Why the Future of AI Isn't Just Bigger Models. It's Models That Evolve | Risto Miikkulainen of Cognizant

[00:00:00] [SPEAKER_01] Let's define evolutionary AI. Search space and some of these problems are so large that humans certainly can't navigate.

[00:00:09] [SPEAKER_00] What's really different about evolution is that it's a population-based method. So you don't have just a single agent, you have 30 or you have 100 or maybe 1000 agents. And you spread them out around the space of solutions as widely as possible. I think that that's where the future of AI is, this creativity. And that's where evolutionary optimization here comes in because evolution thrives on diversity.

[00:00:30] [SPEAKER_01] If this kind of strategy works for stock trading, what are the implications for the economy?

[00:00:37] [SPEAKER_00] Hi, I'm Risto Miikkulainen. I'm a professor at the University of Texas at Austin in computer science. I've been there for quite a while. And also now I'm a VP of AI research at the Cognizant AI Lab in San Francisco. And this is very typical now. AI is partly academic, partly industrial, and they both benefit from it. And I've been working on evolutionary computation and neural networks my whole career since the 80s basically.

[00:01:07] [SPEAKER_00] And they come together very nicely. They address different needs, different aspects of AI. Very early on, we came up with an algorithm called NEAT. That was Ken Stanley, who was a PhD student at the time. That was his dissertation. And one of the successes of NEAT is that it's very robust. It's easy to get to run. It's easy to apply to new tasks.

[00:01:31] [SPEAKER_00] You don't have to optimize a lot of parameters. It gives you neural networks that exhibit behavior very easily for robots, for virtual game characters, for various decision-making tasks, sequential decision-making tasks. So people found it a very convenient tool to do things that were fun. And that is actually what new revolution is about, a lot of it. It's about behavior.

[00:01:55] [SPEAKER_00] Whether it's controlling rockets, cars, whether it's robots or some kind of agents in virtual worlds that exhibit personality, maybe. It is possible to use this technology to create that, discover that. So this is really what makes it interesting to me. I've always been fascinated by computers since I saw the movie, 2001 Space Odyssey. That's what's the motivation for a lot of us people of my generation.

[00:02:27] [SPEAKER_00] Something that I don't just program and the computer does it, but the computer does more. It surprises you. It discovers something that you did not put in. And that's exactly what new revolution is about. I often get something that was surprising and was better than I anticipated, clever solutions. And I think that that's where the future of AI is, this creativity.

[00:02:48] [SPEAKER_01] So let's define evolutionary AI. And we were talking earlier, I did an episode a couple of years ago with Julian Togelius, who's also deeply involved in the field. At that time, we were talking about evolving. He was talking about evolving game environments. A lot of people who were talking about the world, no other problems in the field. And what that means was, is a practical application for.

[00:03:14] [SPEAKER_01] And as he described it, you put a bunch of neural nets out there, were agents. And they each, and you give them the problem, and they each solve the problem. and you take the top five or so, you throw out the rest and you then take those five and have them evolve more solutions

[00:03:44] [SPEAKER_01] and you keep on iterating until you come up with the optimal. Yeah. Does that sound right?

[00:03:50] [SPEAKER_00] Yeah, that's a pretty good description of it, but maybe putting it to the context where we have now in AI, where we have multi-gradient descent-based methods. Yeah. And there you have an individual solution and you're using gradient descent to make it better. Right. Step by step, gradually, towards some goal that you know where it is. You know what the right behavior is. Now, even in reinforcement learning, you are doing exploration, but you are trying to sense where that gradient is.

[00:04:19] [SPEAKER_00] Even if it's not specifically specified, you're doing some exploration in order to find where the good areas are and then you move in that direction. What's really different about evolution is that it's a population-based method. So you don't have just a single agent. You have 30 or you have 100 or maybe 1,000 agents and you spread them out around the space of solutions as widely as possible. So you are exploring a lot more. You're exploring in areas where you would otherwise never really get to in reinforcement learning.

[00:04:48] [SPEAKER_00] So that's one aspect. The other aspect is that then you are making modifications. You are not just making small changes like following the gradient and small steps, but you're doing recombination. So you take two good agents that may be very different, different parts of the space, and you form an offspring that's a combination, recombination of their encoding and typically their abilities. And in this sense, you can make these large jumps in the solution space. So your search is much more exploratory

[00:05:17] [SPEAKER_00] and broader than it would be if you're just following the gradient. So those make evolutionary optimization and evolutionary search different from other mechanisms. And what it results in is creativity, really. Exploration and it results in creativity, and especially these surprises. So while it's usually been really hard to come up with AI that would do better than humans and surprise humans, this is actually routine. So in the Evolution Computation Conference,

[00:05:47] [SPEAKER_00] there's a competition on human competitive results. Results that are at least as good and hopefully better than humans. So it's like routinely surprising human designers with something that they didn't know. And that, I think, is where the real power comes from. That's the role of Evolution Computation. Where we need that kind of solutions, Evolution Computation is a pretty good technique for that.

[00:06:09] [SPEAKER_01] Yeah. Is this really an architecture issue? Or is it, are you using different algorithms? I mean, are you using transformer-based neural nets, models, and allowing them come up with answers and then choosing? Or are you looking at different algorithms within the neural networks?

[00:06:35] [SPEAKER_00] Yeah. So Evolution Optimization itself is very general. You can evolve just about anything that you can cross over or mutate. So it could be strings, or it could be trees, and they can represent programs, they can represent designs, physical designs. When we talk about neural evolution in particular, so you are evolving neural network encodings. You still have to encode them in some way, but they can be graphs, and they can be just concatenation of weights

[00:07:04] [SPEAKER_00] on a standard architecture, like fully connected network. So people have evolved all kinds of architectures. It's easiest to evolve, say, just a simple feed-forward network, because you have the structure, so you just have to modify the weights. NEAT, in particular, evolves the entire connectivity, the architecture itself. But they are still relatively small networks, not that many nodes, not that many connections, that architecture is customized for the task.

