進化は「ランダム探索」ではない:Akarsh Kumar が語る人工生命・ASAL・Core War

I just want to clarify, evolution is anything but random. So artificial life is really about, like, understanding, like you said, life as it could be, rather than life as it is. We need to understand the space of all possible intelligences, the space of all possible life forms that can be. Conway's Game of Life is a super famous simulation by John Conway where you have this grid of cells, this 2D grid of cells, and they can either be on or off, alive or dead. That's it. And from this super, super simple world, it's like possibly the simplest possible thing you can think of, you run this, you get insane amounts of complexity that no one would have thought it could do until you actually run it. You can do counterfactual analysis in other universes. In what worlds do we get this open-ended creation of complexity and emergence and in what worlds do we not get that? Why does this physics, when you run it,

it produces this amazing phenomenon and when you run this physics, it doesn't do that. It produces just random noise. Instead of hard coding a single simulation, you parameterize a space of simulations, which is much, much easier. Just simulate it, see what happens, plug it into a foundation model and see what the foundation model thinks about what happened. There's 260,000 rules. We plotted all of them. So we found that there's literally like a big island of solutions which are not open-ended, and there's a small island of solutions where all of the coolest simulations lie. Some of them actually end up creating little patterns which move around that look like cells. This object, this cell-like object, is exploiting the physics in a way to create a persisting object in this world, right? If you train them on calculus, they'll learn calculus. You don't have to train them on arithmetic first. And it's such a weird, alien way that we train these models. I really like to call them a statistical intelligence. And they're, like, basically perfect statistically.

But this FER idea and this pick-breeder line of work, it's basically like a different paradigm of intelligence that we're trying to study. So it's the difference between, like, a statistical intelligence and a regularity-based intelligence. And you can see this in, like, the pick-breeder skulls. It's in the solutions, like the neural circuits that we find. the regularities are baked into the circuits. We wanted to see what would happen if you used LLMs to evolve basically small little assembly programs that compete against each other. So there's this programming game invented in like the 1980s that people used to play. It's called Core War. What my program needs to do in order to do that is it needs to basically inject an invalid operation in front of your program's execution thread. It's going to halt your program by crashing it. program will crash and can we create like this open-ended evolutionary arms race of programs from this it's like game of thrones cheering edition yeah yeah exactly that's the right way to put it the llm is not good at red code if you just try to zero shot this language it's terrible

but if you do this evolution and you have this verifier it turns out to be really good evolution is probably the most powerful algorithm that we know to solve like discrete combinatorial problems. If you solve, like, even one part of a puzzle, and you hang on to that, and you search for the other ones, you can turn your exponential search problem into a linear search problem. We're basically trying to search for cellular automata, like neural cellular automata, which go from self-replicating molecules, to cells floating in primordial soup, to predator-prey dynamics, to alien animals roaming around. You can imagine that if you did this with a big particle simulation with like quadrillions of particles and you had the right foundation model and you searched for this sequence of events then what would happen maybe you end up with some physics rule for this big particle system which really does give you this entire sequence you get animals at them just like what happened in our universe right the field of artificial life i really think is like a very long-term bet on intelligence

the paper is ASIL automating the search for artificial life with foundation models Akash and his co-authors use vision language models to find interesting simulations across game of life linear boids and more early on in the paper you said you quoted artificial life is not just as we know it but also as it can be what do you mean by that so artificial life is really about like understanding like you said life as it could be rather than life as it is if we really just wanted to understand life on earth and like how it is right now we'd be studying biology right it's the same reason that we don't just study neuroscience to understand intelligence we're create ais because we want a more mechanistic understanding and not only that But me personally, I like to I'm going to borrow this quote from Jeff Klune that I liked a lot. We really want to understand if we want to claim that we understand intelligence or we understand life.

We can't just talk about this specific form of life. We need to understand the space of all possible intelligences, the space of all possible life forms that can be. And only then can you talk about general principles for what governs intelligence or what governs life. Right. You can't just look at one car and be like, this is how all cars are. right that's the same way you can't just look at gpt5 and be like all intelligences are like this you can't just look at the human brain and be like all intelligences are like this there's a space of things and we need to study the whole space if you look at how like humans learn and like animals they have like a natural curriculum they're like i'm gonna learn this and then this and then this there's no other way but if you look at how like our models are trained you can train them on anything and it'll just work and the question is a lot of people think that that's fine and that Maybe it's better than how humans learn. And that's a possibility. It could just be better, right? But there's reasons to believe that there is something fundamentally wrong, and the universe doesn't work like that. The universe doesn't work like you can just learn anything at any given time.

