Matt Botvinick: Neuroscience, Psychology, and AI at DeepMind

00:00:00.000 | The following is a conversation with Matt Botmanick,

00:00:03.440 | Director of Neuroscience Research at DeepMind.

00:00:06.840 | He's a brilliant, cross-disciplinary mind

00:00:09.360 | navigating effortlessly between cognitive psychology,

00:00:12.480 | computational neuroscience, and artificial intelligence.

00:00:16.780 | Quick summary of the ads.

00:00:18.320 | Two sponsors, The Jordan Harbinger Show

00:00:21.060 | and Magic Spoon Cereal.

00:00:23.880 | Please consider supporting the podcast

00:00:25.600 | by going to jordanharbinger.com/lex

00:00:29.340 | and also going to magicspoon.com/lex

00:00:33.800 | and using code LEX at checkout

00:00:36.120 | after you buy all of their cereal.

00:00:39.080 | Click the links, buy the stuff.

00:00:40.920 | It's the best way to support this podcast

00:00:43.040 | and the journey I'm on.

00:00:44.760 | If you enjoy this podcast, subscribe on YouTube,

00:00:47.680 | review it with Five Stars on Apple Podcast,

00:00:49.920 | follow on Spotify, support on Patreon,

00:00:52.360 | or connect with me on Twitter @lexfriedman,

00:00:55.600 | spelled surprisingly without the E, just F-R-I-D-M-A-N.

00:01:00.600 | As usual, I'll do a few minutes of ads now

00:01:03.900 | and never any ads in the middle

00:01:05.180 | that can break the flow of the conversation.

00:01:07.620 | This episode is supported by The Jordan Harbinger Show.

00:01:11.740 | Go to jordanharbinger.com/lex.

00:01:15.220 | It's how he knows I sent you.

00:01:16.900 | On that page, subscribe to his podcast

00:01:19.380 | on Apple Podcast, Spotify, and you know where to look.

00:01:24.340 | I've been binging on his podcast.

00:01:26.100 | Jordan is a great interviewer

00:01:28.400 | and even a better human being.

00:01:30.280 | I recently listened to his conversation with Jack Barsky,

00:01:32.760 | former sleeper agent for the KGB in the '80s

00:01:36.120 | and author of "Deep Undercover,"

00:01:38.880 | which is a memoir that paints yet another

00:01:40.760 | interesting perspective on the Cold War era.

00:01:43.440 | I've been reading a lot about the Stalin

00:01:46.720 | and then Gorbachev and Putin eras of Russia,

00:01:49.280 | but this conversation made me realize

00:01:50.800 | that I need to do a deep dive into the Cold War era

00:01:53.680 | to get a complete picture of Russia's recent history.

00:01:57.100 | Again, go to jordanharbinger.com/lex.

00:02:01.100 | Subscribe to his podcast.

00:02:02.820 | It's how he knows I sent you.

00:02:04.400 | It's awesome.

00:02:05.240 | You won't regret it.

00:02:06.700 | This episode is also supported by Magic Spoon,

00:02:10.280 | low-carb, keto-friendly, super amazingly delicious cereal.

00:02:15.280 | I've been on a keto or very low-carb diet

00:02:18.260 | for a long time now.

00:02:19.440 | It helps with my mental performance.

00:02:21.260 | It helps with my physical performance,

00:02:22.800 | even during this crazy push-up,

00:02:25.280 | pull-up challenge I'm doing, including the running.

00:02:27.760 | It just feels great.

00:02:29.640 | I used to love cereal.

00:02:31.280 | Obviously, I can't have it now

00:02:33.800 | because most cereals have a crazy amount of sugar,

00:02:36.760 | which is terrible for you.

00:02:38.360 | So I quit it years ago.

00:02:40.100 | But Magic Spoon, amazingly, somehow,

00:02:44.200 | is a totally different thing.

00:02:45.880 | Zero sugar, 11 grams of protein,

00:02:48.300 | and only three net grams of carbs.

00:02:50.880 | It tastes delicious.

00:02:53.080 | It has a lot of flavors, two new ones,

00:02:55.160 | including peanut butter.

00:02:56.700 | But if you know what's good for you,

00:02:58.480 | you'll go with cocoa, my favorite flavor,

00:03:01.520 | and the flavor of champions.

00:03:04.160 | Click the magicspoon.com/lex link in the description

00:03:07.840 | and use code LEX at checkout for free shipping

00:03:11.000 | and to let them know I sent you.

00:03:13.080 | They've agreed to sponsor this podcast for a long time.

00:03:16.440 | They're an amazing sponsor and an even better cereal.

00:03:19.920 | I highly recommend it.

00:03:21.760 | It's delicious.

00:03:22.600 | It's good for you.

00:03:23.420 | You won't regret it.

00:03:24.680 | And now, here's my conversation with Matt Botvinick.

00:03:28.460 | How much of the human brain do you think we understand?

00:03:32.360 | - I think we're at a weird moment

00:03:36.920 | in the history of neuroscience in the sense that

00:03:40.040 | I feel like we understand a lot about the brain

00:03:47.320 | at a very high level,

00:03:49.400 | but a very coarse level.

00:03:52.600 | - When you say high level, what are you thinking?

00:03:54.280 | Are you thinking functional?

00:03:55.480 | Are you thinking structurally?

00:03:57.000 | - So in other words, what is the brain for?

00:04:00.760 | You know, what kinds of computation does the brain do?

00:04:03.880 | What kinds of behaviors would we have to explain

00:04:10.040 | if we were gonna look down at the mechanistic level?

00:04:15.000 | And at that level, I feel like we understand

00:04:18.520 | much, much more about the brain than we did

00:04:20.120 | when I was in high school.

00:04:22.080 | But it's almost like we're seeing it through a fog.

00:04:25.240 | It's only at a very coarse level.

00:04:26.600 | We don't really understand what the neuronal mechanisms are

00:04:30.200 | that underlie these computations.

00:04:32.500 | We've gotten better at saying,

00:04:34.600 | what are the functions that the brain is computing

00:04:36.720 | that we would have to understand

00:04:38.400 | if we were gonna get down to the neuronal level?

00:04:40.240 | And at the other end of the spectrum,

00:04:45.520 | in the last few years, incredible progress has been made

00:04:49.600 | in terms of technologies that allow us to see,

00:04:54.600 | actually literally see in some cases,

00:04:57.220 | what's going on at the single unit level,

00:05:01.040 | even the dendritic level.

00:05:02.640 | And then there's this yawning gap in between.

00:05:05.800 | - Well, that's interesting.

00:05:06.640 | So at the high level,

00:05:07.460 | so that's almost a cognitive science level.

00:05:09.600 | And then at the neuronal level,

00:05:11.880 | that's neurobiology and neuroscience,

00:05:14.580 | just studying single neurons,

00:05:16.040 | the synaptic connections and all the dopamine,

00:05:19.800 | all the kind of neurotransmitters.

00:05:21.560 | - One blanket statement I should probably make is that,

00:05:24.260 | as I've gotten older,

00:05:27.760 | I have become more and more reluctant

00:05:30.200 | to make a distinction between psychology and neuroscience.

00:05:33.400 | To me, the point of neuroscience

00:05:37.240 | is to study what the brain is for.

00:05:41.780 | If you're a nephrologist

00:05:44.360 | and you wanna learn about the kidney,

00:05:46.540 | you start by saying, what is this thing for?

00:05:49.960 | Well, it seems to be for taking blood on one side

00:05:54.960 | that has metabolites in it that shouldn't be there,

00:06:00.220 | sucking them out of the blood

00:06:03.280 | while leaving the good stuff behind,

00:06:05.120 | and then excreting that in the form of urine.

00:06:07.020 | That's what the kidney is for, it's like obvious.

00:06:10.200 | So the rest of the work is deciding how it does that.

00:06:13.160 | And this, it seems to me,

00:06:14.760 | is the right approach to take to the brain.

00:06:17.040 | You say, well, what is the brain for?

00:06:19.080 | The brain, as far as I can tell, is for producing behavior.

00:06:22.720 | It's for going from perceptual inputs to behavioral outputs,

00:06:26.960 | and the behavioral outputs should be adaptive.

00:06:30.200 | So that's what psychology is about.

00:06:33.600 | It's about understanding the structure of that function.

00:06:35.880 | And then the rest of neuroscience is about figuring out

00:06:38.880 | how those operations are actually carried out

00:06:41.880 | at a mechanistic level.

00:06:44.200 | - That's really interesting, but,

00:06:45.920 | so unlike the kidney, the brain,

00:06:49.960 | the gap between the electrical signal and behavior,

00:06:53.060 | so you truly see neuroscience as the science

00:06:59.080 | that touches behavior, how the brain generates behavior,

00:07:03.280 | or how the brain converts raw visual information

00:07:07.440 | into understanding.

00:07:09.000 | Like, you basically see cognitive science,

00:07:12.520 | psychology, and neuroscience as all one science.

00:07:15.880 | - Yeah.

00:07:16.720 | - Is that-- - It's a personal statement.

00:07:19.240 | I don't mean to-- - Is that a hopeful,

00:07:20.240 | is that a hopeful or a realistic statement?

00:07:22.920 | So certainly you will be correct in your feeling

00:07:26.880 | in some number of years, but that number of years

00:07:29.240 | could be 200, 300 years from now.

00:07:31.440 | - Oh, well, there's a--

00:07:33.400 | - Is that aspirational, or is that pragmatic engineering

00:07:37.600 | feeling that you have?

00:07:39.360 | - It's both in the sense that this is what I hope

00:07:44.360 | and expect will bear fruit over the coming decades,

00:07:51.520 | but it's also pragmatic in the sense that I'm not sure

00:07:57.560 | what we're doing in either psychology or neuroscience

00:08:02.560 | if that's not the framing.

00:08:04.920 | I don't know what it means to understand the brain

00:08:09.760 | if part of the enterprise is not about understanding

00:08:14.760 | the behavior that's being produced.

00:08:20.040 | - I mean, yeah, but I would compare it to maybe

00:08:24.240 | astronomers looking at the movement of the planets

00:08:27.200 | and the stars without any interest

00:08:30.120 | of the underlying physics, right?

00:08:32.360 | And I would argue that at least in the early days,

00:08:35.560 | there's some value to just tracing the movement

00:08:37.760 | of the planets and the stars without thinking

00:08:41.680 | about the physics too much because it's such a big leap

00:08:44.080 | to start thinking about the physics

00:08:45.600 | before you even understand even the basic

00:08:48.040 | structural elements of--

00:08:49.520 | - Oh, I agree with that, I agree.

00:08:51.080 | - But you're saying in the end,

00:08:52.280 | the goal should be to deeply understand.

00:08:54.760 | - Well, right, and I think, so I thought about this

00:08:58.080 | a lot when I was in grad school 'cause a lot of what

00:08:59.840 | I studied in grad school was psychology,

00:09:01.840 | and I found myself a little bit confused

00:09:06.040 | about what it meant to, it seems like what we were talking

00:09:10.080 | about a lot of the time were virtual causal mechanisms.

00:09:14.720 | Like, oh, well, attentional selection then selects

00:09:19.720 | some object in the environment and that is then passed on

00:09:24.280 | to the motor, information about that is passed on

00:09:26.840 | to the motor system, but these are virtual mechanisms.

00:09:29.680 | These are, they're metaphors, there's no reduction

00:09:34.680 | going on in that conversation to some physical mechanism

00:09:39.920 | that, which is really what it would take

00:09:43.160 | to fully understand how behavior is rising.

00:09:47.240 | But the causal mechanisms are definitely neurons

00:09:50.200 | interacting, I'm willing to say that

00:09:51.840 | at this point in history.

00:09:53.280 | So in psychology, at least for me personally,

00:09:56.160 | there was this strange insecurity about trafficking

00:10:00.080 | in these metaphors, which were supposed to explain

00:10:04.120 | the function of the mind.

00:10:05.640 | If you can't ground them in physical mechanisms,

00:10:09.320 | then what is the explanatory validity of these explanations?

00:10:14.320 | And I managed to soothe my own nerves by thinking

00:10:24.160 | about the history of genetics research.

00:10:29.160 | So I'm very far from being an expert

00:10:32.400 | on the history of this field, but I know enough to say

00:10:36.240 | that Mendelian genetics preceded Watson and Crick.

00:10:41.240 | And so there was a significant period of time

00:10:45.480 | during which people were productively investigating

00:10:52.760 | the structure of inheritance using what was essentially

00:10:56.200 | a metaphor, the notion of a gene.

00:10:58.600 | Oh, genes do this and genes do that,

00:11:00.760 | but where are the genes?

00:11:02.520 | They're sort of an explanatory thing that we made up

00:11:06.080 | and we ascribed to them these causal properties.

00:11:08.880 | Oh, there's a dominant, there's a recessive,

00:11:10.600 | and then they recombine it.

00:11:12.800 | And then later, there was a kind of blank there

00:11:17.440 | that was filled in with a physical mechanism.

00:11:21.620 | That connection was made.

00:11:22.880 | But it was worth having that metaphor

00:11:26.800 | because that gave us a good sense

00:11:29.360 | of what kind of causal mechanism we were looking for.

00:11:34.360 | - And the fundamental metaphor of cognition,

00:11:38.400 | you said, is the interaction of neurons.

00:11:40.440 | What is the metaphor?

00:11:42.680 | - No, no, the metaphors we use in cognitive psychology

00:11:47.680 | are things like attention, the way that memory works.

00:11:52.680 | I retrieve something from memory.

00:11:59.500 | A memory retrieval occurs.

00:12:01.940 | What is that?

00:12:02.960 | That's not a physical mechanism

00:12:06.660 | that I can examine in its own right.

00:12:08.940 | But it's still worth having, that metaphorical level.

00:12:13.900 | - Yeah, I misunderstood, actually.

00:12:15.980 | So the higher level abstractions

00:12:17.620 | is the metaphor that's most useful.

00:12:19.620 | - Yes.

00:12:20.460 | - But what about, so how does that connect

00:12:24.380 | to the idea that that arises from interaction of neurons?

00:12:29.380 | - Well--

00:12:33.820 | - Is the interaction of neurons also not a metaphor to you?

00:12:38.060 | Or is it literally, like, that's no longer a metaphor.

00:12:42.340 | That's already the lowest level of abstractions

00:12:46.100 | that could actually be directly studied.

00:12:48.900 | - Well, I'm hesitating because I think

00:12:53.780 | what I wanna say could end up being controversial.

00:12:56.920 | So what I wanna say is, yes, the interactions of neurons,

00:13:01.900 | that's not metaphorical, that's a physical fact.

00:13:04.620 | That's where the causal interactions actually occur.

00:13:08.420 | Now, I suppose you could say, well,

00:13:10.540 | even that is metaphorical relative

00:13:12.660 | to the quantum events that underlie,

00:13:14.740 | you know, I don't wanna go down that rabbit hole.

00:13:17.260 | - It's always turtles on top of turtles, yeah.

00:13:18.900 | There's turtles all the way down.

00:13:20.220 | - There is a reduction that you can do.

00:13:22.500 | You can say these psychological phenomena

00:13:24.580 | can be explained through a very different

00:13:28.120 | kind of causal mechanism,

00:13:29.060 | which has to do with neurotransmitter release.

00:13:31.380 | And so what we're really trying to do

00:13:33.740 | in neuroscience writ large, as I say,

00:13:37.040 | which for me includes psychology,

00:13:39.700 | is to take these psychological phenomena

00:13:44.340 | and map them onto neural events.

00:13:48.380 | I think remaining forever at the level of description

00:13:54.900 | that is natural for psychology,

00:14:00.420 | for me personally, would be disappointing.

00:14:02.200 | I wanna understand how mental activity

00:14:05.580 | arises from neural activity.

00:14:10.300 | But the converse is also true.

00:14:12.940 | Studying neural activity without any sense

00:14:15.820 | of what you're trying to explain,

00:14:18.440 | to me feels like at best,

00:14:23.900 | groping around at random.

00:14:27.220 | - Now, you've kind of talked about this

00:14:29.820 | bridging of the gap between psychology and neuroscience,

00:14:32.800 | but do you think it's possible,

00:14:33.980 | like my love is,

00:14:36.740 | like I fell in love with psychology

00:14:38.300 | and psychiatry in general with Freud

00:14:40.140 | when I was really young,

00:14:41.780 | and I hope to understand the mind.

00:14:43.540 | And for me, understanding the mind,

00:14:45.260 | at least at a young age,

00:14:47.100 | before I discovered AI and even neuroscience,

00:14:49.500 | is psychology.