[00:07:34] [SPEAKER_00] Recurrency, in particular, what you remember, how you take tasks into account, that's customized for the task when you evolve in the structure, recurrence structure. Now, then there's this whole other area of neural evolution where you are taking, say, deep learning architecture. It's like transformers or just convolutional networks or something else like that. And you are taking advantage of the gradient descent in that architecture that you actually evolve. So you optimize

[00:08:02] [SPEAKER_00] the architecture for the task, but then you use gradient descent to set the weights. And that's neural architecture search. It can be called meta-learning, and you might evolve activation functions, loss functions, the modularity, the layering structure, the channels, many different design aspects of the neural network. And then you take advantage of these advances like transformers and convolutional nets and diffusion networks. And you are, instead of being a human who tries to figure out what the optimal architecture is, you use evolution

[00:08:32] [SPEAKER_00] to pick out the architecture.

[00:08:33] [SPEAKER_01] And how many, is there a scale issue in this? The more models or agents that you have working in this evolutionary system, the stronger the results will be? I mean, or can you, are there ways to reduce the number?

[00:08:55] [SPEAKER_00] Yeah, so it's interesting. There's no single answer. Sometimes you are better off with a large population, especially if your search space is very convoluted, so that if you have fewer population members, you don't really see the global picture, where good the solutions are. You might need the larger population in order to sense that. But in many other cases, it's enough to have a smaller population and just evolve

[00:09:24] [SPEAKER_00] a lot of our strength. But there's a scale issue in other ways, and that is that if you are evolving the entire neural network, like you're doing with NEED, for instance, the topology, as well as the weights, maybe other aspects too. Modern architectures have a lot of parameters, billions of numbers. NEED may evolve networks with thousands perhaps, tens of thousands of values. So how do you actually optimize billions of parameters?

[00:09:53] [SPEAKER_00] Gradient descent does it because gradient propagates through the entire network. The gradient tells you exactly how to change every single weight or parameter. But evolution would have to have a mutation or crossover get those right. So it seems like it's a challenge. And there were some ways of encoding them so that you could actually optimize millions of parameters. But this is a really interesting recent result. And our research group at Cognizant AI Labs came up with that and then others

[00:10:23] [SPEAKER_00] have also done that later, Oxford and NVIDIA. You can use a particular form of evolutionary optimization called evolution strategy that actually optimizes billions of parameters. It does the search in a parameter space instead of the action space. When you're modifying actions, you're modifying probabilities of actions, you're following a gradient, and you're doing less exploration, and you are maybe missing some systematic changes. Now when you are optimizing the parameter space, you might find

[00:10:52] [SPEAKER_00] a small modification that actually gets you a principal change in the behavior. Now that's still a lot of hand-waving. You don't fully understand how it happens, but it seems like it's doing the optimization in a different manner from this kind of gradient following policy search and reinforcement learning, and it might actually solve some problems better. So what used to be a challenge in scale now actually has turned out to be an opportunity to solve certain problems.

[00:11:22] [SPEAKER_00] And this is what makes it exciting. I mean, we will come up with these new ideas, new opportunities every now and then, and I think here's one right now.

[00:11:30] [SPEAKER_01] So let me understand that. So you have, you know, maybe a 5 billion or I don't know, 70 billion or, you know, parameter model. And, well, I'll let you explain. You're starting with a model with 70 billion parameters.

[00:11:52] [SPEAKER_00] Yeah. So the evolution strategy is one version of evolution computation. And it was discovered a long time ago in the 50s and 60s. And occasionally it comes up again. There's a version called CMAS, which is using covariance matrix to figure out what direction to change the variance. But evolution strategy itself is simpler. It's simply taking the current best, creating a population around it,

[00:12:22] [SPEAKER_00] not very far, just around it, and then finding how well each one of those performs and then moving in the direction of the best one or kind of average direction of good solutions. So it's repeatedly doing this kind of a relatively local search or exploration. And it can actually, and when it does it, it does it in the space of these billions of parameters. So every weight can be changed using this evolutionary optimization. Are the weights changed?

[00:12:51] [SPEAKER_01] Is this during training?

[00:12:53] [SPEAKER_00] Well, we applied it to fine tuning and other groups have applied to even pre-training. But yeah, I mean, you are actually changing billions of weights.

[00:13:01] [SPEAKER_01] And are you in that process? Is that after the initial pre-training? We did it.

[00:13:08] [SPEAKER_00] There were models like Quinn and Lama and others that are open source and you can get them after they've been trained and then you fine tune them to a particular task like a particular math kind of reasoning or maybe making the answers shorter and to the point so you can modify the behavior some way. You have to have a now and you have to have a training set that tells you how you want to change the behavior. So you reward those changes that are what you want in that training set.

[00:13:37] [SPEAKER_00] That can be done after pre-training. And it's very common I mean in LLM training that you have a base model and then you make many versions of it that are specialists in some way. So that's what we did. But indeed, there's now exploration of starting from scratch and actually doing the entire training using Everless In Search. Very cutting edge, very new. We don't quite have it down yet but it's an interesting future direction.

[00:14:02] [SPEAKER_01] And when you say starting from scratch how do you point it if you're doing it in pre-training how do you point the model toward what are you optimizing for?

[00:14:18] [SPEAKER_00] You have a training training examples and you are evaluating performance on those examples and you reward the solutions in your population that actually does well on those samples. So it's still not gradient based directly because you have just this identification of good solutions. We don't tell it how it's good. It's just that this solution is better. On average it does perform better than other solutions in that training set. but gradient based methods will actually tell you how you change

[00:14:47] [SPEAKER_00] the weights so that you go directly towards those answers that are correct. And this is a fundamental difference again and we need to understand it better. What is there in this population based search that allows it to find solutions that are apparently more principled than just following the gradient? Very big question for the future.

[00:15:10] [SPEAKER_01] I mean one of the interesting things about this is as you said that the search space in some of these problems are so large that humans certainly can't navigate. Yeah. And this when you were talking about it being creative it finds solutions that you may not find using a simple gradient to set That's right.

