Risto Miikkulainen: Neuroevolution and Evolutionary Computation

00:00:00.000 | The following is a conversation with Risto Michelinan,

00:00:02.880 | a computer scientist at University of Texas at Austin

00:00:06.000 | and Associate Vice President

00:00:07.880 | of Evolutionary Artificial Intelligence at Cognizant.

00:00:11.480 | He specializes in evolutionary computation,

00:00:14.440 | but also many other topics in artificial intelligence,

00:00:17.640 | cognitive science, and neuroscience.

00:00:19.920 | Quick mention of our sponsors,

00:00:21.920 | Jordan Harbinger Show, Grammarly, Belcampo, and Indeed.

00:00:26.600 | Check them out in the description to support this podcast.

00:00:30.600 | As a side note, let me say that nature-inspired algorithms

00:00:34.160 | from ant colony optimization to genetic algorithms

00:00:36.840 | to cellular automata to neural networks

00:00:39.600 | have always captivated my imagination,

00:00:41.920 | not only for their surprising power

00:00:43.960 | in the face of long odds,

00:00:45.600 | but because they always opened up doors

00:00:47.800 | to new ways of thinking about computation.

00:00:50.720 | It does seem that in the long arc of computing history,

00:00:54.200 | running toward biology, not running away from it,

00:00:57.600 | is what leads to long-term progress.

00:01:00.440 | This is the Lex Friedman Podcast,

00:01:03.240 | and here is my conversation with Risto Michelinan.

00:01:06.800 | If we ran the earth experiment,

00:01:10.240 | this fun little experiment we're on,

00:01:12.560 | over and over and over and over a million times

00:01:15.280 | and watch the evolution of life as it pans out,

00:01:19.240 | how much variation in the outcomes of that evolution

00:01:22.000 | do you think we would see?

00:01:23.240 | Now, we should say that you are a computer scientist.

00:01:27.440 | - That's actually not such a bad question

00:01:29.440 | for computer scientists,

00:01:30.440 | because we are building simulations of these things,

00:01:34.080 | and we are simulating evolution,

00:01:36.240 | and that's a difficult question to answer in biology,

00:01:38.520 | but we can build a computational model

00:01:40.760 | and run it a million times

00:01:42.200 | and actually answer that question,

00:01:43.600 | how much variation do we see when we simulate it?

00:01:47.040 | And that's a little bit beyond what we can do today,

00:01:50.640 | but I think that we will see some regularities,

00:01:54.160 | and it took evolution also a really long time

00:01:56.560 | to get started,

00:01:57.760 | and then things accelerated really fast towards the end.

00:02:02.240 | But there are things that need to be discovered,

00:02:04.280 | and they probably will be over and over again,

00:02:06.480 | like manipulation of objects, opposable thumbs,

00:02:11.200 | and also some way to communicate,

00:02:16.040 | maybe orally, like, why will you have speech?

00:02:18.280 | It might be some other kind of sounds.

00:02:20.800 | And decision-making, but also vision.

00:02:24.080 | Eye has evolved many times,

00:02:26.200 | various vision systems have evolved.

00:02:28.160 | So we would see those kinds of solutions,

00:02:30.760 | I believe, emerge over and over again.

00:02:32.880 | They may look a little different,

00:02:34.240 | but they get the job done.

00:02:36.280 | The really interesting question is,

00:02:37.520 | would we have primates?

00:02:39.000 | Would we have humans,

00:02:40.800 | or something that resembles humans?

00:02:43.280 | And would that be an apex of evolution after a while?

00:02:47.000 | We don't know where we're going from here,

00:02:48.440 | but we certainly see a lot of tool use

00:02:51.000 | and building, constructing our environment.

00:02:54.000 | So I think that we will get that.

00:02:56.320 | We get some evolution producing some agents

00:02:59.680 | that can do that, manipulate the environment and build.

00:03:02.480 | - What do you think is special about humans?

00:03:04.080 | Like, if you were running the simulation

00:03:06.040 | and you observe humans emerge,

00:03:08.680 | like these tool makers, they start a fire,

00:03:10.440 | and all that stuff, start running around,

00:03:11.720 | building buildings, and then running for president,

00:03:13.560 | and all those kinds of things.

00:03:15.560 | What would be, how would you detect that?

00:03:19.160 | 'Cause you're like really busy

00:03:20.360 | as the creator of this evolutionary system,

00:03:23.160 | so you don't have much time to observe,

00:03:25.680 | like detect if any cool stuff came up, right?

00:03:28.920 | How would you detect humans?

00:03:31.240 | - Well, you are running the simulation,

00:03:33.280 | so you also put in visualization

00:03:37.480 | and measurement techniques there.

00:03:39.680 | So if you are looking for certain things,

00:03:43.240 | like communication, you'll have detectors

00:03:46.400 | to find out whether that's happening,

00:03:48.000 | even if it's a lot simulation.

00:03:50.120 | And I think that that's what we would do.

00:03:53.520 | We know roughly what we want,

00:03:56.360 | intelligent agents that communicate, cooperate, manipulate,

00:04:01.200 | and we would build detections

00:04:03.160 | and visualizations of those processes.

00:04:05.560 | Yeah, and there's a lot of,

00:04:08.040 | we'd have to run it many times,

00:04:09.520 | and we have plenty of time to figure out

00:04:11.920 | how we detect the interesting things.

00:04:13.520 | But also, I think we do have to run it many times

00:04:16.680 | because we don't quite know what shape those will take,

00:04:21.160 | and our detectors may not be perfect for them

00:04:23.880 | to begin with.

00:04:24.720 | - Well, that seems really difficult

00:04:25.760 | to build a detector of intelligent

00:04:28.680 | or intelligent communication.

00:04:32.800 | If we take an alien perspective, observing Earth,

00:04:37.160 | are you sure that they would be able to detect humans

00:04:40.280 | as the special thing?

00:04:41.360 | Wouldn't they be already curious about other things?

00:04:43.800 | There's way more insects by body mass, I think,

00:04:47.080 | than humans by far, and colonies.

00:04:50.880 | Obviously, dolphins is the most intelligent creature

00:04:53.880 | on Earth, we all know this.

00:04:55.240 | So, it could be the dolphins that they detect.

00:04:58.400 | It could be the rockets that we seem to be launching.

00:05:00.840 | That could be the intelligent creature they detect.

00:05:03.800 | It could be some other trees.

00:05:06.680 | Trees have been here a long time.

00:05:07.960 | I just learned that sharks have been here

00:05:10.560 | 400 million years, and that's longer

00:05:13.240 | than trees have been here.

00:05:15.000 | So, maybe it's the sharks, they go by age.

00:05:17.400 | Like, there's a persistent thing.

00:05:18.960 | Like, if you survive long enough,

00:05:20.800 | especially through the mass extinctions,

00:05:22.320 | that could be the thing your detector is detecting.

00:05:25.360 | Humans have been here a very short time,

00:05:27.840 | and we're just creating a lot of pollution,

00:05:30.640 | but so is the other creatures.

00:05:31.920 | So, I don't know.

00:05:33.320 | Do you think you would be able to detect humans?

00:05:35.680 | Like, how would you go about detecting,

00:05:37.680 | in the computational sense,

00:05:39.120 | maybe we can leave humans behind,

00:05:40.920 | in the computational sense, detect interesting things?

00:05:44.600 | Do you basically have to have a strict objective function

00:05:48.760 | by which you measure the performance of a system,

00:05:51.800 | or can you find curiosities and interesting things?

00:05:55.400 | - Yeah, well, I think the first measurement

00:05:59.480 | would be to detect how much of an effect

00:06:02.240 | you can have in your environment.

00:06:03.600 | So, if you look around, we have cities,

00:06:06.880 | and that is constructed environments,

00:06:08.800 | and that's where a lot of people live, most people live.

00:06:11.960 | So, that would be a good sign of intelligence,

00:06:15.120 | that you don't just live in an environment,

00:06:17.920 | but you construct it to your liking.

00:06:20.080 | And that's something pretty unique.

00:06:21.880 | I mean, certainly birds build nests and all,

00:06:24.240 | but they don't build quite cities.

00:06:25.520 | Termites build mounds and hives and things like that,

00:06:29.080 | but the complexity of the human construction cities,

00:06:32.120 | I think, would stand out, even to an external observer.

00:06:34.960 | - Of course, that's what a human would say.

00:06:36.920 | (Luke laughs)

00:06:38.280 | - Yeah, and you can certainly say that sharks

00:06:41.000 | are really smart because they've been around so long,

00:06:43.280 | and they haven't destroyed their environment,

00:06:45.080 | which humans are about to do,

00:06:46.640 | which is not a very smart thing.

00:06:48.920 | But we'll get over it, I believe.

00:06:52.080 | And we can get over it by doing some construction

00:06:55.320 | that actually is benign, and maybe even enhances

00:06:58.160 | resilience of nature.

00:07:02.520 | - So, you mentioned the simulation

00:07:04.480 | that we run over and over might start,

00:07:07.160 | it's a slow start.

00:07:09.000 | So, do you think how unlikely, first of all,

00:07:12.640 | I don't know if you think about this kind of stuff,

00:07:14.240 | but how unlikely is step number zero,

00:07:18.240 | which is the springing up, like the origin of life on Earth?

00:07:23.040 | And second, how unlikely is anything interesting

00:07:28.040 | happening beyond that?

00:07:30.560 | Sort of like the start that creates all the rich complexity

00:07:35.560 | that we see on Earth today?

00:07:36.760 | - Yeah, there are people who are working

00:07:38.640 | on exactly that problem from primordial soup,

00:07:42.320 | how do you actually get self-replicating molecules?

00:07:45.880 | And they are very close.

00:07:48.800 | With a little bit of help, you can make that happen.

00:07:51.960 | So, of course, we know what we want,

00:07:55.740 | so they can set up the conditions

00:07:57.200 | and try out conditions that are conducive to that.

00:07:59.840 | For evolution to discover that, it took a long time.

00:08:04.160 | For us to recreate it probably won't take that long.

00:08:07.720 | And the next steps from there,

00:08:09.880 | I think also with some hand-holding,

00:08:12.920 | I think we can make that happen.

00:08:14.520 | But with evolution, what was really fascinating

00:08:18.600 | was eventually the runaway evolution of the brain

00:08:22.680 | that created humans and created,

00:08:24.480 | well, also other higher animals.

00:08:27.280 | That was something that happened really fast.

00:08:29.760 | And that's a big question.

00:08:32.440 | Is that something replicable?

00:08:33.760 | Is that something that can happen?

00:08:35.800 | And if it happens, does it go in the same direction?

00:08:39.200 | That is a big question to ask.

00:08:40.760 | Even in computational terms,

00:08:43.000 | I think that it's relatively possible to come up here,

00:08:47.360 | create an experiment where we look at the primordial soup

00:08:49.840 | and the first couple of steps

00:08:51.280 | of multicellular organisms even.

00:08:53.480 | But to get something as complex as the brain,

00:08:55.760 | we don't quite know the conditions for that

00:08:59.680 | and how to even get started

00:09:01.420 | and whether we can get this kind of

00:09:02.480 | runaway evolution happening.

00:09:03.960 | - From a detector perspective,

00:09:09.120 | if we're observing this evolution,

00:09:10.800 | what do you think is the brain?

00:09:12.360 | What do you think is the, let's say, what is intelligence?

00:09:15.960 | So in terms of the thing that makes humans special,

00:09:18.360 | we seem to be able to reason,

00:09:20.080 | we seem to be able to communicate,

00:09:23.520 | but the core of that is this something

00:09:26.000 | in the broad category we might call intelligence.

00:09:29.600 | So if you put your computer scientist hat on,

00:09:33.000 | is there favorite ways you like to think about

00:09:37.560 | that question of what is intelligence?

00:09:39.780 | - Well, my goal is to create agents that are intelligent.

00:09:46.320 | - Not to define what.

00:09:49.520 | (laughing)

00:09:50.600 | - And that is a way of defining it.

00:09:52.720 | And that means that it's some kind of an object

00:09:57.720 | or a program that has limited sensory

00:10:02.720 | and effective capabilities interacting with the world,

00:10:08.000 | and then also a mechanism for making decisions.

00:10:11.720 | So with limited abilities like that, can it survive?

00:10:15.860 | Survival is the simplest goal,

00:10:18.800 | but you could also give it other goals.

00:10:20.500 | Can it multiply?

00:10:21.400 | Can it solve problems that you give it?

00:10:24.440 | And that is quite a bit less than human intelligence.

00:10:27.220 | There are, animals would be intelligent, of course,

00:10:29.720 | with that definition.

00:10:31.080 | And you might have even some other forms of life.

00:10:34.360 | Even, so what, so intelligence in that sense

00:10:37.840 | is a survival skill given resources that you have

00:10:42.840 | and using your resources so that you will stay around.

00:10:46.080 | - Do you think death, mortality is fundamental to an agent?

00:10:52.880 | So like there's, I don't know if you're familiar,

00:10:55.040 | there's a philosopher named Ernest Becker

00:10:56.880 | who wrote "The Denial of Death" and his whole idea.

00:11:01.220 | And there's folks, psychologists, cognitive scientists

00:11:04.020 | that work on terror management theory.

00:11:06.600 | And they think that one of the special things about humans

00:11:10.020 | is that we're able to sort of foresee our death, right?

00:11:13.940 | We can realize not just as animals do,

00:11:16.620 | sort of constantly fear in an instinctual sense,

00:11:19.420 | respond to all the dangers that are out there,

00:11:21.600 | but like understand that this ride ends eventually.

00:11:25.200 | And that in itself is the most,

00:11:28.520 | is the force behind all of the creative efforts

00:11:31.160 | of human nature.

00:11:32.600 | That's the philosophy.

00:11:33.680 | - I think that makes sense, a lot of sense.

00:11:35.280 | I mean, animals probably don't think of death the same way,

00:11:38.680 | but humans know that your time is limited

00:11:40.680 | and you wanna make it count.

00:11:42.080 | And you can make it count in many different ways,

00:11:45.020 | but I think that has a lot to do with creativity

00:11:47.780 | and the need for humans to do something

00:11:50.100 | beyond just surviving.

00:11:52.680 | And now going from that simple definition

00:11:55.560 | to something that's the next level,

00:11:57.400 | I think that that could be the second decision,

00:12:00.200 | the second level of definition

00:12:01.520 | that intelligence means something

00:12:04.320 | and you do something that stays behind you

00:12:06.240 | that's more than your existence.

00:12:10.080 | Something you create something that is useful for others,

00:12:13.280 | is useful in the future, not just for yourself.

00:12:16.200 | And I think that's a nice definition of intelligence

00:12:18.880 | in a next level.

00:12:20.840 | And it's also nice 'cause it doesn't require

00:12:23.480 | that they are humans or biological.

00:12:25.240 | They could be artificial agents that are intelligence.

00:12:28.240 | They could achieve those kind of goals.

00:12:30.360 | - So particular agent, the ripple effects of their existence

00:12:35.360 | on the entirety of the system is significant.

00:12:38.560 | So like they leave a trace where there's like a,

00:12:41.800 | yeah, like ripple effects.

00:12:43.920 | But see, then you go back to the butterfly

00:12:46.080 | with the flap of a wing,

00:12:47.400 | and then you can trace a lot of like nuclear wars

00:12:50.840 | and all the conflicts of human history

00:12:52.720 | somehow connected to that one butterfly

00:12:54.560 | that created all the chaos.

00:12:56.280 | So maybe that's not,

00:12:57.960 | maybe that's a very poetic way to think.

00:13:02.160 | That's something we humans in a human centric way

00:13:04.520 | wanna hope we have this impact.

00:13:09.080 | Like that is the secondary effect of our intelligence.

00:13:12.200 | We've had that long lasting impact on the world,

00:13:14.600 | but maybe the entirety of physics in the universe

00:13:19.600 | has a very long lasting effect.

00:13:22.720 | - Sure, but you can also think of it,

00:13:25.640 | what if like the wonderful life,

00:13:28.200 | what if you're not here?

00:13:30.000 | Will somebody else do this?

00:13:31.640 | Is it something that you actually contributed

00:13:34.600 | because you had something unique to contribute?

00:13:37.480 | That's a pretty high bar though.

00:13:39.480 | - Uniqueness.

00:13:40.720 | Yeah, so you have to be Mozart or something

00:13:44.600 | to actually reach that level.

00:13:46.360 | Nobody would have developed that,

00:13:48.080 | but other people might have solved this equation

00:13:50.840 | if you didn't do it.

00:13:53.240 | But also within limited scope.

00:13:56.240 | I mean, during your lifetime or next year,

00:14:00.440 | you could contribute something that unique

00:14:02.840 | that other people did not see.

00:14:04.560 | And then that could change the way things move forward

00:14:09.800 | for a while.

00:14:11.400 | So I don't think we have to be Mozart

00:14:14.080 | to be called intelligence,

00:14:15.400 | but we have this local effect that is changing.

00:14:18.160 | If you weren't there, that would not have happened.

00:14:20.200 | And it's a positive effect, of course,

00:14:21.560 | you want it to be a positive effect.

