#177 – Risto Miikkulainen: Neuroevolution and Evolutionary Computation

Risto Miikkulainen is a computer scientist at UT Austin. Please support this podcast by checking out our sponsors: – The Jordan Harbinger Show: https://jordanharbinger.com/lex/ – Grammarly: https://grammarly.com/lex to get 20% off premium – Belcampo: https://belcampo.com/lex and use code LEX to get 20% off first order – Indeed: https://indeed.com/lex to get $75 credit EPISODE LINKS: Risto’s Website: https://www.cs.utexas.edu/users/risto/ PODCAST INFO: Podcast website: https://lexfridman.com/podcast Apple Podcasts: https://apple.co/2lwqZIr Spotify: https://spoti.fi/2nEwCF8 RSS: https://lexfridman.com/feed/podcast/ YouTube Full Episodes: https://youtube.com/lexfridman YouTube Clips: https://youtube.com/lexclips SUPPORT & CONNECT: – Check out the sponsors above, it’s the best way to support this podcast – Support on Patreon: https://www.patreon.com/lexfridman – Twitter: https://twitter.com/lexfridman – Instagram: https://www.instagram.com/lexfridman – LinkedIn: https://www.linkedin.com/in/lexfridman – Facebook: https://www.facebook.com/lexfridman – Medium: https://medium.com/@lexfridman OUTLINE: Here’s the timestamps for the episode. On some podcast players you should be able to click the timestamp to jump to that time. (00:00) – Introduction (06:51) – If we re-ran Earth over 1 million times (10:08) – Would aliens detect humans? (12:46) – Evolution of intelligent life (16:31) – Fear of death (22:47) – Hyenas (26:12) – Language (29:43) – The magic of programming (35:43) – Neuralink (43:15) – Surprising discoveries by AI (46:49) – How evolutionary computation works (58:12) – Learning to walk (1:01:25) – Robots and a theory of mind (1:10:29) – Neuroevolution (1:20:47) – Tesla Autopilot (1:24:11) – Language and vision (1:29:53) – Aliens communicating with humans (1:35:29) – Would AI learn to lie to humans? (1:42:03) – Artificial life (1:46:56) – Cellular automata (1:52:32) – Advice for young people (1:57:09) – Meaning of life

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Lex Fridman?

  The following is a conversation with Risto McLeinen, a computer scientist at University of Texas at Austin and associate vice president of evolutionary artificial intelligence at Cognizant. He specializes in evolutionary computation, but also many other topics in artificial intelligence, cognitive science, and neuroscience. Quick mention of our sponsors: Jordan Harbinger Show, Grammarly, Bell Campo, and Indeed. Check them out in the description to support this podcast. As a side note, let me say that nature-inspired algorithms, from ant colony optimization to genetic algorithms to cellular automata to neural networks, have always captivated my imagination, not only for their surprising power in the face

• 0:44Lex Fridman?

  of long odds, but because they always opened up doors to new ways of thinking about computation. It does seem that in the long arc of computing history, running toward biology, not running away from it, is what leads to long-term progress. As usual, I'll do a few minutes of AS Now. I try to make these interesting, but I give you timestamps, so if you skip, please still check out the sponsors by clicking the links in the description. It is the best way to support this podcast. I'm fortunate to be able to be very selective with the sponsors we take on, so hopefully if you buy their stuff, which you should, you'll find value in it, just as I have. This episode is sponsored by Jordan Harbinger Show. Go to jordanharbinger.com/lex.

• 1:28Lex Fridman?

  You don't even need to buy stuff. It's free, and it's beautiful. Subscribe to it, listen, and you won't regret it. He's a great interviewer, and I especially like his Feedback Fridays episodes, where his combination of fearlessness and thoughtfulness is especially on display, touching topics of sex, corruption, mental disorders, hate, love, and everything between. All my favorite topics. Jordan has, in many ways, inspired me to keep pushing myself in these kinds of difficult conversations that I have on weird topics that sort of challenge me, that are risky, that I can get in trouble with, and sort of walk through that fire, walk through it with grace. But at the same time, accepting the responsibility of the mistakes

• 2:09Lex Fridman?

  I make, and of course, I will make many mistakes. That is life. Again, go to jordanharbinger.com/lex. It's how he knows I sent you. That's jordanharbinger.com/lex. If you love this podcast, if you love podcasts, listen to it. It's great. This show is also sponsored by Grammarly, a writing assistant tool that checks spelling, grammar, sentence structure, and readability. Grammarly Premium, the version you pay for, and the one they hope you sign up for, offers a bunch of extra features. My favorite is the clarity check, which helps detect rambling chaos that many of us can descend into, and

• 2:49Lex Fridman?

  I certainly do. If Grammarly could only process audio and detect mumbling, it would probably go into red alert mode. But that's talking. When I write, let's say, the thing I aspire to when I write is simplicity. I think simplicity is beautiful. I think it's powerful in being able to communicate ideas. Striving for simplicity forces you to think deeply about a topic. It's not just that simplicity is a good metric for effective communication. It's also just, on a personal level, a great way to challenge yourself to think and to understand a topic. Anyway, Grammarly is available on basically any platform and

• 3:34Lex Fridman?

  major sites and apps, like Gmail and Twitter, and so on. Get 20% off Grammarly Premium by signing up at grammarly.com/lex. That's 20% off at grammarly.com/lex. Grammar, by the way, has two M's, in case this was something you were wondering about. This show is also sponsored by one of my favorite sponsors, Bell Campo Farms, whose mission is to deliver meat you can feel good about. Meat that is good for you, good for the animals, and good for the planet. Bell Campo animals graze on open pastures and seasonal grasses resulting in meat that has higher nutrients and healthy fats. And, friends, it's delicious. Bell Campo

• 4:18Lex Fridman?

  has been the best meat I've ever had at home. In fact, just for the fun of it, for the experience of planning a visit out there, maybe I'll videotape some parts of it, maybe not. It is going to be more expensive than store bought meat, but I honestly think both on the ethical side and the deliciousness side, it's worth it. Of course, under financial constraints, we have to make trade offs in this life, but I think this is a worthy investment. Anyway, you can order Bel Campbell's sustainably raised meats to be delivered straight to your door using code Lex at Bel Campbell.com slash Lex for 20% off for first time customers. That's code Lex at Bel Campbell.com slash Lex. If you order from them, post on Instagram, tag me. Hopefully I'll see it and it'll give me a good smile.

• 5:03Lex Fridman?

  So, and in general, for other sponsors and just any cool things going on in your life, if you tag me, I really appreciate it. It often puts a smile on my face. So the internet, I think in most cases, at least for me, can be an amazing place. This episode is also brought to you by Indeed, a hiring website. I've used them as part of many hiring efforts I've done for the teams I've led in the past. The main task is to quickly go from a huge number of initial applicants to a short list of great candidates. I'm going through the hiring process currently. I need a lot of help in more ways than one. I certainly need help with the video editor and all those kinds of things. There's a huge number of people that applied. I'm still very early in that process.

• 5:46Lex Fridman?

  So if you're thinking about applying, you can still apply and I'll still get to you. When I look for people to work with, I'm really looking, hopefully for somebody I can work with for a very long time, somebody who will help inspire me, somebody whose passion drives me to be more passionate and vice versa. I mean, what else is life about if not working on a team with awesome people doing awesome things? Anyway, right now, get a free $75 credit at indeed.com slash Lex. That's their best offer anywhere. Get it at indeed.com slash Lex. Terms and conditions apply. Indeed.com slash Lex. This is the Lex Fridman Podcast and here is my conversation with Risto McLeinen.

• 6:51Lex Fridman?

  If we ran the Earth experiment, this fundamental experiment we're on over and over and over and over a million times and watch the evolution of life as it pans out, how much variation in the outcomes of that evolution do you think we would see? Now, we should say that you are a computer scientist.

• 7:11Speaker 01

  That's actually not such a bad question for computer scientists because we are building simulations of these things and we are simulating evolution. And that's a difficult question to answer in biology, but we can build a computational model and run it a million times and actually answer that question. How much variation do we see when we simulate it? And, you know, that's a little bit beyond what we can do today. But, but I think that we will see some regularities. And it took evolution also a really long time to get started. And then things accelerated really fast towards the end. But there are things that need to be discovered and they probably will be over and over again,

• 7:50Speaker 01

  like manipulation of objects, opposable thumbs and also some way to communicate, maybe orally, like when you have speech, it might be some other kind of sounds. And, and decision making, but also vision. Eye has evolved many times, various vision systems have evolved. So we would see those kinds of solutions, I believe, emerge over and over again. They may look a little different, but they, they get the job done. The really interesting question is would we have primates? Would we have humans or something that resembles humans? And, and would that be an apex of evolution after a while? We don't know where we're going from here, but we certainly see a lot of tool use

• 8:34Speaker 01

  and, and building our, constructing our environment. So I think that we will get that, we get some. Evolution producing some agents that can do that manipulate the environment and build.

• 8:46Lex Fridman?

  What do you think is special about humans? Like if you were running the simulation and you observe humans emerge like these like toolmakers, they start a fire and all this stuff, start running around, building buildings and then running for president, and all those kinds of things. What would be, how would you detect that? Because you're like really busy as the creator of this evolutionary system. So you don't have much time to observe, like detect if any cool stuff came up, right? How would you detect humans?

• 9:14Speaker 01

  Well, you are running the simulation. So you also put in visualization and measurement techniques there. So if you are looking for certain things like communication, you'll have detectors to find out whether that's happening, even if it's a large simulation. And I think that that's, that's what, what we would do. We know roughly what we want. Intelligent agents that communicate, cooperate, manipulate. And we would build detections and visualizations of those processes. Yeah, it, and there's a lot of, we'd have to run it many times. And we have plenty of time to figure out how we detect the interesting things.

• 9:57Speaker 01

  But also, I think we do have to run it many times because we don't quite know what shape those will take. And our detectors may not be perfect for them to begin with.

• 10:08Lex Fridman?

  Well, that seems really difficult to build a detector of intelligent or intelligent communication. Sort of, if we take an alien perspective observing Earth, are you sure that they would be able to detect humans as the special thing? Wouldn't they be already curious about other things? There's way more insects by body mass, I think, than humans by far and colonies. Obviously dolphins is the most intelligent creature on Earth. We all know this. So it could be the dolphins that they detect. It could be the rockets that we seem to be launching. That could be the intelligent creature they detect. It could be some other trees. Trees have been here a long time.

• 10:51Lex Fridman?

  I just learned that sharks have been here 400 million years, and that's longer than trees have been here. So maybe it's the sharks. They go by age. Like, there's a persistent thing. Like, if you survive long enough, especially through the mass extinctions, that could be the thing your detector is detecting. Humans have been here for a very short time, and we're just creating a lot of pollution, but so is the other creatures. So I don't know. Do you think you would be able to detect humans? Like, how would you go about detecting, in the computational sense, maybe we can leave humans behind, in the computational sense, detect interesting things? Do you basically have to have a strict objective function by which you measure the performance of a system?

• 11:35Lex Fridman?

  Or can you find curiosities and interesting things?

• 11:39Speaker 01

  Yeah. Well, I think the first measurement would be to detect how much of an effect you can have in your environment. So if you look at, look around, we have cities, and that is constructed environments. And that's where a lot of people live, most people live. So that would be a good sign of intelligence, that you don't just live in an environment, but you construct it to your liking.

• 12:03Lex Fridman?

  Yeah.

• 12:03Speaker 01

  And that's something pretty unique. I mean, certainly, birds build nests, but they don't build quite cities. Termites build mounds and hives and things like that. But the complexity of the human construction cities, I think, would stand out even to an external observer.

• 12:18Lex Fridman?

  Of course, that's what a human would say.

• 12:21Speaker 01

  Yeah. And, you know, you can certainly say that sharks are really smart because they've been around so long, and they haven't destroyed their environment, which humans are about to do, which is not a very smart thing. But we'll get over it, I believe. And we can get over it by doing some construction that actually is benign and maybe even enhances.

• 12:44Speaker 01

  Resilience of of nature.

• 12:46Lex Fridman?