[00:15:40] [SPEAKER_01] Right. And why is that though? I mean because what you're talking about sounds like a kind of gradient to set because you're evaluating the local and then you're moving toward whatever But there's

[00:16:01] [SPEAKER_00] still a landscape you're still in a surface of this landscape that different solutions give you different rewards and you're absolutely right you don't fully understand it. but likely the space is still jagger it's still rough and if you follow the gradient you cannot quite see which way is up because you are in these troughs and you are misled by local very small changes so your population size and this cloud of solutions still has to be larger than those

[00:16:31] [SPEAKER_00] smallest features in the landscape so that it can actually jump make larger jumps over those areas that are not so good and you have to get it right you have to have large enough exploration that fits the problem that allows it to find those large areas so I think that the fundamental principle is still the same that it's doing exploration because it's not following the gradient but it actually is decided they're looking for solutions in a broader area but evolutionary

[00:17:01] [SPEAKER_00] strategy is more limited whereas in population like genic algorithms you might start by having solutions everywhere and then trying to do crossover so there's no crossover in evolutionary strategy there's just this cloud of solutions that propagates

[00:17:16] [SPEAKER_01] there's no crossover so explain that because what Julian was talking about is you give the problem to different models and then you evaluate the solutions and you throw out the models that didn't get a good solution but then you combine those models yes that's

[00:17:41] [SPEAKER_00] recombination so I've been talking about two different mechanisms of evolutionary computation there are several flavors of it and these are say maybe two extremes and one of them is this genetic algorithm type of approach where you like I said have a population that covers as much as you can of the space and you find good solutions and recombine them and they may be very different and your offspring is a combination of representations of both but evolutionary strategy is different it does not do that kind of recombination

[00:18:11] [SPEAKER_00] it has this cloud of solutions it's more local but still like I said large enough that it gets over those rough features of the landscape but it's still a population you still need a population because you don't quite know where those solutions should be placed you just know that you should search in this area of this size so both are population based methods but they are a little different in how they're implemented now we still don't know it's possible that once

[00:18:41] [SPEAKER_00] we get evolutionary strategy to work on this optimizing of billions of parameters there may still be a mechanism that would benefit from recombination so we may be able to improve upon the vanilla kind of evolution strategy search by taking elements from these other areas of evolution computation we'll see

[00:18:59] [SPEAKER_01] right so combining the two I was mentioning to you that there was recently this alpha arena competition for stock trading using LLMs for stock trading and there was one model that was named it was just called the mystery model that

[00:19:28] [SPEAKER_01] outperformed everybody and there's some forensic footprints that lead to evolutionary and neuro evolutionary AI in particular Julian has written a paper called profit which applies this strategy to stock trading

[00:19:58] [SPEAKER_01] so I I'd have to look it up but it's it's specifically for for financial trading how new is that and if that's the case what are the implications if this kind of strategy works for stock trading what are the implications for the economy but in

[00:20:27] [SPEAKER_01] particular it doesn't involve the strategy it involves the code that outputs the strategy

[00:20:36] [SPEAKER_00] again like I said evolution computation can be applied to many different things so one thing that you can evolve is code and that's then it's called evolutionary programming or genetic programming and that's a very big area of evolution computation so I don't think there's a whole lot of difference whether you evolve code or you evolve a neural network that represents a strategy you're still if you're doing trading stock trading or any kind of market

[00:21:06] [SPEAKER_00] trading in order to really make money you have to be a little different from others so if you're doing the same thing as everybody else you don't really gain a whole lot and that's where evolution optimization here comes in because evolution thrives on diversity your population has to have different solutions that's how you discover things you don't know and the same applies in stock trading that you discover strategies that are different from strategies that everybody else is following

[00:21:36] [SPEAKER_00] or what you would learn if you just use the data set and try to optimize day to day you

[00:22:07] [SPEAKER_00] and do to of opportunities on that. And new technologies are always incorporated. But this is the real

[00:22:34] [SPEAKER_00] interesting thing. If you want diversity, then that is where evolution actually steps in and can

[00:22:40] [SPEAKER_01] give you that. Yeah. And diversity is interesting. I saw a talk yesterday from one of the Cognizant, Yuxin. Yeah, Xinshu. About diversity and data sets for training. I don't know if you saw that talk. I don't think so. Yeah. But maybe I'm getting it wrong. But the point was that you need diversity

[00:23:10] [SPEAKER_01] in order to discover new things. Yeah. That's absolutely true. So in evolution computation,

[00:23:18] [SPEAKER_00] you often even have to implement special mechanisms just to maintain diversity. Now, one of the most interesting developments in that field in the last 10 years or so has been novelty search and quality diversity. So you are explicitly rewarding solutions that are different from everything that you've seen before. You don't even care about performance. You just want things

[00:23:42] [SPEAKER_00] that are new. And it turns out that if you do that, you end up discovering things that are somehow stepping stones. Like they are unique. They allow a lot of other things to evolve from them. And now by recombining these stepping stones, you could discover truly another solutions. And this has been, I think, a really powerful discovery. And many techniques today, many applications of evolution computation utilize that. And they combine it with performance. And that's what

[00:24:11] [SPEAKER_00] is called quality diversity. So you have some aspect that rewards just being novel and another aspect that rewards performance. And this is how you discover surprises.

[00:24:21] [SPEAKER_01] Before I ask, Google DeepMind came out with Alpha Evolve. Yeah. Were you involved tangentially? I mean, certainly it was built on your research.

[00:24:33] [SPEAKER_00] So I wasn't involved in that project at all. But this is a great example of what's been happening, that evolution computation is discovered by people who were not even in the field, but it actually fills the niche, something that they know that they have been missing. And the people, I think they got it right. They know how to do it. But it was discovered really outside of the field of evolution computation. It worked quite well.

[00:24:56] [SPEAKER_01] Your co-authors of the book, many of them are from Sakana AI. Yes. And Sakana is working on evolutionary or neuroevolutionary AI. And they've made quite a splash earlier this year with a system that initiated a problem,

[00:25:21] [SPEAKER_01] designed experiments, wrote a paper, which was submitted to, I think, ICML, or I can't remember one of those conferences, and was accepted. They then withdrew it. And it wasn't a groundbreaking paper. It's interesting that it was accepted. But what are your thoughts about how this can be applied to research?

