Machine learning and biological evolution have this in common | Stephen Wolfram

[00:00:04]Biology has traditionally not been a theoretical science.

[00:00:07]Biology has been a science where you just observe.

[00:00:10]This is how organisms work.

[00:00:12]This is how molecular biology works and so on.

[00:00:16]Now, when it comes to even biological evolution, it's not been completely clear why it works.

[00:00:21]Why doesn't it get stuck?

[00:00:22]Why can biological evolution go on and produce more and more elaborate forms?

[00:00:27]You know, Darwin, he thought that there would be some kind of abstract law or kind of law like the laws of physics that would determine the progress of biological evolution.

[00:00:39]But nobody found that.

[00:00:40]I came back to that a couple of years ago and tried to understand sort of in computational terms, how does biological evolution work?

[00:00:48]And I had one very important new data point, which was machine learning and the fact that machine learning works.

[00:00:59]You know, I had played around with neural nets and kind of the foundations of machine learning back in the early 1980s.

[00:01:05]I'd never managed to get neural nets to do anything interesting.

[00:01:08]What was discovered in the 2010s was the surprising fact that if you take a neural net and you just bash it really, really hard, eventually it will learn stuff.

[00:01:17]And so I thought, well, let's try that same idea for thinking about biological evolution.

[00:01:22]The biological evolution and machine learning turn out to be extremely related kinds of things.

[00:01:28]And so I did try that.

[00:01:30]And to my considerable surprise, yes, these simple computational systems could kind of evolve to get more and more complex forms, to achieve more and more elaborate fitness objectives, and so on.

[00:01:43]Whether it's the way that a mollusk shell kind of grows in a spiral, or whether it's the way the pigmentation on mollusk shells work, yes, it really is the case that there's this kind of thing that is probably the secret that nature uses to make all this complexity that it makes, which is that in the computational universe of possible programs, even a very simple program can produce extremely complicated behavior.

[00:02:07]And that kind of led me to this concept of computational irreducibility that's been an important one for a lot of what I've done. Kind of this question of when you have a simple program, if you want to know what it does, one thing you can do is just run it step by step.

[00:02:21]Let's say you want to know what it's going to do after a million steps, where you can run those million steps and see what it does.

[00:02:26]The question is, can you jump ahead and work out what it's going to do without having to go through all those steps?

[00:02:32]And that's the thing, you know, in the tradition of the exact sciences, the idea that you can predict things, that you can say, I don't need to follow a million orbits of the Earth around the sun or some idealized sun.

[00:02:44]can just use a formula and jump ahead and say what's going to happen.

[00:02:49]To know what the whole critter is going to do, you have to just follow those rules and see what happens.

[00:02:55]You can't say, oh, I know those rules, so therefore I know the critter is going to stick its head up at just this moment or something.

[00:03:03]That is the thing.

[00:03:04]Computational irreducibility is what provides a sort of irreducible gap between the underlying deterministic rules and the actual behavior of a system.

[00:03:13]Now, you know, there are many, many consequences of computational irreducibility.

[00:03:17]If you start saying, you know, I'm going to build an AI and I want it only to think good thoughts and do good things.

[00:03:23]Well, the problem is, as soon as you build an AI that is actually making sort of deep use of computation, it's going to have computational irreducibility.

[00:03:32]And it's going to have this feature that it can always surprise us.

[00:03:35]It can always do things that are the result where you can tell what it does by just following through the steps and seeing what it does.

[00:03:42]But you can never say, I know you're never going to do the wrong thing or whatever else.

[00:03:46]It's a trade-off, actually.

[00:03:48]It's something that I think will be a feature of sort of societal decisions, is do you go for computational reducibility or you go for computational irreducibility?

[00:03:59]We've had the experience sort of after the Industrial Revolution of having machines where we can kind of understand how they work.

[00:04:05]They've got gears and levers and things like this.

[00:04:08]Before the Industrial Revolution, lots of things we use, we didn't understand.

[00:04:12]You know, you ride a horse.

[00:04:13]The horse is, we know what we can do with the horse.

[00:04:16]We don't know how the horse works inside.

[00:04:19]And then, you know, post Industrial Revolution, we did have sort of a way of understanding how our machines work.

[00:04:24]But that's something as we get into this sort of domain of computational machines, that's no longer the case.

[00:04:31]And we can either say we insist on knowing how the machine works inside.

[00:04:35]It's got to be computationally reducible in its behavior.

[00:04:38]If it's computationally reducible, it will be very limited in what it can do.

[00:04:42]If we say, no, it can be computationally irreducible, then we can make use of its computational capabilities to the fullest extent.

[00:04:50]But then it has the problem that in principle, it can have things can happen, which will be surprising.