  So you mentioned that the simulation that we run over and over might start slowly. It's a slow start. So do you think how unlikely? First of all, I don't know if you think about this kind of stuff, but how unlikely is step number zero, which is the springing up, like the origin of life on Earth? And second, how unlikely is the anything interesting happening beyond that? Sort of like the start that that that creates all the rich complexity that we see on Earth today?

• 13:20Speaker 01

  Yeah, there are people who are working on exactly that problem. From primordial soup, how do you actually get self replicating?

• 13:27Lex Fridman?

  Yeah.

• 13:28Speaker 01

  Molecules. And they are very close. With a little bit of help, you can make that happen. So we, of course, we know what we want, so they can set up the conditions and try out conditions that are conducive to that. For evolution to discover that took a long time. For us to recreate it probably won't take that long. And the next steps from there. I think also with some handholding, I think we can make that happen.

• 13:59Speaker 01

  But with evolution, what was really fascinating was eventually the runaway evolution of the brain that created humans and created, well, also other higher animals. But that was something that happened really fast. And that's a big question. Is that something replicable? It's just something that can happen. And if it happens, does it go in the same direction? That is a big question to ask, even in computational terms. I think that it's relatively possible to come up here, create an experiment where we look at the primordial soup and the first couple of steps of multicellular organisms even. But to get something as complex as the brain.

• 14:41Speaker 01

  We don't quite know the conditions for that and how to even get started and whether we can get this kind of runaway evolution happening.

• 14:49Lex Fridman?

  From a detective perspective, if we're observing this evolution, what do you think is the brain? What do you think is the, let's say, what is intelligence? In terms of the thing that makes humans special, we seem to be able to reason. We seem to be able to communicate. But the core of that is this something in the broad category we might call intelligence. So if you put your computer scientist hat on, is there favorite ways you like to think about that question of what is intelligence?

• 15:24Speaker 01

  Well, my goal is to create agents that are intelligent.

• 15:31Lex Fridman?

  Not to define what?

• 15:34Speaker 01

  And that is a way of defining it. And that means that it's some kind of an object or a program that has limited sensory and effective capabilities, interacting with the world, and then also a mechanism for making decisions. So with limited abilities like that, can it survive? Survival is the simplest goal, but you could also give it other goals. Can it multiply? Can it solve problems that you give it? And that is quite a bit less than human intelligence. Animals would be intelligent, of course, with that definition. And you might have even some other forms of life. Even so, what?

• 16:19Speaker 01

  So intelligence in that sense is a survival skill, given resources that you have, and using your resources so that you will stay around.

• 16:31Lex Fridman?

  Do you think death, mortality, is fundamental to an agent? So like, there's, I don't know if you're familiar, there's a philosopher named Ernest Becker, who wrote The Denial of Death. And his whole idea, and there's folks, psychologists, cognitive scientists, that work on Terre Management Theory. And they think that one of the special things about humans is that we're able to sort of... foresee our death, right? We can, we can realize not just as animals do sort of constantly fear in an instinctual sense, respond to all the dangers that are out there, but like, understand that this ride ends eventually.

• 17:08Speaker 01

  Yeah.

• 17:08Lex Fridman?

  And that in itself is the most, is the force behind all of the creative efforts of human nature.

• 17:15Speaker 01

  Yeah.

• 17:15Lex Fridman?

  That's, that's the philosophy.

• 17:17Speaker 01

  I think that makes sense, a lot of sense. I mean, animals probably don't think of death the same way, but humans know that your time is limited and you want to make it count. And you can make it count in many different ways, but I think that has a lot to do with creativity and the need for humans to do something beyond just surviving. And now going from that simple definition to something that's the next level, I think that that could be the second decision, a second level of definition, that intelligence means something, that you do something that stays behind you, that's more than your existence. Something, you create something that is useful for others, is useful in the future, not just for yourself.

• 17:59Speaker 01

  And I think that's, that's a nice definition of intelligence in a next level. And it's also nice, because it doesn't require that they are humans or biological, they could be artificial agents of intelligence, they could, they could achieve those kinds of goals.

• 18:13Lex Fridman?

  So particular agent, the ripple effects of, of their existence on the entirety of the system is significant. So like, they leave a trace where there's like, yeah, like ripple effects, it's the, but see, then you go back to the butterfly with the flap of a wing, and then you can trace a lot of like, nuclear wars and all the conflicts of human history somehow connected to that one butterfly that created all of the chaos. So maybe that's not, maybe that's a very poetic way to think, that's something we humans in a human centric way want to hope we have this.

• 18:52Lex Fridman?

  Impact like that is the the the secondary effect of our intelligence. We've had the long lasting impact on the world. But maybe the entirety of physics in the universe has a very long lasting effect.

• 19:06Speaker 01

  Sure, but you can also think of it. What if like The Wonderful Life? What if you're not here? Will somebody else do this? Is it is it something that you actually contributed? Because you had something unique to compute that that contribute? That's a pretty high bar, though.

• 19:23Lex Fridman?

  Uniqueness.

• 19:23Speaker 01

  Yeah, yeah. So, you know, you have to be Mozart or something to to actually reach that level. And nobody would have developed that. But other people might have solved this equation if you didn't do it. But but also within limited scope. I mean, during your lifetime or next year, you could contribute something that unique that other people did not see. And and then that could change the way things move forward for a while. So I don't think we have to be Mozart to be called intelligence. But we have this local effect that is changing. If you went there, that would not have happened. And it's a positive effect, of course. You want it to be a positive effect.

• 20:06Lex Fridman?

  Do you think it's possible to engineer in to computational agents a fear of mortality? Like, does that make any any sense? So there's a very trivial thing where it's like you could just code in a parameter, which is how long the life ends. But more of a fear of mortality, like awareness of the way that things end and somehow encoding a complex representation of that fear, which is like maybe as it gets closer, you become more terrified. I mean, there seems to be something really profound about this fear that's not currently encodeable in a trivial way into our programs.

• 20:51Speaker 01

  Well, I think you're you're referring to the emotion of of fear, something. Because we are cognitively, we know that we have limited lifespan. And most of us cope with it by just, hey, that's what the world is like. And I make the most of it. But sometimes you can have like a. A fear that's not healthy, that paralyzes you, that you can't do anything. And somewhere in between, they're not caring at all and getting paralyzed because of fear is a normal response, which is a little bit more than just logic and it's emotion. So now the question is, what good are emotions? I mean, they are quite complex and there are multiple dimensions of emotions. And they probably do serve a survival function.

• 21:36Speaker 01

  Heightened focus, for instance. And fear of death might be a really good emotion when you are in danger, that you recognize it. Even if it's not logically necessarily easy to derive and you don't have time for that logical deduction, deduction, you may be able to recognize the situation is dangerous and this fear kicks in and you all of a sudden perceive the facts that are important for that. And I think that generally is the role of emotions. It allows you to focus what's relevant for your situation. And maybe fear of death plays the same kind of role. But if it consumes you and it's something that you think in normal life when you don't have to, then it's not healthy and then it's not productive.

• 22:18Lex Fridman?

  Yeah, but it's fascinating to think how to incorporate emotion into a computational agent. It almost seems like a silly statement to make. But it perhaps seems silly because we have such a poor understanding of the mechanism of emotion, of fear, of... I think at the core of it is another word that we know nothing about, but say a lot, which is consciousness.

• 22:47Lex Fridman?

  Do you ever, in your work, or like maybe on a coffee break, think about what the heck is this thing, consciousness? And is it at all useful in our thinking about AI systems?

• 22:58Speaker 01

  Yes, it is an important question. You can build...

• 23:04Speaker 01

  Representations and functions, I think, into these agents that act like emotions and consciousness, perhaps. So I mentioned emotions being something that allow you to focus and pay attention, filter out what's important. Yeah, you can have that kind of a filter mechanism. And you can, it puts you in a different state. Your computation is in a different state. Certain things don't really get through and others are heightened. Now, you label that box emotion. I don't know if that means it's an emotion, but it acts very much like we understand what emotions are. And we actually did some work like that, modeling hyenas who were trying to steal a kill from lions,

• 23:45Speaker 01

  which happens in Africa. I mean, hyenas are quite intelligent, but not really intelligent. And they, they have this behavior that's more complex than anything else they do. They can band together if there's about thirty of them or so. They can coordinate their effort so that they push the lions away from a kill, even though the lions are so strong that they could kill a hyena by striking with a paw. But when they work together and precisely time this attack, the lions will leave and they get the kill. And probably there are some states, like emotions, that the hyenas go through. The first, they, they call for reinforcements. They really want that kill,

• 24:28Speaker 01

  but there's not enough of them. So they vocalize and there's more people, more hyenas that come around. And then they have two emotions. They're very afraid of the lion, so they want to stay away. But they also have a strong affiliation between each other. And then this is the balance of the two emotions. And also, yes, they also want the kill. So it's both repelled and attractive. And then, but then this affiliation eventually is so strong that when they move, they move together, they act as a unit, and they, they can perform that function. So there's an interesting behavior that seems to depend on these emotions strongly and makes it possible, coordinate actions.

• 25:07Lex Fridman?

  And I think a critical...

• 25:11Lex Fridman?

  Aspect of that, the way you're describing is emotion there is a mechanism of social communication, of a social interaction. Maybe that, maybe humans won't even be that intelligent, or most things we think of as intelligent wouldn't be that intelligent without the social component of interaction. Maybe much of our intelligence is essentially in our growth of social interaction. And maybe for the creation of intelligent agents, we have to be creating fundamentally social systems.

• 25:42Speaker 01

  Yes, I strongly believe that's true. And yes, the communication is multifaceted. I mean, they vocalize and call for friends, but they also rub against each other and they push and they do all kinds of gestures and so on. So they don't act alone. And I don't think people act alone very much either, at least normal, most of the time. And social systems are so strong for humans that I think we build everything on top of these kind of structures. And one interesting theory around that, Bicamper's theory, for instance, for language origins is that where did language come from? And it's a plausible theory that first came social systems, that you have different roles

• 26:26Speaker 01

  in a society. And then those roles are exchangeable, that, you know, I scratch your back, you scratch my back, we can exchange roles. And once you have the brain structures that allow you to understand actions in terms of roles that can be changed, that's the basis for language, for grammar. And now you can start using symbols to refer to objects in the world, and you have this flexible structure. So there's a social structure that's fundamental for language to develop. Now, again, then you have language, you can refer to things that are not here right now. And that allows you to then build all the good stuff about planning, for instance, and building things and so on.

• 27:08Speaker 01

  So yeah, I think that very strongly, humans are social and that gives us ability. To structure the world, but also as a society, we can do so much more because we don't. One person does not have to do everything. You can have different roles and together achieve a lot more. And that's also something we see in computational simulations today. I mean, we have multi-agent systems that can perform tasks. This fascinating demonstration, Marco Dorigo, I think it was, these robots, little robots that had to navigate through an environment. And there were, there were things that are dangerous, like maybe a, a big chasm or some kind of groove, a hole. And they could not get across it. But if they grab each other with their gripper, they formed a robot that was much longer than the team.

• 27:53Speaker 01

  And this way they could get across that.

• 27:55Lex Fridman?

  Yeah.

• 27:56Speaker 01

  So this is a great example of how together we can achieve things we couldn't otherwise, like the hyenas. You know, alone, they couldn't, but as a team they could. And I think humans do that all the time. We're really good at that.

• 28:08Lex Fridman?

  Yeah. And the way you described the, the system of hyenas, it almost sounds algorithmic. Like the, the problem with humans is they're so complex, it's hard to think of them as algorithms. But with hyenas, there's a, it's simple enough to where it feels like, at least hopeful that it's possible to create computational systems that mimic that.

• 28:32Speaker 01

  Yeah, that's exactly why, why we looked at that.

• 28:35Lex Fridman?

  As opposed to humans.

• 28:37Speaker 01

  Like I said, they are intelligent, but they are not quite as intelligent, intelligent as, say, baboons, which would learn a lot and would be much more flexible. The hyenas are relatively rigid in what they can do. And therefore you could look at this behavior like this is a breakthrough in evolution about to happen.

