The following is a conversation with Demis Hassabis, CEO and co-founder of DeepMind,
a company that has published and built some of the most incredible artificial intelligence
systems in the history of computing, including AlphaZero that learned all by itself to play
the game of go better than any human in the world and AlphaFold 2 that solved protein folding.
Both tasks consider nearly impossible for a very long time. Demis is widely considered to be one
of the most brilliant and impactful humans in the history of artificial intelligence and science
and engineering in general. This was truly an honor and a pleasure for me to finally sit down
with him for this conversation and I'm sure we will talk many times again in the future.
This is the Lux Friedman podcast. To support it, please check out our sponsors in the description
and now, dear friends, here's Demis Hassabis. Let's start with a bit of a personal question.
Am I an AI program you wrote to interview people until I get good enough to interview you?
Well, I'd be impressed if you were. I'd be impressed with myself if you were.
I don't think we're quite up to that yet, but maybe you're from the future, Lux.
If you did, would you tell me, is that a good thing to tell a language model that's tasked
with interviewing that it is, in fact, AI? Maybe we're in a meta-turing test.
Probably it would be a good idea not to tell you, so it doesn't change your behavior, right?
This is a kind of...
Eisenberg uncertainty principle situation. If I told you, you behaved differently. Maybe
that's what's happening with us, of course. This is a benchmark from the future where they
replay 2022 as a year before AIs were good enough yet, and now we want to see, is it going to pass?
Exactly.
If I was such a program, would you be able to tell, do you think?
So, to the Turing Test question, you've talked about the benchmark for solving
intelligence. What would be the impressive thing? You talked about winning a Nobel Prize
in AI system, winning a Nobel Prize, but I still returned to the Turing Test as a compelling test,
the spirit of the Turing Test as a compelling test.
Yeah, the Turing Test, of course, it's been unbelievably influential, and Turing's one
of my all-time heroes, but I think if you look back at the 1950 papers, original paper, and read
the original, you'll see I don't think he meant it to be a rigorous formal test. I think it was
more like a thought experiment, almost a bit of philosophy he was writing if you look at the style
of the paper. And you can see he didn't specify it very rigorously. So, for example, he didn't
specify the knowledge that the expert or judge would have, not how much time would they have to
investigate this. So, these are important parameters if you were going to make it a true
sort of formal test. And by some measures, people claim the Turing Test passed a decade ago.
I remember someone claiming that with a very bog standard, normal logic model, because they
pretended it was a kid. So, the judges thought that the machine was a child. So, that would be
very different from an expert AI person interrogating a machine and knowing how it was built and so
on. So, I think we should probably move away from that as a formal test and move more towards
a general test where we test the AI capabilities on a range of tasks and see if it reaches human
level or above performance on maybe thousands, perhaps even millions of tasks eventually,
and cover the entire sort of cognitive space. So, I think for its time it was an amazing
thought experiment and also 1950s, obviously, it was barely the dawn of the computer age.
So, of course, he only thought about text and now we have a lot more different inputs.
So, yeah, maybe the better thing to test is the generalizability so across multiple tasks,
but I think it's also possible as systems like God will show that eventually that might map
right back to language. So, you might be able to demonstrate your ability to generalize across
tasks by then communicating your ability to generalize across tasks, which is kind of what
we do through conversation anyway when we jump around. Ultimately, what's in there in that
conversation is not just you moving around knowledge, it's you moving around like these
entirely different modalities of understanding that ultimately map to your ability to
operate successfully in all of these domains, which you can think of as tasks.
Yeah, I think certainly we as humans use language as our main generalization communication tool. So,
I think we end up thinking in language and expressing our solutions in language. So,
it's going to be very powerful mode in which to explain the system to explain what it's doing,
but I don't think it's the only modality that matters. So, I think there's going to be a lot
of, you know, there's a lot of different ways to express capabilities other than just language.
Yeah, visual, robotics, body language,
yeah, actions, the interactive aspect of all that, that's all part of it.
But what's interesting with GATO is that it's sort of pushing prediction to the maximum in
terms of like, you know, mapping arbitrary sequences to other sequences and sort of just
predicting what's going to happen next. So, prediction seems to be fundamental to intelligence.
And what you're predicting doesn't so much matter.
Yeah, it seems like you can generalize that quite well. So, obviously, language models predict the
next word. GATO predicts potentially any action or any token. And it's just the beginning really.
It's our most general agent one could call it so far. But, you know, that itself can be scaled
up massively more than we've done so far. And obviously, we're in the middle of doing that.
But the big part of solving AGI is creating benchmarks that help us get closer and closer,
sort of creating benchmarks that test the generalizability. And it's just still interesting
that this fella, Alan Turing, was one of the first and probably still one of the only people that
was trying, maybe philosophically, but was trying to formulate a benchmark that could be followed.
It is, even though it's fuzzy, it's still sufficiently rigorous to where you can run that
test. And I still think something like the Turing test will, at the end of the day,
be the thing that truly impresses other humans so that you can have a close friend
who's an AI system. And for that friend to be a good friend, they're going to have to be able to
play StarCrafts. And they're going to have to do all of these tasks, get you a beer,
so the robotics tasks, play games with you, use language, humor, all of those kinds of things.
But that ultimately can blow down to language. It feels like not in terms of the AI community,
but in terms of the actual impact of general intelligence on the world. It feels like language
will be the place where it truly shines. I think so, because it's such an important kind of input
output for us. I think you're right. I think the Turing test, the philosophy behind it,
which is the idea of can a machine mimic the behaviors of a human? And I would say,
wider than just language and text, then in terms of actions and everything else,
creativity, all these things, then if it can match or exceed human cognitive capabilities,
then I think we have a true intelligence. So from that perspective, you're right. I think he did
formulate the right kind of setup. I think there'll be a kind of humor
in the AI systems of the future looking back to this conversation and thinking about the Turing
test and also thinking about by that time, they would know which year they were finally able to
sort of cross the threshold of human level intelligence and think how funny it is that
we humans were still confused about this whole problem that's been long ago been solved.
Well, hello to the AI agents of the future. Anyway, so going back to your journey,
when did you fall in love with programming first? Well, it was pretty young age, actually. So
I started off, actually games was my first love. So starting to play chess when I was around four
years old. And then it's actually with winnings from a chess competition that I managed to buy
my first chess computer when I was about eight years old was a ZX Spectrum, which was hugely
popular in the UK at the time. And it was an amazing machine because I think it trained a
whole generation of programmers in the UK because it was so accessible. You literally switched it
on and there was the basic prompt and you could just get going. And my parents didn't really know
anything about computers. But because it was my money from a chess competition, I could say I
wanted to buy it. And then I just went to bookstores, got books on programming and started typing in
the programming code. And then of course, once you start doing that, you start adjusting it and
then making your own games. And that's when I fell in love with computers and realized that
they were a very magical device. In a way, I don't want to have been able to explain this
at the time, but I felt that there was almost a magical extension of your mind. I always had this
feeling and I've always loved this about computers that you can set them off doing something,
some task for you, you can go to sleep, come back the next day and it's solved.
That feels magical to me. All machines do that to some extent. They all enhance our
natural capabilities. Obviously, cars make us allow us to move faster than we can run.
But this was a machine to extend the mind. And then of course, AI is the ultimate expression
of what a machine may be able to do or learn. So very naturally for me, that thought extended
into AI quite quickly. Do you remember the programming language that was first started
in? Was it special to the machine? No, it was just basic. I think it was just basic
on the ZX Spectrum. I don't know what specific form it was. And then later on, I got a Commodore
Amiga, which was a fantastic machine. Now you're just showing off.
So yeah, well, lots of my friends had Atari STs and I managed to get Amiga's. It was a bit
more powerful and that was incredible. I used to do programming in Assembler and also AMOS
Basic, this specific form of basic. It was incredible actually. So all my coding skills.
And when did you fall in love with AI? So when did you first start to gain an understanding
that you can not just write programs that do some mathematical operations for you while you sleep,
but something that's akin to bringing an entity to life? Sort of a thing that can figure out
something more complicated than a simple mathematical operation?
Yeah. So there was a few stages for me all while I was very young. So first of all,
as I was trying to improve at playing chess, I was captaining various England junior chess teams
and at the time when I was about maybe 10, 11 years old, I was going to become a professional
chess player. That was my first thought. So that dream was there to try to get to the highest
level of chess? Yeah. So when I was about 12 years old, I got to master standard and I was
second highest rated player in the world to Judith Polger, who obviously ended up being
an amazing chess player and a world women's champion. And when I was trying to improve at
chess, what you do is you obviously, first of all, you're trying to improve your own thinking
processes. So that leads you to thinking about thinking. How is your brain coming up with these
ideas? Why is it making mistakes? How can you improve that thought process? But the second
thing is that it was just the beginning. This was like in the early 80s, mid 80s of chess
computers. If you remember, they were physical balls like the one we have in front of us and
you pressed down the squares. And I think Kasparov had a branded version of it that I got.