00:14:52.860 | And do you think it's possible to understand the mind

00:14:55.860 | without getting into all the messy details of neuroscience?

00:14:59.900 | Like you kind of mentioned,

00:15:01.380 | to you it's appealing to try to understand the mechanisms

00:15:05.140 | at the lowest level, but do you think that's needed,

00:15:07.580 | that's required, to understand how the mind works?

00:15:10.260 | - That's an important part of the whole picture,

00:15:14.780 | but I would be the last person on Earth

00:15:18.500 | to suggest that that reality

00:15:23.500 | renders psychology in its own right unproductive.

00:15:27.420 | I trained as a psychologist.

00:15:31.180 | I am fond of saying that I have learned much more

00:15:34.980 | from psychology than I have from neuroscience.

00:15:38.460 | To me, psychology is a hugely important discipline.

00:15:43.460 | And one thing that warms my heart is that

00:15:47.340 | ways of investigating behavior

00:15:54.020 | that have been native to cognitive psychology

00:15:57.940 | since its dawn in the '60s

00:16:01.580 | are starting to become,

00:16:04.180 | they're starting to become interesting to AI researchers

00:16:07.620 | for a variety of reasons.

00:16:09.460 | And that's been exciting for me to see.

00:16:11.660 | - Can you maybe talk a little bit about what you see as

00:16:15.220 | beautiful aspects of psychology,

00:16:19.260 | maybe limiting aspects of psychology?

00:16:21.900 | I mean, maybe just start it off as a science, as a field.

00:16:25.620 | - To me, it was when I understood what psychology is,

00:16:29.780 | analytical psychology,

00:16:30.860 | like the way it's actually carried out,

00:16:32.780 | it was really disappointing to see two aspects.

00:16:36.220 | One is how small the N is,

00:16:39.180 | how small the number of subject is in the studies.

00:16:43.040 | And two, it was disappointing to see

00:16:45.300 | how controlled the entire, how much it was in the lab.

00:16:49.660 | It wasn't studying humans in the wild.

00:16:52.660 | There was no mechanism for studying humans in the wild.

00:16:54.980 | So that's where I became a little bit disillusioned

00:16:57.660 | to psychology.

00:16:59.480 | And then the modern world of the internet

00:17:01.700 | is so exciting to me,

00:17:02.980 | the Twitter data or YouTube data,

00:17:05.740 | data of human behavior on the internet becomes exciting

00:17:08.300 | because the N grows and then in the wild grows.

00:17:11.940 | But that's just my narrow sense.

00:17:13.860 | Like, do you have a optimistic or pessimistic,

00:17:16.580 | cynical view of psychology?

00:17:18.180 | How do you see the field broadly?

00:17:19.840 | - When I was in graduate school,

00:17:22.740 | it was early enough that there was still a thrill

00:17:27.740 | in seeing that there were ways of doing,

00:17:32.820 | there were ways of doing experimental science

00:17:35.660 | that provided insight to the structure of the mind.

00:17:40.060 | One thing that impressed me most

00:17:42.060 | when I was at that stage in my education

00:17:44.420 | was neuropsychology,

00:17:46.020 | looking at, looking at the,

00:17:48.380 | analyzing the behavior of populations

00:17:51.460 | who had brain damage of different kinds

00:17:55.540 | and trying to understand what the specific deficits were

00:18:00.540 | that arose from a lesion in a particular part of the brain.

00:18:06.620 | And the kind of experimentation that was done

00:18:08.940 | and that's still being done to get answers in that context

00:18:13.540 | was so creative and it was so deliberate.

00:18:18.160 | It was good science.

00:18:21.340 | An experiment answered one question but raised another

00:18:24.380 | and somebody would do an experiment

00:18:25.580 | that answered that question.

00:18:26.580 | And you really felt like you were narrowing in on

00:18:29.340 | some kind of approximate understanding

00:18:31.740 | of what this part of the brain was for.

00:18:34.820 | - Do you have an example from memory

00:18:36.860 | of what kind of aspects of the mind

00:18:39.540 | could be studied in this kind of way?

00:18:41.380 | - Oh, sure.

00:18:42.220 | I mean, the very detailed neuropsychological studies

00:18:45.820 | of language, language function,

00:18:49.700 | looking at production and reception

00:18:52.020 | and the relationship between, you know,

00:18:54.700 | visual function, you know,

00:18:57.060 | reading and auditory and semantic.

00:19:00.100 | And there were these, and still are,

00:19:02.860 | these beautiful models that came out

00:19:04.340 | of that kind of research that really made you feel

00:19:07.060 | like you understood something

00:19:08.420 | that you hadn't understood before

00:19:10.300 | about how, you know,

00:19:12.460 | language processing is organized in the brain.

00:19:15.260 | But having said all that,

00:19:17.240 | you know, I think, you know,

00:19:22.240 | I think you are, I mean, I agree with you

00:19:25.320 | that the cost of doing highly controlled experiments

00:19:30.320 | is that you, by construction, miss out on the richness

00:19:35.880 | and complexity of the real world.

00:19:39.080 | One thing that, so I was drawn into science

00:19:42.280 | by what in those days was called connectionism,

00:19:44.920 | which is of course the, you know,

00:19:46.760 | what we now call deep learning.

00:19:48.680 | And at that point in history,

00:19:50.840 | neural networks were primarily being used

00:19:54.200 | in order to model human cognition.

00:19:56.400 | They weren't yet really useful for industrial applications.

00:20:00.200 | - So you always found neural networks

00:20:02.080 | in biological form beautiful.

00:20:04.080 | - Oh, neural networks were very concretely the thing

00:20:07.160 | that drew me into science.

00:20:09.120 | I was handed, are you familiar with the PDP books

00:20:13.320 | from the '80s?

00:20:14.920 | So when I was in, I went to medical school

00:20:16.520 | before I went into science.

00:20:18.240 | And-- - Really?

00:20:19.720 | - Yeah. - Interesting.

00:20:20.800 | Wow. - I also did a graduate degree

00:20:23.280 | in art history, so I kind of explore it.

00:20:26.460 | - Well, art history, I understand.

00:20:28.560 | That's just a curious, creative mind.

00:20:31.280 | But medical school, with the dream of what?

00:20:33.940 | If we take that slight tangent,

00:20:35.840 | did you want to be a surgeon?

00:20:39.120 | - I actually was quite interested in surgery.

00:20:41.680 | I was interested in surgery and psychiatry.

00:20:44.200 | And I thought, I must be the only person

00:20:47.280 | on the planet who was torn between those two fields.

00:20:52.720 | And I said exactly that to my advisor in medical school,

00:20:56.840 | who turned out, I found out later,

00:20:59.460 | to be a famous psychoanalyst.

00:21:01.920 | And he said to me, "No, no, it's actually not so uncommon

00:21:05.120 | "to be interested in surgery and psychiatry."

00:21:07.760 | And he conjectured that the reason

00:21:10.480 | that people develop these two interests

00:21:12.600 | is that both fields are about going beneath the surface

00:21:15.480 | and kind of getting into the kind of secret.

00:21:19.120 | I mean, maybe you understand this

00:21:20.640 | as someone who was interested in psychoanalysis

00:21:22.560 | at a younger stage.

00:21:23.440 | There's sort of a, there's a cliche phrase

00:21:26.200 | that people use now on NPR,

00:21:28.400 | the secret life of blankety-blank, right?

00:21:31.400 | And that was part of the thrill of surgery,

00:21:33.560 | was seeing the secret activity

00:21:38.120 | that's inside everybody's abdomen and thorax.

00:21:40.560 | - That's a very poetic way to connect it to disciplines

00:21:43.880 | that are very, practically speaking,

00:21:45.560 | different from each other, I think.

00:21:46.400 | - That's for sure, that's for sure, yes.

00:21:48.480 | - So how did we get onto medical school?

00:21:52.480 | - So I was in medical school

00:21:53.740 | and I was doing a psychiatry rotation

00:21:57.360 | and my kind of advisor in that rotation

00:22:01.000 | asked me what I was interested in.

00:22:04.720 | And I said, "Well, maybe psychiatry."

00:22:07.800 | He said, "Why?"

00:22:09.320 | And I said, "Well, I've always been interested

00:22:11.080 | "in how the brain works.

00:22:13.020 | "I'm pretty sure that nobody's doing scientific research

00:22:16.100 | "that addresses my interests, which are,"

00:22:20.600 | I didn't have a word for it then,

00:22:21.920 | but I would have said, "about cognition."

00:22:25.000 | And he said, "Well, you know, I'm not sure that's true.

00:22:27.640 | "You might be interested in these books."

00:22:29.600 | And he pulled down the PDB books from his shelf

00:22:32.480 | and they were still shrink-wrapped.

00:22:34.000 | He hadn't read them, but he handed them to me.

00:22:36.960 | He said, "You feel free to borrow these."

00:22:38.680 | And that was, you know, I went back to my dorm room

00:22:41.480 | and I just, you know, read them cover to cover.

00:22:43.520 | - What's PDP?

00:22:44.960 | - Parallel Distributed Processing,

00:22:46.520 | which was one of the original names for deep learning.

00:22:50.840 | - And so, I apologize for the romanticized question,

00:22:55.000 | but what idea in the space of neuroscience,

00:22:58.360 | in the space of the human brain,

00:22:59.880 | is to you the most beautiful, mysterious, surprising?

00:23:03.880 | - What had always fascinated me,

00:23:07.160 | even when I was a pretty young kid, I think,

00:23:12.320 | was the paradox that lies in the fact

00:23:17.320 | that the brain is so mysterious and seems so distant.

00:23:26.360 | But at the same time,

00:23:32.520 | it's responsible for the full transparency

00:23:37.360 | of everyday life.

00:23:39.040 | The brain is literally what makes everything obvious

00:23:41.520 | and familiar.

00:23:42.880 | And there's always one in the room with you.

00:23:47.280 | - Yeah.

00:23:48.120 | - I used to teach, when I taught at Princeton,

00:23:50.520 | I used to teach a cognitive neuroscience course.

00:23:53.000 | And the very last thing I would say to the students was,

00:23:56.720 | you know, when people think of scientists

00:24:01.960 | as scientific inspiration, the metaphor is often,

00:24:05.760 | well, look to the stars, you know?

00:24:08.040 | The stars will inspire you to wonder at the universe

00:24:11.600 | and think about your place in it and how things work.

00:24:16.360 | And I'm all for looking at the stars,

00:24:18.320 | but I've always been much more inspired.

00:24:21.560 | And my sense of wonder comes from the,

00:24:25.320 | not from the distant, mysterious stars,

00:24:28.480 | but from the extremely intimately close brain.

00:24:33.480 | - Yeah.

00:24:35.200 | - There's something just endlessly fascinating

00:24:38.600 | to me about that.

00:24:40.240 | - Like Jessica said, the one that's close and yet distant,

00:24:45.240 | in terms of our understanding of it,

00:24:47.920 | do you, are you also captivated by the fact

00:24:52.920 | that this very conversation is happening

00:24:55.960 | because two brains are communicating?

00:24:57.480 | - Yes, exactly.

00:24:58.560 | - I guess what I mean is the subjective nature

00:25:03.800 | of the experience, if we can take a small tangent

00:25:06.360 | into the mystical of it, the consciousness,

00:25:10.280 | or when you're saying you're captivated

00:25:13.360 | by the idea of the brain, are you talking

00:25:15.520 | about specifically the mechanism of cognition?

00:25:18.240 | Or are you also just, like, at least for me,

00:25:23.240 | it's almost like paralyzing the beauty and the mystery

00:25:26.640 | of the fact that it creates the entirety of the experience,

00:25:29.520 | not just the reasoning capability, but the experience?

00:25:32.920 | - Well, I definitely resonate with that latter thought.

00:25:37.920 | And I often find discussions of artificial intelligence

00:25:43.920 | to be disappointingly narrow.

00:25:49.120 | Speaking as someone who has always had an interest in art,

00:25:55.960 | - Right, I was just gonna go there,

00:25:57.360 | 'cause it sounds like somebody who has an interest in art.

00:26:00.160 | - Yeah, I mean, there are many layers

00:26:04.000 | to full bore human experience.

00:26:08.160 | And in some ways, it's not enough to say,

00:26:11.960 | "Oh, well, don't worry, we're talking about cognition,

00:26:14.960 | but we'll add emotion."

00:26:16.240 | - Yeah.

00:26:17.080 | - There's an incredible scope to what humans go through

00:26:24.040 | in every moment.

00:26:25.360 | And yes, so that's part of what fascinates me,

00:26:33.440 | is that our brains are producing that,

00:26:37.320 | but at the same time, it's so mysterious to us, how?

00:26:43.880 | - Yeah.

00:26:44.720 | - Like, we literally, our brains are literally

00:26:48.000 | in our heads producing this experience.

00:26:50.200 | - Producing the experience.

00:26:52.040 | - And yet, it's so mysterious to us,

00:26:55.120 | and the scientific challenge of getting at

00:26:57.680 | the actual explanation for that is so overwhelming.

00:27:02.680 | That's just, I don't know.

00:27:05.600 | Certain people have fixations on particular questions,

00:27:08.480 | and that's always, that's just always been mine.

00:27:11.720 | - Yeah, I would say the poetry of that is fascinating.

00:27:14.040 | And I'm really interested in natural language as well.

00:27:16.760 | And when you look at artificial intelligence community,

00:27:19.480 | it always saddens me how much,

00:27:23.880 | when you try to create a benchmark for the community

00:27:26.840 | to gather around, how much of the magic of language

00:27:30.280 | is lost when you create that benchmark.

00:27:33.240 | That there's something, we talk about experience,

00:27:35.920 | the music, the language, the wit,

00:27:38.600 | the something that makes a rich experience,

00:27:41.080 | something that would be required to pass

00:27:43.800 | the spirit of the Turing test is lost in these benchmarks.

00:27:47.660 | And I wonder how to get it back in,

00:27:50.240 | 'cause it's very difficult.

00:27:51.920 | The moment you try to do real good rigorous science,

00:27:55.160 | you lose some of that magic.

00:27:57.000 | When you try to study cognition

00:28:00.200 | in a rigorous scientific way,

00:28:01.600 | it feels like you're losing some of the magic.

00:28:03.840 | - Mm-hmm, mm-hmm.

00:28:04.680 | - The seeing cognition in a mechanistic way

00:28:07.560 | that AI, at this stage in our history, okay.

00:28:10.080 | - Well, I agree with you, but at the same time,

00:28:13.740 | one thing that I found really exciting

00:28:18.100 | about that first wave of deep learning models in cognition

00:28:23.020 | was the fact that the people who were building these models

00:28:28.020 | were focused on the richness and complexity

00:28:33.020 | of human cognition.

00:28:34.860 | So an early debate in cognitive science,

00:28:39.860 | which I sort of witnessed as a grad student,

00:28:41.900 | was about something that sounds very dry,

00:28:44.260 | which is the formation of the past tense.

00:28:47.260 | But there were these two camps.

00:28:49.260 | One said, well, the mind encodes certain rules,

00:28:54.260 | and it also has a list of exceptions,

00:28:57.960 | because of course, the rule is add E-D,

00:29:00.460 | but that's not always what you do,

00:29:01.900 | so you have to have a list of exceptions.

00:29:03.900 | And then there were the connectionists

00:29:07.020 | who evolved into the deep learning people,

00:29:10.420 | who said, well, if you look carefully at the data,

00:29:13.860 | if you actually look at corpora, like language corpora,

00:29:18.660 | it turns out to be very rich,

00:29:20.140 | because yes, there are most verbs,

00:29:25.140 | and you just tack on E-D, and then there are exceptions,

00:29:28.680 | but there are rules that,

00:29:32.420 | the exceptions aren't just random.

00:29:36.140 | There are certain clues to which verbs should be exceptional,

00:29:41.140 | and then there are exceptions to the exceptions,

00:29:44.100 | and there was a word that was kind of deployed

00:29:47.720 | in order to capture this, which was quasi-regular.

00:29:51.720 | In other words, there are rules, but it's messy,

00:29:54.700 | and there's structure even among the exceptions,

00:29:58.740 | and it would be, yeah, you could try to write down

00:30:03.180 | the structure in some sort of closed form,

00:30:04.820 | but really, the right way to understand

00:30:07.580 | how the brain is handling all this,

00:30:09.060 | and by the way, producing all of this,

00:30:11.460 | is to build a deep neural network

00:30:14.000 | and train it on this data

00:30:15.220 | and see how it ends up representing all of this richness.

