eliezer_s_yudkowsky's picture
Research Fellow; Co-founder, Machine Intelligence Research Institute
The Value Loading Problem


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.