00:14:23.280 | - Do you think it's possible to engineer in

00:14:26.080 | to computational agents, a fear of mortality?

00:14:29.800 | Like, does that make any sense?

00:14:35.640 | So there's a very trivial thing where it's like,

00:14:38.280 | you could just code in a parameter,

00:14:39.720 | which is how long the life ends,

00:14:41.360 | but more of a fear of mortality,

00:14:45.480 | like awareness of the way that things end

00:14:48.960 | and somehow encoding a complex representation of that fear,

00:14:53.960 | which is like, maybe as it gets closer,

00:14:57.000 | you become more terrified.

00:14:58.880 | I mean, there seems to be something really profound

00:15:01.640 | about this fear that's not currently encodable

00:15:04.880 | in a trivial way into our programs.

00:15:08.240 | - Well, I think you're referring to the emotion of fear,

00:15:11.880 | something, 'cause we have cognitively,

00:15:13.600 | we know that we have limited lifespan

00:15:16.360 | and most of us cope with it by just,

00:15:18.080 | hey, that's what the world is like,

00:15:19.680 | and I make the most of it.

00:15:20.600 | But sometimes you can have like a fear that's not healthy,

00:15:25.600 | that paralyzes you, that you can't do anything.

00:15:29.320 | And somewhere in between there,

00:15:32.000 | not caring at all and getting paralyzed because of fear

00:15:36.160 | is a normal response,

00:15:37.280 | which is a little bit more than just logic and it's emotion.

00:15:41.440 | So now the question is what good are emotions?

00:15:43.680 | I mean, they are quite complex

00:15:46.160 | and there are multiple dimensions of emotions

00:15:48.480 | and they probably do serve as a survival function,

00:15:52.560 | heightened focus, for instance.

00:15:55.840 | And fear of death might be a really good emotion

00:15:59.680 | when you are in danger, that you recognize it.

00:16:02.640 | Even if it's not logically necessarily easy to derive

00:16:06.360 | and you don't have time for that logical deduction,

00:16:10.400 | you may be able to recognize the situation is dangerous

00:16:12.720 | and this fear kicks in and you all of a sudden perceive

00:16:16.240 | the facts that are important for that.

00:16:18.440 | And I think that's generally is the role of emotions.

00:16:20.640 | It allows you to focus what's relevant for your situation.

00:16:24.520 | And maybe if fear of death plays the same kind of role,

00:16:27.800 | but if it consumes you and it's something that you think

00:16:30.600 | in normal life when you don't have to,

00:16:32.080 | then it's not healthy and then it's not productive.

00:16:34.440 | - Yeah, but it's fascinating to think

00:16:36.640 | how to incorporate emotion into a computational agent.

00:16:41.640 | It almost seems like a silly statement to make,

00:16:44.280 | but it perhaps seems silly

00:16:47.720 | because we have such a poor understanding

00:16:49.800 | of the mechanism of emotion, of fear, of...

00:16:53.560 | I think at the core of it is another word

00:16:58.520 | that we know nothing about, but say a lot,

00:17:01.160 | which is consciousness.

00:17:02.440 | Do you ever in your work or like maybe on a coffee break,

00:17:08.600 | think about what the heck is this thing consciousness

00:17:11.640 | and is it at all useful in our thinking about AI systems?

00:17:15.000 | - Yes, it is an important question.

00:17:17.400 | You can build representations and functions,

00:17:23.160 | I think into these agents that act like emotions

00:17:26.760 | and consciousness perhaps.

00:17:28.680 | So I mentioned emotions being something

00:17:31.960 | that allow you to focus and pay attention,

00:17:34.240 | filter out what's important.

00:17:35.400 | Yeah, you can have that kind of a filter mechanism

00:17:38.320 | and it puts you in a different state.

00:17:40.360 | Your computation is in a different state.

00:17:42.120 | Certain things don't really get through

00:17:43.600 | and others are heightened.

00:17:45.120 | Now you label that box emotion.

00:17:48.480 | I don't know if that means it's an emotion,

00:17:49.880 | but it acts very much like we understand what emotions are.

00:17:54.280 | And we actually did some work like that,

00:17:56.920 | modeling hyenas who were trying to steal a kill from lions,

00:18:01.920 | which happens in Africa.

00:18:03.520 | I mean, hyenas are quite intelligent,

00:18:05.960 | but not really intelligent.

00:18:08.280 | And they have this behavior that's more complex

00:18:13.160 | than anything else they do.

00:18:14.040 | They can band together if there's about 30 of them or so,

00:18:17.680 | they can coordinate their effort

00:18:20.040 | so that they push the lions away from a kill.

00:18:22.520 | Even though the lions are so strong

00:18:24.040 | that they could kill a hyena by striking with a paw.

00:18:28.440 | But when they work together and precisely time this attack,

00:18:31.640 | the lions will leave and they get the kill.

00:18:34.080 | And probably there are some states like emotions

00:18:39.080 | that the hyenas go through.

00:18:40.840 | The first day they call for reinforcements.

00:18:43.640 | They really want that kill, but there's not enough of them.

00:18:45.640 | So they vocalize and there's more people,

00:18:48.480 | more hyenas that come around.

00:18:50.680 | And then they have two emotions.

00:18:52.280 | They're very afraid of the lion.

00:18:54.360 | So they want to stay away,

00:18:55.600 | but they also have a strong affiliation between each other.

00:18:59.840 | And then this is the balance of the two emotions.

00:19:02.160 | And also, yes, they also want the kill.

00:19:04.880 | So it's both repelled and attractive.

00:19:07.360 | But then this affiliation eventually is so strong

00:19:10.600 | that when they move, they move together,

00:19:12.280 | they act as a unit and they can perform that function.

00:19:15.400 | So there's an interesting behavior

00:19:18.440 | that seems to depend on these emotions strongly

00:19:21.400 | that makes it possible, coordinate actions.

00:19:24.280 | - And I think a critical aspect of that,

00:19:28.880 | the way you're describing is emotion there

00:19:30.600 | is a mechanism of social communication,

00:19:34.320 | of a social interaction.

00:19:35.960 | Maybe humans won't even be that intelligent

00:19:40.520 | or most things we think of as intelligent

00:19:42.440 | wouldn't be that intelligent

00:19:43.500 | without the social component of interaction.

00:19:47.040 | Maybe much of our intelligence is essentially

00:19:50.240 | in an outgrowth of social interaction.

00:19:52.880 | And maybe for the creation of intelligent agents,

00:19:55.720 | we have to be creating fundamentally social systems.

00:19:59.000 | - Yes, I strongly believe that's true.

00:20:01.160 | And yes, the communication is multifaceted.

00:20:05.520 | I mean, they vocalize and call for friends,

00:20:08.120 | but they also rub against each other and they push

00:20:11.200 | and they do all kinds of gestures and so on.

00:20:14.280 | So they don't act alone.

00:20:15.760 | And I don't think people act alone very much either,

00:20:18.400 | at least normal most of the time.

00:20:21.120 | And social systems are so strong for humans

00:20:25.040 | that I think we build everything

00:20:26.800 | on top of these kinds of structures.

00:20:28.320 | And one interesting theory around that,

00:20:30.880 | because this theory, for instance, for language,

00:20:32.520 | but language origins is that, where did language come from?

00:20:36.200 | And it's a plausible theory that first came social systems

00:20:41.200 | that you have different roles in a society.

00:20:44.180 | And then those roles are exchangeable,

00:20:47.380 | that I scratch your back, you scratch my back,

00:20:49.960 | we can exchange roles.

00:20:51.520 | And once you have the brain structures

00:20:53.480 | that allow you to understand actions in terms of roles

00:20:55.760 | that can be changed, that's the basis for language,

00:20:59.080 | for grammar.

00:20:59.960 | And now you can start using symbols to refer

00:21:02.520 | to objects in the world,

00:21:04.800 | and you have this flexible structure.

00:21:06.720 | So there's a social structure that's fundamental

00:21:10.960 | for language to develop.

00:21:12.420 | Now, again, then you have language,

00:21:13.880 | you can refer to things that are not here right now,

00:21:17.160 | and that allows you to then build all the good stuff

00:21:20.880 | about planning, for instance,

00:21:22.580 | and building things and so on.

00:21:24.600 | So yeah, I think that very strongly humans are social,

00:21:28.240 | and that gives us ability to structure the world.

00:21:32.960 | But also as a society, we can do so much more,

00:21:35.480 | 'cause one person does not have to do everything.

00:21:37.960 | You can have different roles

00:21:39.760 | and together achieve a lot more.

00:21:41.680 | And that's also something we see

00:21:43.140 | in computational simulations today.

00:21:44.800 | I mean, we have multi-agent systems

00:21:46.320 | that can perform tasks.

00:21:47.760 | This fascinating demonstration,

00:21:50.000 | Marco Dorigo, I think it was,

00:21:51.880 | these robots, little robots that had to navigate

00:21:54.000 | through an environment,

00:21:54.840 | and there were things that are dangerous,

00:21:57.660 | like maybe a big chasm or some kind of groove,

00:22:01.160 | a hole, and they could not get across it.

00:22:03.600 | But if they grab each other with their gripper,

00:22:06.480 | they formed a robot that was much longer,

00:22:09.000 | like a team, and this way they could get across that.

00:22:12.360 | So this is a great example

00:22:14.240 | of how together we can achieve things

00:22:16.680 | we couldn't otherwise, like the hyenas.

00:22:18.920 | Alone they couldn't, but as a team they could.

00:22:21.440 | And I think humans do that all the time.

00:22:23.200 | We're really good at that.

00:22:24.840 | - Yeah, and the way you described the system of hyenas,

00:22:28.000 | it almost sounds algorithmic.

00:22:29.720 | Like the problem with humans is they're so complex,

00:22:32.840 | it's hard to think of them as algorithms.

00:22:35.080 | But with hyenas, it's simple enough

00:22:39.040 | to where it feels like,

00:22:41.860 | at least hopeful, that it's possible

00:22:43.260 | to create computational systems that mimic that.

00:22:48.260 | - Yeah, that's exactly why we looked at that.

00:22:52.020 | - As opposed to humans.

00:22:53.180 | - Like I said, they are intelligent,

00:22:55.260 | but they are not quite as intelligent as, say, baboons,

00:22:59.420 | which would learn a lot and would be much more flexible.

00:23:02.140 | The hyenas are relatively rigid in what they can do.

00:23:05.700 | And therefore, you could look at this behavior,

00:23:08.100 | like this is a breakthrough in evolution about to happen,

00:23:11.540 | that they've discovered something

00:23:12.860 | about social structures, communication, about cooperation,

00:23:17.560 | and it might then spill over to other things too

00:23:20.580 | in thousands of years in the future.

00:23:22.660 | - Yeah, I think the problem with baboons and humans

00:23:24.960 | is probably too much is going on inside the head,

00:23:27.860 | where we won't be able to measure it

00:23:29.180 | if we're observing the system.

00:23:30.340 | With hyenas, it's probably easier to observe

00:23:34.260 | the actual decision-making

00:23:35.500 | and the various motivations that are involved.

00:23:38.660 | - Yeah, they are visible.

00:23:40.060 | - And we can even quantify possibly their emotional state

00:23:45.060 | because they leave droppings behind.

00:23:48.180 | And there are chemicals there

00:23:49.580 | that can be associated with neurotransmitters.

00:23:52.980 | And we can separate what emotions they might have

00:23:55.660 | experienced in the last 24 hours.

00:23:58.980 | - What to you is the most beautiful, speaking of hyenas,

00:24:04.240 | what to you is the most beautiful nature-inspired algorithm

00:24:08.740 | in your work that you've come across?

00:24:10.460 | Something maybe earlier on in your work or maybe today?

00:24:13.780 | - I think evolutionary computation

00:24:17.140 | is the most amazing method.

00:24:19.860 | So what fascinates me most is that, with computers,

00:24:24.380 | is that you can get more out than you put in.

00:24:27.660 | I mean, you can write a piece of code

00:24:29.900 | and your machine does what you told it.

00:24:32.540 | I mean, this happened to me in my freshman year.

00:24:35.380 | It did something very simple and I was just amazed.

00:24:37.780 | I was blown away that it would get the number

00:24:40.300 | and it would compute the result

00:24:42.220 | and I didn't have to do it myself.

00:24:44.080 | Very simple.

00:24:45.180 | But if you push that a little further,

00:24:47.540 | you can have machines that learn

00:24:49.580 | and they might learn patterns.

00:24:51.580 | And already, say, deep learning neural networks,

00:24:54.040 | they can learn to recognize objects, sounds, patterns

00:24:59.020 | that humans have trouble with.

00:25:00.620 | And sometimes they do it better than humans.

00:25:02.540 | And that's so fascinating.

00:25:04.300 | And now if you take that one more step,

00:25:06.180 | you get something like evolutionary algorithms

00:25:08.220 | that discover things, they create things.

00:25:10.540 | They come up with solutions that you did not think of.

00:25:13.500 | And that just blows me away.

00:25:15.180 | It's so great that we can build systems, algorithms,

00:25:18.680 | that can be, in some sense, smarter than we are,

00:25:21.560 | that they can discover solutions that we might miss.

00:25:24.940 | A lot of times it is because we have, as humans,

00:25:26.680 | we have certain biases.

00:25:27.900 | We expect the solutions to be a certain way.

00:25:30.060 | And you don't put those biases into the algorithm

00:25:32.300 | so they are more free to explore.

00:25:34.140 | And evolution is just absolutely fantastic explorer.

00:25:37.820 | And that's what really is fascinating.

00:25:40.420 | - Yeah, I think I get made fun of a bit

00:25:43.860 | 'cause I currently don't have any kids.

00:25:45.920 | But you mentioned programs.

00:25:47.700 | I mean, do you have kids?

00:25:50.740 | - Yeah.

00:25:51.580 | - So maybe you could speak to this.

00:25:52.660 | But there's a magic to the creative process.

00:25:55.740 | With Spot, the Boston Dynamics Spot,

00:25:59.800 | but really any robot that I've ever worked on,

00:26:02.440 | it just feels like the similar kind of joy,

00:26:04.540 | I imagine, I would have as a father.

00:26:06.600 | Not the same, perhaps, level,

00:26:08.420 | but the same kind of wonderment.

00:26:10.220 | Like, there's exactly this, which is like,

00:26:12.980 | you know what you had to do initially

00:26:16.500 | to get this thing going.

00:26:19.540 | Let's speak on the computer science side,

00:26:21.700 | like what the program looks like.

00:26:23.860 | But something about it doing more

00:26:27.900 | than what the program was written on paper,

00:26:30.900 | is like, that somehow connects to the magic

00:26:34.700 | of this entire universe.

00:26:36.140 | Like, that's like, I feel like I found God.

00:26:39.220 | Every time I like, it's like,

00:26:41.120 | 'cause you've really created something that's living.

00:26:45.700 | - Yeah.

00:26:46.540 | - Even if it's a simple program.

00:26:47.380 | - It has a life of its own, has the intelligence of its own.

00:26:48.780 | It's beyond what you actually thought.

00:26:51.060 | - Yeah.

00:26:51.900 | - And that is, I think it's exactly, Spot on,

00:26:53.460 | that's exactly what it's about.

00:26:55.500 | You created something and has an ability

00:26:57.860 | to live its life and do good things.

00:27:00.940 | And you just gave it a starting point.

00:27:03.260 | So in that sense, I think it's,

00:27:04.420 | that may be part of the joy, actually.

00:27:06.500 | - But you mentioned creativity in this context,

00:27:11.020 | especially in the context of evolutionary computation.

00:27:14.160 | So, you know, we don't often think of algorithms

00:27:17.420 | as creative, so how do you think about creativity?

00:27:20.320 | - Yeah, algorithms absolutely can be creative.

00:27:25.020 | They can come up with solutions that you don't think about.

00:27:28.380 | I mean, creativity can be defined.

00:27:29.820 | A couple of requirements have to, has to be new.

00:27:32.740 | It has to be useful and it has to be surprising.

00:27:35.580 | And those certainly are true with, say,

00:27:38.020 | evolutionary computation, discovering solutions.

00:27:41.580 | So maybe an example, for instance,

00:27:44.340 | we did this collaboration with MIT Media Lab,

00:27:47.500 | Caleb Harvest Lab, where they had a hydroponic food computer

00:27:53.820 | they called it, environment that was completely

00:27:55.900 | computer controlled, nutrients, water, light, temperature,

00:27:59.580 | everything's controlled.

00:28:00.940 | Now, what do you do if you can't control everything?

00:28:05.580 | Farmers know a lot about how to do,

00:28:07.420 | how to make plants grow in their own patch of land.

00:28:10.340 | But if you can control everything, it's too much.

00:28:13.140 | And it turns out that we don't actually

00:28:14.660 | know very much about it.

00:28:16.100 | So we built a system, evolutionary optimization system,

00:28:20.380 | together with a surrogate model of how plants grow.

00:28:23.740 | And let this system explore recipes on its own.

00:28:28.740 | And initially we were focusing on light,

00:28:32.260 | how strong, what wavelengths, how long the light was on.

00:28:37.060 | And we put some boundaries,

00:28:38.740 | which we thought were reasonable.