00:30:18.520 | So the way that deep learning

00:30:21.420 | was deployed in cognitive psychology

00:30:23.720 | was that was the spirit of it.

00:30:25.940 | It was about that richness,

00:30:28.060 | and that's something that I always found very compelling.

00:30:31.940 | Still do.

00:30:33.140 | - Is there something especially interesting

00:30:36.180 | and profound to you in terms of our current deep learning

00:30:39.780 | neural network, artificial neural network approaches,

00:30:42.620 | and whatever we do understand

00:30:46.300 | about the biological neural networks in our brain?

00:30:48.780 | There's quite a few differences.

00:30:52.420 | Are some of them to you either interesting

00:30:55.940 | or perhaps profound in terms of the gap

00:31:02.180 | we might want to try to close

00:31:04.500 | in trying to create a human-level intelligence?

00:31:07.580 | - What I would say here is something

00:31:08.820 | that a lot of people are saying,

00:31:10.740 | which is that one seeming limitation

00:31:15.740 | of the systems that we're building now

00:31:18.940 | is that they lack the kind of flexibility,

00:31:21.840 | the readiness to sort of turn on a dime

00:31:25.980 | when the context calls for it

00:31:28.220 | that is so characteristic of human behavior.

00:31:32.180 | So--

00:31:33.020 | - Is that connected to you to the,

00:31:34.900 | like which aspect of the neural networks in our brain

00:31:37.740 | is that connected to?

00:31:39.140 | Is that closer to the cognitive science level of...

00:31:42.820 | Now again, see, like my natural inclination

00:31:47.340 | is to separate into three disciplines

00:31:49.620 | of neuroscience, cognitive science, and psychology,

00:31:54.260 | and you've already kind of shut that down

00:31:56.380 | by saying you're kind of seeing them as separate,

00:31:58.380 | but just to look at those layers, I guess,

00:32:02.020 | where is there something about the lowest layer

00:32:05.300 | of the way the neurons interact

00:32:09.140 | that is profound to you in terms of its difference

00:32:13.340 | to the artificial neural networks?

00:32:15.460 | Or is all the key differences

00:32:17.220 | at a higher level of abstraction?

00:32:19.240 | - One thing I often think about is that,

00:32:24.420 | if you take an introductory computer science course

00:32:27.100 | and they are introducing you

00:32:28.980 | to the notion of Turing machines,

00:32:31.420 | one way of articulating what the significance

00:32:36.420 | of a Turing machine is, is that it's a machine emulator.

00:32:41.540 | It can emulate any other machine.

00:32:45.240 | And that to me,

00:32:52.900 | that way of looking at a Turing machine

00:32:54.900 | really sticks with me.

00:32:57.580 | I think of humans as maybe sharing

00:33:01.900 | in some of that character.

00:33:04.940 | We're capacity limited, we're not Turing machines,

00:33:06.980 | obviously, but we have the ability to adapt behaviors

00:33:10.980 | that are very much unlike anything we've done before,

00:33:15.340 | but there's some basic mechanism

00:33:17.660 | that's implemented in our brain

00:33:18.900 | that allows us to run software.

00:33:22.340 | - But just on that point, you mentioned Turing machine,

00:33:24.580 | but nevertheless, it's fundamentally our brains

00:33:27.340 | are just computational devices in your view?

00:33:29.700 | Is that what you're getting at?

00:33:32.060 | It was a little bit unclear to this line you drew.

00:33:35.620 | Is there any magic in there

00:33:37.780 | or is it just basic computation?

00:33:40.660 | - I'm happy to think of it as just basic computation,

00:33:43.300 | but mind you, I won't be satisfied

00:33:46.100 | until somebody explains to me

00:33:47.460 | what the basic computations are

00:33:49.820 | that are leading to the full richness of human cognition.

00:33:53.860 | - Yes.

00:33:54.700 | - I mean, it's not gonna be enough for me

00:33:56.660 | to understand what the computations are

00:33:58.860 | that allow people to do arithmetic or play chess.

00:34:02.180 | I want the whole thing.

00:34:06.340 | - And a small tangent,

00:34:07.780 | because you kind of mentioned coronavirus,

00:34:10.500 | there's group behavior.

00:34:12.420 | - Oh, sure.

00:34:13.460 | - Is there something interesting

00:34:14.940 | to your search of understanding the human mind

00:34:18.700 | where behavior of large groups

00:34:21.540 | or just behavior of groups is interesting?

00:34:24.260 | Seeing that as a collective mind,

00:34:25.620 | as a collective intelligence,

00:34:27.180 | perhaps seeing the groups of people

00:34:28.940 | as a single intelligent organisms,

00:34:31.100 | especially looking at the reinforcement learning work

00:34:34.220 | you've done recently.

00:34:35.660 | - Well, yeah, I can't, I mean,

00:34:37.340 | I have the honor of working

00:34:41.820 | with a lot of incredibly smart people

00:34:43.700 | and I wouldn't wanna take any credit

00:34:46.140 | for leading the way on the multi-agent work

00:34:48.820 | that's come out of my group or DeepMind lately,

00:34:51.380 | but I do find it fascinating.

00:34:53.860 | And I mean, I think there,

00:34:57.340 | I think it can't be debated.

00:35:01.540 | The human behavior arises within communities.

00:35:06.020 | That just seems to me self-evident.

00:35:08.940 | - But to me, it is self-evident,

00:35:11.380 | but that seems to be a profound aspects

00:35:14.700 | of something that created.

00:35:16.060 | That was like, if you look at like 2001 Space Odyssey

00:35:19.180 | when the monkeys touched the,

00:35:21.980 | like that's the magical moment.

00:35:23.420 | I think Yuval Harari argues that the ability

00:35:26.660 | of our large numbers of humans to hold an idea,

00:35:30.220 | to converge towards idea together,

00:35:31.900 | like you said, shaking hands versus bumping elbows,

00:35:34.380 | somehow converge like without even like,

00:35:38.340 | like without being in a room altogether,

00:35:40.860 | just kind of this like distributed convergence

00:35:43.380 | towards an idea over a particular period of time

00:35:46.700 | seems to be fundamental to just every aspect

00:35:51.500 | of our cognition of our intelligence,

00:35:53.420 | because humans, we'll talk about reward,

00:35:56.740 | but it seems like we don't really have

00:35:58.700 | a clear objective function under which we operate,

00:36:01.340 | but we all kind of converge towards one somehow.

00:36:04.140 | And that to me has always been a mystery

00:36:06.740 | that I think is somehow productive

00:36:09.820 | for also understanding AI systems.

00:36:13.620 | But I guess that's the next step.

00:36:16.540 | The first step is try to understand the mind.

00:36:18.780 | - Well, I don't know.

00:36:19.700 | I mean, I think there's something to the argument

00:36:22.540 | that that kind of bottom,

00:36:25.780 | like strictly bottom-up approach is wrong-headed.

00:36:29.940 | In other words, there are basic phenomena that,

00:36:34.940 | basic aspects of human intelligence

00:36:36.860 | that can only be understood in the context of groups.

00:36:41.860 | I'm perfectly open to that.

00:36:44.700 | I've never been particularly convinced by the notion

00:36:48.700 | that we should consider intelligence to adhere

00:36:53.700 | at the level of communities.

00:36:55.620 | I don't know why.

00:36:57.180 | I just, I'm sort of stuck on the notion

00:36:58.740 | that the basic unit that we want to understand

00:37:01.380 | is individual humans.

00:37:02.700 | And if we have to understand that

00:37:05.860 | in the context of other humans, fine.

00:37:07.700 | But for me, intelligence is just,

00:37:11.300 | I'm stubbornly, I stubbornly define it as something

00:37:14.660 | that is an aspect of an individual human.

00:37:18.820 | That's just my, I don't know if that's my take.

00:37:20.180 | - I'm with you, but that could be the reductionist dream

00:37:22.860 | of a scientist because you can understand a single human.

00:37:26.420 | It also is very possible that intelligence can only arise

00:37:30.760 | when there's multiple intelligences.

00:37:32.860 | When there's multiple sort of, it's a sad thing,

00:37:37.500 | if that's true, because it's very difficult to study.

00:37:39.900 | But if it's just one human,

00:37:42.500 | that one human would not be homo sapien,

00:37:44.900 | would not become that intelligent.

00:37:46.540 | That's a real, that's a possibility.

00:37:48.540 | - I'm with you.

00:37:49.860 | One thing I will say along these lines

00:37:52.860 | is that I think,

00:37:58.580 | I think a serious effort to understand human intelligence

00:38:03.420 | and maybe to build a human-like intelligence

00:38:09.700 | needs to pay just as much attention

00:38:11.860 | to the structure of the environment

00:38:14.020 | as to the structure of the, you know,

00:38:17.660 | the cognizing system, whether it's a brain or an AI system.

00:38:22.200 | That's one thing I took away actually from my early studies

00:38:26.780 | with the pioneers of neural network research,

00:38:29.900 | people like Jay McClelland and John Cohen.

00:38:32.220 | The structure of cognition is really,

00:38:38.580 | it's only partly a function of the, you know,

00:38:42.820 | the architecture of the brain

00:38:44.500 | and the learning algorithms that it implements.

00:38:46.980 | What it's really a function,

00:38:48.260 | what really shapes it is the interaction of those things

00:38:51.500 | with the structure of the world

00:38:54.460 | in which those things are embedded, right?

00:38:56.680 | - And that's especially important for,

00:38:58.300 | that's made most clear in reinforcement learning

00:39:00.880 | where a simulated environment is,

00:39:03.700 | you can only learn as much as you can simulate,

00:39:05.820 | and that's what made, what DeepMind made very clear

00:39:09.340 | with the other aspect of the environment,

00:39:11.080 | which is the self-play mechanism

00:39:13.580 | of the other agent of the competitive behavior,

00:39:16.840 | which the other agent becomes the environment essentially.

00:39:19.980 | And that's, I mean, one of the most exciting ideas in AI

00:39:24.060 | is the self-play mechanism

00:39:25.720 | that's able to learn successfully.

00:39:27.920 | So there you go.

00:39:28.760 | There's a thing where competition is essential for learning,

00:39:33.440 | at least in that context.

00:39:35.040 | So if we can step back into another sort of beautiful world,

00:39:37.960 | which is the actual mechanics,

00:39:42.040 | the dirty mess of it, of the human brain,

00:39:44.680 | is there something for people who might not know,

00:39:48.520 | is there something you can comment on

00:39:51.120 | or describe the key parts of the brain

00:39:53.960 | that are important for intelligence,

00:39:55.520 | or just in general, what are the different parts

00:39:58.260 | of the brain that you're curious about,

00:39:59.900 | that you've studied, and that are just good to know about

00:40:03.940 | when you're thinking about cognition?

00:40:06.300 | - Well, my area of expertise, if I have one,

00:40:11.260 | is prefrontal cortex.

00:40:14.260 | So-- - What's that?

00:40:16.560 | (laughing)

00:40:17.780 | Where do we-- - It depends on who you ask.

00:40:19.580 | The technical definition is anatomical.

00:40:24.580 | There are parts of your brain

00:40:30.640 | that are responsible for motor behavior,

00:40:32.500 | and they're very easy to identify.

00:40:34.640 | And the region of your cerebral cortex,

00:40:40.760 | the sort of outer crust of your brain

00:40:43.960 | that lies in front of those

00:40:46.460 | is defined as the prefrontal cortex.

00:40:49.360 | - And when you say anatomical, sorry to interrupt,

00:40:51.960 | so that's referring to sort of the geographic region,

00:40:56.960 | as opposed to some kind of functional definition.

00:41:00.160 | - Exactly, so this is kind of the coward's way out.

00:41:04.420 | I'm telling you what the prefrontal cortex is

00:41:06.020 | just in terms of what part of the real estate it occupies.

00:41:09.680 | - The thing in the front of the brain.

00:41:10.720 | - Yeah, exactly.

00:41:11.720 | And in fact, the early history of neuroscientific research

00:41:16.720 | of neuroscientific investigation

00:41:20.880 | of what this front part of the brain does

00:41:23.520 | is sort of funny to read because

00:41:26.400 | it was really World War I that started people

00:41:33.880 | down this road of trying to figure out

00:41:35.520 | what different parts of the brain, the human brain do

00:41:38.880 | in the sense that there were a lot of people

00:41:41.640 | with brain damage who came back from the war

00:41:43.800 | with brain damage.

00:41:44.800 | And that provided, as tragic as that was,

00:41:47.720 | it provided an opportunity for scientists

00:41:49.920 | to try to identify the functions of different brain regions.

00:41:53.520 | And that was actually incredibly productive.

00:41:56.160 | But one of the frustrations that neuropsychologists faced

00:41:59.440 | was they couldn't really identify exactly

00:42:02.120 | what the deficit was that arose from damage

00:42:05.000 | to these most kind of frontal parts of the brain.

00:42:08.420 | It was just a very difficult thing to pin down.

00:42:13.640 | There were a couple of neuropsychologists

00:42:16.040 | who identified through a large amount

00:42:20.560 | of clinical experience and close observation,

00:42:22.960 | they started to put their finger on a syndrome

00:42:26.200 | that was associated with frontal damage.

00:42:27.640 | Actually, one of them was a Russian neuropsychologist

00:42:30.440 | named Luria, who students of cognitive psychology still read.

00:42:35.100 | And what he started to figure out

00:42:40.200 | was that the frontal cortex was somehow involved

00:42:43.560 | in flexibility, in guiding behaviors

00:42:48.560 | that required someone to override a habit

00:42:54.080 | or to do something unusual,

00:42:57.560 | or to change what they were doing in a very flexible way

00:43:01.000 | from one moment to another.

00:43:02.520 | - So focused on like new experiences.

00:43:05.040 | And so the way your brain processes

00:43:08.760 | and acts in new experiences.

00:43:10.900 | - Yeah, what later helped bring this function

00:43:14.700 | into better focus was a distinction

00:43:17.180 | between controlled and automatic behavior.

00:43:19.940 | In other literatures, this is referred to

00:43:23.660 | as habitual behavior versus goal-directed behavior.

00:43:28.100 | So it's very, very clear that the human brain

00:43:33.100 | has pathways that are dedicated to habits,

00:43:36.580 | to things that you do all the time.

00:43:39.340 | And they need to be automatized

00:43:42.420 | so that they don't require you to concentrate too much.

00:43:45.100 | So that leaves your cognitive capacity

00:43:47.820 | for you to do other things.

00:43:49.780 | Just think about the difference between driving

00:43:54.700 | when you're learning to drive

00:43:55.960 | versus driving after you're fairly expert.

00:43:59.180 | There are brain pathways that slowly absorb

00:44:03.540 | those frequently performed behaviors

00:44:07.820 | so that they can be habits, so that they can be automatic.

00:44:12.340 | - That's kind of like the purest form of learning,

00:44:14.900 | I guess, is happening there, which is why,

00:44:18.380 | I mean, this is kind of jumping ahead,

00:44:20.020 | which is why that perhaps is the most useful

00:44:22.160 | for us to focusing on and trying to see

00:44:24.120 | how artificial intelligence systems can learn.

00:44:27.380 | Is that the way you think? - It's interesting.

00:44:28.300 | I do think about this distinction

00:44:30.060 | between controlled and automatic,

00:44:31.460 | or goal-directed and habitual behavior a lot

00:44:34.620 | in thinking about where we are in AI research.

00:44:38.680 | But just to finish the kind of dissertation here,

00:44:46.500 | the role of the prefrontal cortex

00:44:51.380 | is generally understood these days

00:44:54.620 | sort of in contradistinction to that habitual domain.

00:44:59.620 | In other words, the prefrontal cortex

00:45:02.320 | is what helps you override those habits.

00:45:05.840 | It's what allows you to say,

00:45:07.360 | "Whoa, whoa, what I usually do in this situation is X,

00:45:10.720 | "but given the context, I probably should do Y."

00:45:14.160 | I mean, the elbow bump is a great example, right?

00:45:18.080 | Reaching out and shaking hands

00:45:19.300 | is probably a habitual behavior,

00:45:22.520 | and it's the prefrontal cortex that allows us

00:45:26.000 | to bear in mind that there's something unusual

00:45:28.760 | going on right now, and in this situation,

00:45:31.360 | I need to not do the usual thing.

00:45:33.440 | The kind of behaviors that Luria reported,

00:45:38.560 | and he built tests for detecting these kinds of things,

00:45:42.020 | were exactly like this.

00:45:43.440 | So in other words, when I stick out my hand,

00:45:46.640 | I want you instead to present your elbow.

00:45:49.740 | A patient with frontal damage

00:45:51.080 | would have a great deal of trouble with that.