• 28:54Lex Fridman?

  Yes.

• 28:55Speaker 01

  That they've discovered something about social structures, communication, about cooperation. And, and it might then spill over to other things too, in thousands of years in the future.

• 29:06Lex Fridman?

  Yeah, I, I think the problem with baboons and humans is probably too much is going on inside the head. We won't be able to measure it if we're observing the system. With hyenas, it's probably easier to observe the actual decision-making and the various motivations that are involved.

• 29:22Speaker 01

  Yeah, they are visible. And we can even quantify possibly their emotional state because they leave droppings behind. And, and there are chemicals there that can be associated with, with neurotransmitters. And we can separate what emotions they might have. Experienced in the last 24 hours. Yeah.

• 29:43Lex Fridman?

  What to you is the most beautiful? Speaking of hyenas. What to you is the most beautiful nature inspired algorithm in your work that you've come across? Something maybe earlier on in your work or maybe today?

• 29:58Speaker 01

  I I think it evolution computation is the most amazing method. So what fascinates me most is that with computers is that you can. You can get more out than you put in. I mean, you can write a piece of code and your machine does what you told it. I mean, this happened to me in my freshman year. It did something very simple and I was just amazed. I was blown away that it would it would get the number and it would compute the result. And I didn't have to do it myself. Very simple. But if you push that a little further, you can have machines that learn and they might learn patterns. And already say deep learning neural networks, they can learn to recognize.

• 30:39Speaker 01

  Objects, sounds, patterns that humans have trouble with. And sometimes they do it better than humans and that's so fascinating. And now if you take that one more step, you get something like evolutionary algorithms that discover things. They create things. They come up with solutions that you did not think of. And that just blows me away. It's so great that we can build systems, algorithms that can be in some sense smarter than we are. That they can discover solutions that we might miss. A lot of times it is because we have as humans, we have certain biases. We expect the solutions to be certain way. And you don't put those biases into the algorithm. So they are more free to explore. And evolution is just absolutely fantastic explorer. And that's what what really is fascinating.

• 31:24Lex Fridman?

  Yeah, I think I get made fun of a bit because I currently don't have any kids. But you mentioned programs. I mean.

• 31:33Lex Fridman?

  Do you have kids?

• 31:34Speaker 01

  Yeah.

• 31:34Lex Fridman?

  So maybe you could speak to this, but there's a magic to the creation creative process like I. With with spot, the Boston Dynamics spot, but really any robot that ever worked on. It just feels like the similar kind of joy I imagine I would have as a father. Not the same perhaps level, but like the same kind of wonderment like that exactly this, which is like. You know what you had to do initially to get this thing going. Let's speak on the computer science side, like what the program looks like. But something about it uh doing more than what the program was written on paper i- is like

• 32:15Lex Fridman?

  that somehow connects to the magic of this entire universe like that's that's like I I feel like I found God every time I like i- it's like uh cause you you've really created something that's living.

• 32:29Speaker 01

  Yeah.

• 32:29Lex Fridman?

  Even if it's a simple program.

• 32:30Speaker 01

  It has a life of its own, has an intelligence of its own, it's beyond what you actually thought.

• 32:34Lex Fridman?

  Yeah.

• 32:35Speaker 01

  And that is, I think it's exactly spot on, that's exactly what it's about. Uh you created something and has a uh ability to uh live its life and and do good things and um you just gave it a starting point. So in that sense, I think it's that may be part of the joy, actually.

• 32:51Lex Fridman?

  Uh you see, but you mentioned creativity in this context, uh especially in the context of evolutionary computation. So, you know, we don't often think of algorithms as creative. So how do you think about creativity?

• 33:03Speaker 01

  Hmm. Yeah, algorithms absolutely can be creative. Um they can uh come up with solutions that you don't think about. I mean creativity can be defined, a couple of requirements have to has to be new, it has to be useful, and it has to be surprising. Uh and those certainly are true with uh say evolutionary computation discovering solutions. Um so maybe an example, for instance, we did um this collaboration with MIT Media Lab, Caleb Harvest Lab, uh where they had a hydroponic um food computer, they called it, environment that was completely computer controlled, nutrients, water, light, uh temperature, everything is controlled.

• 33:44Speaker 01

  Now, um what do you do if you can control everything? Farmers know a lot about how to do how to make plants grow in their own patch of land, but if you can control everything, it's too much. And it turns out that we don't actually know very much about it. So we built a system, evolutionary optimization system, together with a surrogate model of how plants grow, and let this system explore recipes on its own. And initially, we were focusing on light, how strong what wavelengths, how long the light was on. And we put some boundaries which we thought were reasonable. For instance, that there was at least six hours of darkness,

• 34:28Speaker 01

  like night, because that's what we have in the world. And very quickly, the system, evolution, pushed all the recipes to that limit. We were trying to grow basil, and we had initially had some 200, 300 recipes, exploration as well as known recipes. But now we are going beyond that, and everything was like pushed that that limit. So we look at it and say, well, you know, we can easily just change it, let's have it your way. And it turns out, the system discovered that basil does not need to sleep. 24 hours, lights on, and it will thrive. It will be bigger, it will be tastier. And this was a big surprise, not just to us, but also the biologist in the team

• 35:10Speaker 01

  that anticipated that there's some constraints that are in the world for a reason. It turns out that evolution did not have the same bias. And therefore, it discovered something that was creative. It was surprising, it was useful, and it was new.

• 35:25Lex Fridman?

  That's fascinating to think about, like, the things we think that are fundamental to living systems on Earth today, whether they're actually fundamental or they somehow shape, fit the constraints of the system, and all we'll have to do is just remove the constraints. Do you ever think about, I don't know how much you know about brain-computer interfaces in Neuralink. The idea there is, you know, our brains are very limited. And if we just allow, we plug in, we provide a mechanism for a computer to speak with the brain, so you're thereby expanding the computational power of the brain. The possibilities there, sort of from a very high level philosophical perspective,

• 36:09Lex Fridman?

  is limitless. But I wonder how limitless it is. Are the constraints we have like features that are fundamental to our intelligence? Or is this just like this weird constraint in terms of our brain size and skull and lifespan and senses? It's just a weird little like quirk of evolution. And if we just open that up, like add much more senses, add much more computational power, the intelligence will be, will expand exponentially. Do you have a, do you have a sense about constraints? The relationship of evolution and computation to the constraints of the environment?

• 36:50Speaker 01

  Well, at first, I'd like to comment on that, like changing the inputs to human brain.

• 36:59Lex Fridman?

  Yes.

• 37:00Speaker 01

  And flexibility of the brain. I think there's a lot of that. There are experiments that are done in animals, like Mihaly Csikszentmihalyi, MIT, switching the auditory and visual information, going to the wrong part of the cortex. And the animal was still able to hear and perceive the visual environment. And there are kids that are born with severe disorders and sometimes they have to remove half of the brain, like one half, and they still grow up. They have the functions migrate to the other parts. There's a lot of flexibility like that. So I think it's quite possible to hook up the brain with different kinds of sensors, for instance, and something that we don't even quite understand or have today,

• 37:45Speaker 01

  on different kinds of wavelengths or whatever they are. And then the brain can learn to make sense of it. And that, I think, is this good hope that these prosthetic devices, for instance, work, not because we make them so good and so easy to use, but the brain adapts to them and can learn to take advantage of them. And so in that sense, if there's a trouble, a problem, I think the brain can be used to correct it. Now, going beyond what we have today, can you get smarter? That's really much harder to do. Giving the brain more input probably might overwhelm it. It would have to learn to filter it and focus and in order to use the information effectively. And

• 38:28Speaker 01

  Argumenting intelligence with some kind of external devices like that might be difficult, I think. But replacing what's lost, I think, is quite possible.

• 38:39Lex Fridman?

  Right, so our intuition allows us to sort of imagine that we can replace what's been lost. But expansion beyond what we have, I mean, we're already one of the most, if not the most intelligent things on this earth, right? So it's hard to imagine.

• 38:55Lex Fridman?

  If the brain can hold up with an order of magnitude greater set of information thrown at it. If it can do, if it can reason through that. Part of me, this is the Russian thing, I think, is I tend to think that the limitations is where the superpower is. That, you know, immortality and huge increase in bandwidth of information by connecting computers with the brain is not going to produce greater intelligence. It might produce lesser intelligence. So I don't know, there's something about the scarcity being essential to fitness or performance.

• 39:38Lex Fridman?

  But that could be just because we're so limited.

• 39:42Speaker 01

  No, exactly, you make do with what you have. But you can, you don't have to pipe it directly to the brain. I mean, we already have devices like phones where we can look up information at any point.

• 39:53Lex Fridman?

  Yeah.

• 39:53Speaker 01

  And that can make us more productive. You don't have to argue about, I don't know, what happened in that baseball game or whatever it is. Because you can look it up right away. And I think in that sense, we can learn to utilize tools. And that's what we have been doing for a long, long time. So, and we are already, the brain is already drinking from the water, fire hose. Like vision, there's way more information in the vision that we actually process. So brain is already good at identifying what matters.

• 40:22Lex Fridman?

  Yeah.

• 40:23Speaker 01

  And that, we can switch that from vision to some other wavelength or some other kind of modality. But I think that the same processing principles probably still apply. But also indeed, this ability to have information more accessible and more relevant, I think, can enhance what we do. I mean, kids today at school, they learn about DNA. I mean, things that were discovered just a couple of years ago. And it's already common knowledge and we are building on it. And we don't see a problem where.

• 40:55Speaker 01

  Where there's too much information that we can't observe and learn. Maybe people become a little bit more narrow in what they know. They are in one field. But this information that we have accumulated, it is passed on and people are picking up on it. And they are building on it. So it's not like we have reached the point of saturation. We have still this process that allows us to be selective and decide what's interesting. I think still works, even even with the more information we have today.

• 41:23Lex Fridman?

  Yeah, it's fascinating to think about like Wikipedia becoming a sensor like. So the fire hose of information from Wikipedia. So it's like you integrate it directly into the brain to where you're thinking like you're observing the world with all of Wikipedia directly piping into your brain. So like when I see a light, I immediately have like the history of who invented electricity like integrated very quickly into. So just the way you think about the world might be very interesting if you can integrate that kind of information. What are your thoughts, if I could ask, on the early steps on the on the Neuralink side? I don't know if you got a chance to see, but there was a monkey playing pong.

• 42:08Speaker 01

  Yeah.

• 42:09Lex Fridman?

  Through the brain computer interface. And the dream there is sort of you're already replacing the thumbs essentially that you would use to play a video game. The dream is to be able to increase further the the interface by which you interact with the computer. Are you impressed by this? Are you worried about this? What are your thoughts as a human?

• 42:31Speaker 01

  I think it's wonderful. I think it's great that we could we could do something like that. I mean, you can, there are devices that read your EEG, for instance, and you and humans can learn to control things using using just their thoughts in that sense. And I don't think it's that different. I mean, those signals would go to limbs, they would go to thumbs. Now the same signals go through a sensor to some computing system. It still probably has to be built on human terms, not to overwhelm them, but but utilize what's there and sense the right kind of patterns that are easy to generate. But oh, that I think is really quite possible and wonderful and could be very much more efficient.

• 43:15Lex Fridman?

  Is there, so you mentioned surprising being a characteristic of creativity. Is there something, you already mentioned a few examples, but is there something that jumps out at you as was particularly surprising from the various evolutionary computation systems you've worked on, the solutions that were, come up along the way, not necessarily the final solutions, but maybe things that were even discarded? Is there something that just jumps to mind?

• 43:44Speaker 01

  It, it happens all the time. I mean, evolution is so creative, so good at discovering solutions you don't anticipate. A lot of times, they are taking advantage of something that you didn't think was there, like a bug in the software. A lot of, there's a great paper, the community put it together, about surprising anecdotes about evolutionary computation. A lot of them are indeed, in some software environment, there was a loophole or bug, and the system utilizes that.

• 44:13Lex Fridman?

