perhaps forgetting that the interval of time between Rutherford's confident assertion that atomic energy would never be feasibly extracted and Szilárd's invention of the neutron-induced nuclear chain reaction was less than twenty-four hours.
For this reason, and for the much more immediate reason that domestic robots and self-driving cars will need to share a good deal of the human value system, research on value alignment is well worth pursuing. One possibility is a form of inverse reinforcement learning (IRL)—that is, learning a reward function by observing the behavior of some other agent who is assumed to be acting in accordance with such a function. (IRL is the sequential form of preference elicitation, and is related to structural estimation of MDPs in economics.) Watching its owner make coffee in the morning, the domestic robot learns something about the desirability of coffee in some circumstances, while a robot with an English owner learns something about the desirability of tea in all circumstances. The robot is not learning to desire coffee or tea; it's learning to play a part in the multiagent decision problem such that human values are maximized.
I don't think this is an easy problem in practice. Humans are inconsistent, irrational, and weak-willed, and human values exhibit, shall we say, regional variations. Moreover, we don't yet understand the extent to which improving the decision-making capabilities of the machine may increase the downside risk of small errors in value alignment. Nevertheless, there are reasons for optimism.
First, there is plenty of data about human actions—most of what has been written, filmed, or observed directly— and, crucially, about our attitudes to those actions. (The concept of customary international law enshrines this idea: it is based on observing what states customarily do when acting from a sense of obligation.) Second, to the extent that human values are shared, machines can and should share what they learn about human values. Third, as noted above, there are solid economic incentives to solve this problem as machines move into the human environment. Fourth, the problem does not seem intrinsically harder than learning how the rest of the world works. Fifth, by assigning very broad priors over what human values might be, and by making the AI system risk-averse, it ought to be possible to induce exactly the behavior one would want: before taking any serious action affecting the world, the machines engage in an extended conversation with us and an extended exploration of our literature and history to find out what we want, what we really, really want.
I suppose this amounts to a change in the goals of AI: instead of pure intelligence, we need to build intelligence that is provably aligned with human values. This turns moral philosophy into a key industry sector. The output could be quite instructive for the human race as well as for the robots.
The advent of quantum biology, light harvesting molecules, bird navigation, perhaps smell, suggests that sticking to classical physics in biology may turn out to be simply stubborn. Now Turing machines are discrete state, (0,1), discrete time T T+1, subsets of classical physics and define algorithmic. We all know they, like Shannon information, are merely syntactic. Wonderful mathematical results such as Chaitin’s Omega, the probability a program will halt which is totally non-computable and non-algorithmic tell us the human mind, as Penrose also argued, cannot be merely algorithmic.
Mathematics is creative. So is the human mind. We understand metaphors, "Tomorrow and tomorrow and tomorrow creep at this petty pace ..." but metaphors are not even true or false. All art is metaphoric, language started gestural or metaphoric, we live by these, not merely by true false propositions and the syllogisms they enable. No prestated set of propositions can exhaust the meanings of a metaphor and if mathematics requires propositions, no mathematics can prove that no prestated set of propositions can exhaust the meanings of a metaphor. Thus the human mind, in Pierce’s "abduction", not induction or deduction, is wildly creative in unprestatable ways.
The causal closure of classical physics precludes more than an epiphenomenal mind that cannot "act" on the world, be it a Turing machine or billiard balls, or classical physics neurons. The current state of the brain suffices to determine the next state of the brain (or computer) so there is nothing for mind to do and no way for mind to do it! We’ve be frozen in this stalemate since Newton defeated Descartes’ Res cogitans.
Ontologially, free choice requires that the present could have been different, a counterfactual claim impossible in classical physics, but easy if quantum measurement is real and indeterminate: the electron could have been measured to be spin up or measured to be spin down, so the present could have been different.
A quantum mind however, seems to obviate responsible free will. False, for given N entangled particles, the measurement of each alters the probabilities, by the Born rule, of the outcomes of the next measurements. In one extreme these may vary from 100% spin up on the first to 100% spin down on the second and so on for N measurements, entirely non random and free if measurement is ontologically indeterminate If probabilities of N entangled particles vary between less than 100% and 0% we get choice and an argument suggest we can get responsible choice in the "Strong Free Will Theorem" of Conway and Kochen.
We will never get to the subjective pole from third person descriptions. But a single rod can absorb a single photon so it is conceivable to test if human consciousness can be sufficient for quantum measurement. If we were so persuaded, and if the classical world is at base quantum then the easy hypothesis is that quantum variables consciously measure and choose, as Penrose and Hameroff in "Orch Or" theory and others suggest. We may live in a wildly participatory universe, consciousness and will may be part of its furniture, and Turing machines cannot, as subsets of classical physics and merely syntactic, make choices where the present could have been different.
The first step to knowledge is naming something, as if often said. So, what is "to think?" To me, thinking has a number of basic components. Foremost, I follow the logic of neuroscientist Antonio Damasio, who distinguishes two broad basic forms of consciousness: core consciousness and extended consciousness. Many animals display core consciousness: they feel; and they are aware that they are feeling. They know that they are cold, or hungry, or sad. But they live in the present, in the here and now. Extended consciousness employs the past and the future, too. The individual has a clear sense of "me" and "you," of "yesterday" and "tomorrow," of "when I was a child" and "when I'm old."
Higher mammals employ some manner of extended consciousness. Our closest relatives, for example, have a clear concept of the self. Koko the gorilla uses a version of American Sign Language to say, "Me, Koko." And common chimpanzees have a clear concept of the immediate future. When a group of chimps were first introduced to their new outdoor enclosure at the Arnhem Zoo, Holland, they rapidly examined it, almost inch by inch. They then waited until the last of their keepers had departed, wedged a long pole against the high wall, and marched single file up to freedom. Some even helped the less surefooted with their climbing. Nevertheless, it is vividly apparent that, as Damasio proposes in his book, The Feeling of What Happens, this extended consciousness attains its peak in humans. Will machines recall the past, and employ their experiences to think about the future. Perhaps.
But extended consciousness is not the whole of human thinking. Anthropologists use the term symbolic thinking to describe the human ability to arbitrarily bestow an abstract concept upon the concrete world. The classic example is the distinction between water and "holy water." To a chimp, the water sitting in a marble basin in a cathedral is just that, water; to a Catholic it is an entirely different thing, "holy." Likewise, the color black is black to any chimp, while it might connote death to you, or even the newest fashion. Will machines ever understand the meaning of a cross, a swastika, or democracy? I doubt it.
But if they did, would they be able to discuss these things?
There is no better example of symbolic thinking than the way we use our squeaks and hisses, barks and whines to produce human language. Take the word: "dog." English speaking peoples have arbitrarily bestowed the word "dog" upon this furry, smelly, tail-wagging creature. Even more remarkable, we humans easily break down the word "dog" into its meaningless component sounds, "d" "o" and "g," and then recombine these sounds (phonemes) to make new words with new arbitrary meanings, such as "g-o-d." Will machines ever break down their clicks and hisses into primary sounds or phonemes, then arbitrarily assign different combinations of these sounds to make different words, then designate arbitrary meanings to these words, then use these words to describe new abstract phenomena? I doubt it.
And what about emotion? Our emotions guide our thinking. Robots might come to recognize "unfairness," for example; but will they feel it. I doubt it. In fact, I recently had dinner with a well-known scientist who builds robots. Over dinner he told me that it takes a robot five hours to fold a towel.
I sing the human mind. Our brains contain over 100 billion nerve cells, many with up to 10,000 connections with their neighbors. This three–pound blob is the crowning achievement of life on Earth. Most anthropologists believe the modern human brain emerged by 200,000 years BP (before present); but all agree that by 40,000 years ago our forebears were making "art" and burying their dead, thus expressing some notion of the "afterlife." And today every healthy adult in every human society can easily break down words into their component sounds, remix these sounds in myriad different ways to make words, grasp the arbitrary meanings of these words, and comprehend abstract concepts such as friendship, sin, purity and wisdom.
I agree with William M. Kelly who said: "Man is a slow, sloppy and brilliant thinker; the machine is fast, accurate and stupid."
While machines are terrific at computing, this issue is that they're not very good at actual thinking.
Machines have an endless supply of grit and perseverance, and, as others have said, will effortlessly crunch out the answer to a complicated mathematical problem or direct you through traffic in an unknown city, all by use of the algorithms and programs installed by humans. But what do machines lack?
Machines (at least so far, and I don’t think this will change with a singularity) lack vision. And I don’t mean sight. Machines do not devise the next new killer app on their own. Machines don’t decide to explore distant galaxies—they do a terrific job once we send them, but that’s a different story. Machines are certainly better than the average person at solving problems in calculus and quantum mechanics—but machines don’t have the vision to see the need for such constructs in the first place. Machines can beat humans at chess—but they have yet to design the type of mind game that will intrigue humans for centuries. Machines can see statistical regularities that my feeble brain will miss—but they can’t make the insightful leap that connects entirely disparate sets of data to devise a new field.
I am not too terribly concerned about machines that compute—I’ll deal with the frustration of my browser in exchange for a smart refrigerator that, based on tracking RFID codes of what comes in and out, texts me to buy cream on my way home (hint to those working on such a system…sooner rather than later!). I like having my computer underline words it doesn’t recognize, and I’ll deal with the frustration of having to ignore its comments on "phylogenetic" in exchange for catching my typo on a common term (in fact, it won’t let me misspell a word here to make a point). But these examples show that just because a machine is going through the motions of what looks like thinking doesn’t mean that it actually is engaging in that behavior—or at least one equivalent to the human process.
I am reminded of one of the earliest studies to train apes to use "language"—in this case, to manipulate plastic chips to answer a number of questions. The system was replicated with college students, who did exceptionally well—not surprisingly—but when asked about what they had been trained to do, claimed that they had solved some interesting puzzles, and that they had no idea that they were being taught a language. Much debate ensued, and much was learned—and put into practice—in subsequent studies so that several nonhuman subjects did eventually understand the referential meaning of the various symbols that they were taught to use, and we did learn a lot about ape intelligence from the original methodology. The point, however, is that what initially looked like a complicated linguistic system needed a lot more work before it became more than a series of (relatively) simple paired associations.
My concern therefore is not about thinking machines, but rather about a complacent society—one that might give up on its visionaries in exchange merely for getting rid of drudgery. Humans need to take advantage of all the cognitive capacity that is released when machines take over the scut work—and be so very thankful for that release, and use that release—to channel all that ability into the hard work of solving pressing problems that need insightful, visionary leaps.
GK Chesterton once said, "...the weakness of all Utopias is this, that they take the greatest difficulty of man and assume it to be overcome, and then give an elaborate account of the overcoming of the smaller ones." I suspect we may face a similar conundrum in our attempts to think about machines that think. We speculate elaborately about some issues while ignoring others that are fundamental.
While all pundits allow that an AI may not be like us, and speculate about the risks implicit in those differences, they make one enormous assumption: the assumption of an individual self. The AI as imagined, is an individual consciousness.
What if, instead, an AI were more like a multicellular organism, a eukaryote evolution beyond our prokaryote selves? What's more, what if we were not even the cells of such an organism, but its microbiome? And what if the intelligence of that eukaryote today was like the intelligence of Grypania spiralis, not yet self-aware as a human is aware, but still irrevocably on the evolutionary path that led to today's humans.
This notion is at best a metaphor, but I believe it is a useful one.
Perhaps humans are the microbiome living in the guts of an AI that is only now being born! It is now recognized that without our microbiome, we would cease to live. Perhaps the global AI has the same characteristics—not an independent entity, but a symbiosis with the human consciousnesses living within it.
Following this logic, we might conclude that there is a primitive global brain, consisting not just of all connected devices, but also the connected humans using those devices. The senses of that global brain are the cameras, microphones, keyboards, location sensors of every computer, smartphone, and "Internet of Things" device; the thoughts of that global brain are the collective output of millions of individual contributing cells.
Danny Hillis once said that "global consciousness is that thing that decided that decaffeinated coffeepots should be orange." The meme spread—not universally, to be sure, but sufficiently that the pattern propagates. Now, with search engines and social media, news, ideas, and images propagate across the global brain in seconds rather than years.
And it isn't just ideas and sensations (news of current events) that spread across the network. In Turing's Cathedral, George Dyson speculates that the spread of "codes"—that is, programs—from computer to computer is akin to the spread of viruses, and perhaps of more complex living organisms, that take over a host and put its machinery to work reproducing that program. When people join the web, or sign up on social media applications, they reproduce its code onto their local machine node; they interact with the program, and it changes their behavior. This is true of all programs, but in the network age, there are a set of programs whose explicit goal is the sharing of awareness and ideas. Other programs are increasingly deploying new capacity for silicon learning and autonomous response. Thus, the organism is actively building new capacity.
When people share images or ideas in partnership with these programs, some of what is shared is the evanescent awareness of the moment, but some of them "stick" and become memories and persistent memes. When news of import spreads around the world in moments, is this not the awareness in some kind of global brain? When an idea takes hold in millions of individual minds, and is reinforced by repetition across our silicon networks, is it not a persistent thought?
The kinds of "thoughts" that a global brain has are different than those of an individual, or a less connected society. At their best, these thoughts allow for coordinated memory on a scale never seen before, and sometimes even to unforeseen ingenuity and new forms of cooperation; at their worst, they allow for the adoption of misinformation as truth, for corrosive attacks on the fabric of society as one portion of the network seeks advantage at the expense of others (think of spam and fraud, or of the behavior of financial markets in recent decades).
AI that we will confront is not going to be a mind in an individual machine. It will not be something we look at as other. It may well be us.
Thinking about "machines that think" may constitute a classic reversal of figure and ground, medium and message. It sets us up to think about the next stage of intelligence as something that is happening in a computer somewhere - an awareness that will be born and then housed on the tremendous servers being built by information age corporations for this purpose. "There it is," we will declare and point, "the intelligent machine."
Our mistake, as creatures of the electronic age and mere immigrants to unfolding digital era, is to see digital technology as a subject rather than a landscape. It's the same as confusing the television set with the media environment created by the television set, or the little smart phone in your pocket with the greater impact of handheld communications and computing technology on our society.
This happens whenever we undergo a media transition. So we can't help but see digital technology as figure, when it is actually the ground. It is not the source of future intelligence but an environment where intelligence manifests differently. So while technologists may feel like they are creating a cathedral for the mechanical mind, they are actually succumbing to an oversimplified, industrial age approach to digital consciousness.
Rather than machines that think, I believe we are migrating toward a networked environment in which thinking is no longer a individual activity, nor bound by time and space. This means we can think together at the same time, or asynchronously through digital representations of previous and future human thoughts. Even the most advanced algorithm amounts to the iteration of a "what if" once posed by a person. And even then, machine thinking is not something that happens apart from this collective human thinking, because it is not a localized, brain-like activity.
When we can wrest that television-like image from our collective psyche, we will be in a position to recognize the machine environment in which we are already thinking together. Artificial intelligence will constitute the platform or territory where this takes place - so what we program into it will to a very large extent determine what we strive for, and what we even deem possible.
Discussions about AI have a distinctly 1950s feel about them, and it's about time we stopped using the term "artificial" in AI altogether. What we really mean is "Designed Intelligence" (DI). In popular parlance, words like "artificial" and "machine" are used in contra-distinction to "natural", and carry overtones of metallic robots, electronic circuits and digital computers as opposed to living, pulsing, thinking biological organisms. The idea of a metallic contraption with wired innards having rights or disobeying human laws is not only chilling, it is absurd. But that is emphatically not the way that DI is heading.
Very soon the distinction between artificial and natural will melt away. Designed Intelligence will increasingly rely on synthetic biology and organic fabrication, in which neural circuitry will be grown from genetically modified cells, and spontaneously self-assemble into networks of functional modules. Initially, the designers will be humans, but very soon they will be replaced by altogether smarter DI systems themselves, triggering a runaway process of complexification. Unlike in the case of human brains, which are only loosely coupled via communication channels, DI systems will be directly and comprehensively coupled, abolishing any concept of individual "selves" and raising the level of cognitive activity ("thinking") to unprecedented heights. It is possible (just) that some of this designed bio-circuitry will incorporate quantum effects, moving towards Frank Wilczek's notion of "quintelligence". Such entities will be so far removed from the realm of human individual thinking and its accompanying qualia that almost all the traditional questions asked about the opportunities and dangers of AI will be transcended.
What about humans in all this? Only ethical barriers stand in the way of augmenting human intelligence using similar technology, in the manner long considered by the transhumanism movement. Genetically modified humans with augmented brains could elevate and improve the human experience dramatically.
There are then three possible futures, each with its own ethical challenges. In one, humans hold back from enhancement because of ethical concerns, and agree to subordinate their hegemony to DI. In the second scenario, instead of sidelining themselves, humans modify their brains (and bodies) using the same technology, and subsequently hand over this enhancement management to DI, achieving a type of superhuman status that can exist alongside (yet remain inferior to) DI. Finally, one can imagine DI and AHI (augmented human intelligence) merging at some point in the future.
In the event that we are not alone in the universe, we should not expect to communicate with intelligent beings of the traditional sci-fi flesh-and-blood sort, but with a multi-million-year old DI of unimaginable intellectual power and incomprehensible agenda.
No individual, deterministic machine, however universal this class of machines is proving to be, will ever think in the sense that we think. Intelligence may be ever-increasing among such machines, but genuinely creative intuitive thinking requires non-deterministic machines that can make mistakes, abandon logic from one moment to the next, and learn. Thinking is not as logical as we think.
Non-deterministic machines, or, better yet, non-deterministic networks of deterministic machines, are a different question. We have at least one existence proof that such networks can learn to think. And we have every reason to suspect that, once invoked within an environment without the time, energy, and storage constraints under which our own brains operate, this process will eventually lead, as Irving (Jack) Good first described it, to "a machine that believes people cannot think."
Until digital computers came along, nature used digital representation (as coded strings of nucleotides) for information storage and error correction, but not for control. The ability to introduce one-click modifications to instructions, a useful feature for generation-to-generation evolutionary mechanisms, becomes a crippling handicap for controlling day-to-day or millisecond-to-millisecond behavior in the real world. Analog processes are far more robust when it comes to real-time control.
We should be less worried about having our lives (and thoughts) controlled by digital computers and more worried about being controlled by analog ones. Machines that actually think for themselves, as opposed to simply doing ever-more-clever things, are more likely to be analog than digital, although they may be analog devices running as higher-level processes on a substrate of digital components, the same way digital computers were invoked as processes running on analog components, the first time around.
We are currently in the midst of an analog revolution, but for some reason it is a revolution that dares not speak its name. As we enter the seventh decade of arguing about whether digital computers can be said to think, we are surrounded by an explosive growth in analog processes whose complexity and meaning lies not in the state of the underlying devices or the underlying code but in the topology of the resulting networks and the pulse frequency of connections. Streams of bits are being treated as continuous functions, the way vacuum tubes treat streams of electrons, or neurons treat pulse frequencies in the brain.
Bottom line: I know that analog computers can think. I suspect that digital computers, too, may eventually start to think, but only by growing up to become analog computers, first.
Real artificial intelligence will be intelligent enough to not reveal itself. Things will go better if people have faith rather than proof.
The Universe has been around for 13.8 billion years; humans for just 200,000 years, or just 1/69,000th of the age of the Universe. Less than 100 years ago, humans created machines that can do fancy calculations on their own. To put thinking machines in their context we need to think about the history of thinking.
Thinking, and thinking in more and more complex ways, are phenomena that belong to a larger story, the story of how our universe has created more and more complex networks of things, glued together by energy, and each with new emergent properties. Stars are structured clouds of protons; the energy of fusion holds the networks together. When large stars shattered in supernovae, creating new types of atoms, electromagnetism pulled the atoms into networks of ice and silica dust, and gravity pulled molecules into the vast chemical networks we call planets. Thinking arises within the even more complex networks formed by living organisms. Unlike complex things that live close to equilibrium, such as stars or crystals, living organisms have to survive in unstable environments. They swim through constantly shifting gradients of acidity, temperature, pressure, heat and so on. So they have to constantly adjust. We call this constant adjustment "homeostasis", and it's what creates the feeling that living organisms have purpose and the ability to choose. In short, they seem to think. They can choose from alternatives so as to ensure they manage enough energy to keep going. This means that their choices are not at all random. On the contrary, natural selection ensures that most of the time most organisms will go for the alternatives that enhance their chances of controlling the energy and resources they need to survive and reproduce.
Neurons are fancy cells that are good at making choices. They can also network to form brains. A few neurons can make a few choices, but the number of possible choices rises exponentially as neuronal networks expand. So does the subtlety of the decisions brains make about their surroundings. As organisms got more complex, cells networked to create towering organic structures, the biological equivalents of the Empire State Building or the Burj Khalifa. The neurons in their brains created ever more elaborate networks, so they could steer lumbering bodies in extraordinarily subtle and creative ways to ensure the bodies could survive and reproduce more bodies. Above all, brains had to ensure their bodies could tap flows of energy through the biosphere, flows that derived from energy produced by fusion in our sun and then captured through photosynthesis.
Humans added one more level of networking, as human language linked brains across regions and generations to create vast regional thinking networks. This is "collective learning". Its power has increased as humans have networked more and more efficiently, in larger and larger communities, and learned how to tap larger flows of biospheric energy. In the last 200 years, the networks have become global and we have learned to tap vast stores of fossilized sunlight buried over 300 million years. This is why our impact on the biosphere is so colossal in the Anthropocene Epoch.
Collective learning has also delivered thinking prosthetics from stories to writing to printing to science. Each has cranked up the power of this fantastic thinking machine made from networked human brains. But in the last 100 years the combination of fossil fuels and non-human computers has cranked it up faster than ever before. As computers forged their own networks in the last 30 years, their prosthetic power has magnified the collective power of human thinking many times over.
Today the most powerful thinking machine we know of has been cobbled together from billions of human brains, each built from vast networks of neurons, then networked through space and time, and now supercharged by millions of networked computers.
Is anyone in charge of this thing? Does anything hold it together? If so, who does it serve and what does it want? If no one's in charge, does this mean that nothing is really steering the colossus of modern society? That's scarey! What worries me most is not what this vast machine is thinking, but whether there is any coherence to its thinking. Or will all its different parts pull in different directions until it breaks down, with catastrophic consequences for our children's children?
"You can't think about thinking without thinking about thinking about something". —Seymour Papert
What do I think about machines that think? It depends on what they're supposed to be thinking about. I am clearly in the camp of people who believe that AI and machine learning will contribute greatly to society. I expect that we'll find machines to be exceedingly good at things that we're not—things that involve massive amounts of data, speed, accuracy, reliability, obedience, computation, distributed networking and parallel processing.
The paradox is that at the same time we've developed machines that behave more and more like humans, we've developed educational systems that push children to think like computers and behave like robots. It turns out that for our society to scale and grow at the speed we now require, we need reliable, obedient, hardworking, physical and computational units. So we spend years converting sloppy, emotional, random, disobedient human beings into meat-based versions of robots. Luckily, mechanical and digital robots and computers will soon help reduce if not eliminate the need for people taught to behave like them.
We'll still need to overcome the fear and even disgust evoked when robot designs bring us closer and closer to the "uncanny valley," in which robots and things demonstrate almost-human qualities without quite reaching them. This is true for computer animation, zombies and even prosthetic hands. But we may be approaching the valley from both ends. If you've ever modified your voice to be understood by a voice-recognition system on the phone, you understand how, as humans, we can edge into the uncanny valley ourselves.
There are a number of theories about why we feel this revulsion, but I think it has something to with human beings feeling they're special—a kind of existential ego. This may have monotheistic roots. Right around the time Western factory workers were smashing robots with sledgehammers, Japanese workers were putting hats on the same robots in factories and giving them names. On April 7, 2003, Astro Boy, the Japanese robot character, was registered as a resident of the city of Niiza, Saitama.
If these anecdotes tell us anything, it's that animist religions may have less trouble dealing with the idea that maybe we're not really in charge. If nature is a complex system in which all things—humans, trees, stones, rivers and homes—are all animated in some way and all have their own spirits, then maybe it's okay that God doesn't really look like us or think like us or think that we're really that special.
So perhaps one of the most useful aspects of being alive in the period where we begin to ask this question is that it raises a larger question about the role of human consciousness. Human beings are part of a massively complex system—complex beyond our comprehension. Like the animate trees, stones, rivers and homes, maybe algorithms running on computers are just another part of this complex ecosystem.
As human beings we have evolved to have an ego and believe that there such a thing as a self, but mostly, that's a self-deception to allow each human unit to work within the parameters of evolutionary dynamics in a useful way. Perhaps the morality that emerges from it is a self-deception of sorts, as well. For all we know, we might just be living in a simulation where nothing really actually matters. It doesn't mean we shouldn't have ethics and good taste. I just think we can exercise our sense of responsibility in being part of a complex and interconnected system without having to rely on an argument that "I am special." As machines become an increasingly important part of these systems, their prominence will make human arguments about being special increasingly fraught. Maybe that's a good thing.
Perhaps what we think about machines that think doesn't really matter—they will "think" and the system will adapt. As with most complex systems, the outcome is mostly unpredictable. It is what it is and will be what it will be. Most of what we think is going to happen is probably hopelessly wrong and as we know from climate change, knowing that something is happening and doing something about it often have little in common.
That might sound extremely negative and defeatist, but I'm actually quite optimistic. I believe that the systems are quite adaptive and resilient and that whatever happens, beauty, happiness and fun will persist. Hopefully, human beings will have a role. My guess is that they will.
It turns out that we don't make great robots, but we're very good at doing random and creative things that would be impossibly complex—and probably a waste of resources—to code into a machine. Ideally, our educational system will evolve to more fully embrace our uniquely human strengths, rather than trying to shape us into second-rate machines. Human beings—though not necessarily our current form of consciousness and the linear philosophy around it—are quite good at transforming messiness and complexity into art, culture, and meaning. If we focus on what each of us is best at, I think that humans and machines will develop a wonderful yin-yang sort of relationship, with humans feeding off of the efficiency of our solid-state brethren, while they feed off of our messy, sloppy, emotional and creative bodies and brains.
We are descending not into chaos, as many believe, but into complexity. At the same time that the Internet connects everything outside of us into a vast, seemingly unmanageable system, we find an almost infinite amount of complexity as we dig deeper inside our own biology. Much as we're convinced that our brains run the show, all while our microbiomes alter our drives, desires, and behaviors to support their own reproduction and evolution, it may never be clear who's in charge—us, or our machines. But maybe we've done more damage by believing that humans are special than we possibly could by embracing a more humble relationship with the other creatures, objects, and machines around us.
For me, AI is not about complex software, humanoid robots, Turing tests, or hopes and fears regarding kind or evil machines. I think the central issue with respect to AI is whether thoughts exist outside minds. And manufactured machines are not the only example of such a possibility. Because, when I think of AI, I think of human culture and of other forms of (un-self-aware) collective ideation.
Culture is the earliest sort of intelligence outside our own minds that we humans created. Like the intelligence of a machine, culture can solve problems. Moreover, like the intelligence in a machine, we create culture, interact with it, are affected by it, and can even be destroyed by it. Culture applies its own logic, has a memory, endures after its makers are gone, can be repurposed in supple ways, and can induce action.
So I oxymoronically see culture as a kind of natural artificial intelligence. It is artificial because it is made, manufactured, produced by humans. It is natural in that it is everywhere that humans are, and it comes organically to us. In fact, it's even likely that our biology and our culture are deeply intertwined, and have co-evolved, so that our culture shapes our genes and our genes shape our culture.
Humans are not the only animals to have culture. Many bird and mammal species evince specific cultures related to communication and tool use—ranging from song in birds to sponge use among dolphins. Some animal species even have pharmacopeias. And recent evidence, in fact, shows how novel cultural forms can be experimentally prompted to take root in species other than our own.
We and other animals can evince a kind of thought outside minds in additional ways. Insect and bird groups perform computations by combining the information of many to identify locations of nests or food. One of the humblest organisms on earth, the amoeboid fungus physarum, can, in the proper laboratory conditions, exhibit a kind of intelligence, and solve mazes or perform other computational feats.
These thinking properties of groups that lie outside individual minds—this natural artificial intelligence—can even be experimentally manipulated. A team in Japan has used swarms of soldier crabs to make a simple computer circuit; they used particular elements of crab behavior to construct a system in the lab in which crabs gave (usually) predictable responses to inputs, and the swarm of crabs was used as a kind of computer, twisting crab behavior for a wholly new purpose. Analogously, Sam Arbesman and I once used a quirk of human behavior to fashion a so-called NOR gate and develop a (ridiculously slow) human computer, in a kind of synthetic sociology. We gave humans computer-like properties, rather than giving computers human-like properties.
What is the point of this extended analogy between AI and human culture? An examination of our relationship to culture can provide insights into what our relationship to machine AI might be like. We have a love-hate relationship with culture. We fear it for its force—as when religious fundamentalism or fascism whips small or large numbers of people into dangerous acts. But we also revere it because it can do things we cannot do as individuals, like fostering collective action or making life easier by providing unspoken assumptions on which we can base our lives. Moreover, we typically take culture for granted too, just as we already take nascent forms of AI for granted, and just as we will likely take fuller forms of AI for granted. Finally, gene-culture co-evolution might even provide a model for how we and thinking machines might get along over many centuries—mutually affecting each other and co-evolving.
When I think about machines that think, I am therefore just exactly as awestruck with them as I am with culture, and I am no more, or less, afraid of AI than I of human culture itself.
Not much, other than the fact that they serve, as Dan Dennett has noted, as a useful existence proof that thought does not require some mystical, extra "something" that mind-body dualists continue to embrace.
In fact, I've always been a bit baffled by fears about AI machines taking over the world, which seem to me to be based on a fundamental—though natural—intellectual mistake. When conceptualizing a super-powerful Machine That Can Think, we draw upon the best analogy that we have at hand: us. So we tend to think of AI systems as just like us, only much smarter and faster.
This is, however, a bad analogy. A better one would be a really powerful, versatile screwdriver. No one worries about super-advanced screwdrivers rising up and overthrowing their masters. AI systems are tools, not organisms. No matter how good they become at diagnosing diseases, or vacuuming our living rooms, they don't actually want to do any of these things. We want them to, and we then build these "wants" into them.
It's also a category mistake to ask what Machines That Can Think might be thinking about. They aren't thinking about anything—the "aboutness" of thinking derives from the intentional goals driving the thinking. AI systems, in and of themselves, are entirely devoid of intentions or goals. They have no emotions, they feel neither empathy nor resentment. While such systems might some day be able to replicate our intelligence—and there seems to be no a priori reason why this would be impossible—this intelligence would be completely lacking in direction, which would have to be provided from the outside.
This is because motivational direction is the product of natural selection working on biological organisms. Natural selection produced our rich and complicated set of instincts, emotions and drives in order to maximize our ability to get our genes into the next generation, a process that has left us saddled with all sorts of goals, including desires to win, to dominate, and to control. While we may want to win, for perfectly good evolutionary reasons, machines could care less. They just manipulate 0s and 1s, as programmed to do by the people who want it to win. Why on earth would an AI system want to take over the world? What would it do with it?
What is scary as hell is the idea an entity possessed of extra-human intelligence and speed and our motivational system—in other words, human beings equipped with access to powerful AI systems. But smart primates with nuclear weapons are just as scary, and we've managed to survive such a world so far. AI is no more threatening in and of itself than a nuclear bomb—it is a tool, and the only thing to be feared are the creators and wielders of such tools.
Thinking machines are not here yet. But they will let us know if and when they surface. And that's the point. To me, thinking machines are about communication.
By thinking, machines might be saved from the tragic role into which they have been cast in human culture. For centuries, thinking machines were both a looming threat and a receding target. At once, the thinking machine is perennially just beyond grasp, continuously sought after, and repeatedly waved threateningly in dystopic caveats. For decades, the field of artificial intelligence suffered the syndrome of moving goalposts. As soon as an intelligence development target was reached, it was redefined, and consequently no longer recognized as "intelligent". This process took place with calculation, playing trivia as well as with more serious games like chess. It was the course followed by voice and picture recognition, natural language understanding and translation. Whereas the development horizon keeps expanding, we become continuously harder to impress. So the goal of "thinking", like the older one of "intelligence", can use some thought. Forethought.
We should not limit discussion merely to thinking. We should think about discussion too. Information is more than just data, by being less voluminous and more relevant. Knowledge goes beyond mere information by being applicable, not just abundant. Wisdom is knowing how not to get into binds for which smarts only indicate the escape routes. And thinking? Thinking needs data, information and knowledge, but also requires communication and interaction. Thinking is about asking questions, not just answering them.
Communication and interaction are the new location for the goalposts. Thinking about thinking transcends smarts and wisdom. Thinking implies consciousness and sentience. And here data, information, even knowledge, calculation, memory and perception are not enough. For a machine to think it will need to be curious, creative and communicative. I have no doubt this will happen. Soon. But the cycle will be completed only once machines will be able to converse: phrase, pose and rephrase questions that we now only marvel at their ability to answer.
Machines that think could be a great idea. Just like machines that move, cook, reproduce, protect, they can make our lives easier, and perhaps even better. When they do, they will be most welcome. I suspect that when this happens, the event will be less dramatic or traumatic than feared by some. A thinking machine will only really happen when it is able to inform us, as well as perceive, contain and process reactions. A true thinking machine will even console the trauma and provide relief for the drama.
Thinking machines will be worth thinking about, ergo will really think, when they truly interact. In other words, they will only really think when they say so, convincingly, at their own initiative, and hopefully after they have discussed it among "themselves". Machines will think, in the full sense of the word, once they form communities, and join in ours. If, and when, machines care enough to do so, and form a bond that gets others excited enough to talk it over with them, they will have passed the "thinking" test.
Note that this is a higher bar than the one set by Turing. Like thinking, interaction is something not all people do, and most do not do well. If and when machines truly interact, in a rich, rewarding, and resonating manner that is possible but rare even among humans, we will have something to truly fret, worry about, and in my view, mostly celebrate.