00:45:53.520 | Somebody proffering their hand would elicit a handshake.

00:45:57.760 | The prefrontal cortex is what allows us to say,

00:46:00.840 | "Hold on, hold on, that's the usual thing,

00:46:03.720 | "but I have the ability to bear in mind

00:46:07.000 | "even very unusual contexts,

00:46:09.180 | "and to reason about what behavior is appropriate there."

00:46:13.240 | - Just to get a sense, is us humans special

00:46:17.560 | in the presence of the prefrontal cortex?

00:46:20.680 | Do mice have a prefrontal cortex?

00:46:22.640 | Do other mammals that we can study?

00:46:25.420 | If no, then how do they integrate new experiences?

00:46:30.040 | - Yeah, that's a really tricky question,

00:46:33.760 | and a very timely question,

00:46:35.920 | because we have revolutionary new technologies

00:46:41.840 | for monitoring, measuring,

00:46:48.280 | and also causally influencing neural behavior

00:46:52.040 | in mice and fruit flies.

00:46:57.000 | And these techniques are not fully available

00:47:00.640 | even for studying brain function in monkeys,

00:47:05.640 | let alone humans.

00:47:07.240 | And so it's a very,

00:47:11.320 | sort of, for me at least, a very urgent question

00:47:14.040 | whether the kinds of things that we wanna understand

00:47:16.960 | about human intelligence can be pursued

00:47:20.080 | in these other organisms.

00:47:21.960 | And to put it briefly, there's disagreement.

00:47:27.880 | You know, people who study fruit flies will often tell you,

00:47:32.880 | "Hey, fruit flies are smarter than you think."

00:47:35.160 | And they'll point to experiments

00:47:36.880 | where fruit flies were able to learn new behaviors,

00:47:40.320 | were able to generalize from one stimulus to another

00:47:44.180 | in a way that suggests that they have abstractions

00:47:47.520 | that guide their generalization.

00:47:49.340 | I've had many conversations in which

00:47:54.600 | I will start by observing, you know,

00:47:58.160 | recounting some observation about mouse behavior,

00:48:03.160 | where it seemed like mice were taking an awfully long time

00:48:09.040 | to learn a task that for a human would be profoundly trivial.

00:48:13.640 | And I will conclude from that

00:48:16.440 | that mice really don't have the cognitive flexibility

00:48:18.800 | that we want to explain.

00:48:20.100 | And then a mouse researcher will say to me,

00:48:21.680 | "Well, you know, hold on.

00:48:24.360 | "That experiment may not have worked

00:48:26.200 | "because you asked a mouse to deal with stimuli

00:48:31.120 | "and behaviors that were very unnatural for the mouse.

00:48:34.120 | "If instead you kept the logic of the experiment the same,

00:48:38.600 | "but kind of put it in a, you know,

00:48:41.740 | "presented the information in a way

00:48:44.280 | "that aligns with what mice are used to dealing with

00:48:46.700 | "in their natural habitats,

00:48:48.320 | "you might find that a mouse actually has more intelligence

00:48:50.900 | "than you think."

00:48:52.400 | And then they'll go on to show you videos

00:48:54.960 | of mice doing things in their natural habitat,

00:48:57.460 | which seem strikingly intelligent, you know,

00:49:00.040 | dealing with, you know, physical problems, you know,

00:49:02.960 | I have to drag this piece of food back to my, you know,

00:49:06.200 | back to my lair, but there's something in my way

00:49:08.600 | and how do I get rid of that thing?

00:49:10.440 | So I think these are open questions to put it,

00:49:13.960 | you know, to sum that up.

00:49:15.440 | - And then taking a small step back related to that

00:49:18.560 | is you kind of mentioned we're taking a little shortcut

00:49:21.480 | by saying it's a geographic part of the,

00:49:25.240 | the prefrontal cortex is a region of the brain,

00:49:28.320 | but if we, what's your sense in a bigger philosophical view,

00:49:33.320 | prefrontal cortex and the brain in general?

00:49:36.300 | Do you have a sense that it's a set of subsystems

00:49:38.880 | in the way we've kind of implied

00:49:41.240 | that are pretty distinct?

00:49:43.800 | Or to what degree is it that,

00:49:46.360 | or to what degree is it a giant interconnected mess

00:49:49.560 | where everything kind of does everything

00:49:51.440 | and it's impossible to disentangle them?

00:49:53.920 | - I think there's overwhelming evidence

00:49:57.080 | that there's functional differentiation,

00:50:00.120 | that it's clearly not the case,

00:50:03.540 | that all parts of the brain are doing the same thing.

00:50:07.180 | This follows immediately from the kinds of studies

00:50:11.160 | of brain damage that we were chatting about before.

00:50:14.700 | It's obvious from what you see

00:50:18.140 | if you stick an electrode in the brain

00:50:19.680 | and measure what's going on

00:50:21.280 | at the level of neural activity.

00:50:24.600 | Having said that, there are two other things to add,

00:50:30.760 | which kind of, I don't know,

00:50:32.800 | maybe tug in the other direction.

00:50:34.400 | One is that it's when you look carefully

00:50:39.400 | at functional differentiation in the brain,

00:50:42.280 | what you usually end up concluding,

00:50:45.840 | at least this is my observation of the literature,

00:50:48.180 | is that the differences between regions are graded

00:50:52.820 | rather than being discrete.

00:50:55.240 | So it doesn't seem like it's easy

00:50:57.500 | to divide the brain up into true modules

00:51:01.780 | that have clear boundaries

00:51:06.780 | and that have clear channels of communication between them.

00:51:12.780 | - And this applies to the prefrontal cortex?

00:51:18.020 | - Yeah, oh yeah.

00:51:18.860 | Yeah, the prefrontal cortex is made up

00:51:20.200 | of a bunch of different sub-regions,

00:51:22.060 | the functions of which are not clearly defined

00:51:27.380 | and the borders of which seem to be quite vague.

00:51:30.800 | And then there's another thing that's popping up

00:51:34.420 | in very recent research,

00:51:36.020 | which involves application of these new features.

00:51:41.020 | Application of these new techniques,

00:51:43.660 | which there are a number of studies that suggest

00:51:47.300 | that parts of the brain

00:51:49.660 | that we would have previously thought

00:51:51.540 | were quite focused in their function

00:51:56.540 | are actually carrying signals

00:51:59.100 | that we wouldn't have thought would be there.

00:52:01.340 | For example, looking in the primary visual cortex,

00:52:04.500 | which is classically thought of as basically

00:52:07.900 | the first cortical way station

00:52:09.380 | for processing visual information.

00:52:10.900 | Basically what it should care about is,

00:52:12.980 | where are the edges in this scene that I'm viewing?

00:52:15.840 | It turns out that if you have enough data,

00:52:19.460 | you can recover information from primary visual cortex

00:52:22.220 | about all sorts of things,

00:52:23.220 | like what behavior the animal is engaged in right now

00:52:26.900 | and how much reward is on offer

00:52:29.340 | in the task that it's pursuing.

00:52:31.340 | So it's clear that even regions

00:52:35.180 | whose function is pretty well defined at a coarse grain

00:52:40.520 | are nonetheless carrying some information

00:52:42.860 | about information from very different domains.

00:52:47.060 | So the history of neuroscience is sort of this oscillation

00:52:51.540 | between the two views that you articulated,

00:52:54.020 | the kind of modular view and then the big mush view.

00:52:57.740 | And I think, I guess we're gonna end up

00:53:01.580 | somewhere in the middle,

00:53:02.780 | which is unfortunate for our understanding

00:53:05.580 | because there's something about our conceptual system

00:53:10.020 | that finds it easy to think about a modularized system

00:53:12.740 | and easy to think about a completely undifferentiated system,

00:53:15.480 | but something that kind of lies in between is confusing,

00:53:19.980 | but we're gonna have to get used to it, I think.

00:53:21.620 | - Unless we can understand deeply

00:53:23.380 | the lower level mechanism of neuronal communication.

00:53:25.900 | - Yeah, yeah.

00:53:26.740 | - So on that topic, you kind of mentioned information.

00:53:29.680 | Just to get a sense, I imagine something

00:53:31.880 | that there's still mystery and disagreement on

00:53:34.620 | is how does the brain carry information and signal?

00:53:38.060 | Like what in your sense is the basic mechanism

00:53:43.060 | of communication in the brain?

00:53:46.420 | - Well, I guess I'm old fashioned in that

00:53:50.100 | I consider the networks that we use

00:53:53.300 | in deep learning research to be a reasonable approximation

00:53:57.040 | to the mechanisms that carry information in the brain.

00:54:02.040 | So the usual way of articulating that is to say,

00:54:06.200 | what really matters is a rate code.

00:54:08.560 | What matters is how quickly is an individual neuron spiking?

00:54:13.560 | What's the frequency at which it's spiking?

00:54:16.360 | - So the timing of the spike.

00:54:17.880 | - Yeah, is it firing fast or slow?

00:54:20.340 | Let's put a number on that.

00:54:22.760 | And that number is enough to capture what neurons are doing.

00:54:26.060 | There's still uncertainty about whether

00:54:31.120 | that's an adequate description of how information

00:54:35.400 | is transmitted within the brain.

00:54:39.900 | There are studies that suggest that the precise timing

00:54:44.200 | of spikes matters.

00:54:46.080 | There are studies that suggest that there are computations

00:54:50.680 | that go on within the dendritic tree, within a neuron

00:54:54.520 | that are quite rich and structured

00:54:57.120 | and that really don't equate to anything that we're doing

00:54:59.980 | in our artificial neural networks.

00:55:02.840 | Having said that, I feel like we're getting somewhere

00:55:07.840 | by sticking to this high level of abstraction.

00:55:11.640 | - Just the rate, and by the way,

00:55:13.400 | we're talking about the electrical signal.

00:55:16.200 | I remember reading some vague paper somewhere recently

00:55:20.040 | where the mechanical signal, like the vibrations

00:55:23.400 | or something of the neurons also communicates information.

00:55:28.400 | - I haven't seen that.

00:55:30.640 | - So there's somebody was arguing that

00:55:33.040 | the electrical signal, this is in Nature paper,

00:55:36.840 | something like that, where the electrical signal

00:55:38.800 | is actually a side effect of the mechanical signal.

00:55:43.720 | But I don't think that changes the story.

00:55:46.080 | But it's almost an interesting idea

00:55:49.040 | that there could be a deeper, it's always like in physics

00:55:52.400 | with quantum mechanics, there's always a deeper story

00:55:55.720 | that could be underlying the whole thing.

00:55:57.480 | But you think it's basically the rate of spiking

00:56:00.560 | that gets us, that's like the lowest hanging fruit

00:56:02.800 | that can get us really far.

00:56:04.040 | - This is a classical view.

00:56:06.600 | I mean, this is not, the only way in which this stance

00:56:10.720 | would be controversial is in the sense that

00:56:14.000 | there are members of the neuroscience community

00:56:17.080 | who are interested in alternatives.

00:56:18.840 | But this is really a very mainstream view.

00:56:21.400 | The way that neurons communicate is that

00:56:25.440 | neurotransmitters arrive,

00:56:27.480 | they wash up on a neuron,

00:56:32.800 | the neuron has receptors for those transmitters.

00:56:35.880 | The meeting of the transmitter with these receptors

00:56:39.840 | changes the voltage of the neuron.

00:56:42.320 | And if enough voltage change occurs,

00:56:45.000 | then a spike occurs, right?

00:56:46.840 | One of these like discrete events.

00:56:48.640 | And it's that spike that is conducted down the axon

00:56:52.280 | and leads to neurotransmitter release.

00:56:54.520 | This is just like neuroscience 101.

00:56:56.800 | This is like the way the brain is supposed to work.

00:56:59.280 | Now, what we do when we build artificial neural networks

00:57:03.640 | of the kind that are now popular in the AI community

00:57:06.760 | is that we don't worry about those individual spikes.

00:57:11.760 | We just worry about the frequency

00:57:14.200 | at which those spikes are being generated.

00:57:16.960 | And we consider, people talk about that

00:57:19.360 | as the activity of a neuron.

00:57:22.280 | And so the activity of units in a deep learning system

00:57:27.120 | is broadly analogous to the spike rate of a neuron.

00:57:32.120 | There are people who believe

00:57:37.000 | that there are other forms of communication in the brain.

00:57:39.160 | In fact, I've been involved in some research recently

00:57:41.200 | that suggests that the voltage fluctuations

00:57:46.200 | that occur in populations of neurons

00:57:49.680 | that are sort of below the level of spike production

00:57:54.680 | may be important for communication.

00:57:57.680 | But I'm still pretty old school in the sense

00:58:00.640 | that I think that the things that we're building

00:58:03.120 | in AI research constitute reasonable models

00:58:07.440 | of how a brain would work.

00:58:08.760 | - Let me ask just for fun a crazy question, 'cause I can.

00:58:14.680 | Do you think it's possible we're completely wrong

00:58:17.040 | about the way this basic mechanism

00:58:20.080 | of neuronal communication,

00:58:22.320 | that the information is stored

00:58:24.120 | in some very different kind of way in the brain?

00:58:26.360 | - Oh, heck yes.

00:58:27.600 | I mean, look, I wouldn't be a scientist

00:58:29.960 | if I didn't think there was any chance we were wrong.

00:58:32.520 | But I mean, if you look at the history

00:58:36.440 | of deep learning research

00:58:38.680 | as it's been applied to neuroscience,

00:58:41.200 | of course, the vast majority of deep learning research

00:58:43.840 | these days isn't about neuroscience.

00:58:45.440 | But if you go back to the 1980s,

00:58:49.120 | there's sort of an unbroken chain of research

00:58:52.800 | in which a particular strategy is taken,

00:58:55.000 | which is, hey, let's train a deep learning system.

00:59:00.000 | Let's train a multilayer neural network

00:59:04.120 | on this task that we trained our rat on,

00:59:09.120 | or our monkey on, or this human being on.

00:59:12.320 | And then let's look at what the units

00:59:15.720 | deep in the system are doing.

00:59:17.720 | And let's ask whether what they're doing

00:59:20.800 | resembles what we know about

00:59:22.400 | what neurons deep in the brain are doing.

00:59:24.640 | And over and over and over and over,

00:59:28.560 | that strategy works in the sense that

00:59:31.120 | the learning algorithms that we have access to,

00:59:34.360 | which typically center on back propagation,

00:59:37.760 | they give rise to patterns of activity,

00:59:42.080 | patterns of response,

00:59:44.080 | patterns of neuronal behavior in these artificial models

00:59:48.760 | that look hauntingly similar to what you see in the brain.

00:59:53.680 | And is that a coincidence?

00:59:57.400 | - At a certain point, it starts looking like

00:59:59.720 | such coincidence is unlikely to not be deeply meaningful.

01:00:03.000 | Yeah. - Yeah.

01:00:03.840 | The circumstantial evidence is overwhelming.

01:00:07.160 | But it could-- - But you're always open

01:00:08.680 | to a total flipping of the table.

01:00:10.480 | - Hey, of course.

01:00:11.640 | So you have co-authored several recent papers

01:00:15.160 | that sort of weave beautifully between the world

01:00:17.880 | of neuroscience and artificial intelligence.

01:00:20.640 | And maybe if we could,

01:00:24.800 | can we just try to dance around and talk about some of them,

01:00:27.480 | maybe try to pick out interesting ideas

01:00:29.720 | that jump to your mind from memory.

01:00:32.280 | So maybe looking at,

01:00:33.720 | we were talking about the prefrontal cortex,

01:00:35.520 | the 2018, I believe, paper called

01:00:38.960 | "Prefrontal Cortex as a Meta-Reinforcement Learning System."

01:00:42.240 | What, is there a key idea

01:00:44.400 | that you can speak to from that paper?

01:00:46.720 | - Yeah, I mean, the key idea is about meta-learning.

01:00:52.680 | So-- - What is meta-learning?

01:00:54.880 | - Meta-learning is, by definition,

01:00:58.680 | a situation in which you have a learning algorithm,

01:01:06.160 | and the learning algorithm operates in such a way

01:01:09.800 | that it gives rise to another learning algorithm.

01:01:14.080 | In the earliest applications of this idea,

01:01:17.180 | you had one learning algorithm sort of adjusting

01:01:20.360 | the parameters on another learning algorithm.

01:01:23.080 | But the case that we're interested in this paper

01:01:25.120 | is one where you start with just one learning algorithm,

01:01:29.200 | and then another learning algorithm kind of emerges

01:01:33.040 | out of thin air.

01:01:35.200 | I can say more about what I mean by that.

