Demis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity

00:00:00.000 | The following is a conversation with Demis Hassabis,

00:00:03.480 | CEO and co-founder of DeepMind,

00:00:06.720 | a company that has published and built

00:00:08.600 | some of the most incredible artificial intelligence systems

00:00:12.200 | in the history of computing,

00:00:14.120 | including AlphaZero that learned all by itself

00:00:18.040 | to play the game of Go better than any human in the world,

00:00:21.000 | and AlphaFold2 that solved protein folding.

00:00:25.760 | Both tasks considered nearly impossible for a very long time.

00:00:30.320 | Demis is widely considered to be one of the most brilliant

00:00:33.960 | and impactful humans in the history

00:00:36.560 | of artificial intelligence

00:00:38.120 | and science and engineering in general.

00:00:41.240 | This was truly an honor and a pleasure for me

00:00:44.600 | to finally sit down with him for this conversation,

00:00:47.360 | and I'm sure we will talk many times again in the future.

00:00:50.560 | This is the Lux Freedman Podcast.

00:00:53.280 | To support it, please check out our sponsors

00:00:55.480 | in the description.

00:00:56.760 | And now, dear friends, here's Demis Hassabis.

00:01:00.560 | Let's start with a bit of a personal question.

00:01:04.000 | Am I an AI program you wrote to interview people

00:01:07.800 | until I get good enough to interview you?

00:01:10.080 | - Well, I'd be impressed if you were.

00:01:13.080 | I'd be impressed with myself if you were.

00:01:14.840 | I don't think we're quite up to that yet,

00:01:16.480 | but maybe you're from the future, Lex.

00:01:18.760 | - If you did, would you tell me?

00:01:20.360 | Is that a good thing to tell a language model

00:01:23.040 | that's tasked with interviewing

00:01:25.040 | that it is in fact AI?

00:01:27.400 | - Maybe we're in a kind of meta-Turing test.

00:01:29.640 | Probably it would be a good idea not to tell you

00:01:32.400 | so it doesn't change your behavior, right?

00:01:33.880 | - This is a kind of--

00:01:34.720 | - Heisenberg uncertainty principle situation.

00:01:37.080 | If I told you, you'd behave differently.

00:01:39.040 | Maybe that's what's happening with us, of course.

00:01:40.960 | - This is a benchmark from the future

00:01:42.760 | where they replay 2022 as a year

00:01:46.520 | before AIs were good enough yet,

00:01:49.400 | and now we want to see, is it gonna pass?

00:01:52.080 | - Exactly.

00:01:53.960 | - If I was such a program,

00:01:56.000 | would you be able to tell, do you think?

00:01:57.960 | So to the Turing test question,

00:01:59.960 | you've talked about the benchmark for solving intelligence.

00:02:04.960 | What would be the impressive thing?

00:02:07.320 | You've talked about winning a Nobel Prize

00:02:09.120 | and AI system winning a Nobel Prize,

00:02:11.440 | but I still return to the Turing test as a compelling test.

00:02:14.400 | The spirit of the Turing test is a compelling test.

00:02:17.280 | - Yeah, the Turing test, of course,

00:02:18.520 | it's been unbelievably influential,

00:02:20.200 | and Turing's one of my all-time heroes,

00:02:22.120 | but I think if you look back at the 1950 papers,

00:02:25.040 | original paper and read the original,

00:02:27.000 | you'll see, I don't think he meant it

00:02:28.840 | to be a rigorous formal test.

00:02:30.880 | I think it was more like a thought experiment,

00:02:32.880 | almost a bit of philosophy he was writing

00:02:34.640 | if you look at the style of the paper.

00:02:36.440 | And you can see he didn't specify it very rigorously.

00:02:38.680 | So for example, he didn't specify the knowledge

00:02:41.840 | that the expert or judge would have,

00:02:44.040 | how much time would they have to investigate this.

00:02:48.320 | So these important parameters,

00:02:49.480 | if you were gonna make it a true sort of formal test.

00:02:53.200 | And by some measures,

00:02:56.600 | people claim the Turing test passed several,

00:02:59.040 | a decade ago, I remember someone claiming that

00:03:00.920 | with a kind of very bog standard normal logic model,

00:03:05.920 | because they pretended it was a kid.

00:03:08.440 | So the judges thought that the machine was a child.

00:03:13.280 | So that would be very different from an expert AI person

00:03:17.600 | interrogating machine and knowing how it was built

00:03:19.880 | and so on.

00:03:20.720 | So I think, we should probably move away from that

00:03:24.560 | as a formal test and move more towards a general test

00:03:28.800 | where we test the AI capabilities on a range of tasks

00:03:32.040 | and see if it reaches human level or above performance

00:03:35.320 | on maybe thousands,

00:03:37.160 | perhaps even millions of tasks eventually

00:03:39.200 | and cover the entire sort of cognitive space.

00:03:41.920 | So I think for its time,

00:03:44.120 | it was an amazing thought experiment and also 1950s,

00:03:46.960 | obviously it was barely the dawn of the computer age.

00:03:49.440 | So of course he only thought about text

00:03:51.480 | and now we have a lot more different inputs.

00:03:54.600 | - So yeah, maybe the better thing to test

00:03:57.080 | is the generalizability.

00:03:58.360 | So across multiple tasks,

00:03:59.680 | but I think it's also possible as systems like God or show

00:04:04.560 | that eventually that might map right back to language.

00:04:08.320 | So you might be able to demonstrate your ability

00:04:10.800 | to generalize across tasks by then communicating

00:04:14.760 | your ability to generalize across tasks,

00:04:17.080 | which is kind of what we do through conversation anyway,

00:04:19.200 | when we jump around.

00:04:20.800 | Ultimately what's in there in that conversation

00:04:23.720 | is not just you moving around knowledge,

00:04:27.000 | it's you moving around like these entirely different

00:04:30.360 | modalities of understanding that ultimately map

00:04:34.920 | to your ability to operate successfully

00:04:38.920 | in all of these domains, which you can think of as tasks.

00:04:42.600 | - Yeah, I think certainly we as humans,

00:04:44.680 | use language as our main generalization communication tool.

00:04:48.440 | So I think we end up thinking in language

00:04:51.280 | and expressing our solutions in language.

00:04:54.320 | So it's gonna be very powerful mode

00:04:58.000 | in which to explain the system, to explain what it's doing.

00:05:03.000 | But I don't think it's the only modality that matters.

00:05:07.480 | So I think there's gonna be a lot of,

00:05:09.960 | there's a lot of different ways to express capabilities

00:05:13.400 | other than just language.

00:05:15.600 | - Yeah, visual, robotics, body language.

00:05:19.040 | Yeah, action is the interactive aspect of all that.

00:05:23.800 | That's all part of it.

00:05:24.800 | - But what's interesting with GATO is that

00:05:26.800 | it's sort of pushing prediction to the maximum

00:05:30.240 | in terms of like, mapping arbitrary sequences

00:05:33.440 | to other sequences and sort of just predicting

00:05:35.480 | what's gonna happen next.

00:05:36.440 | So prediction seems to be fundamental to intelligence.

00:05:41.040 | - And what you're predicting doesn't so much matter.

00:05:44.160 | - Yeah, it seems like you can generalize that quite well.

00:05:46.840 | So obviously language models predict the next word.

00:05:49.640 | GATO predicts potentially any action or any token.

00:05:53.880 | And it's just the beginning really.

00:05:55.320 | It's our most general agent one could call it so far.

00:05:58.080 | But that itself can be scaled up massively more

00:06:01.240 | than we've done so far.

00:06:02.160 | And obviously we're in the middle of doing that.

00:06:04.240 | - But the big part of solving AGI

00:06:06.880 | is creating benchmarks that help us get closer and closer.

00:06:11.020 | Sort of creating benchmarks that test the generalizability.

00:06:14.880 | And it's just still interesting that this fella,

00:06:17.400 | Alan Turing, was one of the first

00:06:20.480 | and probably still one of the only people

00:06:22.560 | that was trying, maybe philosophically,

00:06:25.000 | but was trying to formulate a benchmark

00:06:26.800 | that could be followed.

00:06:27.840 | It is, even though it's fuzzy,

00:06:30.920 | it's still sufficiently rigorous

00:06:32.480 | to where you can run that test.

00:06:33.960 | And I still think something like the Turing test

00:06:36.600 | will at the end of the day be the thing

00:06:39.320 | that truly impresses other humans.

00:06:42.500 | So that you can have a close friend who's in the AI system.

00:06:46.380 | For that friend to be a good friend,

00:06:48.300 | they're going to have to be able to play StarCraft.

00:06:53.120 | And they're gonna have to do all of these tasks.

00:06:56.580 | Get you a beer, so the robotics tasks.

00:06:59.540 | Play games with you.

00:07:02.060 | Use language, humor, all of those kinds of things.

00:07:04.760 | But that ultimately can boil down to language.

00:07:07.980 | It feels like, not in terms of the AI community,

00:07:11.180 | but in terms of the actual impact

00:07:13.100 | of general intelligence on the world,

00:07:14.780 | it feels like language will be the place

00:07:16.620 | where it truly shines.

00:07:18.460 | - I think so, because it's such an important

00:07:20.620 | kind of input/output for us.

00:07:22.480 | I think you're right.

00:07:23.320 | I think the Turing test, the kind of,

00:07:25.780 | the philosophy behind it, which is the idea of,

00:07:28.860 | can a machine mimic the behaviors of a human?

00:07:34.980 | And I would say wider than just language and text.

00:07:38.700 | Then, in terms of actions and everything else,

00:07:41.320 | creativity, all these things,

00:07:42.700 | then if it can sort of match or exceed

00:07:45.320 | human cognitive capabilities,

00:07:47.340 | then I think we have a true intelligence.

00:07:49.900 | So from that perspective, you're right.

00:07:51.660 | I think he did formulate the right kind of setup.

00:07:55.980 | - I just, I think there'll be a kind of humor

00:07:59.280 | in the AI systems of the future,

00:08:00.920 | looking back to this conversation,

00:08:03.540 | and thinking about the Turing test,

00:08:06.020 | and also thinking about, by that time,

00:08:09.380 | they would know which year they were finally able

00:08:11.900 | to sort of cross the threshold of human-level intelligence,

00:08:15.580 | and think how funny it is that we humans

00:08:18.820 | were still confused about this whole problem.

00:08:21.620 | (Lex laughs)

00:08:22.460 | - Absolutely. - That's been long ago

00:08:23.460 | been solved.

00:08:24.740 | Well, hello to the AI agents of the future.

00:08:27.260 | Anyway, so going back to your journey,

00:08:30.300 | when did you fall in love with programming first?

00:08:33.780 | - Well, I was pretty young age, actually.

00:08:35.940 | So I started off, actually, games was my first love,

00:08:40.860 | so starting to play chess when I was around four years old,

00:08:43.680 | and then it was actually with winnings

00:08:46.160 | from a chess competition that I managed

00:08:48.400 | to buy my first chess computer

00:08:49.780 | when I was about eight years old.

00:08:50.820 | It was a ZX Spectrum, which was hugely popular

00:08:53.180 | in the UK at the time.

00:08:54.740 | And it was an amazing machine,

00:08:56.540 | because I think it trained a whole generation

00:08:59.320 | of programmers in the UK, because it was so accessible.

00:09:02.540 | You literally switched it on,

00:09:03.820 | and there was the basic prompt,

00:09:05.140 | and you could just get going.

00:09:06.660 | And my parents didn't really know anything about computers,

00:09:10.460 | but because it was my money from a chess competition,

00:09:12.580 | I could say I wanted to buy it.

00:09:14.900 | And then I just went to bookstores,

00:09:17.940 | got books on programming,

00:09:19.860 | and started typing in the programming code.

00:09:23.500 | And then, of course, once you start doing that,

00:09:26.460 | you start adjusting it, and then making your own games.

00:09:29.140 | And that's when I fell in love with computers

00:09:30.840 | and realised that they were a very magical device.

00:09:33.880 | In a way, I wouldn't have been able

00:09:36.440 | to explain this at the time,

00:09:37.420 | but I felt that they were sort of

00:09:38.600 | almost a magical extension of your mind.

00:09:40.920 | I always had this feeling,

00:09:42.120 | and I've always loved this about computers,

00:09:43.860 | that you can set them off doing something,

00:09:46.160 | some task for you, you can go to sleep,

00:09:48.500 | come back the next day, and it's solved.

00:09:50.500 | That feels magical to me.

00:09:53.080 | So, I mean, all machines do that to some extent.

00:09:55.280 | They all enhance our natural capabilities.

00:09:57.660 | Obviously, cars make us, allow us to move faster

00:10:00.120 | than we can run, but this was a machine to extend the mind.

00:10:04.560 | And then, of course, AI is the ultimate expression

00:10:08.520 | of what a machine may be able to do or learn.

00:10:11.400 | So, very naturally for me,

00:10:13.560 | that thought extended into AI quite quickly.

00:10:16.080 | - Do you remember the programming language

00:10:18.600 | that was first started?

00:10:20.400 | - Yeah.

00:10:21.240 | - Was it special to the machine?

00:10:22.080 | - No, it was just a basic.

00:10:23.600 | I think it was just basic on the ZX Spectrum.

00:10:25.920 | I don't know what specific form it was.

00:10:27.520 | And then later on, I got a Commodore Amiga,

00:10:29.660 | which was a fantastic machine.

00:10:32.820 | - Now you're just showing off.

00:10:33.820 | - So, yeah, well, lots of my friends had Atari STs,

00:10:36.540 | and I managed to get Amigas.

00:10:37.860 | It was a bit more powerful, and that was incredible.

00:10:40.720 | And used to do programming in Assembler,

00:10:43.820 | and also Amos Basic, this specific form of basic.

00:10:48.140 | It was incredible, actually.

00:10:49.320 | So, I learned all my coding skills.

00:10:51.060 | - And when did you fall in love with AI?

00:10:53.060 | So, when did you first start to gain an understanding

00:10:56.900 | that you can not just write programs

00:10:58.880 | that do some mathematical operations for you

00:11:01.640 | while you sleep, but something that's akin

00:11:05.400 | to bringing an entity to life?

00:11:08.800 | Sort of a thing that can figure out something

00:11:11.840 | more complicated than a simple mathematical operation.

00:11:15.920 | - Yeah, so there was a few stages for me,

00:11:17.600 | all while I was very young.

00:11:18.940 | So, first of all, as I was trying to improve

00:11:21.720 | at playing chess, I was captaining

00:11:23.160 | various England junior chess teams.

00:11:24.860 | At the time, when I was about maybe 10, 11 years old,

00:11:27.500 | I was gonna become a professional chess player.

00:11:29.380 | That was my first thought.

00:11:31.280 | - So, that dream was there to try to get

00:11:34.700 | to the highest level of chess.

00:11:35.540 | - Yeah, so I was, when I was about 12 years old,

00:11:39.180 | I got to Master Standard, and I was second highest rated

00:11:41.380 | player in the world to Judith Polgar,

00:11:42.740 | who obviously ended up being an amazing chess player,

00:11:45.780 | and a world women's champion.

00:11:48.600 | And when I was trying to improve at chess,

00:11:50.780 | what you do is, obviously, first of all,

00:11:52.780 | you're trying to improve your own thinking processes.

00:11:55.140 | So, that leads you to thinking about thinking.

00:11:58.120 | How is your brain coming up with these ideas?

00:12:00.420 | Why is it making mistakes?

00:12:01.940 | How can you improve that thought process?

00:12:04.580 | But the second thing is that you,

00:12:06.380 | it was just the beginning, this was like in the early 80s,

00:12:09.700 | mid 80s, of chess computers.

00:12:11.280 | If you remember, they were physical balls,

00:12:12.800 | like the one we have in front of us,

00:12:14.020 | and you press down the squares.

00:12:17.020 | And I think Kasparov had a branded version of it

00:12:19.660 | that I got.

00:12:21.020 | And you were, you know, you used to,

00:12:22.980 | they're not as strong as they are today,

00:12:24.700 | but they were pretty strong,

00:12:26.380 | and you used to practice against them

00:12:29.060 | to try and improve your openings and other things.

00:12:31.460 | And so I remember, I think I probably got my first one,

00:12:33.500 | I was around 11 or 12.

00:12:34.940 | And I remember thinking, this is amazing,

00:12:37.420 | you know, how has someone programmed

00:12:39.260 | this chess board to play chess?

00:12:42.940 | And it was a very formative book I bought,

00:12:45.660 | which was called The Chess Computer Handbook by David Levy.

00:12:49.020 | This thing came out in 1984 or something,

00:12:50.700 | so I must have got it when I was about 11, 12.

00:12:52.380 | And it explained fully how these chess programs were made.

00:12:56.140 | And I remember my first AI program being,

00:12:58.940 | programming my Amiga,

00:13:00.420 | it couldn't, it wasn't powerful enough to play chess,

00:13:02.900 | I couldn't write a whole chess program,

00:13:04.220 | but I wrote a program for it to play Othello,

00:13:06.620 | or Reversi, it's sometimes called, I think, in the US.

00:13:09.340 | And so a slightly simpler game than chess,

00:13:11.780 | but I used all of the principles that chess programs had,

00:13:14.380 | alpha, beta, search, all of that.

00:13:16.020 | And that was my first AI program,

00:13:17.420 | I remember that very well, I was around 12 years old.

00:13:19.420 | So that brought me into AI.

00:13:21.660 | And then the second part was later on,

00:13:24.100 | I was around 16, 17,

00:13:25.540 | and I was writing games professionally, designing games,

00:13:28.820 | writing a game called Theme Park,

00:13:30.620 | which had AI as a core gameplay component

00:13:34.020 | as part of the simulation.

00:13:35.660 | And it sold millions of copies around the world,

00:13:38.460 | and people loved the way that the AI,

00:13:40.980 | even though it was relatively simple by today's AI standards,

00:13:44.460 | was reacting to the way you as the player played it.

00:13:47.700 | So it was called a sandbox game,

00:13:49.220 | so it was one of the first types of games like that,

00:13:51.340 | along with SimCity,

00:13:52.660 | and it meant that every game you played was unique.

00:13:55.700 | - Is there something you could say,

00:13:57.100 | just on a small tangent,

00:13:58.860 | about really impressive AI

00:14:02.180 | from a game design, human enjoyment perspective,

00:14:06.540 | really impressive AI that you've seen in games,

00:14:09.660 | and maybe what does it take to create AI system,

00:14:12.500 | and how hard of a problem is that?

00:14:14.220 | So a million questions, just as a brief tangent.

00:14:18.340 | - Well, look, I think games have been significant

00:14:22.660 | in my life for three reasons.

00:14:23.700 | So first of all, I was playing them

00:14:26.100 | and training myself on games when I was a kid.

00:14:28.780 | Then I went through a phase of designing games

00:14:31.460 | and writing AI for games.

00:14:32.980 | So all the games I professionally wrote

00:14:35.100 | had AI as a core component.

00:14:37.700 | And that was mostly in the '90s,

00:14:40.060 | and the reason I was doing that in games industry

00:14:42.980 | was at the time, the games industry,

00:14:45.100 | I think, was the cutting edge of technology.

00:14:47.180 | So whether it was graphics with people like John Carmack

00:14:49.820 | and Quake and those kind of things, or AI,

00:14:53.060 | I think actually all the action was going on in games.

00:14:56.140 | And we're still reaping the benefits of that,

00:14:58.460 | even with things like GPUs, which I find ironic,

00:15:01.500 | was obviously invented for graphics, computer graphics,

00:15:03.700 | but then turns out to be amazingly useful for AI.

00:15:06.220 | It just turns out everything's a matrix multiplication,

00:15:08.380 | it appears in the whole world.

00:15:11.140 | So I think games at the time had the most cutting edge AI,

00:15:15.780 | and a lot of the games, I was involved in writing,

00:15:19.780 | so there was a game called "Black and White,"

00:15:21.220 | which was one game I was involved with

00:15:22.700 | in the early stages of, which I still think

00:15:24.820 | is the most impressive example of reinforcement learning

00:15:29.340 | in a computer game.

00:15:30.500 | So in that game, you trained a little pet animal.

00:15:33.620 | - It's a brilliant game.

00:15:35.500 | - Yeah, and it sort of learned

00:15:36.340 | from how you were treating it.

00:15:37.620 | So if you treated it badly, then it became mean,

00:15:40.660 | and then it would be mean to your villagers

00:15:42.900 | and your population, the sort of,

00:15:45.020 | the little tribe that you were running.

00:15:47.220 | But if you were kind to it, then it would be kind.

00:15:49.380 | And people were fascinated by how that worked,

00:15:51.060 | and so was I, to be honest,

00:15:52.300 | with the way it kind of developed.

00:15:54.100 | And-- - Especially the mapping

00:15:55.940 | to good and evil. - Yeah.

00:15:57.500 | - It made you realize, made me realize

00:15:59.740 | that you can sort of, in the way,

00:16:02.300 | in the choices you make, can define where you end up,

00:16:07.300 | and that means all of us are capable of the good, evil.

00:16:12.620 | It all matters in the different choices

00:16:15.260 | along the trajectory to those places that you make.

00:16:18.220 | It's fascinating.

00:16:19.060 | I mean, games can do that philosophically to you,

00:16:21.380 | and it's rare, it seems rare.

00:16:22.540 | - Yeah, well, games are, I think, a unique medium,

00:16:24.680 | because you as the player,

00:16:26.580 | you're not just passively consuming the entertainment,

00:16:30.080 | right, you're actually actively involved as an agent.