And you were used to, they're not as strong as they are today, but they were pretty strong
and used to practice against them to try and improve your openings and other things.
And so I remember, I think I probably got my first one, I was around 11 or 12. And I remember
thinking, this is amazing, how someone programmed this chess board to play chess. And it was very
formative book I bought, which was called The Chess Computer Handbook by David Levy.
It came out in 1984 or something. So I must have got it when I was about 11, 12. And it
explained fully how these chess programs were made. And I remember my first AI program being
programming my Amiga, it wasn't powerful enough to play chess. I couldn't write a whole chess
program, but I wrote a program for it to play Othello, or Reverse. It's sometimes called, I
think, in the US. And so a slightly simpler game than chess. But I used all of the principles
that chess programs had, alpha, beta, search, all of that. And that was my first AI program.
I remember that very well. I was around 12 years old. So that brought me into AI.
And then the second part was later on, around 16, 17, and I was writing games professionally,
designing games, writing a game called Theme Park, which had AI as a core gameplay component
as part of the simulation. And it sold millions of copies around the world. And people loved
the way that the AI, even though it was relatively simple by today's AI standards,
was reacting to the way you as the player played it. So it was called a sandbox game. So it was
one of the first types of games like that, along with SimCity. And it meant that every game you
played was unique. Is there something you could say, just on a small tangent, about really impressive
AI from a game design, human enjoyment perspective, really impressive AI that you've seen in games?
And maybe what does it take to create AI system? And how hard of a problem is that?
So a million questions, just as a brief tangent.
Well, look, I think games have been significant in my life for three reasons. So first of all,
I was playing them and training myself on games when I was a kid.
Then I went through a phase of designing games and writing AI for games. So all the games I
professionally wrote had AI as a core component. And that was mostly in the 90s. And the reason
I was doing that in games industry was at the time, the games industry I think was the cutting
edge of technology. So whether it was graphics with people like John Carmack and Quake and those
kind of things or AI, I think actually all the action was going on in games. And we're still
reaping the benefits of that, even with things like GPUs, which I find ironic was obviously
invented for graphics, computer graphics, but then turns out to be amazingly useful for AI.
Just turns out everything's a matrix multiplication. It appears in the whole world.
So I think games at the time had the most cutting edge AI. And a lot of the games I was involved
in writing. So there was a game called Black and White, which was one game I was involved with
in the early stages of, which I still think is the most impressive example of reinforcement
learning in a computer game. So in that game, you trained a little pet animal and it's sort of
learned from how you were treating it. So if you treated it badly, then it became mean. And then
it would be mean to your villagers and your population, the sort of the little tribe that
you were running. But if you were kind to it, then it would be kind. And people fascinated by how
that works. And so as I had to be honest with the way it kind of developed. Especially the mapping
to good and evil, it made you realize, made me realize that you can sort of in the way, in the
choices you make can define where you end up. And that means all of us are capable of the good
evil. It all matters in the different choices along the trajectory to those places that you make.
It's fascinating. I mean, games can do that philosophically to you. And it's rare. It seems
rare. Yeah. Well, games are I think a unique medium because you as the player, you're not just
passively consuming the entertainment, right? You're actually actively involved as an agent.
So I think that's what makes it in some ways can be more visceral than other mediums like films
and books. So the second, so that was designing AI and games. And then the third use we've used
of AI is indeed mind from the beginning, which is using games as a testing ground for proving
out AI algorithms and developing AI algorithms. And that was a sort of a core component of our
vision at the start of DeepMind was that we would use games very heavily as our main testing ground,
certainly to begin with, because it's super efficient to use games. And also, it's very easy
to have metrics to see how well your systems are improving and what direction your ideas are going
in and whether you're making incremental improvements. And because those games are often
rooted in something that humans did for a long time beforehand, there's already a strong set of
rules like it's already a damn good benchmark. Yes, it's really good for so many reasons because
you've got clear measures of how good humans can be at these things. And in some cases like Go,
we've been playing it for thousands of years. And often they have scores or at least win conditions.
So it's very easy for reward learning systems to get a reward. It's very easy to specify what
that reward is. And also at the end, it's easy to test externally how strong is your system
by, of course, playing against the world's strongest players at those games. So it's so good
for so many reasons. And it's also very efficient to run potentially millions of simulations
in parallel on the cloud. So I think there's a huge reason why we were so successful back in
the starting out 2010, how come we were able to progress so quickly because we'd utilize games.
And at the beginning of DeepMind, we also hired some amazing game engineers who I knew from my
previous lives in the games industry. And that helped to bootstrap us very quickly.
And plus it's somehow super compelling, almost at a philosophical level of man versus machine
over over a chessboard or a Go board. And especially given that the entire history of AI
is defined by people saying it's going to be impossible to make a machine that beats a human
being in chess. And then once that happened, people were certain when I was coming up in AI,
that Go is not a game that can be solved because of the combinatorial complexity. It's just too,
it's, you know, no matter how much Moore's law you have, compute is just never going to be able
to crack the game of Go. And so then there's something compelling about facing sort of
taking on the impossibility of that task from the AI researcher perspective,
engineer perspective. And then as a human being just observing this whole thing,
your beliefs about what you thought was impossible being broken apart. It's humbling
to realize we're not as smart as we thought. It's humbling to realize that the things we
think are impossible now perhaps will be done in the future. There's something really powerful
about a game, AI system being a human being in a game that drives that message home for like
millions, billions of people, especially in the case of Go. Sure. Well, look, I think it's,
I mean, it has been a fascinating journey. And especially as I think about it from, I can
understand it from both sides, both as the AI creators of the AI, but also as a games player
originally. So it was a really interesting, I mean, it was a fantastic, but also somewhat
bittersweet moment, the AlphaGo match for me seeing that and being obviously heavily
involved in that. But as you say, Chess has been the, I mean, Kasparov, I think rightly called it
the Drosophila of intelligence, right? So it's sort of, I love that phrase. And I think he's
right because Chess has been hand in hand with AI from the beginning of the whole field, right?
So I think every AI practitioner starting with Turing and Claude Shannon and all those,
the sort of forefathers of the field, tried their hand at writing a chess program. I've got
an original edition of Claude Shannon's first chess program. I think it was 1949, the original
sort of paper. And they all did that and Turing famously wrote a chess program that all the
computers around them were obviously too slow to run it. So he had to run, he had to be the
computer, right? So he literally, I think spent two or three days running his own program by hand
with pencil and paper and playing a friend of his with his chess program. So of course,
Deep Blue was a huge moment beating Kasparov. But actually, when that happened, I remember that very,
very vividly, of course, because it was Chess and computers and AI, all the things I loved. And I
was at college at the time. But I remember coming away from that, being more impressed by Kasparov's
mind than I was by Deep Blue. Because here was Kasparov with his human mind, not only could he
play chess more or less to the same level as this brute of a calculation machine. But of course,
Kasparov can do everything else humans can do, ride a bike, talk many languages, do politics,
all the rest of the amazing things that Kasparov does. And so with the same brain. And yet Deep
Blue, brilliant as it was at chess, it'd been hand coded for chess and actually had distilled
the knowledge of chess grandmasters into a cool program. But it couldn't do anything else. It
couldn't even play a strictly simpler game like tic-tac-toe. So something to me was missing from
intelligence from that system that we would regard as intelligence. And I think it was this idea of
generality and also learning. So that's what we tried to do with AlphaGo.
Yeah, with AlphaGo and AlphaZero, MuZero, and then Godo, all the things that we'll get into some
parts of there's just a fascinating trajectory here. But let's just stick on chess briefly,
on the human side of chess. You've proposed that from a game design perspective, the thing
that makes chess compelling as a game is that there's a creative tension between a bishop
and the knight. Can you explain this? First of all, it's really interesting to think about what
makes a game compelling. It makes it stick across centuries.
Yeah, I was sort of thinking about this. And actually a lot of even amazing chess players
don't think about it necessarily from a game's designer point of view. So it's with my game
design hat on that I was thinking about this. Why is chess so compelling?
And I think a critical reason is the dynamicness of the different kind of chess positions you can
have, whether they're closed or open and other things, comes from the bishop and the knight.
So if you think about how different the capabilities of the bishop and knight are in
terms of the way they move, and then somehow chess has evolved to balance those two capabilities
more or less equally. So they're both roughly worth three points each.
