Morphogenesis for the Design of Design

Morphogenesis for the Design of Design

Neil Gershenfeld [7.31.19]

As we work on the self-reproducing assembler, and writing software that looks like hardware that respects geometry, they meet in morphogenesis. This is the thing I’m most excited about right now: the design of design. Your genome doesn’t store anywhere that you have five fingers. It stores a developmental program, and when you run it, you get five fingers. It’s one of the oldest parts of the genome. Hox genes are an example. It’s essentially the only part of the genome where the spatial order matters. It gets read off as a program, and the program never represents the physical thing it’s constructing. The morphogenes are a program that specifies morphogens that do things like climb gradients and symmetry break; it never represents the thing it’s constructing, but the morphogens then following the morphogenes give rise to you.

What’s going on in morphogenesis, in part, is compression. A billion bases can specify a trillion cells, but the more interesting thing that’s going on is almost anything you perturb in the genome is either inconsequential or fatal. The morphogenes are a curated search space where rearranging them is interesting—you go from gills to wings to flippers. The heart of success in machine learning, however you represent it, is function representation. The real progress in machine learning is learning representation. How you search hasn’t changed all that much, but how you represent search has. These morphogenes are a beautiful way to represent design. Technology today doesn’t do it. Technology today generally doesn’t distinguish genotype and phenotype in the sense that you explicitly represent what you’re designing. In morphogenesis, you never represent the thing you’re designing; it's done in a beautifully abstract way. For these self-reproducing assemblers, what we’re building is morphogenesis for the design of design. Rather than a combinatorial search over billions of degrees of freedom, you search over these developmental programs. This is one of the core research questions we’re looking at.

NEIL GERSHENFELD is the director of MIT’s Center for Bits and Atoms; founder of the global fab lab network; the author of FAB; and co-author (with Alan Gershenfeld & Joel Cutcher-Gershenfeld) of Designing Reality. Neil Gershenfeld's Edge Bio Page


NEIL GERSHENFELD: I’d like to end this interesting long day by explaining why I think computer science was one of the worst things ever to happen to computers or science, why I believe that, and what that leads me to. I believe that because it’s fundamentally unphysical. It’s based on maintaining a fiction that digital isn’t physical and happens in a disconnected virtual world.

One of my students built and runs all the computers Facebook runs on, one of my students used to run all the computers Twitter runs on—this is because I taught them to not believe in computer science. In other words, their job is to take billions of dollars, hundreds of megawatts, and tons of mass, and make information while also not believing that the digital is abstracted from the physical. Some of the other things that have come out from this lineage were the first quantum computations, or microfluidic computing, or part of creating some of the first minimal cells.

Stephen made the observation that we’re surrounded by computation, most of which we don’t use. This is what leads me to wanting a do-over. I view the current state of computer science as a bit like Metropolis, where it's training people to frolic in the garden while somebody in the basement moves the levers. What I want to talk about is how you bring them together.

First of all, I’ve come to the conclusion that this is a historical accident. I could ask Marvin what John von Neumann was thinking, and I could ask Andy Gleason what Turing was thinking, and neither of them intended us to be living in these channels. Von Neumann wrote beautifully about many things, but computer architecture wasn’t one of them. We’ve been living with the legacy of the EDVAC and the machines around us, and much of the work of computers is not computationally useful because it’s just shuttling stuff. The Turing machine was never meant to be an architecture. In fact, I'd argue it has a very fundamental mistake, which is that the head is distinct from the tape. And the notion that the head is distinct from the tape—meaning, persistence of tape is different from interaction—has persisted. The computer in front of Rod Brooks here is spending about half of its work just shuttling from the tape to the head and back again.

There’s a whole parallel history of computing, from Maxwell to Boltzmann to Szilard to Landauer to Bennett, where you represent computation with physical resources. You don’t pretend digital is separate from physical. Computation has physical resources. It has all sorts of opportunities, and getting that wrong leads to a number of false dichotomies that I want to talk through now. One false dichotomy is that in computer science you’re taught many different models of computation and adherence, and there’s a whole taxonomy of them. In physics there’s only one model of computation: A patch of space occupies space, it takes time to transit, it stores state, and states interact—that’s what the universe does. Anything other than that model of computation is physics and you need epicycles to maintain the fiction, and in many ways that fiction is now breaking.