00:28:40.340 | For instance, that there was at least six hours of darkness

00:28:44.580 | like night, because that's what we have in the world.

00:28:47.380 | And very quickly, the system evolution pushed

00:28:51.700 | all the recipes to that limit.

00:28:54.340 | We were trying to grow basil,

00:28:56.100 | and we had initially had some 200, 300 recipes,

00:29:00.260 | exploration as well as known recipes.

00:29:02.380 | But now we are going beyond that.

00:29:04.300 | And everything was like pushed at that limit.

00:29:06.700 | So we look at it and say, well, you know,

00:29:08.540 | we can easily just change it.

00:29:09.540 | Let's have it your way.

00:29:10.980 | And it turns out the system discovered

00:29:13.700 | that basil does not need to sleep.

00:29:15.660 | 24 hours, lights on, and it will thrive.

00:29:19.700 | It will be bigger, it will be tastier.

00:29:21.700 | And this was a big surprise, not just to us,

00:29:24.700 | but also the biologists in the team that anticipated

00:29:28.940 | that there's some constraints that are in the world.

00:29:32.380 | For a reason, it turns out that evolution

00:29:34.540 | did not have the same bias.

00:29:36.180 | And therefore it discovered something that was creative.

00:29:38.980 | It was surprising, it was useful, and it was new.

00:29:41.540 | - That's fascinating to think about,

00:29:42.900 | like the things we think that are fundamental

00:29:45.580 | to living systems on Earth today,

00:29:48.420 | whether they're actually fundamental

00:29:49.940 | or they somehow fit the constraints of the system,

00:29:53.940 | and all we'll have to do is just remove the constraints.

00:29:56.700 | Do you ever think about, I don't know how much you know

00:30:00.580 | about brain-computer interfaces and Neuralink.

00:30:03.540 | The idea there is, you know, our brains are very limited.

00:30:08.540 | And if we just allow, we plug in,

00:30:11.740 | we provide a mechanism for a computer

00:30:13.980 | to speak with the brain, so you're thereby expanding

00:30:17.140 | the computational power of the brain,

00:30:19.500 | the possibilities there, sort of from a very high level

00:30:22.940 | philosophical perspective, is limitless.

00:30:27.020 | But I wonder how limitless it is.

00:30:30.700 | Are the constraints we have like features

00:30:33.460 | that are fundamental to our intelligence?

00:30:36.020 | Or is this just like this weird constraint

00:30:38.420 | in terms of our brain size and skull and lifespan

00:30:42.700 | and the senses, it's just the weird little

00:30:46.260 | like a quirk of evolution, and if we just open that up,

00:30:50.020 | like add much more senses,

00:30:51.540 | add much more computational power,

00:30:53.740 | the intelligence will expand exponentially.

00:30:57.860 | Do you have a sense about constraints,

00:31:02.860 | the relationship of evolution and computation

00:31:05.420 | to the constraints of the environment?

00:31:07.320 | - Well, at first I'd like to comment on that,

00:31:12.420 | like changing the inputs to human brain.

00:31:15.820 | - Yes, that would be great. - And flexibility

00:31:17.300 | of the brain, I think there's a lot of that.

00:31:20.740 | There are experiments that are done in animals,

00:31:22.380 | like migangas are, but they might be switching

00:31:25.740 | the auditory and visual information

00:31:29.220 | and going to the wrong part of the cortex,

00:31:31.500 | and the animal was still able to hear

00:31:34.180 | and perceive the visual environment.

00:31:36.540 | And there are kids that are born with severe disorders,

00:31:41.180 | and sometimes they have to remove half of the brain,

00:31:43.980 | like one half, and they still grow up,

00:31:46.180 | they have the functions migrate to the other parts.

00:31:48.380 | There's a lot of flexibility like that.

00:31:50.420 | So I think it's quite possible to hook up the brain

00:31:55.020 | with different kinds of sensors, for instance,

00:31:57.660 | and something that we don't even quite understand

00:32:00.340 | or have today, and different kinds of wavelengths

00:32:02.580 | or whatever they are, and then the brain can learn

00:32:05.700 | to make sense of it.

00:32:07.380 | And that, I think, is this good hope

00:32:10.020 | that these prosthetic devices, for instance, work,

00:32:12.760 | not because we make them so good and so easy to use,

00:32:15.740 | but the brain adapts to them and can learn

00:32:17.860 | to take advantage of them.

00:32:19.160 | And so in that sense, if there's a trouble, a problem,

00:32:23.460 | I think the brain can be used to correct it.

00:32:26.220 | Now, going beyond what we have today, can you get smarter?

00:32:29.140 | That's really much harder to do.

00:32:31.600 | Giving the brain more input probably might overwhelm it.

00:32:35.540 | It would have to learn to filter it and focus

00:32:39.740 | in order to use the information effectively.

00:32:43.340 | And augmenting intelligence with some kind

00:32:47.160 | of external devices like that might be difficult, I think.

00:32:51.560 | But replacing what's lost, I think, is quite possible.

00:32:55.680 | - Right, so our intuition allows us to sort of imagine

00:32:59.360 | that we can replace what's been lost,

00:33:01.360 | but expansion beyond what we have.

00:33:03.480 | I mean, we're already one of the most,

00:33:05.360 | if not the most intelligent things on this earth, right?

00:33:07.800 | So it's hard to imagine if the brain can hold up

00:33:12.800 | with an order of magnitude greater set

00:33:15.500 | of information thrown at it,

00:33:18.060 | if it can reason through that.

00:33:20.740 | Part of me, this is the Russian thing, I think,

00:33:22.580 | is I tend to think that the limitations

00:33:25.420 | is where the superpower is, that immortality

00:33:30.420 | and huge increase in bandwidth of information

00:33:37.440 | by connecting computers with the brain

00:33:40.200 | is not going to produce greater intelligence.

00:33:42.680 | It might produce lesser intelligence.

00:33:44.320 | So I don't know, there's something

00:33:46.080 | about the scarcity being essential

00:33:51.080 | to fitness or performance, but that could be just

00:33:56.080 | 'cause we're so limited.

00:33:59.000 | - No, exactly, you make do with what you have.

00:34:00.680 | But you don't have to pipe it directly to the brain.

00:34:04.280 | I mean, we already have devices like phones

00:34:07.560 | where we can look up information at any point.

00:34:10.160 | And that can make us more productive.

00:34:12.320 | You don't have to argue about, I don't know,

00:34:14.000 | what happened in that baseball game or whatever it is,

00:34:16.400 | because you can look it up right away.

00:34:17.680 | And I think in that sense, we can learn to utilize tools.

00:34:22.040 | And that's what we have been doing for a long, long time.

00:34:25.240 | So, and we are already, the brain is already drinking

00:34:29.000 | from the fire hose, like vision.

00:34:32.440 | There's way more information in vision

00:34:34.520 | than we actually process.

00:34:35.680 | So brain's already good at identifying what matters.

00:34:38.960 | And that, we can switch that from vision

00:34:42.880 | to some other wavelength or some other kind of modality.

00:34:45.040 | But I think that the same processing principles

00:34:47.120 | probably still apply.

00:34:49.040 | But also, indeed, this ability to have information

00:34:53.760 | more accessible and more relevant, I think,

00:34:56.000 | can enhance what we do.

00:34:57.760 | I mean, kids today at school, they learn about DNA.

00:35:00.960 | I mean, things that were discovered

00:35:02.640 | just a couple of years ago,

00:35:04.640 | and it's already common knowledge,

00:35:06.480 | and we are building on it.

00:35:07.600 | And we don't see a problem where there's too much information

00:35:12.600 | that we can't absorb and learn.

00:35:15.160 | Maybe people become a little bit more narrow

00:35:17.560 | in what they know, they are in one field.

00:35:20.920 | But this information that we have accumulated,

00:35:23.760 | it is passed on, and people are picking up on it,

00:35:26.120 | and they are building on it.

00:35:27.600 | So it's not like we have reached the point of saturation.

00:35:31.040 | We have still this process that allows us to be selective

00:35:34.520 | and decide what's interesting, I think still works,

00:35:37.600 | even with the more information we have today.

00:35:40.120 | - Yeah, it's fascinating to think about

00:35:42.200 | like Wikipedia becoming a sensor,

00:35:45.360 | like so the fire hose of information from Wikipedia.

00:35:49.080 | So it's like you integrate it directly into the brain

00:35:51.800 | to where you're thinking, like you're observing the world

00:35:54.240 | with all of Wikipedia directly piping into your brain.

00:35:57.840 | So like when I see a light,

00:35:59.920 | I immediately have like the history of who invented

00:36:03.640 | electricity, like integrated very quickly into.

00:36:07.560 | So just the way you think about the world

00:36:09.840 | might be very interesting

00:36:11.240 | if you can integrate that kind of information.

00:36:13.280 | What are your thoughts, if I could ask,

00:36:15.640 | on the early steps on the Neuralink side,

00:36:20.360 | I don't know if you got a chance to see,

00:36:21.520 | but there's a monkey playing pong

00:36:24.720 | - Mm, yeah.

00:36:25.960 | - through the brain computer interface.

00:36:27.800 | And the dream there is sort of,

00:36:30.680 | you're already replacing the thumbs essentially

00:36:33.560 | that you would use to play a video game.

00:36:35.920 | The dream is to be able to increase further

00:36:38.960 | the interface by which you interact with the computer.

00:36:43.440 | Are you impressed by this?

00:36:44.640 | Are you worried about this?

00:36:46.480 | What are your thoughts as a human?

00:36:48.000 | - I think it's wonderful.

00:36:48.880 | I think it's great that we could do something like that.

00:36:51.600 | I mean, you can, there are devices that read your EEG,

00:36:55.080 | for instance, and humans can learn to control things

00:37:00.080 | using just their thoughts in that sense.

00:37:02.760 | And I don't think it's that different.

00:37:04.920 | I mean, those signals would go to limbs,

00:37:06.720 | they would go to thumbs.

00:37:08.320 | Now the same signals go through a sensor

00:37:11.200 | to some computing system.

00:37:12.920 | It still probably has to be built on human terms,

00:37:17.520 | not to overwhelm them, but utilize what's there

00:37:20.000 | and sense the right kind of patterns

00:37:23.760 | that are easy to generate.

00:37:24.840 | But, oh, that I think is really quite possible

00:37:27.760 | and wonderful and could be very much more efficient.

00:37:30.720 | - Is there, so you mentioned surprising

00:37:34.160 | being a characteristic of creativity.

00:37:37.080 | Is there something, you already mentioned a few examples,

00:37:39.800 | but is there something that jumps out at you

00:37:41.920 | as was particularly surprising

00:37:44.600 | from the various evolutionary computation systems

00:37:48.680 | you've worked on, the solutions that were

00:37:50.880 | come up along the way, not necessarily the final solutions,

00:37:55.320 | but maybe things that were even discarded.

00:37:58.720 | Is there something that just jumps to mind?

00:38:00.760 | - It happens all the time.

00:38:02.200 | I mean, evolution is so creative,

00:38:05.680 | so good at discovering solutions you don't anticipate.

00:38:09.280 | A lot of times they are taking advantage of something

00:38:12.720 | that you didn't think was there,

00:38:13.840 | like a bug in the software, for instance.

00:38:16.000 | A lot of, there's a great paper,

00:38:17.640 | the community put it together,

00:38:19.160 | about surprising anecdotes about evolutionary computation.

00:38:22.960 | A lot of them are indeed, in some software environment,

00:38:25.640 | there was a loophole or a bug,

00:38:28.120 | and the system utilizes that.

00:38:30.600 | - By the way, for people who want to read it,

00:38:31.960 | it's kind of fun to read.

00:38:33.120 | It's called "The Surprising Creativity of Digital Evolution,

00:38:36.040 | "A Collection of Anecdotes from the Evolutionary Computation

00:38:39.240 | "and Artificial Life Research Communities."

00:38:41.600 | And there's just a bunch of stories

00:38:43.200 | from all the seminal figures in this community.

00:38:45.880 | You have a story in there that relates to you,

00:38:48.560 | at least, on the tic-tac-toe memory bomb.

00:38:51.040 | So can you, I guess, describe that situation,

00:38:54.800 | if you think that's--

00:38:55.720 | - Yeah, that's a quite a bit smaller scale

00:38:59.680 | than our basic doesn't need to sleep surprise,

00:39:03.080 | but it was actually done by students in my class,

00:39:06.680 | in a neural nets evolutionary computation class.

00:39:09.480 | There was an assignment.

00:39:11.760 | It was perhaps a final project

00:39:13.920 | where people built game-playing AI.

00:39:17.720 | It was an AI class.

00:39:18.840 | And it was for tic-tac-toe or five in a row

00:39:23.080 | in a large board.

00:39:24.600 | And this one team evolved a neural network

00:39:28.200 | to make these moves.

00:39:29.960 | And they set it up, the evolution.

00:39:32.760 | They didn't really know what would come out,

00:39:35.280 | but it turned out that they did really well.

00:39:37.040 | Evolution actually won the tournament.

00:39:38.920 | And most of the time when it won,

00:39:40.560 | it won because the other teams crashed.

00:39:43.480 | And then when we look at it, like what was going on,

00:39:45.760 | was that evolution discovered that if it makes a move

00:39:48.240 | that's really, really far away,

00:39:49.960 | like millions of squares away,

00:39:53.440 | the other teams, the other programs just expanded memory

00:39:57.800 | in order to take that into account

00:39:59.160 | until they ran out of memory and crashed.

00:40:01.200 | And then you win a tournament

00:40:03.200 | by crashing all your opponents.

00:40:05.720 | - I think that's quite a profound example,

00:40:08.920 | which probably applies to most games

00:40:13.200 | from even a game theoretic perspective,

00:40:16.920 | that sometimes to win,

00:40:18.280 | you don't have to be better within the rules of the game.

00:40:22.720 | You have to come up with ways to break your opponent's brain

00:40:27.720 | if it's a human, like not through violence,

00:40:31.360 | but through some hack where the brain just is not,

00:40:36.440 | you're basically, how would you put it?

00:40:39.360 | You're going outside the constraints

00:40:43.080 | of where the brain is able to function.

00:40:45.080 | - Expectations of your opponent.

00:40:46.560 | I mean, this was even Kasparov pointed that out

00:40:49.560 | that when Deep Blue was playing against Kasparov,

00:40:51.760 | that it was not playing the same way as Kasparov expected.

00:40:55.440 | And this has to do with not having the same biases.

00:40:59.720 | And that's really one of the strengths of the AI approach.

00:41:06.320 | - Can you at a high level say,

00:41:08.120 | what are the basic mechanisms

00:41:10.400 | of evolutionary computation algorithms

00:41:12.800 | that use something that could be called

00:41:15.800 | an evolutionary approach?

00:41:17.680 | Like how does it work?

00:41:18.880 | What are the connections to the,

00:41:21.720 | what are the echoes of the connection to his biological?

00:41:24.840 | - A lot of these algorithms really do take motivation

00:41:27.120 | from biology, but they are carry catches.

00:41:29.600 | You try to essentialize it

00:41:31.320 | and take the elements that you believe matter.

00:41:33.600 | So in evolutionary computation,

00:41:35.920 | it is the creation of variation

00:41:38.080 | and then the selection upon that.

00:41:40.720 | So the creation of variation,

00:41:41.880 | you have to have some mechanism

00:41:43.120 | that allow you to create new individuals

00:41:44.760 | that are very different from what you already have.

00:41:47.120 | That's the creativity part.

00:41:49.040 | And then you have to have some way of measuring

00:41:50.760 | how well they are doing and using that measure to select

00:41:55.600 | who goes to the next generation and you continue.

00:41:58.160 | - So first you have to have some kind

00:42:00.720 | of digital representation of an individual

00:42:03.160 | that can be then modified.

00:42:04.520 | So I guess humans in biological systems have DNA

00:42:08.760 | and all those kinds of things.

00:42:09.720 | And so you have to have similar kind of encodings

00:42:12.160 | in a computer program.

00:42:13.400 | - Yes, and that is a big question.

00:42:15.040 | How do you encode these individuals?

00:42:16.960 | So there's a genotype, which is that encoding

00:42:19.560 | and then a decoding mechanism,

00:42:21.400 | which gives you the phenotype,

00:42:23.040 | which is the actual individual that then performs the task

00:42:26.400 | and in an environment can be evaluated how good it is.

00:42:31.280 | So even that mapping is a big question

00:42:33.120 | and how do you do it?

00:42:34.920 | But typically the representations are either

00:42:37.320 | they are strings of numbers or they are some kind of trees.

00:42:39.760 | Those are something that we know very well

00:42:41.720 | in computer science and we try to do that.

00:42:43.560 | But they, and DNA in some sense is also a sequence

00:42:48.000 | and a string.

00:42:49.480 | So it's not that far from it,

00:42:52.000 | but DNA also has many other aspects

00:42:54.840 | that we don't take into account necessarily

00:42:56.720 | like there's folding and interactions that are other

00:43:01.120 | than just the sequence itself.

00:43:03.560 | And lots of that is not yet captured

00:43:06.000 | and we don't know whether they are really crucial.