So you think that dynamics is always there, and then the rest of the rules are kind of trying
to stabilize the game? Well, maybe. I mean, it's sort of, I don't know if chicken and egg situation
probably both came together. But the fact that it's got to this beautiful equilibrium
where you can have the bishop and knight, they're so different in power, but so equal in value
across the set of the universe of all positions, right? Somehow they've been balanced by humanity
over hundreds of years. I think gives the game the creative tension that you can swap the bishop
and knights for a bishop for a knight, and they're more or less the worth the same. But now you aim
for a different type of position. If you have the knight, you want a closed position. If you have the
bishop, you want an open position. So I think that creates a lot of the creative tension in chess.
So some kind of controlled creative tension. From an AI perspective, do you think AI systems
convention design games that are optimally compelling to humans?
Well, that's an interesting question. Sometimes I get asked about AI and creativity, and the
way I answer that is relevant to that question, which is that I think they're different levels
of creativity, one could say. So I think if we define creativity as coming up with something
original that's useful for a purpose, then I think the kind of lowest level of creativity
is like an interpolation, so an averaging of all the examples you see. So maybe a very basic
AI system could say you could have that. So you show it millions of pictures of cats,
and then you say, give me an average looking cat, generate me an average looking cat.
I would call that interpolation. Then there's extrapolation, which something like AlphaGo
showed. So AlphaGo played millions of games of Go against itself, and then it came up with brilliant
new ideas like Move 37 in game two, bringing a motif, strategies in Go that no humans had
ever thought of, even though we've played it for thousands of years and professionally for hundreds
of years. So that I call that extrapolation. But then there's still a level above that,
which is you could call out of the box thinking or true innovation, which is could you invent Go?
Right? Could you invent chess and not just come up with a brilliant chess move or brilliant Go move,
but can you actually invent chess or something as good as chess or Go? And I think one day
AI could, but what's missing is how would you even specify that task to a program right now?
And the way I would do it if I was telling a human to do it or a game designer, a human game
designer to do it, is I would say something like Go, I would say, come up with a game that only
takes five minutes to learn, which Go does because it's got simple rules, but many lifetimes to master,
right, or impossible to master in one lifetime because it's so deep and so complex.
And then it's aesthetically beautiful. And also it can be completed in three or four hours of
gameplay time, which is useful for us in a human day. And so you might specify these
sort of high level concepts like that. And then with that, and maybe a few other things,
one could imagine that Go satisfies those constraints. But the problem is, is that we're
not able to specify abstract notions like that, high level abstract notions like that yet to our
AI systems. And I think there's still something missing there in terms of high level concepts or
abstractions that they truly understand and they're combinable and compositional. So for the moment,
I think AI is capable of doing interpolation and extrapolation, but not true invention.
So coming up with rule sets and optimizing with complicated objectives around those rule sets
we can't currently do, but you could take a specific rule set and then run a kind of self-play
experiment to see how long, just observe how an AI system from scratch learns. How long is that
journey of learning? And maybe if it satisfies some of those other things you mentioned in terms of
quickness to learn and so on, and you could see a long journey to master for even an AI system,
then you could say that this is a promising game. But it would be nice to do almost like
alpha codes or programming rules. So generating rules that automate even that part of the generation
of rules. So I have thought about systems actually that I think would be amazing in for a games
designer. If you could have a system that takes your game, plays it tens of millions of times,
maybe overnight, and then self-balances the rules better. So it tweaks the rules and maybe the
equations and the parameters so that the game is more balanced, the units in the game, or
some of the rules could be tweaked. So it's a bit of like giving a base set and then allowing
Monte Carlo Tree Search or something like that to sort of explore it. And I think that would be
super powerful tool actually for balancing, auto-balancing a game, which usually takes
thousands of hours from hundreds of human games testers normally to balance some game
like StarCraft, which is, you know, Blizzard are amazing at balancing their games, but it
takes them years and years and years. So one could imagine at some point when this stuff becomes
efficient enough to, you know, you might better do that like overnight.
Do you think a game that is optimal designed by an AI system would look very much like
Planet Earth? Maybe, maybe it's only the sort of game I would love to make is, and I've tried,
you know, in my game's career, the game's design career, you know, my first big game was designing
a theme park, an amusement park. Then with games like Republic, I tried to, you know,
have games where we designed whole cities and allowed you to play in. So, and of course,
people like Will Wright have written games like SimEarth trying to simulate the whole of Earth.
Pretty tricky, but I think- SimEarth, I haven't actually played that one. So what is it?
Does it incorporate evolution? Yeah, it has evolution and it sort of tries to,
it sort of treats it as an entire biosphere, but from quite high level.
So- It'd be nice to be able to sort of zoom in, zoom out and zoom in.
Exactly, exactly. So obviously it couldn't do, that was in the night, I think he wrote that in
the 90s, so it couldn't, you know, it wasn't able to do that. But that would be, obviously,
the ultimate sandbox game, of course. On that topic, do you think we're living in a simulation?
Yes, well, so, okay, so- We're gonna jump around from the absurdly philosophical to the-
Sure, sure, very, very happy to. So I think my answer to that question is a little bit complex,
because there is simulation theory, which obviously Nick Bostrom, I think, famously first proposed.
And I don't quite believe it in that sense. So in the sense that are we in some sort of computer
game, or have our descendants somehow recreated Earth in the 21st century, and for some kind
of experimental reason, I think that- But I do think that we might be- That the best way to
understand physics and the universe is from a computational perspective. So understanding it
as an information universe, and actually information being the most fundamental unit of reality,
rather than matter or energy. So physicists would say, you know, matter or energy, you know,
E equals MC squared, these are the things that are the fundamentals of the universe. I'd actually
say information, which of course itself can be, can specify energy or matter, right? Matter is
actually just, you know, we're just out the way our bodies and all the molecules in our body
arrange is information. So I think information may be the most fundamental way to describe
the universe. And therefore, you could say we're in some sort of simulation because of that.
But I don't- I'm not really a subscriber to the idea that, you know, these are sort of throw away
billions of simulations around. I think this is actually very critical and possibly unique,
this simulation. This particular one. Yes. But and you just mean treating the universe as a computer
that's processing and modifying information is a good way to solve the problems of physics,
of chemistry, of biology, and perhaps of humanity and so on. Yes. I think understanding physics
in terms of information theory might be the best way to really understand what's going on here.
From our understanding of a universal term machine, from our understanding of a computer,
do you think there's something outside of the capabilities of a computer that is present
in our universe? You have a disagreement with Roger Penrose about the nature of consciousness.
He thinks that consciousness is more than just a computation. Do you think all of it,
the whole shebangs, can be a computation? Yeah, I've had many fascinating debates with
Sir Roger Penrose. And obviously, he's famously, and I read, you know, Emperor's New Mind and
his books, his classical books, and they were pretty influential in the, you know, in the 90s.
And he believes that there's something more, you know, something quantum
that is needed to explain consciousness in the brain. I think about what we're doing,
actually, at DeepMind, and what my career is being, we're almost like Turing's champion.
So we are pushing Turing machines or classical computation to the limits. What are the limits
of what classical computing can do? Now, and at the same time, I've also studied neuroscience
to see, and that's why I did my PhD in, was to see also to look at, you know, is there anything
quantum in the brain from a neuroscience or biological perspective? And so far, I think most
neuroscientists and most mainstream biologists and neuroscientists would say there's no evidence of any
quantum systems or effects in the brain. As far as we can see, it can be mostly explained by classical
theories. And then so there's sort of the search from the biology side. And then at the same time,
there's the raising of the water at the bar from what classical Turing machines can do,
and, you know, including our new AI systems. And as you alluded to earlier, you know, I think AI,
especially in the last decade plus, has been a continual story now of surprising events
and surprising successes, knocking over one theory after another, what was thought to be
impossible, you know, from go to protein folding and so on. And so I think I would be very hesitant
to bet against how far the universal Turing machine and classical computation paradigm
can go. And my betting would be that all of certainly what's going on in our brain can probably
be mimicked or approximated on a classical machine, not, you know, not requiring something metaphysical
or quantum. And we'll get there with some of the work with AlphaFold, which I think begins the
journey of modeling this beautiful and complex world of biology. So you think all the magic of
the human mind comes from this, just a few pounds of mush, a biological computational mush that's
akin to some of the neural networks, not directly, but in spirit that deep mind has been working with.