I’ve been working with people on exascale computer architecture, the biggest super computer architecture. If you look at what it costs to move data to memory, what it costs to do interconnect, and what it costs to have all the processors working usefully, all of those things are breaking. We did a study for DARPA of what would happen if you rewrote from scratch a computer software and hardware so that you represented space and time physically. So, if you zoom from a transistor up to an application, you change representations—completely unrelated ones—about five different times. If you zoom the building we’re in from city, state, country, it’s hierarchical, but you respect the geometry. It turns out you can do that to make computer architectures where software and hardware are aligned and not in disconnected worlds. One of the places that I’ve been involved in pushing that is in exascale high-performance computing architecture, really just a fundamental do-over to make software look like hardware and not to be in an abstracted world.

Right now, we’re in deep-learning mania as one of the things pushing computing. Depending on how you count, this is now the fifth boom-bust cycle. From a distance it looks like we’re now in a boom cycle. This is the good thing. A quiet trend that’s been emerging is that scaling is what's driving the current AI boom—networks gathering more data, bigger memories storing the data, more processing cycles. It's a quiet, really interesting trend as it turns out. Most of what’s getting the attention on the deep-learning architectures don’t matter much. Many different approaches work equally well. There’s nothing magic about the deep-learning architectures. The magic is there’s more data with more memory with more cycles. It’s a cargo cult, the obsession with the acronym zoo of deep learning. It’s just an exercise in scaling that’s been making that possible.

Analog versus digital are not two distinct choices where you can pick one or the other. What’s interesting is what lies between them. As an example, my lab spun off a chip company that uses analog degrees of freedom to solve digital problems. A digital system lives on the corner of a hypercube, but what we did in that chip company was use the analog device degrees of freedom to go through the interior of the hypercube, not to stay on the corners. It saves power and speeds and has all these performance benefits.

That’s not a new idea in the context of optimization. Like many of the largest scale computations, what's used are things called interior point methods, or relaxations, where you have a discrete answer you want—like routing an airplane or which way to turn a car—but the way you get through it is to relax the discrete constraints and use internal degrees of freedom. These interior point methods are the most important algorithms for solving large-scale computational problems. If you just took one of my chips doing a physical version of this, a neurobiologist would have absolutely no idea what was going on in it, but it would make perfect sense in an introductory optimization class.

Digital isn’t ones and zeroes. One of the hearts of what Shannon did is threshold theorems. A threshold theorem says I can talk to you as a wave form or as a symbol. If I talk to you as a symbol, if the noise is above a threshold, you’re guaranteed to decode it wrong; if the noise is below a threshold, for a linear increase in the physical resources representing the symbol there’s an exponential reduction in the fidelity to decode it. That exponential scaling means unreliable devices can operate reliably.

The real meaning of digital is that scaling property. But the scaling property isn’t one and zero; it’s the states in the system. In the end, what these interior point and relaxation methods do is drive to an outcome that’s a discrete state, but you pass through continuous degrees of freedom. It’s very naïve to say digital is ones and zeroes. It’s state restoration, but you can use continuous degrees of freedom. In many different areas this is done to do the state restoration.

Now: threshold theorems. It was first proved by Shannon. Von Neumann applied Shannon to computing to show how reliable computers can operate with unreliable devices, but the thing that excites me is threshold theorems were invented four billion years ago, which is the evolutionary age of the ribosome. The connection there is if you mix chemicals and make a chemical reaction, a yield of a part per 100 is good. When the ribosome—the molecular assembler that makes your proteins—elongates, it makes an error of one in 104. When DNA replicates, it adds one extra error-correction step, and that makes an error in 10-8, and that’s exactly the scaling of threshold theorem. The exponential complexity that makes you possible is by error detection and correction in your construction. It’s everything Shannon and von Neumann taught us about codes and reconstruction, but it’s now doing it in physical systems.