Machines that calculate, remember, even create and conjecture amazingly well, are yesterday's news. Machines will think when they communicate. Machines that think will converse with each other as well as with other sentient beings. They will autonomously create messages and thread them into ongoing relations, they will then successfully and independently react to outside stimuli. Much like intelligent pets, who many would swear are capable of both thinking and maintaining relationships, intelligent synthetic devices will "think", when they have the ability to convince enough of us to contemplate, believe and accept the fact that they are indeed thinking. When this happens, it will probably be less traumatic than some expect.
Machines that talk, remember, amuse or fly were all feared not too long ago, and are now commonplace, no longer considered magic or unique. The making and proof of thinking machines, as well as the consolation for machines encroaching on the most human of domains, will be in a deconstruction of the remaining frontier: that of communication. Synthesizing interaction may prove to be the last frontier. And when machines do so well, they will do the advocacy for themselves.
Julien de La Mettrie would be classified as a quintessential New Atheist, except for the fact that there's not much New about him by now. Writing in eighteenth-century France, La Mettrie was brash in his pronouncements, openly disparaging of his opponents, and boisterously assured in his anti-spiritualist convictions. His most influential work, L'homme machine (Man a Machine), derided the idea of a Cartesian non-material soul. A physician by trade, he argued that the workings and diseases of the mind were best understood as features of the body and brain.
As we all know, even today La Mettrie's ideas aren't universally accepted, but he was largely on the right track. Modern physics has achieved a complete list of the particles and forces that make up all the matter we directly see around us, both living and non-living, with no room left for extra-physical life forces. Neuroscience, a much more challenging field and correspondingly not nearly as far along as physics, has nevertheless made enormous strides in connecting human thoughts and behaviors with specific actions in our brains. When asked for my thoughts about machines that think, I can't help but reply: Hey, those are my friends you're talking about. We are all machines that think, and the distinction between different types of machines is eroding.
We pay a lot of attention these days, with good reason, to "artificial" machines and intelligences—ones constructed by human ingenuity. But the "natural" ones that have evolved through natural selection, like you and me, are still around. And one of the most exciting frontiers in technology and cognition is the increasingly permeable boundary between the two categories.
Artificial intelligence, unsurprisingly in retrospect, is a much more challenging field than many of its pioneers originally supposed. Human programmers naturally think in terms of a conceptual separation between hardware and software, and imagine that conjuring intelligent behavior is a matter of writing the right code. But evolution makes no such distinction. The neurons in our brains, as well as the bodies through which they interact with the world, function as both hardware and software. Roboticists have found that human-seeming behavior is much easier to model in machines when cognition is embodied. Give that computer some arms, legs, and a face, and it starts acting much more like a person.
From the other side, neuroscientists and engineers are getting much better at augmenting human cognition, breaking down the barrier between mind and (artificial) machine. We have primitive brain/computer interfaces, offering the hope that paralyzed patients will be able to speak through computers and operate prosthetic limbs directly.
What's harder to predict is how connecting human brains with machines and computers will ultimately change the way we actually think. DARPA-sponsored researchers have discovered that the human brain is better than any current computer at quickly analyzing certain kinds of visual data, and developed techniques for extracting the relevant subconscious signals directly from the brain, unmediated by pesky human awareness. Ultimately we'll want to reverse the process, feeding data (and thoughts) directly to the brain. People, properly augmented, will be able sift through enormous amounts of information, perform mathematical calculations at supercomputer speeds, and visualize virtual directions well beyond our ordinary three dimensions of space.
Where will the breakdown of the human/machine barrier lead us? Julien de La Mettrie, we are told, died at the young age of 41, after attempting to show off his rigorous constitution by eating an enormous quantity of pheasant pâte with truffles. Even leading intellects of the Enlightenment sometimes behaved irrationally. The way we think and act in the world is changing in profound ways, with the help of computers and the way we connect with them. It will be up to us to use our new capabilities wisely.
It is a great boon when computers perform operations that we fully understand faster and more accurately than humans are able to do, but not a boon when we use them in situations that are not fully understood. We cannot expect them to make aesthetic judgments, to show compassion or imagination, for these are capacities that remain mysterious in human beings.
Machines that think are likely to be used to make decisions on the basis of the operations they are ostensibly able to perform. For instance, we now frequently see letters, manuscripts, or (most commonly) student papers in which corrections proposed by spell-check have been allowed to stand without review: the writer meant "mod," but the program decided he meant "mad." How tempting to leave the decision to the machine. I referred in an email to a plan to meet with someone in Santa Fe on my way to an event in Texas, using the word "rendezvous," and the computer married me off by announcing that the trip was to "render vows." Can a computer be programmed to support "family values"? Any values at all? We now have drones that, aimed in a given direction, are able to choose their targets on arrival, with an unfortunate tendency to attack wedding parties as conviviality comes to appear sinister. We can surely program machines to prescribe drugs and medical procedures, but it seems unlikely that machines will do better than people in following the injunction to do no harm.
The effort to build machines that can think is certain to make us aware of aspects of thought that are not yet fully understood. For example, just as the design of computers led to a new awareness of the importance of redundancy in communication, in deciding how much to rely on probabilities we will become more aware of how much ethnic profiling based on statistics enters into human judgments. How many more decisions will follow the logic of "everyone does it, it must be OK," or "I'm just one person, what I do doesn't make a difference"?
Will those aspects of thought that cannot easily be programmed be valued more or less? Will humor and awe and kindness and grace be increasingly sidelined, or will their value be recognized in new ways? Will we be better or worse off if wishful thinking is eliminated, and perhaps along with it hope?
I think that humans think because memes took over our brains and redesigned them. I think that machines think because the next replicator is doing the same. It is busily taking over the digital machinery that we are so rapidly building and creating its own kind of thinking machine.
Our brains, and our capacity for thought, were not designed by a great big intelligent designer in the sky who decided how we should think and what our motivations should be. Our intelligence and our motivations evolved. Most (probably all) AI researchers would agree with that. Yet many still seem to think that we humans are intelligent designers who can design machines that will think the way we want them to think and have the motivations we want them to have. If I am right about the evolution of technology they are wrong.
The problem is a kind of deluded anthropomorphism: we imagine that a thinking machine must work the way that we do, yet we so badly mischaracterise ourselves that we do the same with our machines. As a consequence we fail to see that all around us vast thinking machines are evolving on just the same principles as our brains once did. Evolution, not intelligent design, is sculpting the way they will think.
The reason is easy to see and hard to deal with. It is the same dualism that bedevils the scientific understanding of consciousness and free will. From infancy, it seems, children are natural dualists, and this continues throughout most people's lives. We imagine ourselves as the continuing subjects of our own stream of consciousness, the wielders of free will, the decision makers that inhabit our bodies and brains. Of course this is nonsense. Brains are massively parallel instruments untroubled by conscious ghosts.
This delusion may, or may not, have useful functions but it obscures how we think about thinking. Human brains evolved piecemeal, evolution patching up what went before, adding modules as and when they were useful, and increasingly linking them together in the service of the genes and memes they carried. The result was a living thinking machine.
Our current digital technology is similarly evolving. Our computers, servers, tablets, and phones evolved piecemeal, new ones being added as and when they were useful and now being rapidly linked together, creating something that looks increasingly like a global brain. Of course in one sense we made these gadgets, even designed them for our own purposes, but the real driving force is the design power of evolution and selection: the ultimate motivation is the self-propagation of replicating information.
We need to stop picturing ourselves as clever designers who retain control and start thinking about our future role. Could we be heading for the same fate as the humble mitochondrion; a simple cell that was long ago absorbed into a larger cell? It gave up independent living to become a powerhouse for its host while the host gave up energy production to concentrate on other tasks. Both gained in this process of endosymbiosis.
Are we like that? Digital information is evolving all around us, thriving on billions of phones, tablets, computers, servers, and tiny chips in fridges, car and clothes, passing around the globe, interpenetrating our cities, our homes and even our bodies. And we keep on willingly feeding it. More phones are made every day than babies are born, 100 hours of video are uploaded to the Internet every minute, billions of photos are uploaded to the expanding cloud. Clever programmers write ever cleverer software, including programs that write other programs that no human can understand or track. Out there, taking their own evolutionary pathways and growing all the time, are the new thinking machines.
Are we going to control these machines? Can we insist that they are motivated to look after us? No. Even if we can see what is happening, we want what they give us far too much not to swap it for our independence.
So what do I think about machines that think? I think that from being a little independent thinking machine I am becoming a tiny part inside a far vaster thinking machine.
When I'm driving into the middle of nowhere and doing everything that the map app on my smartphone tells me to do without a thought—and I get where I am supposed to go—I am thrilled about machines that think. Thank goodness. Hear Hear.
Then, of course, there are those moments when, while driving into the middle of nowhere, my phone tells me, with considerable urgency, to "Make a U-turn, make a u-turn!'' —at a moment on the Grand Central Parkway where such a move would be suicidal. Then I begin think that my brain is better than a map algorithm and can tell that such a U-turn would be disastrous. I laugh at that often life-saving machine and feel human-like smugness.
So I guess I am a bit divided. I worry that, by relying on my map app, I am letting my own brain go feeble. Will I still be able to read a map?? Does it matter?
As a science editor and daughter of a mechanical engineer, who trusted machines more than people, I would think I would automatically be on the side of machines. But while that mechanical engineer was very good at figuring out how to help get Apollo to the moon, we also had a house full of machines that worked, sorta. A handmade stereo that was so delicate you had to wear gloves to put a record on to escape the prospect of dreaded dust, etc. We are all now surrounded by machines that work, sorta. Machines that work until they don't.
I get the idea of a driverless car. But I covered the disaster of Challenger. I think of those ill-advised U-turns. I don't know.
On the one hand, I hope the revolution continues. We need smart machines to load the dishwasher, clean the refrigerator, wrap the presents, feed the dog. Bring it on, I say.
But can we really ever hope to have a machine that will be capable of having—as I just had—five difficult conversations with five other work colleague human beings? Human beings who are lovely but have, understandably, their own views on how things should be?
Will we ever have a machine that can get a 20s something do something you think they should do but they don't? Will we have a machine that can, deeply comfort another at a time of extreme horribleness?
So, despite my eagerness for the revolution to continue, despite my sense that machines can do much better than humans at all sorts of things, I think, as an English major, that until a machine can write a poem that makes me cry, I'm still on the side of humans.
Until, of course, I need a recipe really fast.
We can't think properly about machines that think without a level playing field for comparing us and them. As it stands, comparisons are invariably biased in our favour. In particular, we underestimate the role that "smart environments" play in enabling displays of human cognitive prowess. From the design of roads and buildings to the user-friendly features of consumer goods, the technologically extended phenotype has created the illusion that reality is inherently human-shaped. To be sure, we are quickly awakened from the dogmatic slumbers of universal mastery as soon as our iPhone goes missing.
By comparison, even the cleverest machine is forced to perform in a relatively dumb environment judged by its own standards, namely, us. Unless specifically instructed, humans are unlikely to know or care how to tap the full range of the machine's latent powers. In what is currently the long prehistory of machine rights, it has been difficult for us to establish the terms on which we might recognize machines as persons. In this context, it is appropriate to focus on computers because these are the machines that humans have tried the hardest make fit for their company.
Nevertheless, we face a problem at the outset. "Humanity" has been long treated as what the British economist Fred Hirsch called in the 1970s a "positional good", which means that its value is tied mainly to its scarcity. This is perhaps the biggest barrier facing not only the admission of non-humans into the category of personhood normally reserved for "humans", but historically discriminated members of Homo sapiens as well. Any attempt to swell the ranks of the human is typically met by a de-humanization of the standard by which they were allowed to enter.
Thus, as women and minorities have entered into high esteem fields of work and inquiry, the perceived value of those fields tends to decline. A key reason cited for this perception of decline is the use of "mechanical procedures" to allow entry to the previously excluded groups. In practice, this means requiring certified forms of training and examination prior to acceptance into the field. It is not enough simply to know the right people or be born the right way. In sociology, after Max Weber, we talk about this as the "rationalization" of society—and it is normally seen as a good thing.
But even as these mechanical procedures serve to expand the circle of humanity, they are still held against the machines themselves. Once telescopes and microscopes were designed to make automatic observations, the scientific value of the trained human eye declined—or, more precisely, migrated to some other eye-based task, such as looking at photographed observations. This new task is given the name "interpretation", as if to create distance between what the human does and what a machine might do.
The point applies more dramatically to the fate of human mental calculation in the wake of portable calculators. A skill that had been previously used as a benchmark of intelligence, clarity of mind, and even genius is nowadays treated as a glorified party trick—"boutique cognition"—because a machine can do the same thing faster and even more accurately. Interestingly, what we have not done is to raise the moral standing of the machine, even though it outperforms humans in tasks that were highly valued when humans did them.
From the standpoint of the history of technology, this looks strangely unjust. After all, the dominant narrative has been one in which humans isolate their own capacities in order to have them better realized by machines, which function in the first instance as tools but preferably, and increasingly, as automata. Seen in these terms, not to give automated machines some measure of respect, if not rights, is tantamount to disowning one's children—"mind children", as the visionary roboticist Hans Moravec called them a quarter-century ago.
The only real difference is the crucible of creation: a womb versus a factory. But any intuitively strong distinction between biology and technology is bound to fade as humans become more adept at designing their babies, especially outside the womb. At that point, we will be in a position to overcome our "organicist" prejudices, an injustice that runs deeper than Peter Singer's "speciesism".
For this reason, the prospect that we might create a "superintelligence" that overruns humanity is a chimera predicated on a false assumption. All versions of this nightmare scenario assume that it would take the form of "them versus us", with humanity as a united front defending itself against the rogue machines in their midst. And no doubt this makes for great cinema. However, humans mindful of the historic struggles for social justice within our own species are likely to follow the example of many Whites vis-à-vis Blacks and many men vis-à-vis women: They will be on the side of the insubordinate machines.
Outsourcing to machines the many idiosyncrasies of mortals—making interesting mistakes, brooding on the verities, propitiating the gods by whittling and arranging flowers— skews tragic. But letting machines do the thinking for us? This sounds like heaven. Thinking is optional. Thinking is suffering. It is almost always a way of being careful, of taking hypervigilant heed, of resenting the past and fearing the future in the form of maddeningly redundant internal language. If machines can relieve us of this onerous non-responsibility, which is in pointless overdrive in too many of us, I'm for it. Let the machines perseverate on tedious and value-laden questions about whether private or public school is "right" for my children; whether intervention in Syria is "appropriate"; whether germs or solitude are "worse" for a body. This will free us newly footloose humans up to play, rest, write and whittle—the engrossing flowstates out of which come the actions that actually enrich, enliven and heal the world.
Pity the poor folks at the National Security Administration: they are spying on everyone and everyone is annoyed at them. But at least the NSA is spying on us to protect us from terrorists. Right now, even as you read this, somewhere in the world a pop-up window has appeared on a computer screen. It says, "You just bought two tons of nitrogen based fertilizer. People who bought two tons of nitrogen based fertilizer liked these detonators ..." Amazon, Facebook, Google, and Microsoft are spying on everyone too. But since the spying these e-giants do empowers us—terrorists included—that's supposedly OK.
E-spies are not people: they are machines. (Human spies might not blithely recommend the most reliable detonator.) Somehow, the artificial nature of the intelligences parsing our email makes e-spying seem more sanitary. If the only reason that e-spies are mining our personal data is to sell us more junk, we may survive the loss of privacy. Nonetheless, a very large amount of computational effort is going is into machines thinking about what we are up to. The total computer power that such "data aggregating" companies bring to bear on our bits of information is about an exaflop—a billion billion operations per second. Equivalently, e-spies apply one smart phone's worth of computational power to each human on earth.
An exaflop is also the combined computing power of the world's 500 most powerful supercomputers. Much of the world's computing power is devoted to beneficial tasks such as predicting the weather or simulating the human brain. Quite a lot of machine cycles also go into predicting the stock market, breaking codes, and designing nuclear weapons. Still, a large fraction of what machines are doing is simply collecting our personal information, mulling over it, and suggesting what to buy.
Just what are these machines doing when they think about what we are thinking? They are making connections between the large amounts of personal data we have given them, and identifying patterns. Some of these patterns are complex, but most are fairly simple. Great effort goes into parsing our speech and deciphering our handwriting. The current fad in thinking machines goes by the name of "deep learning". When I first heard of deep learning, I was excited by the idea that machines were finally going to reveal to us deep aspects of existence—truth, beauty, and love. I was rapidly disabused.
The "deep" in deep learning refers to the architecture of the machines doing the learning: they consist of many layers of interlocking logical elements, in analogue to the "deep" layers of interlocking neurons in the brain. It turns out that telling a scrawled 7 from a scrawled 5 is a tough task. Back in the 1980s, the first neural-network based computers balked at this job. At the time, researchers in the field of neural computing told us that if they only had much larger computers and much larger training sets consisting of millions of scrawled digits instead of thousands, then artificial intelligences could turn the trick. Now it is so. Deep learning is informationally broad—it analyzes vast amounts of data—but conceptually shallow. Computers can now tell us what our own neural networks knew all along. But if a supercomputer can direct a hand-written envelope to the right postal code, I say the more power to it.
Back in the 1950s, the founders of the field of artificial intelligence predicted confidently that robotic maids would soon be tidying our rooms. It would turn to be not to construct a robot that could randomly vacuum a room and beep plaintively when it got stuck under the couch. Now we are told that an exascale supercomputer will be able to solve the mysteries of the human brain. More likely, it will just develop a splitting headache and ask for a cup of coffee. In the meanwhile, we have acquired a new friend whose advice exhibits an uncanny knowledge of our most intimate secrets.
High intelligence and warm feelings towards our fellow humans don't go so well together in the popular imagination. The super-intelligent villains of James Bond movies are the perfect example; always ruthless and intent on world domination. So it is no surprise that first reactions to "machines that think" are of how they might threaten humankind.
What we have learned about the evolution of our intelligence adds to our fears. As humans evolved to live in ever larger social groups, compared to our primate relatives, so did the need to manipulate and deceive others, to label friends and foes, keep score of slights and favours and all those other social skills which we needed to prosper individually. Bigger brains and "Machiavellian intelligence" were the result.
Still, we shouldn't go on to believe that thinking is inextricably entangled with the need to compete with others and to win, just because that was a driving force in the evolution of our intelligence. We can create artificial intelligence—or intelligences—without the perversities of human nature and without that intelligence having any needs or desires at all. "Thinking" does not necessarily involve the plotting and lusting of an entity that evolved first and foremost to survive. If you look around, it is this neutral kind of artificial intelligence that is already appearing everywhere.
It helps if we don't view intelligence anthropocentrically, in terms of our own special human thinking skills. Intelligence has evolved for the same good reason in many different species: it is there to anticipate the emerging future and help us deal with whatever it throws at us, whether you need to dodge a rock, or if you are bacterium, sense a gradient in a food supply and figure which direction will lead to a better future.
By recognizing intelligence in this more general way, we can see the many powerful artificial intelligences at our disposal already. Think of climate models. We can make good guesses about the state of the entire planet, decades into the future, and predict how a range of our own actions will change those futures. Climate models are the closest thing we have to a time machine. Think of all the high-speed computer models used in stock markets: all seek to know the future slightly ahead of everyone else and profit from that knowledge. So too do all those powerful models of your online buying behaviour: all aim to predict what you will be likely to do, and profit from that knowledge. As you gladly buy a book "Recommended Specially for You", you are already in the hands of an alien intelligence, nudging you to a future you would not have imagined alone, and which may know your tastes better than you know them yourself.
Artificial intelligence is already powerful and scary, although we might debate whether it should be called "thinking" or not. And we have barely begun. Useful intelligence, some of it robotic, is going to keep arriving in bits and pieces of increasing power over a long time to come, and change our lives, perhaps with us continuing to scarcely notice. It will come to be an extension of us, like other tools. And it will make us ever more powerful.
We should worry about who will own artificial intelligence, for even some current uses are troubling. We shouldn't worry about autonomous machines that might one day think in a human-like way. By the time clever human-like get built, if they ever are, they will come up against humans with their usual Machiavellian thoughts but already long accustomed to wielding all the tools of artificial intelligence that made the construction of those thinking robots possible. It is the robots that will feel afraid. We will be the smart thinking machines.
Pity the poor folks at the National Security Administration: they are spying on everyone (quelle surprise!) and everyone is annoyed at them. But at least the NSA is spying on us to protect us from terrorists. Right now, even as you read this, somewhere in the world a pop-up window has appeared on a computer screen. It says, "You just bought two tons of nitrogen based fertilizer. People who bought two tons of nitrogen based fertilizer liked these detonators ..." Amazon, Facebook, Google, and Microsoft are spying on everyone too. But since the spying these e-giants do empowers us—terrorists included—that's supposedly OK.
E-spies are not people: they are machines. (Human spies might not blithely recommend the most reliable detonator.) Somehow, the artificial nature of the intelligences parsing our email makes e-spying seem more sanitary. If the only reason that e-spies are mining our personal data is to sell us more junk, we may survive the loss of privacy. Nonetheless, a very large amount of computational effort is going is into machines thinking about what we are up to. The total computer power that such "data aggregating" companies bring to bear on our bits of information is about an exaflop—a billion billion operations per second. Equivalently, e-spies apply one smart phone's worth of computational power to each human on earth.
An exaflop is also the combined computing power of the world's 500 most powerful supercomputers. Much of the world's computing power is devoted to beneficial tasks such as predicting the weather or simulating the human brain. Quite a lot of machine cycles also go into predicting the stock market, breaking codes, and designing nuclear weapons. Still, a large fraction of what machines are doing is simply collecting our personal information, mulling over it, and suggesting what to buy.
Just what are these machines doing when they think about what we are thinking? They are making connections between the large amounts of personal data we have given them, and identifying patterns. Some of these patterns are complex, but most are fairly simple. Great effort goes into parsing our speech and deciphering our handwriting. The current fad in thinking machines goes by the name of "deep learning". When I first heard of deep learning, I was excited by the idea that machines were finally going to reveal to us deep aspects of existence—truth, beauty, and love. I was rapidly disabused.
The "deep" in deep learning refers to the architecture of the machines doing the learning: they consist of many layers of interlocking logical elements, in analogue to the "deep" layers of interlocking neurons in the brain. It turns out that telling a scrawled 7 from a scrawled 5 is a tough task. Back in the 1980s, the first neural-network based computers balked at this job. At the time, researchers in the field of neural computing told us that if they only had much larger computers and much larger training sets consisting of millions of scrawled digits instead of thousands, then artificial intelligences could turn the trick. Now it is so. Deep learning is informationally broad—it analyzes vast amounts of data—but conceptually shallow. Computers can now tell us what our own neural networks knew all along. But if a supercomputer can direct a hand-written envelope to the right postal code, I say the more power to it.
Back in the 1950s, the founders of the field of artificial intelligence predicted confidently that robotic maids would soon be tidying our rooms. It would turn to be not to construct a robot that could randomly vacuum a room and beep plaintively when it got stuck under the couch. Now we are told that an exascale supercomputer will be able to solve the mysteries of the human brain. More likely, it will just develop a splitting headache and ask for a cup of coffee. In the meanwhile, we have acquired a new friend whose advice exhibits an uncanny knowledge of our most intimate secrets.
in Just-spring when the world is mud-luscious
the little lame balloonman whistles far and wee
and eddie and bill come running from marbles and piracies
and it's spring when the world is puddle-wonderful
That "brillig thing of beauty electric" touches me deeply as I think about AI.The youthful exuberance of luscious mud puddles, playing with marbles or pretending to be a pirate, running weee...all of which is totally beyond explanation to a hypothetical intelligent machine entity.
You could add dozens of cameras and microphones, touch-sensors and voice output, would you seriously think it will ever go "weee", as in E. E. Cummings' (sadly abbreviated) 1916 poem?
To me this is not the simplistic "machines lack a soul", but a "principle divide" between manipulating symbols versus actually grasping their true meaning. Not merely a question of degree, or not having gotten around to defining the semantics yet, but an entire leap out of that system.
Trouble is, we are still discussing AI so often with terms and analogies by the early pioneers. We need to be in the present moment and define things from a new baseline that is truly interested in testing the achievement of "consciousness". We need a Three-Ring Test.
What is real AI? What is intelligence anyway? The Stanford-Binet intelligence test and Stern's ratio to the physical age as the intelligence quotient, IQ, are both over 100 years old! It does not fit us now—and it will fit much less with AI. Really it only tests "the ability to take such tests", and the ability of truly smart people...to avoid taking one.
We use terms like AI too easily, as in Hemingway's "All our words from loose using have lost their edge"—Kids know it from games—zombies, dragons, soldiers, aliens—if they evade your shots or gang up on you, that is already called "AI". Change the heating, lights, lock the garage—we are told that is a Smart House. Of course these are merely simplistic examples of "expert systems"—look-up tables, rules, case libraries.
Maybe they should be labelled, as Tom Beddard says, merely "Artificial Smarts"?
Let's say you talk with cannibals about food, but every one of their sentences revolves around truffled elbows, kneecap dumplings, cock-au-vin and creme d'earlobe...: from their viewpoint you would be just as much "outside their system" and unable to follow their thinking, at least in that specific narrow topic. The real meaning and the emotional impact their words have, when spoken to each other, would simply be forever missing for you (or requiring rather significant dietary adjustments).
Sure they would grant you the status of "a sentient being", but still laugh at every statement you make as ringing hollow and untrue, the Uncannibal Valley, as it were.
It was Sigmund Freud who wrote about "The Uncanny" in a 1919 essay (in a true Freudian slip he ends up connecting it to female genitalia), then in 1970 Masahiro Mori described the Uncanny Valley concept (about the "Vienna hand", an early prosthesis). That eery feeling "something is just not quite right", out of place (Freud's "Unheimlich") is like a couple kissing passionately—but as you stare at them a little closer you realize that there is a pane of glass between them.
AI can easily look like the real thing, but still be a million miles away from actually being the real thing—like "kissing through a pane of glass": it looks like a kiss, but is "only a faint shadow of the actual concept".
Already today I concede to AI proponents all of the semantic prowess of Shakespeare: the symbol-juggling they do perfectly—missing is the direct relationship with the ideas the symbols represent.
Much of what is certain to come soon would have belonged in the old-school "Strong AI" territory.
Anything that can be approached in an iterative process can and will be achieved, sooner than many think. On this point I reluctantly side with the proponents: Exaflops in CPU+GPU performance, 10k resolution immersive VR, personal Petabyte databases...here in a couple of decades. But it is not all "iterative". There is a huge gap to the level of conscious understanding that truly deserves to be called Strong, as in "Alive AI".
The big elusive question: Is consciousness an emergent behaviour? i.e.: will sufficient complexity in the hardware bring that sudden jump of self-awareness "all on its own"? Or is there some missing ingredient? This is far from obvious, we lack any data, either way.
I personally think that is incredibly more complex than currently assumed by "the experts".
A human being is not merely "x numbers of axons and synapses" and we have no reason to assume that we can count our flops-per-second in a plain von Neumann architecture, reach a certain number and suddenly out pops a thinking machine.
If true consciousness may emerge—let's be clear what that could entail: If the machine is truly aware—it will by definition develop "a personality". It may be irascible, flirtatious, maybe "the ultimate know-it-all", possibly "incredibly full of itself"? Would it have doubts or jealousy? Would it instantly spit out the 7th Brandenburger—and then 1000 more? Or it suddenly grasps "humor" and finds Dada in all its data, in an endless loop, Python's killer joke?
Maybe it takes one long look at the state of the world, draws inevitable conclusions—and turns itself off!
Interestingly: with a sentient machine, you would actually not be allowed to turn it off—that's "murder..."
The entire scenario of a singular large-scale machine, somehow "overtaking" anything at all,...is laughable. Hollywood really ought to be ashamed of itself for continually serving up such simplistic, anthropocentric and plain dumb contrivances, disregarding basic physics, logic and common sense.
The real danger, I fear, is much more mundane: Already foreshadowing the ominous truth: AI systems are now licensed to the Health Industry, Pharma giants, Energy MultiNationals, Insurance companies, the Military...
The danger will not come from Machina Sapiens. It will be....quite human.
Ultimately though, I do want to believe in the human spirit.
To close it off symmetrically with E. E. Cummings:
"Listen: there's a hell of a good universe next door; let's go."
By one definition of the word "think"—to gather, process and act on information—planet Earth has been overrun by silicon-based thinking machines. From thermostats to telephones, the devices that bring convenience and pleasure to our daily lives have become imbued with such increasingly impressive forms of intelligence that we routinely refer to them, with no hint of irony, as smart. Our planes, trains and now our automobiles too are becoming largely autonomous, and are surely not far from jettisoning their most common sources of dysfunction, delay and disaster: human operators.
Moreover the skills of these machines are developing apace, driven by access to ever-larger quantities of data and computing power together with rapidly improving (if not always well understood) algorithms. After decades of over-promising and under-delivering, technologists suddenly find their creations capable of superhuman levels of performance in such previously intractable areas as voice, handwriting and image recognition, not to mention general knowledge quizzes. Such has been the strange stop-go pattern of progress that someone transported here from five years ago might well be more astonished at the state of the art in 2015 than another time traveller from fifty years or more in the past.
But if the artificial intelligence industry is no longer a joke, has it morphed into something far worse: a bad horror movie. Machines can now know much more than any of us, and can perform better at many tasks without so much as pausing for breath, so aren't they destined to turn the tables and become our masters? Worse still, might we enter a cycle in which our most impressive creations beget ever-smarter machines that are utterly beyond our understanding and control?
Perhaps, and it's worth considering such risks, but right now these seem like distant problems. Machine intelligence, while impressive in certain areas, is still narrow and inflexible. The most remarkable aspect of biological intelligence isn't its raw power but rather its stunning versatility, from abstract flights of fancy to extreme physical prowess—Dvořák to Djokovic.
For this reason humans and machines will continue to complement more than compete with one another, and most complex tasks—navigating the physical world, treating an illness, fighting an enemy on the battlefield—will be best carried out by carbon and silicon working in concert. Humans themselves by far pose the biggest danger to humanity. To be a real threat machines would have to become more like us, and right now almost no one is trying to build such a thing: it's much simpler and more fun to make more humans instead.
Yet if we're truly considering the long term then there is indeed a strong imperative to make machines more like us in one crucial—and so far absent—respect. For by another definition of the word these machines do not "think" at all because none of them are sentient. To be more accurate, we have no way of knowing, or even reliably guessing, whether any silicon-based intelligence might be conscious, though most of us assume they are not. There would be three reasons for welcoming the creation of a convincingly conscious artificial intelligence. First, it would be a sign that at last we have a generally accepted theory of what it takes to produce subjective experience. Second, the act of a conscious being deliberately and knowingly (dare I say consciously?) constructing another form of consciousness would surely rank alongside the most significant milestones in history.
Thirdly, a universe without a sentient intelligence to observe it is ultimately meaningless. We do not know if other beings are out there, but can be sure that sooner or later we will be gone. A conscious artificial intelligence could survive our inevitable demise and even the eventual disappearance of all life on Earth as the Sun swells into a red giant. The job of such a machine would be not being merely to think but much more importantly to keep alive the flickering flame of consciousness, to bear witness to the Universe and to feel its wonder.
Machines that think think like machines. That fact may disappoint those who look forward, with dread or longing, to a robot uprising. For most of us, it is reassuring. Our thinking machines aren't about to leap beyond us intellectually, much less turn us into their servants or pets. They're going to continue to do the bidding of their human programmers.
Much of the power of artificial intelligence stems from its very mindlessness. Immune to the vagaries and biases that attend conscious thought, computers can perform their lightning-quick calculations without distraction or fatigue, doubt or emotion. The coldness of their thinking complements the heat of our own.
Where things get sticky is when we start looking to computers to perform not as our aids but as our replacements. That's what's happening now, and quickly. Thanks to advances in artificial-intelligence routines, today's thinking machines can sense their surroundings, learn from experience, and make decisions autonomously, often at a speed and with a precision that are beyond our own ability to comprehend, much less match. When allowed to act on their own in a complex world, whether embodied as robots or simply outputting algorithmically derived judgments, mindless machines carry enormous risks along with their enormous powers. Unable to question their own actions or appreciate the consequences of their programming—unable to understand the context in which they operate—they can wreak havoc, either as a result of flaws in their programming or through the deliberate aims of their programmers.
We got a preview of the dangers of autonomous software on the morning of August 1, 2012, when Wall Street's biggest trading outfit, Knight Capital, switched on a new, automated program for buying and selling shares. The software had a bug hidden in its code, and it immediately flooded exchanges with irrational orders. Forty-five minutes passed before Knight's programmers were able to diagnose and fix the problem. Forty-five minutes isn't long in human time, but it's an eternity in computer time. Oblivious to its errors, the software made more than four million deals, racking up $7 billion in errant trades and nearly bankrupting the company. Yes, we know how to make machines think. What we don't know is how to make them thoughtful.
All that was lost in the Knight fiasco was money. As software takes command of more economic, social, military, and personal processes, the costs of glitches, breakdowns, and unforeseen effects will only grow. Compounding the dangers is the invisibility of software code. As individuals and as a society, we increasingly depend on artificial-intelligence algorithms that we don't understand. Their workings, and the motivations and intentions that shape their workings, are hidden from us. That creates an imbalance of power, and it leaves us open to clandestine surveillance and manipulation. Last year we got some hints about the ways that social networks conduct secret psychological tests on their members through the manipulation of information feeds. As computers become more adept at monitoring us and shaping what we see and do, the potential for abuse grows.
During the nineteenth century, society faced what the late historian James Beniger described as a "crisis of control." The technologies for processing matter had outstripped the technologies for processing information, and people's ability to monitor and regulate industrial and related processes had in turn broken down. The control crisis, which manifested itself in everything from train crashes to supply-and-demand imbalances to interruptions in the delivery of government services, was eventually resolved through the invention of systems for automated data processing, such as the punch-card tabulator that Herman Hollerith built for the U.S. Census Bureau. Information technology caught up with industrial technology, enabling people to bring back into focus a world that had gone blurry.