00:43:09.000 | Evolution, biological evolution has produced

00:43:12.600 | wonderful things, but if you look at them,

00:43:16.000 | it's not necessarily the case that every piece

00:43:18.560 | is irreplaceable and essential.

00:43:20.880 | There's a lot of baggage 'cause you have to construct it

00:43:23.680 | and it has to go through various stages

00:43:25.360 | and we still have appendix and we have tailbones

00:43:29.360 | and things like that that are not really that useful.

00:43:31.360 | If you try to explain them now, it would make no sense,

00:43:34.320 | it would be very hard.

00:43:35.200 | But if you think of us as productive evolution,

00:43:38.200 | you can see where they came from.

00:43:39.240 | They were useful at one point perhaps and no longer are,

00:43:42.120 | but they're still there.

00:43:43.400 | So that process is complex

00:43:47.080 | and your representation should support it.

00:43:50.800 | And that is quite difficult if we are limited

00:43:56.280 | with strings or trees and then we are pretty much limited

00:44:01.280 | what can be constructed.

00:44:03.680 | And one thing that we are still missing

00:44:05.600 | in evolutionary computation in particular

00:44:07.520 | is what we saw in biology, major transitions.

00:44:11.400 | So that you go from, for instance, single cell

00:44:14.520 | to multicell organisms and eventually societies.

00:44:17.160 | There are transitions of level of selection

00:44:19.560 | and level of what a unit is.

00:44:22.080 | And that's something we haven't captured

00:44:24.200 | in evolutionary computation yet.

00:44:26.000 | - Does that require a dramatic expansion

00:44:28.640 | of the representation?

00:44:29.960 | Is that what that is?

00:44:31.600 | - Most likely it does, but it's quite,

00:44:34.400 | we don't even understand it in biology very well

00:44:36.840 | where it's coming from.

00:44:37.680 | So it would be really good to look at major transitions

00:44:40.480 | in biology, try to characterize them a little bit more

00:44:43.280 | in detail, what the processes are.

00:44:45.360 | How does a, so like a unit, a cell

00:44:48.600 | is no longer evaluated alone, it's evaluated

00:44:51.480 | as part of a community, a multicell organism.

00:44:54.720 | Even though it could reproduce, now it can't alone.

00:44:57.320 | It has to have this environment.

00:44:59.360 | So there's a push to another level, at least the selection.

00:45:03.400 | - And how do you make that jump to the next level?

00:45:04.760 | - Yes, how do you make the jump?

00:45:06.080 | - As part of the algorithm.

00:45:07.240 | - Yeah, yeah.

00:45:08.160 | So we haven't really seen that in computation yet.

00:45:12.080 | And there are certainly attempts

00:45:13.880 | to have open-ended evolution.

00:45:15.800 | Things that could add more complexity

00:45:18.400 | and start selecting at a higher level,

00:45:20.840 | but it is still not quite the same

00:45:24.680 | as going from single to multi to society,

00:45:27.080 | for instance, in biology.

00:45:29.000 | - So there essentially would be,

00:45:31.720 | as opposed to having one agent,

00:45:33.440 | those agent all of a sudden spontaneously decide

00:45:36.240 | to then be together, and then your entire system

00:45:41.000 | would then be treating them as one agent.

00:45:43.600 | - Something like that.

00:45:44.720 | - Some kind of weird merger.

00:45:46.320 | But also, so you mentioned, I think you mentioned selection.

00:45:49.200 | So basically there's an agent,

00:45:51.080 | and they don't get to live on if they don't do well.

00:45:54.240 | So there's some kind of measure

00:45:55.320 | of what doing well is and isn't.

00:45:57.320 | And does mutation come into play at all

00:46:01.920 | in the process, and what role does it serve?

00:46:04.200 | - Yeah, so, and again, back to what

00:46:06.600 | the computational mechanisms of evolution computation are.

00:46:08.640 | So the way to create variation,

00:46:12.720 | you can take multiple individuals, two usually,

00:46:15.120 | but you could do more.

00:46:17.200 | And you exchange the parts of the representation.

00:46:20.880 | You do some kind of recombination,

00:46:22.720 | could be crossover, for instance.

00:46:25.000 | In biology, you do have DNA strings

00:46:28.920 | that are cut and put together again.

00:46:32.120 | We could do something like that.

00:46:34.320 | And it seems to be that in biology,

00:46:36.480 | the crossover is really the workhorse

00:46:39.520 | in biological evolution.

00:46:42.160 | In computation, we tend to rely more on mutation.

00:46:47.080 | And that is making random changes

00:46:50.120 | into parts of the chromosome.

00:46:51.320 | You could try to be intelligent

00:46:53.200 | and target certain areas of it,

00:46:55.800 | and make the mutations also follow some principle.

00:47:00.800 | Like you collect statistics of performance

00:47:03.400 | and correlations, and try to make mutations

00:47:05.600 | you believe are going to be helpful.

00:47:07.960 | That's where evolution computation has moved

00:47:10.440 | in the last 20 years.

00:47:12.160 | I mean, evolution computation has been around for 50 years,

00:47:14.000 | but a lot of the recent--

00:47:16.160 | - Success comes from mutation.

00:47:17.680 | - Comes from using statistics.

00:47:20.240 | It's like the rest of machine learning,

00:47:22.120 | based on statistics.

00:47:23.040 | We use similar tools to guide evolutionary computation.

00:47:26.160 | And in that sense, it has diverged a bit

00:47:28.760 | from biological evolution.

00:47:30.680 | And that's one of the things I think

00:47:32.240 | we could look at again,

00:47:34.200 | having a weaker selection, more crossover,

00:47:38.520 | large populations, more time,

00:47:40.840 | and maybe a different kind of creativity

00:47:42.840 | would come out of it.

00:47:43.960 | We are very impatient in evolutionary computation today.

00:47:47.000 | We want answers right now, right quickly.

00:47:49.600 | And if somebody doesn't perform, kill it.

00:47:52.120 | And biological evolution doesn't work quite that way.

00:47:56.480 | - It's more patient.

00:47:57.840 | - Yes, much more patient.

00:48:00.040 | - So I guess we need to add some kind of mating,

00:48:03.680 | some kind of dating mechanisms,

00:48:05.960 | like marriage maybe in there,

00:48:07.400 | so into our algorithms to improve the combination,

00:48:13.680 | as opposed to all mutation doing all of the work.

00:48:16.000 | - Yeah, and many ways of being successful.

00:48:18.920 | Usually in evolutionary computation, we have one goal,

00:48:21.600 | play this game really well compared to others.

00:48:25.920 | But in biology, there are many ways of being successful.

00:48:28.680 | You can build niches, you can be stronger, faster,

00:48:32.240 | larger, or smarter, or eat this or eat that.

00:48:36.800 | So there are many ways to solve the same problem of survival.

00:48:40.600 | And that then breeds creativity.

00:48:43.840 | And it allows more exploration.

00:48:46.760 | And eventually you get solutions

00:48:48.720 | that are perhaps more creative,

00:48:51.160 | rather than trying to go from initial population directly,

00:48:54.120 | or more or less directly to your maximum fitness,

00:48:57.400 | which you measure as just one metric.

00:49:00.840 | - So in a broad sense,

00:49:03.840 | before we talk about neuroevolution,

00:49:06.360 | do you see evolutionary computation

00:49:11.200 | as more effective than deep learning in certain contexts?

00:49:14.160 | Machine learning, broadly speaking.

00:49:16.640 | Maybe even supervised machine learning.

00:49:18.680 | I don't know if you want to draw any kind of lines

00:49:21.000 | and distinctions and borders where they rub up

00:49:23.760 | against each other kind of thing,

00:49:25.400 | or one is more effective than the other

00:49:27.000 | in the current state of things.

00:49:28.440 | - Yes, of course, they are very different

00:49:30.240 | and they address different kinds of problems.

00:49:32.280 | And the deep learning has been really successful

00:49:36.720 | in domains where we have a lot of data.

00:49:39.800 | And that means not just data about situations,

00:49:42.440 | but also what the right answers were.

00:49:45.120 | So labeled examples, or there might be predictions,

00:49:47.840 | might be weather prediction

00:49:48.840 | where the data itself becomes labeled.

00:49:51.720 | What happened, what the weather was today,

00:49:53.160 | and what it will be tomorrow.

00:49:55.480 | So they are very effective,

00:49:58.360 | deep learning methods on that kind of tasks.

00:50:01.400 | But there are other kinds of tasks

00:50:03.400 | where we don't really know what the right answer is.

00:50:06.360 | Game playing, for instance,

00:50:07.520 | but many robotics tasks and actions in the world,

00:50:12.520 | decision-making, and actual practical applications

00:50:17.680 | like treatments and healthcare,

00:50:19.440 | or investment in stock market.

00:50:21.360 | Many tasks are like that.

00:50:22.680 | We don't know and we'll never know

00:50:24.840 | what the optimal answers were.

00:50:26.640 | And there you need different kinds of approach.

00:50:28.600 | Reinforcement learning is one of those.

00:50:30.840 | Reinforcement learning comes from biology as well.

00:50:33.760 | Agents learn during their lifetime.

00:50:35.400 | They buries and sometimes they get sick

00:50:37.560 | and then they don't and get stronger.

00:50:40.280 | And then that's how you learn.

00:50:42.280 | And evolution is also a mechanism like that,

00:50:46.040 | but a different timescale because you have a population.

00:50:48.920 | Not an individual during his lifetime,

00:50:50.840 | but an entire population as a whole can discover what works.

00:50:55.200 | And there you can afford individuals that don't work out.

00:50:58.960 | They learn, everybody dies and you have a next generation

00:51:02.080 | and it will be better than the previous one.

00:51:04.120 | So that's the big difference between these methods.

00:51:07.640 | They apply to different kinds of problems.

00:51:09.840 | And in particular, there's often a comparison

00:51:15.120 | that's kind of interesting and important

00:51:16.640 | between reinforcement learning and evolutionary computation.

00:51:20.080 | And initially, reinforcement learning

00:51:23.400 | was about individual learning during their lifetime.

00:51:25.960 | And evolution is more engineering.

00:51:28.160 | You don't care about the lifetime.

00:51:29.720 | You don't care about all the individuals that are tested.

00:51:32.600 | You only care about the final result.

00:51:34.520 | The last one, the best candidate that evolution produced.

00:51:39.080 | And that sense, they also apply

00:51:40.520 | to different kinds of problems.

00:51:42.520 | And another boundary starting to blur a bit.

00:51:46.160 | You can use evolution as an online method

00:51:48.680 | and reinforcement learning to create engineering solutions,

00:51:51.520 | but that's still roughly the distinction.

00:51:55.280 | And from the point of view, what algorithm you wanna use,

00:52:00.280 | if you have something where there is a cost

00:52:02.280 | for every trial, reinforcement learning might be your choice.

00:52:06.080 | Now, if you have a domain

00:52:07.760 | where you can use a surrogate perhaps,

00:52:10.240 | so you don't have much of a cost for trial,

00:52:13.600 | and you want to have surprises,

00:52:16.480 | you want to explore more broadly,

00:52:18.640 | then this population-based method is perhaps a better choice

00:52:23.360 | because you can try things out that you wouldn't afford

00:52:26.960 | when you're doing reinforcement learning.

00:52:28.560 | - There's very few things as entertaining

00:52:31.680 | as watching either evolution computation

00:52:33.760 | or reinforcement learning teaching a simulated robot

00:52:36.600 | to walk.

00:52:37.440 | Maybe there's a higher level question

00:52:42.400 | that could be asked here,

00:52:43.640 | but do you find this whole space of applications

00:52:47.560 | in the robotics interesting for evolution computation?

00:52:51.720 | - Yeah, yeah, very much.

00:52:53.520 | And indeed, there are fascinating videos of that.

00:52:56.480 | And that's actually one of the examples

00:52:58.360 | where you can contrast the difference.

00:53:00.560 | - Between reinforcement learning and evolution.

00:53:03.200 | - Yes, so if you have a reinforcement learning agent,

00:53:06.320 | it tries to be conservative

00:53:08.000 | because it wants to walk as long as possible and be stable.

00:53:11.880 | But if you have evolutionary computation,

00:53:13.720 | it can afford these agents that go haywire.

00:53:17.280 | They fall flat on their face,

00:53:19.200 | and they take a step, and then they jump,

00:53:21.640 | and then again fall flat.

00:53:23.240 | And eventually what comes out of that

00:53:25.280 | is something like a falling that's controlled.

00:53:28.280 | You take another step, another step,

00:53:30.440 | and you no longer fall.

00:53:32.320 | Instead, you run, you go fast.

00:53:34.200 | So that's a way of discovering

00:53:36.200 | something that's hard to discover step by step,

00:53:38.400 | incrementally, because you can afford

00:53:40.480 | these evolutionist dead ends,

00:53:43.680 | although they are not entirely dead ends

00:53:45.520 | in the sense that they can serve as stepping stones.

00:53:47.760 | When you take two of those, put them together,

00:53:49.880 | you get something that works even better.

00:53:52.440 | And that is a great example of this kind of discovery.

00:53:55.920 | - Yeah, learning to walk is fascinating.

00:53:58.160 | I talk quite a bit to Russ Tedrake

00:53:59.920 | 'cause at MIT, there's a community of folks

00:54:03.120 | who just, roboticists, who love the elegance

00:54:06.600 | and beauty of movement.

00:54:09.760 | And walking, bipedal robotics, is beautiful,

00:54:14.760 | but also exceptionally dangerous

00:54:19.440 | in the sense that you're constantly falling,

00:54:22.400 | essentially, if you want to do elegant movement.

00:54:25.360 | And the discovery of that is,

00:54:28.920 | (sighs)

00:54:31.080 | I mean, it's such a good example

00:54:33.800 | of that the discovery of a good solution

00:54:37.520 | sometimes requires a leap of faith and patience

00:54:39.760 | and all those kinds of things.

00:54:41.520 | I wonder what other spaces where you had

00:54:43.600 | to discover those kinds of things in.

00:54:45.400 | - Yeah, yeah.

00:54:47.520 | Another interesting direction is learning

00:54:51.320 | for virtual creatures, learning to walk.

00:54:56.360 | We did a study in simulation, obviously,

00:54:59.920 | that you create those creatures,

00:55:02.640 | not just their controller, but also their body.

00:55:05.240 | So you have cylinders, you have muscles,

00:55:07.880 | you have joints and sensors,

00:55:11.160 | and you're creating creatures that look quite different.

00:55:14.080 | Some of them have multiple legs,

00:55:15.440 | some of them have no legs at all.

00:55:17.640 | And then the goal was to get them to move,

00:55:20.080 | to walk, to run.

00:55:21.960 | And what was interesting is that when you evolve

00:55:25.440 | the controller together with the body,

00:55:28.400 | you get movements that look natural

00:55:30.480 | because they're optimized for that physical setup.

00:55:33.600 | And these creatures, you start believing them,

00:55:36.160 | that they're alive because they walk in a way

00:55:38.080 | that you would expect somebody

00:55:39.560 | with that kind of a setup to walk.

00:55:41.920 | - Yeah, there's something subjective also about that.

00:55:45.760 | I've been thinking a lot about that,

00:55:47.120 | especially in the human-robot interaction context.

00:55:53.800 | I mentioned Spot, the Boston Dynamics robot.

00:55:57.400 | There is something about human-robot communication.

00:56:01.420 | Let's say, let's put it in another context,

00:56:03.380 | something about human and dog context,

00:56:07.940 | like a living dog,

00:56:10.120 | where there's a dance of communication.

00:56:13.360 | First of all, the eyes, you both look at the same thing

00:56:15.520 | and dogs communicate with their eyes as well.

00:56:18.080 | Like if you and a dog want to deal with a person,

00:56:23.200 | deal with a particular object,

00:56:24.560 | you will look at the person,

00:56:26.200 | the dog will look at you and then look at the object

00:56:28.080 | and look back at you, all those kinds of things.

00:56:30.320 | But there's also just the elegance of movement.

00:56:33.240 | I mean, there's the, of course, the tail

00:56:35.840 | and all those kinds of mechanisms of communication.

00:56:38.080 | It all seems natural and often joyful.

00:56:41.880 | And for robots to communicate that

00:56:44.360 | is really difficult how to figure that out

00:56:47.240 | because it almost seems impossible to hard-code in.

00:56:50.800 | You can hard-code it for a demo purpose,

00:56:53.800 | something like that, but it's essentially choreographed.

00:56:58.120 | Like if you watch some of the Boston Dynamics videos

00:57:00.300 | where they're dancing,

00:57:01.760 | all of that is choreographed by human beings.

00:57:05.640 | But to learn how to, with your movement,

00:57:09.380 | demonstrate a naturalness, an elegance, that's fascinating.

00:57:14.380 | Of course, in the physical space,

00:57:15.720 | that's very difficult to do,

00:57:16.840 | to learn the kind of scale that you're referring to,

00:57:20.100 | but the hope is that you could do that in simulation

00:57:23.080 | and then transfer it into the physical space

00:57:25.360 | if you're able to model the robots efficiently, naturally.