One of the projects I’m working on in my lab that I’m most excited about is making an assembler that can assemble assemblers from the parts that it’s assembling—a self-reproducing machine. What it's based on is us. We're made from 20 parts, amino acids, and what’s interesting about amino acids is they’re not interesting. They have simple properties like hydrophobic and hydrophilic and basic and acidic, but you can compose them to make muscles and motors and sensors. In the same way, we’re taking 20 inorganic properties like conducting and insulating to show you can compose them hierarchically. In fact, the project funding was a proposal to the DoD to reduce their whole supply chain to 20 parts, these fundamental building blocks, and they’re based on digitizing the materials.

Compare state of the art manufacturing with a Lego brick or a ribosome: When a kid plays with Lego, you don’t need a ruler because the metrology comes from the parts. It's the same thing for the amino acids. The Lego tower is more accurate than the motor control of the child because you detect and correct errors in their construction. It’s the same thing with the amino acid. There’s no trash with Lego because there’s information in the construction that lets you deconstruct it and use it again. It’s the same thing with the amino acids. It’s everything we understand as digital, but now the digital is in the construction. It’s digitizing the materials. The core project of assembling an assembler is, in part, a paradigmatic challenge. If you look at scaling coding construction by assembly, ribosomes are slow—they run at one hertz, one amino acid a second—but a cell can have a million, and you can have a trillion cells. As you were sitting here listening, you’re placing 1018 parts a second, and it’s because you can ring up this capacity of assembling assemblers. The heart of the project is the exponential scaling of self-reproducing assemblers.

As we work on the self-reproducing assembler, and writing software that looks like hardware that respects geometry, they meet in morphogenesis. This is the thing I’m most excited about right now: the design of design. Your genome doesn’t store anywhere that you have five fingers. It stores a developmental program, and when you run it, you get five fingers. It’s one of the oldest parts of the genome. Hox genes are an example. It’s essentially the only part of the genome where the spatial order matters. It gets read off as a program, and the program never represents the physical thing it’s constructing. The morphogenes are a program that specifies morphogens that do things like climb gradients and symmetry break; it never represents the thing it’s constructing, but the morphogens then following the morphogenes give rise to you.

What’s going on in morphogenesis, in part, is compression. A billion bases can specify a trillion cells, but the more interesting thing that’s going on is almost anything you perturb in the genome is either inconsequential or fatal. The morphogenes are a curated search space where rearranging them is interesting—you go from gills to wings to flippers. The heart of success in machine learning, however you represent it, is function representation. The real progress in machine learning is learning representation. How you search hasn’t changed all that much, but how you represent search has. These morphogenes are a beautiful way to represent design. Technology today doesn’t do it. Technology today generally doesn’t distinguish genotype and phenotype in the sense that you explicitly represent what you’re designing. In morphogenesis, you never represent the thing you’re designing; it's done in a beautifully abstract way. For these self-reproducing assemblers, what we’re building is morphogenesis for the design of design. Rather than a combinatorial search over billions of degrees of freedom, you search over these developmental programs. This is one of the core research questions we’re looking at.

I started off this diatribe by complaining about computer science, but von Neumann and Turing ended exactly here. The last thing von Neumann worked on—and this is something he wrote beautifully about—was self-reproducing machines. If you’ve ever read it, his memo on the EDVAC is a mess. The programming manual where the von Neumann architecture emerged is a dreadful document. It’s a mess. What he wrote about self-reproducing machines was exquisite. It’s a beautiful posthumous document asking how a thing can communicate a computation for its construction, how to abstract a self-reproducing thing. He was asking it to get at the heart of what is life. It was a theoretical thing at that time. That’s what he ended his life doing. The last thing Turing ended his life doing was studying morphogenesis. What it’s casually known for is Turing spots and patterns, but that was the detail. What he was really asking was bits from atoms or atoms from bits. He was asking, how do genes give rise to us?