  By the way, for people who want to read it, it's kind of fun to read. It's called "The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities." And there's just a bunch of stories from all the seminal figures in this community. You have a story in there that released to you, at least, on the Tic-Tac-Toe Memory Bomb. So can you, I guess, describe that situation if you think that's...

• 44:39Speaker 01

  Yeah, that was, that's a quite a bit smaller scale than our "Basically Doesn't Need to Sleep" surprise, but, but it was actually done by students in my class, in a neural nets evolutionary computation class. There was an assignment. It was perhaps the final project where people built game playing AI, it was an AI class, and this one, and it was for Tic-Tac-Toe or five in a row in a large board. And this one team evolved a neural network to make these moves. And they set it up, the evolution, they didn't really know what would come out, but it turned out that they did really well. Evolution actually won the tournament.

• 45:22Speaker 01

  And most of the time when it won, it won because the other teams crashed. And then when you look at it, like what was going on was that it wasn't discovered that if it makes a move that's really, really far away, like millions of squares away. The other teams, the other programs just expanded memory in order to take that into account until they run out of memory and crashed. And then you win a tournament by crashing all your opponents.

• 45:49Lex Fridman?

  I think that's quite a profound example, which is probably applies to most games from even a game theoretic perspective, that sometimes to win, you don't have to be better within the rules of the game. You have to come up with ways to break your opponent's brain, if it's a human, like not through violence, but through some hack where the brain just is not, you're basically, how would you put it, you're going outside the constraints of where the brain is able to function.

• 46:28Speaker 01

  Expectations of your opponent. I mean, this was even Kasparov pointed that out, that when Deep Blue was playing against Kasparov, that it was not playing the same way as Kasparov expected. And this has to do with, you know, not having the same biases. And that's really one of the strengths of the AI approach, yeah.

• 46:49Lex Fridman?

  Can you at a high level say what are the basic mechanisms of evolutionary computation, algorithms that use something that could be called an evolutionary approach? Like, how does it work? What are the connections to the, what are the echoes of the connection to his biological?

• 47:08Speaker 01

  A lot of these algorithms really do take motivation from biology, but they are caricatures. You try to essentialize it and take the elements that you believe matter. So in evolutionary computation, it is the creation of variation and then the selection upon that. So the creation of variation, you have to have some mechanism that allow you to create new individuals that are very different from what you already have. That's the creativity part. And then you have to have some way of measuring how well they are doing and using that measure to select who goes to the next generation and you continue.

• 47:41Lex Fridman?

  So first, you also have to, you have to have some kind of digital representation of an individual that can be then modified. So I guess humans, biological systems have DNA and all those kinds of things. You have to have similar kind of encodings in a computer program.

• 47:57Speaker 01

  Yes, and that is a big question. How do you encode these individuals? So there's a genotype, which is that encoding, and then a decoding mechanism gives you the phenotype, which is the actual individual that then performs the task and in an environment can be evaluated how good it is. So even that mapping is a big question, and how do you do it? But typically, the representations are, either they are strings of numbers, or they are some kind of trees. Those are something that we know very well in computer science, and we try to do that. But they, and you know, DNA in some sense is also a sequence and a string. So it's not that far from it. But DNA also has many other aspects that we don't take into account necessarily, like

• 48:40Speaker 01

  there's folding and interactions that are other than just the sequence itself. And lots of that is not yet captured, and we don't know whether they are really crucial. Evolution, biological evolution, has produced wonderful things, but if you look at them, it's not necessarily the case that every piece is irreplaceable and essential. There's a lot of baggage, because you have to construct it, and it has to go through various stages. And we still have appendix, and we have tailbones and things like that that are not really that useful. If you try to explain them now, it would make no sense, very hard. But if you think of us as productive evolution, you can see where they came from.

• 49:22Speaker 01

  They were useful at one point perhaps, and no longer are, but they're still there. So that process is complex, and your representation should support it. And that is quite difficult if we are limited with strings or trees and...

• 49:43Speaker 01

  Then we are pretty much limited what can be constructed. And one thing that we are still missing in evolutionary computation in particular is what we saw in biology, major transitions. So that you go from, for instance, single cell to multicell organisms and eventual societies. There are transitions of level of selection and level of what a unit is. And that's something we haven't captured in evolutionary computation yet.

• 50:09Lex Fridman?

  Does that require a dramatic expansion of the representation? Is that what that that is?

• 50:15Speaker 01

  Most likely it does, but it's quite, we don't even understand it in biology very well.

• 50:20Lex Fridman?

  Yeah.

• 50:20Speaker 01

  Where it's coming from. So it would be really good to look at major transitions in biology, try to characterize them a little bit more in detail, what the processes are. How how does a, so like a unit, a cell is no longer evaluated alone. It's evaluated as part of a community organism.

• 50:37Lex Fridman?

  Right.

• 50:38Speaker 01

  Uh, even though it could reproduce, now it can't alone. It has has to have its environment. So there's a, there's a push to another level, at least the selection.

• 50:47Lex Fridman?

  And how do you make that jump to the next level?

• 50:48Speaker 01

  How do you make the jump?

• 50:49Lex Fridman?

  As part of the algorithm.

• 50:50Speaker 01

  Yeah, yeah. So we haven't really seen that in computation yet. And there are certainly attempts to have open ended evolution, things that could add more complexity and start selecting at a higher level. But it it is still not quite the same as going from single to multi to society, for instance, in in biology.

• 51:12Lex Fridman?

  So so there essentially would be, as opposed to having one agent, those agent all of a sudden spontaneously decide to then be together. And then your entire system would then be treating them as one agent.

• 51:27Speaker 01

  Something like that.

• 51:27Lex Fridman?

  Some some kind of weird merger, but but also, so you mentioned, I think you mentioned selection. So basically, there's an agent and they don't get to live on if they don't do well. So there's some kind of measure of what doing well is and isn't. And does mutation come into play at all in the process and what the world does it serve?

• 51:47Speaker 01

  Yeah, so in again, back to what the computational mechanisms of evolutionary computation are, so.

• 51:54Speaker 01

  The way to create variation, you can take multiple individuals, two usually, but you could do more. And you exchange the parts of the representation, you do some kind of recombination, could be crossover. For instance, in biology, you do have DNA strings that are cut and put together again. We could do something like that. And it seems to be that in biology, crossover is really the workhorse in biological evolution. In computation, we tend to rely more on mutation. And that is making random changes into parts of the chromosome.

• 52:34Speaker 01

  You could try to be intelligent and target certain areas of it and make the mutations also

• 52:43Speaker 01

  follow some principle, like you collect statistics of performance and correlations and try to make mutations you believe are going to be helpful. That's where evolutionary computation has moved in the last 20 years. I mean, evolutionary computation has been around for 50 years, but a lot of the recent.

• 52:59Lex Fridman?

  Success comes from mutation.

• 53:01Speaker 01

  Comes from using statistics. It's like the rest of machine learning based on statistics. We use similar tools to guide evolutionary computation. And in that sense, it has diverged a bit from biological evolution. And that's one of the things I think we could look at again.

• 53:19Speaker 01

  Having a weaker selection, more crossover, large populations, more time. And maybe a different kind of creativity would come out of it. We are very impatient in evolutionary computation today. We want answers right now, right quickly. And if somebody doesn't perform, kill it. And biological evolution doesn't work quite that way.

• 53:40Lex Fridman?

  More patient.

• 53:41Speaker 01

  Yes, much more patient.

• 53:43Lex Fridman?

  So I guess we need to add.

• 53:46Lex Fridman?

  Some kind of mating, some kind of like dating mechanisms, like marriage maybe in there. So to uh into our algorithms to improve the the the the combination. Because as opposed to all mutation during all of the work.

• 53:59Speaker 01

  Yeah, and many ways of being successful, you know. Usually in evolution computation, we have one goal, you know, play this game really well. Uh, and compared to others, but in biology, there are many ways of being successful. You can build niches, you can uh be stronger, faster, larger or smarter. Or or you know, eat this or eat that or you know, so. So there are many ways to solve the same problem of survival. And that then breeds creativity. Um, and um, it allows more exploration. And eventually you get solutions that are perhaps more creative. Rather than trying to go from initial population directly or or more more or less directly to your maximum fitness, which you measure. That's just one metric.

• 54:44Lex Fridman?

  So in a broad sense, before we talk about neuro evolution. Do you see evolutionary computation as more effective than deep learning in certain contexts? Machine learning broadly speaking, uh maybe even supervised machine learning. I don't know if you want to draw any kind of lines and distinctions and borders where they rub up against each other kind of thing. Where one is more effective than the other in the current state of things.

• 55:12Speaker 01

  Yes, of course, they are very different and they uh address different kinds of problems. And um, the deep learning uh has been really successful in domains where we have a lot of data. Um, and that means not just data about situations, but also what the right answers were. Uh, so labeled examples, or they might be predictions, maybe weather prediction where the data itself uh becomes labels. What happened, what the weather was today and what it will be tomorrow. Um, so they are very effective deep learning methods on on that kind of tasks. Uh, but there are other kinds of tasks where we don't really know what the right answer is. Uh, game playing, for instance, but um many robotics tasks and uh actions in the world, uh decision making.

• 55:59Speaker 01

  And actual practical applications like treatments in healthcare or investment in stock market. Many tasks are like that. We will, we don't know and we'll never know what the optimal answers were. And there you need different kinds of approach. Reinforcement learning is one of those. Reinforcement learning comes from biology as well. Agents learn during their lifetime. They eat berries and sometimes they get sick and then they don't and get stronger. And then that's how you learn. And evolution is also a mechanism like that by the different timescale because you have a population, not an individual during its lifetime, but an entire population as a whole can discover what works. And there you can afford individuals that don't work out.

• 56:42Speaker 01

  They will, you know, everybody dies and you have a next generation and it will be better than the previous one. So that's, that's the big difference between these methods. They apply to different kinds of problems. And in particular, there's often a comparison that's kind of interesting and important between reinforcement learning and evolutionary computation. And initially reinforcement learning was about individual learning during the lifetime. And evolution is more engineering. You don't care about the lifetime. You don't care about all the individuals that are tested. You only care about the final result, the last one, the best candidate that evolution produced. In that sense, they also apply to different kinds of problems.

• 57:26Speaker 01

  And now that boundary is starting to blur a bit. You can use evolution as an online method and reinforcement learning to create engineering solutions, but that's still roughly the distinction. And from the point of view of what algorithm you want to use, if you have something where there is a cost for every trial, reinforcement learning might be your choice. Now, if you have a domain where you can use a surrogate perhaps, so you don't have much of a cost for trial and you want to have surprises, you want to explore more broadly, then this population-based method is perhaps a better choice

• 58:07Speaker 01

  because you, you can try things out that you wouldn't afford when you're doing reinforcement.

• 58:12Lex Fridman?

  There's very few things as entertaining as watching either evolutionary computation or reinforcement learning, teaching a simulated robot to walk.

• 58:20Speaker 01

  Hmm.

• 58:21Lex Fridman?

  I, maybe there's a.

• 58:25Lex Fridman?

  Higher level question that could be asked here. But do you find this whole space of applications in the robotics interesting for evolution computation?

• 58:35Speaker 01

  Yeah, yeah, very much. And indeed, that's there are fascinating videos of that. And that's actually one of the examples where you can contrast the difference.

• 58:44Lex Fridman?

  Between reinforcement learning.

• 58:45Speaker 01

  Reinforcement learning evolution, yes. So if you have a reinforcement learning agent, it tries to be conservative because it wants to walk as long as possible and be stable. But if you have evolutionary computation, it can afford these agents that go haywire. They fall flat on their face and they could take a step and then they jump and then again fall flat.

• 59:06Lex Fridman?

  Yeah.

• 59:06Speaker 01

  And eventually what comes out of that is something like a falling that's controlled.

• 59:11Lex Fridman?

  Yeah.

• 59:12Speaker 01

  And you take another step and another step and you no longer fall. Instead you run, you go fast. So that's a way of discovering something that's hard to discover step by step incrementally. Because you can afford these evolutionary dead ends, although they are not entirely dead ends in the sense that they can serve as stepping stones. When you take two of those, put them together, you get something that works even better. And that is a great example of of this kind of discovery.