00:16:34.300 | So I think that's what makes it, in some ways,

00:16:36.180 | can be more visceral than other mediums

00:16:38.420 | like films and books.

00:16:40.020 | So the second, so that was designing AI in games,

00:16:42.660 | and then the third use we've used of AI

00:16:46.460 | is in "Deep Mind," from the beginning,

00:16:48.460 | which is using games as a testing ground

00:16:50.940 | for proving out AI algorithms and developing AI algorithms.

00:16:55.020 | And that was a sort of a core component of our vision

00:16:59.100 | at the start of "Deep Mind,"

00:17:00.360 | was that we would use games very heavily

00:17:03.220 | as our main testing ground, certainly to begin with,

00:17:06.420 | because it's super efficient to use games,

00:17:08.600 | and also, you know, it's very easy to have metrics

00:17:11.500 | to see how well your systems are improving

00:17:14.100 | and what direction your ideas are going in

00:17:15.900 | and whether you're making incremental improvements.

00:17:18.420 | - And because those games are often rooted

00:17:20.420 | in something that humans did for a long time beforehand,

00:17:23.420 | there's already a strong set of rules,

00:17:26.540 | like it's already a damn good benchmark.

00:17:28.280 | - Yes, it's really good for so many reasons,

00:17:30.220 | because you've got clear measures

00:17:32.860 | of how good humans can be at these things.

00:17:35.580 | And in some cases, like "Go,"

00:17:36.860 | we've been playing it for thousands of years,

00:17:39.760 | and often they have scores or at least win conditions.

00:17:43.340 | So it's very easy for reward learning systems

00:17:45.660 | to get a reward.

00:17:46.500 | It's very easy to specify what that reward is.

00:17:49.340 | And also at the end, it's easy to test externally,

00:17:53.620 | you know, how strong is your system,

00:17:56.120 | by of course playing against, you know,

00:17:58.140 | the world's strongest players at those games.

00:18:00.220 | So it's so good for so many reasons,

00:18:02.700 | and it's also very efficient

00:18:03.940 | to run potentially millions of simulations

00:18:06.660 | in parallel on the cloud.

00:18:08.260 | So I think there's a huge reason why we were so successful

00:18:12.820 | back in, you know, starting out 2010,

00:18:14.740 | how come we were able to progress so quickly,

00:18:16.660 | because we've utilized games.

00:18:18.880 | And, you know, at the beginning of "DeepMind,"

00:18:21.300 | we also hired some amazing game engineers

00:18:24.580 | who I knew from my previous lives in the games industry,

00:18:28.020 | and that helped to bootstrap us very quickly.

00:18:30.940 | - And plus it's somehow super compelling,

00:18:33.860 | almost at a philosophical level of man versus machine

00:18:38.060 | over chess board or a Go board.

00:18:41.220 | And especially given that the entire history of AI

00:18:43.620 | is defined by people saying it's gonna be impossible

00:18:45.980 | to make a machine that beats a human being in chess.

00:18:50.980 | And then once that happened, people were certain

00:18:54.580 | when I was coming up in AI that Go is not a game

00:18:58.020 | that can be solved because of the combinatorial complexity.

00:19:01.220 | It's just too, it's, you know,

00:19:04.140 | no matter how much Moore's law you have,

00:19:06.640 | compute is just never going to be able

00:19:08.580 | to crack the game of Go.

00:19:10.180 | And so then there's something compelling about facing,

00:19:14.900 | sort of taking on the impossibility of that task

00:19:18.140 | from the AI researcher perspective, engineer perspective,

00:19:23.140 | and then as a human being just observing this whole thing,

00:19:27.020 | your beliefs about what you thought was impossible

00:19:31.500 | being broken apart, it's humbling to realize

00:19:37.500 | we're not as smart as we thought.

00:19:40.500 | It's humbling to realize that the things we think

00:19:43.140 | are impossible now perhaps will be done in the future.

00:19:46.980 | There's something really powerful about a game,

00:19:50.820 | AI system beating a human being in a game

00:19:52.900 | that drives that message home

00:19:55.700 | for like millions, billions of people,

00:19:58.020 | especially in the case of Go.

00:19:59.340 | - Sure.

00:20:00.520 | Well, look, I think it's a, I mean,

00:20:01.940 | it has been a fascinating journey,

00:20:03.740 | and especially as I think about it from,

00:20:06.900 | I can understand it from both sides,

00:20:08.780 | both as the AI, you know, creators of the AI,

00:20:13.100 | but also as a games player originally.

00:20:15.660 | So, you know, it was a really interesting,

00:20:17.540 | you know, I mean, it was a fantastic,

00:20:20.380 | but also somewhat bittersweet moment,

00:20:22.100 | the AlphaGo match for me, seeing that,

00:20:25.340 | and being obviously heavily, heavily involved in that.

00:20:28.500 | But, you know, as you say, chess has been the,

00:20:32.500 | I mean, Kasparov, I think rightly called it

00:20:34.380 | the Drosophila of intelligence, right?

00:20:37.300 | So it's sort of, I love that phrase,

00:20:39.500 | and I think he's right, because chess has been

00:20:42.940 | hand in hand with AI from the beginning

00:20:45.380 | of the whole field, right?

00:20:47.420 | So I think every AI practitioner,

00:20:49.580 | starting with Turing and Claude Shannon and all those,

00:20:52.420 | the sort of forefathers of the field,

00:20:55.380 | tried their hand at writing a chess program.

00:20:58.820 | I've got original edition of Claude Shannon's

00:21:01.140 | first chess program, I think it was 1949,

00:21:03.980 | the original sort of paper, and they all did that.

00:21:08.620 | And Turing famously wrote a chess program

00:21:11.060 | that all the computers around then

00:21:12.460 | were obviously too slow to run it.

00:21:13.740 | So he had to run, he had to be the computer, right?

00:21:16.020 | So he literally, I think, spent two or three days

00:21:18.900 | running his own program by hand with pencil and paper

00:21:21.340 | and playing a friend of his with his chess program.

00:21:24.940 | So of course, Deep Blue was a huge moment beating Kasparov.

00:21:29.940 | But actually, when that happened,

00:21:31.820 | I remember that very, very vividly, of course,

00:21:34.060 | because it was chess and computers and AI,

00:21:36.580 | all the things I loved, and I was at college at the time.

00:21:39.180 | But I remember coming away from that

00:21:40.740 | being more impressed by Kasparov's mind

00:21:43.020 | than I was by Deep Blue, because here was Kasparov

00:21:46.420 | with his human mind, not only could he play chess

00:21:48.860 | more or less to the same level

00:21:50.140 | as this brute of a calculation machine,

00:21:53.140 | but of course, Kasparov can do

00:21:54.580 | everything else humans can do,

00:21:55.780 | ride a bike, talk many languages, do politics,

00:21:58.140 | all the rest of the amazing things that Kasparov does.

00:22:00.860 | And so with the same brain, and yet Deep Blue,

00:22:04.500 | brilliant as it was at chess,

00:22:07.020 | it'd been hand-coded for chess,

00:22:09.380 | and actually had distilled the knowledge

00:22:13.140 | of chess grandmasters into a cool program,

00:22:16.420 | but it couldn't do anything else.

00:22:17.820 | Like, it couldn't even play a strictly simpler game

00:22:20.100 | like tic-tac-toe.

00:22:21.300 | So something to me was missing from intelligence

00:22:25.900 | from that system that we would regard as intelligence.

00:22:28.500 | And I think it was this idea of generality

00:22:30.900 | and also learning.

00:22:32.180 | So, and that's obviously what we tried to do with AlphaGo.

00:22:36.140 | - Yeah, with AlphaGo and AlphaZero and MuZero

00:22:38.580 | and then Gato and all the things

00:22:40.460 | that we'll get into some parts of,

00:22:43.220 | there's just a fascinating trajectory here.

00:22:45.660 | But let's just stick on chess briefly.

00:22:48.540 | On the human side of chess,

00:22:50.180 | you've proposed that from a game design perspective,

00:22:53.420 | the thing that makes chess compelling as a game

00:22:56.460 | is that there's a creative tension

00:22:59.540 | between a bishop and the knight.

00:23:02.940 | Can you explain this?

00:23:04.020 | First of all, it's really interesting

00:23:05.580 | to think about what makes a game compelling,

00:23:08.620 | makes it stick across centuries.

00:23:10.980 | - Yeah, I was sort of thinking about this,

00:23:13.460 | and actually a lot of even amazing chess players

00:23:15.460 | don't think about it necessarily

00:23:16.860 | from a games designer point of view.

00:23:18.300 | So it's with my game design hat on

00:23:20.260 | that I was thinking about this.

00:23:21.260 | Why is chess so compelling?

00:23:23.100 | And I think a critical reason is the dynamicness

00:23:27.580 | of the different kind of chess positions you can have,

00:23:29.980 | whether they're closed or open and other things

00:23:32.220 | comes from the bishop and the knight.

00:23:33.500 | So if you think about how different

00:23:36.500 | the capabilities of the bishop and knight are

00:23:39.260 | in terms of the way they move,

00:23:40.860 | and then somehow chess has evolved

00:23:43.100 | to balance those two capabilities

00:23:44.860 | more or less equally.

00:23:46.100 | So they're both roughly worth three points each.

00:23:48.740 | - So you think that dynamics is always there,

00:23:50.580 | and then the rest of the rules

00:23:51.660 | are kind of trying to stabilize the game?

00:23:53.780 | - Well, maybe.

00:23:54.620 | I mean, it's sort of, I don't know,

00:23:55.460 | it's chicken and egg situation.

00:23:56.540 | Probably both came together.

00:23:57.700 | But the fact that it's got to this beautiful equilibrium

00:24:00.540 | where you can have the bishop and knight,

00:24:02.380 | they're so different in power,

00:24:04.420 | but so equal in value across the set

00:24:06.900 | of the universe of all positions, right?

00:24:09.500 | Somehow they've been balanced by humanity

00:24:11.580 | over hundreds of years.

00:24:13.500 | I think gives the game the creative tension

00:24:16.900 | that you can swap the bishop and knight,

00:24:19.020 | for a bishop for a knight,

00:24:20.140 | and they're more or less worth the same,

00:24:22.100 | but now you aim for a different type of position.

00:24:24.020 | If you have the knight, you want a closed position.

00:24:26.060 | If you have the bishop, you want an open position.

00:24:28.180 | So I think that creates a lot

00:24:29.220 | of the creative tension in chess.

00:24:30.940 | - So some kind of controlled creative tension.

00:24:34.060 | From an AI perspective,

00:24:35.980 | do you think AI systems could eventually design games

00:24:38.820 | that are optimally compelling to humans?

00:24:41.620 | - Well, that's an interesting question.

00:24:42.940 | Sometimes I get asked about AI and creativity,

00:24:45.980 | and the way I answer that is relevant to that question,

00:24:48.860 | which is that I think there are different levels

00:24:51.260 | of creativity, one could say.

00:24:52.900 | So I think if we define creativity

00:24:55.300 | as coming up with something original, right,

00:24:57.260 | that's useful for a purpose,

00:24:59.300 | then I think the kind of lowest level of creativity

00:25:02.220 | is like an interpolation,

00:25:03.700 | so an averaging of all the examples you see.

00:25:06.260 | So maybe a very basic AI system could say

00:25:08.300 | you could have that.

00:25:09.140 | So you show it millions of pictures of cats,

00:25:11.380 | and then you say, "Give me an average-looking cat," right?

00:25:13.820 | "Generate me an average-looking cat."

00:25:15.460 | I would call that interpolation.

00:25:17.180 | Then there's extrapolation,

00:25:18.700 | which something like AlphaGo showed.

00:25:20.420 | So AlphaGo played millions of games of Go against itself,

00:25:24.300 | and then it came up with brilliant new ideas

00:25:26.580 | like move 37 in game two,

00:25:28.180 | brilliant motif strategies in Go

00:25:30.740 | that no humans had ever thought of,

00:25:32.820 | even though we've played it for thousands of years

00:25:34.780 | and professionally for hundreds of years.

00:25:36.580 | So that, I call that extrapolation.

00:25:38.820 | But then there's still a level above that,

00:25:41.060 | which is, you know, you could call out-of-the-box thinking

00:25:44.020 | or true innovation, which is, could you invent Go, right?

00:25:47.540 | Could you invent chess?

00:25:48.500 | And not just come up with a brilliant chess move

00:25:50.260 | or a brilliant Go move,

00:25:51.340 | but can you actually invent chess

00:25:53.700 | or something as good as chess or Go?

00:25:55.900 | And I think one day, AI could,

00:25:58.900 | but what's missing is how would you even specify that task

00:26:02.260 | to a program right now?

00:26:04.500 | And the way I would do it,

00:26:05.420 | if I was telling a human to do it or a games designer,

00:26:08.820 | a human games designer to do it,

00:26:10.260 | is I would say something like Go,

00:26:11.460 | I would say, come up with a game

00:26:14.180 | that only takes five minutes to learn,

00:26:16.140 | which Go does 'cause it's got simple rules,

00:26:17.940 | but many lifetimes to master, right?

00:26:20.300 | Or impossible to master in one lifetime

00:26:22.140 | 'cause it's so deep and so complex.

00:26:23.980 | And then it's aesthetically beautiful.

00:26:27.500 | And also it can be completed in three or four hours

00:26:30.820 | of gameplay time, which is, you know, useful for our,

00:26:33.660 | us, you know, in a human day.

00:26:35.980 | And so you might specify these sort of high-level concepts

00:26:39.300 | like that, and then, you know, with that,

00:26:41.380 | and then maybe a few other things,

00:26:43.460 | one could imagine that Go satisfies those constraints.

00:26:48.260 | But the problem is, is that we're not able

00:26:50.300 | to specify abstract notions like that,

00:26:53.740 | high-level abstract notions like that yet to our AI systems.

00:26:57.500 | And I think there's still something missing there

00:26:59.500 | in terms of high-level concepts or abstractions

00:27:02.460 | that they truly understand

00:27:03.700 | and that are, you know, combinable and compositional.

00:27:07.180 | And so for the moment, I think AI is capable

00:27:10.380 | of doing interpolation and extrapolation,

00:27:12.380 | but not true invention.

00:27:14.140 | - So coming up with rule sets and optimizing

00:27:18.500 | with complicated objectives around those rule sets,

00:27:21.220 | we can't currently do, but you could take a specific rule set

00:27:26.060 | and then run a kind of self-play experiment

00:27:28.900 | to see how long, just observe how an AI system

00:27:32.580 | from scratch learns, how long is that journey of learning?

00:27:36.540 | And maybe if it satisfies some of those other things

00:27:39.700 | you mentioned in terms of quickness to learn and so on,

00:27:42.300 | and you could see a long journey to master

00:27:44.700 | for even an AI system,

00:27:46.820 | then you could say that this is a promising game.

00:27:49.860 | But it would be nice to do almost like alpha codes

00:27:52.300 | or programming rules.

00:27:53.940 | So generating rules that kind of,

00:27:55.740 | that automate even that part of the generation of rules.

00:28:00.420 | - So I have thought about systems actually,

00:28:02.980 | I think it'd be amazing for a games designer

00:28:05.660 | if you could have a system that takes your game,

00:28:09.180 | plays it tens of millions of times, maybe overnight,

00:28:11.940 | and then self-balances the rules better.

00:28:13.820 | So it tweaks the rules and maybe the equations

00:28:18.060 | and the parameters so that the game is more balanced,

00:28:22.700 | the units in the game or some of the rules could be tweaked.

00:28:26.260 | So it's a bit of like giving a base set

00:28:28.300 | and then allowing a Monte Carlo tree search

00:28:30.780 | or something like that to sort of explore it.

00:28:33.380 | And I think that would be super powerful tool actually

00:28:37.100 | for balancing, auto balancing a game,

00:28:39.740 | which usually takes thousands of hours

00:28:42.140 | from hundreds of games, human games testers normally

00:28:44.540 | to balance some game like StarCraft,

00:28:47.500 | which is, you know, Blizzard are amazing

00:28:49.700 | at balancing their games,

00:28:50.660 | but it takes them years and years and years.

00:28:52.580 | So one could imagine at some point

00:28:54.140 | when this stuff becomes efficient enough to,

00:28:57.260 | you know, you might better do that like overnight.

00:28:59.540 | - Do you think a game that is optimal

00:29:02.740 | designed by an AI system

00:29:04.980 | would look very much like a Planet Earth?

00:29:08.260 | - Maybe, maybe.

00:29:10.740 | It's only the sort of game I would love to make is,

00:29:13.060 | and I've tried, you know, in my games career,

00:29:15.980 | the games design career, you know,

00:29:17.460 | my first big game was designing a theme park,

00:29:20.260 | an amusement park.

00:29:21.380 | Then with games like Republic,

00:29:23.700 | I tried to, you know, have games where we designed

00:29:25.660 | whole cities and allowed you to play in.

00:29:28.460 | So, and of course people like Will Wright

00:29:30.260 | have written games like Sim Earth,

00:29:32.620 | trying to simulate the whole of Earth.

00:29:34.300 | Pretty tricky, but I think--

00:29:36.140 | - Sim Earth, I haven't actually played that one.

00:29:37.580 | So what is it, does it incorporate of evolution or?

00:29:40.300 | - Yeah, it has evolution and it sort of,

00:29:42.660 | it tries to, it sort of treats it as an entire biosphere,

00:29:45.340 | but from quite high level.

00:29:47.260 | So--

00:29:48.100 | - It'd be nice to be able to sort of zoom in,

00:29:50.300 | zoom out or zoom in. - Exactly, exactly.

00:29:51.740 | So obviously it couldn't do, that was in the,

00:29:53.460 | I think he wrote that in the 90s.

00:29:54.940 | So it couldn't, you know, it wasn't able to do that.

00:29:57.580 | But that would be obviously

00:29:59.220 | the ultimate sandbox game, of course.

00:30:01.460 | - On that topic, do you think we're living in a simulation?

00:30:04.780 | - Yes, well, so, okay, so I--

00:30:06.620 | - We're gonna jump around from the absurdly philosophical

00:30:09.300 | to the technical. - Sure, sure.

00:30:10.740 | Very, very happy to.

00:30:11.900 | So I think my answer to that question

00:30:13.780 | is a little bit complex because there is simulation theory,

00:30:17.620 | which obviously Nick Bostrom, I think,

00:30:19.100 | famously first proposed.

00:30:20.600 | And I don't quite believe it in that sense.

00:30:24.700 | So in the sense that are we in some sort of computer game

00:30:29.580 | or have our descendants somehow recreated

00:30:32.660 | Earth in the 21st century

00:30:35.540 | and for some kind of experimental reason?

00:30:38.420 | I think that, but I do think that we might be,

00:30:43.340 | that the best way to understand physics and the universe

00:30:46.940 | is from a computational perspective.

00:30:49.260 | So understanding it as an information universe

00:30:52.380 | and actually information being the most fundamental

00:30:55.700 | unit of reality rather than matter or energy.

00:30:59.900 | So a physicist would say, matter or energy,

00:31:02.660 | E equals MC squared, these are the things

00:31:04.420 | that are the fundamentals of the universe.

00:31:07.340 | I'd actually say information, which of course itself

00:31:10.980 | can specify energy or matter, right?

00:31:13.540 | Matter is actually just, we're just out the way our bodies

00:31:16.860 | and all the molecules in our body are arranged

00:31:18.700 | as information.

00:31:19.700 | So I think information may be the most fundamental way

00:31:23.060 | to describe the universe and therefore you could say

00:31:26.580 | we're in some sort of simulation because of that.

00:31:29.820 | But I'm not really a subscriber to the idea

00:31:32.780 | that these are sort of throwaway billions

00:31:35.900 | of simulations around.

00:31:36.860 | I think this is actually very critical

00:31:39.380 | and possibly unique, this simulation.

00:31:41.700 | - This particular one?

00:31:42.580 | - Yes.

00:31:43.420 | - And you just mean treating the universe as a computer

00:31:48.420 | that's processing and modifying information

00:31:52.180 | is a good way to solve the problems of physics,

00:31:54.820 | of chemistry, of biology and perhaps of humanity and so on.

00:31:59.660 | - Yes, I think understanding physics

00:32:02.180 | in terms of information theory might be the best way

00:32:05.860 | to really understand what's going on here.

00:32:09.340 | - From our understanding of a universal Turing machine,

00:32:13.500 | from our understanding of a computer,

00:32:15.220 | do you think there's something outside

00:32:17.380 | of the capabilities of a computer

00:32:19.400 | that is present in our universe?

00:32:20.940 | You have a disagreement with Roger Penrose

00:32:23.500 | about the nature of consciousness.

00:32:25.900 | He thinks that consciousness is more

00:32:27.740 | than just a computation.

00:32:28.980 | Do you think all of it, the whole shebang,

00:32:32.660 | can be a computation?

00:32:33.980 | - Yeah, I've had many fascinating debates

00:32:35.740 | with Sir Roger Penrose and obviously he's famously,

00:32:39.560 | and I read "Emperors of the New Mind"

00:32:41.380 | and his books, his classical books,

00:32:45.220 | and they were pretty influential in the '90s.

00:32:47.660 | And he believes that there's something more,

00:32:50.820 | something quantum that is needed

00:32:52.900 | to explain consciousness in the brain.

00:32:55.740 | I think about what we're doing actually at DeepMind

00:32:58.220 | and what my career is being,

00:32:59.820 | we're almost like Turing's champion.

00:33:01.820 | So we are pushing Turing machines

00:33:03.580 | or classical computation to the limits.