[00:25:45] [SPEAKER_00] Yeah. Tremendous opportunity. Absolutely. Again, it's the creativity and discovery. Now, that particular paper also had many other aspects of LLMs, utilizing them to the language and the description. Yes, but the core discovery engine can be evolution optimization. You find, again, things that you don't know, but then you have to evaluate them and bring in other technologies

[00:26:11] [SPEAKER_00] technologies like LLMs to do the evaluation. But it's absolutely right. This scientific discovery, I think it's one of the most promising areas. Because now we can represent the knowledge of science, the problems of science, using these LLMs. And then we can combine them with evolutionary search to find new things. They actually work together. We can use LLMs to implement evolutionary operators, crossover, mutation. You can ask LLM to do, you know, here are two parents,

[00:26:40] [SPEAKER_00] give me an offspring. And you can evolve anything that way. And that can be applied to scientific representations, discover new molecules, discover new machine learning methods. So all of a sudden, anything is fair game. You can evolve anything. So now the challenge is to be able to pose these scientific problems so that this kind of a surge can actually work. And I think there's a lot of

[00:27:06] [SPEAKER_00] opportunity. There's an event at NeurIPS about that panel discussion. And indeed, David, a co-author on the NeurIPS and book, David Ha, is part of that panel discussing it. And there are now companies that are starting. Their model is to do scientific discovery, like Lila Science, for instance. And in Lila Science, Ken Stanley is author of NEAT, is one of the scientists there. So I think this is going

[00:27:34] [SPEAKER_00] in a very interesting direction. And it's all, it's also AI for good. I mean, we need that. We need better science. We need solutions. We need better medicine. We need better rockets, you know, fusion energy. There are all these challenges. And I think that this can be part of the solution for those big

[00:27:50] [SPEAKER_01] challenges. But again, you're evolving. You're either, I'm sorry, there were two terms you used. You're either using it as a strategy, or you're actually reading new code or new architectures or

[00:28:14] [SPEAKER_01] new models. On that side, if you're looking at architectures and you want to evolve architectures, let's look at continual learning. Could you apply this strategy to solving that problem? To saying, here are my neural nets or a neural net. I want it to be able to acquire new learning without

[00:28:42] [SPEAKER_01] forgetting or to be able to expand. I'm going to talk to Sebastian about this, his cellular... Yes. I've forgotten what it's called, but this idea that neural nets can grow on their own. Can this strategy apply to those things to find new architectures to do those things?

[00:29:07] [SPEAKER_00] From what I was talking about before, where we are optimizing architectures for deep learning, transformers. We're talking about architectures that are really novel, fundamentally different. And this is, I think, really important. We should be able to... But we recognize first that we are missing some things today. Continual learning is one. Another one that I really like is metacognition, networks that know what they know. And we don't... It's unlikely that we get it from these architectures.

[00:29:35] [SPEAKER_00] Instead, we probably have to look at the brain and biological neural networks, circuitry there, and organization, and strive towards something like that. Now, how do we actually get that to work? That's a neuroevolution problem. We set the constraints, we set the elements, the primitives, and we let evolution discover architectures that, say, answer questions about their own performance, metacognitive questions, or put them into an environment that's continuously changing and

[00:30:05] [SPEAKER_00] requires this continuous adaptation. And architecture should evolve that actually cope with that. And that's exactly what, say, Sebastian is working on. And we're starting to work on the metacognition part. And I think this is... And it might actually require also actual neuroscientists, from which we can get ideas about what the directions are and what the constraints are. Because neuroscience has solved a neural network, biological neural networks know very well how to do this. And we don't really have any

[00:30:34] [SPEAKER_00] other examples. So it would make sense to look at neuroscience and other disciplines related to that. And then bring in neuroevolution in order to start small and not too ambitious, but build towards

[00:30:45] [SPEAKER_01] those architectures that duplicate that kind of behavior. Can you explain sequence of experiments that you would undertake to reach that goal? So you start small. Anything that you do, you start...

[00:31:00] [SPEAKER_00] I always tell my students that your first experiment should be so small that you don't want to tell anybody that you did it. It's embarrassing that you do something that small. So for instance, you just simply try to train a network to answer a question and then answer a question where you write in that. Is that correct? Do you actually know this fact or you just made it out? And you can do it. I mean, and then you may be fine that the standard neural network architecture doesn't handle it, or maybe

[00:31:28] [SPEAKER_00] reinforcement learning wouldn't. But if you evolve the circuitry, you might get it. This is an experiment that hasn't been done yet. And I'm more or less proposing it. And, but how do you evolve it? What are your, again, elements, primitives, constraints? There you have to talk to a neuroscientist that might tell you that you need feedback. You need a reverberating surface. You need spiking neural networks. I don't know. We'll have to think about it.

[00:31:56] [SPEAKER_00] But it's not just, it's not just one thing. We can do exploration, pretty wild exploration. We can open up all kinds of parameters for evolution to optimize beyond what we can even think of perhaps. And neuroscience is also somewhat limited on what they consider as computation because it depends so much on what we can see and measure. So you have to have electrodes, you have MRI, you have a couple of tools.

[00:32:22] [SPEAKER_00] And our theories always depend on what we can see and measure. So it might even be useful to let evolution explore a wider area and even give ideas for the neuroscientists of what might be going on.

[00:32:35] [SPEAKER_01] What are the inputs? So you, you have this evolutionary, neuroevolutionary model and you, you wanted to explore uh, architectures that would understand uh, its own outputs or that would uh, would not uh, forget as it

[00:32:59] [SPEAKER_01] learn new things. That's a model. Yeah. Is that right? And then you're feeding in inputs. Do you have multiple instances of that model and you're feeding them different inputs? I mean, how, how is the experiment?