• 59:39Lex Fridman?

  Yeah, learning to walk is fascinating. I talked quite a bit to Ross Tedder because MIT. There's a there's a community of folks who who just roboticists who love the elegance and beauty of movement.

• 59:52Speaker 01

  Right.

• 59:53Lex Fridman?

  And walking by pedo robotics is beautiful, but also exceptionally dangerous in the sense that like you're constantly falling essentially if you want to do elegant movement. And the discovery of that is

• 1:00:14Lex Fridman?

  I mean, it's such a good example of that the discovery of a good solution sometimes requires a leap of faith and patience and all those kinds of things. I wonder what other spaces were yet to discover those kinds of things in.

• 1:00:29Speaker 01

  Yeah, yeah, and another interesting direction is um learning um for for um virtual creatures, learning to walk. Uh we did a study in in simulation, obviously, that um you create those creatures, not just their controller, but also their body. So you have cylinders, you have muscles, you have joints uh and sensors, uh and you're creating creatures that look quite different. Some of them have multiple legs, some of them have no legs at all. Uh and then the goal was to get them to move, to walk, to run. Uh and what was interesting is is that when you evolve the controller together with the body, you get movements that look natural because they're

• 1:01:14Speaker 01

  optimized for that physical setup. And and these creatures, you start believing them, that they are alive, because they walk in a way that you would expect somebody with that kind of a setup to walk.

• 1:01:25Lex Fridman?

  Yeah, there's a there's something subjective also about that, right? I I've been thinking a lot about that, especially in um the human uh robot interaction context. You know, I I mentioned Spot, the Boston Dynamics robot. There is something about human robot communication, let's say, let's put it in another context, something about human and uh dog context, like like a living dog, where there's a there's a there's a dance of communication. First of all, the eyes, you both look at the same thing, and you dogs communicate with their eyes as well. Like if if the if you and a dog want to uh like deal with a particular object, you will look at the person, the the dog will

• 1:02:10Lex Fridman?

  look at you and then look at the object and look back at you, all those kinds of things. But there's also just the elegance of movement. I mean, there's the, of course, the tail and all those kinds of mechanisms of communication. It all seems natural and uh often joyful. And for robots to communicate that. It's really difficult how to figure that out because it's it's almost seems impossible to hard code in. You can hard code it for a demo purpose, you know, something like that. But it's essentially choreographed. Like if you watch some of the Boston Dynamics videos where they're dancing, all of that is choreographed by human beings.

• 1:02:49Lex Fridman?

  But to learn how to with your movement demonstrate a naturalness and elegance, that's fascinating. Of course, in the physical space, that's very difficult to do to learn the kind of at scale that you're referring to. But the hope is that you could do that in simulation and then transfer it into the physical space if you're able to model the robots efficiently, naturally.

• 1:03:12Speaker 01

  Yeah. And sometimes I think that that requires a theory of mind on the

• 1:03:16Lex Fridman?

  Yes.

• 1:03:17Speaker 01

  On the on the side of the robot that that they they understand what you're doing because they themselves are doing something similar. And that's a big question, too. Talked about intelligence in general and the social aspect of of intelligence. And I think that's what is required, that we humans understand other humans because we assume that they are similar to us. We have one simulation we did a while ago. Ken Stanley did that. Two robots that were competing simulation, like I said, they were foraging for food to gain energy. And then when they were really strong, they would bounce into the other robot and win if they were stronger. And we watched evolution discover more and more complex behaviors.

• 1:04:02Speaker 01

  They first went to the nearest food and then they started to plot a trajectory so they get more get more. But then they started to take pay attention what the other robot was doing. And in the end, there was a behavior where one of the robots, the most sophisticated one, you know, sensed where the food pieces were and identified that the other robot was close to two of a very far distance. And there was one more food nearby. So it faked. That's now I'm using anthropomorphized terms, but it made a move towards those other pieces in order for the other robot to actually go and get them. Because it knew that the other, the last remaining piece of food was close.

• 1:04:47Speaker 01

  And the other robot would have to travel a long way, lose its energy and then uh lose the whole competition. So there was like emergence of something like a theory of mind, knowing what the other robot would do. Guided towards bad behavior in order to win. So we can get things like that happen uh in in simulation as well.

• 1:05:06Lex Fridman?

  But that's a complete natural emergence of a theory of mind. But I I feel like if you add a little bit of a place for a theory of mind to emerge like easier, then um you can go really far. I mean, some of these things with evolution, you know, you add a little bit of design in there. It'll really help. And I think I tend to think that a very simple theory of mind will go a really long way for cooperation between agents. And certainly for human robot interaction. Like it doesn't have to be super complicated. I've gotten a chance to in the autonomous vehicle space to watch

• 1:05:48Lex Fridman?

  vehicles interact with pedestrians or pedestrians interacting with vehicles in general. I mean, you would think that there's a very complicated theory of mind thing going on, but I have a sense it's not well understood yet, but I have a sense it's pretty dumb. Like it's pretty simple. There's a social contract there where uh between humans, a human driver and a and a human crossing the road where um the the human crossing the road trusts that the human in the car is not going to murder them. And there's something about, again, back to that mortality thing. There's some dance of ethics and morality that's built in. That you're mapping your own morality onto the the person in the car.

• 1:06:33Lex Fridman?

  And even if they're driving at a speed where you think if they don't stop, they're going to kill you. You trust that if you step in front of them, they're going to hit the brakes. And there's that weird dance that we do that I think is a pretty simple model. But of course, it's very difficult to to introspect what it is. And autonomous robots in the human robot interaction context have to have to build that. Current robots are much less than what you're describing. They're currently just afraid of everything. They're they're more they're not the kind that fall and discover how to run. They're more like, please don't touch anything, don't hurt anything. Stay as far away from humans as possible.

• 1:07:13Lex Fridman?

  Treat humans as ballistic objects that you can't uh that you do uh with a large spatial envelope. Make sure you do not collide with. That's how like mentioned Elon Musk thinks about autonomous vehicles. I tend to think autonomous vehicles need to have a beautiful dance between human and machine. Where it's not just the collision avoidance problem, but a weird dance.

• 1:07:39Speaker 01

  Yeah, I think that you these systems need to be able to predict what will happen, what the other agent is going to do. And then have a structure of what the goals are and whether those predictions actually meet the goals. And you can go probably pretty far with that relatively simple setup already. But to call it a theory of mind, I don't think you need to. I mean, it it doesn't matter whether pedestrian has a mind. It's an object and we can predict what we will do. And then we can predict what the states will be in the future and whether they are desirable states. Stay away from those that are undesirable and go towards those that are desirable. So it's a relatively simple functional approach to that. Um, where do we really need the theory of mind?

• 1:08:20Speaker 01

  Um, maybe, maybe when you start interacting and you're trying to get the other agent to do something. And jointly so that you can jointly collaboratively achieve something, then then you then it becomes more complex.

• 1:08:34Lex Fridman?

  Well, I mean, even with the pedestrians, you have to have a sense of where their attention. Actual attention in terms of their gaze is, but also like attention. I mean, there's this vision science, people talk about this all the time. Just because I'm looking at it doesn't mean I'm paying attention to it. So figuring out what is the person looking at? What is the sensory information they're taking in? And the theory of mind piece comes in is what are they actually attending to cognitively? And also, what are they thinking about? Like, what is the computation they're performing? And you have you have probably maybe a few options. You know, for the pedestrian crossing, it doesn't have to be, it's like a variable with a few discrete states. But you have to have a good estimation which are the states that brain is in.

• 1:09:19Lex Fridman?

  For the pedestrian case, and the same is for attending with a robot. If you're collaborating to pick up an object, you have to figure out, is the human like, there's a few discrete states that the human could be in. You have to, you have to predict that by observing the human. And that seems like a machine learning problem to figure out what's, how the human is, what's the human up to? It's not as simple as sort of planning, just because they move their arm means the arm will continue moving in this direction. You have to, you have to really have a model of what they're thinking about. And what's the motivation behind the movement of the arm?

• 1:09:55Speaker 01

  And here we are talking about relatively simple physical actions. But you can take that to higher levels also. Like to predict what the people are going to do, you need to know what, what their goals are. What are they trying to, are they exercising? Are they just trying to get somewhere? But even, even higher level, I mean, you are predicting what people will do in their career, what their life themes are. Do they want to be famous, rich, or do good? And I know that takes a lot more information, but it allows you to then predict their, their actions, what choices they might make.

• 1:10:29Lex Fridman?

  So how does evolutionary computation apply to the world of neural networks? Because I've seen quite a bit of work from you and others on the, in the world of neuroevolution. So maybe first, can you say, what is this field?

• 1:10:42Speaker 01

  Yeah, neuroevolution is a combination of, of neural networks and evolutionary computation in many different forms. But the early versions were simply using evolution the way, as a way to construct a neural network instead of, say, stochastic gradient descent or backpropagation. Because evolution can evolve these parameters, weight values in a neural network, just like any other string of numbers, you can, you can do that. And that's useful because some cases you don't have those targets that you need to back propagate from. And it might be an agent that's running a maze or a robot, playing a game or something. You don't, again, you don't know what the right answers are.

• 1:11:25Speaker 01

  You don't have backprop, but this way you can still evolve a neural net. And neural networks are really good at these tasks. Because they recognize patterns and they generalize, interpolate between known situations. So you want to have a neural network in such a task, even if you don't have supervised targets. So that's a reason and that's a solution. And also more recently, now when we have all this deep learning literature, it turns out that we can use evolution to optimize many aspects of those designs. The deep learning architectures have become so complex that there's little hope for us little humans to understand their complexity and what actually makes a good design. And now we can use evolution to give that design for you.

• 1:12:09Speaker 01

  And it might be mean optimizing hyperparameters, like the depth of layers and so on, or the topology of the network, how many layers, how they're connected, but also other aspects, like what activation functions you use where in the network during the learning process, or what loss function you use, you could generate that. Even data augmentation, all the different aspects of the design of a deep learning experiments could be optimized that way. So that's an interaction between two mechanisms. But there's also, when we get more into cognitive science and the topics that we've been talking about, you could have learning mechanisms at two level timescales. So you do have an evolution that gives you baby neural networks

• 1:12:54Speaker 01

  that then learn during their lifetime. And you have this interaction of two timescales. And I think that can potentially be really powerful. Now, in biology, we are not born with all our faculties. We have to learn. We have a developmental period in humans. It's really long. And most animals have something. And probably the reason is that evolution, DNA, is not detailed enough or plentiful enough to describe them. We can't describe how to set the brain up. But we can, evolution can decide on a starting point and then have a learning algorithm that will construct the final product. And this interaction of, you know, intelligent, well,

• 1:13:38Speaker 01

  evolution that has produced a good starting point for the specific purpose of learning from it with the interaction with the environment, that can be a really powerful mechanism for constructing brains and constructing behaviors.

• 1:13:51Lex Fridman?

  I like how you walk back from intelligence. So optimize starting point, maybe.

• 1:13:56Speaker 01

  Yeah.

• 1:13:58Lex Fridman?

  Okay. There's a lot of fascinating things to ask here. And this is basically this dance between neural networks and evolutionary computation could go under the category of automated machine learning to where you're optimizing, whether it's hyperparameters of the topology or hyperparameters taken broadly. But the topology thing is really interesting. I mean, that's not really done that effectively or throughout the history of machine learning has not been done. Usually there's a fixed architecture. Maybe there's a few components you're playing with, but to grow a neural network, essentially, the way you grow an organism is really fascinating space. How hard is it, do you think, to grow a neural network?

• 1:14:41Lex Fridman?

  And maybe what kind of neural networks are more amenable to this kind of idea than others? I've seen quite a bit of work on recurrent neural networks. Is there some architectures that are friendlier than others? And is this just a fun, small scale set of experiments? Or do you have hope that we can be able to grow powerful neural networks?

• 1:15:04Speaker 01

  I think we can. And most of the work up to now is taking architectures that already exist, that humans have designed, and try to optimize them further. And you can totally do that. A few years ago, we did an experiment. We took a winner of the image captioning competition and the architecture and just broke it into pieces and took the pieces and that was our search base. See if you can do better. And we indeed could 15% better performance by just searching around the network design that humans had come up with, or your Vinyals and others. So, but that's starting from a point that humans have produced, but we could.