00:33:05.920 | What are the limits of what classical computing can do?

00:33:09.340 | Now, and at the same time, I've also studied neuroscience

00:33:13.560 | to see, and that's why I did my PhD in,

00:33:15.460 | was to see, also to look at,

00:33:17.660 | is there anything quantum in the brain

00:33:19.160 | from a neuroscience or biological perspective?

00:33:21.260 | And so far, I think most neuroscientists

00:33:24.460 | and most mainstream biologists and neuroscientists

00:33:26.380 | would say there's no evidence of any quantum systems

00:33:29.420 | or effects in the brain.

00:33:30.740 | As far as we can see, it can be mostly explained

00:33:32.980 | by classical theories.

00:33:35.860 | So, and then, so there's sort of the search

00:33:39.300 | from the biology side.

00:33:40.620 | And then at the same time, there's the raising of the water,

00:33:44.300 | the bar, from what classical Turing machines can do.

00:33:48.220 | And including our new AI systems.

00:33:51.680 | And as you alluded to earlier,

00:33:54.000 | I think AI, especially in the last decade plus,

00:33:57.760 | has been a continual story now,

00:33:59.880 | surprising events and surprising successes,

00:34:03.920 | knocking over one theory after another

00:34:05.800 | of what was thought to be impossible,

00:34:07.760 | from Go to protein folding and so on.

00:34:10.080 | And so I think I would be very hesitant

00:34:14.760 | to bet against how far the universal Turing machine

00:34:19.520 | and classical computation paradigm can go.

00:34:23.400 | And my betting would be that all of, certainly,

00:34:27.280 | what's going on in our brain can probably be mimicked

00:34:30.680 | or approximated on a classical machine,

00:34:34.720 | not requiring something metaphysical or quantum.

00:34:38.400 | - And we'll get there with some of the work with AlphaFold,

00:34:41.720 | which I think begins the journey of modeling

00:34:45.080 | this beautiful and complex world of biology.

00:34:48.160 | So you think all the magic of the human mind

00:34:50.160 | comes from this, just a few pounds of mush,

00:34:53.520 | of biological computational mush

00:34:57.480 | that's akin to some of the neural networks,

00:35:00.560 | not directly, but in spirit,

00:35:03.800 | that DeepMind has been working with?

00:35:06.200 | - Well, look, I think it's, you say it's a few,

00:35:08.360 | you know, of course, this is, I think,

00:35:10.040 | the biggest miracle of the universe is that

00:35:12.760 | it is just a few pounds of mush in our skulls,

00:35:15.040 | and yet it's also, our brains are the most complex objects

00:35:18.560 | that we know of in the universe.

00:35:20.220 | So there's something profoundly beautiful

00:35:22.360 | and amazing about our brains,

00:35:23.920 | and I think that it's an incredibly,

00:35:28.640 | incredible efficient machine,

00:35:30.640 | and it's a phenomenon, basically.

00:35:35.560 | And I think that building AI,

00:35:37.480 | one of the reasons I wanna build AI,

00:35:38.920 | and I've always wanted to, is,

00:35:40.440 | I think by building an intelligent artifact like AI,

00:35:43.800 | and then comparing it to the human mind,

00:35:46.480 | that will help us unlock the uniqueness

00:35:49.560 | and the true secrets of the mind

00:35:50.960 | that we've always wondered about since the dawn of history,

00:35:53.480 | like consciousness, dreaming, creativity, emotions.

00:35:57.800 | What are all these things, right?

00:36:00.760 | We've wondered about them since the dawn of humanity,

00:36:04.200 | and I think one of the reasons,

00:36:05.940 | and, you know, I love philosophy and philosophy of mind,

00:36:08.280 | is, we found it difficult,

00:36:09.920 | is there haven't been the tools for us to really,

00:36:11.960 | other than introspection,

00:36:13.400 | to, from very clever people in history,

00:36:15.880 | very clever philosophers,

00:36:17.200 | to really investigate this scientifically.

00:36:19.360 | But now, suddenly we have a plethora of tools.

00:36:21.720 | Firstly, we have all of the neuroscience tools,

00:36:23.240 | fMRI machines, single-cell recording, all of this stuff,

00:36:25.900 | but we also have the ability, computers and AI,

00:36:29.000 | to build intelligent systems.

00:36:31.640 | So I think that, you know,

00:36:34.720 | I think it is amazing what the human mind does,

00:36:37.320 | and I'm kind of in awe of it, really,

00:36:41.120 | and I think it's amazing that, without human minds,

00:36:44.440 | we're able to build things like computers,

00:36:46.780 | and actually even, you know,

00:36:48.240 | think and investigate about these questions.

00:36:49.880 | I think that's also a testament to the human mind.

00:36:52.720 | - Yeah, the universe built the human mind

00:36:56.200 | that now is building computers

00:36:57.620 | that help us understand both the universe

00:37:00.380 | and our own human mind.

00:37:01.480 | - That's right, that's exactly it.

00:37:02.660 | I mean, I think that's one, you know,

00:37:03.920 | one could say we are,

00:37:05.760 | maybe we're the mechanism by which the universe

00:37:08.160 | is going to try and understand itself.

00:37:09.840 | - Yeah.

00:37:10.680 | (laughing)

00:37:12.520 | It's beautiful.

00:37:13.360 | So let's go to the basic building blocks of biology

00:37:16.960 | that I think is another angle

00:37:19.400 | at which you can start to understand the human mind,

00:37:21.440 | the human body, which is quite fascinating,

00:37:23.400 | which is, from the basic building blocks,

00:37:26.640 | start to simulate, start to model

00:37:28.960 | how, from those building blocks,

00:37:30.480 | you can construct bigger and bigger, more complex systems,

00:37:33.080 | maybe one day the entirety of the human biology.

00:37:35.820 | So here's another problem that thought

00:37:39.680 | to be impossible to solve, which is protein folding.

00:37:42.720 | And AlphaFold, or specifically AlphaFold2,

00:37:47.720 | did just that.

00:37:48.840 | It solved protein folding.

00:37:50.320 | I think it's one of the biggest breakthroughs,

00:37:53.400 | certainly in the history of structural biology,

00:37:55.140 | but in general, in science.

00:37:58.200 | Maybe from a high level,

00:38:02.280 | what is it and how does it work?

00:38:04.840 | And then we can ask some fascinating questions after.

00:38:08.680 | - Sure.

00:38:09.960 | So maybe to explain it to people

00:38:12.360 | not familiar with protein folding is,

00:38:14.400 | first of all, explain proteins,

00:38:15.780 | which is, proteins are essential to all life.

00:38:18.840 | Every function in your body depends on proteins.

00:38:21.520 | Sometimes they're called the workhorses of biology.

00:38:23.920 | And if you look into them,

00:38:24.880 | and obviously as part of AlphaFold,

00:38:26.640 | I've been researching proteins and structural biology

00:38:30.200 | for the last few years,

00:38:31.760 | they're amazing little bio-nano-machines proteins.

00:38:34.760 | They're incredible if you actually watch

00:38:35.920 | little videos of how they work,

00:38:37.200 | animations of how they work.

00:38:39.000 | And proteins are specified by their genetic sequence,

00:38:42.600 | called their amino acid sequence.

00:38:44.280 | So you can think of it as their genetic makeup.

00:38:47.040 | And then in the body, in nature,

00:38:50.320 | when they fold up into a 3D structure.

00:38:53.340 | So you can think of it as a string of beads,

00:38:55.280 | and then they fold up into a ball.

00:38:57.120 | Now the key thing is,

00:38:58.280 | you want to know what that 3D structure is,

00:39:01.100 | because the 3D structure of a protein

00:39:04.480 | is what helps to determine what does it do,

00:39:06.720 | the function it does in your body.

00:39:08.560 | And also, if you're interested in drugs or disease,

00:39:12.280 | you need to understand that 3D structure.

00:39:13.920 | Because if you want to target something

00:39:15.800 | with a drug compound,

00:39:17.040 | about to block something the protein's doing,

00:39:20.240 | you need to understand where it's gonna bind

00:39:21.960 | on the surface of the protein.

00:39:23.400 | So obviously, in order to do that,

00:39:24.900 | you need to understand the 3D structure.

00:39:26.680 | - So the structure's mapped to the function.

00:39:28.600 | - The structure's mapped to the function.

00:39:29.820 | And the structure is obviously somehow

00:39:31.840 | specified by the amino acid sequence.

00:39:34.840 | And that's, in essence, the protein folding problem is,

00:39:37.420 | can you just from the amino acid sequence,

00:39:39.620 | the one-dimensional string of letters,

00:39:42.560 | can you immediately computationally predict

00:39:45.600 | the 3D structure?

00:39:47.160 | And this has been a grand challenge in biology

00:39:50.040 | for over 50 years.

00:39:51.540 | So I think it was first articulated

00:39:53.200 | by Christian Anfiensen, a Nobel Prize winner in 1972,

00:39:56.840 | as part of his Nobel Prize winning lecture.

00:39:59.280 | And he just speculated this should be possible

00:40:01.900 | to go from the amino acid sequence to the 3D structure.

00:40:05.000 | But he didn't say how.

00:40:06.100 | So it's been described to me as equivalent

00:40:09.480 | to Fermat's last theorem, but for biology.

00:40:12.320 | - You should, as somebody that very well might win

00:40:15.180 | the Nobel Prize in the future,

00:40:16.600 | but outside of that,

00:40:18.360 | you should do more of that kind of thing.

00:40:20.020 | In the margin, just put random things.

00:40:22.200 | That'll take like 200 years to solve.

00:40:24.480 | - Set people off for 200 years.

00:40:26.040 | - It should be possible.

00:40:27.760 | - And just don't give any details.

00:40:29.240 | - Exactly, I think everyone should, exactly.

00:40:30.640 | It should be, I'll have to remember that for future.

00:40:33.520 | So yeah, so he set off with this one throwaway remark,

00:40:36.280 | just like Fermat, he set off this whole 50-year field,

00:40:40.280 | really, of computational biology.

00:40:44.400 | And they got stuck.

00:40:46.240 | They hadn't really got very far with doing this.

00:40:48.520 | And until now, until AlphaFold came along,

00:40:52.500 | this is done experimentally, very painstakingly.

00:40:55.500 | So the rule of thumb is,

00:40:56.580 | and you have to crystallize the protein,

00:40:58.720 | which is really difficult.

00:40:59.840 | Some proteins can't be crystallized like membrane proteins.

00:41:03.080 | And then you have to use very expensive electron microscopes

00:41:05.960 | or X-ray crystallography machines,

00:41:08.200 | really painstaking work to get the 3D structure

00:41:10.680 | and visualize the 3D structure.

00:41:12.420 | So the rule of thumb in experimental biology

00:41:14.880 | is that it takes one PhD student,

00:41:16.860 | their entire PhD, to do one protein.

00:41:19.400 | And with AlphaFold2, we're able to predict

00:41:23.440 | the 3D structure in a matter of seconds.

00:41:26.400 | And so over Christmas, we did the whole human proteome,

00:41:30.240 | or every protein in the human body, all 20,000 proteins.

00:41:33.280 | So the human proteome's like the equivalent

00:41:34.760 | of the human genome, but on protein space.

00:41:37.560 | And sort of revolutionized, really,

00:41:40.240 | what structural biologists can do.

00:41:43.300 | Because now, they don't have to worry

00:41:45.720 | about these painstaking experimentals.

00:41:47.960 | Should they put all of their effort in or not?

00:41:49.560 | They can almost just look up the structure

00:41:51.080 | of their proteins like a Google search.

00:41:53.280 | - And so there's a data set on which it's trained

00:41:56.880 | and how to map this amino acid sequence.

00:41:58.800 | First of all, it's incredible that a protein,

00:42:00.760 | this little chemical computer,

00:42:01.920 | is able to do that computation itself

00:42:03.760 | in some kind of distributed way and do it very quickly.

00:42:07.800 | That's a weird thing.

00:42:08.840 | And they evolved that way 'cause in the beginning,

00:42:11.720 | I mean, that's a great invention,

00:42:13.160 | just the protein itself.

00:42:14.640 | - Yes, I mean-- - And then there's,

00:42:16.320 | I think, probably a history of,

00:42:18.900 | like they evolved to have many of these proteins.

00:42:22.720 | And those proteins figure out how to be computers themselves

00:42:26.560 | in such a way that you can create structures

00:42:28.520 | that can interact in complex ways with each other

00:42:30.520 | in order to form high-level functions.

00:42:32.620 | I mean, it's a weird system that they figured it out.

00:42:35.480 | - Well, for sure.

00:42:36.320 | I mean, maybe we should talk about the origins of life too.

00:42:38.960 | But proteins themselves, I think, are magical

00:42:41.140 | and incredible, as I said, little bio-nano machines.

00:42:45.760 | And actually, Leventhal, who is another scientist,

00:42:51.000 | a contemporary of Amundsen, he coined this Leventhal,

00:42:55.080 | what became known as Leventhal's paradox,

00:42:56.800 | which is exactly what you're saying.

00:42:58.280 | He calculated roughly an average protein,

00:43:01.540 | which is maybe 2,000 amino acids bases long,

00:43:04.900 | can fold in maybe 10 to the power 300

00:43:09.960 | different conformations.

00:43:11.480 | So there's 10 to the power 300 different ways

00:43:13.320 | that protein could fold up.

00:43:14.800 | And yet somehow, in nature, physics solves this

00:43:19.120 | in a matter of milliseconds.

00:43:20.540 | So proteins fold up in your body

00:43:22.320 | in sometimes in fractions of a second.

00:43:25.600 | So physics is somehow solving that search problem.

00:43:29.080 | - And just to be clear, in many of these cases,

00:43:31.200 | maybe you can correct me if I'm wrong,

00:43:33.040 | there's often a unique way for that sequence to form itself.

00:43:37.680 | So among that huge number of possibilities,

00:43:41.240 | it figures out a way how to stably,

00:43:43.560 | in some cases, there might be a misfunction, so on,

00:43:47.780 | which leads to a lot of the disorders and stuff like that.

00:43:50.040 | But most of the time, it's a unique mapping.

00:43:52.760 | And that unique mapping's not obvious.

00:43:54.840 | - No, exactly.

00:43:55.920 | - It's just what the problem is.

00:43:57.120 | - No, exactly.

00:43:57.960 | So there's a unique mapping, usually,

00:43:59.640 | in a healthy, if it's healthy.

00:44:01.880 | And as you say, in disease, so for example, Alzheimer's,

00:44:05.440 | one conjecture is that it's because of misfolded protein,

00:44:09.040 | a protein that folds in the wrong way, amyloid beta protein.

00:44:12.040 | So, and then because it folds in the wrong way,

00:44:14.600 | it gets tangled up, right, in your neurons.

00:44:17.640 | So it's super important to understand

00:44:20.560 | both healthy functioning and also disease,

00:44:23.600 | is to understand what these things are doing

00:44:26.480 | and how they're structuring.

00:44:27.600 | Of course, the next step is sometimes proteins change shape

00:44:30.560 | when they interact with something.

00:44:32.160 | So they're not just static, necessarily, in biology.

00:44:35.940 | - Maybe you can give some interesting,

00:44:39.800 | sort of beautiful things to you

00:44:41.360 | about these early days of alpha fold,

00:44:44.160 | of solving this problem.

00:44:46.160 | Because unlike games, this is real physical systems

00:44:51.160 | that are less amenable to self-play type of mechanisms.

00:44:55.640 | - Sure.

00:44:56.600 | - The size of the data set is smaller

00:44:58.440 | than you might otherwise like,

00:44:59.760 | so you have to be very clever about certain things.

00:45:01.780 | Is there something you could speak to

00:45:03.640 | what was very hard to solve

00:45:06.680 | and what are some beautiful aspects about the solution?

00:45:09.920 | - Yeah, I would say alpha fold is the most complex

00:45:12.800 | and also probably most meaningful system

00:45:14.600 | we've built so far.

00:45:15.860 | So it's been an amazing time actually in the last,

00:45:18.400 | you know, two, three years to see that come through

00:45:20.520 | because as we talked about earlier,

00:45:23.000 | you know, games is what we started on,

00:45:25.480 | building things like AlphaGo and AlphaZero.

00:45:27.900 | But really the ultimate goal was to,

00:45:30.400 | not just to crack games, it was just to build,

00:45:33.160 | use them to bootstrap general learning systems

00:45:35.320 | we could then apply to real world challenges.

00:45:37.460 | Specifically, my passion is scientific challenges

00:45:40.640 | like protein folding.

00:45:41.920 | And then alpha fold, of course,

00:45:43.280 | is our first big proof point of that.

00:45:45.360 | And so, you know, in terms of the data

00:45:49.000 | and the amount of innovations that had to go into it,

00:45:50.920 | we, you know, it was like more than 30

00:45:53.060 | different component algorithms needed to be put together

00:45:55.540 | to crack the protein folding.

00:45:57.960 | I think some of the big innovations were the

00:46:00.800 | kind of building in some hard coded constraints

00:46:04.220 | around physics and evolutionary biology

00:46:07.760 | to constrain sort of things like the bond angles

00:46:10.440 | in the protein and things like that.

00:46:15.160 | But not to impact the learning system.

00:46:18.040 | So still allowing the system to be able to learn

00:46:21.000 | the physics itself from the examples that we had.

00:46:25.520 | And the examples, as you say,

00:46:26.620 | there are only about 150,000 proteins,

00:46:28.840 | even after 40 years of experimental biology,

00:46:31.240 | only around 150,000 proteins have been,

00:46:33.880 | the structures have been found out about.

00:46:35.920 | So that was our training set,

00:46:37.120 | which is much less than normally we would like to use.

00:46:41.140 | But using various tricks, things like self-distillation.

00:46:43.840 | So actually using alpha fold predictions,

00:46:48.280 | some of the best predictions

00:46:49.480 | that it thought was highly confident in,

00:46:51.000 | we put them back into the training set, right?

00:46:53.320 | To make the training set bigger.

00:46:55.440 | That was critical to alpha fold working.

00:46:58.400 | So there was actually a huge number of different

00:47:00.840 | innovations like that that were required

00:47:03.560 | to ultimately crack the problem.

00:47:06.080 | Alpha fold one, what it produced was a distogram.

00:47:09.720 | So a kind of a matrix of the pair wise distances

00:47:13.620 | between all of the molecules in the protein.

00:47:17.760 | And then there had to be a separate optimization process

00:47:20.440 | to create the 3D structure.

00:47:23.640 | And what we did for alpha fold two

00:47:25.120 | is make it truly end to end.

00:47:26.900 | So we went straight from the amino acid sequence of bases

00:47:31.720 | to the 3D structure directly,

00:47:33.880 | without going through this intermediate step.

00:47:36.080 | And in machine learning, what we've always found is that

00:47:39.040 | the more end to end you can make it, the better the system.

00:47:42.160 | And it's probably because we, you know,

00:47:45.640 | in the end, the system's better at learning

00:47:47.480 | what the constraints are

00:47:48.520 | than we are as the human designers of specifying it.

00:47:51.880 | So anytime you can let it flow end to end

00:47:54.000 | and actually just generate what it is

00:47:55.360 | you're really looking for, in this case, the 3D structure,

00:47:58.400 | you're better off than having this intermediate step,

00:48:00.520 | which you then have to handcraft the next step for.

00:48:03.320 | So it's better to let the gradients and the learning

00:48:06.120 | flow all the way through the system from the endpoint,

00:48:08.960 | the end output you want to the inputs.

00:48:10.840 | - So that's a good way to start.

00:48:11.960 | I mean, you problem handcraft a bunch of stuff,

00:48:14.320 | add a bunch of manual constraints

00:48:16.640 | with a small end to end learning piece

00:48:18.640 | or a small learning piece and grow that learning piece

00:48:21.560 | until it consumes the whole thing.

00:48:22.840 | - That's right.

00:48:23.680 | And so you can also see, you know,

00:48:25.320 | this is a bit of a method we've developed

00:48:26.960 | over doing many sort of successful outfits.

00:48:29.600 | We call them Alpha X projects, right?

00:48:31.920 | Is, and the easiest way to see that

00:48:33.800 | is the evolution of AlphaGo to AlphaZero.

00:48:36.720 | So AlphaGo was a learning system,

00:48:39.640 | but it was specifically trained to only play Go, right?

00:48:42.280 | So, and what we wanted to do

00:48:44.200 | in the first version of AlphaGo

00:48:45.360 | is just get to world champion performance

00:48:47.520 | no matter how we did it, right?

00:48:49.200 | And then of course, AlphaGo Zero,

00:48:51.360 | we removed the need to use human games

00:48:54.080 | as a starting point, right?

00:48:55.480 | So it could just play against itself

00:48:57.960 | from random starting point from the beginning.

00:49:00.280 | So that removed the need for human knowledge about Go.

00:49:03.720 | And then finally AlphaZero then generalized it

00:49:05.960 | so that any things we had in there, the system,

00:49:08.920 | including things like symmetry of the Go board were removed.

00:49:12.240 | So that AlphaZero could play from scratch

00:49:14.600 | any two-player game.

00:49:15.600 | And then MuZero, which is the final,

00:49:17.480 | our latest version of that set of things,

00:49:19.640 | was then extending it so that you didn't even have to give it

00:49:22.120 | the rules of the game.

00:49:23.200 | It would learn that for itself.

00:49:24.880 | So it could also deal with computer games

00:49:26.600 | as well as board games.