And thus, I think the fact that humans learn in this sequential order, and evolution works in this path-dependent way, is the reason that it creates such nice, regular structures in the representations that it uses. And that's the reason it's so robust, adaptable, generalizable. is for those reasons, I think. We see this cast out in the form of curriculum learning, which is this idea that we learn step-by-step and we learn in kind of with an increasing complexity, right? And you were just pointing to when you learn step-by-step, increasing the complexity, the representations you learn are better. But how would you describe that in a kind of no-nonsense way? What's the intuition? The curriculum is not the only thing that matters because people have been doing curriculum learning in RL and I was obsessed with curriculum learning at RL. I think the key thing is that, I mean, this goes back to the why greatness cannot be planned lesson. You need like a serendipitous curriculum, which is not a pre-planned, like I'm going to do this, then this and this. It has to be much more serendipitous and embrace the chaos of our nature,

of the universe, right? It has to embrace this. So you have to embrace this idea that not just like going down this curriculum, but any possible curriculum, which might lead me to somewhere interesting. So that's just a shout out to like you need serendipitous curriculums. and to answer the question of why curriculums actually end up mattering, I think that's a fundamental question that I'm really interested in answering. And I don't know the short-form answer for that, but it's like a fundamental property of our universe that if you want to capture, as you like to say, the carve of the universe and understand it by the joints, right? That you carve it up by the joints. I think in order to do that, you need to have this process of complexification where you build regularities on top of other regularities in this sequential way, right? And I'm not exactly sure why it has to be like that, but I know that natural evolution works like that, and I know that a lot of theories, like assembly theory, they work like that. That's like an evolutionary theory in biochemistry and stuff. And human learning also works like that.

It works in a sequential way. So I'm not exactly sure why it works like that, but it definitely seems something like it's something fundamental to our universe and we really need to understand that better right if someone can answer that question why that is important or prove that it's not in general necessary that would be a huge discovery right yeah it would and the only analogy i can think of is stephen wolfram so what we're talking about here mapped onto his ideas is that we need to have a constructivist approach to representations which means that we build them step by step. And he noted that there's this phenomenon called computational irreducibility, which is that basically there are no shortcuts. And I guess the way that catches out in my mind is that it's saying that if you want to learn a complex representation, you want to acquire a complex representation, you can't actually jump ahead. You actually need to synthesize these building blocks, and you need to take the path. Yeah, and one of the things I like, whenever ken explains his philosophy is that whenever pete whenever he says like oh we need

novelty search we need this thing and um people say oh but that's just like searching in the dark that's like random it's actually anything but random because randomness is like just doing random actions novelty you know where what you did before and you want to avoid that so i guess what you're saying is that if you want to build if you want to respect the path of where you previously came from whether it's like by avoiding it via novelty or because you're respecting the symmetries that you learned along the way in a representation you do need to respect it and that gives you an idea that get constraints where you can go and i question is like how much does that constrain you there's a good chance it constrains you a lot and it constrains you towards in a right way towards the right path right that's why whenever i think about the fer hypothesis and this line of work i really like looking at it through the lens of regularization what like like people talk about regularization and deep learning with like weight decay or some sort of other regularization mechanism and i love that i think everything is just regularization the data is a form of regularization it regularizes um what like you need to model the world that's a

form of like it's like it's like basically molding the clay in a way that you need to model the world l2 regularization like weight decay is another form of that basically it's saying you want to find solutions with low weight norm right and then i think that this curriculum learning basically you want to respect the symmetries is just another form of regularization but i think it's extremely extremely extremely strong regularization that you're subjecting this model to and maybe it's the right regularization that we really really need and here's a quick refresher our fractured entangled representations documentary that we published last year explained why the path is more important than the destination when it comes to representations in neural networks this is kenneth stanley this hierarchical locking tells us something really important about how representations emerge we think it's about finding the rights building blocks now but weirdly it's about making future discoveries more it matters not just where you get but how you got there you know and that's like something missing right now you know because we tend to just care where you get like we look at

the benchmark score you know in the field and it's like that's the result it's doing really well it just passed the math olympiad but what if it matters how you got there that that kind of structured evolution, I think, is a completely different paradigm to doing it the statistical way. Exactly, exactly. And I think you said the right word. It's like a paradigm to understand intelligence. But this FER idea and this pick-breeder line of work, it's basically like a different paradigm of intelligence that we're trying to study. We're not studying it from the standpoint of statistics. We studying it from the standpoint of regularities right So it the difference between like a statistical intelligence and a regularity intelligence And I really like the idea of statistical intelligence I think we did a good job like going down this path. We should still go down this path, but we should also not put all our eggs in one basket and really try to understand this other thing because there's lots of evidence that this other thing has lots of potential here, right? And it really ties into what all the criticisms of deep learning, they're basically talking about regularities. I wonder