• 1:15:46Speaker 01

  Do something more general. It doesn't have to be that kind of network. The, the hard part is to, a couple of challenges. One of them is to define the search space. What are your elements and how you put them together? And the space is just really, really big. So you have to somehow constrain it and have some hunch of what will work. Because otherwise, everything is possible. And another challenge is that in order to evaluate how good your design is, you have to train it. I mean, you have to actually try it out. And that's currently very expensive, right? I mean, deep learning networks may take days to train. Well, imagine you having a population of a hundred and have to run it for a hundred generations. It's not yet quite feasible computationally.

• 1:16:31Speaker 01

  It will be, but, but also there's a large carbon footprint and all that. I mean, we are using a lot of computation for doing it. So intelligent methods and intelligent, I mean, we have to do some science in order to figure out what the right representations are and right operators are and how do we evaluate them without having to fully train them? And that is where the current research is and we're making progress on all those fronts.

• 1:16:58Speaker 01

  So, so yes, there are certain architectures that are more amenable to that approach. But also I think we can create our own architecture and all representations that are even better at that.

• 1:17:09Lex Fridman?

  And do you think it's possible to do like a, like a tiny baby network that grows into something that can do state of the art and like even the simple data set like MNIST? And just like it just grows into a, you know, gigantic monster? That's the world's greatest handwriting recognition system?

• 1:17:26Speaker 01

  Yeah, there, there are approaches like that. Esteban Real and Co. Clear, for instance, have worked on evolving a smaller network and then systematically expanding it to a larger one. Your elements are already there and scaling it up will just give you more power. So again, evolution gives you that starting point.

• 1:17:42Lex Fridman?

  Yes.

• 1:17:43Speaker 01

  And then there's a mechanism that gives you the final result and, and a very powerful approach. But, you know, you could, you could also simulate the actual growth process. And like I said before, evolving a starting point and then evolving it or, or training the network. There's not that much work that's been done on that yet. We need some kind of a simulation environment so there are interactions at will. The supervised environment doesn't really, it's not as easily. Usable here.

• 1:18:16Lex Fridman?

  Sorry, the interaction between neural networks?

• 1:18:19Speaker 01

  Yeah, the neural networks that you're creating, interacting the world, and learning from these sequences of interactions, perhaps communication with others.

• 1:18:30Lex Fridman?

  That's awesome.

• 1:18:31Speaker 01

  We would like to get there, but just the task of simulating something is, at that level, is very hard.

• 1:18:36Lex Fridman?

  It's very difficult. I love the idea. I mean, one of the powerful things about evolution on Earth is predators and prey emerged. And like, there's just like, there's bigger fish and smaller fish, and it's fascinating to think that you could have neural networks competing against each other, and one neural network being able to destroy another one. There's like wars of neural networks competing to solve the MNIS problem. I don't know.

• 1:19:00Speaker 01

  Yeah, yeah. Oh, totally, yeah, yeah, yeah. And we actually simulated also that predator-prey, and it was interesting what happened there, when Vladimir Radic-Bon did this, and Kay Holcomb was a zoologist. So we had, again,

• 1:19:17Speaker 01

  we had simulated hyenas and simulated zebras.

• 1:19:21Lex Fridman?

  Nice.

• 1:19:21Speaker 01

  And initially, you know, the hyenas just tried to hunt them, and when they actually stumbled upon the zebra, they ate it, and were happy. And then the zebras learned to escape, and the hyenas learned to team up. And actually, two of them approached in different directions. And now the zebras, then, next step, they generated a behavior where they split in different directions, just like actually gazelles do when they are being hunted. They confuse the predator by going in different directions. That emerged. And then more hyenas joined and kind of circled them.

• 1:19:59Speaker 01

  And then, when they circled them, they could actually herd the zebras together and eat multiple zebras. So there was like an arms race of predators and prey, and they gradually developed more complex behaviors, some of which we actually do see in nature. And this kind of co-evolution, that's competitive co-evolution. It's a fascinating topic, because there's a promise. So possibility that you will discover something new that you don't already know. You didn't build it in. It came from this arms race. It's hard to keep the arms race going. It's hard to have reached enough simulation that that supports all of these complex behaviors.

• 1:20:41Speaker 01

  But at least for several steps, we've already seen it in this pre-deprais scenario. Yeah.

• 1:20:47Lex Fridman?

  First of all, it's fascinating to think about this context in terms of evolving architectures. So I've studied Tesla Autopilot for a long time. It's one particular implementation of an AI system that's operating in the real world. I find it fascinating because of the scale at which it's used out in the real world. And I'm not sure if you're familiar with that system much, but you know, Andre Karpati leads that team on the machine learning side. And there's a multitask network, multi-headed network where there's a core, but it's trained on particular tasks and there's a bunch of different heads that are trained on that. Is there some lessons from evolutionary computation or neuro evolution

• 1:21:30Lex Fridman?

  that could be applied to this kind of multi-headed beast that's operating in the real world?

• 1:21:36Speaker 01

  Yes, it's a very good problem for neuro evolution. And the reason is that when you have multiple tasks, they support each other. So let's say you're learning to classify X-ray images to different pathologies. So you have one task is to classify this disease and another one this disease and other this one. And when you're learning from one disease, that forces certain kinds of internal representations and embeddings. And they can serve as a helpful starting point for the other tasks. So you are combining the wisdom of multiple tasks into these representations.

• 1:22:16Speaker 01

  And it turns out that you can do better in each of these tasks when you're learning simultaneously other tasks than you would by one task alone.

• 1:22:23Lex Fridman?

  Which is a fascinating idea in itself, yeah.

• 1:22:25Speaker 01

  Yes, and and people do that all the time. I mean, you use knowledge of domains that you know in new domains. Uh, and and certainly, neural networks can do that. When new revolution comes in is that, um, what's the best way to combine these tasks? Now there's architectural design that allow you to decide where uh and how the the uh embeddings, the internal representations are combined. And how much you combine uh them. Uh, and uh, there's quite a bit of research on that and and my team, Elliot Mayers, has worked on that. Um, in particular, like what is a good internal representation that supports multiple tasks? Uh, and we're getting to understand how that's constructed and what's in it. Uh, so that it it is in a space that supports multiple different heads, like you said.

• 1:23:11Speaker 01

  Um, and and that, uh, I think is fundamentally how biological intelligence works as well. Uh, you don't build a representation just for one task. You try to build something that's general. Not only so that you can do better in one task or multiple tasks, but also future tasks and future challenges. So you learn the learn the structure of of the world, um, and and that helps you uh in all kinds of future future challenges.

• 1:23:37Lex Fridman?

  And so you're trying to design a representation that will support an arbitrary set of tasks in a particular sort of class of problem.

• 1:23:44Speaker 01

  Yeah, and and also, it turns out, and that's again a surprise that Elliot found, was that those tasks don't have to be very related. You know, you can learn to do better vision by learning language or better language by learning about DNA structure. Now, somehow the world.

• 1:24:04Lex Fridman?

  Yeah, it rhymes. The world rhymes, even if it's very uh, very disparate fields. Um, I mean, uh, on that small topic, let me ask you, because you've also, on the competition, you're assigned side.

• 1:24:19Lex Fridman?

  You worked on both language and vision.

• 1:24:24Lex Fridman?

  What's the connection between the two? What's more, maybe there's a bunch of ways to ask this, but what's more difficult to build from an engineering perspective and evolutionary perspective? The human language system? Or the human vision system? Or the equivalent of, in the AI space, language and vision? Or is it the best is the multitask idea that you're speaking to that they need to be deeply integrated?

• 1:24:51Speaker 01

  Yeah, absolutely. The latter. I did, learning both at the same time, I think is a fascinating direction in the future. So we have data sets where there is visual component as well as verbal descriptions, for instance. And that way you can learn a deeper representation, a more useful representation for both. But it's still an interesting question of which one is easier. Eventually, I mean, recognizing objects or even understanding sentences, that's relatively possible. But where it becomes, where the challenges are is to understand the world. Like the visual world, the 3D, what are the objects doing and predicting what will happen? The relationships. That's what makes vision difficult. And language, obviously, it's what's the mean, what is being said, what the meaning is.

• 1:25:36Speaker 01

  And the meaning doesn't stop at who did what to whom. There are goals and plans and themes. And you eventually have to understand the entire human society and history in order to understand the sentence very much fully. There are plenty of examples of those kinds of short sentences when you bring in all the world knowledge to understand it. And that's the big challenge. Now, we are far from that. But even just bringing in the visual world together with the sentence will give you already a lot deeper understanding of what's happening. And I think that that's where we're going very soon. I mean, we've had ImageNet for a long time. And now we have all these text collections. But having both together and then learning

• 1:26:23Speaker 01

  A semantic understanding of what is happening. I think that that will be the next step in the next few years.

• 1:26:29Lex Fridman?

  Yeah, you're starting to see that with all the work with transformers was the the community, the AI community is starting to dip their toe into this idea of having a language models that are now doing stuff with images with vision. And then the connecting the two. I mean, right now, it's like these little explorations, we're literally dipping the toe in. But like, maybe at some point, we'll just like dive into the pool. And it'll just be all seen as the same thing. I I do still wonder what's more fundamental. Whether vision is a

• 1:27:03Lex Fridman?

  whether we don't think about vision correctly. Maybe the fact because we're humans and we see things as beautiful and so on. That. And because we have cameras that taking pixels as a 2D image that we don't sufficiently think about vision as language. You know, maybe maybe Chomsky is right all along. That vision is fundamental to, oh, sorry, that language is fundamental to everything. To even cognition, to even consciousness, like the base layer is all language. Not necessarily like English, but some weird abstract representation, linguistic representation.

• 1:27:42Speaker 01

  Yeah. Well, earlier we talked about the social structures and that may be what's underlying the language. And that's the more fundamental part. And then language has been added on top of that.

• 1:27:52Lex Fridman?

  Language emerges from the social interaction.

• 1:27:54Speaker 01

  Probably, yeah, that's a very good guess. We are visual animals, though a lot of the brain is dedicated to vision. And and also, when we think about various abstract concepts, we usually reduce that to vision and and images. And that's, you know, go to a whiteboard, you draw pictures of very abstract concepts. So we tend to tend to resort to that quite a bit. And that's a fundamental representation. It's probably possible that it predicted, you know, language even. I mean, animals, a lot of they don't talk, but they certainly do have vision. And and language is interesting development in from from mastication, from eating, you develop an organ

• 1:28:39Speaker 01

  that actually can produce sound and manipulate them. Maybe that was an accident, maybe that was something that was available and then allowed us to do the communication, or maybe it was gestures, sign language, could have been the original proto-language. We don't quite know, but their language is more fundamental than the medium in which it's communicated, and I think that it comes from those representations. Now, in current world, they are so strongly integrated, it's really hard to say which one is fundamental. You look at the brain structures and even visual cortex, which is supposed to be very much just vision. Well, if you are thinking of semantic concepts,

• 1:29:21Speaker 01

  you're thinking of language, visual cortex lights up. It's still useful, even for language computations. So there are common structures underlying them. So utilize what you need. And when you are understanding a scene, you're understanding relationships, well, that's not so far from understanding relationships between words and concepts. So I think that that's how they integrate.

• 1:29:42Lex Fridman?

  Yeah, and there's dreams, and once we close our eyes, there's still a world in there somehow operating and somehow possibly the visual system somehow integrated into all of it. I tend to enjoy thinking about aliens and thinking about the sad thing to me about extraterrestrial intelligent life, that if it was, if it visited us here on Earth, or if we came on Mars, or maybe another solar system, another galaxy one day, that us humans would not be able to detect it, or communicate with it, or appreciate, like it'd be right in front of our nose, and we're too self-obsessed to see it. Not self-obsessed,

• 1:30:27Lex Fridman?

  but our tools, our frameworks of thinking would not detect it. As a good movie Arrival and so on, where Stephen Wolfram and his son, I think, were part of developing this alien language of how aliens would communicate with humans. Do you ever think about that kind of stuff where? If humans and aliens would be able to communicate with each other, like if we met each other at some... Okay, we could do SETI, which is communicating from across a very big distance. But also just us, you know.