Well, look, I think it's, you say it's a few, you know, of course, this is the, I think,
the biggest miracle of the universe is that it's just a few pounds of mush in our skulls. And yet
it's also our brains are the most complex objects that we know of in the universe. So there's
something profoundly beautiful and amazing about our brains. And I think that it's an incredibly,
incredible efficient machine. And it's, you know, a phenomenon, basically. And I think that building
AI, one of the reasons I want to build AI, and I've always wanted to is, I think by building
an intelligent artifact like AI, and then comparing it to the human mind, that will help us unlock
the uniqueness and the true secrets of the mind that we've always wondered about since the dawn
of history, like consciousness, dreaming, creativity, emotions, what are all these things,
right? We've wondered about them since the dawn of humanity. And I think one of the reasons and,
you know, I love philosophy and philosophy of mind is we found it difficult is there haven't
been the tools for us to really other than introspection to, from very clever people in history,
very clever philosophers to really investigate this scientifically. But now, suddenly, we have a
plethora of tools. Firstly, we have all the neuroscience tools, fMRI machines, single cell
recording, all of this stuff. But we also have the ability computers and AI to build intelligent
systems. So I think that, you know, I think it is amazing what the human mind does. And
I'm kind of in awe of it really. And I think it's amazing that without human minds, we're able to
build things like computers and actually even, you know, think and investigate about these
questions. I think that's also a testament to the human mind. Yeah. The universe built
the human mind that now is building computers that help us understand both the universe and our
own human mind. That's right. It's exactly it. I mean, I think that's one, you know, one could say
we are maybe we're the mechanism by which the universe is going to try and understand itself.
Yeah. It's beautiful. So let's let's go to the basic building blocks of biology that I think
is another angle at which you can start to understand the human mind, the human body,
which is quite fascinating, which is from the basic building blocks, start to simulate, start to model
how from those building blocks, you can construct bigger and bigger, more complex systems, maybe
one day, the entirety of the human biology. So here's another problem that thought to be
impossible to solve, which is protein folding and Alpha Fold or a specific Alpha Fold two
did just that. It's solved protein folding. I think it's one of the biggest breakthroughs,
certainly in the history of structural biology, but in general, in science,
maybe from a high level, what is it and how does it work? And then we can ask some fascinating
questions after. Sure. So maybe to explain it to people not familiar with protein folding is,
you know, first of all, explain proteins, which is, you know, proteins are essential to all life.
Every function in your body depends on proteins. Sometimes they're called the workhorses of
biology. And if you look into them, and I've, you know, obviously as part of Alpha Fold,
I've been researching proteins and structural biology for the last few years, you know, they're
amazing little bio nanomachines proteins. They're incredible if you actually watch little videos
of how they work, animations of how they work. And proteins are specified by their genetic sequence
called the amino acid sequence. So you can think of it as their genetic makeup. And then in the body,
in nature, they fold up into a 3D structure. So you can think of it as a string of beads,
and then they fold up into a ball. Now, the key thing is you want to know what that 3D structure
is, because the structure, the 3D structure of a protein is what helps to determine what does it
do, the function it does in your body. And also, if you're interested in drug drugs or disease,
you need to understand that 3D structure, because if you want to target something with a drug compound
about to block something the protein's doing, you need to understand where it's going to bind
on the surface of the protein. So obviously, in order to do that, you need to understand the 3D
structure. So the structure is mapped to the function? The structure is mapped to the function.
And the structure is obviously somehow specified by the amino acid sequence. And that's the, in
essence, the protein folding problem is, can you just from the amino acid sequence, the one-dimensional
string of letters, can you immediately computationally predict the 3D structure?
Right. And this has been a grand challenge in biology for over 50 years. So I think it was
first articulated by Christian Anfinsen, a Nobel Prize winner in 1972, as part of his Nobel Prize
winning lecture. And he just speculated, this should be possible to go from the amino acid
sequence to the 3D structure. But he didn't say how. So it's been described to me as equivalent
to Fermat's last theorem, but for biology. You should, as somebody that very well might win
the Nobel Prize in the future, but outside of that, you should do more of that kind of thing.
In the margins, just put random things. That will take like 200 years to solve.
Set people off for 200 years. It should be possible.
Exactly. And just don't give any details.
Exactly. I think everyone's, exactly. It should be, I'll have to remember that for future.
So yeah. So he set off, you know, with this one throwaway remark, just like Fermat, you know,
he set off this whole 50-year field, really, of computational biology. And they had, you know,
they got stuck. They hadn't really got very far with doing this. And until now, until alpha
fold came along, this is done experimentally, very painstakingly. So the rule of thumb is,
and you have to like crystallize the protein, which is really difficult. Some proteins can't
be crystallized like membrane proteins. And then you have to use very expensive
electron microscopes or x-ray crystallography machines, really painstaking work to get the
3D structure and visualize the 3D structure. So the rule of thumb in experimental biology is that
it takes one PhD student, their entire PhD, to do one protein. And with alpha fold 2, we were able
to predict the 3D structure in a matter of seconds. And so we were, you know, over Christmas,
we did the whole human proteome. Or every protein in the human body will 20,000 proteins.
So the human proteome is like the equivalent of the human genome, but on protein space.
And sort of revolutionize, really, what structural biologists can do. Because now,
they don't have to worry about these painstaking experimentals. You know,
should they put all of that effort in or not, they can almost just look up the structure
of their proteins like a Google search. And so there's a data set on which it's
trained and how to map this amino acid sequence. First of all, it's incredible that approaching
this little chemical computer is able to do that computation itself in some kind of distributed way
and do it very quickly. That's a weird thing. And they evolved that way because, you know,
in the beginning, I mean, that's a great invention, just the protein itself.
And then there's, I think, probably a history of like, they evolved to have many of these proteins
and those proteins figure out how to be computers themselves in such a way that you can create
structures that can interact in complexes with each other in order to form high-level functions.
I mean, it's a weird system that they've figured it out.
Well, for sure. I mean, we, you know, maybe we should talk about the origins of life, too,
but proteins themselves, I think, are magical and incredible, as I said, little bio-nanomachines.
And actually, Leventhal, who is another scientist, a contemporary of Anfinson,
he coined this Leventhal, what became known as Leventhal's paradox, which is exactly what
you're saying. He calculated roughly an average protein, which is maybe 2,000 amino acids base
as long, can fold in maybe 10 to the power 300 different conformations. So there's 10 to the
power 300 different ways that protein could fold up. And yet somehow, in nature, physics solves this,
solves this in a matter of milliseconds. So proteins fold up in your body in, you know,
sometimes in fractions of a second. So physics is somehow solving that search problem.
And just to be clear, in many of these cases, maybe you can correct me if I'm wrong,
there's often a unique way for that sequence to form itself. So among that huge number of
possibilities, it figures out a way how to stably, in some cases, there might be a misfunction,
so on, which leads to a lot of the disorders and stuff like that. But most of the time it's a unique
mapping. And that unique mapping is not obvious. No, exactly. Which is what the problem is.
No, exactly. So there's a unique mapping usually in a healthy, if it's healthy. And as you say, in
disease, so for example, Alzheimer's, one conjecture is that it's because of a misfolder protein,
a protein that folds in the wrong way, amyloid beta protein. And then because it folds in the
wrong way, it gets tangled up in your neurons. So it's super important to understand both
healthy functioning and also disease is to understand what these things are doing and
how they're structuring. Of course, the next step is sometimes proteins change shape when
they interact with something. So they're not just static necessarily in biology.
Maybe you can give some interesting, sort of beautiful things to you about these early days
of AlphaFold, of solving this problem. Because unlike games, this is real physical systems that
are less amenable to self-play type of mechanisms. The size of the data set is smaller that you
might otherwise like. So you have to be very clever about certain things. Is there something you
could speak to, what was very hard to solve? And what are some beautiful aspects about the
solution? Yeah, I would say AlphaFold is the most complex and also probably most meaningful
system we've built so far. So it's been an amazing time actually in the last two, three years to
see that come through. Because as we talked about earlier, games is what we started on,
building things like AlphaGo and AlphaZero. But really, the ultimate goal was to not just to crack
games, it was just to use them to bootstrap general learning systems, we could then apply to real
world challenges. Specifically, my passion is scientific challenges like protein folding.
And then AlphaFold, of course, is our first big proof point of that. And so in terms of the data
and the amount of innovations that had to go into it, it was like more than 30 different
component algorithms needed to be put together to crack the protein folding. I think some of the
big innovations were that kind of building in some hard-coded constraints around physics
and evolutionary biology to constrain sort of things like the bond angles in the protein and
things like that, but not to impact the learning system. So still allowing the system to be able
to learn the physics itself from the examples that we had. And the examples, as you say,
there are only about 150,000 proteins, even after 40 years of experimental biology,
only around 150,000 proteins have been, the structures have been found out about. So that was
our training set, which is much less than normally we would like to use. But using various tricks,
things like self-distillation, so actually using AlphaFold predictions, some of the best predictions
that it thought was highly confident in, we put them back into the training set to make the training
set bigger. That was critical to AlphaFold working. So there was actually a huge number of
different innovations like that that were required to ultimately crack the problem.