Looking at exactly this question of how a code and a gene give rise to form. Turing and von Neumann both completely understood that the interesting place in computation is how computation becomes physical, how it becomes embodied and how you represent it. That’s where they both ended their life. That’s neglected in the canon of computing, but we’re now at this interesting point where I’m on the hook to deliver on a research program to make a self-reproducing von Neumann assembler. We can think about making these things now, of embodying it. It is a third digital revolution. There is communication, then computation, now fabrication. It’s not a separate one, but it merges them because it merges them in a thing that communicates its construction to fabricate.

At MIT the first real-time computer was the Whirlwind. Then came the PDP as the mini computers, and there were thousands of those. Then came the hobbyist computers like the Altair, and there are millions of those. Then came the personal computers and smart phones, and there are billions of those. Now, there are the Internet of Things devices, and there are trillions of those.

The Nest thermostat, roughly, has the capacity of the PDP—computing scale from one to a thousand to a million to a billion to a trillion. You could see all of that lurking in 1965 when Gordon Moore made his first plot of Moore’s law that scaled for fifty years. In the same way, if you take digital fabrication, it’s been scaling for about a decade in the same way. You can make a Moore’s law-like plot for performance and scaling of digital fabrication, and there’s a close historical parallel.

MIT made the first NC mill in 1952. That’s like the mainframe. For NSF I started setting up FAB labs, which are mini versions of the big lab I run. With current digital fab tools, they would fit in a room like this—and that’s like the PDP version. There’s a thousand of those today. We’re using those to make machines that make machines, not self-reproducing assemblers but rapid prototyping tools that make rapid prototyping tools, and that’s moving towards a million of them.

In the lab, we’re developing these assemblers that I described and then working toward the self-assemblers. All those things exist in some form today, but they’re going to be emerging between now and fifty years from now, but you can see the thousand, million, billion, trillion scaling happening for digital fabrication.

We're at an interesting point now where it makes as much sense to take seriously that scaling as it did to take Moore’s law scaling in 1965 when he made his first graph. We started doing these FAB labs just as outreach for NSF, and then they went viral, and they let ordinary people go from consumers to producers. It’s leading to very fundamental things about what is work, what is money, what is an economy, what is consumption.

There’s legislation in the Senate and House right now for universal access to digital fabrication, like there was for communication and computation. We're also working with Bhutan’s prime minister—the country is based on gross national happiness, but they buy crap trucked in from India—on how to make gross national happiness physical.

We’re working with a number of cities around the world that have failed economies on how to turn consumption into creation. In the same way that the Internet emerged in the mini computer era, this fifty-year scaling of digital fabrication is emerging today, and the equivalent of "how does the Internet work?" is growing up around it. A surprising fraction of my time has just gone into working with all these governments and organizations and social groups on if anybody can make anything anywhere, how does that reinvent societies and economies?  

I started with complaining that computer science was the worst thing to happen to computers or science because it’s unphysical, and pointed out that you can have a do-over of computer science that’s much more aligned with physics. It has all kinds of benefits ranging from computing with very different physical systems to limits of high-performance computing but, ultimately, reuniting computer science and physical science leads to merging the bits and atoms. Fabrication merges with communication and computation. Most fundamentally, it leads to things like morphogenesis and self-reproducing an assembler. Most practically, it leads to almost anybody can make almost anything, which is one of the most disruptive things I know happening right now. Think about this range I talked about as for computing the thousand, million, billion, trillion now happening for the physical world, it's all here today but coming out on many different link scales.

The last time we gathered, there was a suggestion to turn it into a book, which was a lovely exercise. Coming here, John asked me what I thought we should do coming out from this. I had three suggestions that he thought were all horrible, so I’ll end with those. The baseline is we have a lovely weekend, we admire each other, and then we go home. So that’s the default.