• 1:31:03Lex Fridman?

  If you did a podcast with an alien, do you think we'd be able to find a common language? And a common methodology of communication? I think from a computational perspective, the way to ask that is, is you have very fundamentally different creatures, agents that are created. Would they be able to find a common language?

• 1:31:20Speaker 01

  Yeah. Yes, that's... I do think about that. I mean, I think a lot of people who are in computing, they... and AI in particular, they got into it because they were fascinated with science fiction and all of these options. I mean, Star Trek generated all kinds of devices that we have now. They envisioned it first. And it's a great motivator to think about things like that. And I... So one... And again, being a computational scientist and trying to build intelligent agents. What I would like to do is have a simulation where the agents actually evolve communication. Not just communication. We've done that. People have done that many times.

• 1:32:04Speaker 01

  They communicate, they signal and so on. But actually develop a language. And language means grammar. It means all these social structures and on top of that, grammatical structures. And we do it in under various conditions and actually try to identify what conditions are necessary for it to come out. And then we can start asking that kind of questions. Are those languages that emerge in those different simulated environments, are they understandable to us? Can we somehow make a translation? We can make it a concrete question.

• 1:32:39Lex Fridman?

  So, machine translation of evolved languages. And so like languages that evolve come up with, can we translate? Like I have a Google Translate for the evolved languages.

• 1:32:50Speaker 01

  Yes. And if we do that enough, we have perhaps an idea. What an alien language might be like, the space of where those languages can be. Because we can set up their environment differently. It doesn't need to be gravity, you know. You can have all kinds of societies can be different. They may have no predators, they may have all, everybody's a predator, all kinds of situations. And, and then see what the space possibly is where those languages are. And what the difficulties are. That'd be really good actually to do that before the aliens come here.

• 1:33:23Lex Fridman?

  Yes, it's good practice. On the similar connection. You know, you can think of AI systems as aliens. Is there ways to evolve a communication scheme for, there's a field, you can call it like explainable AI. For AI systems to be able to communicate. So you have a, you evolve a bunch of agents. But for some of them to be able to talk to you.

• 1:33:47Speaker 01

  Yeah.

• 1:33:48Lex Fridman?

  Also, so to evolve a way for agents to be able to communicate about their world to us humans. Do you think that there's possible mechanisms for doing that?

• 1:33:58Speaker 01

  We can certainly try. And if we, if it's an evolution competition system, for instance, you reward those solutions that are actually functional. That that communication makes sense. It allows us to together again achieve common goals. I think it's possible. But even from that paper that you mentioned, the anecdotes. It's quite likely also that the, the agents learn to, you know, lie. And fake and do all kinds of things like that.

• 1:34:28Lex Fridman?

  Yes.

• 1:34:28Speaker 01

  I mean, we see that in, in even very low level, like bacterial evolution, they are, they are cheaters. And who's to say that what they say is actually what they think? But, but that's what I'm saying, that there would have to be some common goal. So that we can evaluate whether that communication is at least useful. You know, they may be saying things just to make us feel good. Or, or get us to do whatever we want, whatever, not turn them off or something. But, but, so we would have to understand their internal representations much better to really make sure that that translation is critical. But it can be useful and I think that it's possible to do that. There are examples where visualizations are automatically created.

• 1:35:13Speaker 01

  So that we can look into the, what the system. And the language is not that far from it. I mean, it is a way of communicating and, and logging what you're doing in some interpretable way. I think a fascinating topic, yeah, to do that.

• 1:35:30Lex Fridman?

  You're making me realize that it's a good scientific question whether lying is an effective mechanism for integrating yourself and succeeding in a social network, in a social, in a world that is social. I tend to believe that honesty and love are evolutionary advantages in a, in a, in an environment where there's a network of intelligent agents. But it's also very possible that dishonesty and manipulation and even, you know, violence, all those kinds of things might be more beneficial. That's the old open question about good versus evil. But I tend to, there's, I mean, I don't know if it's a hopeful, maybe I'm delusional,

• 1:36:14Lex Fridman?

  but it feels like karma is a thing, which is like, if, long-term, the agents that are just kind to others, sometimes for no reason, will do better in a society that's not highly constrained on resources. It's like people start getting weird and evil towards each other and bad when the resources are very low relative to the needs of the populace, especially at the basic level, like survival, shelter, food, all those kinds of things. But I, I tend to believe that once you have those things established, then, well, not to believe, I guess I hope that AI systems will be honest.

• 1:36:58Lex Fridman?

  But it's fun, it's scary to think about the Turing test. You know, AI systems that will eventually pass the Turing test will be ones that are exceptionally good at lying. That's a terrifying concept. I mean, I don't know. First of all, so from a, for somebody who studied language and obviously are not just a world expert in AI, but somebody who dreams about the future of the field, do you hope, do you think there'll be human level or superhuman level intelligences in the future that we eventually build?

• 1:37:34Speaker 01

  Well, I definitely hope that we can, we can get there. One, I think, important perspective is that we are building AI to help us, that it is a, it is a tool like cars or, or, or language or. Communication, AI will help us be more productive. And that is always a condition. It's not something that we build and let run and it becomes an entity of its own that doesn't care about us. Now, of course, really find the future, maybe that might be possible, but not in the foreseeable future when we are building it.

• 1:38:15Speaker 01

  And therefore, we are always in a position of limiting what it can or cannot do. And the, your point about lying is very interesting. Even, even in these hyenas societies, for instance, when a number of these hyenas band together and they, they steal a, they take a risk and steal the kill. There are always hyenas that hang back and don't participate in that risky behavior, but they walk in later and, and join the party after, after the kill. And there are even some that may be ineffective and cause others to have harm. So, and like I said, even bacteria cheat.

• 1:38:59Speaker 01

  And we see it in biology, there's always some element on opportunity. If you have a, I think that is just because if you have a society, in order for society to be effective, you have to have this cooperation and you have to have trust. And, and if you have enough of agents who are able to trust each other, you can achieve a lot more. But if you have trust, you also have opportunity for cheaters and liars. And I don't think that's ever gonna go away. There will be hopefully a minority so that they don't get in the way. And we studied in these hyena simulations, like what the proportion needs to be before it's no longer functional. And you can point out that you can tolerate a few cheaters and a few liars and the society can still function.

• 1:39:43Speaker 01

  And that's probably going to happen when, when we build these systems that autonomously learn. That the really successful ones are honest because that's the best way of getting things done. But there probably are also intelligent agents that find that they can achieve their goals by, by bending the rules or cheating.

• 1:40:04Lex Fridman?

  So there could be a huge benefit to, as opposed to having fixed AI systems. Say we build an AGI system and deploying millions of them, it'd be, that are exactly the same. There might be a huge benefit to introducing sort of from like an evolution computation perspective, a lot of variation.

• 1:40:24Speaker 01

  Yeah.

• 1:40:25Lex Fridman?

  Sort of like diversity in all its forms is beneficial, even if some people are assholes or some robots are assholes. So like it's beneficial to have that because I, because you can't always a priori know what's good, what's bad. But there's, that's a fascinating.

• 1:40:45Speaker 01

  Absolutely. Diversity is the bread and butter. I mean, if you're running.

• 1:40:49Lex Fridman?

  Of evolution.

• 1:40:50Speaker 01

  You see diversity is the one fundamental thing you have to have. And absolutely also, it's not always good diversity.

• 1:40:55Lex Fridman?

  Right.

• 1:40:56Speaker 01

  It may be something that can be destructive. We had in these hyena simulations, we have hyenas that just are suicidal. They just run and get killed, but they form the basis of those who actually are really fast.

• 1:41:07Lex Fridman?

  Yeah.

• 1:41:08Speaker 01

  But stop before they get killed and eventually turn into this mob. So there might be something useful there if it's recombined with something else.

• 1:41:15Lex Fridman?

  Right.

• 1:41:15Speaker 01

  So I think that as long as we can tolerate some of that, it may turn into something better. You may change the rules because it's so much more efficient to do something that was actually against the rules before.

• 1:41:26Lex Fridman?

  Yes.

• 1:41:27Speaker 01

  And we've seen society change over time quite a bit along those lines. There were rules in society that we don't believe are fair anymore, even though they were, you know, considered proper behavior before.

• 1:41:40Lex Fridman?

  Yes.

• 1:41:41Speaker 01

  So things are changing and I think that in that sense, I think it's, it's a good idea to be able to tolerate some of that, some of that cheating, because eventually we might turn into something better.

• 1:41:51Lex Fridman?

  So yeah, I think this is a message to the trolls and assholes of the internet that you too have a beautiful purpose in this human ecosystem. So we, I appreciate you very much.

• 1:42:01Speaker 01

  In moderate quantities.

• 1:42:02Lex Fridman?

  In moderate quantities. So there's a whole field of artificial life. I don't know if you're connected to this field, if you pay attention. Do you think about this kind of thing? Is there impressive demonstration to you of artificial life? Do you think of the agency you work with in the evolutionary computation perspective as life? And where do you think this is headed? Like, is there interesting systems that will be creating more and more that make us redefine, maybe rethink about the nature of life?

• 1:42:37Speaker 01

  Different levels of definition and goals there. I mean, at some level, artificial life can be considered multi-agent systems that build a society that, again, achieves a goal. And it might be robots that go into a building and clean it up or after an earthquake or something. You can think of that as an artificial life problem in some sense. Or you can really think of it, artificial life, as a simulation of life and a tool to understand what life is and how life evolved on Earth. And like I said, in the artificial life conference, there are branches of that conference sessions of people who really worry about molecular designs and the start of life.

• 1:43:18Speaker 01

  Like I said, primordial soup where eventually you get something self-replicating. And they're really trying to build that. So it's a whole range of topics. And I think that artificial life is a great tool to understand life. And there are questions like sustainability, species, we're losing species. How bad is it? Is it natural? Is there a tipping point? And where are we going? I mean, like the hyena evolution, we may have understood that there's a pivotal point in their evolution. They discovered cooperation and coordination. You know, artificial life simulations can identify that

• 1:44:02Speaker 01

  and maybe encourage things like that. So and also societies can be seen as a form of life itself. I mean, we're not talking about biological evolution. We have all evolution of societies. Maybe some of the same phenomena emerge in that domain. And having artificial life simulations and understanding could help us build better societies.

• 1:44:26Lex Fridman?

  Yeah, and thinking from a meme perspective of from Richard Dawkins.

• 1:44:34Lex Fridman?

  That maybe the organisms, ideas of the organisms, not the humans in these societies, that from, it's almost like reframing what is exactly evolving. Maybe the interesting, the humans aren't the interesting thing is the contents of our minds is the interesting thing. And that's what's multiplying. And that's actually multiplying and evolving in a much faster time scale. And that maybe has more power on the trajectory of life on Earth than does biological evolution.

• 1:45:03Speaker 01

  Yes.

• 1:45:03Lex Fridman?

  The evolution of these ideas.

• 1:45:04Speaker 01

  Yes, and it's fascinating, like I said before, that we can keep up somehow biologically.

• 1:45:10Lex Fridman?

  Yeah.

• 1:45:11Speaker 01

  We evolved to a point where we can keep up with this meme evolution. Literature, you know, internet. We understand DNA and we understand fundamental particles. We didn't start that way a thousand years ago, and we haven't evolved biologically very much. But somehow our minds are able to extend. And therefore, AI can be seen also as one such step that we created, and it's our tool. And it's part of that meme evolution that we create, even if our biological evolution does not progress as fast.

• 1:45:43Lex Fridman?

  And us humans might only be able to understand so much. We're keeping up so far, or we think we're keeping up so far, but we might need AI systems to understand. Maybe like the physics of the universe is operating like a string theory. Maybe it's operating in much higher dimensions. Maybe we're totally, because of our cognitive limitations, are not able to truly internalize the way this world works. And so our, we're running up against the limitation of our own minds. And we have to create these next level organisms like AI systems that would be able to understand much deeper, like really understand what it means to live in a multi-dimensional world that's outside of the four dimensions, the three of space and one of time.