[00:33:16] [SPEAKER_00] You have to have a couple of things. You have to have a, a domain of interaction and you have to have a way of measuring how well you're doing. I mean, you still need that feedback. You have to decide which ones of your candidates are good and which ones are not so good. So you can run uh, this recombination, crossover and mutation. Uh, now, but there are probably places that are better starting points. So if you look at neuroscience, um, memory systems, hippocampus is a pretty good start because we know

[00:33:42] [SPEAKER_00] a lot about the circuitry in hippocampus and we know a lot about like place cells for instance. And it's also involved in memory, space and memory. Um, so we have a task of perhaps navigation and remembering where things are. Um, and then you can start asking questions about what do you really know? Do you really remember something that's there or did you just make it up? Um, so we can, we can then pose the question of, do you actually know your own knowledge in,

[00:34:08] [SPEAKER_00] in that constraint area of hippocampus and memory systems? And then there's of course really big questions like how does the memories, how do memories transfer from hippocampus to the rest of the cortex for the lifetime? Uh, because they are no longer in hippocampus. And we could also expand that way and find out what kind of circuitry allows you to do that transfer indexing or whatever it is. So we don't have to stop there, but the starting point neuroscience motivated. Um, I think

[00:34:35] [SPEAKER_00] as, as, as said, such a starting point, hippocampus and the memory systems, it'd start.

[00:34:39] [SPEAKER_01] But how is it inputted? I mean, is this, do you have an LLM interface that you can provide the data

[00:34:49] [SPEAKER_00] in natural language or, or are you? Yeah. Yeah. So initially perhaps not because, and the, and the benefit of having hippocampus is that there's the space representation. So it will be probably navigation task of some kind, uh, location based on, like I said, some object in a location, go there, you know, find it, go there again, remember where it is. So we don't actually have to start with language, which is on the top of the cognition in humans, but we can start with something that's even very

[00:35:17] [SPEAKER_00] simple animals know how to do. I mean, hippocampus is one of the oldest areas of the brain, but, but there's a lot of the elements that we are missing in current LLMs are actually there. So we can address the really challenging questions, continual learning, as well as the, um, introspection and, and metacognition in that domain of navigation, space for representations and memory. And that's

[00:35:40] [SPEAKER_01] why I'm proposing it. That would be a good starting point. It seems pretty greenfield that, that you, you develop these tools and it's now a matter of seeing how well they work on different problem spaces.

[00:35:57] [SPEAKER_00] Computation is a big part of it. I mean, we were working on these problems of natural language processing in the eighties and nineties, and the ideas were pretty good, but we just did not have the data. We did not have the compute and we didn't even understand that if we did, it would actually work. Um, and now we're kind of in the same situation with evolution computation that we've had a lot of ideas, but haven't been able to scale them up. And what would happen if you really did and took advantage of this compute?

[00:36:26] [SPEAKER_00] We could run these experiments on evolving a hippocampus for instance. Uh, and even beyond that, my, my pet project would be understanding how language evolved. Now, what are the, because now we can have simulations of complex environments, world simulations, uh, and we can put these agents in there and have them involved and develop abilities and eventually maybe communication and even language. So that's, that's a big, big challenge for

[00:36:51] [SPEAKER_00] longer term, but we can now, we have a compute and we have a way of simulating this environment. So things like,

[00:36:57] [SPEAKER_01] uh, questions like that could be addressed. I mean, you've been working on this for a long time. Why have you not, or have you worked at that scale scaled up? Sure. We tried. I mean, there are many

[00:37:09] [SPEAKER_00] studies by us and others on evolving communication in, in virtual creatures, but that's all always where it stops. I mean, you can signal and they can read each other's signals and behave, uh, um, in a meaningful manner. But language is still a big step beyond that because you have to have linguistic structure, grammar, uh, and flexible structure and roles and pillars and so on. And how, and, and we know

[00:37:37] [SPEAKER_00] that animals, some animals are pretty good at that and they can even adopt features of human language if you really train them, but they don't do it naturally. But, um, animals communicate, uh, but not with language, not with grammar and they can, um, adapt a lot of that if you really, really train them. So the question is why did that evolve and how did it evolve and why only humans and, uh, what other communication systems are possible and is this the best one? And very interesting questions come up.

[00:38:04] [SPEAKER_00] Now that we have the compute and we have an ability to simulate, we can start asking those questions.

[00:38:09] [SPEAKER_01] Scale it up. Yeah. And, and is that what Cognizant's doing? Is that what Cogna is doing? Um, uh, maybe

[00:38:16] [SPEAKER_00] someday, uh, we hope to hope to get there. Yes. It's a, is it a financing issue or? It is, um, there's no immediate necessarily an application of it because it's really a fundamental scientific issue. Like, it's one of those, those, uh, fundamental challenges of science is how, how did language evolve? Now, well, I'm not talking

[00:38:38] [SPEAKER_01] about that particular project, but just scaling. Oh, scale. Absolutely. Yeah, absolutely. Uh,

[00:38:45] [SPEAKER_00] and there are immediate applications, decision-making, anything decision-making in society, business, healthcare, medicine, uh, science we talked about, uh, is an application now. And, and we can take many steps before we get to these questions that are really kind of flying in the sky.

[00:39:01] [SPEAKER_01] Yeah. I mean, I can see with finance, it's a data rich domain. Could this, this could be applied to anything, but, but you need the data to evolve. You need to be able to evaluate how well you're doing.

[00:39:16] [SPEAKER_00] Uh, it doesn't need to be database. Like I said, simulation could be just as good. So you simulate behavior and you evaluate how well they perform. Well in stock market, you could actually trade stocks and see how well it works. In healthcare medicine, you could see whether your treatments effect the, have good effects, but you don't want to, you don't want to do that in the real world. You have to have some kind of a surrogate model of the world. That's where you need the data. If you have data,

[00:39:43] [SPEAKER_00] you can train a surrogate model. If you have surrogate model, then you can discover the system strategies. And that I think is a really good approach that we are ready for.

[00:39:51] [SPEAKER_01] Is, is that why a lot of the people involved in neuro evolutionary AI are also focused on world

[00:39:58] [SPEAKER_00] models? I think the reason is perhaps that now we are transferring or transforming or, or making progress and moving from imitation, like models that just imitate the statistics of the world that we already know to agentic AI. I mean, that's a big word right now. Agents that interact with the world. Right. They, they make decisions, they affect the world, the world changes.

[00:40:23] [SPEAKER_00] And that's where you need a world model. So I think it's more of a result of us going from modeling to agentic AI. And we need this, this world models to do it.