00:57:28.680 | - Yeah, and sometimes I think that it requires

00:57:31.680 | a theory of mind on the side of the robot

00:57:35.000 | that they understand what you're doing

00:57:38.920 | because they themselves are doing something similar.

00:57:41.440 | And that's a big question too.

00:57:44.360 | We talked about intelligence in general

00:57:47.440 | and the social aspect of intelligence

00:57:50.040 | and I think that's what is required,

00:57:52.040 | that we humans understand other humans

00:57:53.840 | because we assume that they are similar to us.

00:57:57.040 | We have one simulation we did a while ago,

00:57:59.120 | Ken Stanley did that.

00:58:01.440 | Two robots that were competing, simulation, like I said,

00:58:06.440 | they were foraging for food to gain energy.

00:58:09.320 | And then when they were really strong,

00:58:10.680 | they would bounce into the other robot

00:58:12.680 | and win if they were stronger.

00:58:14.860 | And we watched evolution discover

00:58:17.320 | more and more complex behaviors.

00:58:18.920 | They first went to the nearest food

00:58:21.040 | and then they started to plot a trajectory

00:58:24.320 | so they get more, but then they started to pay attention

00:58:28.440 | what the other robot was doing.

00:58:30.320 | And in the end, there was a behavior

00:58:32.720 | where one of the robots, the more sophisticated one,

00:58:35.840 | sensed where the food pieces were

00:58:40.200 | and identified that the other robot

00:58:42.080 | was close to two of a very far distance

00:58:46.000 | and there was one more food nearby.

00:58:48.720 | So it faked, now I'm using anthropomorphized terms,

00:58:53.400 | but it made a move towards those other pieces

00:58:55.900 | in order for the other robot to actually go and get them.

00:58:59.080 | Because it knew that the last remaining piece of food

00:59:02.400 | was close and the other robot would have to travel

00:59:04.960 | a long way, lose its energy,

00:59:06.960 | and then lose the whole competition.

00:59:10.440 | So there was like an emergence

00:59:12.200 | of something like a theory of mind,

00:59:13.640 | knowing what the other robot would do

00:59:16.640 | to guide it towards bad behavior in order to win.

00:59:19.520 | So we can get things like that happen

00:59:21.800 | in simulation as well.

00:59:23.040 | - But that's a complete natural emergence

00:59:25.360 | of a theory of mind.

00:59:26.180 | But I feel like if you add a little bit

00:59:29.480 | of a place for a theory of mind to emerge easier

00:59:34.480 | then you can go really far.

00:59:37.240 | I mean, some of these things with evolution,

00:59:39.760 | you add a little bit of design in there,

00:59:44.000 | it'll really help.

00:59:45.560 | And I tend to think that a very simple theory of mind

00:59:50.560 | will go a really long way for cooperation between agents

00:59:54.880 | and certainly for human-robot interaction.

00:59:57.520 | Like it doesn't have to be super complicated.

00:59:59.760 | I've gotten a chance in the autonomous vehicle space

01:00:03.520 | to watch vehicles interact with pedestrians

01:00:07.040 | or pedestrians interacting with vehicles in general.

01:00:09.920 | I mean, you would think that there's a very complicated

01:00:13.000 | theory of mind thing going on,

01:00:14.520 | but I have a sense, it's not well understood yet,

01:00:17.000 | but I have a sense it's pretty dumb.

01:00:19.480 | Like it's pretty simple.

01:00:21.060 | There's a social contract there where between humans,

01:00:25.560 | a human driver and a human crossing the road

01:00:28.160 | where the human crossing the road trusts

01:00:32.000 | that the human in the car is not going to murder them.

01:00:34.600 | And there's something about,

01:00:36.000 | again, back to that mortality thing,

01:00:38.220 | there's some dance of ethics and morality

01:00:44.000 | that's built in that you're mapping your own morality

01:00:47.560 | onto the person in the car.

01:00:50.040 | And even if they're driving at a speed where you think

01:00:54.080 | if they don't stop, they're going to kill you,

01:00:56.200 | you trust that if you step in front of them,

01:00:58.140 | they're going to hit the brakes.

01:00:59.480 | And there's that weird dance that we do

01:01:02.180 | that I think is a pretty simple model,

01:01:04.680 | but of course it's very difficult

01:01:06.840 | to introspect what it is.

01:01:08.520 | And autonomous robots in the human-robot interaction context

01:01:12.040 | have to build that.

01:01:13.800 | Current robots are much less than what you're describing.

01:01:17.320 | They're currently just afraid of everything.

01:01:19.360 | They're not the kind that fall and discover how to run.

01:01:24.080 | They're more like, please don't touch anything,

01:01:26.800 | don't hurt anything, stay as far away

01:01:29.000 | from humans as possible.

01:01:30.200 | Treat humans as ballistic objects that you can't,

01:01:34.120 | that you do with a large spatial envelope,

01:01:38.760 | make sure you do not collide with.

01:01:40.800 | - That's how like you mentioned Elon Musk

01:01:43.440 | thinks about autonomous vehicles.

01:01:45.360 | I tend to think autonomous vehicles

01:01:47.680 | need to have a beautiful dance between human and machine,

01:01:50.640 | where it's not just the collision avoidance problem,

01:01:53.320 | but a weird dance.

01:01:55.920 | - Yeah, I think these systems need to be able to predict

01:02:00.000 | what will happen, what the other agent is going to do,

01:02:02.320 | and then have a structure of what the goals are

01:02:06.440 | and whether those predictions actually meet the goals.

01:02:08.440 | And you can go probably pretty far

01:02:10.880 | with that relatively simple setup already.

01:02:13.600 | But to call it a theory of mind,

01:02:15.080 | I don't think you need to.

01:02:16.200 | I mean, it doesn't matter whether the pedestrian has a mind,

01:02:19.280 | it's an object and we can predict what we will do.

01:02:21.840 | And then we can predict what the states will be

01:02:23.720 | in the future and whether they are desirable states.

01:02:26.160 | Stay away from those that are undesirable

01:02:27.960 | and go towards those that are desirable.

01:02:29.720 | So it's a relatively simple, functional approach to that.

01:02:34.520 | Where do we really need the theory of mind?

01:02:37.920 | - Maybe when you start interacting

01:02:40.960 | and you're trying to get the other agent to do something

01:02:44.160 | and jointly, so that you can jointly,

01:02:46.480 | collaboratively achieve something,

01:02:48.400 | then it becomes more complex.

01:02:50.560 | - Well, I mean, even with the pedestrians,

01:02:51.880 | you have to have a sense of where their attention,

01:02:54.800 | actual attention in terms of their gaze is,

01:02:57.840 | but also like, there's this vision science people

01:03:00.760 | talk about this all the time.

01:03:01.600 | Just because I'm looking at it

01:03:02.800 | doesn't mean I'm paying attention to it.

01:03:04.680 | So figuring out what is the person looking at,

01:03:07.400 | what is the sensory information they've taken in?

01:03:09.840 | And the theory of mind piece comes in is,

01:03:12.500 | what are they actually attending to cognitively?

01:03:16.480 | And also, what are they thinking about?

01:03:19.000 | Like, what is the computation they're performing?

01:03:21.200 | And you have probably maybe a few options,

01:03:24.340 | for the pedestrian crossing.

01:03:28.260 | It doesn't have to be, it's like a variable

01:03:30.180 | with a few discrete states,

01:03:31.800 | but you have to have a good estimation

01:03:33.280 | of which of the states that brain is in

01:03:35.480 | for the pedestrian case.

01:03:36.600 | And the same is for attending with a robot.

01:03:39.240 | If you're collaborating to pick up an object,

01:03:41.980 | you have to figure out, is the human,

01:03:44.700 | like, there's a few discrete states

01:03:47.600 | that the human could be in,

01:03:48.520 | and you have to predict that by observing the human.

01:03:52.200 | And that seems like a machine learning problem

01:03:53.960 | to figure out what's the human up to.

01:03:58.960 | It's not as simple as sort of planning,

01:04:02.160 | just because they move their arm

01:04:03.920 | means the arm will continue moving in this direction.

01:04:06.840 | You have to really have a model

01:04:08.560 | of what they're thinking about,

01:04:09.880 | and what's the motivation behind the movement of the arm.

01:04:12.560 | - Here we are talking about

01:04:13.880 | relatively simple physical actions,

01:04:16.560 | but you can take that to higher levels also,

01:04:19.280 | like to predict what the people are going to do,

01:04:21.760 | you need to know what their goals are,

01:04:26.080 | what are they trying to, are they exercising?

01:04:28.000 | Are they just trying to get somewhere?

01:04:29.440 | But even higher level, I mean,

01:04:30.880 | you are predicting what people will do in their career.

01:04:33.900 | What their life themes are.

01:04:35.120 | Do they want to be famous, rich, or do good?

01:04:37.840 | And that takes a lot more information,

01:04:40.560 | but it allows you to then predict their actions,

01:04:43.360 | what choices they might make.

01:04:44.800 | - So how does evolution and computation

01:04:48.600 | apply to the world of neural networks?

01:04:50.800 | 'Cause I've seen quite a bit of work

01:04:52.500 | from you and others in the world of neuroevolution.

01:04:55.520 | So maybe first, can you say, what is this field?

01:04:58.600 | - Yeah, neuroevolution is a combination

01:05:01.140 | of neural networks and evolutionary computation

01:05:04.320 | in many different forms,

01:05:05.480 | but the early versions were simply using evolution

01:05:10.480 | as a way to construct a neural network

01:05:13.960 | instead of say, stochastic gradient descent

01:05:17.240 | or back propagation.

01:05:18.380 | Because evolution can evolve these parameters,

01:05:22.460 | weight values in a neural network,

01:05:23.980 | just like any other string of numbers, you can do that.

01:05:27.180 | And that's useful because some cases you don't have

01:05:30.760 | those targets that you need to back propagate from.

01:05:34.800 | And it might be an agent that's running a maze

01:05:36.900 | or a robot playing a game or something.

01:05:39.800 | You don't, again, you don't know what the right answer is,

01:05:42.080 | you don't have backprop,

01:05:43.080 | but this way you can still evolve a neural net.

01:05:45.860 | And neural networks are really good at these tasks

01:05:49.000 | because they recognize patterns

01:05:51.400 | and they generalize, interpolate between known situations.

01:05:55.280 | So you want to have a neural network in such a task,

01:05:57.720 | even if you don't have the supervised targets.

01:06:00.480 | So that's a reason and that's a solution.

01:06:02.560 | And also more recently now,

01:06:04.200 | when we have all this deep learning literature,

01:06:06.920 | it turns out that we can use evolution

01:06:08.880 | to optimize many aspects of those designs.

01:06:12.480 | The deep learning architectures have become so complex

01:06:16.400 | that there's little hope for us little humans

01:06:18.760 | to understand their complexity

01:06:20.160 | and what actually makes a good design.

01:06:22.880 | And now we can use evolution to give that design for you.

01:06:25.840 | And it might mean optimizing hyperparameters,

01:06:29.800 | like the depth of layers and so on,

01:06:31.840 | or the topology of the network,

01:06:34.680 | how many layers, how they're connected,

01:06:36.560 | but also other aspects like what activation functions

01:06:39.000 | you use where in the network during the learning process,

01:06:42.040 | or what loss function you use,

01:06:43.800 | you could generate that.

01:06:46.520 | Even data augmentation,

01:06:47.680 | all the different aspects of the design

01:06:50.040 | of deep learning experiments could be optimized that way.

01:06:53.840 | So that's an interaction between two mechanisms.

01:06:57.000 | But there's also, when we get more into cognitive science

01:07:00.880 | and the topics that we've been talking about,

01:07:02.640 | you could have learning mechanisms at two level timescales.

01:07:06.200 | So you do have an evolution

01:07:08.000 | that gives you baby neural networks

01:07:10.680 | that then learn during their lifetime.

01:07:12.960 | And you have this interaction of two timescales.

01:07:15.960 | And I think that can potentially be really powerful.

01:07:19.400 | Now in biology, we are not born with all our faculties.

01:07:23.520 | We have to learn, we have a developmental period.

01:07:25.480 | In humans, it's really long.

01:07:27.360 | And most animals have something.

01:07:29.400 | And probably the reason is that evolution,

01:07:32.200 | a DNA is not detailed enough

01:07:34.840 | or plentiful enough to describe them.

01:07:36.760 | We can't describe how to set the brain up.

01:07:38.880 | But we can, evolution can decide on a starting point

01:07:44.920 | and then have a learning algorithm

01:07:47.160 | that will construct the final product.

01:07:50.000 | And this interaction of intelligent,

01:07:54.200 | well, evolution that has produced a good starting point

01:07:57.720 | for the specific purpose of learning from it

01:08:00.800 | with the interaction of, with the environment.

01:08:03.240 | That can be a really powerful mechanism

01:08:04.720 | for constructing brains and constructing behaviors.

01:08:08.040 | - I like how you walk back from intelligence.

01:08:10.040 | So optimize starting point, maybe.

01:08:12.400 | Okay, there's a lot of fascinating things to ask here.

01:08:18.520 | And this is basically this dance between neural networks

01:08:22.080 | and evolution and computation.

01:08:23.880 | Could go into the category of automated machine learning

01:08:26.240 | to where you're optimizing,

01:08:28.840 | whether it's hyperparameters of the topology

01:08:31.040 | or hyperparameters taken broadly.

01:08:33.560 | But the topology thing is really interesting.

01:08:36.400 | I mean, that's not really done that effectively

01:08:40.240 | or throughout the history of machine learning

01:08:41.920 | has not been done.

01:08:43.280 | Usually there's a fixed architecture.

01:08:45.020 | Maybe there's a few components you're playing with.

01:08:47.280 | But to grow a neural network, essentially,

01:08:50.120 | the way you grow in that organism

01:08:51.720 | is really fascinating space.

01:08:52.960 | How hard is it, do you think, to grow a neural network?

01:08:57.960 | And maybe what kind of neural networks

01:09:00.880 | are more amenable to this kind of idea than others?

01:09:04.680 | I've seen quite a bit of work on recurrent neural networks.

01:09:06.960 | Is there some architectures that are friendlier than others?

01:09:10.920 | And is this just a fun, small scale set of experiments

01:09:15.280 | or do you have hope that we can be able to grow

01:09:18.760 | powerful neural networks?

01:09:20.280 | - I think we can.

01:09:21.760 | And most of the work up to now

01:09:24.840 | is taking architectures that already exist,

01:09:27.080 | that humans have designed,

01:09:28.760 | and try to optimize them further.

01:09:30.880 | And you can totally do that.

01:09:32.840 | A few years ago, we did an experiment.

01:09:34.280 | We took a winner of the image captioning competition

01:09:38.480 | and the architecture, and just broke it into pieces

01:09:42.620 | and took the pieces, and that was our search base.

01:09:45.480 | See if you can do better.

01:09:46.680 | And we indeed could, 15% better performance

01:09:49.280 | by just searching around the network design

01:09:52.760 | that humans had come up with,

01:09:54.000 | Oreo vinyls and others.

01:09:55.880 | But that's starting from a point that humans have produced.

01:10:00.840 | But we could do something more general.

01:10:03.480 | It doesn't have to be that kind of network.

01:10:05.840 | The hard part is, there are a couple of challenges.

01:10:08.840 | One of them is to define the search base.

01:10:10.760 | What are your elements and how you put them together?

01:10:14.680 | And the space is just really, really big.

01:10:18.960 | So you have to somehow constrain it

01:10:21.040 | and have some hunch of what will work,

01:10:23.360 | because otherwise everything is possible.

01:10:25.840 | And another challenge is that in order to evaluate

01:10:28.560 | how good your design is, you have to train it.

01:10:32.280 | I mean, you have to actually try it out.

01:10:35.000 | And that's currently very expensive, right?

01:10:37.320 | I mean, deep learning networks may take days to train.

01:10:40.400 | Well, imagine you having a population of 100

01:10:42.280 | and have to run it for 100 generations.

01:10:44.680 | It's not yet quite feasible computationally.

01:10:48.080 | It will be, but also there's a large carbon footprint

01:10:51.640 | and all that.

01:10:52.480 | I mean, we are using a lot of computation for doing it.

01:10:54.360 | So intelligent methods and intelligent,

01:10:57.560 | I mean, we have to do some science in order to figure out

01:11:01.680 | what the right representations are and right operators are,

01:11:05.320 | and how do we evaluate them

01:11:07.360 | without having to fully train them?

01:11:09.240 | And that is where the current research is

01:11:11.440 | and we're making progress on all those fronts.

01:11:13.720 | So yes, there are certain architectures

01:11:17.920 | that are more amenable to that approach.

01:11:20.960 | But also I think we can create our own architecture

01:11:23.600 | and all representations that are even better at that.

01:11:26.280 | - And do you think it's possible to do like

01:11:28.840 | a tiny baby network that grows into something

01:11:31.640 | that can do state-of-the-art

01:11:32.760 | and like even the simple data set like MNIST,

01:11:35.440 | and just like it just grows into a gigantic monster

01:11:39.960 | that's the world's greatest handwriting recognition system?

01:11:42.520 | - Yeah, there are approaches like that.