AlphaFold 1, what it produced was a histogram, so a kind of a matrix of the pairwise distances
between all of the molecules in the protein. And then there had to be a separate optimization
process to create the 3D structure. And what we did for AlphaFold 2 is make it truly end-to-end.
So we went straight from the amino acid sequence of bases to the 3D structure directly without
going through this intermediate step. And in machine learning, what we've always found is that
the more end-to-end you can make it, the better the system. And it's probably because
in the end, the system is better at learning what the constraints are than we are as the
human designers of specifying it. So anytime you can let it flow end-to-end and actually just
generate what it is you're really looking for, in this case, the 3D structure, you're better off
than having this intermediate step, which you then have to handcraft the next step for. So it's
better to let the gradients and the learning flow all the way through the system from the end
point, the end output you want to the inputs. So that's a good way to start on a new problem.
Handcraft a bunch of stuff, add a bunch of manual constraints with a small end-to-end learning piece
or a small learning piece and grow that learning piece until it consumes the whole thing.
That's right. And so you can also see, you know, this is a bit of a method we've developed over
doing many sort of successful alphabets, we call them AlphaX projects, right? And the easiest
way to see that is the evolution of AlphaGo to AlphaZero. So AlphaGo was a learning system,
but it was specifically trained to only play Go, right? So and what we wanted to do with
the first version of AlphaGo is just get to world champion performance, no matter how we did it,
right? And then, of course, AlphaGo Zero, we removed the need to use human games as a starting
point, right? So it could just play against itself from random starting point from the beginning.
So that removed the need for human knowledge about Go. And then finally, AlphaZero then
generalized it so that any things we had in there, the system, including things like symmetry of the
Go board, were removed. So that AlphaZero could play from scratch any two player game and then
MuZero, which is the final, our latest version of that set of things, was then extending it so
that you didn't even have to give it the rules of the game. It would learn that for itself. So
it could also deal with computer games as well as board games. So that line of AlphaGo AlphaGo
Zero AlphaZero, MuZero, that's the full trajectory of what you can take from
imitation learning to full self supervised learning. Yeah, exactly. And learning the
entire structure of the environment you put in from scratch, right? And bootstrapping it
through self play yourself. But the thing is, it would have been impossible, I think, or very
hard for us to build AlphaZero or MuZero first out of the box. Even psychologically, because you
have to believe in yourself for a very long time, you're constantly dealing with doubt,
because a lot of people say that it's impossible. Exactly. So it was hard enough just to do Go,
as you were saying, everyone thought that was impossible, or at least a decade away from when
we did it back in 2015, 2016. And so yes, it would have been psychologically probably very
difficult, as well as the fact that, of course, we learn a lot by building AlphaGo first.
Right. So I think this is why I call AI an engineering science. It's one of the most
fascinating science disciplines. But it's also an engineering science in the sense that,
unlike natural sciences, the phenomenon you're studying doesn't exist out in nature. You have
to build it first. So you have to build the artifact first, and then you can study and
pull it apart on how it works. This is tough to ask you this question, because you probably
will say it's everything. But let's try to think through this, because you're in a very
interesting position where DeepMind is a place of some of the most brilliant ideas in the history
of AI, but it's also a place of brilliant engineering. So how much of solving intelligence,
this big goal for DeepMind? How much of it is science? How much is engineering? So how much
is the algorithms? How much is the data? How much is the hardware compute infrastructure?
How much is the software compute infrastructure? What else is there? How much is the human
infrastructure? And just the humans interacting certain kinds of ways in the space of all those
ideas? And how much is maybe like philosophy? What's the key? If you were to sort of look back,
like if we go forward 200 years and look back, what was the key thing that solved intelligence?
Is it the ideas or the engineering? I think it's a combination. First of all,
of course, it's a combination of all those things, but the ratios of them changed over time.
So even in the last 12 years, we started DeepMind in 2010, which is hard to imagine now,
because 2010, it's only 12 short years ago, but nobody was talking about AI. I don't even remember
back to your MIT days. No one was talking about it. I did a postdoc at MIT back around then,
and it was sort of thought of as, well, look, we know AI doesn't work. We tried this hard in
the 90s at places like MIT, mostly using logic systems and old fashioned sort of good old fashioned
AI, we would call it now. People like Minsky and Patrick Winston, and you know all these characters,
right? And used to debate a few of them, and they used to think I was mad thinking about that
some new advance could be done with learning systems. I was actually pleased to hear that,
because at least you know you're on a unique track at that point, right? Even if all of your
professors are telling you you're mad. And of course, in industry, we couldn't get,
you know, it was difficult to get two cents together, which is hard to imagine now as well,
given that it's the biggest sort of buzzword in VCs and fundraising is easy and all these
kind of things today. So back in 2010, it was very difficult. And the reason we started then,
and Shane and I used to discuss what were the sort of founding tenants of DeepMind.
And it was various things. One was algorithmic advances. So deep learning, you know, Jeff Hinton
and Ko had just sort of invented that in academia, but no one in industry knew about it. We love
reinforcement learning. We thought that could be scaled up. But also understanding about the human
brain had advanced quite a lot in the decade prior with fMRI machines and other things. So we
could get some good hints about architectures and algorithms and sort of representations maybe
that the brain uses. So at a systems level, not at an implementation level. And then the other
big things were compute and GPUs, right? So we could see a compute was going to be really useful
and it got to a place where it become commoditized mostly through the games industry. And that could
be taken advantage of. And then the final thing was also mathematical and theoretical definitions
of intelligence. So things like AIXI, AIXE, which Shane worked on with his supervisor Marcus Hutto,
which is this sort of theoretical proof really of universal intelligence, which is actually a
reinforcement learning system in the limit. I mean, it assumes infinite compute and infinite memory
in the way, you know, like a Turing machine proves. But I was also waiting to see something like that
too. You know, like Turing machines and computation theory that people like Turing and Shannon came
up with underpins modern computer science. You know, I was waiting for a theory like that to
sort of underpin a GI research. So when I, you know, met Shane and saw he was working on something
like that, you know, that to me was a sort of final piece of the jigsaw. So in the early days,
I would say that ideas were the most important. You know, for us, it was deep reinforcement
learning, scaling up deep learning. Of course, we've seen transformers. So huge leaps, I would
say, three or four from, if you think from 2010 till now, huge evolutions, things like AlphaGo.
And maybe there's a few more still needed. But as we get closer to AI, AGI,
I think engineering becomes more and more important and data. Because scale and of course,
the recent, you know, results of GPT-3 and all the big language models and large models,
including our ones, has shown that scale is and large models are clearly going to be necessary,
but perhaps not sufficient part of an AGI solution. And throughout that, like you said,
and I'd like to give you a big thank you, you're one of the pioneers in this is sticking by ideas
like reinforcement learning, that this can actually work, given actually limited success in the past.
And also, which we still don't know, but proudly, having the best researchers in the world and
talking about solving intelligence. So talking about whatever you call it, AGI or something like
this, that speaking of MIT, that's just something you wouldn't bring up. Not maybe you did in
like 40, 50 years ago. But that was AI was a place where you do tinkering, very small scale,
not very ambitious projects. And maybe the biggest ambitious projects were in the space of robotics
and doing like the DARPA challenge. But the task of solving intelligence and believing you can,
that's really, really powerful. So in order for engineering to do its work, to have great engineers
build great systems, you have to have that belief, that threads throughout the whole thing,
that you can actually solve some of these impossible challenges.
Yeah, that's right. And back in 2010, our mission statement, and still is today,
is it used to be solving step one, solve intelligence, step two, use it to solve everything
else. So if you can imagine pitching that to a VC in 2010, the kind of looks we got,
we managed to find a few kooky people to back us, but it was tricky. And I got to the
point where we wouldn't mention it to any of our professors, because they would just eye roll and
think we committed career suicide. And so there's a lot of things that we had to do,
but we always believed it. And one reason, by the way, one reason I've always believed in
reinforcement learning is that if you look at neuroscience, that is the way that the
primate brain learns. One of the main mechanisms is the dopamine system implements some form of
TD learning, a very famous result in the late 90s, where they saw this in monkeys. And as a
propagating prediction error. So again, in the limit, this is what I think you can use
neuroscience for is at mathematics. When you're doing something as ambitious as trying to solve
intelligence, and it's blue sky research, no one knows how to do it, you need to use any evidence
or any source of information you can to help guide you in the right direction or give you
confidence you're going in the right direction. So that was one reason we pushed so hard on that.