[00:40:34] [SPEAKER_01] Does Cognizant work on world models at all? I know Sakana is doing.

[00:40:38] [SPEAKER_00] We are not developing right now a world model ourselves, but we are definitely using them in

[00:40:42] [SPEAKER_01] order to do agentic AI. In whose world models are you using?

[00:40:47] [SPEAKER_00] Yeah. Minecraft has been one of those that, that, that's existed until now as a, perhaps the most versatile such world model. But now they are just now like within the last few months, things have changed and more miles are coming out. We'll have to take a look and see what we can do with it.

[00:41:01] [SPEAKER_01] Have you seen, I have Fei-Fei Li, who I, from what I understand could generate sort of...

[00:41:08] [SPEAKER_00] Yep. That's a great opportunity to take advantage of that. Yes.

[00:41:11] [SPEAKER_01] So where are you guys going from here? Well, let's talk about the book first. Okay. The book is kind of the, the intent is to create an authoritative guide or textbook to neuroevolution in much the way that reinforcement learning by Sutton and Barto did.

[00:41:41] [SPEAKER_00] Yeah. It's very much like that. Yeah. Yeah. And there hasn't been one today? There's been books about evolution computation, but not neuroevolution.

[00:41:51] [SPEAKER_01] And is this, is the book intended, is your hope that you'll generate a following and a shift the research community?

[00:42:06] [SPEAKER_00] Yeah. Attract more attention. I mean, we are providing not just the book, which is, first of all, historical overview where all these ideas came from, but also what's currently happening, what's breaking through, like these combinations of neuroevolution and reinforcement learning and deep learning and biology and generative AI, LLMs. All of these are like breaking out right now. And if we can give people the background where they came from, if they understand the strengths of these different

[00:42:34] [SPEAKER_00] approaches, then they can contribute. And I think that's the role of the book, like give everybody the basic knowledge so that they can take it further. And we believe that this is the time in science where this is happening, that there will be a lot of expansion. We want to encourage that. And beyond the book, we also make a lot of software and digital resources available. So there's lots of demos that are

[00:43:01] [SPEAKER_00] inspiring exercises for students. And then we have a GitHub site for community to provide software, to provide pointers to their new papers, new developments. So we really trying to build a community that could work together on taking advantage of that. And they would have a common foundation.

[00:43:18] [SPEAKER_01] Yeah. What I find interesting about this is I've been talking to people about, I had Carl Fristen on talking about free energy principle and post-transformer architectures. There's another group in New York working on post-transformer architectures. And Fei-Fei is working on her sort of explicit world models.

[00:43:41] [SPEAKER_01] Jan LeCun is working on kind of internal world models. This is almost a horizontal technology, neuroevolution, that can apply to all of these sub-problems. Is that right? I mean, that's what strikes me is it's not domain specific. Right. There's actually a lot of different things that can be

[00:44:08] [SPEAKER_01] helped or advanced. And I don't mean domain specific. I mean, it's not specific to one school of AI. Right. It can be applied to all of them. Not the solution for everything. I mean,

[00:44:23] [SPEAKER_00] there are certain areas where neuroevolution can really help and others where you have other methods that are better. And indeed, if you have a lot of data and you are interested in modeling the statistics of it, keep learning is perfectly fine. And it's really good. It's really those areas where there's opportunities for finding creative solutions, where we have little idea of what's going on and what the right solutions should be like. And this, like I said, green field,

[00:44:50] [SPEAKER_00] it's a great method for discovery in such fields. And your association with Sakana,

[00:44:58] [SPEAKER_01] I mean, do Cognizant and Sakana work together at all? Or are you very independent and it's just for

[00:45:05] [SPEAKER_00] the purposes of the book? It was a book really. I mean, we could perhaps in the future, but it was really just all people coming together because we all provided a different compatible perspective on neuroevolution and found out that actually if we put those together, it's a pretty good foundation for

[00:45:20] [SPEAKER_01] for the field and for the community. And so Cognizant, what's Cognizant's, tell us of the history of Cognizant, what's its mission statement, who funds it? Well, Cognizant itself is a very large

[00:45:34] [SPEAKER_00] company, 350,000 people, and they have been doing staff augmentation as well as consulting, and now it's turning to an AI company. But the idea is to take AI to the world. They already have a structure to help digital transformation in many different companies. So now it's becoming an AI transformation. And our role, Cognizant AI Lab, is to provide the technology to do that. I mean, they can use any

[00:46:00] [SPEAKER_00] technology they want, and there's many providers of those. But we also spearheading the development of multi-agent systems in particular, and neuroevolution as a decision-making system strategy, decision strategies for business, for instance. And there are many businesses that benefit from that. Like I said, some tasks are good for it, not others, but you can make decisions, for instance,

[00:46:26] [SPEAKER_00] how to allocate resources in marketing or in research itself, or something like that where you really have to be creative in order to be effective. In other areas, other techniques might be used. But multi-agent systems is one area which is also very strongly coming. We talked about deep learning, big data first and deep learning, and now then LLMs and now agentic AI and multi-agent systems, multi-agent

[00:46:55] [SPEAKER_00] AI perhaps, I believe will be in the future. So now we have multiple agents that are talking to each other and solving problems together. And it's flexible, it's modular. You can have agents whose job is to evaluate other agents and make sure that the system stays in check. And there is absolutely a sensual role for some agents that are creative, that implement and learn those decision strategies in such a context. There might be a role for neuroevolution to fine-tune those models, so they

[00:47:24] [SPEAKER_00] become very skilled at certain kinds of behavior. Otherwise, they are just LLMs that are using RAG, they are using documents, but they can also become real experts on how to reason about certain kinds of, I don't know, medical knowledge for instance. So that is what we are doing in the science and AI lab. This multi-agent system is currently a big push, and then neuroevolution as an element in those multi-agent systems is a role for this technology, yes.

[00:47:52] [SPEAKER_01] And so your basic research on neuroevolution is taking place at Austin?