00:49:27.760 | - So that line of AlphaGo, AlphaGo Zero, AlphaZero, MuZero,

00:49:31.840 | that's the full trajectory of what you can take

00:49:34.200 | from imitation learning to full self-supervised learning.

00:49:39.200 | - Yeah, exactly.

00:49:41.640 | And learning the entire structure of the environment

00:49:45.520 | you put in from scratch, right?

00:49:47.640 | And bootstrapping it through self-play yourself.

00:49:51.840 | But the thing is it would have been impossible, I think,

00:49:53.720 | or very hard for us to build AlphaZero or MuZero first

00:49:57.400 | out of the box.

00:49:58.600 | - Even psychologically, because you have to believe

00:50:01.400 | in yourself for a very long time.

00:50:03.040 | You're constantly dealing with doubt

00:50:04.640 | 'cause a lot of people say that it's impossible.

00:50:06.720 | - Exactly, so it was hard enough just to do Go.

00:50:08.640 | As you were saying, everyone thought that was impossible

00:50:10.760 | or at least a decade away from when we did it back in 2015,

00:50:15.200 | 24, you know, 2016.

00:50:17.320 | And so, yes, it would have been psychologically

00:50:20.960 | probably very difficult as well as the fact that,

00:50:23.080 | of course, we learn a lot by building AlphaGo first.

00:50:26.440 | Right, so I think this is why I call AI

00:50:28.520 | an engineering science.

00:50:29.920 | It's one of the most fascinating science disciplines,

00:50:32.280 | but it's also an engineering science in the sense that,

00:50:34.680 | unlike natural sciences, the phenomenon you're studying

00:50:38.200 | doesn't exist out in nature.

00:50:39.440 | You have to build it first.

00:50:40.880 | So you have to build the artifact first

00:50:42.480 | and then you can study and pull it apart and how it works.

00:50:46.480 | - This is tough to ask you this question

00:50:50.000 | 'cause you probably will say it's everything,

00:50:51.480 | but let's try to think through this

00:50:54.360 | because you're in a very interesting position

00:50:56.480 | where DeepMind is a place of some of the most brilliant ideas

00:51:00.320 | in the history of AI,

00:51:01.760 | but it's also a place of brilliant engineering.

00:51:04.580 | So how much of solving intelligence,

00:51:08.040 | this big goal for DeepMind, how much of it is science?

00:51:12.120 | How much is engineering?

00:51:13.320 | So how much is the algorithms?

00:51:14.720 | How much is the data?

00:51:16.160 | How much is the hardware compute infrastructure?

00:51:19.840 | How much is it the software compute infrastructure?

00:51:22.760 | What else is there?

00:51:24.800 | How much is the human infrastructure?

00:51:27.240 | And like just the humans interacting certain kinds of ways.

00:51:30.120 | It's based of all those ideas.

00:51:31.720 | How much is maybe like philosophy?

00:51:34.160 | What's the key?

00:51:35.080 | If you were to sort of look back,

00:51:40.680 | like if we go forward 200 years and look back,

00:51:43.200 | what was the key thing that solved intelligence?

00:51:46.320 | Is it the ideas or the engineering?

00:51:47.760 | - I think it's a combination.

00:51:49.080 | First of all, of course,

00:51:49.920 | it's a combination of all those things,

00:51:51.360 | but the ratios of them changed over time.

00:51:54.760 | So even in the last 12 years,

00:51:57.480 | so we started DeepMind in 2010,

00:51:59.420 | which is hard to imagine now because 2010,

00:52:01.960 | it's only 12 short years ago,

00:52:03.400 | but nobody was talking about AI.

00:52:05.120 | I don't know if you remember back to your MIT days,

00:52:07.600 | no one was talking about it.

00:52:08.720 | I did a postdoc at MIT back around then,

00:52:11.080 | and it was sort of thought of as a,

00:52:12.880 | well, look, we know AI doesn't work.

00:52:14.200 | We tried this hard in the '90s at places like MIT,

00:52:17.040 | mostly using logic systems and old-fashioned sort of,

00:52:20.280 | good old-fashioned AI, we would call it now.

00:52:22.600 | People like Minsky and Patrick Winston,

00:52:25.320 | and you know all these characters, right?

00:52:26.720 | And used to debate a few of them,

00:52:28.280 | and they used to think I was mad,

00:52:29.500 | thinking about that some new advance

00:52:31.020 | could be done with learning systems.

00:52:32.340 | And I was actually pleased to hear that

00:52:34.740 | because at least you know you're on a unique track

00:52:36.940 | at that point, right?

00:52:37.860 | Even if all of your professors are telling you you're mad.

00:52:41.860 | And of course, in industry,

00:52:43.860 | we couldn't get, you know,

00:52:44.700 | it was difficult to get two cents together,

00:52:47.660 | which is hard to imagine now as well,

00:52:48.940 | given that it's the biggest sort of buzzword in VCs

00:52:51.540 | and fundraising's easy and all these kinds of things today.

00:52:54.700 | So back in 2010, it was very difficult.

00:52:57.700 | And the reason we started then,

00:52:59.340 | and Shane and I used to discuss

00:53:01.080 | what were the sort of founding tenets of DeepMind,

00:53:04.900 | and it was various things.

00:53:06.120 | One was algorithmic advances.

00:53:08.680 | So deep learning, you know, Jeff Hinton and co.

00:53:11.140 | had just sort of invented that in academia,

00:53:13.140 | but no one in industry knew about it.

00:53:15.220 | We love reinforcement learning.

00:53:16.660 | We thought that could be scaled up.

00:53:18.260 | But also understanding about the human brain

00:53:20.160 | had advanced quite a lot in the decade prior,

00:53:23.920 | with fMRI machines and other things.

00:53:25.460 | So we could get some good hints

00:53:27.740 | about architectures and algorithms

00:53:29.700 | and sort of representations maybe that the brain uses.

00:53:33.420 | So at a systems level, not at a implementation level.

00:53:36.900 | And then the other big things were compute and GPUs, right?

00:53:41.060 | So we could see a compute was gonna be really useful

00:53:44.180 | and it got to a place where it become commoditized,

00:53:47.020 | mostly through the games industry,

00:53:48.580 | and that could be taken advantage of.

00:53:50.780 | And then the final thing was also mathematical

00:53:52.820 | and theoretical definitions of intelligence.

00:53:55.000 | So things like AIXI, A-I-X-E,

00:53:57.620 | which Shane worked on with his supervisor, Marcus Hutter,

00:54:00.220 | which is this sort of theoretical proof really

00:54:03.420 | of universal intelligence,

00:54:05.340 | which is actually a reinforcement learning system.

00:54:07.940 | In the limit, I mean, it assumes infinite compute

00:54:09.980 | and infinite memory in the way, you know,

00:54:11.460 | like a Turing machine proves.

00:54:12.960 | But I was also waiting to see something like that too,

00:54:15.900 | to, you know, like Turing machines and computation theory

00:54:19.500 | that people like Turing and Shannon came up with

00:54:21.580 | underpins modern computer science.

00:54:24.820 | You know, I was waiting for a theory like that

00:54:26.420 | to sort of underpin AGI research.

00:54:28.880 | So when I met Shane and saw he was working

00:54:30.940 | on something like that, you know,

00:54:32.180 | that to me was a sort of final piece of the jigsaw.

00:54:34.540 | So in the early days,

00:54:36.420 | I would say that ideas were the most important, you know,

00:54:40.460 | for us it was deep reinforcement learning,

00:54:42.460 | scaling up deep learning.

00:54:44.620 | Of course we've seen transformers.

00:54:46.240 | So huge leaps, I would say, you know,

00:54:48.100 | three or four from, if you think from 2010 till now,

00:54:51.500 | huge evolutions, things like AlphaGo.

00:54:54.640 | And maybe there's a few more still needed.

00:54:58.460 | But as we get closer to AI, AGI,

00:55:01.680 | I think engineering becomes more and more important

00:55:05.180 | and data because scale and of course the recent,

00:55:08.360 | you know, results of GPT-3 and all the big language models

00:55:11.020 | and large models, including our ones,

00:55:13.360 | has shown that scale is, and large models

00:55:16.560 | are clearly gonna be a necessary,

00:55:18.660 | but perhaps not sufficient part of an AGI solution.

00:55:22.520 | - And throughout that, like you said,

00:55:25.040 | and I'd like to give you a big thank you.

00:55:27.220 | You're one of the pioneers in this,

00:55:29.300 | is sticking by ideas like reinforcement learning,

00:55:32.660 | that this can actually work,

00:55:34.060 | given actually limited success in the past.

00:55:38.980 | And also, which we still don't know,

00:55:41.980 | but proudly having the best researchers in the world

00:55:46.980 | and talking about solving intelligence.

00:55:49.900 | So talking about whatever you call it,

00:55:51.460 | AGI or something like this, that speaking of MIT,

00:55:55.240 | that's just something you wouldn't bring up.

00:55:57.840 | - No.

00:55:58.680 | - Maybe you did in like 40, 50 years ago,

00:56:04.120 | but that was,

00:56:05.440 | AI was a place where you do tinkering,

00:56:09.880 | very small scale, not very ambitious projects.

00:56:13.120 | And maybe the biggest ambitious projects

00:56:16.720 | were in the space of robotics

00:56:18.020 | and doing like the DARPA challenge.

00:56:19.720 | But the task of solving intelligence

00:56:21.780 | and believing you can, that's really, really powerful.

00:56:24.900 | So in order for engineering to do its work,

00:56:28.020 | to have great engineers, build great systems,

00:56:31.280 | you have to have that belief,

00:56:32.680 | that threads throughout the whole thing,

00:56:34.280 | that you can actually solve

00:56:35.400 | some of these impossible challenges.

00:56:37.000 | - Yeah, that's right.

00:56:37.840 | And back in 2010, our mission statement,

00:56:41.080 | and still is today,

00:56:42.620 | it was used to be solving step one, solve intelligence,

00:56:45.920 | step two, use it to solve everything else.

00:56:47.840 | So if you can imagine pitching that to VC in 2010,

00:56:51.160 | the kind of looks we got,

00:56:52.680 | we managed to find a few kooky people to back us,

00:56:55.880 | but it was tricky.

00:56:57.680 | And I got to the point where we wouldn't mention it

00:57:00.180 | to any of our professors,

00:57:01.560 | because they would just eye roll

00:57:03.120 | and think we committed career suicide.

00:57:05.920 | So it was, there's a lot of things that we had to do,

00:57:10.040 | but we always believed it.

00:57:11.520 | And one reason, by the way,

00:57:13.240 | one reason I've always believed in reinforcement learning

00:57:16.160 | is that if you look at neuroscience,

00:57:19.160 | that is the way that the primate brain learns.

00:57:22.720 | One of the main mechanisms is the dopamine system

00:57:24.880 | implements some form of TD learning.

00:57:26.440 | It was a very famous result in the late '90s,

00:57:28.680 | where they saw this in monkeys,

00:57:31.200 | and as a proper game prediction error.

00:57:34.520 | So again, in the limit,

00:57:36.800 | this is what I think you can use neuroscience for,

00:57:38.800 | is in any, at mathematics,

00:57:41.480 | when you're doing something as ambitious

00:57:43.160 | as trying to solve intelligence,

00:57:44.560 | and you're, you know, it's blue sky research,

00:57:46.480 | no one knows how to do it.

00:57:47.760 | You need to use any evidence

00:57:50.200 | or any source of information you can

00:57:52.160 | to help guide you in the right direction

00:57:54.280 | or give you confidence you're going in the right direction.

00:57:56.680 | So that was one reason we pushed so hard on that.

00:57:59.880 | And just going back to your earlier question

00:58:01.880 | about organization,

00:58:03.160 | the other big thing that I think we innovated with

00:58:05.360 | at DeepMind to encourage invention and innovation

00:58:10.320 | was the multidisciplinary organization we built,

00:58:12.920 | and we still have today.

00:58:14.160 | So DeepMind originally was a confluence

00:58:16.680 | of the most cutting edge knowledge in neuroscience

00:58:19.400 | with machine learning, engineering, and mathematics, right?

00:58:22.840 | And gaming.

00:58:24.120 | And then since then, we've built that out even further.

00:58:26.760 | So we have philosophers here and, you know, ethicists,

00:58:30.280 | but also other types of scientists, physicists, and so on.

00:58:33.160 | And that's what brings together,

00:58:35.120 | I tried to build a sort of new type of Bell Labs,

00:58:38.720 | but in its golden era, right?

00:58:40.360 | And a new expression of that

00:58:43.800 | to try and foster this incredible sort of innovation machine.

00:58:48.440 | So talking about the humans in the machine,

00:58:50.560 | DeepMind itself is a learning machine

00:58:53.040 | with lots of amazing human minds in it

00:58:55.560 | coming together to try and build these learning systems.

00:58:58.880 | - If we return to the big ambitious dream of AlphaFold

00:59:04.920 | that may be the early steps

00:59:06.760 | on a very long journey in biology,

00:59:12.560 | do you think the same kind of approach

00:59:14.120 | can you use to predict the structure and function

00:59:16.360 | of more complex biological systems?

00:59:18.640 | So multi-protein interaction,

00:59:21.400 | and then, I mean, you can go out from there.

00:59:24.320 | Just simulating bigger and bigger systems

00:59:26.840 | that eventually simulate something like the human brain

00:59:29.480 | or the human body, just the big mush,

00:59:32.480 | the mess of the beautiful, resilient mess of biology.

00:59:36.360 | Do you see that as a long-term vision?

00:59:39.520 | - I do.

00:59:40.360 | I think, you know, if you think about what are the things,

00:59:43.440 | top things I wanted to apply AI to

00:59:45.620 | once we had powerful enough systems,

00:59:47.640 | biology and curing diseases and understanding biology

00:59:52.160 | was right up there, you know, top of my list.

00:59:54.080 | That's one of the reasons I personally pushed that myself

00:59:56.720 | and with AlphaFold.

00:59:58.040 | But I think AlphaFold, amazing as it is,

01:00:01.160 | is just the beginning.

01:00:02.960 | And I hope it's evidence of what could be done

01:00:07.120 | with computational methods.

01:00:08.780 | So, you know, AlphaFold solved this huge problem

01:00:12.160 | of the structure of proteins, but biology is dynamic.

01:00:15.200 | So really what I imagine from here,

01:00:16.880 | and we're working on all these things now,

01:00:18.620 | is protein-protein interaction, protein-ligand binding,

01:00:23.120 | so reacting with molecules.

01:00:25.360 | Then you wanna get built up to pathways,

01:00:27.600 | and then eventually a virtual cell.

01:00:29.960 | That's my dream, maybe in the next 10 years.

01:00:32.640 | And I've been talking actually to a lot of biologists,

01:00:34.520 | friends of mine, Paul Nurse, who runs the Crick Institute,

01:00:36.760 | amazing biologist, Nobel Prize-winning biologist.

01:00:39.100 | We've been discussing for 20 years now virtual cells.

01:00:42.100 | Could you build a virtual simulation of a cell?

01:00:44.740 | And if you could, that would be incredible

01:00:46.260 | for biology and disease discovery,

01:00:48.100 | 'cause you could do loads of experiments

01:00:49.500 | on the virtual cell, and then only at the last stage,

01:00:52.420 | validate it in the wet lab.

01:00:53.920 | So you could, you know, in terms of the search space

01:00:56.400 | of discovering new drugs, you know,

01:00:58.020 | it takes 10 years roughly to go from,

01:01:00.620 | to go from, you know, identifying a target

01:01:03.380 | to having a drug candidate.

01:01:06.480 | Maybe that could be shortened to, you know,

01:01:08.100 | by an order of magnitude with,

01:01:09.780 | if you could do most of that work in silico.

01:01:13.140 | So in order to get to a virtual cell,

01:01:15.780 | we have to build up understanding

01:01:18.360 | of different parts of biology and the interactions.

01:01:20.780 | And so, you know, every few years we talk about this,

01:01:24.580 | I talked about this with Paul.

01:01:25.620 | And then finally, last year after AlphaFold,

01:01:27.860 | I said, now's the time, we can finally go for it.

01:01:30.580 | And AlphaFold's the first proof point

01:01:32.380 | that this might be possible.

01:01:33.820 | And he's very excited, and we have some collaborations

01:01:35.940 | with his lab, they're just across the road actually

01:01:38.460 | from us, it's just, you know, wonderful being here

01:01:40.380 | in Kings Cross with the Crick Institute across the road.

01:01:42.900 | And I think the next steps, you know,

01:01:45.980 | I think there's gonna be some amazing advances in biology

01:01:48.660 | built on top of things like AlphaFold.

01:01:50.980 | We're already seeing that with the community doing that

01:01:53.140 | after we've open sourced it and released it.

01:01:56.020 | And, you know, I often say that I think,

01:02:00.140 | if you think of mathematics

01:02:02.300 | is the perfect description language for physics.

01:02:05.020 | I think AI might be, end up being

01:02:06.820 | the perfect description language for biology

01:02:09.180 | because biology is so messy, it's so emergent,

01:02:12.940 | so dynamic and complex.

01:02:15.220 | I think, I find it very hard to believe we'll ever get

01:02:17.460 | to something as elegant as Newton's laws of motions

01:02:20.180 | to describe a cell, right?

01:02:21.660 | It's just too complicated.

01:02:23.500 | So I think AI is the right tool for this.

01:02:26.060 | - You have to start at the basic building blocks

01:02:29.380 | and use AI to run the simulation

01:02:31.580 | for all those building blocks.

01:02:32.820 | So have a very strong way to do prediction

01:02:35.940 | of what given these building blocks,

01:02:37.700 | what kind of biology, how the function

01:02:40.780 | and the evolution of that biological system.

01:02:43.580 | It's almost like a cellular automata.

01:02:45.220 | You have to run it.

01:02:46.060 | You can't analyze it from a high level.

01:02:47.820 | You have to take the basic ingredients,

01:02:49.740 | figure out the rules and let it run.

01:02:51.860 | But in this case, the rules are very difficult

01:02:53.900 | to figure out. - Yes, exactly.

01:02:55.060 | - You have to learn them.

01:02:56.140 | - That's exactly it.

01:02:56.980 | So it's the biology is too complicated

01:02:58.940 | to figure out the rules.

01:03:00.700 | It's too emergent, too dynamic,

01:03:03.500 | say compared to a physics system

01:03:04.980 | like the motion of a planet, right?

01:03:06.940 | And so you have to learn the rules

01:03:09.100 | and that's exactly the type of systems that we're building.

01:03:11.820 | - So you mentioned you've open sourced AlphaFold

01:03:14.660 | and even the data involved.

01:03:16.500 | To me personally, also really happy

01:03:19.940 | and a big thank you for open sourcing with JoCo,

01:03:22.580 | the physics simulation engine

01:03:24.860 | that's often used for robotics research and so on.

01:03:28.980 | So I think that's a pretty gangster move.

01:03:31.060 | So what's the, (laughs)

01:03:33.260 | what's, I mean, this,

01:03:35.140 | very few companies or people do that kind of thing.

01:03:39.020 | What's the philosophy behind that?

01:03:41.140 | - You know, it's a case by case basis

01:03:42.860 | and in both those cases,

01:03:43.980 | we felt that was the maximum benefit to humanity to do that

01:03:47.300 | and the scientific community.

01:03:49.420 | In one case, the robotics physics community with Mojoco.

01:03:52.940 | So-- - We purchased it.

01:03:53.940 | - We purchased it-- - Open source.

01:03:55.540 | - Yes, we purchased it for the express principle

01:03:57.580 | to open source it.

01:03:58.500 | So, you know, I hope people appreciate that.

01:04:02.380 | It's great to hear that you do.

01:04:04.020 | And then the second thing was,

01:04:05.740 | and mostly we did it because the person building it

01:04:08.140 | was not able to cope with supporting it anymore

01:04:11.900 | 'cause it got too big for him.

01:04:13.540 | He's an amazing professor who built it in the first place.

01:04:16.700 | So we helped him out with that.

01:04:18.180 | And then with AlphaFold, even bigger, I would say,

01:04:20.460 | and I think in that case,

01:04:21.900 | we decided that there were so many downstream applications

01:04:25.460 | of AlphaFold that we couldn't possibly even imagine

01:04:29.340 | what they all were.

01:04:30.420 | So the best way to accelerate drug discovery

01:04:34.300 | and also fundamental research would be to give

01:04:38.500 | all that data away and the system itself.

01:04:42.380 | You know, it's been so gratifying to see

01:04:45.220 | what people have done that within just one year,

01:04:46.980 | which is a short amount of time in science.

01:04:49.180 | And it's being used by over 500,000 researchers have used it.

01:04:54.100 | We think that's almost every biologist in the world.

01:04:56.500 | I think there's roughly 500,000 biologists in the world,

01:04:58.820 | professional biologists,

01:04:59.940 | have used it to look at their proteins of interest.

01:05:03.260 | We've seen amazing fundamental research done.

01:05:06.500 | So a couple of weeks ago, front cover,

01:05:08.980 | there was a whole special history of science,

01:05:10.780 | including the front cover,

01:05:11.980 | which had the nuclear pore complex on it,

01:05:13.940 | which is one of the biggest proteins in the body.

01:05:15.740 | The nuclear pore complex is a protein that governs

01:05:18.900 | all the nutrients going in and out of your cell nucleus.

01:05:21.660 | So it's there like little hall gateways

01:05:23.540 | that open and close to let things go in

01:05:25.740 | and out of your cell nucleus.

01:05:27.260 | So they're really important,

01:05:28.700 | but they're huge because they're massive

01:05:29.940 | donut ring shaped things.