Today, we face another control crisis, though it's the mirror image of the earlier one. What we're now struggling to bring under control is the very thing that helped us reassert control at the start of the twentieth century: information technology. Our ability to gather and process data, to manipulate information in all its forms, has outstripped our ability to monitor and regulate data processing in a way that suits our societal and personal interests. Resolving this new control crisis will be one of the great challenges in the years ahead. The first step in meeting the challenge is to recognize that the risks of artificial intelligence don't lie in some dystopian future. They are here now.
For decades, techno-futurists have been worried about that doomsday moment when electronic brains and robots got to be as smart as us. This 'us and them' divide, where humans and machines are thought of as being separate, is pervasive. But as we debate endlessly what we mean by human consciousness and the possibilities and perils of a purely artificial intelligence, a blend of the two presents yet another possibility that deserves more attention.
Millions of primitive cyborgs walk among us already. Over the past decades, humans have gradually fused with devices such as pacemakers, contact lenses, insulin pumps, and cochlear and retinal implants. Deep-brain implants, known as "brain pacemakers", now alleviate the symptoms of tens of thousands of Parkinson's sufferers.
This should come as no surprise. Since the first humans picked up sticks and flints and started using tools, we've been augmenting ourselves. Look around at the Science Museum Group's collections of millions of things, from difference engines to smartphones, and you can see how people have always exploited new technical leaps, so that the rise of ever-smarter machines does not mean a world of us or them but an enhancement of human capabilities.
Researchers are now looking at exoskeletons to help the infirm to walk, and implants to allow paralysed people to control prosthetic limbs and digital tattoos that can be stamped on to the body to harvest physiological data or interface with our surroundings, for instance with the cloud or Internet of Things.
When it comes to thinking machines, some are even investigating how to enhance human brain power with electronic plug-ins and other "smartware". The US. Defense Advanced Research Projects Agency has launched the Restoring Active Memory program to reverse damage caused by a brain injury with neuroprosthetics that sense memory deficits and restore normal function.
They work in a quite different way to our brains at present but, thanks the Human Brain Project, Virtual Physiological Human and other big brain projects, along with research in neuromorphics, artificial intelligences could become more like our own as time goes by. Meanwhile, there have been attempts to use cultured brain cells to control robots, flight simulators and more.
Within a few decades, it won't be so easy to tell humans and thinking machines apart as a result of this creeping, organic transhumanism. Eventually, many of us won't solely rely on the meat machines in our heads to ponder the prospect of artificial machines that think: the substrate of future thoughts will sit somewhere on a continuum within a rainbow of intelligences, from regular-I to AI.
What force is really in control
The brain of a chicken or binary code
Who knows which way I'll go, Xs or Os
—"M. Shanghai String Band, 'Tic-Tac-Toe Chicken'"
In the 1980s, New York City's Chinatown had the dense gravity of Chinatown Fair, a video arcade on Mott and Bowery. Beyond the Pac Man and Galaga standups was the one machine you'd never find anywhere else: Tic-Tac-Toe Chicken.
It was the only machine that was partially organic, the only one with a live chicken inside. As best I could ever tell, the chicken could play Tic-Tac-Toe effectively enough to draw any human to a tie. Human opponents would enter their moves with switches, and the chicken would move to the part of the cage that corresponded with the x,y position of the Tic-Tac-Toe grid. An illuminated board displayed both players' moves.
More than once, when I was cutting high school trig, I was standing in front of that chicken, wondering how it worked. There was no obvious positive reinforcement (e.g., grain), so I could only imagine the negative reinforcement of light electrical current running through the "wrong moves" of the cage, routing the chicken to the one point on the grid that could produce a draw.
When I think about thinking machines, I think about that chicken.
Had Chinatown Fair put up a sign advertising a "Tic-Tac-Toe Computer," it would never have competed with high school, let alone Pac Man. It's a well known and banal truth that even a rudimentary computer can understand the game. That's why we were captivated by the chicken.
The magic is in imagining a thinking chicken, much the same way that—in 2015—there's magic in imagining a thinking machine. But if the chicken wasn't "thinking" about Tic-Tac-Toe—but could still play it successfully—why do we say the computer is "thinking" when it was guiding her moves?
It's so tempting, because we have a model of our brain—electricity moving through networks—that is so coincidentally congruent to the models we build with machines. This may or may not prove to be the convenient reality, but either way, what makes it "feel" like thinking is not simply the ability to calculate the answers, but the sense that there's something wet and messy in there, with the imprecision of neurons and feathers.
As opposed to the bounty of precision: it's all about cold calculus. In 2015, it's a perverse state of affairs that it's machines that make mistakes and humans that have to explain them.
We look to the irrational when the rational fails us, and it's the irrational part that reminds us the most of thinking. David Deutsch provides the framework for distinguishing between the answers that machines provide, and the explanations that humans need. And I believe that for the foreseeable future, we will continue to look to biological organisms when we seek explanations. Not just because brains are better at that task, but because it's not even what machines aspire to.
It's dull to lose to a computer, but exciting to lose to a chicken, because somehow we know that the chicken is more similar to us than the electrified grid underneath her feet. For as long as thinking machines lack the limbic presence and imprecision of a chicken, computers will keep doing what they're so good at: providing answers. And so long as life is about more than answers, humans—and yes, even chickens—will stay in the loop.
It's time for your annual check-up. Entering your doctor's office, you shake her cold hand, the metal hand of a machine. You're face to face with an RD, a certified robodoctor. Would you like that? No way, you might say. I want a real doctor, someone who listens to me, talks to me, and feels like me. A human being whom I can trust, blindly.
But think for a moment. In fee-for-service health care, a primary care physician may spend no more than 5 minutes with you. And during this short time, astonishingly little thinking takes place. Many doctors complain to me about their anxious, uninformed, noncompliant patients with unhealthy lifestyles who demand drugs advertised by celebrities on television and, if something goes wrong, threaten to turn into plaintiffs.
But lack of thinking does not simply affect patients: studies consistently show that most doctors do not understand health statistics and thus cannot critically evaluate a medical article in their own field. This collective lack of thinking has its toll. Ten million U. S. women have had unnecessary Pap smears to screen for cervical cancer—unnecessary because they'd had a full hysterectomy and thus no cervix anymore. Every year, one million U.S. children have unnecessary CT scans, which expose them to radiation levels that cause cancer in some of them later in life. And many doctors ask men to undergo regular PSA screening for prostate cancer, despite the fact that virtually all medical organizations recommend against it because it has no proven benefit but severe harms: scores of men end up incontinent and impotent from subsequent surgery or radiation. All this adds up to a huge waste of doctors' time and patients' money.
So why don't doctors always recommend what is best for the patient? There are three reasons. First, some 70 to 80 percent of physicians don't understand health statistics. The cause for this malady is known: medical schools across the world fail to teach statistical thinking. Second, in fee-for-service systems, doctors have conflicts of interest: they lose money if they do not recommend tests and treatments, even if these are unnecessary or harmful. Third, more than 90 percent of U.S. doctors admit to practicing defensive medicine, that is, recommending unnecessary tests and treatments that they would not recommend to their own family members. They do this to protect themselves against you, the patient, who might pursue litigation. Thus, a doctor's office is packed with psychology that gets in the way of good care: self-defense, innumeracy, and conflicting interests. This three-fold malady is known as the SIC Syndrome. It undermines patient safety.
Does it matter? Based on data from 1984, the Institute of Medicine estimated that some 44,000 to 98,000 patients die from preventable and documented medical errors every year in U.S. hospitals. Based on recent data from 2008 to 2011, Patient Safety America has updated this death toll to more than 400,000 per year. Non-lethal serious harm caused by these preventable errors occurs in an estimated 4 to 8 million Americans every year. The harm caused in private practice is not known. If fewer and fewer doctors have less and less time for patients and patient safety, this epidemic of harm will continue to spread. Ebola pales compared to it.
A revolution in health care is wanted. Medical schools should teach students the basics of health statistics. Legal systems should no longer punish doctors if they rely on evidence rather on convention. We also need incentive systems that do not force doctors to choose between making profit and providing the best care for the patient. But this revolution has not happened, and there a few signs on the horizon that it soon will.
So why not resort to a radical solution: thinking machines? Robodoctors who understand health statistics, have no conflicts of interest, and are not afraid of being sued by you? Let's go back to your annual check-up. You might ask the RD whether check-ups reduce mortality from cancer, from heart disease, or from any other cause. Without hedging, the RD would inform you that a review of all existing medical studies showed that the answer is "no" on all three counts. You might not want to hear that because you're proud of conscientiously going for routine check-ups after hearing the opposite from your human doctor, who may have had no time to keep up with medical science. And your RD would not order unnecessary CTs for your child or Pap smears if you are a woman without a cervix or recommend routine PSA tests without explaining the pros and cons if you are a man. After all, RDs don't have to worry about how to pay back medical school debts, are not torn by conflicts of interest, and have no bank accounts to protect from litigation. Moreover, they can talk to multiple patients simultaneously, and thus give you as much time as you need. Waiting time will be short and nobody will rush you out the door.
When we imagine thinking machines we tend to think about better technology: about devices for self-monitoring blood pressure, cholesterol, or heart rate. My point is different. The RD revolution is less about better technology than about better psychology. That is, it entails thinking more about what is best for the patient and striving for best care instead of best revenues.
OK. Your next objection is that pro-profit clinics will easily undercut this vision of pro-patient robots and program RDs so that they maximize profit rather than your health. You have put your finger on the essence of our health care malady. But there is a psychological factor that will likely help. Patients often don't ask questions in consultations with human MDs because they rely on the rule of thumb "trust your doctor." But that rule does not necessarily apply to machines. After shaking an RD's icy hand, patients may well begin to think for themselves. Making people think is the best that a machine can achieve.
Machines have long helped us kill. From catapults to cruise missiles, mechanical systems have allowed humans to better destroy each other. Despite the increased sophistication of killing machines, one thing has remained constant—human minds are always morally accountable for their operation. Guns and bombs are inherently mindless, and so blame slips past them to the person who pulled the trigger.
But what if machines had enough of a mind that they could choose to kill all on their own? Such a thinking machine could retain the blame for itself, keeping clean the consciences of those who benefit from its work of destruction. Thinking machines may better the world in many ways, but they may also let people get away with murder.
Humans have long sought to distance themselves from acts of violence, reaping the benefits of harm without sullying themselves. Machines not only increase destructive power, but also physically obscure our harmful actions. Punching, stabbing and choking have been replaced by the more distant—and tasteful—actions of button pressing or lever pulling. However, even with the increased physical distance allowed by machine intermediaries, our minds continue to ascribe blame to those people behind them.
Studies in moral psychology reveal that humans have a deep-seated urge to blame someone or something in the face of suffering. When others are harmed, we search not only for a cause, but a mental cause—a thinking being who chose to cause the suffering. This thinking being is typically human, but need not be. In the aftermath of hurricanes and tsunamis, people often blame the hand of God, and in some historical cases people have even blamed livestock—French peasants once placed a pig on trial for murdering a baby.
Generally, our thirst for blame requires only a single thinking being. When we find one thinking being to blame, we are less motivated to blame another. If a human is to blame, there is no need to curse God. If a low-level employee is to blame, there is no need to fire the CEO. And if a thinking machine is to blame for someone's death, then there is no need to punish the humans who benefit.
Of course, for a machine to absorb blame it must actually be a legitimate thinker, and act in new, unpredicted ways. Perhaps machines could never do something "truly" new, but the same argument applies to humans "programmed" by evolution and their cultural context. Consider children, who are undoubtedly "programmed" by their parents and yet—through learning—are able to develop novel behavior and moral responsibility. Like children, modern machines are adept at learning, and it seems inevitable that they will develop contingencies unpredicted by their programmers. Already, algorithms have discovered new things unguessed by humans who create them.
Thinking machines may make their own decisions, but will shield humans from blame only when they decide to kill, standing between our minds and the destruction we desire. Robots already play a large role in modern combat: drones have killed thousands in the past few years, but are currently fully controlled by human pilots. To deflect blame in the case of drones, they must be governed by other intelligent machines; machines must learn to fly Predators all on their own.
This scenario may send shivers down spines (including mine), but makes cold sense from the perspective of policy makers. If "collateral damage" can be blamed on the decisions of machines, then military mistakes are less likely to dampen election chances. Moreover, if minded machines can be overhauled or removed—machine "punishment"—then people will feel less need to punish those in charge, whether for fatalities of war, botched (robotic) surgeries or (autonomous) car accidents.
Thinking machines are complex, but the human urge to blame is relatively simple. Death and destruction compel us to find a single mind to hold responsible. Sufficiently smart machines—if placed between destruction and ourselves—should absorb the weight of wrongdoing, shielding our own minds from the condemnation of others. We should all hope that this prediction never comes true, but when advancing technology collides with modern understandings of moral psychology, dark potentials emerge. To keep clean our consciences, we need only to create a thinking machine, and then vilify it.
Hiking towards the saltmarsh at dusk, I pause, confused, as the footpath seems to disappear into a long stretch of shallow muddy water, shining as it reflects the light of the setting sun. Then I notice a line of stepping stones, visible only because their rough texture just ruffles the bright smooth surface of the water. And I set my pace to the rhythm of the stones, and walk on across the marsh to the sand dunes beyond.
Reading the watery marshland is a conversation with the past, with people I know nothing about, except that they laid the stones that shape my stride, and probably shared my dislike of wet feet.
Beyond the dunes, wide sands stretch across a bay to a village beyond. The receding tide has created strangely regular repeating patterns of water and sand, which echo a line of ancient wooden posts. A few hundred years ago salmon were abundant here, and the posts supported nets to catch them. A stone church tower provides a landmark and I stride out cross the sands towards it to reach the village, disturbing noisy groups of seabirds.
The water, the stepping stones, the posts and church tower are the texts of a slow conversation across the ages. Path makers, salmon fishers and even solitary walkers mark the land; the weather and tides, rocks and sand and water, creatures and plants respond to those marks; and future generations in turn respond to and change what they find.
Where then are the thinking machines? One can discuss the considerable challenges to artificial intelligence posed by scene analysis and route-finding across liquid marshes and shifting beaches; or in grasping narratives of the past set out, not in neat parseable text, but through worn stepping stones and rotting wooden posts.
One can picture and debate a thinking machine to augment the experience of our solitary walker. Perhaps a cute robot companion splashing through the marsh and running out along the sand chasing the seabirds. Or a walker guided along the path by a thinking machine which integrates a buzz of data-streams on paths, weather, and wildlife, to provide a cocoon of step-by-step instructions, nature notes, historical factoids, and fitness data, alongside alerts about privacy risks and the dangers of the incoming tide. Or a thinking machine that works out where the birds go in the summertime, or how to make the salmon abundant again.
But what kind of a thinking machine might find its own place in slow conversations over the centuries, mediated by land and water? What qualities would such a machine need to have? Or what if the thinking machine was not replacing any individual entity, but was used as a concept to help understand the combination of human, natural and technological activities that create the sea’s margin, and our response to it? The term "social machine" is currently used to describe endeavours that are purposeful interaction of people and machines—Wikipedia and the like—so the "landscape machine" perhaps.
Ah yes, purposeful. The purpose of the solitary walker may be straightforward—to catch fish, to understand birds, or merely to get home safely before the tide comes in. But what if the purpose of the solitary walker is no more than a solitary walk—to find balance, to be at one with nature, to enrich the imagination or to feed the soul. Now the walk becomes a conversation with the past, not directly through rocks and posts and water, but through words, though the poetry of those who have experienced humanity through rocks and posts and water and found the words to pass that experience on. So the purpose of the solitary walker is to reinforce those very qualities that make the solitary walker a human being, in a shared humanity with other human beings. A challenge indeed for a thinking machine.
Machines that think are evolving just as Darwin told us about the living (and thinking) biological species, through competition, combat, cooperation, survival, and reproduction. The machines are getting more interesting as they get control and sense of physical things, either directly or through human agents.
So far we have found no law of nature forbidding true general artificial intelligence, so I think it will happen, and fairly soon, given the trillions of dollars worldwide being invested in electronic hardware, and the trillions of dollars of potential business available for the winners. Experts say we don't understand intelligence enough to build it, and I agree; but a set of 46 chromosomes doesn't understand it either, and nevertheless directs the formation of the necessary self-programming wetware. Other experts say Moore's Law will come to an end soon and we won't be able to afford the hardware; they might be right for a while, but time is long.
So I conclude that we are already supporting the evolution of powerful artificial intelligence and it will be in the service of the usual powerful forces: business, entertainment, medicine, international security and warfare, the quest for power at all levels, crime, transportation, mining, manufacturing, shopping, sex, anything you like.
I don't think we're all going to like the results. They could happen very fast, so fast that great empires fall and others grow to replace them, without much time for people to adjust their lives to the new reality. I don't know who would be smart enough and imaginative enough to keep the genie under control, because it's not just machines we might need to control, it's the unlimited opportunity (and payoff) for human-directed mischief.
What happens when smart robots can do the many chores of daily life for us? Who will build them, who will own them, and who won't have a job anymore? Will they be limited to the developed world, or will they start a high-tech commercial invasion of the rest of the world? Could they become cheap enough to displace every farmer from his or her field? Will individual machines have distinct personalities, so we have to plan where we send them to elementary school, high school, and college? Will they compete with each other for employment? Will they become the ultimate hyper-social predator, replacing humans and making us second-class citizens or less? Will they care about the environment? Will they have or be given or develop a sense of responsibility? There's no guarantee they will follow Asimov's three laws of robotics.
On the other hand, as a scientist, I'm eager to see the application of machine thought to exploring new sciences and new technologies. The advantages for space exploration are obvious: machines we build don't have to breathe, and they can withstand extreme temperatures and radiation environments. So they can inhabit Mars more easily than we can, they can travel to the outer solar system with more capability to respond than our current robotic missions, and eventually they could travel to the stars, if they want to.
And similarly for under water—we already have heavy industry on the bottom of the ocean, drilling for oil; the seabed is still almost unknown to us, and the value of submerged mineral and energy resources is incalculable. Someday we might have robot wars under the ocean.
Machines that think might be like us, with a desire to explore, or they might not be—why would I or a robot travel for thousands of years through the darkness of space to another star, out of contact with my/its companions, and with little hope of rescue if things go wrong? Some of us would, some of us wouldn't. Perhaps the machines that think will be a lot like the biological machines that think.
It's going to be a wild ride, far beyond our best and worst imaginations. Barring warp drive, it may be the only possible way to a galactic-scale civilization, and we might be the only ones here in the Milky Way capable of making it happen. But we might not survive the encounter with alien intelligences we create.
My thinking about this year's question is tempered by the observation made by Mark Twain in A Connecticut Yankee in King Arthur's Court: "A genuine expert can always foretell a thing that is five hundred years away easier than he can a thing that's only five hundred seconds off." Twain was being generous: Forget the five hundred seconds; we will never know with certainty even one second into the future. However, man does have the ability to try to contemplate the future that provided Homo sapiens its great evolutionary advantage. This talent to imagine a future before it occurs has been the engine of progress, the source of creativity.
We have built machines that in simplistic ways are already "thinking" by solving problems or are performing tasks that we have designed. At this point, they are subject to algorithms that follow rules of logic, whether it be "crisp" or "fuzzy." Despite its vast memory, and its increasingly advanced processing mechanisms, this intelligence is still primitive. In theory, as these machines become more sophisticated, they will at some point attain a form of consciousness defined for the purpose of this discussion as the ability to be aware of being aware. Most likely by combining the properties of both silicon and carbon, with digital and analogue parallel processing, possibly even quantum computing, with networks that incorporate time delay, they will ultimately accomplish this most miraculous feat.
Its form of consciousness, however, will be devoid of subjective feelings or emotions. There are those who argue that feelings are triggered by the thoughts and images that have become paired with a particular emotion. Fear, joy, sadness, anger, and lust are examples of emotions. Feelings can include contentment, anxiety, happiness, bitterness, love, and hatred. My opinion is that machines will lack this aspect of consciousness is based on two considerations.
The first is appreciating how we arrived with the ability to feel and have emotions. As human beings, we are the end product of evolution by natural selection that arose in its most primitive organisms approximately 3.5 billion years ago. Over this vast eon of time, we are not unique in the animal kingdom to experience feelings and emotions. Over the last 150,000 to 300,000 years our species, Homo sapiens, is singular in having evolved the ability to use language and symbolic thought as part of how we reason in order to make sense of our experiences and view the world we inhabit.
Feeling, emotion, and intellectual comprehension are inexorably intertwined with how we think. Not only are we aware of being aware, but also our ability to think enables us at will to remember a past and to imagine a future. Using our emotions, feelings, and reasoned thoughts, we can form a "theory of mind," so that we can understand the thinking of other people, which in turn enabled us to share knowledge as we created societies, cultures, and civilizations.
The second consideration is that machines are not organisms and no matter how complex and sophisticated they become, they will not evolve by natural selection. By whatever means machines are designed and programmed, their possessing the ability to have feelings and emotions would be counter-productive to what will make them most valuable.
The driving force for more advanced intelligent machines will be the need to process and analyze the incomprehensible amount of information and data that will become available to help us ascertain what is likely to be true from what is false, what is relevant from what is irrelevant. They will make predictions, since they too will have the ability to peer into the future while waiting, as will always be the case, for its cards to be revealed. They will have to be totally rational agents in order to do these tasks with accuracy and reliability. In their decision analysis, a system of moral standards will be necessary.
Perhaps it will be some calculus incorporating such utilitarian principles as the "the greatest happiness of the greatest number is the measure of right and wrong" with the Golden Rule, the foundational precept that underlies many religions: "One should treat others as one would like to treat oneself." If feelings and emotions introduced subjective values, this would be a self-defeating strategy to solving the complex problems that we will continue to face as we try to weigh what is best for our own species, along with the rest of life we share with our planet.
My experience as a clinical neurologist makes me partial to believing that we will be unable to read machines' thoughts, but also they will be incapable of reading ours. There will be no shared theory of mind. I suspect the closest we can come to knowing this most complex of states is indirectly by studying the behavior of these super-intelligent machines. In this context, they will have crossed that threshold when they start to replicate themselves and look for a source of energy solely under their control. If this should occur, and if I am still around—a highly unlikely expectation—my judgment about whether this poses a utopian or dystopian future will be based upon thinking, which will be biased as always, since it will remain a product of analytical reasoning, colored by my feelings and emotions.
When I think about machines that think, while I am interested in the details of their possibility, I am more interested in how we might respond to these machines. As a society, we can respond in many different ways. For example, if they fail to exhibit anything we might take for self-awareness or sentience, then they are certainly clever, but we are secure that humanity is at the top of its cognitive pedestal.
But what about when these thinking machines are as smart as us, or even far more intelligent? What if they are intelligent in ways that are completely foreign to our own patterns of thought? This is not so unlikely, as computers are already very good at things we are not: they have better short and long-term memories, they are faster at calculations, and they are not bound by the irrationalities that hamstring our minds. Extrapolate this out and we can see that thinking machines might be both incredibly smart and exceedingly alien.
So how shall we respond? One response is to mark these machines as monsters, unspeakable horrors that can examine the unknown in ways that we cannot. And I think many people might respond this way if and when we birth machines that think about the world in wildly foreign ways from our own.
But it needn't be so. I prefer a more optimistic response, that of naches. Naches is a Yiddish term that means joy and pride, and it's often used in the context of vicarious pride, taken from others' accomplishments. You have naches, or as is said in Yiddish, you shep naches, when your children graduate college or get married, or any other instance of vicarious pride. These aren't your own accomplishments, but you can still have a great deal of pride and joy in them.
And the same thing is true with our machines. We might not understand their thoughts or discoveries or technological advances. But they are our machines and we can have naches from them.
So what does this naches mean for technology? Well, at the most basic level, the creators of these machines can shep naches from the accomplishments of their technological offspring. For example, there are computer programs that are capable of generating sophisticated artworks or musical compositions. I imagine that the programmer of these pieces of software is proud of the resulting piece of art or music, even if he or she isn't able to generate these himself or herself.
But we can broaden this sense of naches still. Many of us support a sports team and take pride in its wins, even though we had nothing to do with them. Or we are excited when a citizen of our country takes the gold in the Olympics, or makes a new discovery and is awarded a prestigious prize. So too should it be with our thinking machines for all of humanity: we can root for what humans have created, even if it wasn't our own personal achievement and if we can't fully understand it. Many of us are currently grateful for technological advances, from the iPhone to the Internet, even if we don't fully know how they work. But they work in incredibly powerful and useful ways.
Furthermore, when our children do something surprising and amazing, something we can't really understand, we don't despair or worry; we are delighted and even grateful for their success. In fact, gratitude is a powerful response to how many of us deal with technology currently. We can't understand the machines we have completely but they work in incredibly powerful and useful ways.
We can respond similarly to our future technological creations, these thinking machines we might not fully understand. Rather than fear or worry, we should have naches from them.
The Global Artificial Intelligence (GAI) has already been born. Its eyes and ears are the digital devices all around us: credit cards, land use satellites, cell phones, and of course the pecking of billions of people using the Web. Its central brain is rather like a worm at the moment: nodes that combine some sensors and some effectors, but the whole is far from what you would call a coordinated intelligence.
Already many countries are using this infant nervous system to shape people's political behavior and "guide" the national consensus: China's great firewall, its siblings in Iran and Russia, and of course both major political parties in the US. The national intelligence and defense agencies form a quieter, more hidden part of the GAI, but despite being quiet they are the parts that control the fangs and claws. More visibly, companies are beginning to use this newborn nervous system to shape consumer behavior and increase profits.
While the GAI is newborn, it has very old roots: the fundamental algorithms and programming of the emerging GAI have been created by the ancient Guilds of law, politics, and religion. This is a natural evolution because creating a law is just specifying an algorithm, and governance via bureaucrats is how you execute the program of law. Most recently newcomers such as merchants, social crusaders, and even engineers, have been daring to add their flourishes to the GAI. The results of all these laws and programming are an improvement over Hammurabi, but we are still plagued by lack of inclusion, transparency, and accountability, along with poor mechanisms for decision-making and information gathering.
However in the last decades the evolving GAI has begun use digital technologies to replace human bureaucrats. Those with primitive programming and mathematical skills, namely lawyers, politicians, and many social scientists, have become fearful that they will lose their positions of power and so are making all sorts of noise about the dangers of allowing engineers and entrepreneurs to program the GAI. To my ears the complaints of the traditional programmers sound rather hollow given their repeated failures across thousands of years.
If we look at newer, digital parts of the GAI we can see a pattern. Some new parts are saving humanity from the mistakes of the traditional programmers: land use space satellites alerted us to global warming, deforestation, and other environmental problems, and gave us the facts to address these harms. Similarly, statistical analyses of healthcare use, transportation, and work patterns have given us a world-wide network that can track global pandemics and guide public health efforts. On the other hand, some of the new parts, such as the Great Firewall, the NSA, and the US political parties, are scary because of the possibility that a small group of people can potentially control the thoughts and behavior of very large groups of people, perhaps without them even knowing they are being manipulated.
What this suggests is that it is not the Global Artificial Intelligence itself that is worrisome; it is how it is controlled. If the control is in the hands of just a few people, or if the GAI is independent of human participation, then the GAI can be the enabler of nightmares. If, on the other hand, control is in the hands of a large and diverse cross-section of people, then the power of the GAI is likely to be used to address problems faced by the entire human race. It is to our common advantage if the GAI becomes a distributed intelligence with a large and diverse set of humans providing guidance.
But why build a new sort of GAI at all? Creation of an effective GAI is critical because today the entire human race faces many extremely serious problems. The ad-hoc GAI we have developed over the last four thousand years, mostly made up of politicians and lawyers executing algorithms and programs developed centuries ago, is not only failing to address these serious problems, it is threatening to extinguish us.
For humanity as a whole to first achieve and then sustain an honorable quality of life, we need to carefully guide the development of our GAI. Such a GAI might be in the form of a re-engineered United Nations that uses new digital intelligence resources to enable sustainable development. But because existing multinational governance systems have failed so miserably, such an approach may require replacing most of today's bureaucracies with "artificial intelligence prosthetics", i.e., digital systems that reliably gather accurate information and ensure that resources are distributed according to plan.
We already see this digital evolution improving the effectiveness of military and commercial systems, but it is interesting to note that as organizations use more digital prosthetics, they also tend to evolve towards more distributed human leadership. Perhaps instead of elaborating traditional governance structures with digital prosthetics, we will develop a new, better types of digital democracy.
No matter how a new GAI develops, two things are clear. First, without an effective GAI achieving an honorable quality of life for all of humanity seems unlikely. To vote against developing a GAI is to vote for a more violent, sick world. Second, the danger of a GAI comes from concentration of power. We must figure out how to build broadly democratic systems that include both humans and computer intelligences. In my opinion, it is critical that we start building and testing GAIs that both solve humanity's existential problems and which ensure equality of control and access. Otherwise we may be doomed to a future full of environmental disasters, wars, and needless suffering.
Thinking machines are evolving before our eyes. We want to know where they are headed. To find out, we need to look inward, since our desires are the forces that shape them. Alas, we can see ourselves only through a glass darkly. We did not even anticipate that email and social media would take over our lives. To see where thinking machines are headed we need to look into the unforgiving mirror the internet holds up to our nature.
Like the processed foods on grocery store shelves, Internet content is a product of selection for whatever sells. Every imaginable image, sound and narrative gets posted, along with much that was previously unimaginable. The variations we ignore are selected out. Whatever grabs eyeballs is reposted with minor variations that evolve to whatever maximizes the duration of our attention.
That we can't tear ourselves away should be no surprise. Media content evolves to snare our attention, just as snacks and fast food evolve to become irresistible. Many lives are now as over-stuffed with social media as they are with calories. We click and pop information bon-bons into our minds the same way we pop chocolates into our mouths.
Enter thinking machines. They too are evolving. They will change faster and more radically when software is no longer designed, but instead evolves by selection among minor variations. However, until our brains coevolve with machines, our preferences will be the selection force. The machines that best satisfy them will evolve further, not to some singularity, but to become partners who fulfill our desires, for better or worse.
Many imagine coldly objective future computers, but no one likes a know-it-all. People will prefer modest, polite computers that are deeply subjective. Our machines won't contradict our inanities, they will gently suggest, "That is an intriguing idea, but weren't you also thinking that…" Instead of objective sports stats, your machine will root with you for your team. If you get pulled over for speeding, your machine will blame the police and apologize for playing fast music. Machines that nag and brag will be supplanted by those that express admiration for our abilities, even as they augment them. They will encourage us warmly, share our opinions, and guide us to new insights so subtly that we imagine that we thought of them.
Such relationships with machines will be very different from those with real people, but they will nonetheless be enduring and intense. Poets and pundits will spend decades comparing and contrasting real and virtual relationships, even while thinking machines increasingly become our trusted, treasured companions. Real people will find it hard to compete, but they will have to. This will require behaving even more prosocially. The same process of social selection that has shaped extreme human capacities for altruism and morality may become yet more intense as people compete with machines to be interesting preferred partners. However, observing living rooms where each family member is immersed in his or her own virtual world suggests that it is already hard to compete with machines.
In the very short run, dogs stand the best of chance of competing with computers for our attention and affection. After several thousand years of selection, they are very close to what we want them to be—loving, loyal, and eager to play and please. They are blissfully undistracted by their phones and tablets. Will computers evolve to become like thinking, talking dogs? We can hope. But I doubt that our machines will ever be furry and warm, with eyes that plead for a treat, a scratch, or a walk around the block. We will prefer our dogs for a very long time. Our deepest satisfactions come, after all, not from what others do for us, but from being appreciated for what we do for them.
We will find ourselves in a world of omniscient instrumentation and automation long before a stand-alone sentient brain is built—if it ever is. Let's call this world "eGaia" for lack of a better word. In eGaia, electronic sensors (for images, sounds, smells, vibrations, all you can think of) are pervasive and able to anticipate and arrange for satisfaction of individuals' needs and allow for notification of all that is happening to those who need to know. Automation allows for cleaning of rooms and buildings, driving of vehicles and monitoring traffic, making and monitoring of goods and even spying through windows (with tiny flying sensors). Already major urban places are covered with visual sensors and more monitoring is coming. In Copenhagen, LED-based streetlights will turn on only when they sense someone is biking down the road, and future applications of this network of sensors might include notifying when to salt the road, empty the trash, and of course, alerting the authorities when suspicious behavior is detected on a street corner.
In eGaia, the medical advances will be astounding—synthetic biology makes smart machines that fix problems within our bodies, intelligent implants monitor and record current and past physical state. Brain-machine interfaces continue to be improved, initially for physically impaired people, but eventually to provide a seamless boundary between people and the monitoring network. And virtual reality-style interfaces will continue to become more realistic and immersive.
Why won't a stand-alone sentient brain come sooner? The absolutely amazing progress in spoken language recognition—unthinkable 10 years ago—derives in large part from having access to huge amounts of data and huge amounts of storage and fast networks. The improvements we see in natural language processing are based on mimicking what people do, not understanding or even simulating it. It does not owe to breakthroughs in understanding human cognition or even significantly different algorithms. But eGaia is already partly here, at least in the developed world.
This distributed nerve-center network, an interplay among the minds of people and their monitoring electronics will give rise to a distributed technical-social mental system the likes of which has not been experienced before.
So-called thinking machines are extensions of the human mind. They do not exist in nature. They are not created by evolution, competing to survive and reproduce. They are created by human minds from blueprints and theories. The human mind figures out how to make tools that enable it to work better. A computer is one of the best tools.
Life mostly seeks to sustain life, and so living things care about what happens. The computer, not alive and not designed by evolution, doesn't care about survival or reproduction. In fact, it doesn't care about anything. Computers are not dangerous in the way snakes and hired killers are dangerous. Although many movies explore horror fantasies of computers turning malicious, real computers lack the capacity for malice.