One suggestion I have comes from a conversation with my younger brother who led the biggest video game studio, Activision, and he was horrified when he discovered when you write a book it’s good if it sells thousands of copies. He’s used to selling tens of millions of whatever he does. He left Activision. He now has a company that does games for education and social change. The most recent one they did that got a lot of attention was with Alaska native storytellers. There are great traditions, but terrible alcoholism, and suicide, and unemployment, and they worked with Alaska native storytellers to tell narratives in immersive video game experiences. There’s a whole bunch of examples like that. One suggestion I had John hated was we build the world we’re describing as an immersive experience and get it in the hands of millions of people.

I had done a number of friend-of-friend movie advising in Hollywood, and that led through a collaboration where I helped start an office called the Science Entertainment Exchange, which hijacks popular media. It takes movies and TV shows and uses them as covers to put in science teaching, and it’s been working really well embedding science in all kinds of popular shows. The second idea I had was we take everything we’re trying to do and embed it in the popular conscious by hijacking some movies or TV shows.

Then the third one has been working with some interesting groups that put together bit stadium shows, and so this has been lovely, but it’s just for us. We do this on an epic scale was a third suggestion. Those are the three ideas John thought were terrible that I'll conclude with, so now I'll step back and open that for discussion.

* * * *

PETER GALISON: I wonder if we could discuss whether there’s something different about the biological case than, say, the physical properties that lead to snowflakes or crystals. That is to say, elementary atomic forces don’t have encoded this complicated hexagonal form, but you get there. They just make local decisions, and the local decisions add up like Legos to something else. My question is about the physicalization or the embodiment of computation.

I can think of several reasons why you might want to shorten or eliminate the gap between software and hardware. One might be that there’s an aesthetic objection that something’s wrong with hardware that is disjunct from the way we represent it. There are other things that we do in our representation where they’re not matched; for example, differential equations don’t look like the things that they’re often representing. Or another might be efficiency, that if we could somehow have software that matched the physicality of, say, atoms and bits, it would run without the frictional loss of computing power in our everyday devices. Another might be that there are more and more cases where the software is embedded in the hardware itself. If you dig into your Intel chip, there’s a lot of software in them before you get to high-level programming. Suppose we agree that there is this gap between the representation and the things represented, what is it that propels you?

GERSHENFELD: For your first passing point about the snowflake, I’ll make a passing point. The work I’m describing on coding assembly of digital materials isn’t a single-length scale. We’re doing that in molecular biology when we make synthetic cells. We’re doing nanofab to make nanostructures. We’re micromachining microstructures up to where we’re working with Airbus and robots to make jumbo jets and NASA to make spaceships on big scales. What I spoke about isn’t a single-length scale. It’s better to think about it as the dynamic range between the smallest feature you need to control and the size of the system.

Why align computer science and physical science? There are at least five reasons for me. Only lightly is it philosophical. It’s the cracks in the matrix. The matrix is cracking. 1) The fact that whoever has their laptop open is spending about half of its resources shuttling information from memory transistors to processor transistors even though the memory transistors have the same computational power as the processor transistors is a bad legacy of the EDVAC. It’s a bit annoying for the computer, but when you get to things like an exascale supercomputer, it breaks. You just can’t maintain the fiction as you push the scaling. The resource in very largescale computing is maintaining the fiction so the programmers can pretend it’s not true is getting just so painful you need to redo it. In fact, if you look down in the trenches, things like emerging ways to do very largescale GPU program are beginning to inch in that direction. So, it’s breaking in performance.

2) We're just wasting resources. When you look at what’s going on in your Intel chip, it’s right at the edge of analog. They do a lot of work. Inside it’s awfully analog but ends up looking digital on the outside. We’re wasting a lot of the computational power of the transistor. With the chip fab I mentioned, we’re wasting degrees of freedom in the devices that aren’t a simple version of analog versus digital. You can solve digital problems, but by using the analog degrees of freedom, you win speed, power, performance, and all kinds of good stuff.

3) When we were in the early days of quantum computing or the stuff we did on microfluidic logic, you’re computing with fundamentally different physical resources where you need to represent the computation in a way that can describe the physics that you’re working with.