01:05:31.620 | And they've been looking to try and figure out

01:05:33.420 | that structure for decades.

01:05:34.940 | And they have lots of experimental data,

01:05:37.100 | but it's too low resolution, there's bits missing.

01:05:39.540 | And they were able to, like a giant Lego jigsaw puzzle,

01:05:43.060 | use alpha fold predictions plus experimental data

01:05:46.140 | and combined those two independent sources of information,

01:05:49.740 | actually four different groups around the world

01:05:51.220 | were able to put it together more or less simultaneously

01:05:54.580 | using alpha fold predictions.

01:05:56.260 | So that's been amazing to see.

01:05:57.700 | And pretty much every pharma company,

01:05:59.380 | every drug company executive I've spoken to

01:06:01.420 | has said that their teams are using alpha fold

01:06:03.740 | to accelerate whatever drugs they're trying to discover.

01:06:08.020 | So I think the knock on effect has been enormous

01:06:11.420 | in terms of the impact that alpha fold has made.

01:06:15.220 | - And it's probably bringing in, it's creating biologists,

01:06:17.820 | it's bringing more people into the field,

01:06:20.780 | both on the excitement

01:06:21.820 | and both on the technical skills involved.

01:06:24.580 | And it's almost like a gateway drug to biology.

01:06:28.780 | - Yes, it is.

01:06:29.620 | And to get more computational people involved too, hopefully.

01:06:32.660 | And I think for us, the next stage, as I said,

01:06:35.980 | future we have to have other considerations too.

01:06:37.980 | We're building on top of alpha fold

01:06:39.620 | and these other ideas I discussed with you

01:06:41.220 | about protein-protein interactions

01:06:42.780 | and genomics and other things.

01:06:44.820 | And not everything will be open source.

01:06:46.220 | Some of it we'll do commercially

01:06:48.020 | 'cause that will be the best way

01:06:49.020 | to actually get the most resources and impact behind it.

01:06:51.740 | In other ways, some other projects we'll do non-profit style.

01:06:55.300 | And also we have to consider for future things as well,

01:06:58.540 | safety and ethics as well, like synthetic biology,

01:07:01.620 | there is dual use and we have to think about that as well.

01:07:05.100 | With alpha fold, we consulted with 30 different bioethicists

01:07:08.620 | and other people expert in this field

01:07:10.260 | to make sure it was safe before we released it.

01:07:13.300 | So there'll be other considerations in future.

01:07:15.300 | But for right now, I think alpha fold is a kind of a gift

01:07:18.340 | from us to the scientific community.

01:07:20.860 | - So I'm pretty sure that something like alpha fold

01:07:24.220 | would be part of Nobel prizes in the future.

01:07:29.140 | But us humans, of course, are horrible with credit assignment

01:07:32.500 | so we'll of course give it to the humans.

01:07:34.540 | Do you think there will be a day when AI system

01:07:39.380 | can't be denied that it earned that Nobel prize?

01:07:45.140 | Do you think we will see that in 21st century?

01:07:47.460 | - It depends what type of AIs we end up building,

01:07:50.220 | whether they're goal seeking agents

01:07:53.580 | who specifies the goals, who comes up with the hypotheses,

01:07:57.820 | who determines which problems to tackle.

01:08:00.860 | - And tweets about it, announcement of the results.

01:08:02.420 | - Yes, it's announced the results exactly as part of it.

01:08:05.420 | So I think right now, of course,

01:08:07.900 | it's amazing human ingenuity that's behind these systems

01:08:12.180 | and then the system in my opinion is just a tool.

01:08:15.100 | Be a bit like saying with Galileo

01:08:17.260 | and his telescope, the ingenuity,

01:08:19.180 | the credit should go to the telescope.

01:08:21.140 | I mean, it's clearly Galileo building the tool

01:08:23.580 | which he then uses.

01:08:25.180 | So I still see that in the same way today

01:08:27.340 | even though these tools learn for themselves.

01:08:29.860 | I think of things like alpha fold

01:08:32.940 | and the things we're building as the ultimate tools

01:08:35.820 | for science and for acquiring new knowledge

01:08:38.580 | to help us as scientists acquire new knowledge.

01:08:41.140 | I think one day there will come a point

01:08:43.220 | where an AI system may solve

01:08:46.340 | or come up with something like general relativity

01:08:48.780 | of its own bat, not just by averaging everything

01:08:52.020 | on the internet or averaging everything on PubMed.

01:08:55.220 | Although that would be interesting to see

01:08:56.300 | what that would come up with.

01:08:58.500 | So that to me is a bit like our earlier debate

01:09:00.380 | about creativity, inventing Go

01:09:03.220 | rather than just coming up with a good Go move.

01:09:06.220 | And so I think solving, I think to,

01:09:10.380 | if we wanted to give it the credit

01:09:11.780 | of like a Nobel type of thing,

01:09:13.500 | then it would need to invent Go

01:09:15.740 | and sort of invent that new conjecture out of the blue

01:09:19.300 | rather than being specified by the human scientists

01:09:22.740 | or the human creators.

01:09:23.580 | So I think right now it's definitely just a tool.

01:09:26.300 | - Although it is interesting how far you get

01:09:27.900 | by averaging everything on the internet, like you said,

01:09:29.940 | because a lot of people do see science

01:09:33.140 | as you're always standing on the shoulders of giants.

01:09:35.620 | And the question is how much are you really reaching

01:09:40.060 | up above the shoulders of giants?

01:09:42.020 | Maybe it's just a simulating different kinds of results

01:09:46.260 | of the past with ultimately this new perspective

01:09:49.380 | that gives you this breakthrough idea.

01:09:51.100 | But that idea may not be novel

01:09:54.380 | in the way that it can't be already discovered

01:09:56.140 | on the internet.

01:09:56.980 | Maybe the Nobel prizes of the next hundred years

01:10:00.060 | are already all there on the internet to be discovered.

01:10:03.020 | - They could be, they could be.

01:10:04.540 | I mean, I think this is one of the big mysteries I think

01:10:08.940 | is that, first of all, I believe a lot of the big,

01:10:12.900 | new breakthroughs that are gonna come

01:10:14.380 | in the next few decades, and even in the last decade,

01:10:16.580 | are gonna come at the intersection

01:10:18.260 | between different subject areas

01:10:20.140 | where there'll be some new connection that's found

01:10:23.420 | between what seemingly were disparate areas.

01:10:26.140 | And one can even think of DeepMind, as I said earlier,

01:10:28.780 | as a sort of interdisciplinary between neuroscience ideas

01:10:31.700 | and AI engineering ideas originally.

01:10:35.020 | And so I think there's that.

01:10:37.900 | And then one of the things we can't imagine today is,

01:10:40.340 | and one of the reasons I think people,

01:10:41.660 | we were so surprised by how well large models worked,

01:10:44.380 | is that actually it's very hard for our human minds,

01:10:47.860 | our limited human minds to understand

01:10:49.380 | what it would be like to read the whole internet, right?

01:10:51.980 | I think we can do a thought experiment,

01:10:53.460 | and I used to do this, of like,

01:10:54.620 | well, what if I read the whole of Wikipedia?

01:10:57.540 | What would I know?

01:10:58.420 | And I think our minds can just about comprehend

01:11:00.420 | maybe what that would be like,

01:11:01.860 | but the whole internet is beyond comprehension.

01:11:04.380 | So I think we just don't understand what it would be like

01:11:07.380 | to be able to hold all of that in mind, potentially, right?

01:11:10.300 | And then active at once,

01:11:12.900 | and then maybe what are the connections

01:11:14.500 | that are available there?

01:11:15.740 | So I think no doubt there are huge things

01:11:17.460 | to be discovered just like that.

01:11:19.220 | But I do think there is this other type of creativity,

01:11:22.220 | of a true spark of new knowledge,

01:11:24.660 | new idea never thought before about,

01:11:26.620 | can't be averaged from things that are known,

01:11:29.260 | that really, of course, everything come,

01:11:31.900 | nobody creates in a vacuum,

01:11:33.580 | so there must be clues somewhere,

01:11:35.380 | but just a unique way of putting those things together.

01:11:38.260 | I think some of the greatest scientists in history

01:11:40.460 | have displayed that, I would say,

01:11:42.180 | although it's very hard to know, going back to their time,

01:11:45.060 | what was exactly known when they came up with those things.

01:11:48.060 | - Although, you're making me really think,

01:11:51.220 | because just the thought experiment

01:11:53.260 | of deeply knowing 100 Wikipedia pages,

01:11:57.300 | I don't think I can,

01:11:59.140 | I've been really impressed by Wikipedia

01:12:01.740 | for technical topics.

01:12:03.380 | So if you know 100 pages or 1,000 pages,

01:12:07.020 | I don't think you can visually, truly comprehend

01:12:10.100 | what kind of intelligence that is.

01:12:13.180 | That's a pretty powerful intelligence.

01:12:14.700 | If you know how to use that

01:12:16.100 | and integrate that information correctly,

01:12:18.340 | I think you can go really far.

01:12:20.020 | You can probably construct thought experiments

01:12:22.060 | based on that, like simulate different ideas.

01:12:25.860 | So if this is true, let me run this thought experiment,

01:12:28.860 | then maybe this is true.

01:12:30.180 | It's not really invention,

01:12:31.340 | it's just taking literally the knowledge

01:12:34.740 | and using it to construct

01:12:36.060 | a very basic simulation of the world.

01:12:37.900 | I mean, some argue it's romantic in part,

01:12:40.100 | but Einstein would do the same kind of things

01:12:42.380 | with thought experiments, right?

01:12:43.700 | - Yeah, one could imagine doing that systematically

01:12:46.300 | across millions of Wikipedia pages,

01:12:48.420 | plus PubMed, all these things.

01:12:50.340 | I think there are many, many things

01:12:52.980 | to be discovered like that that are hugely useful.

01:12:55.260 | You know, you could imagine,

01:12:56.180 | and I want us to do some of these things

01:12:57.580 | in material science, like room temperature superconductors

01:12:59.980 | is something on my list one day

01:13:01.500 | that I'd like to have an AI system to help build,

01:13:04.980 | better optimized batteries,

01:13:06.620 | all of these sort of mechanical things.

01:13:08.940 | I think a systematic sort of search

01:13:11.580 | could be guided by a model,

01:13:14.340 | could be extremely powerful.

01:13:17.100 | - So speaking of which, you have a paper on nuclear fusion,

01:13:20.160 | magnetic control of tachymeric plasmas

01:13:23.100 | through deep reinforcement learning.

01:13:24.700 | So you're seeking to solve nuclear fusion with deep RL.

01:13:29.780 | So it's doing control of high temperature plasmas.

01:13:31.820 | Can you explain this work?

01:13:33.500 | And can AI eventually solve nuclear fusion?

01:13:36.340 | - (laughs) It's been very fun last year or two,

01:13:39.380 | and very productive because we've been taking off

01:13:41.180 | a lot of my dream projects, if you like,

01:13:43.820 | of things that I've collected over the years

01:13:45.620 | of areas of science that I would like to,

01:13:48.140 | I think could be very transformative

01:13:49.660 | if we helped accelerate,

01:13:51.180 | and are really interesting problems,

01:13:53.580 | scientific challenges in and of themselves.

01:13:55.700 | - So this is energy.

01:13:57.020 | - So energy, yes, exactly.

01:13:58.500 | So energy and climate.

01:13:59.940 | So we talked about disease and biology

01:14:01.740 | as being one of the biggest places I think AI can help with.

01:14:04.500 | I think energy and climate is another one.

01:14:07.100 | So maybe they would be my top two.

01:14:09.220 | And fusion is one area I think AI can help with.

01:14:12.500 | Now, fusion has many challenges,

01:14:15.340 | mostly physics and material science

01:14:17.220 | and engineering challenges as well,

01:14:18.540 | to build these massive fusion reactors

01:14:20.460 | and contain the plasma.

01:14:21.860 | And what we try to do,

01:14:22.700 | and whenever we go into a new field to apply our systems,

01:14:26.220 | is we look for, we talk to domain experts,

01:14:29.180 | we try and find the best people in the world

01:14:30.620 | to collaborate with.

01:14:31.620 | In this case, in fusion,

01:14:34.060 | we collaborated with EPFL in Switzerland,

01:14:36.340 | the Swiss Technical Institute, who are amazing.

01:14:38.180 | They have a test reactor.

01:14:39.580 | They were willing to let us use,

01:14:41.300 | which I double-checked with the team

01:14:43.340 | we were gonna use carefully and safely.

01:14:46.060 | I was impressed they managed to persuade them

01:14:47.700 | to let us use it.

01:14:49.060 | And it's an amazing test reactor they have there.

01:14:53.380 | And they try all sorts of pretty crazy experiments on it.

01:14:56.940 | And what we tend to look at is,

01:14:59.660 | if we go into a new domain like fusion,

01:15:01.700 | what are all the bottleneck problems?

01:15:04.140 | Like thinking from first principles,

01:15:05.940 | what are all the bottleneck problems

01:15:06.980 | that are still stopping fusion working today?

01:15:09.260 | And then we look at, we get a fusion expert to tell us,

01:15:12.100 | and then we look at those bottlenecks,

01:15:13.740 | and we look at the ones,

01:15:14.580 | which ones are amenable to our AI methods today.

01:15:18.100 | - Yes.

01:15:18.940 | - And would be interesting from a research perspective,

01:15:22.220 | from our point of view, from an AI point of view.

01:15:24.420 | And that would address one of their bottlenecks.

01:15:26.740 | And in this case, plasma control was perfect.

01:15:29.700 | So, the plasma, it's a million degrees Celsius,

01:15:32.500 | something like that, hotter than the sun.

01:15:34.660 | And there's obviously no material that can contain it.

01:15:37.660 | So they have to be containing these magnetic,

01:15:39.460 | very powerful superconducting magnetic fields.

01:15:42.540 | But the problem is plasma, it's pretty unstable,

01:15:44.860 | as you imagine.

01:15:45.700 | You're kind of holding a mini sun, mini star in a reactor.

01:15:49.340 | So, you kind of want to predict ahead of time

01:15:52.540 | what the plasma is gonna do,

01:15:54.060 | so you can move the magnetic field within a few milliseconds

01:15:58.140 | to basically contain what it's gonna do next.

01:16:00.940 | So it seems like a perfect problem, if you think of it,

01:16:03.140 | for like a reinforcement learning prediction problem.

01:16:06.300 | So, you got a controller,

01:16:08.140 | you're gonna move the magnetic field.

01:16:09.740 | And until we came along,

01:16:11.380 | they were doing it with traditional operational

01:16:14.460 | research type of controllers,

01:16:16.740 | which are kind of handcrafted.

01:16:18.300 | And the problem is, of course, they can't react in the moment

01:16:20.460 | to something the plasma is doing.

01:16:21.620 | They have to be hard-coded.

01:16:23.020 | And again, knowing that that's normally our go-to solution

01:16:26.020 | is we would like to learn that instead.

01:16:27.940 | And they also had a simulator of these plasma.

01:16:30.300 | So there were lots of criteria

01:16:31.460 | that matched what we like to use.

01:16:34.740 | - So, can AI eventually solve nuclear fusion?

01:16:38.420 | - Well, so we, with this problem,

01:16:39.740 | and we published it in a Nature paper last year,

01:16:42.020 | we held the fusion,

01:16:43.820 | we held the plasma in a specific shapes.

01:16:46.180 | So actually, it's almost like carving the plasma

01:16:48.380 | into different shapes and hold it there

01:16:51.020 | for the record amount of time.

01:16:52.860 | So, that's one of the problems of fusion sort of solved.

01:16:57.580 | - So, have a controller that's able to,

01:16:59.820 | no matter the shape--

01:17:01.460 | - Contain it.

01:17:02.300 | - Contain it. - Yeah, contain it

01:17:03.180 | and hold it in structure.

01:17:04.180 | And there's different shapes that are better

01:17:05.780 | for the energy productions called droplets and so on.

01:17:10.060 | So, that was huge.

01:17:11.860 | And now we're looking,

01:17:12.740 | we're talking to lots of fusion startups

01:17:14.420 | to see what's the next problem we can tackle

01:17:17.420 | in the fusion area.

01:17:19.380 | - So, another fascinating place in a paper titled,

01:17:23.020 | "Pushing the Frontiers of Density Functionals

01:17:25.100 | by Solving the Fractional Electron Problem."

01:17:27.540 | So, you're taking on modeling and simulating

01:17:30.900 | the quantum mechanical behavior of electrons.

01:17:33.340 | - Yes.

01:17:34.180 | - Can you explain this work and can AI model

01:17:39.260 | and simulate arbitrary quantum mechanical systems

01:17:41.580 | in the future?

01:17:42.420 | - Yeah, so this is another problem I've had my eye on

01:17:44.260 | for a decade or more,

01:17:47.180 | which is sort of simulating the properties of electrons.

01:17:51.220 | If you can do that, you can basically describe

01:17:54.300 | how elements and materials and substances work.

01:17:58.060 | So, it's kind of like fundamental

01:18:00.020 | if you wanna advance material science.

01:18:02.860 | And we have Schrodinger's equation

01:18:05.260 | and then we have approximations

01:18:06.500 | to that density functional theory.

01:18:08.420 | These things are famous.

01:18:10.580 | And people try and write approximations

01:18:13.220 | to these functionals and kind of come up with descriptions

01:18:18.220 | of the electron clouds, where they're gonna go,

01:18:20.620 | how they're gonna interact

01:18:22.100 | when you put two elements together.

01:18:24.220 | And what we try to do is learn a simulation,

01:18:26.780 | learn a functional that will describe more chemistry,

01:18:30.540 | types of chemistry.

01:18:31.780 | So, until now, you can run expensive simulations,

01:18:35.580 | but then you can only simulate very small molecules,

01:18:38.780 | very simple molecules.

01:18:40.180 | We would like to simulate large materials.

01:18:43.060 | And so, today there's no way of doing that.

01:18:45.780 | And we're building up towards building functionals

01:18:48.580 | that approximate Schrodinger's equation

01:18:51.220 | and then allow you to describe what the electrons are doing.

01:18:55.580 | And all material sort of science and material properties

01:18:58.420 | are governed by the electrons and how they interact.

01:19:01.340 | - So, have a good summarization of the simulation

01:19:05.820 | through the functional,

01:19:07.100 | but one that is still close

01:19:11.340 | to what the actual simulation would come out with.

01:19:13.180 | So, how difficult is that task?

01:19:16.660 | What's involved in that task?

01:19:17.700 | Is it running those complicated simulations

01:19:20.700 | and learning the task of mapping

01:19:23.260 | from the initial conditions and the parameters

01:19:25.180 | of the simulation, learning what the functional would be?

01:19:27.660 | - Yeah, so it's pretty tricky.

01:19:29.380 | And we've done it with, you know,

01:19:31.260 | the nice thing is we can run a lot of the simulations,

01:19:35.100 | the molecular dynamic simulations on our compute clusters.

01:19:39.060 | And so, that generates a lot of data.

01:19:40.780 | So, in this case, the data is generated.

01:19:42.740 | So, we like those sort of systems.

01:19:44.700 | And that's why we use games, it's simulator generated data.

01:19:48.020 | And we can kind of create as much of it as we want really.

01:19:51.060 | And just let's leave some, you know,

01:19:53.220 | if any computers are free in the cloud,

01:19:55.180 | we just run some of these calculations, right?

01:19:57.580 | Compute cluster calculations.

01:19:59.260 | - I like how the free compute time

01:20:00.980 | is used up on quantum mechanics.

01:20:02.140 | - Yeah, quantum mechanics, exactly.

01:20:03.460 | Simulations and protein simulations and other things.

01:20:06.180 | And so, you know, when you're not searching on YouTube

01:20:09.780 | for video, cat videos,

01:20:11.260 | we're using those computers usefully in quantum chemistry.

01:20:13.900 | It's the idea.

01:20:14.740 | - Finally, it's been used for good.

01:20:15.580 | - And putting them to good use.

01:20:16.940 | And then, yeah, and then all of that computational data

01:20:19.700 | that's generated,

01:20:20.780 | we can then try and learn the functionals from that,

01:20:23.420 | which of course are way more efficient

01:20:25.620 | once we learn the functional

01:20:27.020 | than running those simulations would be.

01:20:30.540 | - Do you think one day AI may allow us to do something like

01:20:34.100 | basically crack open physics?

01:20:36.340 | So, do something like travel faster than the speed of light?

01:20:39.460 | - My ultimate aim has always been with AI is,

01:20:42.940 | the reason I am personally working on AI for my whole life,

01:20:46.260 | it was to build a tool to help us understand the universe.

01:20:50.300 | So I wanted to, and that means physics really,

01:20:53.740 | and the nature of reality.

01:20:54.860 | So I don't think we have systems

01:20:57.980 | that are capable of doing that yet,

01:20:59.340 | but when we get towards AGI,

01:21:00.940 | I think that's one of the first things

01:21:02.860 | I think we should apply AGI to.

01:21:05.260 | I would like to test the limits of physics

01:21:07.100 | and our knowledge of physics.

01:21:08.540 | There's so many things we don't know.

01:21:10.020 | This is one thing I find fascinating about science,

01:21:12.300 | and as a huge proponent of the scientific method

01:21:15.060 | as being one of the greatest ideas humanity's ever had

01:21:17.860 | and allowed us to progress with our knowledge.

01:21:20.140 | But I think as a true scientist,

01:21:21.980 | I think what you find is the more you find out,

01:21:25.180 | the more you realize we don't know.