01:01:36.720 | I don't mean to be, you know,

01:01:38.800 | (Michael laughs)

01:01:39.640 | ascurantist, but that's the idea of meta-learning.

01:01:43.320 | It relates to the old idea in psychology

01:01:46.040 | of learning to learn.

01:01:47.300 | Situations where you have experiences

01:01:54.320 | that make you better at learning something new.

01:01:57.320 | Like a familiar example would be learning a foreign language.

01:02:01.400 | The first time you learn a foreign language,

01:02:02.880 | it may be quite laborious and disorienting and novel.

01:02:07.880 | But if, let's say you've learned two foreign languages,

01:02:12.220 | the third foreign language, obviously,

01:02:14.160 | is gonna be much easier to pick up.

01:02:15.960 | And why?

01:02:16.800 | Because you've learned how to learn.

01:02:18.240 | You know how this goes.

01:02:20.200 | You know, okay, I'm gonna have to learn how to conjugate.

01:02:22.160 | I'm gonna have to...

01:02:23.920 | That's a simple form of meta-learning, right?

01:02:26.360 | In the sense that there's some slow learning mechanism

01:02:30.240 | that's helping you kind of update

01:02:33.040 | your fast learning mechanism.

01:02:34.280 | Does that bring it into focus?

01:02:36.520 | - So how, from our understanding,

01:02:39.160 | from the psychology world, from neuroscience,

01:02:42.360 | our understanding how meta-learning works

01:02:45.320 | might work in the human brain,

01:02:47.200 | what lessons can we draw from that

01:02:50.000 | that we can bring into the artificial intelligence world?

01:02:53.080 | - Well, yeah, so the origin of that paper

01:02:55.960 | was in AI work that we were doing in my group.

01:03:00.200 | We were looking at what happens

01:03:03.720 | when you train a recurrent neural network

01:03:06.280 | using standard reinforcement learning algorithms.

01:03:10.200 | But you train that network, not just in one task,

01:03:12.680 | but you train it in a bunch of interrelated tasks.

01:03:15.200 | And then you ask what happens

01:03:17.640 | when you give it yet another task

01:03:19.920 | in that sort of line of interrelated tasks.

01:03:23.360 | And what we started to realize is that

01:03:29.360 | a form of meta-learning spontaneously happens

01:03:31.840 | in recurrent neural networks.

01:03:33.760 | And the simplest way to explain it is to say

01:03:37.680 | a recurrent neural network has a kind of memory

01:03:43.480 | in its activation patterns.

01:03:45.320 | It's recurrent by definition in the sense

01:03:47.480 | that you have units that connect to other units

01:03:50.120 | that connect to other units.

01:03:51.000 | So you have sort of loops of connectivity,

01:03:53.640 | which allows activity to stick around

01:03:55.720 | and be updated over time.

01:03:57.360 | In psychology, we call, in neuroscience,

01:03:59.000 | we call this working memory.

01:04:00.080 | It's like actively holding something in mind.

01:04:03.000 | And so that memory gives the recurrent neural network

01:04:10.480 | a dynamics, right?

01:04:13.080 | The way that the activity pattern evolves over time

01:04:17.680 | is inherent to the connectivity

01:04:19.960 | of the recurrent neural network, okay?

01:04:21.560 | So that's idea number one.

01:04:23.480 | Now, the dynamics of that network

01:04:25.520 | are shaped by the connectivity, by the synaptic weights.

01:04:29.600 | And those synaptic weights are being shaped

01:04:31.600 | by this reinforcement learning algorithm

01:04:33.800 | that you're training the network with.

01:04:35.960 | So the punchline is,

01:04:39.120 | if you train a recurrent neural network

01:04:41.200 | with a reinforcement learning algorithm

01:04:43.080 | that's adjusting its weights,

01:04:44.160 | and you do that for long enough,

01:04:45.840 | the activation dynamics will become very interesting, right?

01:04:50.800 | So imagine I give you a task

01:04:53.160 | where you have to press one button or another,

01:04:56.040 | left button or right button.

01:04:57.560 | And there's some probability

01:05:00.760 | that I'm gonna give you an M&M

01:05:02.240 | if you press the left button,

01:05:04.160 | and there's some probability I'll give you an M&M

01:05:06.160 | if you press the other button.

01:05:07.560 | And you have to figure out what those probabilities are

01:05:09.320 | just by trying things out.

01:05:10.640 | But as I said before,

01:05:13.760 | instead of just giving you one of these tasks,

01:05:15.480 | I give you a whole sequence.

01:05:16.960 | You know, I give you two buttons

01:05:18.640 | and you figure out which one's best,

01:05:19.840 | and I go, "Good job, here's a new box,

01:05:22.120 | two new buttons, you have to figure out which one's best.

01:05:24.040 | Good job, here's a new box."

01:05:25.400 | And every box has its own probabilities

01:05:27.280 | and you have to figure it.

01:05:28.240 | So if you train a recurrent neural network

01:05:30.360 | on that kind of sequence of tasks,

01:05:32.600 | what happens, it seemed almost magical to us

01:05:37.320 | when we first started kind of realizing what was going on.

01:05:41.120 | The slow learning algorithm

01:05:42.880 | that's adjusting the synaptic weights,

01:05:45.480 | those slow synaptic changes give rise to a network dynamics

01:05:51.320 | that themselves, that, you know,

01:05:52.960 | the dynamics themselves turn into a learning algorithm.

01:05:56.840 | So in other words, you can tell this is happening

01:05:59.040 | by just freezing the synaptic weights, saying,

01:06:00.920 | "Okay, no more learning, you're done.

01:06:03.360 | Here's a new box, figure out which button is best."

01:06:07.600 | And the recurrent neural network will do this just fine.

01:06:09.600 | There's no, like it figures out which button is best.

01:06:13.040 | It kind of transitions from exploring the two buttons

01:06:16.680 | to just pressing the one that it likes best

01:06:18.360 | in a very rational way.

01:06:20.680 | How is that happening?

01:06:21.680 | It's happening because the activity dynamics of the network

01:06:25.840 | have been shaped by the slow learning process

01:06:28.480 | that's occurred over many, many boxes.

01:06:30.720 | And so what's happened is that this slow learning algorithm

01:06:34.680 | that's slowly adjusting the weights

01:06:37.160 | is changing the dynamics of the network,

01:06:39.760 | the activity dynamics into its own learning algorithm.

01:06:43.480 | And as we were kind of realizing that this is a thing,

01:06:49.160 | it just so happened that the group that was working on this

01:06:53.760 | included a bunch of neuroscientists.

01:06:56.040 | And it started kind of ringing a bell for us,

01:06:59.920 | which is to say that we thought,

01:07:01.680 | "This sounds a lot like the distinction

01:07:03.720 | between synaptic learning and activity,

01:07:07.560 | synaptic memory and activity-based memory in the brain."

01:07:10.560 | And it also reminded us of recurrent connectivity

01:07:15.920 | that's very characteristic of prefrontal function.

01:07:18.400 | So this is kind of why it's good to have people working

01:07:22.840 | on AI that know a little bit about neuroscience

01:07:26.240 | and vice versa, because we started thinking about

01:07:29.560 | whether we could apply this principle to neuroscience.

01:07:32.360 | And that's where the paper came from.

01:07:33.680 | - So the kind of principle of the recurrence

01:07:37.560 | they can see in the prefrontal cortex,

01:07:39.560 | then you start to realize that it's possible

01:07:42.920 | to force something like an idea of a learning to learn,

01:07:47.360 | emerging from this learning process,

01:07:50.880 | as long as you keep varying the environment sufficiently.

01:07:54.520 | - Exactly.

01:07:55.360 | So the kind of metaphorical transition

01:07:59.320 | we made to neuroscience was to think,

01:08:00.720 | "Okay, well, we know that the prefrontal cortex

01:08:03.720 | is highly recurrent.

01:08:05.000 | We know that it's an important locus for working memory,

01:08:08.540 | for activation-based memory.

01:08:11.300 | So maybe the prefrontal cortex

01:08:13.720 | supports reinforcement learning."

01:08:15.640 | In other words, what is reinforcement learning?

01:08:19.320 | You take an action, you see how much reward you got,

01:08:21.680 | you update your policy of behavior.

01:08:23.620 | Maybe the prefrontal cortex is doing that sort of thing

01:08:26.900 | strictly in its activation patterns.

01:08:28.520 | It's keeping around a memory in its activity patterns

01:08:31.920 | of what you did, how much reward you got,

01:08:35.400 | and it's using that activity-based memory

01:08:39.040 | as a basis for updating behavior.

01:08:41.120 | But then the question is,

01:08:42.020 | "Well, how did the prefrontal cortex get so smart?"

01:08:44.580 | In other words, where did these activity dynamics come from?

01:08:48.020 | How did that program that's implemented

01:08:50.780 | in the recurrent dynamics of the prefrontal cortex arise?

01:08:54.460 | And one answer that became evident in this work was,

01:08:58.020 | "Well, maybe the mechanisms that operate

01:09:00.940 | on the synaptic level,

01:09:02.820 | which we believe are mediated by dopamine,

01:09:06.460 | are responsible for shaping those dynamics."

01:09:10.180 | - So this may be a silly question,

01:09:12.460 | but because this kind of several temporal classes

01:09:17.460 | of learning are happening,

01:09:20.900 | and the learning-to-learnism emerges,

01:09:24.340 | can you keep building stacks of learning-to-learn-to-learn,

01:09:29.340 | learning-to-learn-to-learn-to-learn-to-learn,

01:09:31.660 | because it keeps, I mean, basically abstractions

01:09:34.420 | of more powerful abilities to generalize

01:09:38.700 | of learning complex rules?

01:09:41.140 | - Yeah.

01:09:41.980 | - Or is this overstretching this kind of mechanism?

01:09:46.100 | - Well, one of the people in AI

01:09:49.460 | who started thinking about meta-learning

01:09:52.460 | from very early on, Juergen and Schmidhuber,

01:09:55.900 | sort of cheekily suggested,

01:09:59.800 | I think it may have been in his PhD thesis,

01:10:03.940 | that we should think about meta, meta, meta,

01:10:06.900 | meta, meta, meta learning.

01:10:08.100 | You know, that's really what's gonna get us

01:10:11.260 | to true intelligence.

01:10:13.140 | - Certainly there's a poetic aspect to it,

01:10:15.420 | and it seems interesting and correct

01:10:19.300 | that that kind of level of abstraction would be powerful,

01:10:21.700 | but is that something you see in the brain?

01:10:23.940 | This kind of, is it useful to think of learning

01:10:27.800 | in these meta, meta, meta way,

01:10:29.660 | or is it just meta learning?

01:10:32.100 | - Well, one thing that really fascinated me

01:10:35.300 | about this mechanism that we were starting to look at,

01:10:39.020 | and other groups started talking about very similar things

01:10:43.220 | at the same time, and then a kind of explosion of interest

01:10:47.020 | in meta-learning happened in the AI community

01:10:48.980 | shortly after that.

01:10:50.580 | I don't know if we had anything to do with that,

01:10:52.060 | but I was gratified to see that a lot of people

01:10:55.620 | started talking about meta-learning.

01:10:57.780 | One of the things that I like about the kind of flavor

01:11:01.380 | of meta-learning that we were studying

01:11:03.320 | was that it didn't require anything special.

01:11:05.960 | It was just, if you took a system

01:11:07.820 | that had some form of memory,

01:11:10.020 | that the function of which could be shaped

01:11:13.260 | by pick your RL algorithm,

01:11:16.880 | then this would just happen.

01:11:18.480 | - Yes.

01:11:19.320 | - I mean, there are a lot of forms of,

01:11:21.320 | there are a lot of meta-learning algorithms

01:11:23.200 | that have been proposed since then

01:11:24.520 | that are fascinating and effective

01:11:26.600 | in their domains of application,

01:11:29.800 | but they're engineered.

01:11:31.640 | There are things that somebody had to say,

01:11:33.160 | well, gee, if we wanted meta-learning to happen,

01:11:34.960 | how would we do that?

01:11:35.800 | Here's an algorithm that would,

01:11:37.080 | but there's something about the kind of meta-learning

01:11:39.520 | that we were studying that seemed to me special

01:11:42.560 | in the sense that it wasn't an algorithm.

01:11:45.000 | It was just something that automatically happened

01:11:48.760 | if you had a system that had memory

01:11:51.080 | and it was trained with a reinforcement learning algorithm.

01:11:54.040 | And in that sense, it can be as meta as it wants to be.

01:11:59.600 | There's no limit on how abstract the meta-learning can get

01:12:04.600 | because it's not reliant on a human engineering

01:12:08.040 | a particular meta-learning algorithm to get there.

01:12:10.820 | And that's, I also, I don't know,

01:12:15.200 | I guess I hope that that's relevant in the brain.

01:12:17.880 | I think there's a kind of beauty

01:12:19.280 | in the ability of this emergent--

01:12:23.440 | - The emergent aspect of it.

01:12:24.720 | - Yeah, it's something that--

01:12:25.560 | - As opposed to engineered.

01:12:26.520 | - Exactly, it's something that just,

01:12:28.280 | it just happens in a sense.

01:12:31.000 | In a sense, you can't avoid this happening.

01:12:33.680 | If you have a system that has memory

01:12:35.880 | and the function of that memory

01:12:37.520 | is shaped by reinforcement learning,

01:12:41.440 | and this system is trained

01:12:42.800 | in a series of interrelated tasks,

01:12:45.040 | this is gonna happen, you can't stop it.

01:12:48.520 | - As long as you have certain properties,

01:12:50.200 | maybe like a recurrent structure to--

01:12:52.600 | - You have to have memory.

01:12:53.440 | It actually doesn't have to be a recurrent neural network.

01:12:55.680 | A paper that I was honored to be involved with even earlier

01:12:59.800 | used a kind of slot-based memory.

01:13:02.280 | - Do you remember the title?

01:13:03.120 | Just for people who watched it.

01:13:05.040 | - It was Memory Augmented Neural Networks.

01:13:08.120 | I think the title was

01:13:10.160 | Meta-Learning in Memory Augmented Neural Networks.

01:13:13.020 | And it was the same exact story.

01:13:17.920 | If you have a system with memory,

01:13:21.080 | here it was a different kind of memory,

01:13:22.800 | but the function of that memory

01:13:24.600 | is shaped by reinforcement learning.

01:13:28.620 | Here it was the reads and writes

01:13:32.560 | that occurred on this slot-based memory.

01:13:36.440 | This'll just happen.

01:13:38.080 | But this brings us back to something I was saying earlier

01:13:42.080 | about the importance of the environment.

01:13:44.500 | This will happen if the system is being trained

01:13:49.920 | in a setting where there's a sequence of tasks

01:13:53.080 | that all share some abstract structure.

01:13:55.220 | Sometimes we talk about task distributions.

01:13:59.020 | And that's something that's very obviously true

01:14:04.020 | of the world that humans inhabit.

01:14:06.380 | If you just kind of think about what you do every day,

01:14:12.340 | you never do exactly the same thing

01:14:16.280 | that you did the day before.

01:14:17.640 | But everything that you do has a family resemblance.

01:14:21.040 | It shares a structure with something that you did before.

01:14:23.480 | And so the real world is

01:14:26.280 | saturated with this property.

01:14:32.680 | It's endless variety with endless redundancy.

01:14:37.680 | And that's the setting

01:14:38.560 | in which this kind of meta-learning happens.

01:14:40.560 | - And it does seem like we're just so good at finding,

01:14:44.840 | just like in this emergent phenomenon

01:14:47.320 | you described, we're really good at finding that redundancy,

01:14:50.040 | finding those similarities, the family resemblance.

01:14:53.480 | Some people call it sort of, what is it?

01:14:56.560 | Melanie Mitchell was talking about analogies.

01:14:59.160 | So we're able to connect concepts together

01:15:01.920 | in this kind of way,

01:15:03.160 | in this same kind of automated emergent way.

01:15:06.120 | There's so many echoes here of psychology and neuroscience.

01:15:10.640 | And obviously now with reinforcement learning

01:15:15.320 | with recurrent neural networks at the core.

01:15:18.280 | If we could talk a little bit about dopamine,

01:15:20.160 | you have really, you're a part of co-authoring

01:15:23.800 | really exciting recent paper, very recent,

01:15:26.440 | in terms of release on dopamine

01:15:28.920 | and temporal difference learning.

01:15:31.040 | Can you describe the key ideas of that paper?

01:15:34.800 | - Sure, yeah.