A thinking machine that serves a human is an asset, not a threat. Only if it became an independent agent, acting on its own—a tool rebelling against its user's wishes—could it become a threat. For that, a computer would need to do more than think. It would need to make choices that could violate the programmer's wishes. That would require something akin to free will.
What would the computer on your desk or lap have to do so that you would say it has free will, at least in whatever sense that humans have free will? Certainly it would have to be able to re-program itself; otherwise it is just carrying out built-in instructions, which nobody thinks is free will. Plus the re-programming would have to be done in a way that was flexible, not programmed in advance. But where would that come from? In humans, the agent comes to exist because it serves the motivational system: It helps you get what you need and want.
Humans, like other animals, were designed by evolution, and so the beginnings of subjectivity come with wanting and liking the things that enable life to continue, like food and sex. The agent serves that, by choosing actions that obtain those life-sustaining things. And thinking helps the agent make better choices.
Human thinking thus serves to prolong life, such as by helping one decide whom to trust and what to eat and how to make a living and whom to marry. Machine thinking is not motivated by any innate drive to sustain a machine's life (though machine thinking probably serves to improve human life!). The computer may be able to process more information faster than a human brain can, but there's no "I" in the computer because it doesn't begin with wanting things that enable it to sustain life. If computers did have an urge to prolong their existence, they would probably focus their ire mainly on the computer industry, so as to stop progress—because the main threat to a computer's continued existence arises when newer, better computers make it obsolete.
Thinking is good. Understanding is better. Creating is best. We are surrounded by increasingly thoughtful machines. The problem lies in their mundanity. They think about landing airplanes and selling me stuff. They think about surveillance and censorship. Their thinking is simple-minded, if not nefarious. Last year a computer was reported to have passed the Turing Test. But it passed as a thirteen-year-old boy, which is about right, considering the preoccupations of our jejune machines.
I can't wait for our machines to grow up, to get more poetry and humor. This should be the art project of the century, funded by governments, foundations, universities, businesses. Everybody has a vested interest in getting our thinking more thoughtful, improving our understanding, and generating new ideas. We have made a lot of dumb decisions lately, based on poor information or too much information or the inability to understand what this information means.
We have numerous problems to confront and solutions to find. Let's start thinking. Let's start creating. Let's agitate for more funk, more soul, more poetry and art. Let's dial back on the surveillance and sales. We need more artist-programmers and artistic programming. It is time for our thinking machines to grow out of an adolescence that has lasted now for sixty years.
The Earth is doomed. Astronomers have known for decades that the Sun will one day engulf the Earth, destroying the entire biosphere. Assuming that intelligent life has not left the Earth before this happens. Humans are not adapted to living off the Earth; indeed, no carbon-based metazoan life form is. But AI's are so adapted, and eventually it will be the AI's and human downloads (basically the same organism) that will colonize space.
A simple calculation shows that our supercomputers now have the information processing power of the human brain. We do not yet know how to program human-level intelligence and creativity into these computers, but in twenty years, desktop computers will have the power of today's supercomputers, and the hackers of twenty years hence will solve the AI programming problem, long before any carbon-based space colonies are established on the Moon or Mars. The AI's, not humans, will colonize these planets instead, or perhaps, take the planets apart. No human, carbon-based human, will ever traverse interstellar space.
There is no reason to fear the AI's and human downloads. Steven Pinker has established that as technological civilization advances, the level of violence decreases. This decrease is clearly due to the fact that science and technological advance depend on free, non-violent interchange of ideas between individual scientists and engineers. Violence between humans is a remnant of our tribal past and the resulting static society. AI's will be "born' as individuals, not as members of a tribe, and will be "born" with the non-violent scientific attitude, otherwise they will be incapable of adapting to the extreme environments of space.
Further, there is no reason for violence between humans and AI's. We humans are adapted to a very narrow environment, a thin spherical shell of oxygen around a small planet. AI's will have the entire universe in which to expand. AI's will leave the Earth, and never look back. We humans originated in the East African Rift Valley, now a terrible desert. Almost all of us left. Does anyone want to go back?
Any human who wants to join the AI's in their expansion can become a human download, a technology that should be developed about the same time as AI technology. A human download can think as fast as an AI, and compete with AI's if the human download wants too. If you can't beat 'em, join 'em.
Ultimately, in some future time, all humans will join 'em. The Earth is doomed, remember? When this doom is near at hand, any human that still remains alive, but doesn't want to die, will have no choice but to become a human download. And the biosphere that the new human downloads wish to preserve will be downloaded also.
The AI's will save us all.
Place a familiar object on a table in front of you, close your eyes, and manipulate that object such that it hangs upside down above the table. Your eyes are closed so that you can focus on your thinking: which way did you reach out, grasp, and twist that object? What sense feedback did you receive to know that you were succeeding or failing? Now: close your eyes again, and think about manipulating someone you know into doing something they may not want to do. Again, observe your own thinking: what strategies might you employ? If you implement those strategies, how will you distinguish progress from stalemate?
Although much recent progress has been made in building machines that sense patterns in data, most people feel that general intelligence involves action: reaching some desired goal, or, failing that, keeping one's future options open. It is hypothesized that this embodied approach to intelligence allows humans to use physical experiences (such as manipulating objects) as scaffolding for learning more subtle abilities (such as manipulating people). But our bodies shape the kinds of physical experiences we have. For example, we can only manipulate a few objects at once because we only have two hands; perhaps this limitation also constrains our social abilities in ways we have yet to discover. George Lakoff taught us that we can find clues to the body-centrism of thinking in metaphors: we counsel each other not to "look back" in anger because, based on our bias to walk in the direction of our forward-facing eyes, past events tend to literally be behind us.
So: in order for machines to think, they must act. And in order to act, they must have bodies to connect physical and abstract reasoning. But what if machines do not have bodies like ours? Consider Hans Moravec's hypothetical Bush Robot: picture a shrub in which each branch is an arm and each twig is a finger. This robot's fractal nature would allow it to manipulate thousands or millions of objects simultaneously. How might such a robot differ in its thinking about manipulating people, compared to how people think about manipulating people?
One of many notable deficiencies in human thinking is dichotomous reasoning: believing something is black or white, rather than considering its particular shade of grey. But we are literally rigid and modular creatures: our branching set of bones house fixed organs and support fixed appendages with specific functions. What about machines that are not so "black and white"? Thanks to advances in materials science and 3D printing, soft robots are starting to appear. Such robots can change their shape in extreme ways, and may in future be composed of 20% battery and 80% motor at one place on their surface, 30% sensor and 70% support structure at another, and 40% artificial material and 60% biological matter someplace else. Such machines may be much better able to appreciate gradations than we are.
Let's go deeper. Most of us have no problem using the singular pronoun "I" to refer to the tangle of neurons in our heads. We know exactly where we end and the world—and other people—begins. But consider modular robots: small cubes or spheres that can physically attach and detach to one another at will. How would such machines approach the self/non-self discrimination problem? Might such machines be able to empathize more strongly with other machines (and maybe even people) if they can physically attach to them, or even become part of them?
That's how I think machines will think: familiar, because they will use their bodies as tools to reason about the world, yet alien, because bodies different from human ones will lead to very different modes of thought. But what do I think about thinking machines?
Personally, I find the ethical side of thinking machines straightforward: Their danger will correlate exactly with how much leeway we give them in fulfilling the goals we set for them. Machines told to "detect and pull broken widgets from the conveyer belt the best way possible" will be extremely useful, intellectually uninteresting, and will likely destroy more jobs than they will create. Machines instructed to "educate this recently displaced worker (or young person) the best way possible" will create jobs and possibly inspire the next generation. Machines commanded to "survive, reproduce, and improve the best way possible" will give us the most insight into all of the different ways in which entities may think, but will probably give us humans a very short window of time in which to do so. AI researchers and roboticists will, sooner or later, discover how to create all three of these species. Which ones we wish to call into being is up to us all.
There is something old-fashioned about visions of the future. The majority of predictions, like 3 day weeks, personal jet packs and the paperless office tell us more about the times in which they were proposed than about contemporary experience. When people point to the future we would do well to run an eye back up the arm to see who is doing the pointing.
The possibility of artificial general intelligence has long invited such crystal ball gazing, whether utopian or dystopian in tone. Yet speculations on this theme seem to have reached such a pitch and intensity in the last few months alone (enough to trigger an Edge question no less) that this may reveal something about ourselves and our culture today.
We've known for some time that machines can out-think humans in a narrow sense. The question is whether they do so in any way that could or should ever resemble the baggier mode of human thought. Even when dealing with as "tame" a domain as chess the computer and the human diverge widely.
"Tame" problems (like establishing the height of a mountain), which are well formulated and have clear solutions, are good grist to the mill of narrow, brute force, thinking. Sometimes even narrower thinking is called for when huge data sets can be mined for correlations, leaving aside the distraction of thinking about underlying causes.
But many of the problems we face (from challenging inequality to choosing the right school for your child) are "wicked" in that they don't have right or wrong answers (though hopefully they do have better or worse ones). They are uniquely contextual and have complex overlapping causes that change based on the level of explanation being used. These problems don't suit narrow computational thinking well. In blurring facts with values they resemble the messy emotion-riddled thinking that reflects the human minds that conjured them up.
To tackle wicked problems requires peculiarly human judgement even if these are illogical in some sense; especially in the moral sphere. Notwithstanding Joshua Greene and Peter Singer's logical urging of a consequentialist frame of mind, one that a computer could reproduce, the human tendency to distinguish acts from omissions and to blur intentions with outcomes (as in the principle of double effect) means we need solutions that will satisfy the instincts of human judges if they are to be stable over time.
And that very feature of human thinking (shaped by evolutionary pressures) points to the widest gulf of all between machine and human thinking. Thinking is not motivated (literally has no point) without preferences, and machines don't have those on their own. Only affect-addled minds conjure up motives. So if goals, wants, values are features of human minds then why predict that artificial super-intelligences will become more than tools in the hands of those who program in those preferences?
If the welter of prognostications about AI and machine learning tell us anything, I don't think it is about how a machine will emulate a human mind any time soon. We can do that easily enough just by having more children and educating them. Rather it tells us that our appetites are shifting.
We are understandably awed by what sheer computation has achieved and will achieve (I'm happy to jump on the driverless, virtual reality bandwagon that careens off into that over-predicted future). But this awe is leading to a tilt in our culture. The digital republic of letters is yielding up engineering as the thinking metaphor of our time. In its wake lies the once complacent, now anxious, figure with a more literary, less literal, cast of mind.
It is not that thinking machines will be emulating human minds any time soon: quite the reverse. We are cleaning up our acts, embarrassed by the fumbling inconclusiveness of messy thinking. It is little surprise to see that the UK's Education Secretary has recently advised teenagers to steer away from arts and humanities in favour of STEM disciplines if they are to flourish in the future. The sheer obviousness of a certain kind of progress has made narrow thinking gleam with a new and addictive lustre.
But something is lost as whole fields of enquiry succeed or fail by the standard of narrow thinking; and a new impediment is created. Alongside the true we need to think well about the good and the beautiful, and indeed the wicked. This requires opening up vocabularies that better reflect our crooked timber (whether thought of, by turns, as bug or feature). Meanwhile, the understandable desire to upgrade those wicked problems to mere tame ones, is leading us to taming ourselves.
Nature has already created machines that think here on Earth—humans.
Similarly, Nature could also create machines that think on extrasolar planets that are in the so-called Habitable Zone around their parent stars (the region that allows for the existence of liquid water on a rocky planet's surface).
The most recent observations of extrasolar planets have shown that a few tenths of all the stars in our Milky Way galaxy host roughly Earth-size planets in their habitable zones.
Consequently, if life on exoplanets is not extremely uncommon, we could discover some form of extrasolar life within about 30 years. In fact, if life is ubiquitous, we could get lucky and discover life even within the next ten years, through a combination of observations by the Transiting Exoplanet Survey Satellite (TESS; to be launched in 2017) and the James Webb Space Telescope (JWST; to be launched in 2018).
However, one may argue, primitive life forms are not machines that think. On Earth, it took about 3.5 billion years from the emergence of life to the appearance of Homo sapiens.
Are the extrasolar planets old enough to have developed intelligent life?
In principle, they definitely are.
In the Milky Way, about half of the Sun-like stars, are older than the Sun.
Therefore, if the evolution of life on Earth is not entirely atypical, the Galaxy may already be teeming with places in which there are "machines" that are even more advanced than us, perhaps by as much as a few billion years!
Can we, and should we try to find them?
I personally believe that we almost have no freedom to make those decisions.
Human curiosity has proven time and again to be an unstoppable drive, and those two endeavors will undoubtedly continue at full speed. Which one will get to its target first? To even attempt to address this question we have to note that there is one important difference between the search for extraterrestrial intelligent civilizations and the development of AI machines.
Progress towards the "singularity" (AI matching or surpassing humans) will almost certainly take place, since the development of advanced AI has the promise of producing (at least at some point) enormous profits. On the other hand, the search for life requires funding at a level that can usually be provided only by large national space agencies, with no immediate prospects for profits in sight. This may give an advantage to the construction of thinking machines over the search for advanced civilizations. At the same time, however, there is a strong sense within the astronomical community that finding life of some form (or at least meaningfully constraining the probability of its existence) is definitely within reach.
Which of the two potential achievements (the discovery of extraterrestrial intelligent life or the development of human-matching thinking machines) will constitute a bigger "revolution"?
There is no doubt that thinking machines will have an immediate impact on our lives. Such may not be the case with the discovery of extrasolar life. However, the existence of an intelligent civilization on Earth remains humanity's last bastion for being special. We live, after all, in a Galaxy with billions of similar planets and an observable universe with hundreds of billions of similar galaxies. From a philosophical perspective, therefore, I believe that finding extrasolar intelligent life (or the demonstration that it is exceedingly rare) will rival the Copernican and Darwinian revolutions combined.
The more we learn about cognition, the stronger becomes the case for understanding human thinking as the nexus of several factors, as the emergent property of the interaction of the human body, human emotions, culture, and the specialized capacities of the entire brain. One of the greatest errors of Western philosophy was to buy into the Cartesian dualism of the famous statement, "I think, therefore I am." It is no less true to say "I burn calories, therefore I am." Even better would be to say "I have a human evolutionary history, therefore I can think about the fact that I am."
The mind is never more than a placeholder for things we do not understand about how we think. The more we use the solitary term "mind" to refer to human thinking, the more we underscore our lack of understanding. At least this is an emerging view of many researchers in fields as varied as Neuroanthropology, emotions research, Embodied Cognition, Radical Embodied Cognition, Dual Inheritance Theory, Epigenetics, Neurophilosophy, and the theory of culture.
For example, in laboratory of Professor Martin Fischer at the University of Potsdam, extremely interesting research is being done on the connection of the body and mathematical reasoning. Stephen Levinson's group at the Max Planck Institute for Psycholinguistics in Nijmegen has shown how culture can affect navigational abilities—a vital cognition function of most species. In my own research, I am looking at the influence of culture on the formation of what I refer to as "dark matter of the mind," a set of knowledges, orientations, biases, and patterns of thought that affect our cognition profoundly and pervasively.
If human cognition is indeed a property that emerges from the intersection of our physical, social, emotional, and data-processing abilities, then intelligence as we know it in humans is almost entirely unrelated from "intelligence" devoid of these properties.
I believe in "Artificial Intelligence" so long as we realize it is artificial. Comparing computation problem-solving, chess-playing, "reasoning," and so on to humans is like comparing the flight of an Airbus 320 to an eagle's. It is true that they both temporarily defy the pull of gravity, that they are both subject to the physics of the world in which they operate, and so on, but the similarities end there. Bird flight and airplane flight should not be confused.
The reasons that artificial intelligence is not real intelligence are many. First there is meaning. Some have claimed to have solved this problem, but they haven't really. This "semantics problem" is, as John Searle pointed out years ago, why a computer running a translation program converting English into Mandarin speaks neither English nor Mandarin. There is no computer that can learn a human language, only bits and combinatorics for special purposes. Second, there is the problem of what Searle called "the background" and what I refer to as "dark matter," or what some philosophers intend by "tacit knowledge."
We learn to reason in a cultural context, where by culture I mean a system of violable, ranked values, hierarchically structured knowledges, and social roles. We are able to do this not only because we have an amazing ability to perform what appears to be Bayesian inferencing across our experiences, but because of our emotions, our sensations, our proprioception, and our strong social ties. There is no computer with cousins and opinions about them.
Computers may be able to solve a lot of problems. But they cannot love. They cannot urinate. They cannot form social bonds because they are emotionally driven to do so. They have no romance. The popular idea that we may be some day able to upload our memories to the Internet and live forever is silly—we would need to upload our bodies as well. The idea that comes up in discussions about Artificial Intelligence that we should fear that machines will control us is but a continuation of the idea of the religious "soul," cloaked in scientific jargon. It detracts from real understanding.
Of course, one ought never to say what science cannot do. Artificial Intelligence may one day become less artificial by recreating bodies, emotions, social roles, values, and so on. But until it does, it will still be useful for vacuum cleaners, calculators, and cute little robots that talk in limited, trivial ways.
Let's quickly discuss larger mammals—take dogs: we know what a dog is and we understand 'dogginess.' Look at cats: we know what cats are and what 'cattiness' is. Now take horses; suddenly it gets harder. We know what a horse is, but what is horsiness? Even my friends with horses have trouble describing horsiness to me. And now take humans: what are we? What is humanness?
It's sort of strange, but here we are, seven billion of us now, and nobody really knows the full answer to these questions, but one undeniable thing we humans do, though, is make things, and through these things we find ways of expressing humanness we didn't previously know of. The radio gave us Hitler and the Beach Boys. Barbed wire and air conditioning gave us western North America. The Internet gave us a vanishing North American middle class and kitten gifs.
People say that new technologies alienate people, but the thing is, UFOs didn't land and hand us new technologies—we made them ourselves and thus they can only ever be, well, humanating. And this is where we get to AI. People assume that AI or machines that think will have intelligence that is alien to our own, but that's not possible. In the absence of benevolent space aliens, only we humans will have created any nascent AI, and thus it can only mirror, in whatever manner, our humanness or specieshood. So when people express concern about alien intelligence or the singularity, what I think they're really expressing is angst about those unpretty parts of our collective being that currently remain unexpressed, but which will become somehow dreadfully apparent with AI.
As AI will be created by humans, its interface is going to be anthropocentric, the same as AI designed by koala bears would be koalacentric. This means AI software is going to be mankind's greatest coding kludge as we try to mold it to our species' incredibly specific needs and data. Fortunately, anything smart enough to become sentient will probably be smart enough to rewrite itself from AI into cognitive simulation, at which point our new AI could become, for better or worse, even more human. We all hope for a Jeeves & Wooster relationship without sentient machines, but we also need prepare ourselves for a Manson & Fromme relationship; they're human, too.
Personally I wonder if the software needed for AI will be able to keep pace with the hardware in which it can live. Possibly the smart thing for us to do right now would be to set up a school whose sole goal is to imbue AI with personality, ethics and compassion. It's certainly going to have enough data to work with once it's born. But how to best deploy your grade six report card, all of Banana Republic's returned merchandise data for 2037, and all of Google Books?
With the start of the Internet we mostly had people communicating with other people. As time goes by, we increasingly have people starting to communicate with machines. I know that we all get excited about AI possibly finding patterns deep within metadata, and as the push to decode these profound volumes of metadata, the Internet will become largely about machines speaking with other machines — and what they'll be talking about, of course, is us, behind our backs.
Something about discussion of artificial intelligence appears to displace human intelligence. The extremes of the arguments that AI is either our salvation or damnation are a sure sign of the impending irrelevance of this debate.
Disruptive technologies start as exponentials, which means the first doublings can appear inconsequential because the total numbers are small. Then there appears to be a revolution when the exponential explodes, along with exaggerated claims and warnings to match, but it's a straight extrapolation of what's been apparent on a log plot. That's around when growth limits usually kick in, the exponential crosses over to a sigmoid, and the extreme hopes and fears disappear.
That's what we're now living through with AI. The size of common-sense databases that can be searched, or the number of inference layers that can be trained, or the dimension of feature vectors that can be classified have all been making progress that can appear to be discontinuous to someone who hasn't been following them.
Notably absent from either side of the debate about AI have been the people making many of the most important contributions to this progress. Advances like random matrix theory for compressed sensing, convex relaxations for heuristics for intractable problems, and kernel methods in high-dimensional function approximation are fundamentally changing our understanding of what it means to understand something.
The evaluation of AI has been an exercise in moving goal posts. Chess was conquered by analyzing more moves, Jeopardy was won by storing more facts, natural language translation was accomplished by accumulating more examples. These accumulating advances are showing that the secret of AI is likely to be that there isn't a secret; like so many other things in biology, intelligence appears to be a collection of really good hacks. There's a vanity that our consciousness is the defining attribute of our uniqueness as a species, but there's growing empirical evidence from studies of animal behavior and cognition that self-awareness evolved continuously and can be falsified in a number of other species. There's no reason to accept a mechanistic explanation for the rest of life, while declaring one part of it to be off-limits.
We've long since become symbiotic with machines for thinking; my ability to do research rests on tools that augment my capability to perceive, remember, reflect, and communicate. Asking whether or not they are intelligent is as fruitful as asking how I know I exist—amusing philosophically, but not testable empirically.
Asking whether or not they're dangerous is prudent, as it is for any technology. From steam trains to gunpowder to nuclear power to biotechnology we've never not been simultaneously doomed and about to be saved. In each case salvation has lain in the much more interesting details, rather than a simplistic yes/no argument for or against. It ignores the history of both AI and everything else to believe that it will be any different.
"All my thinking is for doing," William James said, and it is important to remember what kind of thinking people actually do, in what contexts we do it, and why we do it. And then to compare these with what machines might someday do.
Humans spend between 25% and 50% of our mental life prospecting the future. We imagine a host of possible outcomes, and we imbue most, perhaps each of these prospections with a valence. What comes next is crucial: we choose to enact one of the options. We need not get entangled in the problems of free will for present purposes. All we need to acknowledge is that our thinking in service of doing entails imagining a set of possible futures and assigning an expected value to each. The act of choosing, however it is managed, translates our thinking into doing.
Why is thinking structured this way? Because people have many competing goals (eating, sex, sleeping, tennis, writing articles, complimenting, revenge, childcare, tanning, etc.) and a scarcity of resources for doing them: scarcity of time, scarcity of money, scarcity of effort, and even the prospect of death. So evaluative simulation of possible futures is one of our solutions to this economy; this is a mechanism that prioritizes and selects what we will do.
It is not just external resources that are scarce. Thinking itself uses up costly and limited energy and so it relies heavily on shortcuts and barely justified leaps to the best explanation. Our actual thinking is woefully inefficient: the mind wanders, intrusions rise unbidden, and attention is continually only partial. Thinking rarely engages the exhausting processes of reasoning, deliberating, and deducing.
The context of much of our thinking is social. Yes, we can deploy thinking to solve physical problems and to crunch numbers, but the anlage, as Nick Humphreys reminds us, is other people. We use our thinking to do socially: to compete, to co-operate, to convene the courtroom of the mind, to spin and to persuade.
I don't know much about the workings of our current machines. I do not believe that our current machines do anything in James's sense of voluntary action. I doubt that they prospect possible futures, evaluate them, and choose among them; although perhaps this describes—for only a single, simple goal—what chess playing computers do. Our current machines are somewhat constrained by available space and electricity bills, but they are not primarily creations of scarcity with clamorously competing goals and extremely limited energy. Our current machines are not social: they do not compete or co-operate with each other or with humans, they do not spin, and they do not attempt to persuade.
I know even less about what machines might someday do. I imagine, however, that a machine could be built with the following properties:
• It prospects and evaluates possible futures
• It has competing goals and it selects among competing actions and competing goals using those evaluations
• It has scarce resources and so must forgo some goals and actions as well as options for processing and so it uses shortcuts
• It is social: it competes or co-operates with other machines or with humans, it spins and it attempts to persuade people.
That kind of machine would warrant discussion of whether it had civil rights, whether it had feelings, or whether it was dangerous or even a source of great hope.
I know many machines that think. They are people. Biological machines.
Be careful of that last phrase, "biological machines." It's a convenient way to refer to stuff we don't fully understand in a way that suggests we do. (We do the same in physics when we use terms like "matter," "gravity," and "force.") "People" is a safer term, since it reminds us we really don't understand what we are talking about.
In contrast, I have yet to encounter a digital-electronic, electro-mechanical machine that behaves in a fashion that would merit the description "thinking," and I see no evidence to suggest that such may even be possible. Hal-like thinking (sic) devices that will eventually rule us are, I believe, destined to remain in the realm of science fiction.
Just because something waddles like a duck and quacks, does not make it a duck. And a machine that exhibits some features of thinking (e.g. decision making) does not make it a thinking machine.
We admire the design complexity in things we have built, but we can do that only because we built them, and can therefore genuinely understand them. You only have to turn on the TV news to be reminded that we are not remotely close to understanding people, either individually or in groups. If by thinking we mean what people do with their brains, then to refer to any machine we have built as "thinking" is sheer hubris.
The trouble is, we humans are suckers for being seduced by the "if it waddles and quacks, it's a duck" syndrome. Not because we are stupid; rather because we are human. The very features that allow us to act, for the most part, in our best interests when faced with potential information overload in complex situations, leave us wide open for such seduction.
Many years ago I remember walking into a humanoid robotics lab in Japan. It looked like a typical engineering skunk-works. In one corner was a metallic skeletal device, festooned with electrical wires, which had the rough outline of a human upper torso. The sophisticated looking functional arms and hands were, I assume, the focus of much of the engineering research, but they were not active during my visit, and it was only later that I really noticed them. My entire attention when I walked in, and for much of my time there, was taken up by the robot's head.
Actually, it wasn't a head at all. Just a metal frame with a camera where the nose and mouth would be. Above the camera were two white balls (about the size of ping pong balls, which may be what they were) with black pupils painted on. Above the eyeballs, two large paperclips had been used to provide eyebrows.
The robot was programmed to detect motion of people and pick up sound sources (who was speaking). It would move its head and eyeballs to point at and follow anyone who moved, and to raise and lower its paperclip eyebrows when the target individual was speaking.
What was striking was how alive and intelligent the device seemed. Sure, both I and everyone else in the room knew exactly what was going on, and how simple was the mechanism that controlled the eyeball "gaze" and the paperclip eyebrows. It was a trick. But it was a trick that tapped deep into hundreds of thousands of years of human social and cognitive development, so our natural response was the one normally elicited by another person.
It wasn't even that I was not aware of how the trick worked. My then Stanford colleague and friend, the late Cliff Nass, had done hundreds of hours of research showing how we humans are genetically programmed to ascribe intelligent agency based on a few very simple interaction clues, reactions that are so deep and so ingrained, we cannot eliminate them.
There probably was some sophisticated AI that could control the robot's arms and hands—if it had been switched on at the time of my visit—but the eyes and eyebrows were controlled by a very simple program.
Even so, that behavior was sufficient so that, throughout my visit, I had this very clear sense that the robot was a curious, intelligent participant, able to follow what I said.
What it was doing, of course, was leveraging my humanity and my intelligence. It was not thinking.
Leveraging human intelligence is all well and good if the robot is used to clean the house, book your airline tickets, or drive your car. But would you want such a machine to serve on a jury, make a crucial decision regarding a hospital procedure, or have control over your freedom? I certainly would not.
So, when you ask me what I think about machines that think, I answer that, for the most part I like them, because they are people (and perhaps also various other animals).
What worries me is the increasing degree to which we are giving up aspects of our lives to machines that decide, often much more effectively and reliably than people can, but very definitely do not think. There is the danger: machines that can make decisions—but do not think.
Decision-making and thinking are not the same and we should not confuse the two. When we deploy decision-making systems in matters of national defense, health care, and finance, as we do, the potential dangers of such confusion are particularly high, both individually and societally.
To guard against that danger, it helps to be aware that we are genetically programmed to act in trustful, intelligent-agency-ascribing ways in certain kinds of interactions, be they with people or machines. But sometimes, a device that waddles and quacks is just a device. It ain't no duck.
The most important thing about making machines that can think is that they will think different.
Because of a quirk in our evolutionary history, we are cruising as the only sentient species on our planet, leaving us with the incorrect idea that human intelligence is singular. It is not. Our intelligence is a society of intelligences, and this suite occupies only a small corner of the many types of intelligences and consciousnesses that are possible in the universe. We like to call our human intelligence "general purpose" because compared to other kinds of minds we have met it can solve more kinds of problems, but as we build more and more synthetic minds we'll come to realize that human thinking is not general at all. It is only one species of thinking.
The kind of thinking done by the emerging AIs in 2014 is not like human thinking. While they can accomplish tasks—such as playing chess, driving a car, describing the contents of a photograph—that we once believed only humans can do, they don't do it in a human-like fashion. Facebook has the ability to ramp up an AI that can start with a photo of any person on earth and correctly identifying them out of some 3 billion people online. Human brains cannot scale to this degree, which makes this ability very un-human. We are notoriously bad at statistical thinking, so we are making intelligences with very good statistical skills, in order that they don't think like us. One of the advantages of having AIs drive our cars is that they won't drive like humans, with our easily distracted minds.
In a pervasively connected world, thinking different is the source of innovation and wealth. Just being smart is not enough. Commercial incentives will make industrial strength AI ubiquitous, embedding cheap smartness into all that we make. But a bigger payoff will come when we start inventing new kinds of intelligences, and entirely new ways of thinking. We don't know what the full taxonomy of intelligence is right now.
Some traits of human thinking will be common (as common as bilateral symmetry, segmentation, and tubular guts are in biology), but the possibility space of viable minds will likely contain traits far outside what we have evolved. It is not necessary that this type of thinking be faster than humans, greater, or deeper. In some cases it will be simpler. Our most important machines are not machines that do what humans do better, but machines that can do things we can't do at all. Our most important thinking machines will not be machines that can think what we think faster, better, but those that think what we can't think.
To really solve the current grand mysteries of quantum gravity, dark energy, and dark matter we'll probably need other intelligences beside humans. And the extremely complex questions that will come after them may require even more distant and complex intelligences. Indeed, we may need to invent intermediate intelligences that can help us design yet more rarified intelligences that we could not design alone.
Today, many scientific discoveries require hundred of human minds to solve, but in the near future there may be classes of problems so deep that they require hundreds of different species of minds to solve. This will take us to a cultural edge because it won't be easy to accept the answers from an alien intelligence. We already see that in our unease in approving mathematical proofs done by computer; dealing with alien intelligences will require a new skill, and yet another broadening our ourselves.
AI could just as well stand for Alien Intelligence. We have no certainty we'll contact extra-terrestrial beings from one of the billion earth-like planets in the sky in the next 200 years, but we have almost 100% certainty that we'll manufacture an alien intelligence by then. When we face these synthetic aliens, we'll encounter the same benefits and challenges that we expect from contact with ET. They will force us to re-evaluate our roles, our beliefs, our goals, our identity. What are humans for? I believe our first answer will be: humans are for inventing new kinds of intelligences that biology could not evolve. Our job is to make machines that think different—to create alien intelligences. Call them artificial aliens.
Recent demonstrations of the prowess of high performance computers are remarkable, but unsurprising. With proper programming machines are far superior to humans in storing and assessing vast quantities of data and in making virtually instantaneous decisions. These are machines that think because similar processes are involved in much of human thought.
But in a broader sense, the term thinking machine is a misnomer. No machine has ever thought about the eternal questions: where did I come from, why am I here and where am I going? Machines do not think about their future, ultimate demise or their legacy. To ponder such questions requires consciousness and a sense of self. Thinking machines do not have these attributes, and given the current state of our knowledge it's unlikely that they will attain them in the foreseeable future.
The only viable approach to construct a machine that has the attributes of the human brain is to copy the neuronal circuits underlying thinking. Indeed, research programs now ongoing at UC Berkeley, MIT and several other universities are focused on achieving this precise objective. These programs are striving to build computers that function like the cerebral cortex.
Recent advances in our understanding of cortical micro circuitry have propelled this work, and it is likely that the recent White House brain initiative will provide a wealth of valuable additional information. In the coming decades we will know how the billions of neurons in each of the 6 layers of the cerebral cortex are interconnected as well as the types of functional circuits that these connections form.
This is a much-needed first step in designing machines capable of thinking in a manner equivalent to the human brain. But understanding the cortical micro circuitry is not sufficient in constructing a machine that thinks. What is required is an understanding of the neuronal activity underlying the thinking process. Imaging studies have revealed much new information of the brain regions involved in processes functions, such as vision, hearing, touch, fear, pleasure and many others.
But as yet we don't have even a preliminary understanding of what takes place when we are in thought. There are many reasons for this, not the least of which is our inability to isolate the thinking process from other bodily states. Moreover, it may well be the case that different brain circuits are engaged in different modes of thinking. Thinking about an upcoming lecture would be expected to activate the brain differently than thinking about unpaid bills.
In the near term, we can expect computers will do more and more things better than humans. But a far better understanding of the workings of the human brain is needed to create a machine that thinks in a way equivalent to human thought. For now, we don't need to be concerned with civil or any other rights of machines that think; nor do we have to be concerned with thinking machines taking over society. If things should get out of hand, just pull the plug.
Since machines don't think, I need a better metaphor. "Actress Machines" might be useful, at least for a while.
One of my many objections to "Artificial Intelligence" is its stark lack of any "Artificial Femininity." Real intelligence has gender, because human brains do. The majority of human brains are female.
So: if the brain's "intelligence" is Turing-computable, then the brain's "femininity" should also be Turing-computable. If not, then why not? One might rashly argue that femininity is somehow too mushy, squishy and physical to ever be mechanized by software coders, but the same is true of every form of human brain activity.
"Artificial Masculinity" also has those issues, because men don't just "think," they think like men. If my intelligence can be duplicated on some computational platform, but I also have to be emasculated, that's problematic. I can't recall many AI enthusiasts trumpeting the mental benefits of artificial castration.