01:11:44.400 | Esteban Real and Cochlear, for instance,

01:11:46.040 | that worked on evolving a smaller network

01:11:48.560 | and then systematically expanding it to a larger one.

01:11:52.000 | Your elements are already there and scaling it up

01:11:55.040 | will just give you more power.

01:11:56.600 | So again, evolution gives you that starting point

01:11:59.400 | and then there's a mechanism that gives you the final result

01:12:02.880 | and a very powerful approach.

01:12:04.600 | But you could also simulate the actual growth process.

01:12:11.000 | And like I said before, evolving a starting point

01:12:15.360 | and then evolving or training the network.

01:12:18.440 | There's not that much work that's been done on that yet.

01:12:21.960 | We need some kind of a simulation environment

01:12:24.680 | so that interactions at will,

01:12:27.440 | the supervised environment doesn't really,

01:12:29.520 | it's not as easily usable here.

01:12:33.080 | - Sorry, the interaction between neural networks?

01:12:35.560 | - Yeah, the neural networks that you're creating,

01:12:37.320 | interacting the world and learning from these sequences

01:12:42.120 | of interactions, perhaps communication with others.

01:12:44.760 | (laughing)

01:12:46.920 | - That's awesome.

01:12:47.760 | - We would like to get there,

01:12:48.920 | but just the task of simulating something

01:12:51.640 | at that level is very hard.

01:12:53.280 | - It's very difficult.

01:12:54.120 | I love the idea.

01:12:55.400 | I mean, one of the powerful things about evolution on Earth

01:12:58.760 | is the predators and prey emerged.

01:13:01.320 | And like, there's just like,

01:13:03.560 | there's bigger fish and smaller fish

01:13:05.360 | and it's fascinating to think

01:13:07.080 | that you could have neural networks

01:13:08.360 | competing against each other

01:13:09.480 | and one neural network being able to destroy another one.

01:13:12.280 | There's like wars of neural networks

01:13:14.280 | competing to solve the MNIST problem.

01:13:16.560 | I don't know.

01:13:17.400 | - Yeah, yeah.

01:13:18.240 | Oh, totally, yeah, yeah, yeah.

01:13:19.280 | And we actually simulated also that prey

01:13:22.720 | and it was interesting what happened there.

01:13:24.960 | But I mean, Rajak Pallan did this

01:13:26.920 | and Kay Holcomb was a zoologist.

01:13:29.600 | So we had, again,

01:13:31.080 | we had simulated hyenas and simulated zebras.

01:13:37.480 | - Nice.

01:13:38.320 | - And initially, you know,

01:13:40.480 | the hyenas just tried to hunt them.

01:13:42.880 | And when they actually stumbled upon the zebra,

01:13:45.400 | they ate it and were happy.

01:13:47.760 | And then the zebras learned to escape

01:13:51.800 | and the hyenas learned to team up.

01:13:54.360 | And actually two of them approached

01:13:55.720 | in different directions.

01:13:56.960 | And now the zebras, their next step,

01:13:59.040 | they generated a behavior

01:14:01.880 | where they split in different directions,

01:14:03.920 | just like actually gazelles do

01:14:07.360 | when they are being hunted.

01:14:08.400 | They confuse the predator

01:14:09.600 | by going in different directions.

01:14:10.920 | That emerged.

01:14:11.880 | And then more hyenas joined

01:14:14.360 | and kind of circled them.

01:14:15.880 | And then when they circled them,

01:14:18.840 | they could actually herd the zebras together

01:14:21.040 | and eat multiple zebras.

01:14:23.520 | So there was like an arms race

01:14:26.560 | of predators and prey.

01:14:28.360 | And they gradually developed more complex behaviors,

01:14:31.000 | some of which we actually do see in nature.

01:14:33.880 | And this kind of co-evolution,

01:14:36.840 | that's competitive co-evolution,

01:14:38.080 | it's a fascinating topic

01:14:39.600 | because there's a promise or possibility

01:14:42.960 | that you will discover something new

01:14:45.600 | that you don't already know.

01:14:46.480 | You didn't build it in.

01:14:48.160 | It came from this arms race.

01:14:50.760 | It's hard to keep the arms race going.

01:14:52.560 | It's hard to have rich enough simulation

01:14:55.400 | that supports all of these complex behaviors.

01:14:58.320 | But at least for several steps,

01:15:00.080 | we've already seen it in this predator-prey scenario.

01:15:03.680 | - First of all, it's fascinating to think about this context

01:15:06.360 | in terms of evolving architectures.

01:15:09.640 | So I've studied Tesla Autopilot for a long time.

01:15:12.800 | It's one particular implementation

01:15:16.680 | of an AI system that's operating in the real world.

01:15:18.920 | I find it fascinating because of the scale

01:15:21.000 | at which it's used out in the real world.

01:15:23.400 | And I'm not sure if you're familiar with that system much,

01:15:26.240 | but Andrej Karpathy leads that team

01:15:28.600 | on the machine learning side.

01:15:30.120 | And there's a multi-task network, multi-headed network

01:15:34.880 | where there's a core, but it's trained on particular tasks

01:15:38.920 | and there's a bunch of different heads

01:15:40.280 | that are trained on that.

01:15:41.760 | Is there some lessons from evolutionary computation

01:15:46.240 | or neuroevolution that could be applied

01:15:48.360 | to this kind of multi-headed beast

01:15:50.960 | that's operating in the real world?

01:15:52.440 | - Yes, it's a very good problem for neuroevolution.

01:15:55.680 | And the reason is that when you have multiple tasks,

01:16:00.320 | they support each other.

01:16:01.600 | So let's say you're learning to classify X-ray images

01:16:07.480 | to different pathologies.

01:16:09.120 | So you have one task is to classify this disease

01:16:13.440 | and another one, this disease, another one, this one.

01:16:15.520 | And when you're learning from one disease,

01:16:18.040 | that forces certain kinds of internal representations

01:16:21.280 | and embeddings, and they can serve

01:16:24.480 | as a helpful starting point for the other tasks.

01:16:27.240 | So you are combining the wisdom of multiple tasks

01:16:30.560 | into these representations.

01:16:32.000 | And it turns out that you can do better

01:16:33.920 | in each of these tasks when you are learning

01:16:36.120 | simultaneously other tasks than you would by one task alone.

01:16:39.440 | - Which is a fascinating idea in itself, yeah.

01:16:41.320 | - Yes, and people do that all the time.

01:16:43.440 | I mean, you use knowledge of domains that you know

01:16:45.640 | in new domains, and certainly neural networks can do that.

01:16:49.320 | Where neuroevolution comes in is that

01:16:51.880 | what's the best way to combine these tasks?

01:16:54.760 | Now there's architectural design that allow you to decide

01:16:57.760 | where and how the embeddings,

01:17:01.000 | the internal representations are combined

01:17:02.920 | and how much you combine them.

01:17:05.560 | And there's quite a bit of research on that.

01:17:07.600 | And my team, Elliot Mayerson's worked on that in particular,

01:17:11.640 | like what is a good internal representation

01:17:14.440 | that supports multiple tasks?

01:17:16.720 | And we're getting to understand how that's constructed

01:17:20.160 | and what's in it so that it is in a space

01:17:23.680 | that supports multiple different heads, like you said.

01:17:26.800 | And that, I think, is fundamentally

01:17:31.360 | how biological intelligence works as well.

01:17:33.960 | You don't build a representation just for one task.

01:17:37.600 | You try to build something that's general,

01:17:39.680 | not only so that you can do better in one task

01:17:42.320 | or multiple tasks, but also future tasks

01:17:44.640 | and future challenges.

01:17:45.920 | So you learn the structure of the world,

01:17:50.200 | and that helps you in all kinds of future challenges.

01:17:54.040 | - And so you're trying to design a representation

01:17:56.120 | that will support an arbitrary set of tasks

01:17:58.440 | in a particular sort of class of problem.

01:18:01.040 | - Yeah, and also it turns out,

01:18:03.120 | and that's, again, a surprise that Elliot found,

01:18:06.000 | was that those tasks don't have to be very related.

01:18:09.800 | You can learn to do better vision by learning language

01:18:14.120 | or better language by learning about DNA structure.

01:18:17.920 | - No, somehow the world-- - What?

01:18:20.040 | (laughing)

01:18:21.160 | Yeah. - It rhymes.

01:18:22.680 | - The world rhymes, even if it's very disparate fields.

01:18:28.240 | I mean, on that small topic, let me ask you,

01:18:31.440 | 'cause you've also, on the competition neuroscience side,

01:18:34.340 | you worked on both language and vision.

01:18:38.240 | What's the connection between the two?

01:18:44.440 | What's more, maybe there's a bunch of ways to ask this,

01:18:46.840 | but what's more difficult to build

01:18:48.600 | from an engineering perspective

01:18:50.600 | and evolutionary perspective,

01:18:52.320 | the human language system or the human vision system,

01:18:56.080 | or the equivalent of, in the AI space, language and vision,

01:19:00.600 | or is it the best, is the multitask idea

01:19:03.640 | that you're speaking to,

01:19:04.680 | that they need to be deeply integrated?

01:19:07.440 | - Yeah, absolutely the latter.

01:19:09.980 | Learning both at the same time,

01:19:11.640 | I think, is a fascinating direction in the future.

01:19:15.200 | So you have datasets where there's visual component

01:19:17.480 | as well as verbal descriptions, for instance,

01:19:20.040 | and that way you can learn a deeper representation,

01:19:22.760 | a more useful representation for both.

01:19:25.160 | But it's still an interesting question of

01:19:27.160 | which one is easier.

01:19:29.480 | I mean, recognizing objects or even understanding sentences,

01:19:33.200 | that's relatively possible,

01:19:35.800 | but where it becomes, where the challenges are

01:19:37.860 | is to understand the world.

01:19:39.820 | Like the visual world, the 3D,

01:19:42.560 | what are the objects doing

01:19:43.640 | and predicting what will happen, the relationships.

01:19:46.800 | That's what makes vision difficult.

01:19:48.240 | And language, obviously, it's what is being said,

01:19:51.560 | what the meaning is.

01:19:52.760 | And the meaning doesn't stop at who did what to whom.

01:19:57.340 | There are goals and plans and themes,

01:19:59.780 | and eventually you have to understand

01:20:01.760 | the entire human society and history

01:20:04.720 | in order to understand a sentence very much fully.

01:20:08.200 | There are plenty of examples

01:20:09.480 | of those kinds of short sentences

01:20:10.880 | when you bring in all the world knowledge to understand it.

01:20:14.320 | And that's the big challenge.

01:20:15.920 | Now, we are far from that,

01:20:17.360 | but even just bringing in the visual world

01:20:20.640 | together with the sentence

01:20:22.240 | will give you already a lot deeper understanding

01:20:25.640 | of what's happening.

01:20:26.880 | And I think that that's where we're going very soon.

01:20:29.760 | I mean, we've had ImageNet for a long time,

01:20:33.000 | and now we have all these text collections,

01:20:35.960 | but having both together

01:20:38.280 | and then learning a semantic understanding

01:20:41.600 | of what is happening,

01:20:42.800 | I think that will be the next step in the next few years.

01:20:45.400 | - Yeah, you're starting to see that

01:20:46.360 | with all the work with Transformers,

01:20:48.000 | was the AI community started to dip their toe

01:20:51.920 | into this idea of having language models

01:20:56.920 | that are now doing stuff with images, with vision,

01:21:01.360 | and then connecting the two.

01:21:03.920 | I mean, right now it's like these little explorations,

01:21:05.920 | we're literally dipping the toe in.

01:21:07.800 | But maybe at some point we'll just dive into the pool

01:21:11.800 | and it'll just be all seen as the same thing.

01:21:13.880 | I do still wonder what's more fundamental,

01:21:16.880 | whether vision is,

01:21:17.960 | whether we don't think about vision correctly.

01:21:23.320 | Maybe the fact, 'cause we're humans

01:21:24.720 | and we see things as beautiful and so on,

01:21:26.720 | and because we have cameras that take in pixels

01:21:31.040 | as a 2D image,

01:21:33.280 | that we don't sufficiently think about vision as language.

01:21:38.720 | Maybe Chomsky is right all along,

01:21:41.640 | that vision is fundamental to,

01:21:43.760 | sorry, that language is fundamental to everything,

01:21:46.760 | to even cognition, to even consciousness.

01:21:49.280 | The base layer is all language,

01:21:51.400 | not necessarily like English,

01:21:53.040 | but some weird abstract representation,

01:21:57.200 | the linguistic representation.

01:21:59.360 | - Yeah, well, earlier we talked about the social structures

01:22:02.600 | and that may be what's underlying the language,

01:22:05.400 | and that's the more fundamental part,

01:22:06.720 | and then language has been added on top of that.

01:22:08.720 | - Language emerges from the social interaction.

01:22:11.160 | - Yeah, that's a very good guess.

01:22:13.040 | We are visual animals, though.

01:22:15.280 | A lot of the brain is dedicated to vision,

01:22:17.560 | and also when we think about various abstract concepts,

01:22:22.560 | we usually reduce that to vision and images.

01:22:27.760 | And that's, you know, we go to a whiteboard,

01:22:29.760 | you draw pictures of very abstract concepts.

01:22:33.160 | So we tend to resort to that quite a bit,

01:22:35.880 | and that's a fundamental representation.

01:22:37.520 | It's probably possible that it predated language even.

01:22:41.800 | I mean, animals, a lot of, they don't talk,

01:22:43.960 | but they certainly do have vision.

01:22:45.840 | And language is interesting development from mastication,

01:22:51.720 | from eating, you develop an organ

01:22:55.320 | that actually can produce sound to manipulate them.

01:22:57.880 | Maybe that was an accident,

01:22:59.840 | maybe that was something that was available

01:23:01.520 | and then allowed us to do the communication,

01:23:05.680 | or maybe it was gestures.

01:23:07.480 | Sign language could have been the original proto-language.

01:23:10.680 | We don't quite know, but the language is more fundamental

01:23:13.960 | than the medium in which it's communicated.

01:23:17.360 | And I think that it comes from those representations.

01:23:19.960 | Now, in current world, they are so strongly integrated,

01:23:26.600 | it's really hard to say which one is fundamental.

01:23:28.840 | You look at the brain structures and even visual cortex,

01:23:32.840 | which is supposed to be very much just vision.

01:23:35.200 | Well, if you are thinking of semantic concepts,

01:23:38.040 | if you're thinking of language, visual cortex lights up.

01:23:41.560 | It's still useful, even for language computations.

01:23:44.560 | So there are common structures underlying them.

01:23:47.200 | So utilize what you need.

01:23:49.040 | And when you are understanding a scene,

01:23:51.560 | you're understanding relationships,

01:23:53.200 | well, that's not so far from understanding relationships

01:23:55.440 | between words and concepts.

01:23:56.880 | So I think that that's how they are integrated.

01:23:59.160 | - Yeah, and there's dreams, and once we close our eyes,

01:24:02.400 | there's still a world in there somehow operating

01:24:04.480 | and somehow possibly the visual system

01:24:07.720 | somehow integrate into all of it.

01:24:09.920 | I tend to enjoy thinking about aliens

01:24:12.960 | and thinking about the sad thing to me

01:24:17.400 | about extraterrestrial intelligent life,

01:24:21.080 | that if it visited us here on earth,

01:24:24.840 | or if we came on Mars or maybe in other solar system,

01:24:29.120 | another galaxy one day,

01:24:30.960 | that us humans would not be able to detect it

01:24:34.920 | or communicate with it or appreciate,

01:24:37.100 | like it'd be right in front of our nose

01:24:38.760 | and we're too self-obsessed to see it.

01:24:43.380 | Not self-obsessed, but our tools,

01:24:48.380 | our frameworks of thinking would not detect it

01:24:52.560 | as a good movie "Arrival" and so on,

01:24:55.120 | where Stephen Wolfram and his son,

01:24:56.760 | I think were part of developing this alien language

01:24:59.360 | of how aliens would communicate with humans.

01:25:01.560 | Do you ever think about that kind of stuff

01:25:02.960 | where if humans and aliens would be able

01:25:06.840 | to communicate with each other,

01:25:08.800 | like if we met each other at some,

01:25:11.440 | okay, we could do SETI,

01:25:12.520 | which is communicating from across a very big distance,

01:25:16.000 | but also just us, you know,

01:25:17.600 | if you did a podcast with an alien,

01:25:21.320 | do you think we'd be able to find a common language

01:25:25.440 | and a common methodology of communication?

01:25:28.440 | I think from a computational perspective,

01:25:30.880 | the way to ask that is you have

01:25:32.440 | very fundamentally different creatures,

01:25:34.320 | agents that are created,

01:25:35.440 | would they be able to find a common language?

01:25:38.480 | - Yes, I do think about that.

01:25:41.000 | I mean, I think a lot of people who are in computing,

01:25:43.000 | they, and AI in particular,

01:25:45.360 | they got into it because they were fascinated

01:25:47.320 | with science fiction and all of these options.

01:25:50.740 | I mean, Star Trek generated all kinds of devices

01:25:54.040 | that we have now, they envisioned it first,

01:25:56.560 | and it's a great motivator to think about things like that.