01:15:35.640 | I mean, one thing I want to pause to do

01:15:37.760 | is acknowledge my co-authors on actually both

01:15:40.600 | of the papers we're talking about.

01:15:41.560 | So this dopamine paper--

01:15:43.080 | - I'll just, I'll certainly post all their names.

01:15:45.720 | - Okay, wonderful, yeah.

01:15:46.560 | 'Cause I'm sort of a bash to be the spokesperson

01:15:50.360 | for these papers when I had such amazing collaborators

01:15:53.880 | on both.

01:15:55.200 | So it's a comfort to me to know

01:15:57.000 | that you'll acknowledge them.

01:15:58.600 | - Yeah, there's an incredible team there, but yeah.

01:16:00.440 | - Oh yeah, it's so much fun.

01:16:03.120 | And in the case of the dopamine paper,

01:16:06.400 | we also collaborated with Naoichi at Harvard,

01:16:09.040 | who obviously the paper simply wouldn't have happened

01:16:11.760 | without him.

01:16:12.680 | But so you were asking for like a thumbnail sketch of--

01:16:17.600 | - Yes, a thumbnail sketch or key ideas or things,

01:16:21.320 | the insights that continue on our kind of discussion here

01:16:24.840 | between neuroscience and AI.

01:16:26.920 | - Yeah, I mean, this was another,

01:16:28.920 | a lot of the work that we've done so far

01:16:30.640 | is taking ideas that have bubbled up in AI

01:16:35.400 | and asking the question of whether the brain

01:16:39.680 | might be doing something related,

01:16:41.460 | which I think on the surface sounds like something

01:16:45.440 | that's really mainly of use to neuroscience.

01:16:48.360 | We see it also as a way of validating what we're doing

01:16:54.360 | on the AI side.

01:16:55.320 | If we can gain some evidence that the brain

01:16:57.960 | is using some technique that we've been trying out

01:17:01.760 | in our AI work, that gives us confidence

01:17:05.500 | that it may be a good idea,

01:17:07.800 | that it'll scale to rich, complex tasks,

01:17:11.540 | that it'll interface well with other mechanisms.

01:17:14.820 | - So you see it as a two-way road.

01:17:16.820 | - Yeah, for sure.

01:17:17.660 | - Just because a particular paper is a little bit focused

01:17:19.620 | on from AI from neural networks to neuroscience,

01:17:24.620 | ultimately the discussion, the thinking,

01:17:28.340 | the productive long-term aspect of it

01:17:30.780 | is the two-way road nature of the whole--

01:17:33.180 | - Yeah, I mean, we've talked about the notion

01:17:36.220 | of a virtuous circle between AI and neuroscience.

01:17:39.260 | And the way I see it, that's always been there

01:17:43.940 | since the two fields jointly existed.

01:17:49.000 | There have been some phases in that history

01:17:52.100 | when AI was sort of ahead.

01:17:53.500 | There are some phases when neuroscience was sort of ahead.

01:17:56.300 | I feel like given the burst of innovation

01:18:00.620 | that's happened recently on the AI side,

01:18:03.740 | AI is kind of ahead in the sense that

01:18:06.260 | there are all of these ideas

01:18:07.700 | for which it's exciting to consider

01:18:12.620 | that there might be neural analogs.

01:18:14.680 | And neuroscience, in a sense,

01:18:20.120 | has been focusing on approaches to studying behavior

01:18:23.560 | that come from, that are kind of derived

01:18:26.540 | from this earlier era of cognitive psychology.

01:18:29.540 | And so in some ways,

01:18:32.140 | fail to connect with some of the issues

01:18:34.380 | that we're grappling with in AI,

01:18:36.660 | like how do we deal with large, complex environments?

01:18:40.140 | But I think it's inevitable

01:18:45.180 | that this circle will keep turning

01:18:47.900 | and there will be a moment

01:18:49.500 | in the not too distant future

01:18:51.260 | when neuroscience is pelting AI researchers

01:18:54.600 | with insights that may change the direction of our work.

01:18:58.220 | - Just a quick human question.

01:19:01.820 | You have parts of your brain,

01:19:05.460 | this is very meta,

01:19:06.340 | but they're able to both think about neuroscience and AI.

01:19:10.300 | You know, I don't often meet people like that.

01:19:14.220 | So do you think,

01:19:16.060 | let me ask a metaplasticity question.

01:19:19.780 | Do you think a human being can be both good at AI

01:19:22.660 | and neuroscience?

01:19:23.580 | Is like what, on the team at DeepMind,

01:19:26.500 | what kind of human can occupy these two realms?

01:19:30.180 | And is that something you see?

01:19:31.500 | Everybody should be doing, can be doing,

01:19:34.740 | or is that a very special few can kind of jump?

01:19:37.400 | Just like we talk about art history,

01:19:39.180 | I would think it's a special person

01:19:41.020 | that can major in art history

01:19:43.620 | and also consider being a surgeon.

01:19:46.860 | - Otherwise known as a dilettante.

01:19:48.700 | - A dilettante, yeah.

01:19:50.140 | Easily distracted.

01:19:51.320 | - I think it does take a special kind of person

01:19:58.660 | to be truly world-class at both AI and neuroscience.

01:20:02.700 | And I am not on that list.

01:20:04.460 | I happen to be someone

01:20:08.100 | who's interest in neuroscience and psychology

01:20:11.380 | involved using the kinds of modeling techniques

01:20:17.660 | that are now very central in AI.

01:20:21.180 | And that sort of, I guess,

01:20:23.420 | bought me a ticket to be involved

01:20:25.180 | in all of the amazing things

01:20:26.620 | that are going on in AI research right now.

01:20:28.820 | I do know a few people

01:20:31.140 | who I would consider pretty expert on both fronts,

01:20:34.660 | and I won't embarrass them by naming them,

01:20:36.400 | but there are exceptional people out there

01:20:40.700 | who are like this.

01:20:41.540 | The one thing that I find is a barrier

01:20:46.020 | to being truly world-class on both fronts

01:20:49.420 | is just the complexity of the technology

01:20:55.140 | that's involved in both disciplines now.

01:20:58.340 | So the engineering expertise that it takes

01:21:03.140 | to do truly frontline, hands-on AI research

01:21:08.020 | is really, really considerable.

01:21:10.780 | - The learning curve of the tools,

01:21:12.100 | just like the specifics of just,

01:21:14.320 | whether it's programming or the kind of tools necessary

01:21:16.660 | to collect the data, to manage the data,

01:21:18.460 | to distribute, to compute, all that kind of stuff.

01:21:20.940 | And on the neuroscience, I guess, side,

01:21:22.580 | there'll be all different sets of tools.

01:21:24.760 | - Exactly, especially with the recent explosion

01:21:26.980 | in neuroscience methods.

01:21:28.980 | So having said all that,

01:21:32.380 | I think the best scenario for both neuroscience and AI

01:21:40.180 | is to have people who, interacting,

01:21:45.780 | who live at every point on this spectrum,

01:21:49.100 | from exclusively focused on neuroscience

01:21:52.860 | to exclusively focused on the engineering side of AI.

01:21:56.480 | But to have those people inhabiting a community

01:22:01.480 | where they're talking to people

01:22:03.600 | who live elsewhere on the spectrum.

01:22:05.680 | And I may be someone who's very close to the center

01:22:09.540 | in the sense that I have one foot in the neuroscience world

01:22:13.120 | and one foot in the AI world.

01:22:15.060 | And that central position, I will admit,

01:22:18.100 | prevents me, at least someone

01:22:20.020 | with my limited cognitive capacity,

01:22:22.460 | from having true technical expertise in either domain.

01:22:27.460 | But at the same time, I at least hope

01:22:31.060 | that it's worthwhile having people around

01:22:33.300 | who can kind of see the connections between these two.

01:22:37.220 | - Yeah, the emergent intelligence of the community

01:22:41.060 | when it's nicely distributed is useful.

01:22:44.060 | Okay, so-- - Exactly, yeah.

01:22:45.180 | So hopefully, I mean, I've seen that work out well

01:22:48.220 | at DeepMind.

01:22:50.860 | There are people who, I mean, even if you just focus

01:22:53.500 | on the AI work that happens at DeepMind,

01:22:56.460 | it's been a good thing to have some people around

01:23:00.140 | doing that kind of work whose PhDs are in neuroscience

01:23:03.860 | or psychology.

01:23:05.340 | Every academic discipline has its kind of blind spots

01:23:10.340 | and kind of unfortunate obsessions

01:23:15.360 | and its metaphors and its reference points.

01:23:18.340 | And having some intellectual diversity is really healthy.

01:23:23.340 | People get each other unstuck, I think.

01:23:28.460 | I see it all the time at DeepMind.

01:23:30.660 | And I like to think that the people

01:23:33.100 | who bring some neuroscience background to the table

01:23:35.980 | are helping with that.

01:23:37.500 | - So one of my, probably the deepest passion for me,

01:23:41.460 | what I would say, maybe we kind of spoke off mic

01:23:44.180 | a little bit about it, but that I think

01:23:48.420 | is a blind spot for at least robotics and AI folks

01:23:51.420 | is human-robot interaction, human-agent interaction.

01:23:55.620 | Maybe, do you have thoughts about how we reduce

01:24:00.620 | the size of that blind spot?

01:24:03.060 | Do you also share the feeling that not enough folks

01:24:07.500 | are studying this aspect of interaction?

01:24:10.300 | - Well, I'm actually pretty intensively interested

01:24:14.580 | in this issue now.

01:24:15.920 | And there are people in my group who've actually pivoted

01:24:19.320 | pretty hard over the last few years

01:24:20.980 | from doing more traditional cognitive psychology

01:24:24.220 | and cognitive neuroscience to doing experimental work

01:24:28.100 | on human-agent interaction.

01:24:30.260 | And there are a couple of reasons

01:24:32.540 | that I'm pretty passionately interested in this.

01:24:35.540 | One is, it's kind of the outcome of having thought

01:24:42.500 | for a few years now about what we're up to.

01:24:46.960 | Like, what are we doing?

01:24:49.400 | Like, what is this AI research for?

01:24:53.480 | So what does it mean to make the world a better place?

01:24:57.060 | I think, I'm pretty sure that means

01:24:59.040 | making life better for humans.

01:25:00.520 | And so how do you make life better for humans?

01:25:05.800 | That's a proposition that when you look at it carefully

01:25:10.560 | and honestly is rather horrendously complicated,

01:25:15.560 | especially when the AI systems that you're building

01:25:20.880 | are learning systems.

01:25:25.240 | They're not, you're not programming something

01:25:29.080 | that you then introduce to the world

01:25:31.440 | and it just works as programmed,

01:25:33.160 | like Google Maps or something.

01:25:34.860 | We're building systems that learn from experience.

01:25:39.720 | So that typically leads to AI safety questions.

01:25:43.480 | How do we keep these things from getting out of control?

01:25:45.440 | How do we keep them from doing things that harm humans?

01:25:49.080 | And I mean, I hasten to say,

01:25:51.800 | I consider those hugely important issues.

01:25:54.480 | And there are large sectors of the research community

01:25:58.880 | at DeepMind and of course elsewhere,

01:26:00.800 | who are dedicated to thinking hard all day,

01:26:03.480 | every day about that.

01:26:04.960 | But there's, I guess I would say a positive side to this too

01:26:09.600 | which is to say, well,

01:26:11.160 | what would it mean to make human life better?

01:26:15.880 | And how can we imagine learning systems doing that?

01:26:20.100 | And in talking to my colleagues about that,

01:26:23.460 | we reached the initial conclusion that

01:26:26.160 | it's not sufficient to philosophize about that.

01:26:30.080 | You actually have to take into account

01:26:32.000 | how humans actually work and what humans want

01:26:37.840 | and the difficulties of knowing what humans want

01:26:40.440 | and the difficulties that arise

01:26:43.760 | when humans want different things.

01:26:46.300 | And so human agent interaction has become

01:26:50.760 | a quite intensive focus of my group lately.

01:26:55.080 | If for no other reason that,

01:26:58.060 | in order to really address that issue in an adequate way,

01:27:04.040 | you have to, I mean, psychology becomes part of the picture.

01:27:07.640 | - And so there's a few elements there.

01:27:10.400 | So if you focus on solving,

01:27:12.000 | if you focus on the robotics problem,

01:27:14.720 | let's say AGI without humans in the picture,

01:27:17.800 | you're missing fundamentally the final step.

01:27:22.340 | When you do want to help human civilization,

01:27:24.600 | you eventually have to interact with humans.

01:27:27.360 | And when you create a learning system, just as you said,

01:27:31.400 | that will eventually have to interact with humans,

01:27:34.400 | the interaction itself has to become

01:27:39.040 | part of the learning process.

01:27:40.800 | So you can't just watch, well, my sense is,

01:27:43.840 | it sounds like your sense is,

01:27:45.120 | you can't just watch humans to learn about humans.

01:27:48.280 | You have to also be part of the human world.

01:27:50.280 | You have to interact with humans.

01:27:51.440 | - Yeah, exactly.

01:27:52.280 | And I mean, then questions arise that start imperceptibly,

01:27:57.280 | but inevitably to slip beyond the realm of engineering.

01:28:02.420 | So questions like, if you have an agent

01:28:06.000 | that can do something that you can't do,

01:28:08.980 | under what conditions do you want that agent to do it?

01:28:13.820 | So if I have a robot that can play Beethoven sonatas

01:28:18.820 | better than any human, in the sense that the sensitivity,

01:28:30.760 | the expression is just beyond what any human,

01:28:33.980 | do I wanna listen to that?

01:28:36.340 | Do I wanna go to a concert and hear a robot play?

01:28:38.820 | These aren't engineering questions.

01:28:41.360 | These are questions about human preference

01:28:44.340 | and human culture.

01:28:45.980 | - Psychology bordering on philosophy.

01:28:47.900 | - Yeah.

01:28:48.980 | And then you start asking,

01:28:50.260 | well, even if we knew the answer to that,

01:28:54.660 | is it our place as AI engineers

01:28:57.060 | to build that into these agents?

01:28:59.180 | Probably the agents should interact with humans

01:29:02.140 | beyond the population of AI engineers

01:29:05.620 | and figure out what those humans want.

01:29:07.800 | And then when you start, I referred this the moment ago,

01:29:11.800 | but even that becomes complicated.

01:29:14.340 | Be quote, what if two humans want different things

01:29:19.100 | and you have only one agent that's able to interact with them

01:29:22.380 | and try to satisfy their preferences?

01:29:24.620 | Then you're into the realm of

01:29:26.940 | of like economics and social choice theory

01:29:31.780 | and even politics.

01:29:33.660 | So there's a sense in which,

01:29:35.540 | if you kind of follow what we're doing

01:29:37.980 | to its logical conclusion,

01:29:39.940 | then it goes beyond questions of engineering and technology

01:29:44.940 | and starts to shade in perceptibly into questions about

01:29:49.460 | what kind of society do you want?

01:29:51.660 | And actually that, once that dawned on me,

01:29:55.740 | I actually felt,

01:29:57.300 | I don't know what the right word is,

01:29:59.820 | quite refreshed in my involvement in AI research.

01:30:03.020 | It was almost like building this kind of stuff

01:30:06.300 | is gonna lead us back to asking

01:30:08.300 | really fundamental questions about,

01:30:10.740 | what's the good life and who gets to decide?

01:30:16.700 | And bringing in viewpoints from multiple sub-communities

01:30:23.820 | to help us shape the way that we live.

01:30:26.340 | It started making me feel like doing AI research

01:30:32.700 | in a fully responsible way

01:30:35.180 | could potentially lead to a kind of cultural renewal.

01:30:42.860 | - Yeah, it's the way to understand human beings

01:30:47.860 | at the individual, the societal level.

01:30:50.340 | It may become a way to answer all the

01:30:52.900 | silly human questions of the meaning of life

01:30:54.860 | and all those kinds of things.

01:30:57.100 | - Even if it doesn't give us a way

01:30:58.100 | of answering those questions,

01:30:59.260 | it may force us back to thinking about them.

01:31:02.980 | - Thinking about them.

01:31:03.820 | - And it might restore a certain, I don't know,

01:31:07.700 | a certain depth to,

01:31:10.500 | or even dare I say, spirituality to the world.

01:31:15.500 | I don't know, maybe that's too grandiose.

01:31:19.420 | - Well, I'm with you.

01:31:21.060 | I think AI will be the philosophy of the 21st century,

01:31:26.060 | the way which will open the door.

01:31:28.980 | I think a lot of AI researchers are afraid to open that door

01:31:32.420 | of exploring the beautiful richness

01:31:35.620 | of the human-agent interaction, human-AI interaction.