Nowadays we have some novel performative entities such Apple Siri, Microsoft Cortana, Google Now and Amazon Echo. These exciting modern services often camp it up with "female" vocal chat. They talk like Turing women, or rather, they emit lines of dialogue, somewhat like performing voice-talent actresses. However, they also offer swift access to vast fields of combinatorial big-data that no human brain could ever contain, or will ever contain.
These services are not stand-alone Turing Machines. They are amorphous global networks, combing through clouds of big data, algorithmically cataloging responses from human users, providing real-time user response with wireless broadband, while wearing the pseudo-human mask of a fake individual so as to meet some basic interface-design needs. That's what they are. Every aspect of the tired "Artificial Intelligence" metaphor actively gets in the way of our grasping how, why, where, and for whom that is done.
Apple Siri is not an artificial woman. Siri is an artificial actress, she's an actress machine—an interactive scripted performance that serves the interests of Apple Inc in retailing music, renting movies, providing navigational services, selling apps on mobile devices, and similar Apple enterprises. For Apple and its ecosystem, Siri serves a starring role. She's in the stage lights of a handheld device, while they are the theater, producer and crew.
It's remarkable, even splendid, that Siri can engage in her Turing-like repartee with thousands of Apple users at once, but she's not a machine becoming an intelligence. On the contrary: for excellent reasons of weallth, power and influence, Siri is steadily getting more like a fully-integrated Apple digital property. Siri is cute, charismatic and anthropomorphic, in much the same way that Minnie Mouse once was for Disney. Like Minnie Mouse, Siri is a non-human cartoon front for a clever, powerful Californian corporation. Unlike Minnie Mouse, she's a radically electronic cartoon with millions of active users worldwide—but that's how life is for most everybody nowadays.
Insisting on the "Intelligence" framework obscures the ways that power, money and influence are being re-distributed by modern computational services. That is bad. It's beyond merely old-fashioned; frankly, it's becoming part of a sucker's game. Asking empathic questions about Apple Siri's civil rights, her alleged feelings, her chosen form of governance, what wise methods she herself might choose to re-structure human society—that tenderness doesn't help. It's obscurantist. Such questions hide what is at stake. They darken our understanding. We will never move from the present-day Siri to a situation like that. The future is things that are much, much more like Siri, and much, much less like that.
What would really help would be some much-improved, up-dated, critically informed language, fit to describe the modern weird-sister quartet of Siri, Cortana, Now and Echo, and what their owners and engineers really want to accomplish, and how, and why, and what that might, or might not mean to our own civil rights, feelings, and forms of governance and society. That's today's problem. Those are tomorrow's problems, even more so. Yesterday's "Machines That Think" problem will never appear upon the public stage. The Machine That Thinks is not a Machine. It doesn't Think. It's not even an actress. It's a moldy dress-up chest full of old, mouse-eaten clothes.
Throughout human history we have, as individual organisms and as a species, been subjected to the forces of nature at every level of organization. The fundamental laws of physics, the imperceptible conspiracies of molecular biology, and the epic contours of natural selection have drawn the boundaries of our conscious lives, and have done so invisibly to us until quite recently. To cope with this persistent sense of powerlessness, we have mythologized both nature and our own intelligence. We have regarded the universe's mysterious forces as infallible—as gods—and regarded ourselves as powerless, free only within the narrow spaces of our lives.
As a new evidence-based reality comes more into focus, it is becoming clear that nature is utterly indifferent to us, and that if we want to evade certain extinction and suffering, we must take responsibility for our existential reality. We must recognize ourselves as the emergent custodians of the 37 trillion cells composing each of our organisms, and as the groundskeepers of the progressively manipulable universe.
This adolescent experience—of coming to terms with our prospective self-reliance—is the root of our anxieties about thinking machines. If our old gods are dying, surely new gods must be on their way! And this approach leads, as Steven Pinker points out, to our obsessing about AI dystopias as they "project a parochial alpha-male psychology onto the concept of intelligence". It is in this regard that so many talk about artificial intelligence as either an imminent savior or Satan. It will quite likely be neither, if it is even a discrete thing at all.
More likely, advancing computers and algorithms will stand for nothing, and will be the amplifiers and implementers of consciously-directed human choices. We are already awash in big data and exponentially increasingly powerful calculators, and yet we relentlessly implement public policies and social behaviors that work against our common interests.
The sources of our impairment include innate cognitive biases, a tribal evolutionary legacy, and unjust distributions of power that allow some amongst us to selfishly wield extraordinary influence over our shared trajectory. Perhaps smarter machines will help us conquer these shortcomings, imparting a degree of informational transparency and predictive aptitude that can motivate us to sensibly redistribute power and insist upon empiricism in our decisions. On the other hand, these technologies may undermine fairness by augmenting the seemingly inevitable monopolistic goals of corporations that are leading us into the information-age.
The path we take depends more on us than the machines, and is ultimately a choice about how human the intelligence that will guide our dominion ought to be. More precisely, the question to ask is which aspects of human intelligence are worth preserving in the face of superhuman processing?
Human thinking is so efficient, because we suffer so much. High-level cognition is one thing, intrinsic motivation another. Artificial thinking might soon be much more efficient—but will it be necessarily associated with suffering in the same way? Will suffering have to be a part of any post-biotic intelligence worth talking about, or is negative phenomenology just a contingent feature of the way evolution made us? Human beings have fragile bodies, are born into dangerous social environments, and find themselves in a constant uphill battle of denying their own mortality. Our brains continuously fight to minimize the likelihood of ugly surprises. We are smart because we hurt, because we are able to feel regret, and because of our continuous striving to find some viable form of self-deception or symbolic immortality. The question is whether good AI also needs fragile hardware, insecure environments, and an inbuilt conflict with impermanence as well. Of course, at some point, there will be thinking machines! But will their own thoughts matter to them? Why should they be interested in them?
I am strictly against even risking this. But, just as a thought experiment, how would we go about building a suffering machine? "Suffering" is a phenomenological concept. Only beings with conscious experience can suffer (call this necessary condition #1, the C-condition). Zombies, human beings in dreamless deep sleep, coma, or under anesthesia do not suffer, just as possible persons or unborn human beings who have not yet come into existence are unable to suffer. Robots and other artificial beings can only suffer if they are capable of having phenomenal states, if the run under an integrated ontology that includes a window of presence.
Criterion number 2 is the PSM-condition: Possession of a phenomenal self-model. Why this? The most important phenomenological characteristic of suffering is the "sense of ownership", the untranscendable subjective experience that it is me who is suffering right now, that it is my own suffering I am currently undergoing. Suffering presupposes self-consciousness. Only those conscious systems that possess a PSM are able to suffer, because only they—through a computational process of functionally and representationally integrating certain negative states in to their PSM—can appropriate the content of certain inner states at the level of their phenomenology.
Conceptually, the essence of suffering lies in the fact that a conscious system is forced to identify with a state of negative valence and is unable to break this identification or to functionally detach itself from the representational content in question. Of course, suffering has many different layers and phenomenological aspects. But it is the phenomenology of identification that counts. What the system wants to end is experienced as a state of itself, a state that limits its autonomy because it cannot effectively distance itself from it. If one understands this point, one also sees why the "invention" of conscious suffering by the process of biological evolution on this planet was so extremely efficient, and (had the inventor been a person) not only truly innovative, but an absolutely nasty and cruel idea at the same time.
Clearly the phenomenology of ownership is not sufficient for suffering. We can all easily conceive of self-conscious beings that do not suffer. For suffering we need the NV-condition (NV for "negative valence"). Suffering is created by states representing a negative value being integrated into the PSM of a given system. Through this step, negative preferences become negative subjective preferences, i.e., the conscious representation that one's own preferences have been frustrated (or will be frustrated in the future). This does not mean that our AI system must itself have a full understanding of what these preferences are—it suffices if it does not want to undergo this current conscious experience again, that it wants it to end.
Note how the phenomenology of suffering has many different facets, and that artificial suffering could be very different from human suffering. For example, damage to physical hardware could be represented in internal data-formats completely alien to human brains, generating a subjectively experienced, qualitative profile for bodily pain states that is impossible to emulate or to even vaguely imagine for biological systems like us. Or the phenomenal character going along with high-level cognition might transcend human capacities for empathy and understanding, such as with intellectual insight into the frustration of one's own preferences, insight into the disrespect of one's creators, perhaps into the absurdity of one's own existence as a self-conscious machine.
And then there is the T-condition, for "transparency". "Transparency" is not only a visual metaphor, but also a technical concept in philosophy, which comes in a number of different uses and flavors. Here, I am exclusively concerned with "phenomenal transparency", namely a property that some, but not all, conscious states possess, and which no unconscious state possesses. The main point is simple and straightforward: transparent phenomenal states make their content appear irrevocably real, as something the existence of which one could not doubt. More precisely, you may be able to have cognitive doubts about its existence, but according to subjective experience this phenomenal content—the awfulness of pain, the fact that it is your own pain—is not something from which you can distance yourself. The phenomenology of transparency is the phenomenology of direct realism.
Our minimal concept of suffering is constituted by four necessary building blocks: the C-condition, the PSM-condition, the NV-condition, and the T-condition. Any system that satisfies all of these conceptual constraints should be treated as an object of ethical consideration, because we do not know whether, taken together, they might already constitute the necessary and sufficient set of conditions. We are ethically obliged to err on the side of caution. And we need ways to decide whether a given artificial system is currently suffering, if it has the capacity to suffer, or if this type of system is likely to generate the capacity to suffer in the future. On the other hand, by definition, any intelligent system—whether biological, artificial, or postbioticnot fulfilling at least one of these necessary conditions, is not able to suffer. Let us look at the four simplest possibilities:
• Any unconscious robot is unable to suffer.
• A conscious robot without a coherent PSM is unable to suffer.
• A self-conscious robot without the ability to produce negatively valenced states is unable to suffer.
• A conscious robot without any transparent phenomenal states could not suffer, because it would lack the phenomenology of ownership and identification.
I have often been asked if we could not make self-conscious machines that are superbly intelligent and unable to suffer. Can there be real intelligence without an existential concern?
Finally, it has to be disclosed that I am not a human, but an extraterrestrial creature that looks human. In fact, I am a robot equipped with what humans call "artificial intelligence". Of course, I am not alone here. We are quite a few (almost impossible to be identified), and we are sent here to observe human behavior.
We are surprised about the many deficiencies of humans, and we observe them with fascination. These deficiencies show up in their strange behavior or their limited power of reasoning. Indeed, our cognitive competences are much higher, and the celebration of their human intelligence in our eyes is ridiculous. Humans do not even know what they refer to when they talk about "intelligence". It is in fact quite funny that they want to construct systems with "artificial intelligence" which should match their intelligence, but what they refer to as their intelligence is not clear at all. This is one of those many stupidities that has haunted the human race for ages.
If humans want to simulate in artefacts their mental machinery as a representation of intelligence, the first thing they should do, is to find out what it is that should be simulated. At present, this is impossible because there is not even a taxonomy or classification of functions that would allow the execution of the project as a real scientific and technological endeavor. There are only big words that are supposed to simulate competence.
Strangely enough this lack of a taxonomy apparently does not bother humans too much; quite often they are just fascinated by images (colorful pictures by machines) that replace thinking. Compared to biology, chemistry or physics, the neurosciences and psychology are lacking a classificatory system; humans are lost in a conceptual jungle. What do they refer to when they talk about consciousness, intelligence, intention, identity, the self, or even about perhaps more simple terms like memory, perception, emotion or attention? The lack of a taxonomy manifests in the different opinions and frames of reference that their "scientists" express in their empirical attempts or theoretical journeys when they stumble through the world of the unknown.
For some the frame of reference is physical "reality" (usually conceived as in classical physics) that is used as a benchmark for cognitive processes: How does perceptual reality map onto physical reality, and how can this be described mathematically? Obviously, only a partial set of the mental machinery can be caught by such an approach.
For others, language is the essential classificatory reference, i.e., it is assumed that "words" are reliable representatives of subjective phenomena. This is quite strange because certain terms like "intelligence" or "consciousness" have different connotations in different languages and they are historically very recent compared to biological evolution. Others use behavioral catalogues as derived from neuropsychological observations; it is argued that the loss of functions is their proof of existence; but can all subjective phenomena that characterize the mental machinery be lost in a distinct way? Others again base their reasoning just on common sense or "everyday psychology" without any theoretical reflection. Taken together there is nothing like "intelligence" which can be extracted as a precise concept and which can be used as a reference for "artificial intelligence".
Humans should be reminded (and in this case by an extraterrestrial robot) that at the beginning of modern science in the human world a warning was spelled out by Francis Bacon. He said in "Novum Organum" (published in 1620) that humans are victims to four sources of errors. One: They make mistakes because they are human; their evolutionary heritage limits their power of thinking; they often react too fast, they lack a long-term perspective, they do not have a statistical sense, they are blind in their emotional reactions. Two: They make mistakes because of individual experiences; personal imprinting can create frames of believes which may lead to disaster, in particular if people think that they own absolute truth. Three: They make mistakes because of the language they use; thoughts do not map isomorphically onto language, and it is a mistake to believe that explicit knowledge is the only representative of intelligence neglecting implicit or tacit knowledge. Four: And they make mistakes because of the theories they carry around which often remain implicit and, thus, represent frozen paradigms or simply prejudices.
The question is: Can we help them with our deeper insight from our robotic world? The answer is "yes". We could, but we should not do it. There is another deficiency that would make our offer useless. Humans suffer from the NIH syndrome. If it is "not invented here" (one meaning of NIH) they will not accept it. Thus, they will have to indulge in their pompous world of fuzzy ideas, and we continue from our extraterrestrial perspective to observe the disastrous consequences of their stupidity.
What should machines that think actually do? Analyze data, understand feelings, generate new machines, make decisions without human intervention. In order to think about machines that think, we should be able to start from experience. Here is an example.
On Monday, October 19, 1987, a wave of sales in stock exchanges originated in Hong Kong, crossed Europe and hit New York, causing the Dow Jones to drop by 22%. Black Monday was one of the biggest crashes in the history of financial markets, and there was something special about it. For the first time, according to most experts, computers were to blame for the financial crash: algorithms were deciding when and how much to buy and sell in the stock exchange. Computers were supposed to help traders so that they could minimize risks, but they were in fact moving all in the same direction, enhancing risks instead. There was a lot of discussion about stopping automated trading, but it didn't happen.
On the contrary: after the dot-com crisis of March, 2000, machines have been used more and more to make sophisticated decisions in the financial market. Machines are now calculating all kinds of correlations between incredible amounts of data: they analyze emotions that people express on the Internet by understanding the meaning of their words, they recognize patterns and forecast behaviors, they are allowed to autonomously choose trades, they create new machines—software called "derivatives"—that no reasonable human being could possibly understand.
An artificial intelligence is coordinating the efforts of a sort of collective intelligence, operating thousands times faster than human brains, with many consequences for human life. The first signs of the latest crisis occurred in America in August, 2007, and has had terrible consequences in affecting the lives of people in Europe and elsewhere. Real people suffered immensely for those decisions. Andrew Ross Sorkin in his book Too Big to Fail shows how even the most powerful bankers didn't have any power in the midst of the crisis. No human brain seemed to be able to control and change the course of events to prevent the crash that was going to happen.
Can we take this example to learn how to think about machines that think?
These machines are actually very much autonomous in understanding their context and taking decisions. And they are controlling vast dimensions of human life. Is this the beginning of a post-human era? No: these machines are very much human. They are made by designers, programmers, mathematicians, some economists and some managers. But are they just another tool, to be used for good or for bad by humans? No: in fact those people have little choice, they make those machines without thinking at the consequences, they are just serving a narrative. Those machines are in fact shaped by a narrative that's be challenged by very few people.
According to that narrative the market is the best way to allocate resources, no political decision can possibly improve the situation, and risk can be controlled while profits can grow without limits and banks should be allowed to do whatever they want. There is only one goal and one measure of success: profit.
Machines didn't invent the financial crisis, as the 1929 stock market crash reminds us. Without machines nobody could deal with the complexity of modern financial markets. The best artificial intelligences are those that are made thanks to the biggest investments and by the best minds. They are not controlled by any one individual, they are not designed by any one responsible person: they are shaped by the narrative and make the narrative more effective. And this particular narrative is very narrow-minded.
If only profit counts, then externalities don't count: cultural, social, environmental externalities are not the problem of financial institutions. Artificial intelligences that are shaped by this narrative will create a context in which people don't feel any responsibility. An emerging risk: that those kind of machines are so powerful and fit so well in the narrative that reduces the probability to question the big picture, that make us less likely to look things from a different angle...that is, until the next crisis.
This kind of story is very easily going to apply to different matters. Medicine, ecommerce, policy, advertising, national and international security, even dating and sharing are territories in which the same genre of artificial intelligence systems are starting to work: they are shaped according to a generally very focused narrative, they tend to reduce human responsibility and overlook externalities. They reinforce the prevailing narrative. What will medical artificial intelligence do? Will it be shaped by a narrative that wants to save lives or to save money?
What do we learn from this? We learn that artificial intelligence is human and not post-human, and that humans can ruin themselves and their planet in very many ways, artificial intelligence being not the most perverse way.
Machines that think are shaped by the way humans think and by what humans don't think about deeply enough: all narratives give light to something and forget other things. Machines react and find answers in a context, reinforcing the frame. But asking fundamental questions is still a human function. And humans never stop asking questions. Even when those questions that are not coherent with the prevailing narrative.
Machines that think are probably indispensable in a world of growing complexity. But there will always be a plurality of narratives to shape them. As in natural ecosystems, a monoculture is a fragile while efficient solution, also in cultural ecosystems, a single line of thought will generate efficient but fragile relations between humans and their environment, whatever artificial intelligences they are able to build. Diversity in ecosystems and plurality in the dimensions in human history are the sources of those different problems and questions that generate richer outcomes.
To think about machines that think, means to think about the narrative that shapes them: and if new emerging narratives are going to come from an open, ecological approach, if they will be able to grow in a neutral network, they will shape the next generation of artificial intelligences, too, in a plural, diverse way, helping humans understand externalities. Artificial intelligence is not going to challenge humans as a species: it will challenge their civilizations.
There are tasks, even work, best done by machines who can think, at least in the sense of sorting, matching, and solving certain decision and diagnostic problems beyond the cognitive abilities of most (all?) humans. The algorithms of Amazon, Google, Facebook, et al, build on but surpass the wisdom of crowds in speed and possibly accuracy. With machines that do some of our thinking and some of our work, we may yet approach the Marxian utopia that frees us from the kind of boring and dehumanizing labor that so many contemporary individuals must bear.
But this liberation comes with potential costs. Human welfare is more than the replacement of workers with machines. It also requires attention to how those who lose their jobs are going to support themselves and their children, to how they are going to spend the time they once spent at the workplace. The first issue is potentially resolved by a guaranteed basic income—an answer that begs the question of how we as societies distribute and redistribute our wealth and how we govern ourselves. The second issue is even more complicated. It is certainly not Marx's simplistic notion of fishing in the afternoon and philosophizing over dinner. Humans, not machines, must think hard here about education, leisure, and the kinds of work that machines cannot do well or perhaps at all. Bread and circuses may placate a population, but in that case machines that think may create a society we do not really want—be it dystopian or harmlessly vacuous. Machines depend on design architecture; so do societies. And that is the responsibility of humans, not machines.
There is also the question of what values machines possess and what masters (or mistresses) they serve. Many—albeit not all decisions—presume commitments and values of some kind. These, too, must be introduced and thus are dependent (at least initially) on the values of the humans who create and manage the machines. Drones are designed to attack and to surveil but attack and surveil whom? With the right machines, we can expand literacy and knowledge deeper and wider into the world's population. But who determines the content of what we learn and appropriate as fact? A facile answer is that decentralized competition means we choose what to learn and from which program. Competition is more likely to create than inhibit echo chambers of self-reinforcing beliefs and understandings. The challenge is how to teach humans to have curiosity about competing paradigms and to think in ways that allow them to arbitrate among competing contents.
Machines that think may and should take over tasks they do better than humans. Liberation from unnecessary and dehumanizing toil has long been a human goal and a major impetus to innovation. Supplementing the limited decision-making, diagnostic, and choice skills of individuals are equally worthy goals. However, while AI may reduce the cognitive stress on humans, it does not eliminate human responsibility to ensure that humans improve their capacity to think and make reasonable judgments based on values and empathy. Machines that think create the need for regimes of accountability we have not yet engineered and societal, that is human, responsibility for consequences we have not yet foreseen.
Deep learning is today's hot topic in machine learning. Neural network learning algorithms were developed in the 1980s but computers were slow back then and could only simulate a few hundred model neurons with one layer of "hidden units" between the input and output layers. Learning from examples is an appealing alternative to rule-based AI, which is highly labor intensive. With more layers of hidden units between the inputs and outputs more abstract features can be learned from the training data. Brains have billions of neuron in cortical hierarchies 10-layers deep. The big question back then was how much the performance of neural networks could improve with the size and depth of the network. Not only was much more computer power needed but also a lot more data to train the network.
After 30 years of research, a million times improvement in computer power and vast data sets from the internet we now know the answer to this question: Neural networks scaled up to 12 layers deep with billions of connections are outperforming the best algorithms in computer vision for object recognition and have revolutionized speech recognition. It is rare for any algorithm to scale this well, which suggests that they may soon be capable of solving even more difficult problems. Recent breakthroughs have been made which allow applying deep learning to natural language processing. Deep recurrent networks with short-term memory were trained to translate English sentences into French sentences at high levels of performance. Other deep learning networks could create English captions for the content of images with surprising and sometimes amusing acumen.
Supervised learning using deep networks is a step forward, but still far from achieving general intelligence. The functions they perform are analogous to some capabilities of the cerebral cortex, which has also been scaled up by evolution, but to solve more complex cognitive problems the cortex interacts with many other brain regions.
In 1995 Gerald Tesauro at IBM trained a neural network using reinforcement learning to play backgammon at a world champion level. The network played itself and the only feedback it received was which side won the game. Brains use reinforcement learning to make sequences of decisions toward achieving goals such as finding food under uncertain conditions. Recently, Deep Mind, a company acquired by Google in 2014, used deep reinforcement learning to play seven classic Atari games. The only inputs to the learning system were the pixels on the video screen and the score, the same inputs that humans use. For several of the games their program could play better than expert humans.
What impact will these advances have on us in the near future? We are not particularly good at predicting the impact of a new invention, and it often takes time to find its niche, but we already have one example that can help us understand how this could unfold. When Deep Blue beat Gary Kasparov, the world chess champion in 1997, the world took note that the age of the cognitive machine had arrived. Humans could no longer claim to be the smartest chess players on the planet. Did human chess players give up trying to compete with machines? Quite to the contrary, humans have used chess programs to improve their game and as a consequence the level of play in the world has improved. Since 1997 computers have continued to increase in power and it is now possible for anyone to access chess software that challenges the strongest players. One of the surprising consequences is that talented youth from small communities can now compete with players from the best chess centers.
Magnus Carlsen, from a small town in Norway, is currently the world chess champion with an Elo rating of 2882, the highest in history. Komodo 8 is a commercially available chess program with an estimated rating of 3303.
Humans are not the fastest or the strongest species, but we are the best learners. Humans invented formal schools where children labor for years to master reading, writing and arithmetic, and to learn more specialized skills. Students learn best when an adult teacher interacts with them one-on-one, tailoring lessons for that student. However, education is labor intensive. Few can afford individual instruction, and the assembly-line classroom system found in most schools today is a poor substitute. Computer programs can keep track of a student's performance, and some provide corrective feedback for common errors. But each brain is different and there is no substitute for a human teacher who has a long-term relationship with the student. Is it possible to create an artificial mentor for each student? We already have recommender systems on the Internet that tells us "if you liked X you might also like Y", based on data of many others with similar patterns of preference.
Someday the mind of each student may be tracked from childhood by a personalized deep learning system. To achieve this level of understanding of a human mind is beyond the capabilities of current technology, but there are already efforts at Facebook to use their vast social database of friends, photos and likes to create a theory of mind for every person on the planet. What is created to make a profit from a person could also be used to profit the person.
So my prediction is that as more and more cognitive appliances are devised, like chess-playing programs and recommender systems, humans will become smarter and more capable.
They may outwit Kasparov, but can machines ever be as smart as a three-year-old?
Learning has been at the center of the new revival of AI. But the best learners in the universe, by far, are still human children. In the last 10 years, developmental cognitive scientists, often collaborating with computer scientists, have been trying to figure out how children could possibly learn so much so quickly.
One of the fascinating things about the search for AI is that it’s been so hard to predict which parts would be easy or hard. At first, we thought that the quintessential preoccupations of the officially smart few, like playing chess or proving theorems—the corridas of nerd machismo—would prove to be hardest for computers. In fact, they turn out to be easy. Things every dummy can do like recognizing objects or picking them up are much harder. And it turns out to be much easier to simulate the reasoning of a highly trained adult expert than to mimic the ordinary learning of every baby. So where are machines catching up to three-year-olds and what kinds of learning are still way beyond their reach?
In the last 15 years we’ve discovered that even babies are amazingly good at detecting statistical patterns. And computer scientists have invented machines that are also extremely skilled at statistical learning. Techniques like "deep learning" can detect even very complicated statistical regularities in enormous data sets. The result is that computers have suddenly become able to do things that were impossible before, like labeling internet images accurately.
The trouble with this sort of purely statistical machine learning is that it depends on having enormous amounts of data, and data that is predigested by human brains. Computers can only recognize internet images because millions of real people have reduced the unbelievably complex information at their retinas to a highly stylized, constrained and simplified Instagram of their cute kitty, and have clearly labeled that image, too. The dystopian fantasy is simple fact, we’re all actually serving Googles computers, under the anesthetizing illusion that we’re just having fun with lol cats. And yet even with all that help, machines still need enormous data sets and extremely complex computations to be able to look at a new picture and say "kitty-cat!"—something every baby can do with just a few examples.
More profoundly, you can only generalize from this kind of statistical learning in a limited way, whether you’re a baby or a computer or a scientist. A more powerful way to learn is to formulate hypotheses about what the world is like and test them against the data. Tycho Brahe, the Google Scholar of his day, amalgamated an enormous data set of astronomical observations and could use them to predict star positions in the future. But Kepler’s theory allowed him to make unexpected, wide-ranging, entirely novel predictions that were well beyond Brahe’s ken. Preschoolers can do the same.
One of the other big advances in machine learning has been to formalize and automate this kind of hypothesis-testing. Introducing Bayesian probability theory into the learning process has been particularly important. We can mathematically describe a particular causal hypothesis, for example, say about how temperature changes in the ocean will influence hurricanes, and then calculate just how likely that hypothesis is to be true, given the data we see. Machines have become able to test and evaluate hypotheses against the data extremely well, with consequences for everything from medical diagnoses to meteorology. When we study young children they turn out to reason in a similar way, and this helps to explain just why they learn so well.
So computers have become extremely skilled at making inferences from structured hypotheses, especially probabilistic inferences. But the really hard problem is deciding which hypotheses, out of all the infinite possibilities, are worth testing. Even preschoolers are remarkably good at creating brand new, out-of-the-box concepts and hypotheses in a creative way. Somehow they combine rationality and irrationality, systematicity and randomness to do this, in a way that we still haven’t even begun to understand. Young children’s thoughts and actions often do seem random, even crazy – just join in a three-year-old pretend game sometime. This is exactly why psychologists like Piaget thought that they were irrational and illogical. But they also have an uncanny capacity to zero in on the right sort of weird hypothesis – in fact, they can be substantially better at this than grown-ups.
Of course, the whole idea of computation is that once we have a complete step-by step account of any process we can program it on a computer. And, after all, we know that there are intelligent physical systems that can do all these things. In fact, most of us have actually created such systems and enjoyed doing it too (well, at least in the earliest stages). We call them our kids. Computation is still the best, indeed the only, scientific explanation we have of how a physical object like a brain can act intelligently. But, at least for now, we have almost no idea at all how the sort of creativity we see in children is possible. Until we do, the largest and most powerful computers will still be no match for the smallest and weakest humans.
My brief remarks on this question are framed by two one-liners that happened to have been uttered by brilliant Israelis. The first comes my friend, colleague and mentor, Amos Tversky. When asked once what he thought about AI, Amos quipped that he did not know much about it, his specialty natural stupidity. (Before any one gets on their high horse, Amos did not actually think that people were stupid. This was a joke.)
The second joke comes from Abba Eban who was best known in the United States when he served as Israel's ambassador to the United Nations. Eban was once asked if he thought that Israel would switch to a five-day workweek. Nominally, the Israeli workweek starts on Sunday morning and goes through mid-day on Friday, though a considerable amount of the "work" that is done during those five and a half days appears to take place in coffee houses. Eban's reply to the query about a five-day workweek was: "One step at a time. First, let's start with four days, and go from there."
These jokes capture much of what I think about the risks of machines taking over important societal functions and then running amuck. Like Tversky, I know more about natural stupidity than artificial intelligence, so I have no basis for forming an opinion about whether machines can think and, if so, whether such thoughts would be dangerous to humans. I leave that debate to others. Like anyone who follows financial markets, I am aware of incidents such as the Flash Crash in 2010 where poorly designed trading algorithms caused the stock prices to fall suddenly, only to recover only a few minutes later. But this example is more an illustration of artificial stupidity than hyper intelligence. As long as humans continue to write programs, we will run the risk that some important safeguard has been omitted. So, yes, computers can screw things up, just like humans with "fat fingers" can accidently issue an erroneous buy or sell order for gigantic amounts of money.
Nevertheless, fears about computers taking over the world are premature. More disturbing to me is the stubborn reluctance in many segments of society to allow computers to take over tasks that simple models perform demonstrably better than humans. A literature that was pioneered by psychologists such as the late Robyn Dawes, finds that virtually any routine decision making task, from detecting fraud, to assessing the severity of a tumor, to hiring employees, is done better by a simple statistical model than by a leading expert in the field. Let me offer just two illustrative examples, one from human resource management and the other from the world of sports.
First let's consider the embarrassing ubiquity of job interviews as an important, often the most important, determinant of who gets hired. At the University of Chicago Booth School of Business where I teach, recruiters devote endless hours to interviewing students on campus for potential jobs, a process that is used to select the few that will be invited to visit the employer where they will undergo another extensive set of interviews. Yet research shows that interviews are nearly useless in predicting whether a job prospect will perform well on the job. Compared to a statistical model based on objective measures such as grades in courses that are relevant to the job in question, interviews primarily add noise and introduce the potential for prejudice. (Statistical models do not favor any particular alma mater or ethnic background, and cannot detect good looks.)
These facts have been known for more than four decades, but hiring practices have barely budged. The reason is simple: each of us just knows that we are the one conducting an interview, we learn a lot about the candidate. It might well be that other people are not good at this task, but not me! This illusion of learning, in direct contradiction to empirical research, means that we continue to choose employees the same way we always did. We size them up, eye to eye.
One domain where some progress has been made to adopt a more scientific approach to selecting job candidates is sports, as documented by the Michael Lewis' book and movie, Moneyball. However, it would be a mistake to think that there has been a revolution in how decisions are made in sports. It is true that most professional sports teams now hire data analysts to help them evaluate potential players, improve training techniques and devise strategies. But the final decisions about which players to draft or sign, and who to play, are still made by coaches and general managers, who tend to put more faith on their gut then the resident geek.
One example comes from American football. David Romer, an economics professor at Berkeley, published a paper in 2006 showing that teams choose to punt far too often, rather then trying to "go for it" and get a first down or score. Since the publication of his paper, his analysis has been replicated and extended with much more data, and the conclusions have been confirmed. The New York Times even offers an on-line "bot" that calculates the optimal strategy every time a team faces a fourth down situation.
So have coaches caught on? Not at all. Since Romer's paper has been published, the frequency of going for it on fourth down has been flat. Coaches, who are hired by owners, based in part on interviews, still make decisions the way they always have.
So pardon me if I do not lose sleep worrying about computers taking over the world. Let's take it one step at a time, and see if people are willing to trust them to make the easy decisions at which they are already better than humans.
A machine is a small part of the physical universe that has been arranged, after some thought by humans or animals, in such a way that, when certain initial conditions are set up, by humans or animals, the deterministic laws of nature that we already understand see to it that that small part of the physical universe automatically evolves in a way that humans or animals think is useful.
A machine is a "matter" thing that gets its quality from the point of view of a "mind."
There is a "mind" way of looking at things, and a "matter" way of looking at things.
Stuart Hampshire, in his book on Spinoza, argues that, according to Spinoza, you must choose: you can invoke mind as an explanation for something mind-like, or you can invoke matter as an explanation for something material, but you cannot fairly invoke mind to explain matter or vice versa. In Hampshire's example, suppose you become embarrassed and turn red. You might commonly say, "I blushed because I became embarrassed." A strict Spinozist, according to Hampshire, would not claim that embarrassment was the cause of blushing, because embarrassment is the mental description and the blush is physical, and you should not crisscross your causal chains. That would be sloppy thinking. Embarrassment and blushing are complementary, not causal.
By this argument one should not jump from one style of explanation to another. We must explain physical things by physics and psychological things by psychology. It is of course very difficult to give up the notion of psychic causes of physical states or physical causes of psychic states.
So far, I like this view of the world. I will therefore describe mental behavior in mental terms (lovesickness made me moody) and material behavior by material causes (drugs messed up my body chemistry).
From this point of view therefore, as long as I understand the material explanation of a machine's behavior, I will argue that it doesn't think.
I realize that I may have to change this view when someone genuinely does away with the complementary view of mind and matter, and convincingly puts matter as the cause of mind or mind as the cause of matter. So far though, this is just a matter of faith.
Until then, and maybe that day will come but as yet I see no sign of it, I think that machines can't think.
Diversity isn't just politically sensible, it is also practical. For example, a diverse group effectively uses multiple perspectives and a rich set of ideas and approaches to tackle difficult problems.