[00:47:57] [SPEAKER_00] Yes. Well, both. I mean, I'm also a professor at UT Austin, so I do have a research group there as well and people who are working on mostly, right now, it's mostly on cognitive science, which I also have been working on for a long, long time. So that is modeling, say, human behavior, like patients who

[00:48:21] [SPEAKER_00] suffer the stroke or dementia, how to help them. We've looked at the visual cortex before, but it's all coming together now because we're really starting to see that evolution, which we used mostly to create behavior, now has to look at these cognitive aspects of behavior. So I believe that in the end, they will come together. We will build systems that are cognitive using neuroevolution.

[00:48:45] [SPEAKER_01] Yeah. I'm just curious, how do you manage all of these different, I mean, they're related. Right. But all of these different directions of research.

[00:48:55] [SPEAKER_00] Yeah. Yeah. It's to say that they're all very different directions, but they all related in my head.

[00:49:01] [SPEAKER_01] Yeah. But I mean, do you have a team that's working on this and you spend a few hours with them working out a roadmap and then you switch and you have some people working on something else?

[00:49:16] [SPEAKER_00] Well, sure. But there's this dynamic self-organization. Teams form, if it really works, teams form around an idea that's an opportunity. And I'm lucky to have some of these senior colleagues in Collin State, for instance, Shin Chu was mentioned and Elliot Mason, for instance, who have their own teams and working on different aspects. And we talk all the time. And also internships are great.

[00:49:43] [SPEAKER_00] So I have students who are at UT and they come for internships at College of the A-Lab and learn a lot and maybe stay, or maybe they back and do their dissertations on topics related to that. In general, in the bigger picture, I think what is really wonderful right now is this interaction between industry and academia.

[00:50:00] [SPEAKER_00] So we're all here in St. Conference and we publish papers and we talk the same language. And this openness of industry and the possibility of academic people to have a one-footed industry, that I believe is a big factor of why AI gained what it is now.

[00:50:21] [SPEAKER_01] Is that partly because industry has the financial muscle to provide the...

[00:50:25] [SPEAKER_00] Absolutely. And you can still, you can kind of still see a little bit of specialty for both. So industry has finances and means resources, just computational resources, but resources otherwise too. Resources to access, to get data and to have people who build systems that actually work. You know, in academia, well, maybe if you don't have quite the resources and don't have the engineers,

[00:50:52] [SPEAKER_00] you think and be more creative and test ideas that are further along. So we are actually evaluating or testing or generating, exploring architectures that are not right now feasible. While industry more or less will have to have architectures that are feasible. So there is still a little bit separate role time scale wise and exploration wise and resource wise. But I think it's right now, it's a very synergetic relationship.

[00:51:20] [SPEAKER_01] Yeah, that's fascinating. That you're testing ideas. I mean, the whole process of scientific research fascinates me that you're coming up with hypotheses that you don't know to work. You're right. Industry doesn't, may not want to waste money on it, but academia.

[00:51:45] [SPEAKER_00] A lot of times I say this is the best time ever to be alive. I mean, this is so much excitement now. You have opportunities, you have an idea and it has a very good chance, if it's a good idea, to actually affect the world and make the world a better place, which is what we always used to say. But now it's really reality.

[00:52:04] [SPEAKER_01] Yeah. Where do you see neuroevolution going? I mean, it sounds so powerful and it can be applied to so many things. Is developing it a matter of compute or is it a matter of data? Because you need good data

[00:52:24] [SPEAKER_01] data and presumably a lot of data. I mean, you were talking about decision making in the different problem spaces. I always ask people about the Ukraine, Russia war. If you had enough data, would an evolutionary system

[00:52:51] [SPEAKER_01] be able to find an optimal outcome by trying all these, you know, all the concerns of one side, the concerns of another. I mean, maybe that's not a good example.

[00:53:07] [SPEAKER_00] Oh, no, I understand what we're getting at. And this is really something I often very much like to end my talks with as this opportunity for the society in that we have AI now and it's indeed discovering decision strategies is a big part of it because we can talk about decision making in society.

[00:53:27] [SPEAKER_00] And the dream is that we decide what we want the society to be like. Do we want to maximize profit or production or scientific progress? Or do we want it to maximize equality, distribution of wealth, protection of environment? If we decide what those goals are, then using this kind of technology, we can come up with decision strategies that get us there. And it's important because they are

[00:53:57] [SPEAKER_00] objective. They don't really care about, I don't know, pork and personal agendas and other things that often get in the way in the society, right? Or making really good decisions. So we have, I believe we have the technology, if we have the will to take advantage of it. Now, I'm not quite sure of the, of really big challenges like Russia, Ukraine war or the Middle East, because there are so many of

[00:54:22] [SPEAKER_00] these other agendas and things that come into play. But if we can actually identify some part of society where the goals are clear and agreed upon, I believe now we could do a much better job by

[00:54:34] [SPEAKER_01] utilizing technology. Do you think that you need, as you said, the will to use it? Is this something that governments could take advantage of? And are they? And then are there? Well, yes, absolutely.

[00:54:52] [SPEAKER_00] So one example, pandemic. Everybody wanted to do something about the pandemic when it started, and we also developed a decision-making system for non-pharmaceutical interventions. Should you close the schools, stop the buses, wear masks, do contact tracing? There are many different dimensions that you can try to establish in order to help or prevent the pandemic from spreading.

[00:55:16] [SPEAKER_00] And it was based on data, and I was lucky that the data was available. The cases and deaths and hospitalizations, but also what governments were doing all around the world. And they were doing different things, so there was diversity, and we could learn from it. And it was quite fascinating that we could come up with these suggestions, just training the system overnight. The next morning, the system would

[00:55:40] [SPEAKER_00] make suggestions for any country in the world where we had data. Now, it was really difficult at the time to get anybody to listen. And it wasn't just our group. There are other groups who have been studying, U.T. Austin, for instance, Lauren Meyer's group, has been studying pandemics for decades and had very good understanding. They had actually an audience in Austin, in the city that listened. And Austin did really well think about it. And we had to learn more about the world compared to other cities because they listened to science.