[00:04:55]to us and which we can't foresee in advance.

[00:04:58]And sort of one can ask, what does that mean for sort of us coexisting with the AIs and so on?

[00:05:03]I mean, I think already we're in a situation where in addition to human civilization, there's a civilization of the AIs.

[00:05:09]And the question is, what is it like to have a world in which there's this alien civilization right in front of us doing all these things?

[00:05:18]Well, actually, we have a very common experience of that, which is nature.

[00:05:22]Nature is, we can think of it also like a sort of alien civilization that's doing things, that's computing all these different kinds of things.

[00:05:30]We've learned to coexist with nature.

[00:05:32]You know, we build houses that prevent it, you know, problems when it rains and things like this.

[00:05:38]One of the questions in biology is, what's special about life?

[00:05:44]Even if you just take a piece of living tissue, what kind of a thing is that?

[00:05:49]Is a piece of living tissue liquid?

[00:05:52]Well, it's kind of gooey often.

[00:05:54]Is it solid?

[00:05:54]Well, it's kind of has some, you know, maintains its structure in some way.

[00:05:59]It's really not those things.

[00:06:01]When you look at it microscopically, sort of the big discovery of molecular biology, I suppose, in the last few decades has been that things are very orchestrated.

[00:06:10]Molecules are sort of specifically and actively transported from here to there.

[00:06:15]thing fits exactly into that, which then opens up to do this and so on.

[00:06:20]There's this question of, how are all these pieces kind of orchestrated together to do the things that happen in biology?

[00:06:28]And this notion of sort of bulk orchestration, that we're full of tons of molecules, but they're all doing things in this very kind of orchestrated way.

[00:06:37]That's a phenomenon that seems to be sort of an essential phenomenon of life.

[00:06:42]It's sort of the result of this big technology stack that's been built up through the course of biological evolution.

[00:06:48]The analogy is the organism is trying to do like build a wall.

[00:06:53]Well, if we were doing that by engineering, we would make bricks that are nice shapes and we would arrange them in some simple pattern.

[00:06:59]But what's happening in biology, and by the way, also in machine learning, is that one is picking up these kind of random lumps of irreducible computation.

[00:07:07]They're kind of like rocks lying around on the ground.

[00:07:10]And one's fitting those in and saying, well, this one happens to fit this way and this way. And eventually you build up this wall.

[00:07:16]And the reason that biological evolution works is that the fitness objectives that exist for biological organisms are computationally very simple compared to sort of the power of this underlying irreducible computation. I I mean, it's sort of unsurprising. If every organism, as soon as it was born, had to be able to solve some elaborate mathematical problem, no organisms would survive.

[00:07:36]It's because the sort of fitness objectives are computationally quite simple, particularly relative to sort of the power of these underlying computational elements, that biological evolution is possible and can work, so to speak.

[00:07:50]One might have thought that science, the universe, is sort of a cold, inhuman kind of place.

[00:07:58]And I think what has come out from the science I've done is that an awful lot of science reflects back on us humans in very important ways.

[00:08:07]In other words, there's in a sense nothing to say if there isn't a human somewhere in the middle.

[00:08:12]So when it comes to kind of when we talk about sort of AI and is there something sort of different and special about us humans, the answer is yes.

[00:08:21]The whole bundle of things that make up the human condition is unique.

[00:08:25]It is that whole bundle of things.

[00:08:27]And the AIs that don't have mortality or don't have certain kinds of sensory experiences or whatever, they are different in those ways from us humans.

[00:08:37]Now, interesting question for us as humans.

[00:08:41]You might say at some point, enough is enough.

[00:08:44]You know, we in our technology, what is technology?

[00:08:47]Technology is kind of taking what exists in the world and applying it for human purposes, finding that, you know, that magnetic material, we can use that to, you know, snap things together or make a compass.

[00:09:00]We can use those liquid crystals to make a display.

[00:09:03]We're taking things from the natural world and we're kind of applying them for human purposes.

[00:09:08]And this idea of computational irreducibility and so on tells us we're always going to be able to find more things that we can apply from the natural world.

[00:09:17]And the question is, well, at what point is sort of are we done?

[00:09:21]At what point is, you know, between our computational systems and our AIs and our robotics and all that kind of thing, at what point have we got everything that we need to have?

[00:09:33]I don't think that's the nature of us as biological organisms.

[00:09:38]I think we are, to some extent, we have the vestiges of natural selection, of the sort of struggle for life over the last three billion years.

[00:09:48]We are continually kind of seeking the new, that's been the experience to this point.

[00:09:55]So, you know, the idea that kind of we're done now is unlikely to be what will happen.