01:31:39.500 | I'm really happy that somebody like you

01:31:42.340 | have opened that door.

01:31:44.540 | - One thing I often think about is the usual schema

01:31:49.460 | for thinking about human-agent interaction

01:31:54.460 | is this kind of dystopian,

01:31:56.220 | oh, our robot overlords.

01:32:00.540 | And again, I hasten to say AI safety is hugely important.

01:32:03.580 | And I'm not saying we shouldn't be thinking

01:32:06.500 | about those risks, totally on board for that.

01:32:09.620 | But there's,

01:32:10.460 | having said that,

01:32:14.100 | what often follows for me is the thought

01:32:18.940 | that there's another kind of narrative

01:32:23.060 | that might be relevant, which is when we think of humans

01:32:28.060 | gaining more and more information about human life,

01:32:33.340 | the narrative there is usually that they gain

01:32:37.020 | more and more wisdom and they get closer to enlightenment

01:32:40.780 | and they become more benevolent.

01:32:43.860 | Like the Buddha is like,

01:32:45.540 | that's a totally different narrative.

01:32:47.420 | And why isn't it the case that we imagine

01:32:50.460 | that the AI systems that we're creating

01:32:52.540 | are just gonna, they're gonna figure out more and more

01:32:54.580 | about the way the world works

01:32:55.700 | and the way that humans interact

01:32:56.860 | and they'll become beneficent.

01:32:59.220 | I'm not saying that will happen.

01:33:00.820 | I don't honestly expect that to happen

01:33:05.460 | without some careful, setting things up very carefully.

01:33:08.860 | But it's another way things could go, right?

01:33:11.780 | - Yeah, and I would even push back on that.

01:33:14.180 | I personally believe that the most trajectories,

01:33:19.180 | natural human trajectories will lead us towards progress.

01:33:25.460 | So for me, there is a kind of sense

01:33:28.420 | that most trajectories in AI development

01:33:30.860 | will lead us into trouble.

01:33:32.580 | To me, and we over-focus on the worst case.

01:33:37.180 | It's like in computer science,

01:33:38.500 | theoretical computer science has been

01:33:40.220 | this focus on worst case analysis.

01:33:42.100 | There's something appealing to our human mind

01:33:45.220 | at some lowest level to be,

01:33:47.660 | I mean, we don't wanna be eaten by the tiger, I guess.

01:33:50.260 | So we wanna do the worst case analysis.

01:33:52.340 | But the reality is that shouldn't stop us

01:33:55.700 | from actually building out all the other trajectories

01:33:58.660 | which are potentially leading to all the positive worlds,

01:34:02.060 | all the enlightenment.

01:34:04.580 | There's a book, "Enlightenment Now"

01:34:05.740 | with Steven Pinker and so on.

01:34:07.020 | This is looking generally at human progress.

01:34:09.700 | And there's so many ways that human progress

01:34:12.300 | can happen with AI.

01:34:13.900 | And I think you have to do that research.

01:34:16.340 | You have to do that work.

01:34:17.380 | You have to do the, not just the AI safety work

01:34:20.740 | of the one worst case analysis, how do we prevent that?

01:34:23.500 | But the actual tools and the glue

01:34:27.580 | and the mechanisms of human AI interaction

01:34:31.340 | that would lead to all the positive actions that can go.

01:34:34.180 | - Yeah, right.

01:34:35.020 | - It's a super exciting area, right?

01:34:36.540 | - Yeah, we should be spending,

01:34:38.340 | we should be spending a lot of our time saying

01:34:40.820 | what can go wrong?

01:34:42.860 | I think it's harder to see that there's work to be done

01:34:47.860 | to bring into focus the question of what it would look like

01:34:51.540 | for things to go right.

01:34:53.020 | That's not obvious.

01:34:56.460 | And we wouldn't be doing this

01:34:58.820 | if we didn't have the sense there was huge potential.

01:35:01.980 | We're not doing this for no reason.

01:35:05.940 | We have a sense that AGI would be a major boom to humanity.

01:35:10.300 | But I think it's worth starting now,

01:35:13.940 | even when our technology is quite primitive,

01:35:15.860 | asking, well, exactly what would that mean?

01:35:19.660 | We can start now with applications

01:35:21.260 | that are already gonna make the world a better place,

01:35:22.780 | like solving protein folding.

01:35:25.260 | I think this deep mind has gotten heavy

01:35:28.100 | into science applications lately,

01:35:30.260 | which I think is a wonderful, wonderful move

01:35:34.620 | for us to be making.

01:35:36.260 | But when we think about AGI,

01:35:37.460 | when we think about building fully intelligent agents

01:35:41.100 | that are gonna be able to, in a sense,

01:35:42.700 | do whatever they want,

01:35:44.060 | we should start thinking about

01:35:46.980 | what do we want them to want?

01:35:49.220 | What kind of world do we wanna live in?

01:35:52.460 | That's not an easy question.

01:35:54.540 | And I think we just need to start working on it.

01:35:56.940 | - And even on the path to sort of,

01:35:58.780 | it doesn't have to be AGI,

01:36:00.100 | but just intelligent agents that interact with us

01:36:02.500 | and help us enrich our own existence on social networks,

01:36:06.460 | for example, on recommender systems, various intelligent,

01:36:09.060 | there's so much interesting interaction

01:36:10.580 | that's yet to be understood and studied.

01:36:12.300 | And how do you create,

01:36:14.740 | I mean, Twitter is struggling with this very idea,

01:36:19.500 | how do you create AI systems

01:36:21.460 | that increase the quality and the health of a conversation?

01:36:24.420 | - For sure, yeah.

01:36:25.260 | - That's a beautiful, beautiful human psychology question.

01:36:28.540 | - And how do you do that

01:36:29.780 | without deception being involved,

01:36:34.780 | without manipulation being involved,

01:36:37.860 | maximizing human autonomy?

01:36:42.060 | And how do you make these choices in a democratic way?

01:36:46.820 | How do we face the,

01:36:50.980 | again, I'm speaking for myself here.

01:36:53.500 | How do we face the fact that

01:36:55.380 | it's a small group of people

01:36:58.540 | who have the skillset to build these kinds of systems,

01:37:02.260 | but what it means to make the world a better place

01:37:08.060 | is something that we all have to be talking about.

01:37:11.180 | - Yeah, the world that we're trying to make a better place

01:37:16.180 | includes a huge variety of different kinds of people.

01:37:20.300 | - Yeah, how do we cope with that?

01:37:22.020 | This is a problem that has been discussed

01:37:25.260 | in gory, extensive detail

01:37:27.660 | in social choice theory.

01:37:30.740 | One thing I'm really enjoying

01:37:32.900 | about the recent direction work has taken

01:37:35.180 | in some parts of my team is that,

01:37:36.900 | yeah, we're reading the AI literature,

01:37:38.580 | we're reading the neuroscience literature,

01:37:39.940 | but we've also started reading economics

01:37:42.940 | and, as I mentioned, social choice theory,

01:37:44.820 | even some political theory,

01:37:45.940 | because it turns out that it all becomes relevant.

01:37:50.380 | It all becomes relevant.

01:37:51.580 | But at the same time,

01:37:55.620 | we've been trying not to write philosophy papers, right?

01:38:00.140 | We've been trying not to write physician papers.

01:38:01.980 | We're trying to figure out ways

01:38:03.780 | of doing actual empirical research

01:38:05.740 | that kind of take the first small steps

01:38:07.780 | to thinking about what it really means

01:38:10.820 | for humans with all of their complexity and contradiction

01:38:14.660 | and paradox to be brought into contact

01:38:19.660 | with these AI systems in a way

01:38:22.660 | that really makes the world a better place.

01:38:25.500 | - And often reinforcement learning frameworks

01:38:27.500 | actually kind of allow you to do that machine learning.

01:38:31.580 | And so that's the exciting thing about AI

01:38:33.540 | is it allows you to reduce the unsolvable problem,

01:38:37.220 | philosophical problem,

01:38:38.180 | into something more concrete that you can get a hold of.

01:38:41.660 | - Yeah, and it allows you to kind of define the problem

01:38:43.860 | in some way that allows for growth in the system

01:38:48.860 | that's sort of, you know,

01:38:51.100 | you're not responsible for the details, right?

01:38:54.060 | You say, this is generally what I want you to do.

01:38:56.660 | And then learning takes care of the rest.

01:38:59.540 | Of course, the safety issues are, you know,

01:39:02.340 | arise in that context.

01:39:04.060 | But I think also some of these positive issues

01:39:05.940 | arise in that context.

01:39:06.900 | What would it mean for an AI system

01:39:09.140 | to really come to understand what humans want?

01:39:12.660 | And, you know, with all of the subtleties of that, right?

01:39:18.980 | You know, humans want help with certain things.

01:39:24.740 | But they don't want everything done for them, right?

01:39:27.500 | There is part of the satisfaction that humans get from life

01:39:30.620 | is in accomplishing things.

01:39:32.780 | So if there were devices around that did everything for,

01:39:34.740 | you know, I often think of the movie "Wall-E", right?

01:39:37.580 | That's like dystopian in a totally different way.

01:39:39.420 | It's like, the machines are doing everything for us.

01:39:41.380 | That's not what we wanted.

01:39:43.860 | You know, anyway, I just, I find this, you know,

01:39:46.780 | this opens up a whole landscape of research

01:39:50.620 | that feels affirmative and exciting.

01:39:52.860 | - To me, it's one of the most exciting and it's wide open.

01:39:56.140 | We have to, 'cause it's a cool paper,

01:39:58.380 | talk about dopamine.

01:39:59.420 | - Oh yeah, okay, so I can.

01:40:01.260 | We were gonna, we were gonna,

01:40:02.940 | I was gonna give you a quick summary.

01:40:05.100 | - Yeah, a quick summary of, what's the title of the paper?

01:40:10.100 | - I think we called it a distributional code for value

01:40:15.700 | in dopamine-based reinforcement learning.

01:40:18.980 | - Yes.

01:40:19.820 | So that's another project that grew out of pure AI research.

01:40:24.820 | A number of people at DeepMind and a few other places

01:40:30.340 | had started working on a new version

01:40:33.060 | of reinforcement learning, which was defined

01:40:37.420 | by taking something in traditional reinforcement learning

01:40:41.340 | and just tweaking it.

01:40:42.180 | So the thing that they took

01:40:43.460 | from traditional reinforcement learning was a value signal.

01:40:47.580 | So at the center of reinforcement learning,

01:40:50.260 | at least most algorithms, is some representation

01:40:53.260 | of how well things are going,

01:40:54.860 | your expected cumulative future reward.

01:40:58.300 | And that's usually represented as a single number.

01:41:01.900 | So if you imagine a gambler in a casino

01:41:04.940 | and the gambler's thinking, well,

01:41:07.620 | I have this probability of winning

01:41:09.140 | such and such an amount of money,

01:41:10.220 | and I have this probability of losing

01:41:11.500 | such and such an amount of money,

01:41:13.780 | that situation would be represented as a single number,

01:41:16.100 | which is like the expected,

01:41:17.780 | the weighted average of all those outcomes.

01:41:20.060 | And this new form of reinforcement learning said,

01:41:24.260 | well, what if we generalize that

01:41:26.980 | to a distributional representation?

01:41:28.660 | So now we think of the gambler as literally thinking,

01:41:31.340 | well, there's this probability

01:41:32.780 | that I'll win this amount of money,

01:41:34.100 | and there's this probability

01:41:35.100 | that I'll lose that amount of money,

01:41:36.220 | and we don't reduce that to a single number.

01:41:38.380 | And it had been observed through experiments,

01:41:42.260 | through just trying this out,

01:41:44.180 | that that kind of distributional representation

01:41:47.700 | really accelerated reinforcement learning

01:41:51.380 | and led to better policies.

01:41:54.180 | - What's your intuition about,

01:41:55.420 | so we're talking about rewards.

01:41:56.860 | - Yeah. - So what's your intuition

01:41:58.620 | why that is?

01:41:59.460 | Why does it depend? - Well, it's kind of

01:42:00.980 | a surprising historical note,

01:42:04.420 | at least surprised me when I learned it,

01:42:06.140 | that this had been tried out in a kind of heuristic way.

01:42:09.820 | People thought, well, gee, what would happen if we tried?

01:42:12.580 | And then it had this, empirically,

01:42:14.540 | it had this striking effect.

01:42:17.260 | And it was only then that people started thinking,

01:42:19.300 | well, gee, wait, why?

01:42:20.660 | Why?

01:42:21.500 | Wait, why?

01:42:22.340 | Why is this working?

01:42:23.380 | And that's led to a series of studies

01:42:26.140 | just trying to figure out why it works, which is ongoing.

01:42:29.700 | But one thing that's already clear from that research

01:42:31.740 | is that one reason that it helps

01:42:34.300 | is that it drives richer representation learning.

01:42:39.380 | So if you imagine two situations

01:42:43.020 | that have the same expected value,

01:42:45.260 | the same kind of weighted average value,

01:42:47.260 | standard deep reinforcement learning algorithms

01:42:51.260 | are going to take those two situations

01:42:53.460 | and kind of, in terms of the way

01:42:54.980 | they're represented internally,

01:42:56.420 | they're gonna squeeze them together.

01:42:58.140 | Because the thing that you're trying to represent,

01:43:02.540 | which is their expected value, is the same.

01:43:04.160 | So all the way through the system,

01:43:06.220 | things are gonna be mushed together.

01:43:08.380 | But what if those two situations

01:43:11.020 | actually have different value distributions?

01:43:13.920 | They have the same average value,

01:43:16.880 | but they have different distributions of value.

01:43:19.860 | In that situation, distributional learning

01:43:22.300 | will maintain the distinction between these two things.

01:43:25.060 | So to make a long story short,

01:43:26.820 | distributional learning can keep things separate

01:43:30.000 | in the internal representation

01:43:32.140 | that might otherwise be conflated or squished together.

01:43:35.100 | And maintaining those distinctions

01:43:36.380 | can be useful when the system

01:43:39.060 | is now faced with some other task

01:43:41.100 | where the distinction is important.

01:43:43.260 | - If we look at the optimistic

01:43:44.540 | and pessimistic dopamine neurons.

01:43:46.540 | So first of all, what is dopamine?

01:43:49.560 | - Oh, right. - And why is this,

01:43:53.260 | why is this at all useful to think about

01:43:56.260 | in the artificial intelligence sense?

01:44:00.740 | But what do we know about dopamine in the human brain?

01:44:04.180 | What is it, why is it useful, why is it interesting?

01:44:07.460 | What does it have to do with the prefrontal cortex

01:44:09.380 | and learning in general?

01:44:10.300 | - Yeah, so, well, this is also a case

01:44:15.300 | where there's a huge amount of detail and debate.

01:44:19.660 | But one currently prevailing idea

01:44:24.660 | is that the function of this neurotransmitter dopamine

01:44:29.060 | resembles a particular component

01:44:33.460 | of standard reinforcement learning algorithms,

01:44:36.860 | which is called the reward prediction error.

01:44:39.860 | So I was talking a moment ago

01:44:41.580 | about these value representations.

01:44:44.200 | How do you learn them?

01:44:45.160 | How do you update them based on experience?

01:44:46.880 | Well, if you made some prediction about future reward,

01:44:51.820 | and then you get more reward than you were expecting,

01:44:54.420 | then probably retrospectively,

01:44:55.980 | you wanna go back and increase the value representation

01:45:00.700 | that you attached to that earlier situation.

01:45:03.780 | If you got less reward than you were expecting,

01:45:06.140 | you should probably decrement that estimate.

01:45:08.460 | - And that's the process of temporal difference.

01:45:10.220 | - Exactly, this is the central mechanism

01:45:11.960 | of temporal difference learning,

01:45:12.820 | which is one of several kind of,

01:45:15.180 | you know, kind of back,

01:45:16.620 | sort of the backbone of our armamentarium in RL.

01:45:20.360 | And it was this connection

01:45:22.540 | between the reward prediction error and dopamine

01:45:25.900 | was made, you know, in the 1990s.

01:45:30.900 | And there's been a huge amount of research

01:45:33.420 | that, you know, seems to back it up.

01:45:35.860 | Dopamine may be doing other things,

01:45:37.300 | but this is clearly, at least roughly,

01:45:39.860 | one of the things that it's doing.

01:45:42.420 | But the usual idea was that dopamine

01:45:45.100 | was representing these reward prediction errors,

01:45:48.020 | again, in this like kind of single number way,

01:45:52.860 | representing your surprise with a single number.