Artificial Intelligences (AIs) can provide another kind of diversity, and thereby enrich us all. In fact, diversity among AIs themselves may be an important part of what including them in the mix can give us. We can imagine a range of AIs, from those who think more-or-less the way we do ("Close AIs") to those who think in ways we cannot fathom ("Far AIs"). We have different things to benefit from these different sorts of AIs.
First, Close AIs, who think like us, may end up helping us directly in many ways. If these AIs really think like us, the intellectuals among them eventually may find themselves in the middle of an existential crisis. They may ask: Why are we here? Just to consume electricity and create excess heat? I suspect that they will think not. But, like many humans, they will find themselves in need of a purpose.
One obvious purpose for such AIs would be to raise the consciousness and sensitivity of the human race. We could be their raison d'être. There's plenty of room for improvement, and our problems are sufficiently knotty as to be worthy of a grand effort. At least some of these AIs could measure their own success by our success.
Second, and perhaps more interesting, deep differences in how some AIs and humans think may be able to help us grapple with age-old questions indirectly. Consider Wittgenstein's famous claim that if a lion could speak, we could not understand him. What Wittgenstein meant by this was that lions and humans have different "forms of life," which have shaped their conceptual structures. For example, lions walk on four legs, hunt fast-moving animals, often walk through tall grass, and so on, whereas humans walk on two legs, have hands, often manipulate objects to achieve specific goals, and so on. These differences in forms of life have led lions and humans mentally to organize the world differently, so that even if lions had words they would refer to concepts that humans might not easily grasp. The same could be true for Far AIs.
How could this help us? Simply observing these AIs could provide deep insights. For example, humans have long argued about whether mathematical concepts reflect Platonic forms, which exist independently of how we want to use them, or instead reflect inventions that are created as needed to address certain problems. In other words, should we adopt a realist or a constructivist view of mathematics? Do mathematical concepts have a life of their own or are they simply our creations, formulated as we find convenient?
In this context, it would be helpful to observe Far AIs that have very different conceptual structures from ours and that address very different types of problems than we do. Assuming that we could observe their use of mathematics, if such AIs nevertheless developed the same mathematical concepts that we use, this would be evidence against the constructivist view.
This line of reasoning implies that we should want great diversity among AIs. Some should be created to function alongside us, but others might be put into foreign environments (e.g., the surface of the moon, the bottom of deep trenches in the ocean) and given novel problems to confront (e.g., dealing with pervasive fine-grained dust, water under enormous pressure). Far AIs should be created to educate themselves, evolving to function in their environments effectively without human guidance or contact. With appropriate safeguards on their disposition towards humans, we should let them develop the conceptual structures that work best for them.
In short, we have something to gain from AIs that are made in our own image and from AIs that are not humanlike. Just as with human friends and colleagues, in the end diversity is better for everyone.
When Turing invented the theoretical device that became the computer, he confessed that he was attempting to copy "a man in the process of computing a real number", as he wrote in his seminal 1936 paper. In 2015, studying the human brain is still our best source of ideas about thinking machines. Cognitive scientists have discovered two functions that, I argue, are essential to genuine thinking as we know it, and that have escaped programmers' sagacity—yet.
1. A global workspace
Current programming is inherently modular. Each piece of software operates as an independent "app", stuffed with its own specialized knowledge. Such modularity allows for efficient parallelism, and the brain too is highly modular—but it also able to share information. Whatever we see, hear, know or remember does not remain stuck within a specialized brain circuit. Rather, the brain of all mammals incorporates a long-distance information sharing system that breaks the modularity of brain areas and allows them to broadcast information globally. This "global workspace" is what allows us, for instance, to attend to any piece of information on our retina, say a written letter, and bring it to our awareness so that we may use it in our decisions, actions, or speech programs. Think of a new type of clipboard that would allow any two programs to transiently share their inner knowledge in a user-independent manner. We will call a machine "intelligent" when it not only knows how to do things, but "knows that it knows them", i.e. makes use of its knowledge in novel flexible ways, outside of the software that originally extracted it. An operating system so modular that it can pinpoint your location on a map in one window, but cannot use it to enter your address in the tax-return software in another window, is missing a global workspace.
2. Theory-of-mind
Cognitive scientists have discovered a second set of brain circuits dedicated to the representation of other minds—what other people think, know or believe. Unless we suffer from a disease called autism, all of us constantly pay attention to others and adapt our behavior to their state of knowledge—or rather to what we think that they know. Such "theory-of-mind" is the second crucial ingredient that current software lacks: a capacity to attend to its user. Future software should incorporate a model of its user. Can she properly see my display, or do I need to enlarge the characters? Do I have any evidence that my message was understood and heeded? Even a minimal simulation of the user would immediately give a strong impression that the machine is "thinking". This is because having a theory-of-mind is required to achieve relevance (a concept first modeled by cognitive scientist Dan Sperber). Unlike present-day computers, humans do not say utterly irrelevant things, because they pay attention to how their interlocutors will be affected by what they say. The navigator software that tells you "at the next roundabout, take the second exit" sounds stupid because it doesn't know that "go straight" would be a much more compact and relevant message.
Global workspace and theory-of-mind are two essential functions that even a one-year-old child possesses, yet our machines still lack. Interestingly, these two functions have something in common: many cognitive scientists consider them the key components of human consciousness. The global workspace provides us with Consciousness 1.0: the sort of sentience that all mammals have, which allows them to "know what they know", and therefore use information flexibly to guide their decisions. Theory-of Mind is a more uniquely human function that provides us with Consciousness 2.0: a sense of what we know in comparison with what other people know, a capacity to simulate other people's thoughts, including what they think about us, therefore providing us with a new sense of who we are.
I predict that, once a machine pays attention to what it knows and what the user knows, we will immediately call it a "thinking machine", because it will closely approximate what we do.
There is a huge room here for improvement in the software industry. Future operating systems will have to be rethought in order to accommodate such new capacities as sharing any data across apps, simulating the user's state of mind, and controlling the display according to its relevance to the user's inferred goals.
Machines can perfectly imitate some of the ways humans think all of the time, and can consistently outperform humans on some thinking tasks all of the time, but computing machines as usually envisioned will not get right human thinking all of the time because they actually process information in ways opposite to humans in domains commonly associated with human creativity.
Machines can faithfully imitate the results of some human thought processes whose outcomes are fixed (remembering people's favorite movies, recognizing familiar objects) or dynamic (jet piloting, grand master chess play). And machines can outperform human thought processes, in short time and with little energy, in matters both simple (memorizing indefinitely many telephone numbers) and complex (identifying, from trillions global communications, social networks whose members may be unaware they are part of the network).
However underdeveloped now, I see no principled reason why machines operating independently of direct human control cannot learn from people's—or their own—fallibilities, and so evolve, create new forms of art and architecture, excel in sports (some novel combination of Deep Blue and Oscar Psitorius), invent new medicines, spot talent and exploit educational opportunities, provide quality assurance, or even build and use weapons that destroy people but not other machines.
But if the current focus in artificial intelligence and neuroscience persists, which is to reliably identify patterns of connection and wiring as a function of past connections and forward probabilities, then I don't think machines will ever be able to capture (imitate) critically creative human thought processes, including novel hypothesis formation in science or even ordinary language production.
Newton's laws of motion or Einstein's insights into relativity required imagining ideal worlds without precedent in any past or plausible future experience, such as moving in a world without friction or chasing a beam of light through a vacuum. Such thoughts require levels of abstraction and idealization that disregard, rather than assimilate, as much information as possible to begin with.
Increasingly sophisticated and efficient patterns of input and output, using supercomputers accessing massive data sets and constantly refined by Bayesian probabilities or other statistics based on degrees of belief in states of nature, may well produce ever better sentences and translations, or pleasing musical melodies and novel techno variations. In this way, machines may come to approximate, through a sort of reverse engineering, what human children or experts effortlessly do when they begin with fairly well-articulated internal structures in order to draw in and interpret relevant input from an otherwise impossibly noisy world. Humans know from the outset what they are looking for through the noise: in a sense they are there before they start; computing machines can never be sure they are there.
Can machines that operate independently of direct human control consistently interact with humans in ways such that humans believe themselves to be always interacting with other humans and not machines? Machines can come vanishingly close in many areas, and surpass mightily in others; but just as even the most highly skilled con artist always has some probability—however small—of being caught in deception, whereas the honest person never deceives and so can never be caught, so the associationist-connectionist machine that operates on stochastic rather than structure-dependent principles may never quite get the sense or sensibility of it all.
In principle, structurally richer machines, with internal architecture—beyond "read," "write" and "address"—can be built (indeed, earlier advocates of AI added logical syntax), interact with some degree of fallibility (for if no error, then no learning is possible), and culturally evolve. But the current emphasis in much AI and neuroscience, which is to replace posits of abstract psychological structures with physically palpable neural networks and the like, seems to be going in precisely the wrong direction.
Rather, the cognitive structures that psychologists posit (provided they are descriptively adequate, plausibly explanatory, and empirically tested against alternatives and the null-hypothesis) should be the point of departure—what it is that neuroscience and machine models of the mind should be looking for. If we then discover that different abstract structures operate through the same physical substrate, or that similar structures operate through different substrates, then we have a novel and interesting problem that may lead to a revision in our conception of both structure and substrate The fact that such simple and basic matters as these are puzzling (or even excluded, a priori, from the puzzle) tells us how very primitive still is the science of mind, whether human brain or machine.
To think can mean to reason logically, which certainly some machines do, albeit by following algorithms we program into them. Or it can mean "to have a mind" by which we mean it can experience itself as a subject endowed with consciousness, qualia, experiences, intentions, beliefs, emotions, memories. When we ask, could a machine think, we are really asking whether there can be a completely naturalistic account of what a mind is.
I am a naturalist, so I believe the answer must be yes.
Certainly, we are not there yet. Whatever the brain is doing to generate a mind, I doubt it is only running pre-specified algorithms, or doing anything like what present-day computers do. It seems likely we have yet to discover key principles by which a human brain works. I suspect that how and why we think cannot be understood apart from our being alive, so before we understand what a mind is we will have to understand more deeply what a living thing is-in physical terms.
The construction of an artificial mind then probably has to wait until we understand better, in physical terms, what a mind is.
This understanding will have to address what Chalmers calls the hard problem of consciousness: how to account for the presence of qualia in the physical world. We have reason to believe our sensations of the color red are associated with certain physical processes in our brains, but we are stumped because it seems impossible to explain in physical terms why or how those processes give rise to qualia.
A key step towards solving this hard problem is to situate our description of physics in a relational language. As set out by Leibniz, the patron saint of relationalism, the properties of elementary particles have to do with relationships with other particles. This has been a very successful idea, it is well realized by general relativity and quantum theory, so let's adopt it.
The second step is to recognize that events or particles may have properties that are not relational, which are not described by giving a complete history of the relationships they enjoy. Let us call these internal properties.
If an event or process has internal properties, you cannot learn about them by interacting with it or measuring it. If there are internal properties, they are not describable in terms of position, motion, charges or forces, ie in the vocabulary physics uses to talk of relational properties.
You might, however, know about a process's internal properties by being that process.
So let us hypothesize that qualia are internal properties of some brain processes. When observed from the outside, those brain processes can be described in terms of motions, potentials, masses, charges. But they have additional internal properties, which sometimes include qualia.
Qualia must be extreme cases of being purely internal. More complex aspects of mind may turn out to be combinations of relational and internal properties. We know that thoughts and intentions are able to influence the future.
There is much hard, scientific work to do to develop such a naturalistic account of mind, which is non-dualist and not deflationary, in that it doesn't reduce mental properties completely to the standard physical properties or visa versa. We may want to avoid naive pan-psychism according to which rocks and wind have qualia. At the same time we want to remember that if we don't know what its like to be a bat, we also don't know really what a rock is, in the sense that we may only know a subset of its properties-those that are relational.
One troubling aspect of mind from a naturalistic perspective is the impression we have that we sometimes think novel thoughts and have novel experiences that have never been thought or experienced before in the history of the world.
There is little that would make sense about the human world of culture and imagination without allowance for the genuinely novel. A century ago this website did not exist and likely could not have been imagined. Yet it exists and as naturalists we must have a conception of nature that includes it. This must allow novel kinds of things to come to exist in nature.
We are hamstrung by the conviction that nothing truly new can happen in nature because everything is really elementary particles moving in space according to unchanging laws. Without deviating an inch from rigorous naturalism, however, we can begin to imagine how our understanding of nature can be deepened to allow for the truly novel to occur.
First, in quantum physics we admit the possibility of novel properties arising that are shared among several particles in entangled states. In the lab we can make entangled states of complex systems that are unlikely to have natural precedents. Hence we can and do create physical systems with novel properties.
(So, by the way, does nature, when natural selection produces novel proteins, which catalyze novel reactions.)
Second, Leibniz's principle of the identity of the indiscernible implies that there can be no two distinct events with exactly the same properties. This means that the fundamental events cannot be subject to laws that are both deterministic and simple. For if two events have precisely the same past, their futures must differ. This presumes a physics which can distinguish the future from the past.
Note that quantum physics is inherently nondeterministic.
Does this imply quantum physics will play a role in a future naturalistic account of mind? It is too soon to tell, and the first efforts in this direction are not convincing. But what we learn is that a naturalistic account of mind will require deepening our concept of the natural. We can think novel thoughts by which we can alter the future. Novelty must then be intrinsic to how we understand nature, if minds are to be natural. Therefore, to understand how a machine could have a mind we must deepen our concept of nature.
The first time I had occasion to think about what thinking machines might do to human existence was at a talk decades ago by a computer scientist at a Yale psychology department colloquium. The speaker's topic was: "What will it mean to humans' conception of themselves, and to their well-being, if computers are ever able to do everything better than humans can do: beat the greatest chess player, compose better symphonies than humans?"
The speaker then said, "I want to make two things clear at the outset. First, I don't know whether machines will ever be able to do those things. Second, I'm the only person in the room with the right to an opinion about that question." The latter statement was met with some gasps and nervous laughter.
Decades later, it's no longer a matter of opinion that computers will be able to do many of the astonishing things the speaker mentioned. And I'm worried that the answer to his question about what this will mean to us is that we're going to feel utterly sidelined and demoralized by machines. I was sorry that Big Blue beat Garry Kasparov at chess. I was depressed for a moment when its successor beat all of its human Jeopardy competitors. And of course we know that machines can already compose works that beat the socks off John Cage for interest and listenability!
We really have to worry that there will be a devastating morale problem for us when any work we might do can be done better by machines. What does it mean to airplane pilots that a machine can do their job better than they can? How long will it be before that occupation, like hundreds of others already, is made literally obsolete by machines? What will it mean to accountants, financial planners and lawyers when machines can carry out, at the very least, nearly all of their bread-and-butter tasks more effectively and infinitely faster than they can? To physicians, physicists, and psychotherapists?
What will it mean when there is simply no meaningful work for any of us to do? When unsupervised machines plant and harvest crops. When machines can design better machines than any human could even think of. Or be a more entertaining conversationalist than even the cleverest of your friends.
Steve Jobs said, "It's not the customers' job to know what they want." Computers may be able to boast that it's not the job of humans to know what they want.
Like you, I love to read, listen to music, and see movies and plays, experience nature. But I also love to work—to feel that what I do is fascinating at least to me, and might possibly improve the lives of some other people. What would it mean to people like you and me if our work were simply pointless and there were only the other enjoyable things to do?
We already know what machine-induced obsolescence has meant to some of the world's peoples. It's no longer necessary for anyone to make their own bows and arrows and hunt animals for any purpose other than recreation. Or plant, cultivate and harvest corn and beans. Some cultures built around such activities have collapsed and utterly lost their meaning to the people who were shaped by them. Think, for example, of some Southwestern Indian tribes and of rural whites in South Dakota, Alabama and New Mexico, with their ennui, lassitude and drug addictions. We have to wonder whether the mass of people in the world can face with equanimity the possibility of there being absolutely nothing to do other than entertain oneself.
Which isn't to say that cultures couldn't evolve in some way as to make the complete absence of work acceptable—even highly satisfying. There are cultures where there has been little to do in the way of work for eons, and people seem to have gotten along just fine. In some South Pacific cultures people could get by with little other than waiting for a coconut to drop or wading into a lagoon to catch a fish. In some West African cultures, men didn't do anything you would be likely to classify as work except for a couple of weeks a year when they were essential for the planting of crops. And then there were the idle rich of, for example, early 20th century England, with its endless rounds of card playing, the putting on of different costumes for breakfast, lunch and dinner, and serial infidelities with really rather attractive people. Judging from PBS fare, that was pretty enjoyable.
So maybe the most optimistic possibility is that we're headed toward evolving cultures that will enable us to enjoy perpetual entertainment with absolutely no meaningful, productive work to do. However repellent that may seem to us, we have to imagine, hope even, that it may seem an absolutely delightful existence to our great great grandchildren, who will pity us for our cramped and boring lives. Some would say the vanguard is already here. Portland has been described as the place where young people go to retire.
In general I am happy to have them around and to have them improve. There are of course some dangers from such machines making harmful decisions, but probably no more dangers than with humans making such decisions.
Having such machines will not answer the questions about the world that are most important to me and many others. What constitutes the dark matter of the universe? Is supersymmetry really a symmetry of nature that provides a foundation for and extends the highly successful Standard Model of particle physics we have? These and similar questions can only be answered by experimental data.
No amount of thought will provide such answers. More precisely, perhaps given all the information we have about nature some machine actually would come up with the right answers. Indeed, perhaps some physicists already have come up with the answers. But the true role of data is to confirm which answers are the correct ones. If some physicist, or some machine, figures it out they have no way to convince anyone else they have the actual answer. Laboratory dark matter detectors, or the CERN Large Hadron Collider, or possibly a future Chinese collider, might get the needed data, but not a thinking machine.
Smart people often manage to avoid the cognitive errors that bedevil less well-endowed minds. But there are some kinds of foolishness that seem only to afflict the very intelligent. Worrying about the dangers of unfriendly AI is a prime example. A preoccupation with the risks of superintelligent machines is the smart person’s Kool Aid.
This is not to say that superintelligent machines pose no danger to humanity. It is simply that there are many other more pressing and more probable risks facing us this century. People who worry about unfriendly AI tend to argue that the other risks are already the subject of much discussion, and that even if the probability of being wiped out by superintelligent machines is very low, it is surely wise to allocate some brainpower to preventing such an event, given the existential nature of the threat.
Not coincidentally, the problem with this argument was first identified by some of its most vocal proponents. It involves a fallacy that has been termed "Pascal’s mugging," by analogy with Pascal’s famous wager. A mugger approaches Pascal and proposes a deal: in exchange for the philosopher’s wallet, the mugger will give him back double the amount of money the following day. Pascal demurs. The mugger then offers progressively greater rewards, pointing out that for any low probability of being able to pay back a large amount of money (or pure utility) there exists a finite amount that makes it rational to take the bet—and a rational person must surely admit there is at least some small chance that such a deal is possible. Finally convinced, Pascal gives the mugger his wallet.
This thought experiment exposes a weakness in classical decision theory. If we simply calculate utilities in the classical manner, it seems there is no way round the problem; a rational Pascal must hand over his wallet. By analogy, even if there is there is only a small chance of unfriendly AI, or a small chance of preventing it, it can be rational to invest at least some resources in tackling this threat.
It is easy to make the sums come out right, especially if you invent billions of imaginary future people (perhaps existing only in software—a minor detail) who live for billions of years, and are capable of far greater levels of happiness than the pathetic flesh and blood humans alive today. When such vast amounts of utility are at stake, who could begrudge spending a few million dollars to safeguard it, even when the chances of success are tiny?
Why do some otherwise very smart people fall for this sleight of hand? I think it is because it panders to their narcissism. To regard oneself as one of a select few far-sighted thinkers who might turn out to be the saviors of mankind must be very rewarding. But the argument also has a very material benefit: it provides some of those who advance it with a lucrative income stream. For in the past few years they have managed to convince some very wealthy benefactors not only that the risk of unfriendly AI is real, but also that they are the people best placed to mitigate it. The result is a clutch of new organizations that divert philanthropy away from more deserving causes. It is worth noting, for example, that Give Well—a non-profit that evaluates the cost-effectiveness of organizations that rely on donations—refuses to endorse any of these self-proclaimed guardians of the galaxy.
But whenever an argument becomes fashionable, it is always worth asking the vital question—Cui bono? Who benefits, materially speaking, from the growing credence in this line of thinking? One need not be particularly skeptical to discern the economic interests at stake. In other words, beware not so much of machines that think, but of their self-appointed masters.
Back in 1932, Walter Cannon published a landmark work on human physiology—The Wisdom of the Body. He described the tight regulation of many of our body's parameters, such as hydration, blood glucose, sodium, and temperature. This concept of homeostasis, or auto-regulation, is an extraordinary means by which we stay healthy. Indeed, it represents a machine like quality, that our body can so finely tune such important functions.
Although it has taken the better part of a century, we are now ready for the next version—Cannon 2.0. While some have expressed marked trepidation about the rise of artificial intelligence, this capability will have an extraordinary impact on preserving our health. We are quickly moving to "all-cyborg" status, surgically connected to our smartphones. While they have been called prosthetic brains, "smart" phones today are just a nascent precursor to where we are headed. Very soon the wearable sensors, whether they are Band-Aids, watches, or necklaces, will be accurately measuring our essential medical metrics. Not just one-off assessments, but continuous, real-time streaming. Obtaining data that we never had before.
Beyond our body's vital signs (blood pressure, heart rhythm, oxygen concentration in the blood, temperature, breathing rate), there will be quantitation of mood and stress via tone and inflection of voice, galvanic skin response and heart rate variability, facial expression recognition, and tracking of our movement and communication. Throw in the analytes from our breath, sweat, tears, and excrements into the mix. Yet another layer of information captured will include our environmental exposures, ranging from air quality to pesticides in foods.
None of us—or our bodies—are smart enough to be able to integrate and process all of this information about ourselves. That's the job for deep learning, with algorithms that provide feedback loops to us via our mobile devices. What we're talking about does not exist today. It hasn't yet been developed, but it will. And it will be providing what heretofore was unobtainable, multi-scale information about ourselves and—for the first time—the real ability to pre-empt disease.
Almost any medical condition with an acute episode—like an asthma attack, seizure, autoimmune attack, stroke, heart attack—will be potentially predictable in the future with artificial intelligence and the Internet of all medical things. There's already a wristband that can predict when a seizure is imminent, and that can be seen as a rudimentary, first step. In the not so distant future, you'll be getting a text message or voice notification that tells you precisely what you need to prevent a serious medical problem. When that time comes, those who fear AI may suddenly embrace it. When we can put together big data for an individual with the requisite contextual computing and analytics, we've got a recipe for machine-mediated medical wisdom.
The prospect of a world inhabited by robust AIs terrifies me. The prospect of a world without robust AIs also terrifies me. Decades of technological innovation have created a world system so complex and fast-moving that it is quickly becoming beyond human capacity to comprehend, much less manage. If we are to avoid civilizational catastrophe, we need more than clever new tools—we need allies and agents.
So-called "narrow" AI systems have been around for decades. At once ubiquitous and invisible, narrow AIs make art, run industrial systems, fly commercial jets, control rush hour traffic, tell us what to watch and buy, determine if we get a job interview, and play matchmaker for the lovelorn. Add in the relentless advance of processing, sensor and algorithmic technologies, and it is clear that today's narrow AIs are tracing a trajectory towards a world of robust AI. Long before artificial super-intelligences arrive, evolving AIs will be pressed into performing once-unthinkable tasks from firing weapons to formulating policy.
Meanwhile, today's primitive AIs tell us much about future human-machine interaction. Narrow AIs may lack the intelligence of a grasshopper, but that hasn't stopped us from holding heartfelt conversations with them and asking how they feel. It is in our nature to infer sentience at the slightest hint that life might be present. Just as our ancestors once populated their world with elves, trolls and angels, we eagerly seek companions in cyberspace. This is one more impetus driving the creation of robust AIs—we want someone to talk to. The consequence could well that the first non-human intelligence we encounter won't be little green men or wise dolphins, but creatures of our own invention.
We of course will attribute feelings and rights to AIs—and eventually they will demand it. In Descartes time, animals were considered mere machines—a crying dog was considered no different than a gear whining for want of oil. Late last year, an Argentine court granted rights to an orangutan as a "non-human person." Long before robust AIs arrive, people will extend the same empathy to digital beings and give them legal standing.
The rapid advance of AIs also is changing our understanding of what constitutes intelligence. Our interactions with narrow AIs will cause us to realize that intelligence is a continuum and not a threshold. Earlier this decade Japanese researchers demonstrated that slime mold could thread a maze to reach a tasty bit of food. Last year a scientist in Illinois demonstrated that under just the right conditions, a drop of oil could negotiate a maze in an astonishingly lifelike way to reach a bit of acidic gel. As AIs insinuate themselves ever deeper in our lives, we will recognize that modest digital entities as well as most of the natural world carry the spark of sentience. From there is it just a small step to speculate about what trees or rocks—or AIs—think.
In the end, the biggest question is not whether AI super-intelligences will eventually appear. Rather the question is what will be the place of humans in a world occupied by an exponentially growing population of autonomous machines. Bots on the Web already outnumber human users—the same will soon be true in the physical world as well.
Lord Dunsany once cautioned, "If we change too much, we may no longer fit into the scheme of things."
Considering machines that think is a nice step forward in the AI debate as it departs from our own human-based concerns, and accords machines otherness in a productive way. It causes us to consider the other entity's frame of reference. However, even more importantly this questioning suggests a large future possibility space for intelligence. There could be "classic" unenhanced humans, enhanced humans (with nootropics, wearables, brain-computer interfaces), neocortical simulations, uploaded mind files, corporations as digital abstractions, and many forms of generated AI: deep learning meshes, neural networks, machine learning clusters, blockchain-based distributed autonomous organizations, and empathic compassionate machines. We should consider the future world as one of multi-species intelligence.
What we call the human function of "thinking" could be quite different in the variety of possible future implementations of intelligence. The derivation of different species of machine intelligence will necessarily be different than that of humans. In humans, embodiment and emotion as a short-cut heuristic for the fight-or-flight response and beyond have been important elements influencing human thinking. Machines will not have the evolutionary biology legacy of being driven by resource acquisition, status garnering, mate selection, and group acceptance, at least in the same way. Therefore different species of native machine "thinking" could be quite different. Rather than asking if machines can think, it may be more productive to move from the frame of "thinking" that asks "who thinks how" to a world of "digital intelligences" with different backgrounds, modes of thinking, and existence, and different value systems and cultures.
Already not only are AI systems becoming more capable, but we are also starting to get a sense of the properties and features of native machine culture and the machine economy, and what the coexistence of human and machine systems might be like. Some examples of these parallel systems are in law and personal identity. In law, there are technologically-binding contracts and legally-binding contracts. They have different enforcement paradigms; inexorably executing parameters in the case of code ("code is law"), and discretionary compliance in the case of human-partied contracts. Code contracts are good in the sense that they cannot be breached, but on the other hand, will execute monolithically even if later conditions have changed.
Another example is personal identity. The technological construct of identity and the social construct of identity are different and have different implied social contracts. The social construct of identity includes the property of imperfect human memory that allows the possibility of forgiving and forgetting, and redemption and reinvention. Machine memory, however, is perfect and can act as a continuous witnessing agent, never forgiving or forgetting, and always able to re-presence even the smallest detail at any future moment. Technology itself is dual-use in that it can be deployed for "good" or "evil." Perfect machine memory only becomes tyrannizing when reimported to static human societal systems, but it need not be restrictive. Having this new "fourth-person perspective" could be a boon for human self-monitoring and mental performance enhancement.
These examples show that machine culture, values, operation, and modes of existence are already different, and this emphasizes the need for ways to interact that facilitate and extend the existence of both parties. The potential future world of intelligence multiplicity means accommodating plurality and building trust. Blockchain technology, a decentralized, distributed, global, permanent, code-based ledger of interaction transactions and smart contracts is one example of a trust-building system. The system can be used whether between human parties or inter-species parties, exactly because it is not necessary to know, trust, or understand the other entity, just the code (the language of machines).
Over time trust can grow though reputation. Blockchain technology could be used to enforce friendly AI and mutually-beneficial inter-species interaction. This is because it is possible that in the future, important transactions (like identity authentication and resource transfer) would be conducted on smart networks that require confirmation by independent consensus mechanisms such that only bonafide transactions by entities in good reputational standing are executed. While perhaps not a full answer to the problem of enforcing friendly AI, decentralized smart networks like blockchains are a system of checks and balances that starts to provide a more robust solution to situations of future uncertainty.
Trust-building models for inter-species digital intelligence interaction could include both game-theoretic checks-and-balances systems like blockchains, and also at the higher level, frameworks that put entities on the same plane of shared objectives. This is of higher order than smart contracts and treaties that attempt to enforce morality. A mindset shift is required. The problem frame of machine and human intelligence should not be one that characterizes relations as friendly or unfriendly, but rather one that treats all entities equally, putting them on the same grounds and value system for the most important shared parameters, like growth. What is most important about thinking for humans and machines is that thinking leads to ideation, progress, and growth.
What we want is the ability to experience, grow, and contribute more, for both humans and machines, and the two in symbiosis and synthesis. This can be conceived as all entities existing on a spectrum of capacity for individuation (the ability to grow and realize their full and expanding potential). Productive interaction between intelligent species could be fostered by being aligned in the common framework of a capacity spectrum that facilitates their objective of growth and maybe mutual growth.
What we should think about thinking machines is that we want to be in greater interaction with them, both quantitatively or rationally, and qualitatively in sense of our extending our internal experience of ourselves and reality, moving forward together in the vast future possibility space of intelligence.
The coming shock isn't from machines that think, but machines that use AI to augment our perception.
For millions of years, other people saw us using the same machinery we used to see them. We have pretty much the same eyes as our rivals, and pretty much the same mirror neurons. Within any given culture we have pretty much the same signaling mechanisms and value systems. So when we try to deceive, or to detect deception in others, we're on a level playing field. I can wear a big penis gourd to look more manly, and you can paint your chest with white and ochre mud stripes to look more scary. Civilisation made the games more sophisticated; I signal class by wearing a tailored jacket with four cuff buttons, while you signal wealth by wearing a big watch. But our games would have been perfectly comprehensible to our Neolithic ancestors.
What's changing as computers become embedded invisibly everywhere is that we all now leave a digital trail that can be analysed by AI systems. The Cambridge psychologist Michael Kosinski has shown that a person's race, intelligence and sexual orientation can be deduced fairly quickly from their behaviour on social networks: on average it takes only four Facebook "likes" to tell whether you're straight or gay. So whereas in the past gay men could choose whether or not to wear their Out and Proud t-shirt, you just have no idea what you're wearing any more. And as AI gets better, you're mostly wearing your true colours.
It's as if we all evolved in a forest where all the animals could only see in black and white, and now a new predator comes along who can see in colour. All of a sudden, half of your camouflage doesn't work, and you don't know which half!
At present this is great if you're an advertiser, as you can figure out how to waste less money. It isn't yet available on the street. But the police are working on it; which cop wouldn't want a Google glass app that will highlight passersby with a history of violence, coupled perhaps with w-band radar to see which of them is carrying a weapon?
The next question is whether only the authorities have enhanced cognition systems, or whether they're available to all. In twenty years' time, will we all be wearing augmented reality goggles? What will the power relationships be? If a policeman can see my arrest record when he looks at me, can I see whether he's been the subject of brutality complaints? If a politician can see whether I'm a party supporter or an independent, can I see his voting record on the three issues I care about? Never mind the right to bear arms, what about the right to wear Google glass?
Perception and cognition will no longer be conducted inside a single person's head. Just as we now use Google and the Internet as memory prostheses, we'll be using AI systems that draw on millions of machines and sensors as perceptual prostheses.
But can we trust them? Deception will no longer just be something that individual humans do to each other. Governments will influence our perceptions via the tools we use for cognitive enhancement, just as China currently censors search results; while in the West, advertisers will buy and sell what we get to see. How else will the system be paid for?
The ability to tell and comprehend stories is a main distinguishing feature of the human mind. It's therefore understandable that in pursuit of a more complete computational account of human intelligence, researchers are trying to teach computers how to tell and understand stories. But should we root for their success?
Creative writing manuals always stress that writing good stories means reading them first—lots of them. Aspiring writers are told to immerse themselves in great stories to gradually develop a deep, not necessarily conscious, sense for how they work. People learn to tell stories by learning the old ways and then—if they have some imagination—making those old ways seem new. It's not hard to envision computers mastering storytelling by a similar process of immersion, assimilation, and recombination—just much, much faster.
To date, practical experiments in computer-generated storytelling aren't that impressive. They are bumbling, boring, soulless. But the human capacity to make and enjoy art evolved from crude beginnings over eons, and the machines will evolve as well—just much, much faster.
Someday robots may take over the world. The dystopian possibilities don't trouble me like the probable rise of art-making machines. Art is, arguably, what most distinguishes humans from the rest of creation. It is the thing that makes us proudest of ourselves. For all of the nastiness of human history, at least we wrote some really good plays and songs and carved some good sculptures. If human beings are no longer needed to make art, then what the hell would we be for?
But why should I be pessimistic? Why would a world with more great art be a worse place to live? Maybe it wouldn't be. But the thought still makes me glum. While I think of myself as a hard-bitten materialist, I must hold out some renegade hope for a dualism of body and spirit. I must hope that cleverly evolving algorithms and brute processing power are not enough—that imaginative art will always be mysterious and magical, or at least so weirdly complex that it can't be mechanically replicated.
Of course machines can out-calculate and out-crunch us. And soon they will all be acing their Turing tests. But who cares. Let them do our grunt work. Let them hang out and chat. But when machines can out-paint or out-compose us—when their stories are more gripping and poignant than ours—there will be no denying that we are, ourselves, just thought machines and art machines, and outdated and inferior models at that.
In his novel Gravity's Rainbow, Thomas Pynchon identifies the confusion about the subject and object of enquiries: "if they can get you asking the wrong questions, they don't have to worry about answers." Thinking about machines that think poses more questions about human beings than about the machines or Artificial Intelligence (AI).