people talk about like we need symbolic processing whenever we the way we need geometric deep learning we need more inductive biases and people have like all these different fields uh there's whole different fields on like what we need right and they're all forms of adding regularities and symmetries to the system and what i really like about this fer approach is i like to think about it it's like it's like an automatic symmetry learning algorithm right because um so i have I love you. I have an example I like to say. People like CNNs because they're translation invariant, right? People like Transformers because they're permutation invariant and they encode certain symmetries. And that's great. I mean, we need permutation invariant for set operations, right? But can you code in like lighting invariants into your model, like lighting invariants, like if you want your CNN or image model to be invariant to whether it's a nighttime or daytime, there's no equation in the world I'll probably get you there. Maybe there is in the future, but that's impossible for me to imagine that you could come up with a mathematical like inductive bias that's going to guarantee that right so what you really need is some sort of

data-driven symmetry learning thing that automatically respects it and it forms neural circuits which are naturally have that inductive bias built in right and you can see this in like the pick-beater skulls it's like the neural it's not like the inductive bias is in the hypothesis space of things we're searching for it's in the solutions like the neural circuits that we find the regularities are baked into the circuits i suppose another way to look at it is with human intelligence um our knowledge acquisition process is grounded in the world when we build these artificial life simulations they're not grounded in the same way right so they they just and but they do have these properties that we're just talking about so they're actually building these representations and interesting things happen when you look at the phenomena of how these things evolve that and you get certain things that just fall out of the fact that you're doing this iterative computation even in an undirected way yeah so i want to touch on one key point that you said that you said it's not grounded in our world right and that's one of the my favorite things

about artificial life is that you can ground it however you want to in whatever artificial world you want to right so people often think about artificial life as it's only talking about life the reason i like it because there's a subfield in artificial life called artificial chemistry it's about just changing instead of using like the normal hydrogen helium you can change how they interact and see what happens right so that's the power artificial life you can do counterfactual analysis in other universes right and i really like this idea because artificial life is not only about life but it's also about studying like what happens in general in like artificial physics artificial chemistries can we study the emergent properties in each of these different universes when in what universe this is the questions i really want to focus my research agenda around is that in what like worlds do we get this open-ended creation of complexity and emergence and in what worlds do we not get that and why is it and this is the key question is why is this parameter why does this physics when you run it it produces this amazing phenomenon

and when you run this physics it doesn't do that it produces just random noise or something right and that was the key reason i actually worked on automated search for artificial life as acil work is because I wanted to answer that exact question directly. But ASIL was trying to basically parameterize a space of intelligences, find the good ones, find the bad ones, and just do simple stuff like interpolate between them and just see what happens, right? Interpolate the parameter space between them and run the simulation for each intermediate point and just see what happens. Just simple scientific analysis like that. Yeah, and another thing that jumps out to me is that you get these convergent patterns and rule your space. So even though the rules that... And we use this term emergence, and maybe we should talk about that as well. We don't necessarily mean emergence in the strong David Krakauer sense, but, you know, weak emergence is just that when you sequentially apply these rules many, many times, you see these very surprising, unpredictable phenomena. And these phenomena have characteristics which seem to be shared in rule of your space,

like certain types of glider and object and pattern. And, of course, Wolfram believes that the universe itself is computational. But it's almost like we're seeing shadows. What do you mean by that? Like you're seeing shadows of, I guess, the substrate that the universe is running on. Well, I mean, it's only a hypothesis. So, like, we see these convergent patterns in different parts of rural space. And now the skeptical person might just say, okay, well, those are just, they're like design bias. They are just features of the way that you are doing this computation. They are features of the structure of the game of life, maybe because you're using a grid or you're using a certain type of, you know, program interpreter to do it, that those things are influencing these patterns. But they might also be hints of just something else. You know, they could be shadows. I kind of agree with you because in my ASIL work, I span the space of all possible 2D cellular rotunda, for example, right? And in my case, I had a lot of diversity.