01:21:27.020 | And I always think that it's surprising

01:21:29.860 | that more people aren't troubled.

01:21:31.860 | Every night I think about all these things

01:21:33.980 | we interact with all the time,

01:21:35.220 | that we have no idea how they work.

01:21:36.860 | Time, consciousness, gravity, life.

01:21:40.900 | We can't, I mean,

01:21:41.740 | these are all the fundamental things of nature.

01:21:43.860 | - I think the way we--

01:21:44.700 | - We don't really know what they are.

01:21:47.340 | - To live life, we pin certain assumptions on them

01:21:51.500 | and kind of treat our assumptions as if they're fact.

01:21:54.780 | - Yeah.

01:21:55.620 | - That allows us to sort of--

01:21:56.580 | - Box them off somehow.

01:21:57.580 | - Yeah, box them off somehow.

01:21:59.020 | But the reality is when you think of time,

01:22:02.340 | you should remind yourself,

01:22:03.540 | you should take it off the shelf

01:22:06.780 | and realize like, no, we have a bunch of assumptions.

01:22:09.020 | There's still a lot of,

01:22:10.100 | there's even now a lot of debate.

01:22:11.540 | There's a lot of uncertainty about exactly what is time.

01:22:15.540 | Is there an error of time?

01:22:17.500 | You know, there's a lot of fundamental questions

01:22:19.500 | that you can't just make assumptions about.

01:22:21.180 | And maybe AI allows you to not put anything on the shelf.

01:22:26.180 | - Yeah.

01:22:28.540 | - Not make any hard assumptions

01:22:30.220 | and really open it up and see what--

01:22:32.060 | - Exactly, I think we should be truly open-minded about that

01:22:34.660 | and exactly that, not be dogmatic to a particular theory.

01:22:39.060 | It'll also allow us to build better tools,

01:22:42.860 | experimental tools eventually,

01:22:45.420 | that can then test certain theories

01:22:47.300 | that may not be testable today.

01:22:49.020 | About things about like what we spoke about

01:22:51.780 | at the beginning about the computational nature

01:22:53.500 | of the universe, how one might, if that was true,

01:22:56.260 | how one might go about testing that, right?

01:22:58.340 | And how much, you know, there are people

01:23:00.740 | who've conjectured people like Scott Aronson

01:23:03.180 | and others about, you know, how much information

01:23:05.740 | can a specific Planck unit of space and time contain, right?

01:23:10.140 | So one might be able to think about testing those ideas

01:23:13.020 | if you had AI helping you build

01:23:16.460 | some new exquisite experimental tools.

01:23:20.500 | This is what I imagine, you know,

01:23:21.980 | many decades from now we'll be able to do.

01:23:24.140 | - And what kind of questions can be answered

01:23:26.860 | to running a simulation of them?

01:23:29.740 | So there's a bunch of physics simulations

01:23:31.820 | you can imagine that could be run

01:23:33.660 | in some kind of efficient way,

01:23:36.700 | much like you're doing in the quantum simulation work.

01:23:39.540 | And perhaps even the origin of life.

01:23:43.020 | So figuring out how going even back

01:23:46.020 | before the work of AlphaFold begins,

01:23:48.580 | of how this whole thing emerges from a rock.

01:23:53.380 | - Yes.

01:23:54.220 | - From a static thing.

01:23:55.060 | What do you, do you think AI will allow us to,

01:23:57.860 | is that something you have your eye on?

01:23:59.740 | It's trying to understand the origin of life.

01:24:02.340 | First of all, yourself, what do you think,

01:24:04.620 | how the heck did life originate on Earth?

01:24:08.780 | - Yeah, well, maybe I'll come to that in a second,

01:24:11.140 | but I think the ultimate use of AI

01:24:13.820 | is to kind of use it to accelerate science to the maximum.

01:24:18.100 | So I think of it a little bit like the tree of all knowledge.

01:24:22.620 | If you imagine that's all the knowledge

01:24:23.860 | there is in the universe to attain.

01:24:25.820 | And we sort of barely scratched the surface of that so far

01:24:29.340 | and even though we've done pretty well

01:24:31.980 | since the enlightenment, right, as humanity.

01:24:34.340 | And I think AI will turbocharge all of that

01:24:36.860 | like we've seen with AlphaFold.

01:24:38.620 | And I want to explore as much of that tree of knowledge

01:24:41.420 | as is possible to do.

01:24:42.940 | And I think that involves AI helping us

01:24:46.420 | with understanding or finding patterns,

01:24:49.660 | but also potentially designing and building new tools,

01:24:52.220 | experimental tools.

01:24:53.580 | So I think that's all,

01:24:54.840 | and also running simulations and learning simulations,

01:24:58.900 | all of that we're already,

01:25:00.220 | we're sort of doing at a baby steps level here.

01:25:04.980 | But I can imagine that in the decades to come

01:25:08.540 | as what's the full flourishing of that line of thinking.

01:25:12.940 | It's gonna be truly incredible, I would say.

01:25:15.140 | - If I visualize this tree of knowledge,

01:25:17.340 | something tells me that that knowledge,

01:25:19.580 | tree of knowledge for humans is much smaller.

01:25:21.980 | In the set of all possible trees of knowledge,

01:25:25.060 | it's actually quite small,

01:25:26.620 | given our cognitive limitations,

01:25:30.340 | limited cognitive capabilities,

01:25:33.680 | that even with the tools we build,

01:25:35.740 | we still won't be able to understand a lot of things.

01:25:38.100 | And that's perhaps what non-human systems

01:25:41.140 | might be able to reach farther, not just as tools,

01:25:44.820 | but in themselves understanding something

01:25:47.180 | that they can bring back.

01:25:48.500 | - Yeah, it could well be.

01:25:50.180 | So, I mean, there's so many things

01:25:51.780 | that are sort of encapsulated in what you just said there.

01:25:55.020 | I think, first of all, there's two different things.

01:25:58.340 | There's like, what do we understand today?

01:26:00.580 | What could the human mind understand?

01:26:02.680 | And what is the totality

01:26:03.920 | of what is there to be understood?

01:26:06.420 | And so there's three concentric,

01:26:08.620 | you can think of them as three larger and larger trees

01:26:10.740 | or exploring more branches of that tree.

01:26:12.900 | And I think with AI, we're gonna explore that whole lot.

01:26:15.940 | Now, the question is, if you think about

01:26:19.140 | what is the totality of what could be understood,

01:26:21.860 | there may be some fundamental physics reasons

01:26:24.820 | why certain things can't be understood,

01:26:26.340 | like what's outside a simulation or outside the universe.

01:26:29.040 | Maybe it's not understandable from within the universe.

01:26:32.360 | So there may be some hard constraints like that.

01:26:34.900 | - Could be smaller constraints,

01:26:36.020 | like we think of space-time as fundamental.

01:26:40.540 | Our human brains are really used to this idea

01:26:42.900 | of a three-dimensional world with time.

01:26:44.700 | - Right. - Maybe--

01:26:46.100 | - But our tools could go beyond that.

01:26:47.820 | They wouldn't have that limitation necessary.

01:26:49.780 | They could think in 11 dimensions, 12 dimensions,

01:26:51.820 | whatever is needed.

01:26:52.980 | But we could still maybe understand that

01:26:55.700 | in several different ways.

01:26:56.780 | The example I always give is,

01:26:59.100 | when I play Garry Kasparov at Speed Chess

01:27:01.460 | or we've talked about chess and these kinds of things,

01:27:04.160 | if you're reasonably good at chess,

01:27:07.580 | you can't come up with the move Garry comes up with

01:27:11.260 | in his move, but he can explain it to you.

01:27:13.380 | - And you can understand.

01:27:14.220 | - And you can understand post hoc the reasoning.

01:27:16.740 | So I think there's an even further level of like,

01:27:19.460 | well, maybe you couldn't have invented that thing,

01:27:21.700 | but going back to using language again,

01:27:24.340 | perhaps you can understand and appreciate that.

01:27:27.100 | Same way that you can appreciate Vivaldi or Mozart

01:27:30.220 | or something without, you can appreciate the beauty of that

01:27:32.740 | without being able to construct it yourself, right?

01:27:35.860 | Invent the music yourself.

01:27:37.460 | So I think we see this in all forms of life.

01:27:39.340 | So it will be that times a million,

01:27:42.480 | but you can imagine also one sign of intelligence

01:27:45.860 | is the ability to explain things clearly and simply, right?

01:27:49.540 | People like Richard Feynman,

01:27:50.460 | another one of my old time heroes used to say that, right?

01:27:52.420 | If you can't, if you can explain it something simply,

01:27:55.620 | then that's the best sign, a complex topic simply,

01:27:58.640 | then that's one of the best signs of you understanding it.

01:28:00.680 | - Yeah.

01:28:01.520 | I can see myself talking trash in the AI system in that way.

01:28:04.460 | It gets frustrated how dumb I am

01:28:07.780 | in trying to explain something to me.

01:28:09.900 | I was like, well, that means you're not intelligent

01:28:11.580 | because if you were intelligent,

01:28:12.740 | you'd be able to explain it simply.

01:28:14.420 | - Yeah, of course, there's also the other option,

01:28:16.700 | of course, we could enhance ourselves and without devices.

01:28:19.580 | We are already sort of symbiotic with our compute devices,

01:28:23.140 | right, with our phones and other things.

01:28:24.640 | And there's stuff like Neuralink and Acceptra

01:28:27.140 | that could advance that further.

01:28:30.020 | So I think there's lots of really amazing possibilities

01:28:33.900 | that I could foresee from here.

01:28:35.380 | - Well, let me ask you some wild questions.

01:28:37.020 | So out there looking for friends,

01:28:39.900 | do you think there's a lot of alien civilizations out there?

01:28:43.140 | - So I guess this also goes back

01:28:44.980 | to your origin of life question too,

01:28:46.700 | because I think that that's key.

01:28:49.400 | My personal opinion looking at all this,

01:28:51.460 | and it's one of my hobbies, physics, I guess.

01:28:53.780 | So it's something I think about a lot

01:28:56.980 | and talk to a lot of experts on and read a lot of books on.

01:29:00.860 | And I think my feeling currently is that we are alone.

01:29:05.380 | I think that's the most likely scenario

01:29:07.260 | given what evidence we have.

01:29:08.860 | So, and the reasoning is I think that we've tried

01:29:13.860 | since things like SETI program,

01:29:16.220 | and I guess since the dawning of the space age,

01:29:19.960 | we've had telescopes, open radio telescopes

01:29:22.400 | and other things.

01:29:23.400 | And if you think about and try to detect signals.

01:29:27.360 | Now, if you think about the evolution of humans on earth,

01:29:30.220 | we could have easily been a million years ahead

01:29:34.040 | of our time now or a million years behind, right?

01:29:36.560 | Easily with just some slightly different quirk

01:29:39.520 | thing happening hundreds of thousands of years ago,

01:29:42.480 | things could have been slightly different.

01:29:43.740 | If the meteor had hit the dinosaurs a million years earlier,

01:29:46.280 | maybe things would have evolved.

01:29:48.200 | We'd be a million years ahead of where we are now.

01:29:51.000 | So what that means is if you imagine where humanity will be

01:29:54.160 | in a few hundred years, let alone a million years,

01:29:56.800 | especially if we hopefully, you know,

01:29:59.920 | solve things like climate change and other things,

01:30:02.280 | and we continue to flourish and we build things like AI

01:30:05.720 | and we do space traveling and all of the stuff

01:30:08.000 | that humans have dreamed of forever, right?

01:30:10.880 | And sci-fi is talked about forever.

01:30:13.340 | We will be spreading across the stars, right?

01:30:16.820 | And Voigt-Neumann famously calculated, you know,

01:30:19.300 | it would only take about a million years

01:30:20.860 | if you send out Voigt-Neumann probes to the nearest,

01:30:23.300 | you know, the nearest other solar systems.

01:30:26.240 | And then they built, all they did was build

01:30:28.380 | two more versions of themselves and set those two out

01:30:30.500 | to the next nearest systems.

01:30:32.300 | You know, within a million years,

01:30:33.500 | I think you would have one of these probes

01:30:35.100 | in every system in the galaxy.

01:30:36.940 | So it's not actually in cosmological time.

01:30:40.080 | That's actually a very short amount of time.

01:30:42.120 | So, and, you know, we've, people like Dyson

01:30:44.340 | have thought about constructing Dyson spheres around stars

01:30:47.320 | to collect all the energy coming out of the star.

01:30:49.840 | You know, that, there would be constructions like that

01:30:51.820 | would be visible across space,

01:30:54.180 | probably even across a galaxy.

01:30:56.020 | So, and then, you know, if you think about

01:30:57.940 | all of our radio, television emissions

01:31:00.780 | that have gone out since the, you know, 30s and 40s,

01:31:04.220 | imagine a million years of that.

01:31:06.780 | And now hundreds of civilizations doing that.

01:31:10.020 | When we opened our ears at the point

01:31:12.240 | we got technologically sophisticated enough in the space age,

01:31:15.920 | we should have heard a cacophony of voices.

01:31:19.160 | We should have joined that cacophony of voices.

01:31:20.920 | And what we did, we opened our ears and we heard nothing.

01:31:24.520 | And many people who argue that there are aliens would say,

01:31:27.560 | well, we haven't really done exhaustive search yet.

01:31:29.960 | And maybe we're looking in the wrong bands

01:31:31.920 | and we've got the wrong devices

01:31:33.800 | and we wouldn't notice what an alien form was like

01:31:36.120 | 'cause it'd be so different to what we're used to.

01:31:38.280 | But, you know, I don't really buy that,

01:31:40.660 | that it shouldn't be as difficult as that.

01:31:42.620 | Like, I think we've searched enough.

01:31:44.340 | - It should be everywhere.

01:31:45.660 | - If it was, yeah, it should be everywhere.

01:31:47.300 | We should see Dyson spheres being put up,

01:31:49.220 | suns blinking in and out.

01:31:50.620 | You know, there should be a lot of evidence for those things.

01:31:52.900 | And then there are other people who argue,

01:31:54.180 | well, the sort of safari view of like,

01:31:55.980 | well, we're a primitive species still

01:31:57.820 | 'cause we're not space faring yet.

01:31:59.400 | And we're, you know, there's some kind of global,

01:32:01.420 | like universal rule not to interfere,

01:32:03.340 | you know, Star Trek rule.

01:32:04.580 | But like, look, we can't even coordinate humans

01:32:07.340 | to deal with climate change.

01:32:08.660 | And we're one species.

01:32:10.060 | What is the chance that of all of these different

01:32:12.420 | human civilization, you know, alien civilizations,

01:32:14.800 | they would have the same priorities

01:32:16.740 | and agree across, you know, these kinds of matters.

01:32:20.180 | And even if that was true,

01:32:21.860 | and we were in some sort of safari for our own good,

01:32:25.020 | to me, that's not much different

01:32:26.340 | from the simulation hypothesis.

01:32:27.620 | Because what does it mean, the simulation hypothesis?

01:32:29.860 | I think in its most fundamental level,

01:32:31.340 | it means what we're seeing is not quite reality, right?

01:32:34.940 | It's something, there's something more deeper underlying it,

01:32:37.740 | maybe computational.

01:32:39.060 | Now, if we were in a sort of safari park,

01:32:42.540 | and everything we were seeing was a hologram,

01:32:44.420 | and it was projected by the aliens or whatever,

01:32:46.460 | that to me is not much different than thinking

01:32:48.660 | we're inside of another universe.

01:32:50.220 | 'Cause we still can't see true reality, right?

01:32:53.100 | - I mean, there's other explanations.

01:32:55.060 | It could be that the way they're communicating

01:32:57.980 | is just fundamentally different.

01:32:59.260 | That we're too dumb to understand

01:33:01.180 | the much better methods of communication they have.

01:33:03.860 | It could be, I mean, it's silly to say,

01:33:06.580 | but our own thoughts could be the methods

01:33:09.940 | by which they're communicating.

01:33:11.180 | Like, the place from which our ideas,

01:33:13.220 | writers talk about this, like the muse.

01:33:15.180 | - Yeah.

01:33:16.020 | - The, I mean, it sounds like very kind of wild,

01:33:20.900 | but it could be thoughts,

01:33:22.140 | it could be some interactions with our mind

01:33:24.580 | that we think are originating from us

01:33:27.820 | is actually something that is coming

01:33:31.420 | from other life forms elsewhere.

01:33:33.020 | Consciousness itself might be that.

01:33:34.860 | - It could be, but I don't see any sensible argument

01:33:37.340 | to the why would all of the alien species behave this way?

01:33:41.580 | Yeah, some of them would be more primitive,

01:33:43.180 | they would be close to our level.

01:33:44.900 | You know, there should be a whole sort of

01:33:46.700 | normal distribution of these things, right?

01:33:48.700 | Some would be aggressive, some would be, you know,

01:33:50.940 | curious, others would be very historical and philosophical.

01:33:55.380 | Because, you know, maybe they're a million years

01:33:57.140 | older than us, but it's not, it shouldn't be like,

01:34:00.140 | I mean, one alien civilization might be like that,

01:34:02.980 | communicating thoughts and others,

01:34:04.180 | but I don't see why, you know,

01:34:05.780 | potentially the hundreds there should be

01:34:07.700 | would be uniform in this way, right?

01:34:10.020 | - It could be a violent dictatorship that the people,

01:34:13.020 | the alien civilizations that become successful,

01:34:17.620 | become, gain the ability to be destructive,

01:34:22.620 | an order of magnitude more destructive.

01:34:25.020 | But of course the sad thought,

01:34:29.860 | well, either humans are very special,

01:34:32.660 | we took a lot of leaps that arrived

01:34:35.500 | at what it means to be human.

01:34:36.940 | There's a question there, which was the hardest,

01:34:41.140 | which was the most special?

01:34:42.740 | But also if others have reached this level,

01:34:45.220 | and maybe many others have reached this level,

01:34:47.680 | the great filter that's prevented them from going farther,

01:34:52.680 | to becoming a multi-planetary species,

01:34:55.060 | or reaching out into the stars.

01:34:57.740 | And those are really important questions for us,

01:35:00.260 | whether there's other alien civilizations out there or not,

01:35:04.980 | this is very useful for us to think about.

01:35:07.260 | If we destroy ourselves, how will we do it?

01:35:10.500 | And how easy is it to do?

01:35:12.260 | - Yeah, well, you know, these are big questions,

01:35:14.420 | and I've thought about these a lot,

01:35:15.620 | but the interesting thing is that if we're alone,

01:35:19.860 | that's somewhat comforting from the great filter perspective,

01:35:22.280 | because it probably means the great filters are past us,

01:35:25.540 | and I'm pretty sure they are.

01:35:26.540 | So going back to your origin of life question,

01:35:29.340 | there are some incredible things

01:35:30.860 | that no one knows how happened.

01:35:32.060 | Like obviously the first life form from chemical soup,

01:35:35.540 | that seems pretty hard.

01:35:37.100 | But I would guess the multicellular,

01:35:39.020 | I wouldn't be that surprised if we saw single cell

01:35:42.460 | sort of life forms elsewhere, bacteria type things.

01:35:45.740 | But multicellular life seems incredibly hard,

01:35:48.300 | that step of capturing mitochondria,

01:35:50.500 | and then sort of using that as part of yourself,

01:35:53.420 | when you've just eaten it.

01:35:54.260 | - Would you say that's the biggest, the most,

01:35:57.820 | like if you had to choose one,

01:36:00.900 | sort of a hitchhiker's guide to the galaxy,

01:36:03.180 | one sentence summary of like,

01:36:04.660 | oh, those clever creatures did this,

01:36:07.540 | that would be the multicellular.

01:36:08.540 | - I think that was probably the one that's the biggest.

01:36:10.880 | I mean, there's a great book called

01:36:11.860 | "The 10 Great Inventions of Evolution" by Nick Lane,

01:36:15.060 | and he speculates on 10 of these,

01:36:17.300 | you know, what could be great filters.

01:36:20.100 | I think that's one.

01:36:20.980 | I think the advent of intelligence,

01:36:23.860 | and conscious intelligence,

01:36:25.460 | and in order to us to be able to do science

01:36:27.500 | and things like that, is huge as well.

01:36:29.860 | I mean, it's only evolved once as far as,

01:36:31.900 | you know, in Earth history.

01:36:34.820 | So that would be a later candidate,

01:36:37.140 | but there's certainly for the early candidates,

01:36:39.100 | I think multicellular life forms is huge.

01:36:41.420 | - By the way, it's interesting to ask you

01:36:43.540 | if you can hypothesize about

01:36:45.740 | what is the origin of intelligence?

01:36:47.960 | Is it that we started cooking meat over fire?

01:36:53.620 | Is it that we somehow figured out

01:36:55.500 | that we could be very powerful

01:36:57.020 | and we start collaborating?

01:36:58.100 | So cooperation between our ancestors

01:37:03.100 | so that we can overthrow the alpha male?

01:37:05.920 | What is it, Richard?

01:37:07.740 | I talked to Richard Ranham,

01:37:08.900 | who thinks we're all just beta males

01:37:10.740 | who figured out how to collaborate

01:37:12.060 | to defeat the dictator, the authoritarian alpha male

01:37:16.340 | that control the tribe.

01:37:19.140 | Is there other explanation?

01:37:20.900 | Was there a 2001 space odyssey?

01:37:23.340 | - A space odyssey type of monolith

01:37:24.820 | that came down to Earth?

01:37:25.900 | - Well, I think all of those things

01:37:28.100 | you suggested are good candidates.