01:26:00.700 | And I, so one, and again, being a computational scientist

01:26:06.880 | and trying to build intelligent agents,

01:26:11.280 | what I would like to do is have a simulation

01:26:14.520 | where the agents actually evolve communication,

01:26:18.400 | not just communication, we've done that,

01:26:19.960 | people have done that many times,

01:26:21.280 | that they communicate, they signal and so on,

01:26:23.840 | but actually develop a language,

01:26:26.080 | and language means grammar,

01:26:27.200 | it means all this social structures

01:26:28.880 | and on top of that, grammatical structures.

01:26:31.860 | And we do it under various conditions

01:26:36.000 | and actually try to identify what conditions

01:26:37.780 | are necessary for it to come out.

01:26:40.880 | And then we can start asking that kind of questions.

01:26:44.240 | Are those languages that emerge

01:26:46.280 | in those different simulated environments,

01:26:48.880 | are they understandable to us?

01:26:50.800 | Can we somehow make a translation?

01:26:53.560 | We can make it a concrete question.

01:26:55.920 | - So machine translation of evolved languages,

01:26:59.000 | and so like languages that evolve come up with,

01:27:02.040 | can we translate, like I have a Google Translate

01:27:04.960 | for the evolved languages.

01:27:07.200 | - Yes, and if we do that enough,

01:27:09.800 | we have perhaps an idea what an alien language

01:27:14.120 | might be like, the space of where those languages can be.

01:27:17.200 | 'Cause we can set up their environment differently.

01:27:20.000 | It doesn't need to be gravity.

01:27:22.040 | You can have all kinds of, societies can be different,

01:27:24.880 | they may have no predators,

01:27:26.320 | they may have all, everybody's a predator,

01:27:28.520 | all kinds of situations.

01:27:30.120 | And then see what the space possibly is

01:27:32.880 | where those languages are and what the difficulties are.

01:27:35.920 | That'd be really good actually to do that

01:27:37.680 | before the aliens come here.

01:27:39.520 | - Yes, it's good practice.

01:27:40.920 | On the similar connection,

01:27:44.140 | you can think of AI systems as aliens.

01:27:48.240 | Is there a ways to evolve a communication scheme for,

01:27:51.920 | there's a field you can call it like explainable AI,

01:27:55.080 | for AI systems to be able to communicate.

01:27:59.000 | So you evolve a bunch of agents,

01:28:01.640 | but for some of them to be able to talk to you also.

01:28:05.440 | So to evolve a way for agents to be able to communicate

01:28:08.520 | about their world to us humans.

01:28:11.080 | Do you think that there's possible mechanisms

01:28:13.440 | for doing that?

01:28:14.760 | - We can certainly try.

01:28:16.280 | And if it's an evolution competition system,

01:28:20.600 | for instance, you reward those solutions

01:28:22.640 | that are actually functional,

01:28:24.160 | that that communication makes sense,

01:28:25.640 | it allows us to together again, achieve common goals.

01:28:29.500 | I think that's possible.

01:28:30.940 | But even from that paper that you mentioned,

01:28:35.160 | the anecdotes, it's quite likely also

01:28:37.880 | that the agents learn to lie and fake

01:28:42.880 | and do all kinds of things like that.

01:28:45.360 | I mean, we see that in even very low level,

01:28:47.760 | like bacterial evolution, they are cheaters.

01:28:51.800 | And who's to say that what they say

01:28:53.920 | is actually what they think.

01:28:55.400 | But that's one thing that there would have to be

01:28:59.400 | some common goal so that we can evaluate

01:29:01.920 | whether that communication is at least useful.

01:29:04.220 | They may be saying things just to make us feel good

01:29:09.040 | or get us to do what we want,

01:29:10.680 | but it would not turn them off or something.

01:29:12.440 | But so we would have to understand

01:29:15.160 | their internal representations much better

01:29:16.760 | to really make sure that that translation is critical.

01:29:20.120 | But it can be useful.

01:29:21.400 | And I think it's possible to do that.

01:29:24.000 | There are examples where visualizations

01:29:27.680 | are automatically created

01:29:30.000 | so that we can look into the system

01:29:33.600 | and that language is not that far from it.

01:29:35.880 | I mean, it is a way of communicating

01:29:37.440 | and logging what you're doing in some interpretable way.

01:29:41.480 | I think a fascinating topic, yeah, to do that.

01:29:46.600 | - You're making me realize

01:29:47.720 | that it's a good scientific question

01:29:51.080 | whether lying is an effective mechanism

01:29:54.440 | for integrating yourself and succeeding in a social network

01:29:57.320 | in a world that is social.

01:30:00.360 | I tend to believe that honesty and love

01:30:04.540 | are evolutionary advantages in an environment

01:30:09.540 | where there's a network of intelligent agents.

01:30:12.620 | But it's also very possible that dishonesty

01:30:14.840 | and manipulation and even violence,

01:30:19.840 | all those kinds of things might be more beneficial.

01:30:23.100 | That's the old open question about good versus evil.

01:30:25.900 | But I tend to, I mean, I don't know if it's a hopeful,

01:30:29.240 | maybe I'm delusional, but it feels like karma is a thing,

01:30:34.240 | which is like long-term the agents

01:30:39.560 | that are just kind to others,

01:30:41.000 | sometimes for no reason, will do better.

01:30:43.780 | In a society that's not highly constrained on resources.

01:30:47.560 | So like people start getting weird and evil

01:30:50.520 | towards each other and bad

01:30:51.880 | when the resources are very low

01:30:54.200 | relative to the needs of the populace,

01:30:56.960 | especially at the basic level,

01:30:58.320 | like survival, shelter, food, all those kinds of things.

01:31:02.640 | But I tend to believe

01:31:06.200 | that once you have those things established,

01:31:09.200 | then, well, not to believe, I guess I hope

01:31:12.680 | that AI systems will be honest.

01:31:14.900 | But it's scary to think about the Turing test.

01:31:19.900 | AI systems that will eventually pass the Turing test

01:31:23.940 | will be ones that are exceptionally good at lying.

01:31:26.740 | That's a terrifying concept.

01:31:28.380 | I mean, I don't know.

01:31:31.260 | First of all, so from somebody who studied language

01:31:34.220 | and obviously are not just a world expert in AI,

01:31:37.860 | but somebody who dreams about the future of the field,

01:31:41.540 | do you hope, do you think there'll be human level

01:31:45.640 | or superhuman level intelligences in the future

01:31:48.700 | that we eventually build?

01:31:51.240 | - Well, I definitely hope that we can get there.

01:31:56.180 | One, I think, important perspective

01:31:59.240 | is that we are building AI to help us.

01:32:02.240 | That it is a tool like cars

01:32:04.640 | or language or communication.

01:32:10.280 | AI will help us be more productive.

01:32:12.760 | And that is always a condition.

01:32:17.560 | It's not something that we build and let run

01:32:20.340 | and it becomes an entity of its own

01:32:22.500 | that doesn't care about us.

01:32:23.840 | Now, of course, really far into the future,

01:32:27.360 | maybe that might be possible,

01:32:28.780 | but not in the foreseeable future when we are building it.

01:32:32.200 | And therefore we are always in a position

01:32:34.360 | of limiting what it can or cannot do.

01:32:38.840 | And your point about lying is very interesting.

01:32:43.840 | Even in these hyena societies, for instance,

01:32:49.320 | when a number of these hyenas band together

01:32:52.640 | and they take a risk and steal the kill,

01:32:56.240 | there are always hyenas that hang back

01:32:58.560 | and don't participate in that risky behavior,

01:33:02.040 | but they walk in later and join the party after the kill.

01:33:06.960 | And there are even some that may be ineffective

01:33:10.000 | and cause others to have harm.

01:33:12.880 | So, and like I said, even bacteria cheat.

01:33:15.440 | And we see in biology,

01:33:17.320 | there's always some element on opportunity.

01:33:20.520 | If you have a, I think that is this,

01:33:22.160 | because if you have a society,

01:33:24.160 | in order for society to be effective,

01:33:26.040 | you have to have this cooperation

01:33:27.560 | and you have to have trust.

01:33:29.020 | And if you have enough of agents

01:33:32.100 | who are able to trust each other,

01:33:33.960 | you can achieve a lot more.

01:33:36.560 | But if you have trust,

01:33:37.480 | you also have opportunity for cheaters and liars.

01:33:40.600 | And I don't think that's ever gonna go away.

01:33:43.600 | There will be hopefully a minority

01:33:45.200 | so that they don't get in the way.

01:33:46.640 | And we studied in these hyena simulations,

01:33:48.720 | like what the proportion needs to be

01:33:50.480 | before it's no longer functional.

01:33:52.640 | And you can point out that you can tolerate

01:33:55.040 | a few cheaters and a few liars,

01:33:57.240 | and the society can still function.

01:33:59.640 | And that's probably going to happen

01:34:02.280 | when we build these systems that autonomously learn

01:34:06.360 | that the really successful ones are honest

01:34:09.320 | because that's the best way of getting things done.

01:34:12.040 | But there probably are also intelligent agents

01:34:16.000 | that find that they can achieve their goals

01:34:18.040 | by bending the rules of cheating.

01:34:20.960 | - So there could be a huge benefit to,

01:34:22.920 | as opposed to having fixed AI systems,

01:34:25.720 | say we build an AGI system and deploying millions of them,

01:34:30.080 | it'd be that are exactly the same.

01:34:32.500 | There might be a huge benefit to

01:34:36.320 | introducing sort of from like

01:34:37.840 | an evolution competition perspective, a lot of variation.

01:34:41.280 | Sort of like diversity in all its forms is beneficial

01:34:46.280 | even if some people are assholes or some robots are assholes.

01:34:49.920 | So like it's beneficial to have that

01:34:51.920 | because you can't always a priori know

01:34:56.720 | what's good, what's bad.

01:34:58.440 | But that's a fascinating--

01:35:01.360 | - Absolutely.

01:35:02.240 | Diversity is the bread and butter.

01:35:04.320 | I mean, if you're running a competition,

01:35:05.760 | you see diversity is the one fundamental thing

01:35:08.040 | you have to have.

01:35:09.040 | And absolutely, also, it's not always good diversity.

01:35:12.640 | It may be something that can be destructive.

01:35:14.920 | We had in this hyena simulations,

01:35:16.320 | we have hyenas that just are suicidal.

01:35:19.160 | They just run and get killed.

01:35:20.520 | But they form the basis of those

01:35:22.760 | who actually are really fast,

01:35:24.400 | but stop before they get killed

01:35:26.040 | and eventually turn into this mob.

01:35:28.360 | So there might be something useful there

01:35:30.000 | if it's recombined with something else.

01:35:32.160 | So I think that as long as we can tolerate some of that,

01:35:34.960 | it may turn into something better.

01:35:36.840 | You may change the rules

01:35:38.480 | because it's so much more efficient to do something

01:35:40.640 | that was actually against the rules before.

01:35:43.280 | And we've seen society change over time

01:35:46.480 | quite a bit along those lines.

01:35:47.760 | That there were rules in society

01:35:49.920 | that we don't believe are fair anymore,

01:35:52.160 | even though they were considered proper behavior before.

01:35:57.160 | So things are changing.

01:35:58.760 | And I think that in that sense,

01:36:00.000 | I think it's a good idea to be able to tolerate some of that,

01:36:03.960 | some of that cheating,

01:36:05.080 | because eventually we might turn into something better.

01:36:07.480 | - So yeah, I think this is a message to the trolls

01:36:10.040 | and the assholes of the internet

01:36:11.400 | that you too have a beautiful purpose

01:36:13.480 | in this human ecosystem.

01:36:15.640 | So I appreciate you very much.

01:36:17.360 | - In moderate quantities.

01:36:18.200 | - In moderate quantities.

01:36:20.360 | So there's a whole field of artificial life.

01:36:23.120 | I don't know if you're connected to this field,

01:36:24.880 | if you pay attention.

01:36:26.600 | Do you think about this kind of thing?

01:36:29.560 | Is there an impressive demonstration to you

01:36:32.240 | of artificial life?

01:36:33.160 | Do you think of the agency you work with

01:36:35.320 | in the evolutionary computation perspective as life?

01:36:40.320 | And where do you think this is headed?

01:36:43.600 | Like, is there interesting systems

01:36:45.080 | that we'll be creating more and more

01:36:47.040 | that make us redefine, maybe rethink

01:36:50.760 | about the nature of life?

01:36:52.440 | - Different levels of definition and goals there.

01:36:57.000 | I mean, at some level artificial life

01:36:59.760 | can be considered multi-agent systems

01:37:02.360 | that build a society that again, achieves a goal.

01:37:05.080 | And it might be robots that go into a building

01:37:07.000 | and clean it up or after an earthquake or something.

01:37:10.360 | You can think of that as an artificial life problem

01:37:12.880 | in some sense.

01:37:14.600 | Or you can really think of it, artificial life,

01:37:16.880 | as a simulation of life

01:37:19.120 | and a tool to understand what life is

01:37:22.520 | and how life evolved on earth.

01:37:24.640 | And like I said, in artificial life conference,

01:37:26.840 | there are branches of that conference sessions

01:37:29.760 | of people who really worry about molecular designs

01:37:33.480 | and the start of life.

01:37:34.960 | Like I said, primordial soup

01:37:36.760 | where eventually you get something self-replicating

01:37:39.720 | and they're really trying to build that.

01:37:41.960 | So it's a whole range of topics.

01:37:44.840 | And I think that artificial life is a great tool

01:37:50.840 | to understand life.

01:37:53.000 | And there are questions like sustainability,

01:37:56.440 | species, we're losing species.

01:37:59.280 | How bad is it?

01:38:00.880 | Is it natural?

01:38:02.560 | Is there a tipping point?

01:38:04.000 | And where are we going?

01:38:06.480 | I mean, like the hyena evolution,

01:38:08.080 | we may have understood that there's a pivotal point

01:38:11.360 | in their evolution.

01:38:12.200 | They discovered cooperation and coordination.

01:38:14.880 | Artificial life simulations can identify that

01:38:18.680 | and maybe encourage things like that.

01:38:22.920 | And also societies can be seen as a form of life itself.

01:38:27.920 | I mean, we're not talking about biological evolution,

01:38:29.840 | we have all evolution of societies.

01:38:31.920 | Maybe some of the same phenomena emerge in that domain

01:38:36.520 | and having artificial life simulations

01:38:39.440 | and understanding could help us build better societies.

01:38:42.560 | - Yeah, and thinking from a meme perspective

01:38:45.800 | from Richard Dawkins,

01:38:50.880 | that maybe the organisms, ideas of the organisms,

01:38:54.080 | not the humans in these societies,

01:38:56.680 | it's almost like reframing what is exactly evolving.

01:39:01.920 | Maybe the interesting,

01:39:02.920 | the humans aren't the interesting thing

01:39:04.560 | as the contents of our minds is the interesting thing.

01:39:07.360 | And that's what's multiplying.

01:39:09.240 | And that's actually multiplying and evolving

01:39:10.880 | in a much faster timescale.

01:39:13.040 | And that maybe has more power on the trajectory

01:39:16.240 | of life on earth than does biological evolution.

01:39:19.560 | - Yes. - Is evolution of these ideas.

01:39:21.000 | - Yes, and it's fascinating, like I said before,

01:39:23.880 | that we can keep up somehow biologically.

01:39:27.560 | We evolve to a point where we can keep up

01:39:30.120 | with this meme evolution, literature, internet.

01:39:35.120 | We understand DNA and we understand fundamental particles.

01:39:39.040 | We didn't start that way a thousand years ago

01:39:41.400 | and we haven't evolved biologically very much,

01:39:43.360 | but somehow our minds are able to extend.

01:39:47.040 | And therefore AI can be seen also as one such step

01:39:51.280 | that we created and it's our tool.

01:39:53.440 | And it's part of that meme evolution that we created,

01:39:56.360 | even if our biological evolution does not progress as fast.

01:39:59.640 | - And us humans might only be able to understand so much.

01:40:03.720 | We're keeping up so far,

01:40:05.800 | or we think we're keeping up so far,

01:40:07.360 | but we might need AI systems to understand.

01:40:09.520 | Maybe like the physics of the universe is operating,

01:40:13.800 | look at string theory,

01:40:14.760 | maybe it's operating in much higher dimensions.

01:40:17.440 | Maybe we're totally, because of our cognitive limitations,

01:40:21.240 | are not able to truly internalize the way this world works.

01:40:25.720 | And so we're running up against the limitation

01:40:28.920 | of our own minds and we have to create

01:40:30.920 | these next level organisms like AI systems

01:40:34.520 | that would be able to understand much deeper,

01:40:36.280 | like really understand what it means to live

01:40:38.480 | in a multidimensional world

01:40:41.240 | that's outside of the four dimensions,

01:40:42.640 | the three of space and one of time.

01:40:45.400 | - Translation, and generally we can deal with the world,

01:40:48.200 | even if you don't understand all the details,

01:40:49.680 | we can use computers, even though we don't,

01:40:52.120 | most of us don't know all the structures underneath

01:40:55.040 | or drive a car.