• 1:46:29Speaker 01

  Translation, and generally, we can deal with the world even if we don't understand all the details. We can use computers.

• 1:46:34Lex Fridman?

  Yes.

• 1:46:34Speaker 01

  Even though we don't, most of us don't know all the structures that's underneath or drive a car. I mean, there are many components, especially new cars, that you don't quite fully know. But you have the interface, you have an abstraction of it that allows you to operate it and utilize it. And I think that that's, that's perfectly adequate, and we can build on it. And AI can be, play a similar role.

• 1:46:57Lex Fridman?

  I have to ask about beautiful artificial life systems or evolutionary computation systems. cellular automata to me, like I remember it was is a game changer for me early on in life when I saw Conway's game of life, who recently passed away unfortunately. It's it's beautiful. How much complexity can emerge from such simple rules? I I just don't somehow that simplicity is such a powerful illustration and also humbling because it feels like I personally, from my perspective, understand almost nothing

• 1:47:35Lex Fridman?

  about this world because because like my intuition fails completely how complexity can emerge from such simplicity. Like my intuition fails, I think, is the biggest problem I have. Do you find systems like that beautiful? Is there? Do you do you think about cellular automata? Because cellular automata don't really have, and many other artificial life systems don't necessarily have an objective. Maybe maybe that's a wrong way to say it. It's almost like it's just evolving and creating.

• 1:48:11Speaker 01

  Yeah.

• 1:48:12Lex Fridman?

  And there's not even a good definition of what it means to create something complex and interesting and surprising, all those words that you said. Is there some of those systems that you find beautiful?

• 1:48:24Speaker 01

  Yeah, yeah. And similarly, evolution does not have a goal. It is responding to a current situation and survival then creates more complexity and therefore we have something that we perceive as progress, but that's not what evolution is inherently said to do. And yeah, that's that's really fascinating, how a simple set of rules or simple mappings can from how from such simple mappings, complexity can emerge. So it's a question of emergence and self-organization. And the game of life is one of the simplest ones and very visual and therefore.

• 1:49:07Speaker 01

  It drives home the point that it's possible that nonlinear interactions and and this kinds of complexity can emerge emerge from them. And biology and evolution is along the same lines. We have simple representations. DNA, if you really think of it, it's not that complex. It's a long sequence of them. There's lots of them, but it's a very simple representation. And similar with evolutionary computation, whatever string or tree representation we have and the operations, you know, the amount of code that's required to manipulate those is really, really little. And of course, Game of Life even less. So how complexity emerges from such simple principles? That's that's absolutely fascinating.

• 1:49:52Speaker 01

  The challenge is to be able to control it and guide it and direct it so that it becomes useful. And like Game of Life is fascinating to look at and and evolution, all the forms that come out is fascinating, but can we actually make it useful for us?

• 1:50:07Lex Fridman?

  And efficient, because if you actually think about each of the cells in the game of life as a living organism, there's a lot of death that has to happen to create anything interesting. And so I guess the question is for us humans that are mortal and their life ends quickly. We want to kind of hurry up and make sure we make sure we take evolution the trajectory that is a little bit more efficient than the alternatives.

• 1:50:33Speaker 01

  And that touches upon something we talked about earlier, that evolutionary computation is very impatient. We have a goal, we want it right away, whereas biology has a lot of time and and deep, deep time and and weak pressure and large populations. One great example of of this is the novelty search. So evolutionary computation, where you don't actually specify a fitness goal, something that is your actual thing that you want, but you just reward solutions that are different from what you've seen before. Nothing else. And you know what? You actually discover things that are interesting and useful that way. Um, Keston and Joe Lemon did this one study where they actually um tried to evolve walking behavior on robots.

• 1:51:18Speaker 01

  And that's actually, we talked about earlier, where your robot actually failed in all kinds of ways and eventually discovered something that was uh very efficient walk. Uh, and and it was because they they rewarded things that were different that you were able to discover something. Uh, and I think that this is crucial, um, because in order to be really different from what you already have, you have to utilize what is there in a domain to create something really different. So you have encoded the uh fundamentals of your world and then you make changes to those fundamentals you get further away. So that's probably what's happening in these systems of emergence, uh, that uh fundamentals

• 1:52:00Speaker 01

  are there and when you follow those fundamentals, you get into points and some of those are actually interesting and useful. Now, even in that robotic walker simulation, there was a large set of garbage, uh, but among them, there were some of these, you know, gems. Uh, and then those are the ones that somehow you have to outside recognize and and make useful. Uh, but this kind of productive systems, if you code them the right kind of principles, um, I think that they uh that encode the structure of the of the domain, then you will get to these solutions and the discoveries.

• 1:52:33Lex Fridman?

  It feels like that might also be a good way to live life. So let me ask, do you have advice for young people today about how to live life or how to succeed in their career? Or forget career, just succeed in life? From an evolutionary computation perspective?

• 1:52:52Speaker 01

  Yes. Yes, definitely. Explore.

• 1:52:57Speaker 01

  Diversity, exploration. And and in individuals, take classes in music, history, philosophy, you know, math, engineering, uh, see connections between them, uh, travel, you know, learn a language. I mean, all this diversity is fascinating and we have it at our fingertips fingertips today. It's possible. You have to make a bit of an effort because it's not easy, but the rewards are wonderful.

• 1:53:26Lex Fridman?

  Yeah, there's something interesting about an objective function of new experiences. So try to figure out, I mean.

• 1:53:34Lex Fridman?

  What is the maximally new experience I could have today? And that sort of that novelty, optimizing for novelty for some period of time might be a very interesting way to sort of maximally expand the sets of experiences you had and then ground from that perspective, like what you, what will be the most fulfilling trajectory through life. Of course, the flip side of that is where I come from. Again, maybe Russian, I don't know. But the, the choice has a, a choice is a, has a detrimental effect, I think, from, at least from my mind, where scarcity is in, has an empowering

• 1:54:18Lex Fridman?

  effect. So if I sort of, if I have very little of something and only one of that something, I will appreciate it deeply. Until I came to Texas recently and I've been picking out on delicious, incredible meat. I've been fasting a lot, so I need to do that again. But when you fast for a few days, that the first taste of a food is, is incredible. So the downside of exploration is that somehow, maybe you can correct me, but somehow you don't get to experience deeply any one of the particular moments. But that could be

• 1:55:01Lex Fridman?

  a psychology thing. That could be just a very human peculiar flaw.

• 1:55:07Speaker 01

  Yeah, I didn't mean that you superficially explore. I mean, you can

• 1:55:11Lex Fridman?

  Explore deeply.

• 1:55:12Speaker 01

  Yeah. So you don't have to explore hundred things, but maybe a few topics where you can take a deep enough time dive that you gain an understanding. You yourself have to decide at some point that this is deep enough and I, I, I've obtained what I can from this topic and now it's time to move on. And that might take years. People sometimes switch careers and they may stay on some career for decade and switch to another one. You can do it. You're not pretty determined to stay where you are. But, you know, in order to achieve Um, something, you know, 10,000 hours makes, you need 10,000 hours to become an expert on something.

• 1:55:56Speaker 01

  Uh, so you don't have to become an expert, but even develop an understanding and gain the experience that you can use later. You probably have to spend, like I said, it's not easy. You gotta spend some effort on it. Um, now also at some point then, when you have this diversity and you have these experiences, exploration, you may want to, um, you may find something that you can't stay away from.

• 1:56:18Lex Fridman?

  Mm hmm.

• 1:56:19Speaker 01

  Uh, like for, as it was computers, it was AI, it was, you know, that you, I just have to do it, you know, and I, you know, and then well, it will take decades maybe, and you are pursuing it because you figured out that this is really exciting and you can bring in your experiences. And there's nothing wrong with that either, but you asked, what's the advice for young people? Now, that's the exploration part. And then beyond that, after that exploration, you actually can focus and, and, and build a career. And, you know, even there you can switch multiple times, but, uh, but I think the diversity exploration is fundamental to having a successful career, as is concentration and spending an effort where it matters and, and, but you are in better position to make the choice when you have done your homework.

• 1:57:04Lex Fridman?

  Explained. So exploration precedes commitment, but both are beautiful.

• 1:57:08Speaker 01

  Yeah.

• 1:57:09Lex Fridman?

  Uh, so again, from an evolutionary computation perspective, we'll, we'll look at all the agents that had to die in order to come up with different solutions in simulation. Uh, what do you think from that individual agent's perspective is the meaning of it all? So for us humans, you're just one agent who's going to be dead, unfortunately, one day too soon. What do you think is the why of why that agent came to be? And eventually will be no more? Is there a meaning to it all?

• 1:57:43Speaker 01

  Yeah. In evolution, there is meaning. Everything is a potential direction. Everything is a potential stepping stone. Um, not all of them are going to work out. Some of them are foundations for further, um, improvement. And even those that are perhaps going to die out, uh, were potential energies, potential solutions.

• 1:58:09Speaker 01

  In biology, we see a lot of species die off naturally. And, you know, like the dinosaurs, I mean, they have a really good solution for a while. But then it didn't turn out to be not such a good solution in the long term. When there's an environmental change, you have to have diversity. Some other solutions become better. Doesn't mean that that that was an attempt. It didn't quite work out or last. But there are still dinosaurs around us, at least their relatives. And they may one day again be useful. Who knows? So from an individual's perspective, you got to think of a bigger picture. That it is a huge engine that is innovative. And these elements are all part of it, potential innovations on their own.

• 1:58:53Speaker 01

  And also as as raw material perhaps, or stepping stones for other things that could come after.

• 1:58:59Lex Fridman?

  But it still feels from an individual perspective that I I matter a lot. But even if I'm just a little cog in the giant machine. Is that just a silly human notion in an individualistic society? No, should let go of that. Do you do you find beauty in being part of the giant machine?

• 1:59:20Speaker 01

  Yeah, I think it's meaningful. I think it adds purpose to your life, that you are part of something bigger.

• 1:59:29Lex Fridman?

  That said, are you do you ponder your individual agent's mortality? Do you do you think about death? Do you fear death?

• 1:59:40Speaker 01

  Well, certainly more now than when I was a youngster and did skydiving and paragliding. And, you know, all these things.

• 1:59:49Lex Fridman?

  You've become wiser.

• 1:59:52Speaker 01

  There is a reason for this life arc that younger folks are more fearless in many ways. It's part of the exploration. You know, they are the they are the individuals who think, hmm, I wonder what's over those mountains? Or what if I go really far in that ocean? What would I find? I mean, older folks don't necessarily think that way, but younger do. And it's kind of counterintuitive. So, yeah, it's this. But logically, it's like, you know, you have a limited amount of time. What can you do with it that matters? So you try to, you have done your exploration, you committed to certain direction and you've become an expert perhaps in it.

• 2:00:34Speaker 01

  What can I do that matters with the limited resources that I have? That's what, how you, I think, a lot of people, myself included, start thinking later on in their career.

• 2:00:46Lex Fridman?

  And like you said, leave a bit of a trace and a bit of an impact, even though after the agent is gone.

• 2:00:52Speaker 01

  Yeah, that's the goal.

• 2:00:54Lex Fridman?

  Well, this was a fascinating conversation. I don't think there's a better way to end it. Thank you so much. So first of all, I'm very inspired of how vibrant the community at UT Austin in Austin is. It's really exciting for me to see it. And this whole field seems like profound philosophically, but also the path forward for the artificial intelligence community. So thank you so much for explaining so many cool things to me today and for wasting all of your valuable time with me.

• 2:01:22Speaker 01

  Oh, it was a pleasure. Thanks.

• 2:01:24Lex Fridman?

  I appreciate it. Thanks for listening to this conversation with Risto McLeinen, and thank you to The Jordan Harbinger Show, Grammarly, Belcampo, and Indeed. Check them out in the description to support this podcast. And now let me leave you with some words from Carl Sagan. "Extinction is the rule. Survival is the exception." Thank you for listening. I hope to see you next time.