01:37:29.220 | Fire and cooking, right?

01:37:31.300 | So that's clearly important for energy efficiency,

01:37:36.140 | cooking our meat and then being able

01:37:38.540 | to be more efficient about eating it

01:37:41.000 | and consuming the energy.

01:37:42.620 | I think that's huge.

01:37:44.340 | And then utilizing fire and tools.

01:37:46.340 | I think you're right about the tribal cooperation aspects

01:37:49.140 | and probably language is part of that

01:37:51.560 | because probably that's what allowed us

01:37:52.900 | to out-compete Neanderthals

01:37:54.180 | and perhaps less cooperative species.

01:37:56.700 | So that may be the case.

01:37:59.280 | Tool making, spears, axes.

01:38:01.660 | I think that let us, I mean,

01:38:03.180 | I think it's pretty clear now

01:38:04.420 | that humans were responsible

01:38:05.620 | for a lot of the extinctions of megafauna,

01:38:08.340 | especially in the Americas when humans arrived.

01:38:11.360 | So you can imagine once you discover tool usage,

01:38:14.940 | how powerful that would have been

01:38:16.260 | and how scary for animals.

01:38:18.020 | So I think all of those could have been explanations for it.

01:38:21.160 | Now, the interesting thing is

01:38:22.700 | that it's a bit like general intelligence too,

01:38:24.560 | is it's very costly to begin with to have a brain

01:38:28.620 | and especially a general purpose brain

01:38:30.020 | rather than a special purpose one.

01:38:31.420 | 'Cause the amount of energy our brains use,

01:38:32.820 | I think it's like 20% of the body's energy.

01:38:34.860 | And it's massive.

01:38:36.140 | And when you're thinking chess,

01:38:37.140 | one of the funny things that we used to say

01:38:39.460 | is it's as much as a racing driver uses

01:38:42.020 | for a whole Formula One race.

01:38:44.060 | Just playing a game of serious high-level chess,

01:38:46.380 | which you wouldn't think, just sitting there.

01:38:49.340 | Because the brain's using so much energy.

01:38:52.080 | So in order for an animal or an organism to justify that,

01:38:54.800 | there has to be a huge payoff.

01:38:57.880 | And the problem with half a brain

01:39:00.320 | or half intelligence, say an IQs of like a monkey brain,

01:39:05.320 | it's not clear you can justify that evolutionary

01:39:10.280 | until you get to the human level brain.

01:39:12.480 | And so, but how do you do that jump?

01:39:14.800 | It's very difficult,

01:39:15.640 | which is why I think it's only been done once

01:39:17.200 | from the sort of specialized brains that you see in animals

01:39:19.860 | to this sort of general purpose,

01:39:22.540 | cheering powerful brains that humans have.

01:39:25.180 | And which allows us to invent the modern world.

01:39:28.980 | And it takes a lot to cross that barrier.

01:39:33.660 | And I think we've seen the same with AI systems,

01:39:35.660 | which is that maybe until very recently,

01:39:38.260 | it's always been easier to craft a specific solution

01:39:40.960 | to a problem like chess

01:39:42.380 | than it has been to build a general learning system

01:39:44.540 | that could potentially do many things.

01:39:46.340 | 'Cause initially, that system will be way worse

01:39:49.520 | than less efficient than the specialized system.

01:39:52.160 | - So one of the interesting quirks of the human mind

01:39:55.880 | of this evolved system is that it appears to be conscious.

01:40:00.880 | This thing that we don't quite understand,

01:40:02.960 | but it seems very special,

01:40:06.600 | is ability to have a subjective experience

01:40:08.760 | that it feels like something to eat a cookie,

01:40:12.280 | the deliciousness of it or see a color

01:40:14.320 | and that kind of stuff.

01:40:15.560 | Do you think in order to solve intelligence,

01:40:17.940 | we also need to solve consciousness along the way?

01:40:20.700 | Do you think AGI systems need to have consciousness

01:40:23.940 | in order to be truly intelligent?

01:40:27.980 | - Yeah, we thought about this a lot actually.

01:40:29.640 | And I think that my guess is that consciousness

01:40:33.420 | and intelligence are double dissociable.

01:40:35.800 | So you can have one without the other both ways.

01:40:38.360 | And I think you can see that with consciousness in that,

01:40:41.600 | I think some animals, pets,

01:40:44.160 | if you have a pet dog or something like that,

01:40:46.280 | you can see some of the higher animals

01:40:47.920 | and dolphins, things like that,

01:40:50.240 | have self-awareness and are very sociable, seem to dream.

01:40:55.240 | Those kinds of, a lot of the traits one would regard

01:40:59.040 | as being kind of conscious and self-aware.

01:41:01.660 | But yet they're not that smart, right?

01:41:05.120 | So they're not that intelligent by say IQ standards

01:41:08.120 | or something like that.

01:41:08.960 | - Yeah, it's also possible that our understanding

01:41:11.120 | of intelligence is flawed, like putting an IQ to it.

01:41:14.600 | - Sure.

01:41:15.440 | - Maybe the thing that a dog can do

01:41:17.360 | is actually gone very far along the path of intelligence

01:41:20.640 | and we humans are just able to play chess

01:41:23.240 | and maybe write poems.

01:41:24.840 | - Right, but if we go back to the idea of AGI

01:41:27.040 | and general intelligence, dogs are very specialized, right?

01:41:29.480 | Most animals are pretty specialized.

01:41:30.920 | They can be amazing at what they do,

01:41:32.360 | but they're like kind of elite sports people or something.

01:41:35.600 | Right, so they do one thing extremely well

01:41:38.040 | 'cause their entire brain is optimized.

01:41:40.120 | - They have somehow convinced the entirety

01:41:41.920 | of the human population to feed them and service them.

01:41:44.560 | So in some way they're controlling.

01:41:46.440 | - Yes, exactly, well we co-evolved to some crazy degree,

01:41:49.360 | right, including the way the dogs, you know,

01:41:52.320 | even wag their tails and twitch their noses, right?

01:41:55.200 | We find it inexorably cute.

01:41:57.520 | But I think you can also see intelligence on the other side.

01:42:01.880 | So systems like artificial systems that are amazingly smart

01:42:06.200 | at certain things, like maybe playing Go and chess

01:42:08.720 | and other things, but they don't feel at all

01:42:11.760 | in any shape or form conscious in the way that,

01:42:14.480 | you know, you do to me or I do to you.

01:42:17.280 | And I think actually building AI,

01:42:21.320 | these intelligent constructs, is one of the best ways

01:42:25.440 | to explore the mystery of consciousness, to break it down.

01:42:28.040 | Because we're gonna have devices that are pretty smart

01:42:33.040 | at certain things or capable of certain things,

01:42:36.240 | but potentially won't have any semblance

01:42:39.160 | of self-awareness or other things.

01:42:40.800 | And in fact, I would advocate, if there's a choice,

01:42:43.880 | building systems in the first place, AI systems,

01:42:46.560 | that are not conscious to begin with,

01:42:48.640 | are just tools until we understand them better

01:42:52.440 | and the capabilities better.

01:42:53.960 | - So on that topic, just not as the CEO of DeepMind,

01:42:58.320 | just as a human being, let me ask you about this

01:43:01.480 | one particular anecdotal evidence of the Google engineer

01:43:05.320 | who made a comment or believed that there's some aspect

01:43:09.880 | of a language model, the Lambda language model,

01:43:13.240 | that exhibited sentience.

01:43:15.960 | So you said you believe there might be a responsibility

01:43:18.440 | to build systems that are not sentient.

01:43:21.120 | And this experience of a particular engineer,

01:43:23.520 | I think, I'd love to get your general opinion

01:43:25.880 | on this kind of thing, but I think it will happen

01:43:28.000 | more and more and more, which not when engineers,

01:43:31.480 | but when people out there that don't have

01:43:33.120 | an engineer background start interacting

01:43:34.760 | with increasingly intelligent systems,

01:43:37.120 | we anthropomorphize them, they start to have

01:43:40.600 | deep, impactful interactions with us in a way

01:43:45.000 | that we miss them when they're gone.

01:43:47.920 | And we sure as heck feel like they're living entities,

01:43:51.960 | self-aware entities, and maybe even we project sentience

01:43:55.200 | onto them.

01:43:56.040 | So what's your thought about this particular system?

01:43:59.960 | Have you ever met a language model that's sentient?

01:44:04.560 | - No, no, no.

01:44:06.320 | - What do you make of the case of when you kind of feel

01:44:10.160 | that there's some elements of sentience to this system?

01:44:12.920 | - Yeah, so this is an interesting question

01:44:15.040 | and obviously a very fundamental one.

01:44:17.760 | So first thing to say is I think that none of the systems

01:44:20.760 | we have today, I would say, even have one iota

01:44:23.640 | of semblance of consciousness or sentience.

01:44:26.320 | That's my personal feeling, interacting with them every day.

01:44:29.720 | So I think this way premature to be discussing

01:44:32.440 | what that engineer talked about.

01:44:34.160 | I think at the moment it's more of a projection

01:44:36.480 | of the way our own minds work, which is to see

01:44:39.080 | sort of purpose and direction in almost anything that we,

01:44:44.360 | you know, our brains are trained to interpret agency,

01:44:48.200 | basically, in things, even inanimate things sometimes.

01:44:52.280 | And of course, with a language system,

01:44:54.880 | 'cause language is so fundamental to intelligence,

01:44:57.080 | it's gonna be easy for us to anthropomorphize that.

01:45:00.440 | I mean, back in the day, even the first, you know,

01:45:03.840 | the dumbest sort of template chatbots ever,

01:45:05.800 | Eliza and the ilk of the original chatbots

01:45:09.200 | back in the '60s fooled some people

01:45:11.160 | under certain circumstances, right?

01:45:12.600 | It pretended to be a psychologist.

01:45:14.040 | So just basically rabbit back to you

01:45:16.080 | the same question you asked it back to you.

01:45:18.240 | And some people believe that.

01:45:21.280 | So I don't think we can, this is why I think

01:45:23.240 | the Turing test is a little bit flawed as a formal test

01:45:25.400 | because it depends on the sophistication of the judge,

01:45:29.200 | whether or not they are qualified to make that distinction.

01:45:33.240 | So I think we should talk to, you know,

01:45:36.760 | the top philosophers about this,

01:45:38.280 | people like Daniel Dennett and David Chalmers

01:45:40.800 | and others who've obviously thought deeply

01:45:42.480 | about consciousness.

01:45:43.640 | Of course, consciousness itself hasn't been well,

01:45:46.000 | there's no agreed definition.

01:45:47.720 | If I was to, you know, speculate about that,

01:45:51.040 | you know, I kind of, the working definition I like is,

01:45:55.080 | it's the way information feels when, you know,

01:45:57.200 | it gets processed.

01:45:58.040 | I think maybe Max Tegmark came up with that.

01:46:00.120 | I like that idea.

01:46:01.000 | I don't know if it helps us get towards

01:46:02.240 | any more operational thing,

01:46:03.880 | but I think it's a nice way of viewing it.

01:46:07.760 | I think we can obviously see from neuroscience

01:46:09.920 | certain prerequisites that are required,

01:46:11.680 | like self-awareness, I think is necessary,

01:46:14.400 | but not sufficient component.

01:46:16.040 | This idea of a self and other,

01:46:18.120 | and set of coherent preferences that are coherent over time.

01:46:22.480 | You know, these things are maybe memory.

01:46:24.800 | These things are probably needed

01:46:26.200 | for a sentient or conscious being.

01:46:29.320 | But the reason, the difficult thing I think for us

01:46:31.800 | when we get, and I think this is a really interesting

01:46:33.400 | philosophical debate, is when we get closer to AGI

01:46:37.280 | and, you know, and much more powerful systems

01:46:40.680 | than we have today,

01:46:42.240 | how are we going to make this judgment?

01:46:44.440 | And one way, which is the Turing test,

01:46:47.000 | is sort of a behavioral judgment.

01:46:48.640 | Is the system exhibiting all the behaviors

01:46:52.080 | that a human sentient or a sentient being would exhibit?

01:46:56.880 | Is it answering the right questions?

01:46:58.160 | Is it saying the right things?

01:46:59.160 | Is it indistinguishable from a human?

01:47:01.960 | And so on.

01:47:03.360 | But I think there's a second thing

01:47:05.760 | that makes us as humans regard each other as sentient.

01:47:09.040 | Right, why do we think this?

01:47:10.920 | And I debated this with Daniel Dennett.

01:47:12.720 | And I think there's a second reason

01:47:13.880 | that's often overlooked,

01:47:15.600 | which is that we're running on the same substrate.

01:47:18.040 | Right, so if we're exhibiting the same behavior,

01:47:21.120 | more or less, as humans,

01:47:22.680 | and we're running on the same, you know,

01:47:24.400 | carbon-based biological substrate,

01:47:26.200 | the squishy, you know, few pounds of flesh in our skulls,

01:47:29.560 | then the most parsimonious, I think, explanation

01:47:32.800 | is that you're feeling the same thing as I'm feeling.

01:47:35.040 | Right, but we will never have that second part,

01:47:37.840 | the substrate equivalence, with a machine.

01:47:40.680 | Right, so we will have to only judge based on the behavior.

01:47:43.960 | And I think the substrate equivalence

01:47:45.920 | is a critical part of why we make assumptions

01:47:48.200 | that we're conscious.

01:47:49.080 | And in fact, even with animals,

01:47:51.040 | high-level animals, why we think they might be,

01:47:52.680 | 'cause they're exhibiting some of the behaviors

01:47:54.160 | we would expect from a sentient animal,

01:47:55.920 | and we know they're made of the same things,

01:47:57.600 | biological neurons.

01:47:58.680 | So we're gonna have to come up with explanations

01:48:02.880 | or models of the gap between substrate differences

01:48:06.320 | between machines and humans

01:48:07.760 | to get anywhere beyond the behavioral.

01:48:10.840 | But to me, sort of the practical question

01:48:12.920 | is very interesting and very important.

01:48:16.080 | When you have millions, perhaps billions of people

01:48:18.680 | believing that you have a sentient AI,

01:48:20.840 | believing what that Google engineer believed,

01:48:23.080 | which I just see as an obvious,

01:48:26.360 | very near-term future thing,

01:48:28.760 | certainly on the path to AGI,

01:48:31.200 | how does that change the world?

01:48:33.160 | What's the responsibility of the AI system

01:48:35.280 | to help those millions of people?

01:48:37.120 | And also, what's the ethical thing?

01:48:39.800 | Because you can make a lot of people happy

01:48:44.120 | by creating a meaningful, deep experience

01:48:48.240 | with a system that's faking it before it makes it.

01:48:52.520 | - Yeah.

01:48:53.360 | I don't, are we the right,

01:48:56.320 | who is to say what's the right thing to do?

01:48:59.960 | Should AI always be tools?

01:49:01.880 | Like, why? - Sure.

01:49:02.720 | - Why are we constraining AIs to always be tools

01:49:06.080 | as opposed to friends?

01:49:07.920 | - Yeah, I think, well, I mean,

01:49:09.560 | these are fantastic questions and also critical ones.

01:49:14.080 | And we've been thinking about this

01:49:16.480 | since the start of DeepMind and before that

01:49:18.280 | because we planned for success,

01:49:19.800 | and however remote that looked like back in 2010.

01:49:24.800 | And we've always had sort of these ethical considerations

01:49:27.200 | as fundamental at DeepMind.

01:49:28.680 | And my current thinking on the language models

01:49:31.960 | is, and large models, is they're not ready,

01:49:34.160 | we don't understand them well enough yet.

01:49:36.720 | And in terms of analysis tools and guardrails,

01:49:40.480 | what they can and can't do and so on,

01:49:42.320 | to deploy them at scale.

01:49:44.000 | Because I think there are big, still ethical questions,

01:49:47.120 | like should an AI system always announce

01:49:48.920 | that it is an AI system to begin with?

01:49:50.880 | Probably yes.

01:49:51.840 | What do you do about answering those philosophical questions

01:49:55.800 | about the feelings people may have about AI systems,

01:49:59.080 | perhaps incorrectly attributed?

01:50:01.000 | So I think there's a whole bunch of research

01:50:03.120 | that needs to be done first to responsibly,

01:50:06.320 | before you can responsibly deploy these systems at scale.

01:50:09.400 | That will be at least be my current position.

01:50:12.320 | Over time, I'm very confident we'll have those tools,

01:50:15.360 | like interpretability questions,

01:50:18.160 | and analysis questions.

01:50:20.920 | And then with the ethical quandary,

01:50:23.480 | I think there it's important to look beyond just science.

01:50:28.480 | That's why I think philosophy, social sciences,

01:50:31.720 | even theology, other things like that come into it.

01:50:34.720 | Where arts and humanities,

01:50:37.400 | what does it mean to be human

01:50:39.120 | and the spirit of being human

01:50:40.320 | and to enhance that and the human condition, right?

01:50:43.680 | And allow us to experience things

01:50:45.080 | we could never experience before

01:50:46.400 | and improve the overall human condition

01:50:49.080 | and humanity overall, get radical abundance,

01:50:51.640 | solve many scientific problems, solve disease.

01:50:54.120 | So this is the era I think,

01:50:55.240 | this is the amazing era I think we're heading into

01:50:57.520 | if we do it right.

01:50:58.560 | But we've got to be careful.

01:51:00.800 | We've already seen with things like social media,

01:51:02.680 | how dual use technologies can be misused by,

01:51:05.920 | firstly, by bad actors or naive actors or crazy actors, right?

01:51:10.920 | So there's that set of just the common or garden misuse

01:51:15.680 | of existing dual use technology.

01:51:18.000 | And then of course, there's an additional thing

01:51:20.960 | that has to be overcome with AI

01:51:21.960 | that eventually it may have its own agency.

01:51:24.480 | So it could be good or bad in of itself.

01:51:28.720 | So I think these questions have to be approached

01:51:31.480 | very carefully using the scientific method, I would say,

01:51:35.360 | in terms of hypothesis generation, careful control testing,

01:51:38.680 | not live A/B testing out in the world,

01:51:40.720 | because with powerful technologies like AI,

01:51:44.400 | if something goes wrong, it may cause a lot of harm

01:51:47.680 | before you can fix it.

01:51:49.160 | It's not like an imaging app or game app

01:51:52.040 | where if something goes wrong, it's relatively easy to fix

01:51:56.200 | and the harm is relatively small.

01:51:58.000 | So I think it comes with the usual cliche

01:52:02.600 | of like with a lot of power comes a lot of responsibility.

01:52:05.280 | And I think that's the case here with things like AI

01:52:07.840 | given the enormous opportunity in front of us.

01:52:11.120 | And I think we need a lot of voices

01:52:14.120 | and as many inputs into things like the design of the systems

01:52:17.920 | and the values they should have

01:52:19.880 | and what goals should they be put to.

01:52:22.360 | I think as wide a group of voices as possible

01:52:24.520 | beyond just the technologists is needed

01:52:26.760 | to input into that and to have a say in that,

01:52:29.080 | especially when it comes to deployment of these systems,

01:52:31.800 | which is when the rubber really hits the road,

01:52:33.440 | it really affects the general person in the street

01:52:35.440 | rather than fundamental research.

01:52:37.360 | And that's why I say, I think as a first step,

01:52:40.240 | it would be better if we have the choice

01:52:42.360 | to build these systems as tools to give.

01:52:45.120 | And I'm not saying that it should never,

01:52:46.680 | they should never go beyond tools

01:52:47.960 | 'cause of course the potential is there

01:52:50.360 | for it to go way beyond just tools.

01:52:52.960 | But I think that would be a good first step

01:52:55.800 | in order for us to allow us to carefully experiment

01:52:58.880 | and understand what these things can do.

01:53:01.000 | - So the leap between tool to sentient entity being

01:53:05.800 | is one we should take very care of.

01:53:07.240 | - Yes.

01:53:08.280 | - Let me ask a dark personal question.

01:53:11.120 | So you're one of the most brilliant people

01:53:13.480 | in the AI community, also one of the most kind

01:53:16.800 | and if I may say sort of loved people in the community.

01:53:20.860 | That said, creation of a super intelligent AI system

01:53:25.860 | would be one of the most powerful things in the world,

01:53:32.680 | tools or otherwise.

01:53:34.820 | And again, as the old saying goes,

01:53:37.560 | power corrupts and absolute power corrupts absolutely.

01:53:41.620 | You are likely to be one of the people,

01:53:46.620 | I would say probably the most likely person

01:53:50.280 | to be in the control of such a system.

01:53:53.240 | Do you think about the corrupting nature of power

01:53:57.140 | when you talk about these kinds of systems

01:53:59.540 | that as all dictators and people have caused atrocities

01:54:04.540 | in the past always think they're doing good.

01:54:07.780 | But they don't do good because the powers polluted

01:54:10.940 | their mind about what is good and what is evil.

01:54:13.700 | Do you think about this stuff

01:54:14.820 | or are we just focused on language model?

01:54:16.420 | - No, I think about them all the time.

01:54:18.700 | And I think what are the defenses against that?

01:54:22.340 | I think one thing is to remain very grounded

01:54:24.820 | and sort of humble no matter what you do or achieve.

01:54:28.780 | And I try to do that.

01:54:30.380 | My best friends are still my set of friends

01:54:32.180 | from my undergraduate Cambridge days.

01:54:34.700 | My family and friends are very important.

01:54:38.100 | I've always, I think trying to be a multidisciplinary person

01:54:42.380 | it helps to keep you humble

01:54:43.780 | because no matter how good you are at one topic,

01:54:45.920 | someone will be better than you at that.