4) The final reason goes back to where von Neumann ended up. When I make this self-reproducing assembler in the very short term, I’m using conventional computer architectures for the intelligence of it, but what I need to do is overlay the computation as geometry. If I’m doing morphogenesis with a self-reproducing system, I don’t want to then just paste in some lines of code. The computation is part of the construction of the object. I need to represent the computation in the construction, so it forces you to be able to overlay geometry with construction.

There are all different reasons, but they all lead you to the same place. Interestingly, for the do-over I mentioned in DARPA, we took the BLAS, which are the routines that underlie high-performance computing, and we rewrote them in a geometrical spatial computing model. What’s interesting is a lot of the things that are hard—for example, in parallelization and synchronization—come for free. By representing time and space explicitly, you don’t need to do the annoying things like thread synchronization and all the stuff that goes into parallel programming.

DAVID CHALMERS: What you’re saying is, when thinking about software, hardware and physics matter. In some sense everyone has known all along that the hardware matters and the physics matters, and chip makers and everyone else under the sun has been thinking about how to do the best computing you can given the limitations that you have about technology and the resources of physics. One thing you’re saying is that we haven’t done everything we can to take advantage of the hardware possibilities, so we’ve got to push the project harder and faster. Does it go beyond that? The part about fabrication and self-assembly is fundamentally new and different.

GERSHENFELD: Let me help you connect those parts. Communication degraded with distance. Along came Shannon. We now have the Internet. Computation degraded with time. The last great analog computer work was Vannevar Bush's differential analyzer. One of the students working on it was Shannon. He was so annoyed that he invented our modern digital notions in his Master’s thesis to get over the experience of working on the differential analyzer.

Today, in this computer, it’s head-bangingly stupid what’s going on with this accidental legacy of von Neumann architecture. He never talked about the von Neumann architecture long past its due date. Much of the resources are shuttling information from memory transistors to processor transistors, wasting the power of all of this, and then the utilization of it is more inefficient still when you go from the software compilation to the hardware.

So, one of the points was just it’s very inefficient. It doesn’t matter if you’re doing word processing; it does matter if you’re pushing limits of computing performance. So, very low power or very high power, you care about that.

CHALMERS: If it’s so head-bangingly stupid, why didn’t someone from Intel figure this out years ago?

GERSHENFELD: What's interesting is that there’s a whole parallel history. We’ve been lulled into sleep by Gordon Moore. I spent some time with Gordon Moore in the early days of this fabrication scaling I was mentioning, and he was amused by the parallel with what he did at that time. It’s like the matrix. We had a few decades where we could pretend that nobody’s moving the levers in the basement and we can frolic in the garden. There’s been a parallel history all the way through it. It passes through people like Danny. There are a number of device physics. There’s a whole parallel history building this, but you could ignore it. Again, limits of either high performance, low power are pushing it.

I started by mentioning my students who built the computers for Facebook and Twitter, and they’re not doing this at the fundamental physics level, but they had to completely re-architecture how you build a data center with coarse-grain versions of it. You don’t see it, but it percolates in things like how Jason built the Facebook data center. Just to recap the answer, you need to do what I’m describing if you don’t compute with Intel. So, the stuff we did on quantum computing or fluidic or molecular computing, you need to revisit these assumptions.

If you are confused by everything I say, and you take a single thing away, it's the last part I talked about, about digitizing fabrication. It's not about computing and then there’s this other thing here, but it’s the synthesis. When you merge communication with computation with fabrication, it’s not there’s a duopoly of communication and computation and then over here is manufacturing; they all belong together. The heart of how we work is this trinity of communication plus computation and fabrication, and for me the real point is merging them.

W. DANIEL HILLIS: I was going to give just a very specific small example that supports the abstraction that you’re saying. In modern ways of analyzing algorithms, and computers, and the computer science, we count the cost of moving a bit in time. We call that storage, and that’s very carefully measured in the algorithms and things like that. The cost of moving a bit in space is completely invisible, and it just doesn’t come up. There’s no measure of that in the way that we abstract it, but if you look at the megawatts that are dissipated in high-performance computers, it mostly comes from moving bits in space. That’s the big limitation, and that’s also where the errors are and where the cost is. So, our abstraction that we’re thinking about the algorithms in is completely out of sync with where our costs are.