And just going back to your earlier question about organization, the other big thing that
I think we innovated with at DeepMind to encourage invention and innovation was the
multidisciplinary organization we built, and we still have today. So DeepMind originally
was a confluence of the most cutting edge knowledge in neuroscience with machine learning,
engineering and mathematics, and gaming. And then since then, we've built that out even further.
So we have philosophers here and by ethicists, but also other types of scientists, physicists
and so on. And that's what brings together, I tried to build a sort of new type of Bell labs,
but in its golden era, and a new expression of that to try and foster this incredible
sort of innovation machine. So talking about the humans in the machine, DeepMind itself is a
learning machine with a lot of amazing human minds in it coming together to try and build these
learning systems. If we return to the big ambitious dream of AlphaFold that may be the
early steps on a very long journey in biology, do you think the same kind of approach can use
to predict the structure and function of more complex biological systems? So multi-protein,
interaction, and then, I mean, you can go out from there, just simulating bigger and bigger
systems that eventually simulate something like the human brain or the human body, just the big
mush, the mess of the beautiful, resilient mess of biology. Do you see that as a long-term vision?
I do. And I think, if you think about what are the top things I wanted to apply AI to once we
had powerful enough systems, biology and curing diseases and understanding biology was right
up there, you know, top of my list. That's one of the reasons I personally pushed that myself and
with AlphaFold. But I think AlphaFold, amazing as it is, is just the beginning. And I hope it's
evidence of what could be done with computational methods. So, you know, AlphaFold solved this
huge problem of the structure of proteins, but biology is dynamic. So really, what I imagine
from here, and we're working on all these things now, is protein-protein interaction,
protein-ligand binding, so reacting with molecules. Then you want to build up to pathways
and then eventually a virtual cell. That's my dream, maybe in the next 10 years. And I've been
talking actually to a lot of biologists, friends of mine, Paul Nurse, who runs the Crick Institute,
amazing biologists, Nobel Prize-winning biologists. We've been discussing for 20 years now,
virtual cells. Could you build a virtual simulation of a cell? And if you could, that would be
incredible for biology and disease discovery, because you could do loads of experiments on the
virtual cell and then only at the last stage validate it in the wet lab. So you could, you know,
in terms of the search space of discovering new drugs, you know, it takes 10 years roughly to go
from, you know, identifying a target to having a drug candidate. Maybe that could be shortened to,
you know, by an order of magnitude with if you could do most of that work in silico.
So in order to get to a virtual cell, we have to build up understanding of different parts of
biology and the interactions. And so, you know, every few years we talk about this, I talked
about this with Paul. And then finally, last year after AlphaFold, I said, now's the time we can
finally go for it. And AlphaFold's the first proof point that this might be possible. And he's
very exciting. We have some collaborations with his lab. They're just across the road, actually,
from us. It's a wonderful being here in King's Cross with the Crick Institute across the road.
And I think the next steps, you know, I think there's going to be some amazing advances in
biology built on top of things like AlphaFold. We're already seeing that with the community
doing that after we've open sourced it and released it. And, you know, I often say that
I think if you think of mathematics as the perfect description language for physics,
I think AI might end up being the perfect description language for biology. Because
biology is so messy, it's so emergent, so dynamic and complex. I think I find it very hard to believe
we'll ever get to something as elegant as Newton's Laws of Motions to describe a cell.
Right. It's just too complicated. So I think AI is the right tool for this.
You have to start at the basic building blocks and use AI to run the simulation
for all those building blocks. So have a very strong way to do prediction of what,
given these building blocks, what kind of biology, how the function and the evolution of
that biological system. It's almost like a cellular automata. You have to run it. You can't
analyze it from a high level. You have to take the basic ingredients, figure out the rules,
and let it run. But in this case, the rules are very difficult to figure out.
Yes.
You have to learn them.
That's exactly it. So the biology is too complicated to figure out the rules. It's too
emergent, too dynamic, say, compared to a physics system, like the motion of a planet.
Yeah.
Right. And so you have to learn the rules. And that's exactly the type of systems that we're
building.
So you mentioned you've open sourced AlphaFold and even the data involved. To me, personally,
I'm also really happy and a big thank you for open sourcing Majoko, the physics simulation
engine that's often used for robotics research and so on. So I think that's a pretty gangster
move. So what's the... I mean, very few companies or people would do that kind of thing. What's
the philosophy behind that?
You know, it's a case by case basis. And in both those cases, we felt that was the maximum
benefit to humanity to do that. And the scientific community, in one case, the robotics physics
community with Majoko.
We purchased it.
We purchased it for...
Yes, we purchased it for the express principle to open sourced. So I hope people appreciate
that. It's great to hear that you do. And then the second thing was... And mostly we did it
because the person building it was not able to cope with supporting it anymore because
it got too big for him. He's an amazing professor who built it in the first place. So we helped
him out with that. And then with AlphaFold's even bigger, I would say. And I think in that
case, we decided that there were so many downstream applications of AlphaFold that we couldn't
possibly even imagine what they all were. So the best way to accelerate drug discovery
and also fundamental research would be to give all that data away and the system
itself. It's been so gratifying to see what people have done that within just one year,
which is a short amount of time in science. And it's been used by over 500,000 researchers
have used it. We think that's almost every biologist in the world. I think there's roughly
500,000 biologists in the world, professional biologists, have used it to look at their
proteins of interest. We've seen amazing fundamental research done. So a couple of weeks ago,
there was a whole special issue of science, including the front cover,
which had the nuclear pore complex on it, which is one of the biggest proteins in the body. The
nuclear pore complex is a protein that governs all the nutrients going in and out of your cell
nucleus. So they're like little gateways that open and close to let things go in and out of
your cell nucleus. So they're really important. But they're huge because they're massive doughnut
ring shaped things. And they've been looking to try and figure out that structure for decades.
And they have lots of experimental data, but it's too low resolution. There's bits missing.
And they were able to, like a giant Lego jigsaw puzzle, use alpha fold predictions plus experimental
data and combined those two independent sources of information, actually four different groups
around the world were able to put it together more or less simultaneously using alpha fold
predictions. So that's been amazing to see. And pretty much every pharma company, every drug company
executive I've spoken to has said that their teams are using alpha fold to accelerate whatever
drugs they're trying to discover. So I think the knock on effect has been enormous in terms of
the impact that alpha fold has made. And it's probably bringing in, it's creating biologists,
it's bringing more people into the field, both on the excitement and both on the technical skills
involved. And it's almost like a gateway drug to biology. Yes, it is. And more computational
people involved too, hopefully. And I think for us, the next stage, as I said, in future,
we have to have other considerations too. We're building on top of alpha fold and these other
ideas I discussed with you about protein, protein interactions and genomics and other things.
And not everything will be open source. Some of it will do commercially because that will be the
best way to actually get the most resources and impact behind it. In other ways, some other
projects will do a nonprofit style. And also, we have to consider for future things as well,
safety and ethics as well, like synthetic biology, there are, there is dual use.
And we have to think about that as well. With alpha fold, we consulted with 30 different
bioethicists and other people expert in this field to make sure it was safe before we released it.
So there'll be other considerations in future. But for right now, I think alpha fold is a kind
of a gift from us to the scientific community. So I'm pretty sure that something like alpha fold
would be part of Nobel prizes in the future. But us humans, of course, are horrible with credit
assignment. So we'll of course give it to the humans. Do you think there will be a day when
AI system can't be denied that it earned that Nobel prize? Do you think we will see that in
21st century? It depends what type of AI as we end up building, whether they're
goal seeking agents who specifies the goals, who comes up with the hypotheses, who determines
which problems to tackle, right? And tweets about it, announcement of the results.
Yes, it's announcement of the results, exactly as part of it. So I think right now, of course,
it's amazing human ingenuity that's behind these systems. And then the system, in my opinion,
is just a tool. It would be a bit like saying with Galileo and his telescope, the ingenuity,
that the credit should go to the telescope. I mean, it's clearly Galileo building the tool
which he then uses. So I still see that in the same way today, even though these tools learn for
themselves. I think of things like alpha fold and the things we're building as the ultimate tools
for science and for acquiring new knowledge to help us as scientists acquire new knowledge.
I think one day there will come a point where an AI system may solve or come up with something like
general relativity of its own bat, not just by averaging everything on the internet or averaging
everything on PubMed. Although that would be interesting to see what that would come up with.