01:54:47.620 | And always relearning a new topic again from scratch

01:54:50.980 | is a new field is very humbling.

01:54:53.380 | So for me, that's been biology over the last five years.

01:54:56.420 | Huge area topic and it's been,

01:54:59.020 | and I just love doing that,

01:55:00.260 | but it helps to keep you grounded

01:55:01.620 | like and keeps you open-minded.

01:55:04.340 | And then the other important thing

01:55:06.380 | is to have a really group,

01:55:07.660 | amazing set of people around you at your company

01:55:10.820 | or your organization who are also very ethical

01:55:13.660 | and grounded themselves and help to keep you that way.

01:55:16.840 | And then ultimately, just to answer your question,

01:55:18.880 | I hope we're gonna be a big part of birthing AI

01:55:22.020 | and that being the greatest benefit to humanity

01:55:24.460 | of any tool or technology ever

01:55:26.820 | and getting us into a world of radical abundance

01:55:29.540 | and curing diseases

01:55:32.100 | and solving many of the big challenges we have

01:55:34.260 | in front of us and then ultimately,

01:55:36.380 | help the ultimate flourishing of humanity

01:55:38.260 | to travel the stars and find those aliens if they are there.

01:55:41.180 | And if they're not there, find out why they're not there,

01:55:43.500 | what is going on here in the universe.

01:55:45.580 | This is all to come

01:55:47.380 | and that's what I've always dreamed about.

01:55:49.460 | But I think AI is too big an idea.

01:55:53.020 | It's not going to be,

01:55:54.780 | there'll be a certain set of pioneers who get there first.

01:55:56.980 | I hope we're in the vanguard

01:55:58.600 | so we can influence how that goes.

01:56:00.380 | And I think it matters who builds,

01:56:02.460 | which cultures they come from and what values they have,

01:56:06.500 | the builders of AI systems.

01:56:07.860 | 'Cause I think even though the AI system

01:56:09.300 | is gonna learn for itself most of its knowledge,

01:56:11.580 | there'll be a residue in the system of the culture

01:56:14.780 | and the values of the creators of that system.

01:56:17.140 | And there's interesting questions

01:56:18.700 | to discuss about that geopolitically,

01:56:21.580 | different cultures as we're in a more fragmented world

01:56:23.820 | than ever unfortunately,

01:56:24.900 | I think in terms of global cooperation,

01:56:27.500 | we see that in things like climate

01:56:29.220 | where we can't seem to get our act together globally

01:56:31.980 | to cooperate on these pressing matters.

01:56:34.060 | I hope that will change over time.

01:56:35.580 | Perhaps if we get to an era of radical abundance,

01:56:38.600 | we don't have to be so competitive anymore.

01:56:40.420 | Maybe we can be more cooperative

01:56:42.620 | if resources aren't so scarce.

01:56:44.300 | - It's true that in terms of power corrupting

01:56:48.220 | and leading to destructive things,

01:56:50.020 | it seems that some of the atrocities of the past

01:56:52.780 | happen when there's a significant constraint on resources.

01:56:56.660 | - I think that's the first thing.

01:56:57.540 | I don't think that's enough.

01:56:58.380 | I think scarcity is one thing that's led to competition,

01:57:02.420 | sort of zero sum game thinking.

01:57:03.980 | I would like us to all be in a positive sum world.

01:57:06.080 | And I think for that, you have to remove scarcity.

01:57:08.460 | I don't think that's enough unfortunately to get world peace

01:57:10.780 | because there's also other corrupting things

01:57:12.780 | like wanting power over people and this kind of stuff,

01:57:15.460 | which is not necessarily satisfied by just abundance,

01:57:18.980 | but I think it will help.

01:57:20.240 | But I think ultimately AI is not gonna be run

01:57:24.860 | by any one person or one organisation.

01:57:26.740 | I think it should belong to the world,

01:57:28.020 | belong to humanity.

01:57:29.580 | And I think there'll be many ways this will happen.

01:57:33.100 | And ultimately, everybody should have a say in that.

01:57:36.840 | - Do you have advice for young people

01:57:41.300 | in high school and college,

01:57:43.060 | maybe if they're interested in AI

01:57:45.820 | or interested in having a big impact on the world,

01:57:50.700 | what they should do to have a career they can be proud of

01:57:53.220 | or to have a life they can be proud of?

01:57:55.060 | - I love giving talks to the next generation.

01:57:57.460 | What I say to them is actually two things.

01:57:59.180 | I think the most important things to learn about

01:58:02.420 | and to find out about when you're young

01:58:04.540 | is what are your true passions is first of all,

01:58:07.140 | as two things.

01:58:07.960 | One is find your true passions.

01:58:09.740 | And I think you can do that by,

01:58:11.860 | the way to do that is to explore as many things as possible

01:58:14.660 | when you're young and you have the time

01:58:16.540 | and you can take those risks.

01:58:19.180 | I would also encourage people to look at the,

01:58:21.100 | finding the connections between things in a unique way.

01:58:24.620 | I think that's a really great way to find a passion.

01:58:27.300 | Second thing I would say, advise is know yourself.

01:58:30.660 | So spend a lot of time understanding how you work best.

01:58:35.620 | Like what are the optimal times to work?

01:58:37.740 | What are the optimal ways that you study?

01:58:39.900 | What are your, how do you deal with pressure?

01:58:42.300 | Sort of test yourself in various scenarios

01:58:44.580 | and try and improve your weaknesses,

01:58:47.260 | but also find out what your unique skills and strengths are

01:58:50.740 | and then hone those.

01:58:52.180 | So then that's what will be your super value

01:58:54.540 | in the world later on.

01:58:55.900 | And if you can then combine those two things

01:58:57.860 | and find passions that you're genuinely excited about,

01:59:01.220 | that intersect with what your unique strong skills are,

01:59:05.360 | then you're onto something incredible.

01:59:07.860 | And I think you can make a huge difference in the world.

01:59:10.900 | - So let me ask about know yourself.

01:59:12.740 | This is fun.

01:59:13.580 | This is fun.

01:59:14.420 | Quick questions about day in the life, the perfect day,

01:59:18.140 | the perfect productive day in the life of Demis' house.

01:59:21.180 | Maybe these days you're, there's a lot involved.

01:59:26.180 | So maybe a slightly younger Demis' house,

01:59:29.020 | where you could focus on a single project maybe.

01:59:31.420 | How early do you wake up?

01:59:34.460 | Are you a night owl?

01:59:35.620 | Do you wake up early in the morning?

01:59:36.780 | What are some interesting habits?

01:59:39.180 | How many dozens of cups of coffees do you drink a day?

01:59:42.420 | What's the computer that you use?

01:59:46.340 | What's the setup?

01:59:47.180 | How many screens?

01:59:48.000 | What kind of keyboard?

01:59:49.140 | Are we talking Emacs Vim?

01:59:51.420 | Are we talking something more modern?

01:59:53.340 | So there's a bunch of those questions.

01:59:54.500 | So maybe day in the life.

01:59:56.780 | - Yes.

01:59:57.620 | - What's the perfect day involved?

01:59:58.940 | - Well, these days it's quite different

02:00:00.860 | from say 10, 20 years ago.

02:00:02.660 | Back 10, 20 years ago, it would have been a whole day

02:00:06.300 | of research, individual research or programming,

02:00:10.860 | doing some experiment, neuroscience,

02:00:12.540 | computer science experiment,

02:00:14.020 | reading lots of research papers.

02:00:16.620 | And then perhaps at nighttime,

02:00:18.380 | reading science fiction books or playing some games.

02:00:24.700 | - But lots of focus, so deep focused work

02:00:28.340 | on whether it's programming or reading research papers.

02:00:31.820 | - Yes, yes.

02:00:32.660 | So that would be lots of deep focus work.

02:00:35.260 | These days, for the last sort of, I guess,

02:00:38.060 | five to 10 years, I've actually got quite a structure

02:00:40.980 | that works very well for me now,

02:00:42.300 | which is that I'm a complete night owl, always have been.

02:00:46.100 | So I optimize for that.

02:00:47.660 | So I basically do a normal day's work,

02:00:50.740 | get into work about 11 o'clock

02:00:52.540 | and sort of do work till about seven in the office.

02:00:56.380 | And I will arrange back-to-back meetings

02:00:58.940 | for the entire time of that.

02:01:00.900 | And with as many, meet as many people as possible.

02:01:03.180 | So that's my collaboration management part of the day.

02:01:06.460 | Then I go home, spend time with the family and friends,

02:01:10.620 | have dinner, relax a little bit.

02:01:13.580 | And then I start a second day of work.

02:01:15.220 | I call it my second day of work around 10 p.m., 11 p.m.

02:01:18.500 | And that's the time till about the small hours

02:01:20.860 | of the morning, four or five in the morning,

02:01:22.540 | where I will do my thinking and reading research,

02:01:26.500 | writing research papers.

02:01:28.060 | Sadly, I don't have time to code anymore,

02:01:30.980 | but it's not efficient to do that these days,

02:01:34.900 | given the amount of time I have.

02:01:37.140 | But that's when I do, maybe do the long kind of stretches

02:01:40.740 | of thinking and planning.

02:01:42.460 | And then probably, using email or other things,

02:01:45.260 | I would fire off a lot of things to my team

02:01:47.900 | to deal with the next morning.

02:01:49.380 | But actually, thinking about this overnight,

02:01:51.620 | we should go for this project

02:01:53.220 | or arrange this meeting the next day.

02:01:54.860 | - When you're thinking through a problem,

02:01:56.140 | are you talking about a sheet of paper?

02:01:58.140 | Is there some structured process?

02:02:01.060 | - I still like pencil and paper best for working out things,

02:02:04.340 | but these days it's just so efficient

02:02:06.740 | to read research papers just on the screen.

02:02:08.740 | I still often print them out, actually.

02:02:10.220 | I still prefer to mark out things.

02:02:12.500 | And I find it goes into the brain quicker,

02:02:14.460 | better and sticks in the brain better

02:02:15.980 | when you're still using physical pen and pencil and paper.

02:02:19.420 | - So you take notes with the--

02:02:20.780 | - I have lots of notes, electronic ones,

02:02:22.420 | and also whole stacks of notebooks that I use at home, yeah.

02:02:27.420 | - On some of these most challenging next steps,

02:02:29.820 | for example, stuff none of us know about

02:02:32.620 | that you're working on, you're thinking,

02:02:35.540 | there's some deep thinking required there, right?

02:02:37.580 | Like what is the right problem?

02:02:39.380 | What is the right approach?

02:02:41.260 | Because you're gonna have to invest a huge amount of time

02:02:43.860 | for the whole team.

02:02:44.780 | They're going to have to pursue this thing.

02:02:46.700 | What's the right way to do it?

02:02:48.500 | Is RL gonna work here or not?

02:02:50.020 | - Yes.

02:02:50.860 | - What's the right thing to try?

02:02:53.100 | What's the right benchmark to use?

02:02:55.020 | Do we need to construct a benchmark from scratch?

02:02:57.260 | All those kinds of things.

02:02:58.140 | - Yes, so I think of all those kind of things

02:03:00.180 | in the night time phase, but also much more,

02:03:03.420 | I find I've always found the quiet hours of the morning

02:03:07.620 | when everyone's asleep, it's super quiet outside.

02:03:11.420 | I love that time, it's the golden hours,

02:03:13.380 | like between like one and three in the morning.

02:03:16.500 | Put some music on, some inspiring music on,

02:03:18.900 | and then think these deep thoughts.

02:03:21.580 | So that's when I would read my philosophy books

02:03:24.220 | and Spinoza's my recent favorite, Kant, all these things.

02:03:28.820 | And I read about a great scientist of history,

02:03:33.660 | how they did things, how they thought things.

02:03:35.660 | So that's when you do all your create,

02:03:37.220 | that's when I do all my creative thinking.

02:03:39.140 | And it's good, I think people recommend

02:03:41.780 | you do your sort of creative thinking in one block.

02:03:45.100 | And the way I organize the day,

02:03:47.140 | that way I don't get interrupted

02:03:48.540 | 'cause obviously no one else is up at those times.

02:03:51.460 | So I can go, I can sort of get super deep

02:03:55.900 | and super into flow.

02:03:57.580 | The other nice thing about doing it night time wise

02:03:59.620 | is if I'm really onto something

02:04:02.780 | or I've got really deep into something,

02:04:04.940 | I can choose to extend it

02:04:06.860 | and I'll go into six in the morning, whatever,

02:04:08.980 | and then I'll just pay for it the next day.

02:04:10.780 | So I'll be a bit tired and I won't be my best,

02:04:12.980 | but that's fine.

02:04:13.900 | I can decide, looking at my schedule the next day

02:04:16.660 | that I'm given where I'm at with this particular thought

02:04:19.380 | or creative idea that I'm gonna pay that cost the next day.

02:04:22.820 | So I think that's more flexible

02:04:25.380 | than morning people who do that.

02:04:27.660 | They get up at four in the morning,

02:04:28.780 | they can also do those golden hours then,

02:04:31.020 | but then their start of their scheduled day

02:04:32.660 | starts at breakfast, you know, 8 a.m., whatever,

02:04:34.780 | they have their first meeting.

02:04:36.020 | And then it's hard,

02:04:36.860 | you have to reschedule a day if you're in flow.

02:04:38.980 | - Yeah, that could be a true special thread of thoughts

02:04:41.900 | that you're too passionate about.

02:04:45.140 | This is where some of the greatest ideas

02:04:46.740 | could potentially come

02:04:47.740 | is when you just lose yourself late into the night.

02:04:51.380 | And for the meetings, I mean,

02:04:53.180 | you're loading in really hard problems

02:04:54.860 | in a very short amount of time.

02:04:56.500 | So you have to do some kind of first principles thinking

02:04:58.820 | here, it's like, what's the problem?

02:05:00.180 | What's the state of things?

02:05:01.340 | What's the right next step?

02:05:03.140 | - You have to get really good at context switching,

02:05:05.100 | which is one of the hardest things,

02:05:07.220 | 'cause especially as we do so many things,

02:05:09.020 | if you include all the scientific things we do,

02:05:10.780 | scientific fields we're working in,

02:05:12.580 | these are entire complex fields in themselves,

02:05:15.380 | and you have to sort of keep abreast of that.

02:05:18.980 | But I enjoy it.

02:05:20.020 | I've always been a sort of generalist in a way,

02:05:23.860 | and that's actually what happened

02:05:24.780 | with my games career after chess.

02:05:26.420 | One of the reasons I stopped playing chess

02:05:29.260 | was 'cause I got into computers,

02:05:30.340 | but also I started realizing there were many

02:05:32.100 | other great games out there to play too.

02:05:33.900 | So I've always been that way, inclined, multidisciplinary,

02:05:36.940 | and there's too many interesting things in the world

02:05:39.140 | to spend all your time just on one thing.

02:05:41.700 | - So you mentioned Spinoza,

02:05:43.260 | gotta ask the big, ridiculously big question about life.

02:05:47.660 | What do you think is the meaning of this whole thing?

02:05:50.500 | Why are we humans here?

02:05:52.580 | You've already mentioned that perhaps

02:05:54.580 | the universe created us.

02:05:56.740 | Is that why you think we're here?

02:05:58.940 | To understand how the universe--

02:06:00.100 | - Yeah, I think my answer to that would be,

02:06:02.100 | and at least the life I'm living,

02:06:03.980 | is to gain and understand knowledge,

02:06:08.100 | to gain knowledge and understand the universe.

02:06:10.620 | That's what I think,

02:06:12.260 | I can't see any higher purpose than that.

02:06:13.860 | If you think back to the classical Greeks,

02:06:15.700 | the virtue of gaining knowledge,

02:06:17.300 | I think it's one of the few true virtues

02:06:20.460 | is to understand the world around us

02:06:23.580 | and the context and humanity better.

02:06:25.460 | And I think if you do that,

02:06:27.820 | you become more compassionate

02:06:29.140 | and more understanding yourself and more tolerant

02:06:32.060 | and all these, I think all these other things

02:06:33.580 | may flow from that.

02:06:34.740 | And to me, understanding the nature of reality,

02:06:37.660 | that is the biggest question.

02:06:38.740 | What is going on here is sometimes the colloquial way I say,

02:06:41.380 | what is really going on here?

02:06:43.620 | It's so mysterious.

02:06:44.900 | I feel like we're in some huge puzzle.

02:06:46.820 | But the world is also seems to be,

02:06:49.980 | the universe seems to be structured in a way,

02:06:53.100 | why is it structured in a way that science is even possible?

02:06:55.860 | That methods, the scientific method works,

02:06:58.180 | things are repeatable.

02:06:59.300 | It feels like it's almost structured in a way

02:07:02.580 | to be conducive to gaining knowledge.

02:07:05.020 | So I feel like, and why should computers be even possible?

02:07:08.020 | Isn't that amazing that computational electronic devices

02:07:11.900 | can be possible?

02:07:14.100 | And they're made of sand,

02:07:15.300 | our most common element that we have,

02:07:17.460 | silicon on the Earth's crust,

02:07:19.940 | that could be made of diamond or something,

02:07:21.500 | then we would have only had one computer.

02:07:23.820 | So a lot of things are kind of slightly suspicious to me.

02:07:26.540 | - It sure as heck sounds,

02:07:27.740 | this puzzle sure as heck sounds like something

02:07:29.700 | we talked about earlier,

02:07:30.740 | what it takes to design a game that's really fun to play

02:07:35.120 | for prolonged periods of time.

02:07:36.620 | And it does seem like this puzzle, like you mentioned,

02:07:40.460 | the more you learn about it,

02:07:42.320 | the more you realize how little you know.

02:07:44.900 | So it humbles you,

02:07:46.060 | but excites you by the possibility of learning more.

02:07:49.060 | It's one heck of a puzzle we got going on here.

02:07:53.600 | So like I mentioned, of all the people in the world,

02:07:56.460 | you're very likely to be the one who creates the AGI system

02:08:01.460 | that achieves human level intelligence and goes beyond it.

02:08:06.360 | So if you got a chance,

02:08:07.660 | and very well you could be the person

02:08:09.460 | that goes into the room with the system

02:08:11.100 | and have a conversation,

02:08:13.180 | maybe you only get to ask one question.

02:08:15.300 | If you do, what question would you ask her?

02:08:18.160 | - I would probably ask,

02:08:21.500 | what is the true nature of reality?

02:08:23.640 | I think that's the question.

02:08:24.560 | I don't know if I'd understand the answer

02:08:26.000 | 'cause maybe it would be 42 or something like that.

02:08:28.520 | But that's the question I would ask.

02:08:30.980 | - And then there'll be a deep sigh from the systems,

02:08:34.800 | like, all right, how do I explain to this human?

02:08:37.440 | All right, let me, I don't have time to explain.

02:08:41.840 | Maybe I'll draw you a picture.

02:08:43.720 | It is, I mean, how do you even begin

02:08:46.480 | to answer that question?

02:08:49.840 | - Well, I think it would--

02:08:52.760 | - What would you think the answer could possibly look like?

02:08:55.720 | - I think it could start looking like

02:08:58.440 | more fundamental explanations of physics

02:09:02.100 | would be the beginning.

02:09:03.960 | More careful specification of that,

02:09:05.780 | taking you, walking us through by the hand

02:09:07.740 | as to what one would do to maybe prove those things out.

02:09:10.660 | - Maybe giving you glimpses of what things

02:09:13.740 | you totally missed in the physics of today.

02:09:15.680 | - Exactly, exactly.

02:09:16.760 | - Just here's glimpses of, no,

02:09:19.500 | there's a much more elaborate world

02:09:23.660 | or a much simpler world or something.

02:09:25.500 | - A much deeper, maybe simpler explanation of things,

02:09:30.260 | right, than the standard model of physics,

02:09:31.900 | which we know doesn't work, but we still keep adding to.

02:09:34.860 | So, and that's how I think the beginning

02:09:37.940 | of an explanation would look.

02:09:38.940 | And it would start encompassing many of the mysteries

02:09:41.260 | that we have wondered about for thousands of years,

02:09:43.380 | like consciousness, dreaming, life,

02:09:47.140 | and gravity, all of these things.

02:09:48.820 | - Yeah, giving us glimpses of explanations

02:09:51.100 | for those things, yeah.

02:09:52.620 | Well, Demis, you're one of the special human beings

02:09:57.180 | in this giant puzzle of ours,

02:09:59.060 | and it's a huge honor that you would take a pause

02:10:01.020 | from the bigger puzzle to solve this small puzzle

02:10:03.220 | of a conversation with me today.

02:10:04.740 | It's truly an honor and a pleasure.

02:10:06.260 | Thank you so much. - Thank you for having me.

02:10:07.100 | I really enjoyed it.

02:10:07.940 | Thanks, Lex.

02:10:09.100 | - Thanks for listening to this conversation

02:10:10.580 | with Demis Hassabis.

02:10:11.980 | To support this podcast,

02:10:13.180 | please check out our sponsors in the description.

02:10:15.820 | And now, let me leave you with some words

02:10:17.860 | from Edgar Dijkstra.

02:10:20.340 | "Computer science is no more about computers

02:10:23.460 | than astronomy is about telescopes."

02:10:26.340 | Thank you for listening, and hope to see you next time.

02:10:29.980 | (upbeat music)

02:10:32.580 | (upbeat music)

02:10:35.180 | [BLANK_AUDIO]

Demis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity | Lex Fridman Podcast #299

Chapters