CHALMERS: You mean that hardware makers have not been thinking about those costs of moving bits in space?

GERSHENFELD: One more example of the cracks in the matrix is, every few months there’s a headline about a new security vulnerability, and an awful lot of them have to do with things that are supposed to be far away in computation space colliding in physical space, because there’s no way to say things that are far apart computationally should be far apart physically. 

I’ve spent time with the people after Gordon who ran Moore’s law at Intel, the keepers of Moore’s law, and one of the most evocative images came from one of them describing his job as the scene in the Indiana Jones movie when the boulder is running down. All he can do is not get run over by the boulder. They’re running this multibillion-dollar oil tanker, and it’s hard to steer. They have to make sure the boulder doesn’t run over them.

I almost took over running research at Intel. It ended up being a bad idea on both sides, but when I was talking to them about it, I was warned off. It was like the godfather: "You can do that other stuff, but don’t you dare mess with the mainline architecture." We weren't allowed to even think about that. In defense of them, it’s billions and billions of dollars investment. It was a good multi-decade reign. They just weren’t able to do it. 

SETH LLOYD: Maybe we take what Frank and others have been saying about the power of the brain and ask what we would need in a computational device to do that. The brain has, we were saying, 1011 neurons, around 1015 connections, and it operates at the 100-hertz scale. Suppose you wanted to get a silicon device that had similar scale. If the size of the objects were a nanometer and you weren’t worrying about the wiring, you would have to have about one electron per transistor. You’d have to go down to single-electron transistors. This device would be tremendously noisy, the problem of moving information around. If you want to get to the kind of information processing that human beings and other animals have, you would need to go far beyond the paradigms that people have: dealing with noisy computation, making it analog, mapping the way the physical processing is going on onto a chip in a way that’s very different from the way that people do the architecture right now, doing things massively in parallel.

If you wish to fulfill the promise of Moore’s law to get artificial intelligences that are similar in scale to human beings, you’ve got to do something quite different.

GERSHENFELD: Analog doesn’t mean analog. In other words, analog in this context means you have states, and you recover from errors, and you detect states. But states are outcomes of the system, they're not ones and zeroes. One of the things we’re stuck in is this idea that a state is one and a zero. This device in front of me keeps recurring the state not at the high-level thing I’m trying to do, but at the ones and zeroes.

These interior point relaxation methods I was describing both in software optimization and in emerging chips do digitize, but they’re digitizing on high-level outcomes but using the analog degrees of freedom. That was behind my comment that when the brain does a few moves a second, it’s moving through this very high-dimensional space, ending into a discrete outcome. So, the effective number of operations that are done this way is an enormous number.

TOM GRIFFITHS: I wanted to return to Rod’s talk, asking whether any of the things you learn about when you're thinking about scaling should inform the way that we think about neuroscience in terms of getting at some of the inadequacies of classic models of computation for neuroscientists.

GERSHENFELD: I was recently at a retreat of many of the leading neuroscientists for a review of the state of the art of the field, and boy I was horrified. They were horrified. The state of the art of neuroscience is like you throw the watch at the wall and you see the parts that come out. We had a lively discussion about the devices I’m building and the algorithms we’re using they would be completely stumped by. They would have absolutely no idea how to recognize that was going on. We don’t have an easy next step after that, but there’s an interesting dialogue with the neuroscientists about it.

ALISON GOPNIK: There is something that’s a bit puzzling about this, which is that you have these incredibly complex devices—brains—and they can be translated into a bunch of symbols on a piece of paper or a bunch of simple digitally described symbols in a language, and that seems to be able to do a lot of work for human beings. Arguably, a lot of the capacities for intelligence that we have come from things like being able to talk to one another, or write, or use symbols in these ways that from a hardware perspective are completely trivial.