In one world, a glider looked like a Conway's game of life glider. in another world the uh the minimal glider looked like a train going down the tracks right so and it like didn't it left like a long path so um in different worlds it can be different but i guess what you're talking about is that when they're all the same in all the different worlds that points to something fundamental about like not only the world that you're in but like not only the rule that you're in like the rule that you're operating on but like the entire substrate that you're working in maybe all possible rules in that substrate um have this type of thing um maybe we should just talk a little bit about some of these simulators just so that folks in the audience understand what we're talking about just to bring this to life can you can you go through some of these simulators and just explain how they work yeah i can go through some of the simulators so uh i can start off with uh the linear one i'll start with the simplest one conway's game of life so conway's game of life is a super famous um simulation by john conway where you have this grid of cells this 2d grid of cells and they can either be on or off alive or dead that's it

and basically it has super simple rules like if i'm dead and three of my neighbors in like my eight i have eight neighbors if three of them are alive and i'm dead then i come to life right as an act of like um uh coming to life and there's like rules like that there's like two or three rules like that and it's like super simple rules and that's it that's it that's that's that's all you need to know then you start off with a grid of like random um dead or alive cells or you can start off with some pattern that humans have created and you just run this simulation over many many time steps and from this super super simple world it's like possibly the simplest possible thing you can think of you run you get insane amounts of complexity that no one would have thought about that thought it could do until you actually run it and now people are like there's whole discord servers i'm actually in one of them where people try to go through and They just share what they found. They're like, oh, I found this cool spaceship that's a repeating pattern that repeats every 64 time step,

and it moves at this speed. And I really love this community because they call, like, the glider how fast it moves. I think it's at, like, a certain speed, right? And they realize that there's, like, a speed of causality in this system, which is, like, one cell per time step. That's, like, the speed at which information can propagate. So they call that the speed of light C. and then they use their gliders and they're like this goes at c over 16 speed right just like in real physics how people talk about this kind of stuff so it really is like an artificial universe and people a lot of people seem to think it's like a toy system but i guess the point of it is that it's toy enough that we can study it if it were super complex our real world then we wouldn't be able to study it right and but i do think it's a very interesting quest that's one of my quests is to create more compelling universes, which are capture, like Conway's game of life, they capture some properties of our universe, like emergence and this idea that you can build up global patterns from local interactions. But there's other stuff we want to capture too, right? Like the idea of self-organization

and open-ended evolution. So a lot of my research is on trying to find simulations, which have all those properties that we care about. And that was the point of ASIL again, yeah. So then how structured is this space? So, you know, like we think about, you know, many of us have seen visualizations of representation spaces of neural networks like using TISNI and something like that. And when we see islands and clusters and whatnot, is it the same for something like Conway's Game of Life? So, for example, if you change one neighboring cell, does it, you know, is it like a wormhole where you get transported into a completely different part of the space? Or does it make some commensurate small change in the emergent, like high-level domain? Yeah, exactly. That's a great, great question. and this is one of the things I really want to study. So in the initial state of game of life, yes, it's very, very chaotic. In Conway's game of, let's talk about that specifically. It's very, very chaotic. I've had examples where you change one cell, like one pixel in like the 64 by 64 grid. And beforehand, I actually use this in one of my talks. It's like, it basically runs forever.

It's just like, it's just not forever. I shouldn't say forever, but like for like tens of thousands of steps, it just keeps on moving, right? And it's just like, keeps on doing interesting stuff. You change this one thing, it stops it basically halts into an oscillating pattern in like less than 100 steps um so it's like in the initial state uh yeah it's very very chaotic but the interesting thing is that this is not necessarily true for all simulations and um in acil we're really interested in searching over the physics like the actual rules that dictate the dynamical system like what i call the physics what we've been talking about is the initial state right but there's two things initial state of physics and if you do it in the physics uh game of life seems to be a little more robust as in if you change like one there's basically 18 possible rules you can toggle on or off and if you change one of them it doesn't impact it too much in my experience but there are some exceptions to that where yeah very interesting so and and talk about some of the other simulators as well so uh one of the so i can go through a list of the ones we use in our paper in linea by bert chan and he basically created this continuous generalization of game of life

which generalizes to like three channels and you can dictate them to RGB. And they're basically just like a continuous differential equation update rule that dictates how they operate. So I think that one's really cool. Here is Bert Chan, the creator of Linear, on what happens when he made the rules continuous. By doing this kind of making everything continuous, what will happen? When I cook this system in my computer, something creepy happened. a pattern that is creeping across my screen. And then also another cool one is the neurocellular automata. Yeah, by Maud Vincel. Yeah, are you going to interview him? Hopefully. And he famously, in 2020, released this paper called Growing Neurocellular Automata. And my God, we've talked about this on MLSD millions of times, the thing with the gecko. Yeah, definitely. And it was showing this emergentist optimization and morphogenetic engineering, essentially where you train the CNN and then you can kind of think of this a bit like a convolution operator.