01:45:56.740 | And in distributional reinforcement learning,

01:45:58.540 | this kind of new elaboration of the standard approach,

01:46:02.820 | it's not only the value function

01:46:06.100 | that's represented as a single number,

01:46:08.500 | it's also the reward prediction error.

01:46:10.980 | And so what happened was that Will Dabney,

01:46:15.980 | one of my collaborators,

01:46:17.380 | who was one of the first people to work

01:46:19.280 | on distributional temporal difference learning,

01:46:22.340 | talked to a guy in my group, Zeb Kurth-Nelson,

01:46:25.780 | who's a computational neuroscientist,

01:46:27.700 | and said, "Gee, you know, is it possible

01:46:29.580 | that dopamine might be doing something

01:46:31.740 | like this distributional coding thing?"

01:46:33.460 | And they started looking at what was in the literature,

01:46:36.000 | and then they brought me in,

01:46:37.100 | and we started talking to Nao Uchida,

01:46:39.220 | and we came up with some specific predictions about,

01:46:42.140 | you know, if the brain is using

01:46:43.540 | this kind of distributional coding,

01:46:45.180 | then in the tasks that Nao has studied,

01:46:47.380 | you should see this, this, this, and this.

01:46:49.340 | And that's where the paper came from.

01:46:50.660 | We kind of enumerated a set of predictions,

01:46:53.580 | all of which ended up being fairly clearly confirmed,

01:46:56.440 | and all of which leads to at least some initial indication

01:47:00.780 | that the brain might be doing something

01:47:02.220 | like this distributional coding,

01:47:03.460 | that dopamine might be representing surprise signals

01:47:06.820 | in a way that is not just collapsing everything

01:47:10.020 | to a single number, but instead is kind of respecting

01:47:12.220 | the variety of future outcomes, if that makes sense.

01:47:16.660 | - So yeah, so that's showing, suggesting possibly

01:47:19.620 | that dopamine has a really interesting representation scheme

01:47:23.500 | in the human brain for its reward signal.

01:47:27.700 | - Exactly. - That's fascinating.

01:47:29.020 | It's just, that's another beautiful example

01:47:31.780 | of AI revealing something nice about neuroscience.

01:47:34.500 | Potentially, suggesting possibilities.

01:47:36.300 | - Well, you never know, so the minute you publish a paper

01:47:38.900 | like that, the next thing you think is,

01:47:40.940 | "I hope that replicates, like, I hope we see

01:47:43.700 | "that same thing in other data sets."

01:47:44.980 | But of course, several labs now

01:47:47.420 | are doing the follow-up experiment,

01:47:49.180 | so we'll know soon.

01:47:50.020 | - That'd be interesting. - But it has been,

01:47:50.900 | it has been a lot of fun for us to take these ideas from AI

01:47:54.780 | and kind of bring them into neuroscience

01:47:56.780 | and see how far we can get.

01:47:58.940 | - So we kind of talked about it a little bit,

01:48:01.300 | but where do you see the field of neuroscience

01:48:03.980 | and artificial intelligence heading broadly?

01:48:07.740 | Like, what are the possible exciting areas

01:48:12.540 | that you can see breakthroughs in the next, let's get crazy,

01:48:16.340 | not just three or five years, but next 10, 20, 30 years

01:48:20.020 | that would make you excited and perhaps you'd be part of?

01:48:27.300 | - On the neuroscience side,

01:48:31.060 | there's a great deal of interest now

01:48:34.380 | in what's going on in AI.

01:48:36.740 | And at the same time, I feel like,

01:48:44.380 | so neuroscience, especially the part of neuroscience

01:48:49.380 | that's focused on circuits and systems,

01:48:54.180 | kind of like really mechanism-focused,

01:48:56.340 | there's been this explosion in new technology.

01:49:01.980 | And up until recently, the experiments

01:49:06.980 | that have exploited this technology

01:49:08.980 | have not involved a lot of interesting behavior.

01:49:13.380 | And this is for a variety of reasons,

01:49:15.500 | one of which is in order to employ

01:49:18.740 | some of these technologies, you actually have to,

01:49:21.060 | if you're studying a mouse, you have to head fix the mouse.

01:49:23.660 | In other words, you have to immobilize the mouse.

01:49:26.340 | And so it's been tricky to come up with ways

01:49:29.420 | of eliciting interesting behavior

01:49:30.940 | from a mouse that's restrained in this way.

01:49:33.500 | But people have begun to create

01:49:35.700 | very interesting solutions to this,

01:49:39.540 | like virtual reality environments

01:49:41.340 | where the animal can kind of move a trackball

01:49:43.220 | and as people have kind of begun to explore

01:49:48.220 | what you can do with these technologies,

01:49:50.300 | I feel like more and more people are asking,

01:49:52.860 | well, let's try to bring behavior into the picture.

01:49:55.780 | Let's try to like reintroduce behavior,

01:49:58.260 | which was supposed to be what this whole thing was about.

01:50:01.060 | And I'm hoping that those two trends,

01:50:06.660 | the kind of growing interest in behavior

01:50:10.180 | and the widespread interest in what's going on in AI,

01:50:15.020 | will come together to kind of open a new chapter

01:50:18.540 | in neuroscience research where there's a kind of

01:50:22.780 | a rebirth of interest in the structure of behavior

01:50:26.700 | and its underlying substrates,

01:50:28.420 | but that that research is being informed

01:50:32.100 | by computational mechanisms

01:50:34.460 | that we're coming to understand in AI.

01:50:36.460 | If we can do that, then we might be taking

01:50:39.620 | a step closer to this utopian future

01:50:42.540 | that we were talking about earlier,

01:50:43.980 | where there's really no distinction

01:50:45.540 | between psychology and neuroscience.

01:50:47.980 | Neuroscience is about studying the mechanisms

01:50:51.580 | that underlie whatever it is the brain is for,

01:50:55.340 | and what is the brain for?

01:50:56.980 | It's for behavior.

01:50:58.420 | I feel like we could maybe take a step toward that now

01:51:03.060 | if people are motivated in the right way.

01:51:05.160 | You also asked about AI.

01:51:08.700 | So that was a neuroscience question.

01:51:10.380 | - You said neuroscience, that's right.

01:51:12.140 | And especially a place like DeepMind,

01:51:13.700 | I'm interested in both branches.

01:51:15.220 | So what about the engineering of intelligence systems?

01:51:18.660 | - I think one of the key challenges

01:51:24.860 | that a lot of people are seeing now in AI

01:51:28.700 | is to build systems that have the kind of flexibility,

01:51:33.140 | and the kind of flexibility that humans have in two senses.

01:51:38.580 | One is that humans can be good at many things.

01:51:41.860 | They're not just expert at one thing.

01:51:44.300 | And they're also flexible in the sense

01:51:45.620 | that they can switch between things very easily,

01:51:49.660 | and they can pick up new things very quickly

01:51:52.060 | because they very ably see what a new task has in common

01:51:57.060 | with other things that they've done.

01:51:59.380 | And that's something that our AI systems

01:52:05.340 | just blatantly do not have.

01:52:09.020 | There are some people who like to argue

01:52:11.300 | that deep learning and deep RL

01:52:13.660 | are simply wrong for getting that kind of flexibility.

01:52:16.980 | I don't share that belief,

01:52:19.940 | but the simpler fact of the matter

01:52:22.500 | is we're not building things yet

01:52:23.740 | that do have that kind of flexibility.

01:52:25.380 | And I think the attention of a large part

01:52:28.580 | of the AI community is starting to pivot to that question.

01:52:31.380 | How do we get that?

01:52:32.340 | That's gonna lead to a focus on abstraction.

01:52:37.980 | It's gonna lead to a focus on what in psychology

01:52:41.740 | we call cognitive control,

01:52:43.540 | which is the ability to switch between tasks,

01:52:45.840 | the ability to quickly put together a program of behavior

01:52:49.220 | that you've never executed before,

01:52:51.720 | but you know makes sense for a particular set of demands.

01:52:55.260 | It's very closely related to what the prefrontal cortex does

01:52:59.100 | on the neuroscience side.

01:53:01.060 | So I think it's gonna be an interesting new chapter.

01:53:05.380 | - So that's the reasoning side and cognition side,

01:53:07.380 | but let me ask the over romanticized question.

01:53:10.540 | Do you think we'll ever engineer an AGI system

01:53:13.660 | that we humans would be able to love

01:53:17.140 | and then would love us back?

01:53:18.640 | So have that level and depth of connection.

01:53:23.060 | - I love that question.

01:53:27.860 | And it relates closely to things

01:53:31.980 | that I've been thinking about a lot lately,

01:53:33.940 | in the context of this human AI research.

01:53:36.100 | There's social psychology research,

01:53:40.020 | in particular by Susan Fisk at Princeton,

01:53:44.940 | in the department where I used to work,

01:53:47.140 | where she dissects human attitudes

01:53:52.740 | toward other humans into a sort of two-dimensional,

01:53:56.580 | two-dimensional scheme.

01:53:59.940 | And one dimension is about ability.

01:54:03.980 | How able, how capable is this other person?

01:54:08.280 | But the other dimension is warmth.

01:54:12.740 | So you can imagine another person

01:54:15.420 | who's very skilled and capable, but is very cold.

01:54:18.640 | And you wouldn't really highly,

01:54:23.460 | you might have some reservations

01:54:24.900 | about that other person.

01:54:26.100 | But there's also a kind of reservation

01:54:30.020 | that we might have about another person

01:54:31.980 | who elicits in us or displays a lot of human warmth,

01:54:35.860 | but is not good at getting things done.

01:54:38.820 | We reserve our greatest esteem, really,

01:54:45.020 | for people who are both highly capable

01:54:47.980 | and also quite warm.

01:54:50.900 | That's like the best of the best.

01:54:53.700 | This isn't a normative statement I'm making.

01:54:57.380 | This is just an empirical statement.

01:54:59.780 | This is what humans seem.

01:55:01.340 | These are the two dimensions that people seem

01:55:02.860 | to kind of like, along which people size other people up.

01:55:06.740 | And in AI research, we really focus on this capability thing.

01:55:10.680 | We want our agents to be able to do stuff.

01:55:12.700 | This thing can play Go at a superhuman level.

01:55:14.740 | That's awesome.

01:55:16.180 | But that's only one dimension.

01:55:17.820 | What about the other dimension?

01:55:19.340 | What would it mean for an AI system to be warm?

01:55:23.900 | And I don't know, maybe there are easy solutions here.

01:55:26.820 | Like we can put a face on our AI systems.

01:55:29.820 | It's cute, it has big ears.

01:55:31.180 | I mean, that's probably part of it.

01:55:33.060 | But I think it also has to do with a pattern of behavior,

01:55:36.380 | a pattern of, what would it mean for an AI system

01:55:40.140 | to display caring, compassionate behavior

01:55:43.380 | in a way that actually made us feel like it was for real?

01:55:47.700 | That we didn't feel like it was simulated.

01:55:49.900 | We didn't feel like we were being duped.

01:55:51.900 | To me, people talk about the Turing test

01:55:55.660 | or some descendant of it.

01:55:57.820 | I feel like that's the ultimate Turing test.

01:56:01.100 | Is there an AI system that can not only convince us

01:56:05.420 | that it knows how to reason

01:56:07.140 | and it knows how to interpret language,

01:56:09.060 | but that we're comfortable saying,

01:56:12.660 | yeah, that AI system's a good guy.

01:56:14.560 | - On the warmth scale, whatever warmth is,

01:56:18.700 | we kind of intuitively understand it,

01:56:20.820 | but we also wanna be able to, yeah,

01:56:25.060 | we don't understand it explicitly enough yet

01:56:29.140 | to be able to engineer it.

01:56:30.900 | - Exactly.

01:56:32.060 | - And that's an open scientific question.

01:56:33.620 | You kind of alluded to it several times

01:56:35.320 | in the human-AI interaction.

01:56:37.220 | That's a question that should be studied

01:56:38.900 | and probably one of the most important questions

01:56:42.100 | as we move to AGI.

01:56:43.660 | - We humans are so good at it.

01:56:45.980 | - Yeah, weird.

01:56:47.660 | - It's not just that we're born warm.

01:56:50.140 | I suppose some people are warmer than others

01:56:53.060 | given whatever genes they manage to inherit.

01:56:55.720 | But there are also learned skills involved.

01:57:01.640 | There are ways of communicating to other people

01:57:04.760 | that you care, that they matter to you,

01:57:07.800 | that you're enjoying interacting with them.

01:57:09.900 | And we learn these skills from one another.

01:57:14.160 | And it's not out of the question

01:57:16.760 | that we could build engineered systems.

01:57:20.080 | I think it's hopeless, as you say,

01:57:21.480 | that we could somehow hand design

01:57:23.600 | these sorts of behaviors.

01:57:26.120 | But it's not out of the question

01:57:27.120 | that we could build systems

01:57:28.480 | that kind of we instill in them

01:57:32.760 | something that sets them out in the right direction

01:57:36.000 | so that they end up learning

01:57:39.080 | what it is to interact with humans

01:57:40.560 | in a way that's gratifying to humans.

01:57:44.200 | I mean, honestly, if that's not where we're headed,

01:57:47.520 | I want out.

01:57:50.400 | (laughing)

01:57:52.240 | - I think it's exciting as a scientific problem,

01:57:54.960 | just as you described.

01:57:56.840 | I honestly don't see a better way to end it

01:57:59.480 | than talking about warmth and love.

01:58:01.160 | And Matt, I don't think I've ever had

01:58:04.000 | such a wonderful conversation

01:58:06.000 | where my questions were so bad

01:58:07.520 | and your answers were so beautiful.

01:58:09.360 | So I deeply appreciate it.

01:58:10.680 | I really enjoyed it.

01:58:11.520 | - Well, it's been very fun.

01:58:13.240 | As you can probably tell,

01:58:14.520 | there's something I like about

01:58:18.240 | kind of thinking outside the box.

01:58:21.000 | So it's good having the opportunity to do that.

01:58:22.880 | - Awesome.

01:58:23.720 | Thanks so much for doing it.

01:58:25.600 | Thanks for listening to this conversation

01:58:27.160 | with Matt Boppenegg.

01:58:28.400 | And thank you to our sponsors,

01:58:30.520 | the Jordan and Harbinger Show

01:58:32.320 | and Magic Spoon low-carb keto cereal.

01:58:36.120 | Please consider supporting this podcast

01:58:38.000 | by going to jordanharbinger.com/lex

01:58:41.040 | and also going to magicspoon.com/lex

01:58:44.920 | and using code LEX at checkout.

01:58:48.240 | Click the links, buy all the stuff.

01:58:50.900 | It's the best way to support this podcast

01:58:52.840 | and the journey I'm on in my research

01:58:55.360 | and the startup.

01:58:57.280 | If you enjoy this thing, subscribe on YouTube,

01:58:59.600 | review it with the five stars on Apple Podcasts,

01:59:02.360 | support it on Patreon, follow on Spotify

01:59:05.360 | or connect with me on Twitter @LexFriedman.

01:59:08.200 | Again, spelled miraculously without the E,

01:59:12.200 | just F-R-I-D-M-A-N.

01:59:15.040 | And now let me leave you with some words

01:59:17.080 | from urologist B.S. Samachandran.

01:59:20.800 | How can a three pound mass of jelly

01:59:23.320 | that you can hold in your palm

01:59:25.120 | imagine angels, contemplate the meaning of infinity

01:59:28.720 | and even question its own place in the cosmos?

01:59:31.720 | Especially awe-inspiring is the fact that any single brain,

01:59:35.660 | including yours, is made up of atoms

01:59:38.600 | that were forged in the hearts of countless

01:59:41.200 | far-flung stars billions of years ago.

01:59:45.480 | These particles drifted for eons and light years

01:59:48.320 | until gravity and change brought them together here now.

01:59:53.160 | These atoms now form a conglomerate, your brain,

01:59:57.560 | that can not only ponder the very stars that gave it birth,

02:00:00.840 | but can also think about its own ability to think

02:00:04.160 | and wonder about its own ability to wander.

02:00:07.800 | With the arrival of humans, it has been said,

02:00:10.640 | the universe has suddenly become conscious of itself.

02:00:14.600 | This truly is the greatest mystery of all.

02:00:19.960 | Thank you for listening and hope to see you next time.

02:00:23.320 | (upbeat music)

02:00:25.900 | (upbeat music)

02:00:28.480 | [BLANK_AUDIO]

Matt Botvinick: Neuroscience, Psychology, and AI at DeepMind | Lex Fridman Podcast #106

Chapters