I’m not being disingenuous about this. This is a real puzzle, and in some ways what Turing is modeling, what he’s starting out with when he’s thinking about the computer who’s sitting there in Bletchley Park is not anything like this tiny bit of complexity compared to the complexity of what’s going on underneath the hood. It’s puzzling to me about what the relationship is between those two things.

GERSHENFELD: One interesting group I worked with was at Wright-Patterson Air Force Base where among the most sensory overloaded tasks are fighter pilots, and so they wanted to make planes you could fly by thinking. What came out of that after a lot of work is that it's a terrible idea. The reason is, with a lot of work to pull a lot of signals out and do a lot of interpretation, you can barely control anything because all of this just isn’t the right representation. All of this is designed so that this moves and this moves, and the best way to interface with this is to move your fingers. So, this representation is an internal one and then this is an external one.

GOPNIK: It seems to me like it’s an incredibly interesting understudied fact that what this all ends up driving is a bunch of fingers and your larynx. This tiny system with tiny degrees of freedom and very little complexity is the thing that’s doing the work that we think of as being a lot of the work of intelligence.

GERSHENFELD: But again, these kinds of relaxation interior point methods that I keep alluding to, there’s something similar to them in that they’re moving through these billion-dimensional spaces, but what they’re putting outside is not the interior point but statistics of the states that they’re getting driven to. So, there are analogs between unpacking the huge number of internal degrees of freedom versus small numbers of observable degrees of freedom in these engineered systems.

CHALMERS: The brain also has these amazing hardware inefficiencies in it, which are analogous to your hardware cases, like the fact that it uses electrical transmission within neurons, but between cells it’s chemical transmission. So, I guess the brain just got locked into that the way Intel got locked in years ago, and then it couldn’t escape the boulder fast enough.

GERSHENFELD: That’s true. Again, the embodiment of everything we’re talking about, for me, is the morphogenes—the way evolution searches for design by coding for construction. And they’re the oldest part of the genome. They were invented a very long time ago and nobody has messed with them since.

LLOYD: I disagree with that about the brain. The electrical signals use a lot more power, but they go fast and they go a long distance. The synaptic connections, of which there are thousands more, use much less power. I’m talking about just energy, but they go over a very tiny distance and they only use a few hundred molecules. So, it’s pretty efficient.

CAROLINE JONES: They’re chemical, and there’s a kind of redundancy and robustness in those separate things. There’s also a different system of feedback, which is fascinating. The chemicals are regulated by completely different body systems, which allows for all different kinds of intelligence to overlap and reinforce each other.

GOPNIK: It’s worth pointing out that plasticity is expensive. This is one of my favorite factoids: Everyone knows brains are taking about 20 percent of calories. If you look at four-year-olds, it’s 66 to 70 percent of calories are getting used up by brains. It’s not so much that they’re doing the computations, but they’re establishing what the wiring looks like.

GERSHENFELD: I worked with an IBM largescale computer architect on a project to make a computer that can physically remodify itself—taking the kind of assembler I’m describing to make a computer that can rebuild its construction. We’re still discussing that and working on it, but he told me something interesting. They did an early crude version of that, and what they discovered was the computer got configured but never reconfigured, which is very analogous to learning. The configurability was used to adapt the computer to the workload, but they never went back to change it. So, that led us to look at not reconfigurable but just configurable computers, like computers that can build themselves but don’t necessarily need to unbuild themselves.

Get over digital and physical are separate; they can be united. Get over analog as separate from digital; there’s a really profound place in between. We’re at the beginning of fifty years of Moore’s law but for the physical world. We didn’t talk much about it, but it has the biggest impact of anything I know if anybody can make anything.

I’ll leave you with my three questions that John doesn’t like. Do you want to make a video game for millions of people to live in the world we’re in? By the way, I did one of these. It’s fun to build the world you’re trying to create. Do you want to portray it on a large scale? Do you want to do what we’re doing here on a large scale? Any of those have great teams that could help with it rather than just doing a book next.