So you learn the parameters of stochastic gradient descent of the CNN. I guess it was like emergentistically trained to reproduce an image so like a gecko Yeah you training the local rules such that they reorganize in order to self-organize in order to form the lizard. Yeah. Yeah, and then you can delete parts of the lizard, and then because this thing is running continuously, right, so it's just convolution, convolution, convolution, and it reconstructs the lizard. So, and again, this is like what we want in AI systems is, you know, we don't want imposters. We don't want to learn the lizard. We want to learn the ability to reconstruct the lizard from the building blocks. You know, the path dependency things that's coming up again and again. Like that's just so important, you know, rather than just learning what the lizard looks like. Yeah, yeah, yeah. And to build on top of that, I guess it's like you don't really want to learn the intelligence. You want to learn the process which grows the intelligence sort of, right? Yes. It's similar to that. I mean, that's one of the reasons that deep learning took off in the first place,

We're not hard coding all this knowledge we know about the world. We're letting it learn from the data itself, right? But there's other problems to that maybe, but yeah. Going back to the neural cellular automda, just for the audience, it's basically just like you have the cellular automda, but what if you made this three, basically the dynamics rule is like this mapping from a three by three patch into one cell, right? That's what the dynamics rule tells you. How do you go from one time step to the next? And basically what they do there is they just replace this with a small neural network. that dictates this so that gives you a whole space of possible physics rules that you can search over um so i really like neural cell automda um and i can briefly go over the next two the boids and the particle life yeah yeah boids is like a super seminal algorithm inspired by how fish and how birds flock in nature because if you look at i actually just went to the osako aquarium and i was watching these videos uh not these videos i was seeing this fish fish in person and they were like always together and there's like a whale shark coming up behind them and they just like steered to their left a little bit and i was like how did they know how to do that because each fish

they only one of them knew about the whale shark and it tugged the other ones to the left and the other ones all kind of just realized oh we should all go there it's an example of collective intelligence in nature right so voyage is basically an inspiration of that where you have these like bird-like objects which flow through 2d or 3d space and they have super simple rules where they just look at the neighbors like i'm my neighbor to the left and to the right and i basically let's say I have like four neighbors I can see, right? I'm just going to make a decision on whether to go left or right based on where their positions are. That's it. And if you make the right decision and you're all making the same decision, you can get the emergent flocking behavior that we see in nature. And yeah, so what we did is we just like parametrized this with a small neural network and just searched over this decision-making network, this decision-making. If the boys make different decision rules, how does the global behavior of the collective change according to local decision rules that they're making. Particle life is super cool because in grid-based simulations, it feels like this, it's like this spatial thing, but particle life is basically saying that we're

going to have this 3D Euclidean space or 2D Euclidean space, and we're just going to throw particles at it, just like, just like, just vanilla particles, but their physics rule is not going to be like, you know, Newtonian mechanics or something like that, it's just going to be whatever we want it to be, right? So in this case, in particle life, I was inspired by Thomas Moore, he has great YouTube videos on this type of thing, is basically you have this set of particles, and each particle can take on one of six types, right? And these are dictated by their colors. And then what you can have is that you have this six by six matrix, which tells you how do different types of particles interact with each other. So it could be that red and red repel, but red and orange attract, and orange and green, they are neutral, right? And you have the six by matrix and if you change the values of this physics matrix this is the physics of the world again right or you can call it the periodic table or whatever you want to call it as you change this physics how do the actual how does the simulation change right some of them are super bland and

boring and i don't want to look at them but some of them they're like evolving and moving and doing interesting things and some of them actually end up creating little patterns which move around that look like cells so you can literally see them they have like a cell barrier which is like this red coating and inside they'll have like purple and yellow in a triangle and that's like it's inside and it's basically this is a simulation this object this cell-like object is exploiting the physics in a way to create a persisting object in this world right and you just interviewed blaze so this idea that um life is just the persistence right it's just that idea that like life is basically about persistence these organisms are just trying to figure out how to persist into the next time frame and they've exploited the laws of physics in order to do so yeah and blaze was saying that this persistence is kind of like a natural is what you know i was talking about these natural convergent phenomena that might be an example of one because even though these systems are basically ungrounded we're just running this this computational simulation

and isn't it fascinating that we see the emergence of like persistence that things just want to to kind of keep their form over time. Yeah, the emergence of persistence is very fascinating. It doesn't happen in all simulations. It only happens in some. But I think this points to the broader point about what life is and what the second law of thermodynamics tells us. There's a generalization to persistence, which is this, at least how I have it in my mental model of how I think about it, is the second law of thermodynamics and life and persistence, they all basically are saying, and evolution, I should say, they're all saying the same thing they're talking about what exists right now and what is likely to exist in the future so the second law of thermodynamics is saying that you're more likely to go towards higher entropy states right that's just like if you look at the macro states and micro states of any simulation that's how it operates but if you look at it from the standpoint of evolution what is more likely to exist in the future it's self-replicators because they'll naturally exist more often in the future versus non-self-self-applicators so they're really