[00:56:07] [SPEAKER_00] And our goal was really the whole world, the different countries. It was so much harder to call up, you know, Modi in India and say, this is what you should do. We did have one piece of success, and that was Iceland. So they, in the fall of 2021, they were considering what they should do when schools open after the summer. And we had a contact there, Thur Ascolund, who was an academic and had contacts

[00:56:34] [SPEAKER_00] with the government, all the way to the health ministry and even prime minister. And now we could run models that were relevant to them and run these scenarios, make suggestions. And that actually worked. As far as I know, they did communicate to the ministers and they did follow some of those because they did well at that point. So it's possible, but it was a much smaller country than India, for instance, or US.

[00:57:00] [SPEAKER_00] And it is still a big challenge, communication of science. It's magic to many people. And you really have to work hard on getting the message through. I think the technology is ready, but the communication

[00:57:16] [SPEAKER_01] is not. But is the technology ready? I mean, it sounds like it's still in the development phase.

[00:57:23] [SPEAKER_00] But if you pick a topic that is a good match, yes, I think it's ready. Like the pandemic,

[00:57:28] [SPEAKER_01] it would have been, it was. In your work with Cognizant, can you talk, are there, what's the biggest application or problem that you've applied a neuroevolutionary AI?

[00:57:43] [SPEAKER_00] The most immediate ones are things like marketing, budget allocations, transportation, design of clinical trials tend to be a little smaller. Domains where either the current state of the art is relatively uninformed, so you can't really go wrong. Or at least there's a comparison and you can immediately

[00:58:12] [SPEAKER_00] tell that there's value. Now, and this applies to AI in general, why it's really hard to get into the world is because a lot of times AI is very ambitious and AI would just replace what's already there. And this is a very hard decision to make. So we have to build AI that maybe runs alongside or replaces a small piece that doesn't really matter yet. And you gradually gain confidence and you gain ground

[00:58:39] [SPEAKER_00] that way. And humans are always working alongside with this AI and it's empowering these humans to make better decisions. So that's how we have to approach and that's what we are doing too, that you find these small wins, small victories first, and then you gradually expand when everybody gets, becomes more comfortable and the value becomes clear. I had on the podcast a while ago,

[00:59:00] [SPEAKER_01] a company that's Aero Technologies is the name and they're building these decision-making co-pilots. And the idea is, there's a CEO would have this model in his office accessible and could,

[00:59:26] [SPEAKER_01] it could help him work through decisions of what the optimal is. It sounds like you guys-

[00:59:34] [SPEAKER_00] Very good idea. That's exactly what we can build and we actually had the prototype already. And a pandemic decision-making was exactly that. It wasn't a CEO, it was for health officials. And one important aspect of that is that when AI makes you a suggestion, you get immediately a prediction of what its effect is. So you can see the economic cost and pandemic, you can see the economic cost and you can see the number of cases. But not only that, you got to give the CEO or the

[01:00:02] [SPEAKER_00] pandemic decision-maker an ability to change those decisions and see the effects. So they can convince themselves that even though I tried a couple of things that I think might work, they don't work as well. This is the best solution. That's how you convince people and you give them, you make it more in power. They have a better idea that what they're doing is right.

[01:00:21] [SPEAKER_01] Yeah. And you think this kind of decision-making will eventually spread through the society?

[01:00:30] [SPEAKER_00] I certainly hope so. And I think it will spread when people see the value in it. So when people, at first maybe they're afraid of AI and that they don't understand it, and then maybe afraid in the sense that they might think that they've reprises them. But when they see that they can use it to their advantage to do the job better, understanding better, get more reward and satisfaction, I think that's how it starts spreading.

[01:00:51] [SPEAKER_01] Right. And Cognizant, is implementing this largely on a bespoke consultancy basis? Or are you building products that are not?

[01:01:05] [SPEAKER_00] Yes. Both. Well, we're still starting it. So we have developed basic technology, and now we are building teams that can actually go and implement these for customers. And this will expand. So we'll have more teams and maybe at some point also a product that customers can then self use. But AI is still difficult to actually deploy. So we have to have core teams

[01:01:30] [SPEAKER_00] that really know what they're doing and work together with domain experts. And that's where we are now, at this space where it requires expert knowledge that we are building and we have, and we can start with a few successful applications and then scale.

[01:01:46] [SPEAKER_01] On this combining or evolving code or self-recursive improvement of code, how does that work? I mean, if you have two code bases and you evaluate the outputs and decide that there's some good from this one, some good from that one, let's combine them. Mm-hmm. Yeah.

[01:02:16] [SPEAKER_01] How does it work? Yeah. Yeah. Yeah. So... I mean, obviously, first of all, it has to be written in the same programming language. Or is that not a problem?

[01:02:28] [SPEAKER_00] Well, different levels. But I mean, traditionally, genetic programming as a discipline works with certain language. And early on, it was Lisp because it's very easy to combine and recombine Lisp. But now there are languages that are designed for precisely to be evolved. And the idea is that they are compositional. You can take parts from one and another part from another. Now, how you do that, there's a big

[01:02:54] [SPEAKER_00] chance part of... is randomness. So you do it randomly. Okay. But it doesn't mean that it's just a random search. Not at all. Because you are taking parents that are good. And if you recombine them, chances are pretty good that your offspring is good too. But the random recombination is the creativity part. You want it to be something that is not just like, hmm, I think that's good and then that's good. Because then you are putting your

[01:03:23] [SPEAKER_00] preconceived notions in the play and you are limiting your search. But if you do it randomly, anything is possible. And that's how you make those big discoveries that you wouldn't otherwise make. So the idea is a little different from this is good and that's good, so I put them together. No, we don't do that. We say that that parent is good and that parent is good. So they both have good aspects. I wonder if I can recombine them to something even better. And you can do it many times. You can have

[01:03:48] [SPEAKER_00] lots offspring and some of them, if some of them are better than the parents, then you make progress.

[01:03:54] [SPEAKER_01] Combining, you said is composable. Combining functions or units of code?

[01:04:01] [SPEAKER_00] You have to have some way of getting an offspring that's functional, that actually is a viable

[01:04:06] [SPEAKER_01] individual. Yes. Okay. Well, I hope we can. I don't want to keep you longer.