So that to me is a bit like our earlier debate about creativity, you know, inventing go,
rather than just coming up with a good go move. And so I think solving, I think to, you know,
if we wanted to give it the credit of like a Nobel type of thing, then it would need to invent go
and sort of invent that new conjecture out of the blue, rather than being specified by the
human scientists or the human creators. So I think right now it's definitely just a tool.
Although it is interesting how far you get by averaging everything on the internet, like you
said, because, you know, a lot of people do see science as you're always standing on the shoulders
of giants. And the question is how much are you really reaching up above the shoulders of giants?
Maybe it's just assimilating different kinds of results of the past with ultimately this new
perspective that gives you this breakthrough idea. But that idea may not be novel in the way that
it can't be already discovered on the internet. Maybe the Nobel prizes of the next 100 years
are already all there on the internet to be discovered. They could be. They could be. I mean,
I think this is one of the big mysteries, I think, is that I, first of all, I believe a lot of the
big new breakthroughs that are going to come in the next few decades. And even in the last decade
are going to come at the intersection between different subject areas, where there'll be some
new connection that's found between what seemingly were disparate areas. And one can even think of
deep mind, as I said earlier, as a sort of interdiscipline between neuroscience ideas and AI
engineering ideas originally. And so I think there's that. And then one of the things we can't
imagine today is, and one of the reasons I think people, we were so surprised by how well large
models worked is that actually, it's very hard for our human minds, our limited human minds to
understand what it would be like to read the whole internet, right? I think we can do a thought
experiment. And I used to do this of like, well, what if I read the whole of Wikipedia? What would
I know? And I think our minds can just about comprehend maybe what that would be like, but
the whole internet is beyond comprehension. So I think we just don't understand what it would be
like to be able to hold all of that in mind, potentially, right? And then active at once.
And then maybe what are the connections that are available there? So I think no doubt there are huge
things to be discovered just like that. But I do think there is this other type of creativity of
true spark of new knowledge, new idea never thought before about can't be averaged from things that
are known, that really, of course, everything come, you know, nobody creates in a vacuum. So
there must be clues somewhere. But just a unique way of putting those things together, I think
some of the greatest scientists in history have displayed that I would say, although it's very
hard to know, going back to their time, what was exactly known when they came up with those things.
Although you're making me really think because just the thought experiment
of deeply knowing 100 Wikipedia pages, I don't think I can. I've been really impressed by Wikipedia
for technical topics. So if you know 100 pages or 1000 pages, I don't think we can
truly comprehend what kind of intelligence that is. It's a pretty powerful. If you know how to
use that and integrate that information correctly, I think you can go really far. You can probably
construct thought experiments based on that, like simulate different ideas. So if this is true,
let me run this thought experiment that maybe this is true. It's not really invention. It's
like just taking literally the knowledge and using it to construct a very basic simulation of the
world. I mean, some argue it's romantic in part, but Einstein would do the same kind of things
with thought experiments. Yeah, one could imagine doing that systematically across millions of
Wikipedia pages plus PubMed, all these things. I think there are many, many things to be discovered
like that that are hugely useful. You could imagine, and I want us to do some of these things in
material science, like room temperature superconductors or something on my list one day that I'd
like to have an AI system to help build better optimized batteries. All of these sort of mechanical
things, I think a systematic sort of search could be guided by a model, could be extremely
powerful. So speaking of which, you have a paper on nuclear fusion, magnetic control of
tachymic plasmus through deeper enforcement learning. So you're seeking to solve nuclear fusion
with deep RL, so it's doing control of high temperature plasmas. Can you explain this work
and can AI eventually solve nuclear fusion? It's been very fun last year or two and very productive
because we've been ticking off a lot of my dream projects, if you like, of things that I've collected
over the years of areas of science that I would like to, I think could be very transformative
if we helped accelerate and are really interesting problems, scientific challenges in of themselves.
This is energy. So energy, yes, exactly. So energy and climate. So we talked about disease and
biology as being one of the biggest places I think AI can help with. I think energy and climate
is another one. So maybe they would be my top two. And fusion is one area I think AI can help with.
Now, fusion has many challenges, mostly physics and material science and engineering challenges
as well to build these massive fusion reactors and contain the plasma. And what we try to do,
and whenever we go into a new field to apply our systems is we look for, we talk to domain experts,
we try and find the best people in the world to collaborate with. In this case, in fusion,
we collaborate with EPFL in Switzerland, the Swiss Technical Institute, who are amazing.
They have a test reactor. They were willing to let us use, which I double-checked with the team we
were going to use carefully and safely. I was impressed. They managed to persuade them to let
us use it. And it's an amazing test reactor they have there. And they try all sorts of pretty crazy
experiments on it. And what we tend to look at is if we go into a new domain like fusion,
what are all the bottleneck problems? Thinking from first principles, what are all the
bottleneck problems that are still stopping fusion working today? And then we get a fusion expert to
tell us. And then we look at those bottlenecks and we look at the ones which ones are amenable
to our AI methods today. And we'd be interesting from a research perspective, from our point of
view, from an AI point of view. And that would address one of their bottlenecks. And in this case,
plasma control was perfect. So the plasma, it's a million degrees Celsius, something like that's
hotter than the sun. And there's obviously no material that can contain it. So they have to
be containing these magnetic, very powerful superconducting magnetic fields. But the problem
is plasma is pretty unstable, as you imagine. You're kind of holding a mini-sun, mini-star,
in a reactor. So you kind of want to predict ahead of time what the plasma is going to do,
so you can move the magnetic field within a few milliseconds to basically contain what it's going
to do next. So it seems like a perfect problem if you think of it for a reinforcement learning
prediction problem. So you've got a controller, you've got to move the magnetic field. And until
we came along, they were doing it with traditional operational research type of controllers,
which are kind of handcrafted. And the problem is, of course, they can't react in the moment to
something the plasma is doing. They have to be hard coded. And again, knowing that that's normally
our go-to solution is we would like to learn that instead. And they also had a simulator of these
plasma. So there were lots of criteria that matched what we like to use.
So can AI eventually solve nuclear fusion?
Well, so with this problem, and we published it in Nature paper last year, we held the fusion,
that we held the plasma in a specific shape. So actually, it's almost like carving the plasma
into different shapes and hold it there for a record amount of time. So that's one of the
problems of fusion sort of solved. So have a controller that's able to, no matter the shape,
contain it. Yeah, contain it and hold it in structure. And there's different shapes that
are better for the energy productions called droplets and so on. So that was huge. And now
we're looking, we're talking to lots of fusion startups to see what's the next problem we can
tackle in the fusion area. So another fascinating place in a paper title,
pushing the frontiers of density functionals by solving the fractional electron problem.
So you're taking on modeling and simulating the quantum mechanical behavior of electrons.
Can you explain this work and can AI model and simulate arbitrary quantum mechanical
systems in the future? Yeah, so this is another problem I've had my eye on for decade or more,
which is sort of simulating the properties of electrons. If you can do that, you can basically
describe how elements and materials and substances work. So it's kind of like fundamental if you
want to advance material science. And we have Schrodinger's equation, and then we have approximations
to that density functional theory. These things are famous. And people try and write approximations
to these to these functionals and kind of come up with descriptions of the electron clouds,
where they're going to go, how they're going to interact when you put two elements together.
And what we try to do is learn a simulation, learn a functional that will describe more
chemistry, types of chemistry. So until now, you can run expensive simulations, but then you can
only simulate very small molecules, very simple molecules, we would like to simulate large materials.
And so today, there's no way of doing that. And we're building up towards building
functionals that approximate Schrodinger's equation, and then allow you to describe
what the electrons are doing. And all material sort of science and material properties are
governed by the electrons and how they interact. So have a good summarization of the simulation
through the functional, but one that is still close to what the actual simulation would come
out with. So what, how difficult is that task? What's involved in that task? Because they're
running those, those complicated simulations and learning the task of mapping from the initial
conditions and the parameters of the simulation, learning what the functional would be.
Yeah. So it's pretty tricky. And we've done it with, you know, the nice thing is we there are,
we can run a lot of the simulations, the molecular dynamic simulations on our compute clusters.
And so that generates a lot of data. So in this case, the data is generated.
So we like those sort of systems and that's why we use games, it's simulator generator data.
And we can kind of create as much of it as we want really. And just let's leave some,
you know, if any computers are free in the cloud, we just run, we run some of these calculations,
right, compute cluster calculation. I like how the free compute time is used up on quantum
mechanics. Yeah, quantum mechanics, exactly, simulations and protein simulations and other
things. And so, and so, you know, when you're not searching on YouTube for video, cat videos,
we're using those computers usefully and quantum chemistry, the idea and, and putting them for
good use. And then, yeah, and then all of that computational data that's generated,
we can then try and learn the functionals from that, which of course are way more efficient
once we learn the functional than running those simulations would be.