people make it out to be like oh life is like fighting entropy they're really two sides of like the same idea of persistence or what is likely to exist in the future yeah fascinating so in in your paper essentially you are using foundation models as a way as a critic right so you're saying i want to have a principled way for searching this space yeah we use it in three different ways. So at a high level before that, I guess the field of artificial life, I think it's a very cool field, but often people like only like simulate single simulations and they try to draw conclusions from it. And and I feel like it's cool, but like you really need to be studying the space of simulations. Right. So I think our algorithm, what it does is basically you instead of hard coding a single simulation, you parameterize a space of simulations, which is much, much easier. Basically, it's a difference between coding up the periodic table, like how does each element interact with each element, versus just saying that there's going to be like 100 elements, and I don't know how they interact.

You figure it out, telling the algorithm to search for it automatically. So it's a much easier problem to just say there's 100 elements, right? So once you have this space of simulations that you have described, then you can just search for what you want in an automatic way using a foundation model. And the reason I like ASIL is because it captures this principle of computational irreducibility. you don't need to know what this physics rule is going to do and try to predict which one is the best one just simulate it see what happens plug it into a foundation model and see what the foundation model thinks about what happened and then go back and the key thing is that here you're letting the simulation unroll and you're not having like you're not trying to predict what's going to happen right you're kind of embracing this idea of chaos and computational irreducibility yeah and in a sense it's a step past more vinceph's idea with the neurosellular automata because he was doing this emergent disoptimization where you're saying, okay, when the macroscopic thing emerges, I want it to look like this specific thing. And you're now plugging in a foundation model, and you're saying, I want to ask the foundation model

whether it resembles something I'm interested in. Yeah, yeah, exactly, exactly. And there's like two reasons it's better, because one is just easier to specify what you want. You can just be like, I want a cat. I don't need to have a picture of a cat and supervise it in a pixel-wise manner. I can just say I want a cat in like a general way, right? That's one way. But importantly, it also makes the optimization and the search much easier because it's really dang hard to find a self-organizing rule that creates this particular image of this particular cat. That's like really, if you do that in Conway's Game of Life, that's not going to happen. But if you just say, oh, I don't want that specific cat. I just want a cat. Give me any cat. That's a much easier space of things that satisfy that problem. So it's much easier to search for. and we experienced this firsthand that in the open-ended simulation, we're not searching for particular configurations of states that we want. We're just saying we want some open-ended simulation. It can be open-ended in however it wants to be. As long as the foundation model thinks it's open-ended, it'll do.

And the cool thing about, I want to say, the reason we use foundation models is not just because it's like the newest and greatest thing that you can use. it's because we really care about what a human thinks about these simulations because it's really hard to describe what complexity is what emergence is what does it mean to have a very interesting system it's like you can't really mathematize that really easily and if you did and you optimize for those simulations you'll find like good hearts law you'll find something that you asked for but not something that you wanted right so we really want is what a human thinks about this in this subjective, vague way. And we can't put a human in the loop because that's, like, too expensive. So what's the next best thing? Is you try to find a representation space which matches that of a human. And there's a lot of evidence, statistically at least, statistically, that these foundation models are matching human representations in a statistical manner, right? Yes. So I'm going to show a graphic on the screen now. So this is the first experiment you did was when you had a supervised target. And we've got a multi-celled organism,

a network of neurons a virus and a fungal colony and in every single row we can see what the emergent phenomena was for linear and for boids and particle life um but it also leads to the question of there must be some degree of ambiguity because there are so many possible places in computational space where the foundation model would say that this was a match and it's represented differently in the different simulators and to the you know where this concept we're talking about like this imposter. So there are many different paths to produce the phenomena. Would you call these things imposters or do you think that they're different? So these are, that's a good question. So you're basically asking, let's talk about the Boyds one. You're basically saying, is that decision-making rule of each Boyd, is it an imposter or is it the real thing that gives you the real thing, right? Yeah, and what would an imposter mean in this simulator? Yeah, that's a great question. So I think the idea of imposter is I like to ground this idea in its performance on future tasks and how it adapts to things we

want it to adapt to. That's how I like to coin it. I don't really like to ground it in modularity or something specific like that. I like to ground it in set of future tasks. When you continue to run the simulation, does it decohere? Yeah, in this case, they're pretty robust to that sort of thing. But in this case, I mean, it could be that like small changes. I guess what I'd call you, you can imagine that if these boys were really ufr or like good nicely factored the small changes to their network change the global behavior of the collective in like nice ways where it goes from a spiral to like triangles or something maybe not triangles but some other cool shape but in this case they don do that they actually degrade if you sweep the parameters then they become basically kind of random So in that way it kind of is an imposter version of this spiral Exactly, exactly. And the thing is, in this case, they're evaluated on and they're expected to maintain, persist into the future of this simulation. So in that way, they're very, very well structured because they are able to persist indefinitely.