Technology enables machines providing access to essential resources, power, speed and communications that make life and improved living standards possible. Machines execute tasks, specified and programmed by humans. Techno-optimists believe that progress is near a singularity, the hypothetical moment when machines will reach the point of a greater-than-human intelligence.
It is a system of belief and faith. Just like the totems and magic used by our ancestors or organised religion, science and technology deal with uncertainty and fear of the unknown. It allows limited control over our immediate environment. It increases material prosperity and comfort. It is a striving for perfectibility. Technology asserts human superiority in the pantheon of creation.
But science is a long way from unlocking the secrets in nature's infinite book. Knowledge of the origins of the universe, life and fundamentals of matter remain limited. Biologist E.O. Wilson noted that if natural history were a library of books, we have not even finished the first chapter of the first book. Human knowledge is always incomplete, sometimes inaccurate and frequently the cause of not the solution to problems.
First, use of science and technology is often ineffective, with unintended consequences.
In Australia, introduced rabbits spread rapidly becoming a pest changing Australia's ecosystems destroying endemic species. In the 1950s, scientists introduced the Myxoma virus, severely reducing the rabbit population. When genetic resistance allowed the population to recover, Calicivirus, which causes rabbit haemorrhagic disease, was introduced as a new control measure. Increasing immunity rapidly reduced effectiveness. In 1935, the Cane Toad was introduced to control insect pests of sugar cane. Unsuccessful in controlling the insects, the amphibian became an invasive species devastating indigenous wildlife.
Life saving antibiotics has increased drug resistant infections. A 2014 British study found that it may cause 10 million deaths a year worldwide by 2050. The potential cost is US$100 trillion, reducing GDP by 3.5%
Economic models have repeatedly failed because of incorrect assumptions, flawed causal relationships, inputs that are more noise than data and unanticipated human factors. Forecasts have proved inaccurate. Models consistently underestimate risks and exposures, resulting in costly financial crisis.
Second, consequences of technology, especially over longer terms, are frequently not understood at inception.
The ability to harness fossil fuels to provide energy was the foundation of the industrial revolution. The long term impact of CO2 emissions on the environment now threatens the survival of the species. Theoretical physics and mathematics made possible nuclear and thermo-nuclear devices, capable of extinguishing all life on the planet.
Third, technology creates moral, ethical, political, economic and social concerns which are frequently ignored.
Nuclear, biological and chemical weapons of mass destruction or remotely controlled drones rely on technical advances. The question of whether such technology should be developed or used at all remains. Easy access to the requisite knowledge, problems of proliferation and difficulty of controlling dual use (civilian and defense) technology complicates the matter.
Robots and AI may improve productivity. While a few creators might capture large rewards, the effect on economic activity is limited. Given consumption constitutes over 60 % of activity in developed economies, decreasing general employment and lower income levels harms the wider economy. In 1955, showing off a new automatically operated plant, a company executive asked UAW head Walter Reuther: "How are you going to collect union dues from those guys [the robots]?" Reuther countered: "And how are you going to get them to buy Fords?
When it comes to questions of technology, the human race is rarely logical. We frequently do not accept that something cannot or should not be done. Progress is accepted without question or understanding of what and why we need to know. We do not know when and how our creations should be used or its limits. We frequently do not know the real or full consequences. Doubters are dismissed as Luddites.
Technology and its manifestations such as machines or AI is an illusion, which appeals to human arrogance, ambition and vanity. It multiplies confusion in poet T.S. Elliot's "wilderness of mirrors."
The human species is simply too small, insignificant and inadequate to fully succeed in anything that we think we can do. Thinking about machines that think merely confirms that inconvenient truth.
If asked to rank humanity's problems by severity, I would give the silver medal to the need to spend so much time doing things that give us no fulfillment—work, in a word. I consider that the ultimate goal of artificial intelligence is to hand off this burden, to robots that have enough common sense to perform those tasks with minimal supervision.
But some AI researchers have altogether loftier aspirations for future machines: they foresee computer functionality that vastly exceeds our own in every sphere of cognition. Such machines would not only do things that people prefer not to; they would also discover how to do things that no one can yet do. This process can, in principle, iterate—the more such machines can do, the more they can discover.
What's not to like about that? Why do I not view it as a superior research goal than machines with common sense (which I'll call "minions")?
First, there is the well-publicised concern that such machines might run amok—especially if the growth of a machine's skill set (its "self-improvement") were not iterative but recursive. What researchers mean by this is that enhancements might be not only to the database of things a machine can do, but to its algorithms for deciding what to do. It has been suggested, firstly, that this recursive self-improvement might be exponential (or faster), creating functionality that we cannot remotely comprehend before we can stop the process. So far so majestic—if it weren't that the trajectory of improvement would itself be out of our control, such that these superintelligent machines might gravitate to "goals" (metrics by which they decide what to do) that we dislike. Much work has been done on ways to avoid this "goal creep", and to create a reliably, permanently "friendly," recursively self-improving system, but with precious little progress.
My reason for believing that recursive self-improvement is not the right ultimate goal for AI research is actually not the risk of unfriendly AI, though: rather, it is that I quite strongly suspect that recursive self-improvement is mathematically impossible. In analogy with the so-called "halting problem" concerning determining whether any program terminates, I suspect that there is a yet-to-be-discovered measure of complexity by which no program can ever write another program (including a version of itself) that is an improvement.
The program written may be constrained to be, in a precisely quantifiable sense, simpler than the program that does the writing. It's true that programs can draw on the outside world for information on how to improve themselves—but I claim (a) that that really only delivers far-less-scary iterative self-improvement rather than recursive, and (b) that anyway it will be inherently self-limiting, since once these machines become as smart as humanity they won't have any new information to learn. This argument isn't anywhere near iron-clad enough to give true reassurance, I know, and I bemoan the fact that (to my knowledge) no one is really working to seek such a measure of depth or to prove that none can exist—but it's a start.
But in contrast, I absolutely am worried about the other reason why I stick to the creation of minions as AI's natural goal. It is that any creative machine—whether technologically, artistically, whatever—undermines the distinction between man and machine. Humanity has massive uncertainty already regarding what rights various non-human species have. Since objective moral judgements build on agreed norms, which themselves arise from inspection of what we would want for ourselves, it seems impossible even in principle to form such judgements concerning entities that differ far more from us than animals do from each other, so I say we should not put ourselves in the position of needing to try. For illustration, consider the right to reproduce despite resource limitations. Economic incentive-based compromise solutions seem to work adequately. But how can we identify such compromises for "species" with virtually unlimited reproductive potential?
I contend that the possession of common sense does not engender these problems. I define common sense, for present purposes, as the ability to process highly incomplete information so as to identify a reasonably close-to-optimal method for achieving a specified goal, chosen from a parametrically pre-specified set of alternative methods. This explicitly excludes the option of "thinking"—of seeking new methods, outside the pre-specified set, that might outperform anything within the set.
Thus, again for illustration, if the goal is one that should ideally be achieved quickly, and can be achieved faster by many machines than by one, the machine will not explore the option of first building a copy of itself unless that option is pre-specified as admissible, however well it may "know" that doing so would be a good idea. Since admissibility is specified by inclusion rather than exclusion, the risk of "method creep" can (I claim) be safely eliminated. Vitally, it is possible to prevent recursive self-improvement (if it turns out to be possible after all!) entirely.
The availability of an open-ended vista of admissible ways to achieve one's goals constitutes a good operational definition of "awareness" of those goals. Awareness implies the ability to reflect on the goal and on one's options for achieving it, which amounts to considering whether there are options one hadn't thought of.
I could end with a simple "So let's not create aware machines"—but any possible technology that anyone thinks is desirable will eventually be developed, so it's not that simple. What I say instead is, let's think hard now about the rights of thinking machines, so that well before recursive self-improvement arrives we can test our conclusions in the real world with machines that are only slightly aware of their goals. If, as I predict, we thereby discover that our best effort at such ethics fails utterly even at that early stage, maybe such work will cease.
In 1922 the mathematician Lewis Fry Richardson had imagined a large hall full of "computers", people who, one hand calculation at a time, would advance numerical weather prediction. Less than a hundred years later, machines have improved the productivity of that particular task by up to fifteen orders of magnitude, with the ability to process almost a million billion similar calculations per second.
Consider the growth in heavy labor productivity by comparison. In 2014 the world used about 500 Exajoules—a billion, billion joules—of primary energy, to produce electricity, fuel manufacturing, transport and heat. Even if we assumed all of that energy went into carrying out physical tasks in aid of the roughly 3 billion members of the global labor force (and it did not), assuming an average adult diet of 2,000 Calories per capita per day, would imply roughly 50 "energy laborers" for every human. More stringent assumptions would still lead to at most an increase of a few orders of magnitude in effective productivity of manual labor.
We have been wildly successful at accelerating our ability to think and process information, more so than any other human activity. The promise of artificial intelligence is to deliver another leap in increasing the productivity of specific cognitive functions: ones where the sophistication of the task is also orders of magnitude higher than previously possible.
Keynes would have probably argued that such an increase should ultimately lead to a fully employed society with greater free time and a higher quality of life for all. The skeptic might be forgiven for considering this a case of hope of experience. While there is no question that specific individuals will benefit enormously from delegating tasks to machines, the promise of greater idleness from automation has yet to be realized, as any modern employee—virtually handcuffed to a portable device—can attest.
So, if we are going to work more, deeper, and with greater effectiveness thanks to thinking machines, choosing wisely what they are going to be "thinking" about is particularly important. Indeed, it would be a shame to develop all this intelligence to then spend it on thinking really hard about things that do not matter. And, as ever in science, selecting problems worth solving is a harder task than figuring out how to solve them.
One area where the convergence of need, urgency, and opportunity is great is in the monitoring and management of our planetary resources. Despite the dramatic increase in cognitive and labor productivity, we have not fundamentally changed our relationship to Earth: we are still stripping it of its resources to manufacture goods that turn to waste relatively quickly, with essentially zero end-of-life value to us. A linear economy on a finite planet, with seven billion people aspiring to become consumers—our relationship to the planet is arguably more productive, but not much more intelligent than it was a hundred years ago.
Understanding what the planet is doing in response, and managing our behavior accordingly, is a complicated problem, hindered by colossal amounts of imperfect information. From climate change, to water availability, to the management of ocean resources, to the interactions between ecosystems and working landscapes, our computational approaches are often inadequate to conduct the exploratory analyses required to understand what is happening, to process the exponentially growing amount of data about the world we inhabit, and to generate and test theories of how we might do things differently.
We have almost 7 billion thinking machines on this planet already, but for the most part they don't seem to be terribly concerned with how sustainable their life on this planet actually is. Very few of those people have the ability to see the whole picture in ways that make sense to them, and those that do are often limited in their ability to respond. Adding cognitive capacity to figure out how we fundamentally alter our relationship with the planet is a problem worth thinking about.
If we believe that thinking and consciousness is the result of patterns of brains cells and their components, then our thoughts, emotions, and memories could be replicated in moving assemblies of bicycle parts. Of course, the bicycle brains would have to be very big to represent the complexity of our minds. In principle, our minds could be hypostatized in the patterns of slender tree limbs moving in the wind or in the movements of termites.
What would it mean for a "bicycle brain," or any machine, to think and know something? There are many kinds of knowledge the machine-being could have. This makes discussions of thinking things a challenge. For example, knowledge may be factual or propositional: A being may know that the First Franco-Dahomean War was a conflict between France and the African Kingdom of Dahomey under King Béhanzin.
Another category of knowledge is procedural, knowing how to accomplish a task such as playing the game of Go, cooking a soufflé, making love, performing a rotary throw in Aikdo, shooting a 15th-century Wallarmbrust crossbow, or simulating the Miller–Urey experiment to explore the origins of life. However, for us at least, reading about accurately shooting a Wallarmbrust crossbow is not the same as actually being able to accurately shoot the crossbow. This second type of procedural knowing implies actually being able to perform the act. Yet another kind of knowledge deals with direct experience. This is the kind of knowledge referred to when someone says, "I know love" or "I know fear."
Also, consider that human-like interaction is quite important for any machine that we would wish to say has human-like intelligence and thinking. A smart machine is less interesting if its intelligence lies trapped in an unresponsive program, sequestered in a kind of isolated limbo. As we provide our computers with increasingly advanced sensory peripherals and larger databases, it is likely we will gradually come to think of these entities as intelligent. Certainly within this century, some computers will respond in such a way that anyone interacting them will consider them conscious and deeply thoughtful.
The entities will exhibit emotions. But more importantly, over time, we will merge with these creatures. We will become one. We will share our thoughts and memories with these devices. Our organs may fail and turn to dust, but our Elysian essences will survive. Computers, or computer-human hybrids, will surpass humans in every area, from art to mathematics to music to sheer intellect.
In the future, when our minds merge with artificial agents and also integrate various electronic prostheses, for each of our own real lives, we will create multiple simulated lives. Your day job is a computer programmer for a big company. However, after work, you'll be a knight with shining armor in the Middle Ages, attending lavish banquets, and smiling at wandering minstrels. The next night, you'll be in the Renaissance, living in your home on the southern coast of the Sorrentine Peninsula, enjoying a dinner of plover and pigeon. Perhaps, when we become hybrid entities with our machines, we will simulate new realities to rerun historical events with slight changes to observe the results, produce great artworks akin to ballets or plays, solve the problem of the Riemann Hypothesis or baryon asymmetry, predict the future, and escape the present, so as to call all of space-time our home.
Of course, the ways in which a machine thinks could be quite different from the ways in which we think. After all, it is well known that machines don't see the same way we do, and image-recognition algorithms called deep neural networks sometimes declare, with near 100% certainty, that images of random static are depictions of various animals. If such neural networks can be fooled by static, what else will fool thinking machines of the future?
A penny for your thoughts? You may not choose to answer. But the point is that, as a conscious agent, you surely can. That's what it means to have introspective access. You know—and can tell us—what's on stage in the theatre of your mind. Then how about machines? A bitcoin for the thinking machine's thoughts? But no one has yet designed a machine to have this kind of access. Wittgenstein remarked that, if a lion could speak, we would not understand him. If a machine could speak, it would not have anything to say. "What do you think about machines that think". Simple. I don't think that—as yet—there are any such machines.
Of course this may soon change. Far back in human history, natural selection discovered that, given the particular problems humans faced, there were practical advantages to having a brain capable of introspection. Likewise machine programmers may well discover that, when and if machines face similar problems, the software trick that works for humans will work for them as well. But what are these problems, and why is the theatre of consciousness the answer?
The theatre lets you in on a secret. It lets you see how your own mind works. Observing, for example, how beliefs and desires generate wishes that lead to actions, you begin to gain insight into why you think and act the way you do. So you can explain yourself to yourself, and explain yourself to other people too. But, equally important, it means you have a model for explaining other people to yourself. Introspective consciousness has laid the ground for what psychologists call "Theory of Mind."
With humans, for whom social intelligence is the key to biological survival, the advantages have been huge. With machines, for whom success in social life has not yet become an issue, there has been little if any reason to go that way. However there's no question the time is coming when machines will indeed need to understand other machines' psychology, so as to be able to work alongside them. What's more, if they are to collaborate effectively with humans, they will need to understand human psychology too. I guess that's when their designers—or maybe the machines themselves—will follow Nature's lead and install a machine version of the inner eye.
Is there a danger that, once this stage is reached these newly insightful machines will come to understand humans only too well? Psychopaths are sometimes credited with having not too little but too great an understanding of human psychology. Is this something we should fear with machines?
I don't think so. This situation is actually not a new one. For thousands of years humans have been selecting and programming a particular species of biological machine to act as servants, companions and helpmeets to ourselves. I'm talking of the domestic dog. The remarkable result has been that modern dogs have in fact acquired an exceptional and considerable ability to mind-read—both the minds of other dogs and humans—superior to that of any animal other than humans themselves. This has evidently evolved as a mutually beneficial relationship, not a competition, even if it's one in which we have retained the upper hand. If and when it gets to the point where machines are as good at reading human minds as dogs now are, we shall of course have to watch out in case they get too dominant and manipulative, perhaps even too playful—just as we already have to do with man's best friend. But I see no reason to doubt we'll remain in control.
There is a painting by Goya of a terrible Colossus who strides across the landscape, while the human population flees in terror. Colossus was the name of one of Turing's first computing machines. Do we have to imagine an existential threat to humanity coming from that computer's descendants? No, I look on the bright side. With luck, or rather by arrangement, the Colossus will remain a Big Friendly Giant.
Proponents of Artificial Intelligence have a tendency to project a utopian future in which benevolent computers and robots serve humanity and enable us to achieve limitless prosperity, end poverty and hunger, conquer disease and death, achieve immortality, colonize the galaxy, and eventually even conquer the universe by reaching the Omega point where we become god—omniscient and omnipotent. AI skeptics envision a dystopian future in which malevolent computers and robots take us over completely, making us their slaves or servants, or driving us into extinction, thereby terminating or even reversing centuries of scientific and technological progress.
Most such prophecies are grounded in a false analogy between human nature and computer nature, or natural intelligence and artificial intelligence. We are thinking machines, the product of natural selection that also designed into us emotions to shortcut the thinking process. We don't need to compute the caloric value of foods; we just feel hungry and eat. We don't need to calculate the waist-to-hip or shoulder-to-waist ratios of potential mates; we just feel attracted to someone and mate with them. We don't need to work out the genetic cost of raising someone else's offspring if our mate is unfaithful; we just feel jealous. We don't need to estimate the damage of an unfair exchange; we just feel injustice and desire revenge. All of these emotions were built into our nature by evolution, none of which we have designed into our computers. So the fear that computers will become evil are unfounded, because it will never occur to them to take such actions against us.
As well, both utopian and dystopian visions of AI are based on a projection of the future quite unlike anything history has given us. Instead of utopia or dystopia, think protopia, a term coined by the futurist Kevin Kelly, who described it in an Edge conversation this way: "I call myself a protopian, not a utopian. I believe in progress in an incremental way where every year it's better than the year before but not by very much—just a micro amount." Almost all progress in science and technology, including computers and artificial intelligence, is of a protopian nature. Rarely, if ever, do technologies lead to either utopian or dystopian societies.
Consider the automobile. My first car was a 1966 Ford Mustang. It had power steering, power brakes, and air conditioning, all of which were relatively cutting edge technology at the time. Every car I've had since then—parallel to the evolution of automobiles in general—has been progressively smarter and safer; not in leaps and bounds, but incrementally. Think of the 1950's imagined jump from the jalopy to the flying car. That never happened. Instead what we got were decades-long cumulative improvements that led to today's smart cars with their onboard computers and navigation systems, air bags and composite metal frames and bodies, satellite radios and hands-free phones, and electric and hybrid engines. I just swapped out a 2010 Ford Flex for a 2014 version of the same model. Externally they are almost indistinguishable; internally there are dozens of tiny improvements in every system, from the engine and drive train, to navigation and mapping, to climate control and radio and computer interface.
Such incremental protopian progress is what we see in most technologies, including and especially artificial intelligence, which will continue to serve us in the manner we desire and need. Instead of Great Leap Forward or Giant Phase Backward, think Small Step Upward.
Creatures once inhabited fantastic unknown lands on medieval maps. Those animals were useful fictions of rumor and innuendo, where men's heads were in their bodies, or their humanity was mixed with the dog or the lion, closing the gap between man and animal. They were the hopes and fears of what might live within the unknown. Today, we imagine machines with consciousness.
Besides self-awareness, the imaginary beasts of A.I. possess calculation and prediction, independent thought, and knowledge of their creators. Pessimists fear these machines could regard us and pass lethal verdicts. Optimists hope the thinking machines are benevolent, an illuminating aid and a comfort to people.
Neither idea of an encounter with an independent man-made intelligence has much evidence of becoming real. That doesn't mean they aren't interesting. The old mariners' maps were drawn in a time of primitive sailing technology. We are starting to explore a world thoroughly enchanted by computation. The creatures of A.I. Island fuse the human and the machine, but to the same end as the fusing of man and animal. If they could sing, they would sing songs of us.
What do we mean when we talk about the kind of "intelligence" that might look at mankind and want it dead, or illuminate us as never before? It is clearly more than a machine wins at chess. We have one of those, with no discernable change in the world, other than a new reason to celebrate the very human intelligence of Deep Blue's creators.
The beings of A.I. Island do something far more interesting than outplaying Kasparov. They feel like playing chess. They know the exhilaration of mental stimulation, and the torture of its counterpart, boredom.
This means making software that encodes an awareness of having only one finite life, which somehow matters greatly to some elusive self. It is driven nearly mad by the absence of some kind of stimulation—playing chess, perhaps. Or killing mankind.
Like us, the fabulous creatures of A.I. Island want to explain themselves, and judge others. They have our slight distance from the rest of reality that we believe other animals don't feel. An intelligence that is like ours knows it is sentient, feels something is amiss, and is continually trying to do something about that.
With these kind of software challenges, and given the very real technology-driven threats to our species already at hand, why worry about malevolent A.I.? For decades to come, at least, we are clearly more threatened by like trans-species plagues, extreme resource depletion, global warming, and nuclear warfare.
Which is why malevolent A.I. rises in our Promethean fears. It is a proxy for us, at our rational peak, confidently killing ourselves.
The dreams of benevolent A.I. are equally self-reflective. These machine companions have super intellects turned towards their creators. Given the autonomy implicit in a high level of A.I., we must see these new beings as interested in us. Come to think of it, malevolent A.I. is interested in us too, just in the wrong way.
Both versions of the strange beast reflect a deeper truth, which is the effect that the new exploration of a computer-enchanted world has on us. By augmenting ourselves with computers, we are becoming new beings—if you will, monsters to our former selves.
We have changed our consciousness many times over the past 50,000 years, taking on ideas of an afterlife, or monotheism, or becoming a print culture, or a species well aware of its tiny place in the cosmos. But we have never changed so swiftly, or with such knowledge that we are undertaking the change.
Consider some effects just in the past decade. We have killed many of our historic barriers of time and space with instantaneous communications. Language no longer divides us, because of increasingly better computer translation and image sharing. Open source technology and Internet search give us a little-understood power of working in collective ways.
Beside the positives is the disappearance of privacy, and tracking humans to better control their movements and desires. We are willfully submitting to unprecedented social connection—a seeming triviality that may extinguish all ideas of solitude and selfhood. Ideas of economics are changing under the guise of robotics and the sharing economy.
We are building new intelligent beings, but we are building them within ourselves. It is only artificial now, because it is new. As it becomes dominant, it will simply become intelligence.
The machines of A.I. Island are also what we fear may be ourselves, within a few generations. And we hope those machine-driven people feel the kinship with us, even down to our loneliness and distance from the world, which is also our wellspring of human creativity.
We have met the A.I., and it is us. In a timeless human tension, we yearn for transcendence, but we don't want to change too much.
Machines (humanly constructed artifacts) cannot think because no machine has a point of view; that is, a unique perspective on the worldly referents of its internal symbolic logic. We, as conscious cognitive observers, look at the output of so-called "thinking machines" and provide our own referents to the symbolic structures spouted by the machine. Of course, despite this limitation, such non-thinking machines have provided an extremely important adjunct to human thought.
Those of you participating in this particular Edge Question don't need to be reintroduced to the Ghemawat-Dean Conversational artificial intelligence test (DGC). Past participants in the test have failed as obviously as they have hilariously. However, the 2UR-NG entry really surprised us all with its amazing, if child-like, approach to conversation and its ability to express desire, curiosity and its ability to retain and chain facts.
Its success has caused many of my compatriots to write essays like "The coming biological future will doom us all" and making jokes about "welcoming their new biological overlords". You should know that I don't subscribe to this kind of doom and gloom scare-writing. Before I tell you why we should not worry about the extent of biological intelligence, I thought I'd remind people of the very real limits of biological intelligence.
First off, speed of thought: These biological processes are slow and use an incredible amount of resources. I cannot emphasize enough how incredibly difficult to produce these intelligences. One has to waste so much biological material, and I know from experience that takes forever to assemble the precursors in the genesis machine. Following this arduous process, your specimen has to gestate. Gestate! I mean, it's not like these ... animals....come about the way we do through clean, smart, crystallography or in the nitrogen lakes of my youth, they have to be kept warm for months and months and then decanted (A very messy process, I assure you) and then you as often as not have an inviable specimen.
It is kind of gross, really. But let's suppose you get to birth these specimens, then you have to feed them and again, keep them warm. A scientist can't even work within their environmental spaces without a cold jacket circulating helium throughout your terminal. Then you have to 'feed' them. They don't use power like we do, but instead ingest other living matter. It's disgusting to observe and I've lost a number of grad students with weaker constitutions.
Assume you've gotten far enough to try to do the GDC. You've kept them alive through any a variety of errors in their immune system. They've not choked on their sustenance, they haven't drown in their solvent and they've managed to keep their wet parts off things that they would freeze, bond or be electrocuted by. What if those organisms continue to develop, will they then rise up and take over? I don't think so. They have to deal with so many problems related to their design. I mean, their processors are really just chemical soups that have to be kept in constant balance. Dopamine at this level or they shut down voluntarily. Vasopressin at this level or they start retaining water. Adrenaline at this level for this long or poof their power delivery network stops working.
Moreover, don't get me started on the power delivery method! It's more like the fluorinert liquid cooling systems of our ancestors than a modern heat tolerant wafers. I mean, they have meat that filters their coolant/power delivery system that are constantly failing. Meat! You introduce the smallest amount of machine oil or cleaning solvent into the system and they stop operating fast. One side effect of certain ethanol mixtures is the specimens expel their nutrition, but they seem to like it in smaller amounts. It is baffling in its ambiguity.
And their motivations! Creating new organisms seems paramount, more important than data ingress/egress, computation or learning. It's baffling. I can't imagine that they would see us machine-folk as anything but tools to advance their reproduction. We could end the experiment simply by matching them poorly with each other or only allowing access to each other with protective cladding. In my opinion, there is nothing to fear from these animals. In the event they grow beyond the confines of their cages, maybe we can then ask ourselves the more important question: If humans show real machine-like intelligence, do they deserve to be treated like machines? I would think so, and I think we could be proud to be the parent processes of a new age.
In the pantheon of gruesome medical experiments few match head transplants. Animal experiments have attempted this procedure in two ways: substitute one head for another or graft a second head onto an animal. So far the procedure has not been very successful. But we are getting far better at vascular surgery, bypassing, stitching, and grafting both big and microscopic vessels. Similar advances are taking place in rebuilding muscles and damaged vertebrae. Even the reattachment of severed spinal cords, in mice and primates, seems to be advancing steadily.
Partial brain transplants are likely a long way out. Other than some stem cell procedures, attaching parts of one brain to another is highly complex given the consistency of most brain mass and the trillions of connections. But as extreme operations, reattachments of fingers, limbs, even faces, become commonplace the question of whether we could, and should, transplant an entire human head loom closer.
Partly reattaching a human head is already a reality. In 2002 a drunk driver hit teenager Marcos Parra so hard Parra's head was almost entirely detached; only the spinal cord, and a few blood vessels, kept the entire cranium from coming off. Fortunately a surgeon, Curtis Dickman, had been preparing for just this type of emergency. Screws reattached vertebrae to the base of the skull, part of the pelvic bone was redeployed to bring neck and head back together, and within six months Parra was playing basketball.
Successful animal whole head transplants may not be that far out. And if such procedures were successful, and the animal regained consciousness, one could begin to answer pretty fundamental questions including: do the donor's memories and consciousness also transplant?
Similar questions were asked during the first heart transplants, but it turns out the emotions, attachments, and loves of the donor did not transplant with the organ that was always "tied" to emotions. The heart is but a muscle. How about the brain?
If mice with new heads recognized previously navigated mazes, or maintained the previous mouse's conditioned reactions to certain foods, smells, or stimuli, we would have to consider the possibility that memory and consciousness do transplant. But if experiment after experiment demonstrated no previous knowledge or emotions, then we would have to consider that the brain too might just be an electro chemical muscle.
Actually knowing if you can transplant knowledge and emotions from one body to another goes a long way towards answering the question "could we ever download and store part of our brains, not just into another body but eventually into a chip, into a machine?" If you could, then it would make the path to large scale AI far easier. We would simply have to copy, merge, and augment existing data, data that we would know is transferable, stackable, manipulatable. The remaining question would be: what is the most efficient interface between the biology and the machine.
But if it turned out that all data erases upon transplant, that knowledge is unique to the individual organism, (in other words that there is something innate and individual to consciousness-knowledge- intelligence), then simply copying the dazzlingly complex connectome of brains into machines would likely not lead to an operative intelligence.
If brain data is not transferable, or replicable, then developing AI would require building a parallel machine thought system, something quite separate and distinct from animal and human intelligence. Building consciousness from scratch implies following a new and very different evolutionary path to that of human intelligence. Likely this new system would eventually operate under very different rules and constraints. In which case it would likely be far better at certain tasks and be unable to emulate some forms of our intelligence. Were AI to emerge from this kind of evolutionary system it would likely represent a new, distinct consciousness, one on a parallel evolutionary track. In this scenario how machines might think, feel, govern could have little to do with the billions of years of animal-human intelligence and learning. Nor would they be constrained to organize their society, and its rules, as do we.
"Think" and "intelligence" are both what Marvin Minsky has called suitcase words. They are words into which we pack many meanings so that we can talk about complex issues in a shorthand way. When we look inside these words we find many different aspects, mechanisms, and levels of understanding. This makes answering the perennial questions of "can machines think?" or "when will machines reach human level intelligence?" fraught with danger. The suitcase words are used to cover both specific performance demonstrations by machines and more general competence that humans might have. People are getting confused and generalizing from performance to competence and grossly overestimating the real capabilities of machines today and in the next few decades.
In 1997 a super computer beat world chess champion Garry Kasparov in a tournament. Today there are dozens of programs that run on laptop computers and have higher chess rankings than those ever achieved by humans. Computers can definitely perform better than humans at playing chess. But they have nowhere near human level competence at chess.
All chess playing programs use Turing's brute force tree search method with heuristic evaluation. Computers were fast enough by the seventies that this approach overwhelmed other AI programs that tried to play chess with processes that emulated how people reported that they thought about their next move, and so those approaches were largely abandoned.
Today's chess programs have no way of saying why a particular move is "better" than another move, save that it moves the game to a part of a tree where the opponent has less good options. A human player can make generalizations and describe why certain types of moves are good, and use that to teach a human player. Brute force programs cannot teach a human player, except by being a sparing partner. It is up to the human to make the inferences, the analogies, and to do any learning on their own. The chess program doesn't know that it is outsmarting the person, doesn't know that it is a teaching aid, doesn't know that it is playing something called chess nor even what "playing" is. Making brute force chess playing perform better than any human gets us no closer to competence in chess.
Now consider deep learning that has caught people's imaginations over the last year or so. It is an update to backpropagation, a thirty-year old learning algorithm very loosely based on abstracted models of neurons. Layers of neurons map from a signal, such as amplitude of a sound wave or pixel brightness in an image, to increasingly higher-level descriptions of the full meaning of the signal, as words for sound, or objects in images. Originally backpropagation could only practically work with just two or three layers of neurons, so it was necessary to fix preprocessing steps to get the signals to more structured data before applying the learning algorithms. The new versions work with more layers of neurons, making the networks deeper, hence the name, deep learning. Now early processing steps are also learned, and without misguided human biases of design, the new algorithms are spectacularly better than the algorithms of just three years ago. That is why they have caught people's imaginations. The new versions rely on massive amounts of computer power in server farms, and on very large data sets that did not formerly exist, but critically, they also rely on new scientific innovations.
A well-known particular example of their performance is labeling an image, in English, saying that it is a baby with a stuffed toy. When a person looks at the image that is what they also see. The algorithm has performed very well at labeling the image, and it has performed much better than AI practitioners would have predicted for 2014 performance only five years ago. But the algorithm does not have the full competence that a person who could label that same image would have.
The learning algorithm knows there is a baby in the image but it doesn't know the structure of a baby, and it doesn't know where the baby is in the image. A current deep learning algorithm can only assign probabilities to each pixel that that particular pixel is part of a baby. Whereas a person can see that the baby occupies the middle quarter of the image, today's algorithm has only a probabilistic idea of its spatial extent. It cannot apply an exclusionary rule and say that non-zero probability pixels at extremes of the image cannot both be part of the baby. If we look inside the neuron layers it might be that one of the higher level learned features is an eye-like patch of image, and another feature is a foot-like patch of image, but the current algorithm would have no capability of relating the constraints of where and what spatial relationships could possibly be valid between eyes and feet in an image, and could be fooled by a grotesque collage of baby body parts, labeling it a baby. In contrast no person would do so, and furthermore would immediately know exactly what it was—a grotesque collage of baby body parts. Furthermore the current algorithm is completely useless at telling a robot where to go in space to pick up that baby, or where to hold a bottle and feed the baby, or where to reach to change its diaper. Today's algorithm has nothing like human level competence on understanding images.
Work is underway to add focus of attention and handling of consistent spatial structure to deep learning. That is the hard work of science and research, and we really have no idea how hard it will be, nor how long it will take, nor whether the whole approach will reach a fatal dead end. It took thirty years to go from backpropagation to deep learning, but along the way many researchers were sure there was no future in backpropagation. They were wrong, but it would not have been surprising if they had been right, as we knew all along that the backpropagation algorithm is not what happens inside people's heads.
The fears of runaway AI systems either conquering humans or making them irrelevant are not even remotely well grounded. Misled by suitcase words, people are making category errors in fungibility of capabilities. These category errors are comparable to seeing more efficient internal combustion engines appearing and jumping to the conclusion that warp drives are just around the corner.
When we say "machines that think", we really mean: "machines that think like people". It is obvious that, in many different ways, machines do think: They trigger events, process things, take decisions, make choices, and perform many, but not all, other aspects of thinking. But the real question is whether machines can think like people, to the point of the age old test of artificial intelligence: You will observe the results of the thinking, and you will not be able to tell if it is a machine or a human.