01:40:55.880 | I mean, there are many components,

01:40:57.280 | especially new cars that you don't quite fully know,

01:40:59.880 | but you have the interface, you have an abstraction of it

01:41:02.720 | that allows you to operate it and utilize it.

01:41:05.120 | And I think that that's perfectly adequate

01:41:08.200 | and we can build on it.

01:41:09.240 | And AI can play a similar role.

01:41:12.160 | - I have to ask about beautiful artificial life systems

01:41:18.120 | or evolution computation systems, cellular automata to me.

01:41:22.640 | Like I remember it was a game changer for me early on

01:41:26.200 | in life when I saw Conway's Game of Life

01:41:28.800 | who recently passed away, unfortunately.

01:41:31.400 | It's beautiful how much complexity can emerge

01:41:38.640 | from such simple rules.

01:41:40.000 | I just don't, somehow that simplicity

01:41:44.440 | is such a powerful illustration and also humbling

01:41:48.320 | because it feels like I personally, from my perspective,

01:41:50.880 | understand almost nothing about this world

01:41:54.920 | because like my intuition fails completely

01:41:58.400 | how complexity can emerge from such simplicity.

01:42:01.280 | Like my intuition fails, I think,

01:42:02.640 | is the biggest problem I have.

01:42:04.140 | Do you find systems like that beautiful?

01:42:08.520 | Is there, do you think about cellular automata?

01:42:11.400 | Because cellular automata don't really have,

01:42:14.060 | and many other artificial life systems

01:42:17.200 | don't necessarily have an objective.

01:42:18.920 | Maybe that's a wrong way to say it.

01:42:21.640 | It's almost like it's just evolving and creating.

01:42:26.640 | And there's not even a good definition

01:42:29.760 | of what it means to create something complex

01:42:33.080 | and interesting and surprising,

01:42:34.600 | all those words that you said.

01:42:38.000 | Is there some of those systems that you find beautiful?

01:42:41.080 | - Yeah, yeah.

01:42:41.920 | And similarly, evolution does not have a goal.

01:42:44.460 | It is responding to the current situation

01:42:49.520 | and survival then creates more complexity

01:42:52.720 | and therefore we have something that we perceive as progress

01:42:56.080 | but that's not what evolution is inherently set to do.

01:43:00.680 | And yeah, that's really fascinating

01:43:03.800 | how a simple set of rules or simple mappings can,

01:43:08.800 | how from such simple mappings, complexity can emerge.

01:43:14.520 | So it's a question of emergence and self-organization.

01:43:17.720 | And the game of life is one of the simplest ones

01:43:21.500 | and very visual and therefore it drives home the point

01:43:25.680 | that it's possible that nonlinear interactions

01:43:31.240 | and this kinds of complexity can emerge from them.

01:43:34.720 | And biology and evolution is along the same lines.

01:43:37.920 | We have simple representations.

01:43:40.080 | DNA, if you really think of it, it's not that complex.

01:43:43.160 | It's a long sequence of them.

01:43:45.400 | There's lots of them

01:43:46.240 | but it's a very simple representation.

01:43:48.200 | And similar with evolutionary computation,

01:43:49.880 | whatever string or tree representation we have

01:43:52.640 | and the operations, the amount of code that's required

01:43:57.560 | to manipulate those, it's really, really little.

01:44:00.520 | And of course, game of life, even less.

01:44:02.440 | So how complexity emerges from such simple principles,

01:44:06.200 | that's absolutely fascinating.

01:44:08.280 | The challenge is to be able to control it

01:44:11.440 | and guide it and direct it so that it becomes useful.

01:44:15.520 | And like game of life is fascinating to look at

01:44:17.920 | and evolution, all the forms that come out is fascinating

01:44:21.200 | but can we actually make it useful for us?

01:44:24.040 | - And efficient because if you actually think about

01:44:27.040 | each of the cells in the game of life as a living organism,

01:44:30.320 | there's a lot of death that has to happen

01:44:32.600 | to create anything interesting.

01:44:34.360 | And so I guess the question is for us humans

01:44:36.480 | that are mortal and then life ends quickly,

01:44:38.920 | we wanna kinda hurry up and make sure we take evolution,

01:44:43.920 | the trajectory that is a little bit more efficient

01:44:47.400 | than the alternatives.

01:44:49.360 | - And that touches upon something we talked about earlier

01:44:51.240 | that evolutionary computation is very impatient.

01:44:54.600 | We have a goal, we want it right away

01:44:57.160 | versus biology has a lot of time and deep time

01:45:01.080 | and weak pressure and large populations.

01:45:04.520 | One great example of this is the novelty search.

01:45:08.960 | So evolutionary computation where you don't actually

01:45:12.440 | specify a fitness goal, something that is your actual thing

01:45:16.520 | that you want, but you just reward solutions

01:45:19.640 | that are different from what you've seen before.

01:45:22.520 | Nothing else.

01:45:23.800 | And you know what?

01:45:25.160 | You actually discover things that are interesting

01:45:27.400 | and useful that way.

01:45:28.520 | Ken Stanley and Joel Lehman did this one study

01:45:31.120 | where they actually tried to evolve

01:45:33.600 | walking behavior on robots.

01:45:35.320 | And that's actually, we talked about earlier

01:45:36.600 | where your robot actually failed in all kinds of ways

01:45:39.680 | and eventually discovered something

01:45:41.040 | that was a very efficient walk.

01:45:43.920 | And it was because they rewarded things

01:45:47.720 | that were different that you were able to discover something

01:45:50.720 | and I think that this is crucial

01:45:53.000 | because in order to be really different

01:45:55.120 | from what you already have,

01:45:56.600 | you have to utilize what is there in a domain

01:45:59.080 | to create something really different.

01:46:00.760 | So you have encoded the fundamentals of your world

01:46:05.760 | and then you make changes to those fundamentals

01:46:08.120 | you get further away.

01:46:09.720 | So that's probably what's happening

01:46:11.520 | in these systems of emergence,

01:46:14.280 | that the fundamentals are there.

01:46:17.360 | And when you follow those fundamentals,

01:46:18.960 | you get into points and some of those

01:46:21.080 | are actually interesting and useful.

01:46:23.200 | Even in that robotic walker simulation,

01:46:25.160 | there was a large set of garbage,

01:46:28.120 | but among them, there were some of these gems.

01:46:31.600 | And then those are the ones that somehow

01:46:33.160 | you have to outside recognize and make useful.

01:46:36.600 | But this kind of productive systems,

01:46:38.640 | if you code them the right kind of principles,

01:46:41.560 | I think that they encode the structure of the domain,

01:46:45.640 | then you will get to these solutions and you discover it.

01:46:48.480 | - It feels like that might also be a good way to live life.

01:46:52.760 | So let me ask, do you have advice for young people today

01:46:57.760 | about how to live life or how to succeed in their career

01:47:01.520 | or forget career, just succeed in life?

01:47:04.680 | From an evolutionary computation perspective.

01:47:08.800 | - Yes, yes, definitely.

01:47:11.560 | Explore, diversity, exploration.

01:47:15.880 | And individuals take classes in music,

01:47:19.840 | history, philosophy, math, engineering.

01:47:23.800 | See connections between them.

01:47:27.480 | Travel, learn a language.

01:47:30.120 | I mean, all this diversity is fascinating

01:47:32.160 | and we have it at our fingertips today.

01:47:35.480 | It's possible, you have to make a bit of an effort

01:47:37.840 | 'cause it's not easy, but the rewards are wonderful.

01:47:42.000 | - Yeah, there's something interesting

01:47:43.840 | about an objective function of new experiences.

01:47:47.400 | So try to figure out, I mean,

01:47:49.400 | what is the maximally new experience I could have today?

01:47:55.560 | And that sort of, that novelty, optimizing for novelty

01:47:59.360 | for some period of time might be a very interesting way

01:48:01.800 | to sort of maximally expand the sets of experiences

01:48:06.320 | you had and then ground from that perspective,

01:48:10.360 | like what will be the most fulfilling trajectory

01:48:14.520 | through life?

01:48:15.360 | And of course, the flip side of that is where I come from.

01:48:19.200 | Again, maybe Russian, I don't know.

01:48:21.000 | But the choice has a detrimental effect, I think,

01:48:26.000 | at least from my mind,

01:48:31.080 | where scarcity has an empowering effect.

01:48:35.280 | So if I have very little of something

01:48:40.280 | and only one of that something, I will appreciate it deeply

01:48:45.000 | until I came to Texas recently

01:48:48.640 | and I've been pigging out on delicious, incredible meat.

01:48:51.560 | I've been fasting a lot, so I need to do that again.

01:48:53.840 | But when you fast for a few days,

01:48:56.320 | that the first taste of a food is incredible.

01:49:00.760 | So the downside of exploration is that somehow,

01:49:05.760 | maybe you can correct me,

01:49:11.080 | but somehow you don't get to experience deeply

01:49:15.120 | any one of the particular moments.

01:49:17.520 | But that could be a psychology thing.

01:49:19.760 | That could be just a very human, peculiar flaw.

01:49:23.680 | - Yeah, I didn't mean that you superficially explore.

01:49:26.720 | I mean, you can-- - Explore deeply.

01:49:28.360 | - Yeah, so you don't have to explore 100 things,

01:49:31.080 | but maybe a few topics where you can take

01:49:34.080 | a deep enough dive that you gain an understanding.

01:49:39.080 | Yourself have to decide at some point

01:49:42.640 | that this is deep enough.

01:49:44.440 | And I've obtained what I can from this topic

01:49:49.280 | and now it's time to move on.

01:49:51.400 | And that might take years.

01:49:54.040 | People sometimes switch careers

01:49:56.280 | and they may stay on some career for a decade

01:49:59.160 | and switch to another one.

01:50:00.520 | You can do it.

01:50:01.840 | You're not pretty determined to stay where you are.

01:50:04.680 | But in order to achieve something,

01:50:09.120 | 10,000 hours makes, you need 10,000 hours

01:50:11.760 | to become an expert on something.

01:50:13.640 | So you don't have to become an expert,

01:50:15.360 | but to even develop an understanding

01:50:17.200 | and gain the experience that you can use later,

01:50:19.360 | you probably have to spend, like I said, it's not easy.

01:50:21.960 | You gotta spend some effort on it.

01:50:24.440 | Now, also at some point then when you have this diversity

01:50:28.160 | and you have these experiences, exploration,

01:50:30.360 | you may want to, you may find something

01:50:33.720 | that you can't stay away from.

01:50:35.920 | Like for, as it was computers, it was AI,

01:50:38.720 | it was, that you, I just have to do it.

01:50:42.000 | And I, and then it will take decades maybe

01:50:45.240 | and you are pursuing it because you figured out

01:50:47.800 | that this is really exciting

01:50:49.320 | and you can bring in your experiences.

01:50:51.280 | And there's nothing wrong with that either,

01:50:52.760 | but you asked what's the advice for young people.

01:50:55.840 | That's the exploration part.

01:50:57.520 | And then beyond that, after that exploration,

01:51:00.120 | you actually can focus and build a career.

01:51:03.200 | And even there you can switch multiple times,

01:51:05.800 | but I think that diversity exploration is fundamental

01:51:09.120 | to having a successful career as is concentration

01:51:13.320 | and spending an effort where it matters.

01:51:15.520 | And, but you are in better position to make the choice

01:51:18.960 | when you have done your homework.

01:51:20.400 | - Explored.

01:51:21.240 | So exploration precedes commitment, but both are beautiful.

01:51:24.920 | So again, from an evolutionary computation perspective,

01:51:29.440 | we'll look at all the agents that had to die

01:51:32.440 | in order to come up with different solutions in simulation.

01:51:35.780 | What do you think from that individual agent's perspective

01:51:40.280 | is the meaning of it all?

01:51:41.880 | So far as humans, you're just one agent

01:51:43.880 | who's going to be dead, unfortunately, one day too soon.

01:51:47.600 | What do you think is the why of why that agent came to be

01:51:53.760 | and eventually will be no more?

01:51:57.560 | Is there a meaning to it all?

01:52:00.100 | - Yeah, in evolution, there is meaning.

01:52:02.480 | Everything is a potential direction.

01:52:05.640 | Everything is a potential stepping stone.

01:52:07.820 | Not all of them are gonna work out.

01:52:11.400 | Some of them are foundations for further improvement.

01:52:16.400 | And even those that are perhaps going to die out

01:52:20.240 | where potential energy is potential solutions.

01:52:24.680 | In biology, we see a lot of species die off naturally

01:52:28.720 | and like the dinosaurs.

01:52:29.880 | I mean, they were really good solution for a while,

01:52:31.880 | but then it didn't turn out to be not such a good solution

01:52:36.020 | in the longterm.

01:52:37.800 | When there's an environmental change,

01:52:39.440 | you have to have diversity.

01:52:40.660 | Some other solutions become better.

01:52:42.680 | Doesn't mean that there was an attempt.

01:52:45.040 | It didn't quite work out or last,

01:52:47.560 | but there are still dinosaurs and mamas,

01:52:49.360 | at least their relatives,

01:52:51.240 | and they may one day again be useful.

01:52:54.280 | Who knows?

01:52:55.560 | So from an individual's perspective,

01:52:57.200 | you've got to think of a bigger picture

01:52:59.100 | that it is a huge engine that is innovative.

01:53:04.100 | And these elements are all part of it,

01:53:06.760 | potential innovations on their own

01:53:09.360 | and also as raw material perhaps

01:53:12.360 | or stepping stones for other things that could come after.

01:53:16.400 | - But it still feels from an individual perspective

01:53:18.760 | that I matter a lot.

01:53:21.080 | But even if I'm just a little cog in the giant machine,

01:53:24.520 | is that just a silly human notion

01:53:28.200 | in an individualistic society

01:53:29.980 | and they should let go of that?

01:53:31.540 | Do you find beauty in being part of the giant machine?

01:53:35.800 | - Yeah, I think it's meaningful.

01:53:39.060 | I think it adds purpose to your life,

01:53:41.580 | that you are part of something bigger.

01:53:43.600 | - That said, do you ponder your individual agent's mortality?

01:53:50.420 | Do you think about death?

01:53:53.780 | Do you fear death?

01:53:56.760 | - Well, certainly more now than when I was a youngster

01:54:00.740 | and did skydiving and paragliding and all these things.

01:54:05.740 | - You've become wiser.

01:54:06.880 | - There is a reason for this life arc

01:54:13.980 | that younger folks are more fearless in many ways.

01:54:17.220 | It's part of the exploration.

01:54:18.860 | They are the individuals who think,

01:54:22.220 | hmm, I wonder what's over those mountains

01:54:24.900 | or what if I go really far in that ocean?

01:54:27.100 | What would I find?

01:54:28.040 | I mean, older folks don't necessarily think that way,

01:54:32.240 | but younger do and it's kind of counterintuitive.

01:54:34.900 | So yeah, but logically it's like,

01:54:39.200 | you have a limited amount of time,

01:54:40.160 | what can you do with it that matters?

01:54:42.480 | So you try to, you have done your exploration,

01:54:45.400 | you committed to a certain direction

01:54:48.200 | and you become an expert perhaps in it.

01:54:50.440 | What can I do that matters

01:54:52.600 | with the limited resources that I have?

01:54:55.640 | That's how I think a lot of people, myself included,

01:54:59.840 | start thinking later on in their career.

01:55:02.520 | - And like you said, leave a bit of a trace

01:55:05.700 | and a bit of an impact even after the agent is gone.

01:55:08.620 | - Yeah, that's the goal.

01:55:10.240 | - Well, this was a fascinating conversation.

01:55:13.720 | I don't think there's a better way to end it.

01:55:16.000 | Thank you so much.

01:55:17.120 | So first of all, I'm very inspired

01:55:19.520 | of how vibrant the community at UT Austin and Austin is.

01:55:23.040 | It's really exciting for me to see it.

01:55:25.640 | And this whole field seems like profound philosophically,

01:55:30.040 | but also the path forward

01:55:31.400 | for the artificial intelligence community.

01:55:33.400 | So thank you so much for explaining

01:55:35.480 | so many cool things to me today

01:55:36.920 | and for wasting all of your valuable time with me.

01:55:39.280 | - Oh, it was a pleasure.

01:55:40.520 | Thanks, Lex. - I appreciate it.

01:55:42.920 | Thanks for listening to this conversation

01:55:44.560 | with Vristo McAlinan

01:55:46.040 | and thank you to the Jordan and Harbinger Show,

01:55:48.760 | Grammarly, Belcampo, and Indeed.

01:55:52.080 | Check them out in the description to support this podcast.

01:55:55.640 | And now let me leave you with some words from Carl Sagan.

01:55:59.400 | "Extinction is the rule.

01:56:01.800 | Survival is the exception."

01:56:05.000 | Thank you for listening.

01:56:06.120 | I hope to see you next time.

01:56:07.960 | (upbeat music)

01:56:10.560 | (upbeat music)

01:56:13.160 | [BLANK_AUDIO]

Risto Miikkulainen: Neuroevolution and Evolutionary Computation | Lex Fridman Podcast #177

Chapters