You could run this simulation probably for like millions of time steps and they'll still be doing what they're doing. Yeah, but if you want them to do something else, then they like make a triangle by sweeping some weight, then they would fail in that way. But to be fair, they're not trained to do that in the first place. They're not, their evolutionary lineage doesn't, there's no reason for them to want to have to do that, right? I want to point out to, we had this figure in our paper which showed like what the future of this possible direction of searching for simulations via foundation models could look like. And we have this fun example. I have it pulled up right here. The bottom one is particularly fascinating. We're basically trying to search for cellular automata, like neural cellular automata, which go from self-replicating molecules, so cells floating in primordial soup, to predator-prey dynamics, to alien animals roaming around. And this is just like a fun toy example that we use. In this case, the cellular automata just memorized that sequence of frames. But you can imagine that if you did this with a big particle simulation, with like quadrillions of particles and you have the right foundation model and you search for this

sequence of events then what would happen maybe you end up with some physics rule for this big particle system which really does give you this entire sequence of like you get animals at them just like what happened in our universe right yeah but isn't it fascinating that you've taken something which you previously argued was a fractured entangled representation so the foundation model has fractured and tangled representations um but they produce imposters but you only need the imposter because all you're doing is you're acting as a critic on the imposter and then you're using that to bootstrap the creation of something which is a factored representation yeah yeah it's a good point because the the performance of these foundation models for this type of task it's like it will be flawless right because this is the exact type of things where you only care about the output behavior you don't care about internal representation you don't care about creativity in a sense because the humans are the ones kind of injecting those problems and you're using this like statistical intelligence to bootstrap like a real natural like organic

particle simulation which should have all the properties of our real world right yeah yeah exactly and then so the other experiment you did was creating a diversity search you said some parts of the computational space are open-ended. So there's this Goldilocks zone you were speaking to before. Maybe we're in such a region of the computational space where all of this life-like stuff happened and this open-endedness happened. But did you see similar things in the space where these properties that we're talking about existed in a sliver of this computational space? Yeah, definitely. So if you mean in the computational space, I guess we can talk about the laws of physics. That's your space, right? In that way, definitely. I mean, I have this, I have a pull up right here. Found simulations, which are basically in, they look more open ended than Conway's Game of Life. Yeah, that one, that exact one. Yeah. Basically, we found simulations which are more open ended than Game of Life, right? But we wanted to go further than that and ask, what is the space of possible

rules look like? What rules are open ended? What rules are not, right? So if you scroll down to in the paper, we had this graph right here, where we basically tried to plot all possible rules, there's 260,000 rules, we plotted all of them, we colored them, we did, we ran the rules, we put this image into clip, and plotted this clip space in 2d, but we colored each of the points based on how open ended they are. So we found that there's literally like a big island of solutions which are not open ended. And there's a small island of solutions where all of the coolest simulations lie. So in that way, I mean, this graph kind of outlines exactly what we're talking about, is the space of possible simulations, and we want to find what part of this space is like that, yeah. And you can see that there's like this, there's like different islands, there's like one big island, and like one smaller island, and one tiny island with like one simulation that's very different from all the rest of them, and it's behavioral properties. But then we have the problem of, there seems to be a disconnect between building these artificial life simulations,

and we want to build a general purpose intelligence which has utility. So how can we leverage this type of technology to build AGI? Yeah, yeah. So that's the grand question, right? That's how I like to go about it. So the field of artificial life, I really think is like a very long-term bet on intelligence. But there's two caveats. One, they're still interesting to study because they give you lots of insights about our universe and emergence and all these other scientific things that we still care about, even if we don't create intelligence. But second of all, I think Ken talks about this a lot, is that we don't really need to simulate every possible particle at a lower level and get everything to bootstrap. What we really want to do is create an abstraction of natural evolution and an abstraction of these simulations and find the abstraction which is, one, ultra-efficient that we can run in 10 years from now and two it gives this abstract in the right way where you get what you want which is intelligence