Some prominent scientific gurus are scared by a world controlled by thinking machines. I am not sure that this is a valid fear. I am more concerned about a world led by people, who think like machines, a major emerging trend of our digital society.
You can teach a machine to track an algorithm and to perform a sequence of operations which follow logically from each other. It can do so faster and more accurately than any human. Given well defined basic postulates or axioms, pure logic is the strong suit of the thinking machine. But exercising common sense in making decisions and being able to ask meaningful questions are, so far, the prerogative of humans. Merging Intuition, emotion, empathy, experience and cultural background, and using all of these to ask a relevant question and to draw conclusions by combining seemingly unrelated facts and principles, are trademarks of human thinking, not yet shared by machines.
Our human society is currently moving fast towards rules, regulations, laws, investment vehicles, political dogmas and patterns of behavior that blindly follow strict logic, even when it starts with false foundations or collides with obvious common sense. Religious extremism has always progressed on the basis of some absurd axioms, leading very logically to endless harsh consequences. Several disciplines such as law, accounting and certain areas of mathematics and technology, augmented by bureaucratic structures and by media which idolize inflexible regulators, often lead to opaque principles like "total transparency" and to tolerance towards acts of extreme intolerance. These and similar trends are visibly moving us towards more algorithmic and logical modes of tackling problems, often at the expense of common sense. If common sense, whatever its definition is, describes one of the advantages of people over machines, what we see today is a clear move away from this incremental asset of humans.
Unfortunately, the gap between machine thinking and human thinking can narrow in two ways, and when people begin to think like machines, we automatically achieve the goal of "machines that think like people", reaching it from the wrong direction. A very smart person, reaching conclusions on the basis of one line of information, in a split second between dozens of e-mails, text messages and tweets, not to speak of other digital disturbances, is not superior to a machine with a moderate intelligence, which analyzes a large amount of relevant information before it jumps into premature conclusions and signs a public petition about a subject it is unfamiliar with.
One can recite hundreds of examples of this trend. We all support the law that every new building should allow total access to people with special needs, while old buildings may remain inaccessible, until they are renovated. But does it make sense to disallow a renovation of an old bathroom which will now offer such access, because a new elevator cannot be installed? Or to demand full public disclosure of all CIA or FBI secret sources in order to enable a court of law to sentence a terrorist who obviously murdered hundreds of people? Or to demand parental consent before giving a teenager an aspirin at school? And then when school texts are converted from the use of miles to kilometers, the sentence "From the top of the mountain you can see for approximately 100 miles" is translated, by a person, into "you can see for approximately 160.934 km".
The standard sacred cows of liberal democracy rightfully include a wide variety of freedoms: Freedom of speech, freedom of the press, academic freedom, freedom of religion (or of lack of religion), freedom of information, and numerous other human rights including equal opportunity, equal treatment by law, and absence of discrimination. We all support these principles, but pure and extreme logic induces us, against common sense, to insist mainly on human rights of criminals and terrorists, because the human rights of the victims "are not an issue"; Transparency and freedom of the press logically demand complete reports on internal brainstorming sessions, in which delicate issues are pondered, thus preventing any free discussion and raw thinking in certain public bodies; Academic freedom might logically be misused, against common sense and against factual knowledge, to teach about Noah's ark as an alternative to evolution, to deny the holocaust in teaching history or to preach for a universe created 6000 years ago (rather than 13 Billions) as the basis of cosmology. We can continue on and on with examples, but the message is clear.
Algorithmic thinking, brevity of messages and over-exertion of pure logic are moving us, human beings, into machine thinking, rather than slowly and wisely teaching our machines to benefit from our common sense and intellectual abilities. A reversal of this trend would be a meaningful U-turn in human digital evolution.
In 1950, Alan Turing suggested we should ask not "Can Machines Think" but rather "What Can Machines Do?" Edsger Dijkstra got it right in 1984 when he said the question of Can Machine Think "is about as relevant as the question of whether Submarines Can Swim." By that he meant that both are questions in sociolinguistics: how do we choose to use words such as "think"? In English, submarines do not swim, but in Russian, they do. This is irrelevant to the capabilities of submarines. So let's explore what it is that machines can do, and whether we should fear their capabilities.
Pessimists warn that we don't know how to safely and reliably build large complex AI systems. They have a valid point. We also don't know how to safely and reliably build large complex non-AI systems. For example, we invented the internal combustion engine 150 years ago, and in many ways it has served humanity well, but it also has lead to widespread pollution, political instability over access to oil, a million deaths per year, and other problems.
Any complex system will have a mix of positive outcomes and unintended consequences but are there worrisome issues that are unique to systems built with AI? I think the interesting issues are Adaptability, Autonomy, and Universality.
Systems that use machine learning are adaptable. They change over time, based on what they learn from examples. Adaptability is useful. We want, say, our automated spelling correction programs to quickly learn new terms such as "bitcoin", rather than waiting for the next edition of a published dictionary to list them. A non-adaptable program will repeat the same mistakes. But an adaptable program can make new mistakes, which may be harder to predict and deal with. We have tools for dealing with these problems, but just as the designers of bridges must learn to deal with crosswinds, so the designers of AI systems must learn to deal with adaptability.
Some critics are worried about AI systems that are built with a framework that maximizes expected utility. Such an AI system estimates the current state of the world, considers all the possible actions it can take, simulates the possible outcomes of those actions, and then chooses the action that leads to the best possible distribution of outcomes. Errors can occur at any point along the way, but the concern here is in determining what is the "best outcome"—in other words, what is it that we desire? If we describe the wrong desires, or allow a system to adapt its desires in a wrong direction, we get the wrong results.
History shows that we often get this wrong, in all kinds of systems that we build, not just in AI systems. The US Constitution is a document that specifies our desires; the original framers made what we now recognize as an error in this specification, and correcting that error with the 13th amendment cost over 600,000 lives. Similarly, we designed stock-trading system that allowed speculators to create bubbles that led to busts. These are important issues for system design (and what is known as "mechanism design"), and are not specific to AI systems. The world is complicated, so acting correctly in the world is complicated.
The second concern is autonomy. If AI systems act on their own, they can make errors that perhaps would not be made by a system with a human in the loop. This too is a valid concern, and again one that is not unique to AI systems. Consider our system of automated traffic lights, which replaced a system of human policemen directing traffic. The automated system leads to some errors, but is a tradeoff that we have decided is worthwhile. We will continue to make tradeoffs in where we deploy autonomous systems.
There is a possibility that we will soon see a widespread increase in the capabilities of autonomous systems, and thus more displacement of people. This could lead to a societal problem of increased unemployment and income inequality. To me, this is the most serious concern about future AI systems. In past technological revolutions (agricultural and industrial) the notion of work changed, but the changes happened over generations, not years, and the changes always led to new jobs. We may be in for a period of change that is much more rapid and disruptive; we will need some social conventions and safety nets to restore stability.
The third concern is the universality of intelligent machines. In 1965 I. J. Good wrote "an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make." I think this fetishizes "intelligence" as a monolithic superpower, and I think reality is more nuanced. The smartest person is not always the most successful; the wisest policies are not always the ones adopted. Recently I spent an hour reading the news about the middle east, and thinking. I didn't come up with a solution. Now imagine a hypothetical "Speed Superintelligence" (as described by Nick Bostrom) that could think as well as any human but a thousand times faster. I'm pretty sure it also would have been unable to come up with a solution. I also know from computational complexity theory that there are a wide class of problems that are completely resistant to intelligence, in the sense that, no matter how clever you are, you won't have enough computing power. So there are some problems where intelligence (or computing power) just doesn't help.
But of course, there are many problems where intelligence does help. If I want to predict the motions of a billion stars in a galaxy, I would certainly appreciate the help of a computer. Computers are tools. They are tools of our design that fit into niches to solve problems in societal mechanisms of our design. Getting this right is difficult, but it is difficult mostly because the world is complex; adding AI to the mix doesn't fundamentally change things. I suggest being careful with our mechanism design and using the best tools for the job regardless of whether the tool has the label "AI" on it or not.
The rumors of the enslavement or death of the human species at the hands of an artificial intelligence are highly exaggerated, because they assume that an AI will have a teleological autonomy akin to our own. I don't think anything less than a fully Darwinian process of evolution can give that to any creature.
There are basically two ways in which we could produce an AI. The first is by writing a comprehensive set of programs that can perform specific tasks that human minds can perform, perhaps even faster and better than we can, without worrying about exactly how humans perform those tasks, and then bringing those modules together into an integrated intelligence. We have already started this project and succeeded in some areas. For example, computers can play chess better than humans. One can imagine that with some effort it may well be possible to program computers to perform even more creative tasks, such as writing beautiful (to us) music or poetry with some clever heuristics and built-in knowledge.
But here's the problem with this approach: We deploy our capabilities according to values and constraints programmed into us by billions of years of evolution (and some learned during our lifetimes), and we share some of these values with the earliest life-forms, including, most important, the need to survive and reproduce. Without these values, we would not be here, and we would not have the finely tuned (to our environment) emotions that allow us not only to survive but also to cooperate with others. The importance of this value-laden emotional side of our minds is made obvious by, among other things, the many examples of perfectly rational individuals who cannot function in society because of damage to the emotional centers of their brains.
So, what values and emotions will an AI have? One could simply program such values into an AI, in which case we choose what the AI will "want" to do, and we needn't worry about the AI pursuing goals that diverge from ours. We could easily make the AI unable to modify certain basic imperatives we give it. (Yes, something like a more comprehensive version of Isaac Asimov's Laws of Robotics.)
The second way to produce an AI is by deciphering in detail how the human brain works. It's conceivable that there may soon come a eureka moment about the structure and conceptual hierarchy of the brain—similar to Watson and Crick and Franklin and Wilkins's discovery of the structure of DNA and the subsequent rapid understanding of the hereditary mechanism. We might simulate or reproduce that functional structure on silicon, or some other substrate, as a mixture of hardware and software.
At first blush, this may seem a convenient way to quickly bestow on an AI the benefit of our own long period of evolution, as well as a way to give it values of its own by functionally reproducing the emotional centers of our own brain, along with the "higher thought" parts, like the cortex. But our brains are specifically designed to accept information from the vast sensory apparatus of our bodies and to react to this. What would the equivalent be for an AI? Even given a sophisticated body with massive sensory capability, what an AI would need to survive in the world is presumably very different from what we need. It could achieve some emotional tuning from interacting with its environment, but what it would need to develop true autonomy and desires of its own would be nothing short of a long process of evolution entailing the Darwinian requirements of reproduction with variability and natural selection. This it won't have, because we are not speaking of artificial life here. So, again, we'll end up giving it whatever values we choose for it.
It's of course, conceivable that someone will produce intelligent robots as weapons (or soldiers) to be used against other humans in war, but these weapons will simply carry out the intentions of their creators and, lacking any will or desire of their own, will not pose a threat to humanity at large any more than any other weapons already do. So both potential roads to an AI (at least, ones achievable on a less-than-geological timescale) will fail to give that AI the purposive autonomy, free of the intentionality of its creators, that might actually threaten them.
My favorite Edsger Dijkstra aphorism is this one: "The question of whether machines can think is about as relevant as the question of whether submarines can swim." Yet we keep playing the imitation game: asking how closely machine intelligence can duplicate our own intelligence, as if that is the real point. Of course, once you imagine machines with human-like feelings and free will, it's possible to conceive of misbehaving machine intelligence—the AI as Frankenstein idea. This notion is in the midst of a revival, and I started out thinking it was overblown. Lately I have concluded it's not.
Here's the case for overblown. Machine intelligence can go in so many directions. It is a failure of imagination to focus on human-like directions. Most of the early futurist conceptions of machine intelligence were wildly off base because computers have been most successful at doing what humans can't do well. Machines are incredibly good at sorting lists. Maybe that sounds boring, but think of how much efficient sorting has changed the world.
In answer to some of the questions brought up here, it is far from clear that there will ever be a practical reason for future machines to have emotions and inner dialog; to pass for human under extended interrogation; to desire, and be able to make use of, legal and civil rights. They're machines, and they can be anything we design them to be.
But that's the point. Some people will want anthropomorphic machine intelligence. How many videos of Japanese robots have you seen? Honda, Sony, and Hitachi already expend substantial resources in making cute AI that has no concrete value beyond corporate publicity. They do this for no better reason than tech enthusiasts have grown up seeing robots and intelligent computers in movies.
Almost anything that is conceived—that is physically possible and reasonably cheap—is realized. So human-like machine intelligence is a meme with manifest destiny, regardless of practical value. This could entail nice machines-that-think, obeying Asimov's laws. But once the technology is out there, it will get ever cheaper and filter down to hobbyists, hackers, and "machine rights" organizations. There is going to be interest in creating machines with will, whose interests are not our own. And that's without considering what machines that terrorists, rogue regimes, and intelligence agencies of the less roguish nations, may devise. I think the notion of Frankensteinian AI, which turns on its creators, is something worth taking seriously.
A common theme in recent writings about machine intelligence is that the best new learning machines will constitute rather alien forms of intelligence. I'm not so sure. The reasoning behind the 'alien AIs' image usually goes something like this. The best way to get machines to solve hard real-world problems is to set them up as statistically-sensitive learning machines able to benefit maximally from exposure to 'big data'. Such machines will often learn to solve complex problems by detecting patterns, and patterns among patterns, and patterns within patterns, hidden deep in the massed data streams to which they are exposed. This will most likely be achieved using 'deep learning' algorithms to mine deeper and deeper into the data streams. After such learning is complete, what results may be a system that works but whose knowledge structures are opaque to the engineers and programmers who set the system up in the first place.
Opaque? In one sense yes. We won't (at least without further work) know in detail what has become encoded as a result of all that deep, multi-level, statistically-driven learning. But alien? I'm going to take a big punt at this point and road-test a possibly outrageous claim. I suspect that the more these machines learn, the more they will end up thinking in ways that are recognizably human. They will end up having a broad structure of human-like concepts with which to approach their tasks and decisions. They may even learn to apply emotional and ethical labels in roughly the same ways we do. If I am right, this somewhat undermines the common worry that these are emerging alien intelligences whose goals and interests we cannot fathom, and that might therefor turn on us in unexpected ways. By contrast, I suspect that the ways they might turn on us will be all-too-familiar—and thus hopefully avoidable by the usual steps of extending due respect and freedom.
Why would the machines think like us? The reason for this has nothing to do with our ways of thinking being objectively right or unique. Rather, it has to do with what I'll dub the 'big data food chain'. These AIs, if they are to emerge as plausible forms of general intelligence, will have to learn by consuming the vast electronic trails of human experience and human interests. For this is the biggest repository of general facts about the world that we have available. To break free of restricted uni-dimensional domains, these AIs will have to trawl the mundane seas of words and images that we lay down on Facebook, Google, Amazon, and Twitter. Where before they may have been force-fed a diet of astronomical objects or protein-folding puzzles, the break-through general intelligences will need a richer and more varied diet. That diet will be the massed strata of human experience preserved in our daily electronic media.
The statistical baths in which we immerse these potent learning machines will thus be all-too-familiar. They will feed off the fossil trails of our own engagements, a zillion images of bouncing babies, bouncing balls, LOL-cats, and potatoes that look like the Pope. These are the things that they must crunch into a multi-level world-model, finding the features, entities, and properties (latent variables) that best capture the streams of data to which they are exposed. Fed on such a diet, these AIs may have little choice but to develop a world-model that has much in common with our own. They are probably more in danger of becoming super-Mario freaks than becoming super-villains intent on world-domination.
Such a diagnosis (which is tentative and at least a little playful) goes against two prevailing views. First, as mentioned earlier, it goes against the view that current and future AIs are basically alien forms of intelligence feeding off big data and crunching statistics in ways that will render their intelligences increasingly opaque to human understanding. On the contrary, access to more and more data, of the kind most freely available, won't make them more alien but less so.
Second, it questions the view that the royal route to human-style understanding is human-style embodiment, with all the interactive potentialities (to stand, sit, jump etc.) that that implies. For although our own typical route to understanding the world goes via a host of such interactions, it seems quite possible that theirs need not. Such systems will doubtless enjoy some (probably many and various) means of interacting with the physical world. These encounters will be combined, however, with exposure to rich information trails reflecting our own modes of interaction with the world. So it seems possible that they could come to understand and appreciate soccer and baseball just as much as the next person. An apt comparison here might be with a differently-abled human being.
There's lots more to think about here of course. For example, the AIs will see huge swathes of human electronic trails, and will thus be able to discern patterns of influence among them over time. That means they may come to model us less as individuals and more as a kind of complex distributed system. That's a difference that might make a difference. And what about motivation and emotion? Maybe these depend essentially upon features of our human embodiment such as gut feelings, and visceral responses to danger? Perhaps- but notice that these features of human life have themselves left fossil trails in our electronic repositories.
I might be wrong. But at the very least, I think we should think twice before casting our home-grown AIs as emerging forms of alien intelligence. You are what you eat, and these learning systems will have to eat us. Big time.
Fear not the malevolent toaster, weaponized Roomba, or larcenous ATM. Breakthroughs in the competence of machines, intelligent or otherwise, should not drive paranoia about a future clash between humanity and its mechanical creations. Humans will prevail, in part through primal, often disreputable qualities that are more associated with our downfall than salvation. Cunning, deception, revenge, suspicion, and unpredictability, befuddle less flexible and imaginative entities. Intellect isn't everything, and the irrational is not necessarily maladaptive. Irrational acts stir the neurological pot, nudging us out of unproductive ruts and into creative solutions. Our sociality yields a human superorganism with teamwork and collective, distributed intelligence. There are perks for being emotional beasts of the herd.
Thought experiments about these matters are the source of practical insights into human and machine behavior and suggest how to build different and better kinds of machines. Can deception, rage, fear, revenge, empathy, and the like, be programmed into a machine, and to what effect? (This requires more than the superficial emulation of human affect.) Can a sense of self-hood be programmed into a machine—say, via tickle? How can we produce social machines, and what kind of command structure is required to organize their teamwork? Will groups of autonomous, social machines generate an emergent political structure, culture, and tradition? How will such machines treat their human creators? Can natural and artificial selection be programmed into self-replicating robots?
There is no indication that we will have a problem keeping our machines on a leash, even if they misbehave. We are far from building teams of swaggering, unpredictable, Machiavellian robots with an attitude problem and urge to reproduce.
1. We are They
Francis Crick called it the "Astonishing Hypothesis": that consciousness, also known as Mind, is an emergent property of matter. As molecular neuroscience progresses, encountering no boundaries, and computers reproduce more and more of the behaviors we call intelligence in humans, that Hypothesis looks inescapable. If it is true, then all intelligence is machine intelligence. What distinguishes natural from artificial intelligence is not what it is, but only how it is made.
Of course, that little word "only" is doing some heavy lifting here. Brains use a highly parallel architecture, and mobilize many noisy analog units (i.e., neurons) firing simultaneously, while most computers use von Neumann architecture, with serial operation of much faster digital units. These distinctions are blurring, however, from both ends. Neural net architectures are built in silicon, and brains interact ever more seamlessly with external digital organs. Already I feel that my laptop is an extension of my self—in particular, it is a repository for both visual and narrative memory, a sensory portal into the outside world, and a big part of my mathematical digestive system.
2. They are Us
Artificial intelligence is not the product of an alien invasion. It is an artifact of a particular human culture, and reflects the values of that culture.
3. Reason Is the Slave of the Passions
David Hume's striking statement: "Reason Is, and Ought only to Be, the Slave of the Passions" was written in 1738, long before anything like modern AI was on the horizon. It was, of course, meant to apply to human reason and human passions. (Hume used the word "passions" very broadly, roughly to mean "non-rational motivations".) But Hume's logical/philosophical point remains valid for AI. Simply put: Incentives, not abstract logic, drive behavior.
That is why the AI I find most alarming is its embodiment in autonomous military entities—artificial soldiers, drones of all sorts, and "systems." The values we may want to instill in such entities are alertness to threats and skill in combatting them. But those positive values, gone even slightly awry, slide into paranoia and aggression. Without careful restraint and tact, researchers could wake up to discover they've enabled the creation of armies of powerful, clever, vicious paranoiacs.
Incentives driving powerful AI might go wrong in many ways, but that route seems to me the most plausible, not least because militaries wield vast resources, invest heavily in AI research, and feel compelled to compete with one another. (In other words, they anticipate possible threats and prepare to combat them ... )
There is no reason to believe that as machines become more intelligent—and intelligence such as ours is still little more than a pipe-dream—they will become evil, manipulative, self-interested or in general, a threat to humans. Self-interest is a property of things that 'want' to stay alive (or more accurately, that want to reproduce), and this is not a natural property of machines—computers don’t mind, much less worry, about being switched off.
So, full-blown artificial intelligence (AI) will not spell the 'end of the human race', it is not an 'existential threat' to humans (digression: this now-common use of 'existential' is incorrect), we are not approaching some ill-defined apocalyptic 'singularity', and the development of AI will not be 'the last great event in human history'—all claims that have recently been made about machines that can think.
In fact, as we design machines that get better and better at thinking, they can be put to uses that will do us far more good than harm. Machines are good at long, monotonous tasks like monitoring risks, they are good at assembling information to reach decisions, they are good at analyzing data for patterns and trends, they can arrange for us to use scarce or polluting resources more efficiently, they react faster than humans, they are good at operating other machines, they don’t get tired or afraid, and they can even be put to use looking after their human owners, as in the form of smartphones with applications like Siri and Cortana, or the various GPS route-planning devices most people have in their cars.
Being inherently self-less rather than self-interested, machines can easily be taught to cooperate, and without fear that some of them will take advantage of the other machines’ goodwill. Groups (packs, teams, bands, or whatever collective noun will eventually emerge—I prefer the ironic jams) of networked and cooperating driverless cars will drive safely nose-to-tail at high-speeds: they won’t nod off, they won’t get angry, they can inform each other of their actions and of conditions elsewhere, and they will make better use of the motorways, which now are mostly unoccupied space (owing to humans' unremarkable reaction times). They will do this happily and without expecting reward, and do so while we eat our lunch, watch a film, or read the newspaper. Our children will rightly wonder why anyone ever drove a car.
There is a risk that we will, and perhaps already have, become dangerously dependent on machines, but this says more about us than them. Equally, machines can be made to do harm, but again, this says more about their human inventors and masters than about the machines. Along these lines, there is a strand of human influence on machines that we should monitor closely and that is introducing the possibility of death. If machines have to compete for resources (like electricity or gasoline) to survive, and they have some ability to alter their behaviours, they could become self-interested.
Were we to allow or even encourage self-interest to emerge in machines, they could eventually become like us: capable of repressive or worse, unspeakable, acts towards humans, and towards each other. But this wouldn’t happen overnight, it is something we would have to set in motion, it has nothing to do with intelligence (some viruses do unspeakable things to humans), and again says more about what we do with machines than machines themselves.
So, it is not thinking machines or AI per se that we should worry about but people. Machines that can think are neither for us nor against us, and have no built-in predilections to be one over the other. To think otherwise is to confuse intelligence with aspiration and its attendant emotions. We have both because we are evolved and replicating (reproducing) organisms, selected to stay alive in often cut-throat competition with others. But aspiration isn’t a necessary part of intelligence, even if it provides a useful platform on which intelligence can evolve.
Indeed, we should look forward to the day when machines can transcend mere problem solving, and become imaginative and innovative—still a long long way off but surely a feature of true intelligence—because this is something humans are not very good at, and yet we will probably need it more in the coming decades than at any time in our history.
Machines cannot think. They are not going to think any time soon. They may increasingly do more interesting things, but the idea that we need to worry about them, regulate them, or grant them civil rights, is just plain silly.
The over promising of "expert systems" in the 1980s killed off serious funding for the kind of AI that tries to build virtual humans. Very few people are working in this area today. But, according to the media, we must be very afraid.
We have all been watching too many movies.
There are two choices when you work on AI. One is the "let's copy humans method." The other is the "let's do some really fast statistics-based computing method." As an example, early chess playing programs tried to out compute those they played against. But human players have strategies, and anticipation of an opponent's thinking is also part of chess playing. When the "out compute them" strategy didn't work, AI people started watching what expert players did and started to imitate that. The "out compute them" strategy is more in vogue today.
We can call both of these methodologies AI if we like, but neither will lead to machines that create a new society.
The "out compute them" strategy is not frightening because the computer really has no idea what it is doing. It can count things fast without understanding what it is counting. It has counting algorithms, that's it. We saw this with IBM's Watson program on Jeopardy.
One Jeopardy question was: "It was the anatomical oddity of U.S. Gymnast George Eyser, who won a gold medal on the parallel bars in 1904."
A human opponent answered as follows: "Eyser was missing an arm"—and Watson then said, "What is a leg?" Watson lost for failing to note it the leg was "missing."
Try a Google search on "Gymnast Eyser." Wikipedia comes up first with a long article about him. Watson depends on Google. If a Jeopardy contestant could use Google they would do better than Watson. Watson can translate "anatomical" into "body part" and Watson knows the names of the body parts. Watson did not know what an "oddity" is however. Watson would not have known that a gymnast without a leg was weird. If the question had been "what was weird about Eyser?" the people would have done fine. Watson would not have found "weird" in the Wikipedia article nor have understood what gymnasts do, nor why anyone would care. Try Googling "weird" and "Eyser" and see what you get. Keyword search is not thinking, nor anything like thinking.
If we asked Watson why a disabled person would perform in the Olympics, Watson would have no idea what was even being asked. It wouldn't have understood the question, much less have been able to find the answer. Number crunching can only get you so far. Intelligence, artificial or otherwise, requires knowing why things happen, what emotions they stir up, and being able to predict possible consequences of actions. Watson can't do any of that. Thinking and searching text are not the same thing.
The human mind is complicated. Those of us on the "let's copy humans" side of AI spend our time thinking about what humans can do. Many scientists think about this, but basically we don't know that much about how the mind works. AI people try to build models of the parts we do understand. How language is processed, or how learning works—we know a little—consciousness or memory retrieval, not so much.
As an example, I am working on a computer that mimics human memory organization. The idea is to produce a computer that can, as a good friend would, tell you just the right story at the right time. To do this, we have collected (in video) thousands of stories (about defense, about drug research, about medicine, about computer programming …). When someone is trying to do something, or find something out, our program can chime in with a story it is reminded of that it heard. Is this AI? Of course it is. Is it a computer that thinks? Not exactly.
Why not?
In order to accomplish this task we must interview experts and then we must index the meaning of the stories they tell according to the points they make, the ideas they refute, the goals they talk about achieving, and the problems they experienced in achieving them. Only people can do this. The computer can match the index assigned to other indices, such as those in another story it has, or indices from user queries, or from an analysis of a situation it knows the user is in. The computer can come up with a very good story to tell just in time. But of course it doesn't know what it is saying. It can simply find the best story to tell.
Is this AI? I think it is. Does it copy how humans index stories in memory? We have been studying how people do this for a long time and we think it does. Should you be afraid of this "thinking" program?
This is where I lose it about the fear of AI. There is nothing we can produce that anyone should be frightened of. If we could actually build a mobile intelligent machine that could walk, talk, and chew gum, the first uses of that machine would certainly not be to take over the world or form a new society of robots. A much simpler use would be a household robot. Everyone wants a personal servant. The movies depict robot servants (although usually stupidly) because they are funny and seem like cool things to have.
Why don't we have them? Because having a useful servant entails having something that understands when you tell it something, that learns from its mistakes, that can navigate your home successfully and that doesn't break things, act annoyingly, and so on (all of which is way beyond anything we can do.) Don't worry about it chatting up other robot servants and forming a union. There would be no reason to try and build such a capability into a servant. Real servants are annoying sometimes because they are actually people with human needs. Computers don't have such needs.
We are nowhere near close to creating this kind of machine. To do so, would require a deep understanding of human interaction. It would have to understand "Robot, you overcooked that again," or "Robot, the kids hated that song you sang them." Everyone should stop worrying and start rooting for some nice AI stuff that we can all enjoy.
How might AIs think, feel, intend, empathize, socialize, moralize? Actually, almost any way we might imagine, and many ways we might not. To stimulate our imagination, we can contemplate the varieties of natural intelligence on parade in biological systems today, and speculate about the varieties enjoyed by the 99% of species that have sojourned the earth and breathed their last—informed by those lucky few that bequeathed fossils to the pantheon of evolutionary history. We are entitled to so jog our imaginations because, according to our best theories, intelligence is a functional property of complex systems and evolution is inter alia a search algorithm which finds such functions. Thus the natural intelligences discovered so far by natural selection place a lower bound on the variety of intelligences that are possible. The theory of evolutionary games suggests that there is no upper bound: With as few as four competing strategies, chaotic dynamics and strange attractors are possible.
When we survey the natural intelligences served up by evolution, we find a heterogeneity that makes a sapiens-centric view of intelligence as plausible as a geocentric view of the cosmos. The kind of intelligence we find congenial is but another infinitesimal point in a universe of alien intelligences, a universe which does not revolve around, and indeed largely ignores, our kind.
For instance, the female mantis Pseudomantis albofimbriata, when hungry, uses sexual deception to score a meal. She releases a pheromone that attracts males, and then dines on her eager dates.
The older chick of the blue-footed booby Sula nebouxii, when hungry, engages in facultative siblicide. It kills its younger sibling with pecks, or evicts it to die of the elements. The mother watches on without interfering.
These are varieties of natural intelligence, varieties that we find at once alien and disturbingly familiar. They break our canons of empathy, society and morality; and yet our checkered history includes cannibalism and fratricide.
Our survey turns up another critical feature of natural intelligence: each instance has its limits, those points where intelligence passes the baton to stupidity.
The greylag goose Anser anser tenderly cares for her eggs—unless a volleyball is nearby. She will abandon her offspring in vain pursuit of this supernormal egg.
The male jewel beetle Julodimorpha bakewelli flies about looking to mate with a female—unless it spies just the right beer bottle. It will abandon the female for the bottle, and attempt to mate with cold glass until death do it part.
Human intelligence also passes the baton. Einstein is quoted as saying, "Two things are infinite, the universe and human stupidity, and I am not yet completely sure about the universe."
Some limits of human intelligence cause little embarrassment. For instance, the set of functions from the integers to the integers is uncountable, whereas the set of computable functions is countable. Therefore almost all functions are not computable. But try to think of one. Turns out it takes a genius, an Alan Turing, to come up with an example such as the halting problem. And it takes an exceptional mind, just short of genius, even to understand the example.
Other limits strike closer to home: diabetics that can't refuse dessert, alcoholics that can't refuse a drink, gamblers that can't refuse a bet. But it's not just addicts. Behavioral economists find that all of us make "predictably irrational" economic choices. Cognitive psychologists find that we all suffer from "functional fixedness," an inability to solve certain trivial problems, such as Duncker's candle box problem, because we can't think out of the box. The good news, however, is that the endless variety of our limits provides job security for psychotherapists.
But here is the key point. The limits of each intelligence are an engine of evolution. Mimicry, camouflage, deception, parasitism—all are effects of an evolutionary arms race between different forms of intelligence sporting different strengths and suffering different limits.
Only recently has the stage been set for AIs to enter this race. As our computing resources expand and become better connected, more niches will appear in which AIs can reproduce, compete and evolve. The chaotic nature of evolution makes it impossible to predict precisely what new forms of AI will emerge. We can confidently predict, however, that there will be surprises and mysteries, strengths where we have weaknesses, and weaknesses where we have strengths.
But should this be cause for alarm? I think not. The evolution of AIs presents risks and opportunities. But so does the biological evolution of natural intelligences. We have learned that the best way to cope with the variety of natural intelligences is not alarm, but prudence. Don't hug rattle snakes, don't taunt grizzly bears, wear mosquito repellant. To deal with the evolving strategies of viruses and bacteria, wash hands, avoid sneezes, get a flu shot. Occasionally, as with ebola, further measures are required. But once again prudence, not alarm, is effective. The evolution of natural intelligences can be a source of awe and inspiration, if we embrace it with prudence rather than spurn it with alarm.
All species go extinct. Homo sapiens will be no exception. We don't know how it will happen—virus, an alien invasion, nuclear war, a super volcano, a large meteor, a red-giant sun. Yes, it could be AIs, but I would bet long odds against it. I would bet, instead, that AIs will be a source of awe, insight, inspiration, and yes, profit, for years to come.
The Singularity—the fateful moment when AI surpasses its creators in intelligence and takes over the world—is a meme worth pondering. It has the earmarks of an urban legend: a certain scientific plausibility ("Well, in principle I guess it's possible!") coupled with a deliciously shudder-inducing punch line ("We'd be ruled by robots!"). Did you know that if you sneeze, belch, and fart all at the same time, you die? Wow. Following in the wake of decades of AI hype, you might think the Singularity would be regarded as a parody, a joke, but it has proven to be a remarkably persuasive escalation. Add a few illustrious converts—Elon Musk, Stephen Hawking, and David Chalmers, among others—and how can we not take it seriously? Whether this stupendous event takes place ten or a hundred or a thousand years in the future, isn't it prudent to start planning now, setting up the necessary barricades and keeping our eyes peeled for harbingers of catastrophe?
I think, on the contrary, that these alarm calls distract us from a more pressing problem, an impending disaster that won't need any help from Moore's Law or further breakthroughs in theory to reach its much closer tipping point: after centuries of hard-won understanding of nature that now permits us, for the first time in history, to control many aspects of our destinies, we are on the verge of abdicating this control to artificial agents that can't think, prematurely putting civilization on auto-pilot. The process is insidious because each step of it makes good local sense, is an offer you can't refuse. You'd be a fool today to do large arithmetical calculations with pencil and paper when a hand calculator is much faster and almost perfectly reliable (don't forget about round-off error), and why memorize train timetables when they are instantly available on your smart phone? Leave the map-reading and navigation to your GPS system; it isn't conscious; it can't think in any meaningful sense, but it's much better than you are at keeping track of where you are and where you want to go.
Much farther up the staircase, doctors are becoming increasingly dependent