general intelligence and that is an open problem and that's one of my research agendas is finding the principles that we extract from natural evolution artificial life in order to create better ai systems we wanted to see what would happen if you used llms to evolve basically small little assembly programs that compete against each other so i can give you some background so there's this programming game invented in like the 1980s that people used to play it's called core war and basically the idea is it's there's like this assembly like language called red code and how it goes down is i create an assembly program you create an assembly program and everyone creates an assembly program we put them into a virtual machine and we put them in the memory address of this virtual machine in a random location and we just let the machine go it's just going to run the programs and they can do whatever they want it's like a turing complete language and the goal is that my program will be the last one to run. So what my program needs to do in order to do that is it needs to basically inject an invalid operation

in front of your program's execution thread. And when it does that, it's going to halt your program by crashing it. Your program will crash. And then my program will be the last one running and I will win this game. So in order to do this, you need to basically, your program needs to defend itself against other people attacking it and attack other people so it's like this really cool game that you can um just run and what we did was we saw that people have been making these war human warriors for such a long time we asked what would happen is if we launched uh llms to evolve these programs and have them compete against each other and can we create like this open-ended evolutionary arms race of programs from this that's amazing so it's like game of thrones touring edition yeah yeah exactly that's a good way to put it yeah yeah well tell me more so um how many generations how what was the population size how did you set this thing up yeah so the actual algorithm is really really simple so what we do is we start off with a program this can be a human program or you can be like just the llm

generator program and then you basically just optimize a program with an llm as the mutation operator in an evolutionary loop to beat this program so you just try to do that and we actually ended up using map elites because the space that you're optimizing is very deceptive it's very non-local if you just do a vanilla gradient descent and not gradient descent a greedy genetic algorithm it won't work you need to use map elite but then we end up with a warrior these are programs are called warriors in core war you end up with a warrior that beats your initial warrior then what you can do is optimize a third warrior to beat both of these two and then you can optimize a fourth warrior to beat these three and you just keep on doing this and you end up with a long sequence of warriors right and then the question is like where does this like adversarial like rounds of evolution where does it take you and one of the cool results i guess we have two main results is the first one is in this first round of training we can already beat like 96 percent of a data set of human warriors there's like a data set of like 300 human warriors

we can beat 96 percent of them just by using an lm to evolve them and importantly the lm is not good at red code if you just try to zero shot this language it's terrible if you do best of best of n sampling it's terrible but if you do this evolution and you have this verifier it turns out to be really good this is the same lesson from alpha evolve right these elements are not good at zero shotting certain kinds of proofs and stuff but if you let them evolve with a verifier in the loop they can be really really powerful that speaks to the power evolution and the priors from the model but the second result that we have that i'm the most excited about it turns out if you do this rounds of evolution right then these warriors that you end up with more further down the line they become more and more general as in they're more likely to beat human warriors that they haven't seen before so i think that's a pretty cool result but even more than that the cooler part is is if we we have this vector describing the behavior of these warriors

just like their behavioral characteristics and we can get into details of how we measure the behavior but we have this vector the variance in the behavior of these warriors goes down as you increase the number of rounds so that kind of means that they're getting more general and they're trending towards like kind of like reducing variance so they're kind of behaviorally becoming like a single generalist warrior at least that's what the trend is towards exactly and i think uh people in AI, I've noticed, tend to think of evolution in general as this very dumb, naive algorithm that you use when you can't do gradient descent, right? And they basically call it just random search. And I just want to clarify, evolution is anything but random. It's anything but random. And evolution is probably the most powerful algorithm that we know to solve like discrete combinatorial problems where you need to basically get pieces of the puzzle right before you get the whole thing right. And the selection mechanism in evolution is really, really underrated because if you solve like even one part of a puzzle and you hang on to that and you search

for the other ones, you can turn your exponential search problem into a linear search problem. So in that way, evolution is a really, really powerful algorithm. But all it needs is your mutations to be good, like 1% of the time, right? If you just do this with like naively in like program space and just involve i think there was a cool paper from like the before lm era of ai like i think auto ml zero where they did program search um but with random mutations and that's like really hard to get working properly because there's like literally around a zero percent chance that any mutation will be give you some useful code but lms sure they can't give you a hundred percent success but they can give you way more than one percent so then you can basically it whenever you are successful you can hang on to that achievement and build on top of it and that's what gives this combination of evolution plus llms a lot like a significant more amount of power than if you just do llms by themselves it's not even close akash kumar it's been an absolute honor having you on the show awesome thanks for having me yeah it's been amazing this was a great conversation yeah

Thank you.