Do you think one day AI may allow us to do something like basically crack open physics,
so do something like travel faster than the speed of light?
My ultimate aim has always been with AI is the reason I am personally working on AI for my whole
life was to build a tool to help us understand the universe. So I wanted to, and that means physics,
really, and the nature of reality. So I don't think we have systems that are capable of doing
that yet. But when we get towards AGI, I think that's one of the first things I think we should
apply AGI to. I would like to test the limits of physics and our knowledge of physics. There's so
many things we don't know. This is one thing I find fascinating about science and, you know,
as a huge proponent of the scientific method as being one of the greatest ideas humanities ever had
and allowed us to progress with our knowledge. I think as a true scientist, I think what you
find is the more you find out, the more you realize we don't know. And I always think that it's
surprising that more people aren't troubled. Every night, I think about all these things we interact
with all the time, that we have no idea how they work, time, consciousness, gravity, life.
These are all the fundamental things of nature. We don't really know what they are.
To live life, we pin certain assumptions on them and kind of treat our assumptions as if
they're a fact that allows us to sort of box them off somehow. But the reality is, when you think of
time, you should remind yourself, you should take it off the shelf and realize, like, no,
we have a bunch of assumptions. There's even not a lot of debate. There's a lot of uncertainty
about exactly what is time. Is there an error of time? There's a lot of fundamental questions
that you can't just make assumptions about. And maybe AI allows you to not put anything on the
shelf, not make any hard assumptions and really open it up and see what's-
Exactly. I think we should be truly open-minded about that. And exactly that, not be dogmatic
to a particular theory. It'll also allow us to build better tools, experimental tools eventually
that can then test certain theories that may not be testable today about things about what
we spoke about at the beginning, about the computational nature of the universe,
how one might, if that was true, how one might go about testing that. There are people who've
conjectured people like Scott Aronson and others about how much information can a specific
plank unit of space and time contain. So one might be able to think about testing those ideas
if you had AI helping you build some new, exquisite, experimental tools. This is what I
imagine many decades from now will be able to do. And what kind of questions can be answered
through running a simulation of them? There's a bunch of physics simulations you can imagine
that could be run in some kind of efficient way, much like you're doing in the quantum
simulation work. And perhaps even the origin of life. So figuring out how going even back before
the work of AlphaFold begins, of how this whole thing emerges from a rock. Yes.
From a static thing. Do you think AI will allow us to- Is that something you have your eye on?
Is trying to understand the origin of life? First of all, yourself, what do you think?
How the heck did life originate on Earth? Yeah, well, maybe I'll come to that in a second. But
I think the ultimate use of AI is to use it to accelerate science to the maximum. So I
think of it a little bit like the tree of all knowledge. If you imagine that's all the knowledge
there is in the universe to attain. And we barely scratched the surface of that so far
in even though we've done pretty well since the Enlightenment as humanity. And I think AI will
turbocharge all of that like we've seen with AlphaFold. And I want to explore as much of that
tree of knowledge as is possible to do. And I think that involves AI helping us with
understanding or finding patterns, but also potentially designing and building new tools,
experimental tools. So I think that's all and also running simulations and learning simulations.
All of that, we're sort of doing at a baby steps level here. But I can imagine that in
the decades to come as what's the full flourishing of that line of thinking. It's going to be
truly incredible, I would say. If I visualize this tree of knowledge,
something tells me that that tree of knowledge for humans is much smaller in the set of all
possible trees of knowledge. It's actually quite small, giving our cognitive limitations,
limited cognitive capabilities, that even with the tools we build, we still won't be able to
understand a lot of things. And that's perhaps what non-human systems might be able to reach
farther, not just as tools, but in themselves understanding something that they can bring back.
Yeah, it could well be. So there's so many things that are sort of encapsulated in what you just
said there. I think, first of all, there's two different things there. What do we understand
today? What could the human mind understand? And what is the totality of what is there to be
understood? And so you can think of them as three larger and larger trees or exploring
more branches of that tree. And I think with AI, we're going to explore that whole lot.
Now, the question is, if you think about what is the totality of what could be understood,
there may be some fundamental physics reasons why certain things can't be understood,
like what's outside a simulation or outside the universe. Maybe it's not understandable from within
the universe. So there may be some hard constraints like that. Could be smaller constraints like
we think of space-time as fundamental. Our human brains are really used to this idea of a
three-dimensional world with time. But our tools could go beyond that. They wouldn't have that
limitation necessarily. They could think in 11 dimensions, 12 dimensions, whatever is needed.
But we could still maybe understand that in several different ways. The example I always give is,
when I play Gary Kasparov at Speed Chess or we've talked about chess and these kind of things,
if you're reasonably good at chess, you can't come up with the move Gary comes up with in his
move, but he can explain it to you. And you can understand. And you can understand post hoc the
reasoning. So I think there's an even further level of like, well, maybe you couldn't have invented
that thing, but going back to using language again, perhaps you can understand and appreciate that.
Same way, you can appreciate Vivaldi or Mozart or something without, you can appreciate the beauty
of that without being able to construct it yourself, right? Invent the music yourself.
So I think we see this in all forms of life. So it'll be that times a million, but you can imagine
also one sign of intelligence is the ability to explain things clearly and simply, right?
People like Richard Feyn, another one of my all-time heroes used to say that, right? If you
can't, if you can explain it something simply, then that's the best sign, a complex topic simply,
then that's one of the best signs of you understanding it. I can see myself talking trash
in the AI system in that way. It gets frustrated at how dumb I am and trying to explain something
to me. I was like, well, that means you're not intelligent because if you were intelligent,
you'd be able to explain it simply. Yeah, of course, there's also the other option,
of course, we could enhance ourselves and with our devices. We are already sort of symbiotic
with our compute devices, right? With our phones and other things. And there's stuff like Neuralink
and Exeptra that could advance that further. So I think there's lots of really amazing
possibilities that I could foresee from here. Well, let me ask you some wild questions. So out
there, looking for friends, do you think there's a lot of alien civilizations out there?
So I guess this also goes back to your origin of life question too, because I think that's key.
My personal opinion, looking at all this, and it's one of my hobbies, physics, I guess. So
something I think about a lot and talk to a lot of experts on and read a lot of books on.
And I think my feeling currently is that we are alone. I think that's the most likely scenario,
given what evidence we have. And the reasoning is I think that we've tried since things like
SETI program and I guess since the dawning of the space age, we've had telescopes,
open radio telescopes and other things. And if you think about and try to detect signals,
now, if you think about the evolution of humans on Earth, we could have easily been
a million years ahead of our time now or million years behind, right easily, with just some
slightly different quirk thing happening hundreds of thousands years ago, things could have been
slightly different. If the meteor would hit the dinosaurs a million years earlier, maybe things
would have evolved, we'd be a million years ahead of where we are now. So what that means is if you
imagine where humanity will be in a few hundred years, let alone a million years, especially if we
hopefully, you know, solve things like climate change and other things and we continue to flourish
and we build things like AI and we do space traveling and all of the stuff that humans
have dreamed of forever, right, and sci-fi has talked about forever. We will be spreading across
the stars, right, and von Neumann famously calculated, you know, it would only take about a
million years if you send out von Neumann probes to the nearest, you know, the nearest other solar
systems and then they built, all they did was build two more versions of themselves and set
those two out to the next nearest systems. You, you know, within a million years, I think you would
have one of these probes in every system in the galaxy. So it's not actually in cosmological
time, that's actually a very short amount of time. So, and you know, we people like Dyson have
thought about constructing Dyson spheres around stars to collect all the energy coming out of
the star, you know, that there would be constructions like that would be visible across space,
probably even across a galaxy. So, and then, you know, if you think about all of our radio,
television, emissions that have gone out since, since the, you know, 30s and 40s,
imagine a million years of that and now hundreds of civilizations doing that. When we opened our
years, at the point we got technologically sophisticated enough in the space age, we should
have heard a cacophony of voices. We should have joined that cacophony of voices and what,
what we did, we opened our ears and we heard nothing. And many people who argue that there are
aliens would say, well, we haven't really done exhaustive search yet and maybe we're looking in
the wrong bands and, and we've got the wrong devices and we wouldn't notice what an alien form was
like to be so different to what we're used to. But you know, I don't, I don't really buy that,
that it shouldn't be as difficult as that. Like we, I think we've searched enough.
There should be everywhere. If it was, yeah, it should be everywhere.
We should see Dyson's fears being put up, suns blinking in and out. You know,
there should be a lot of evidence for those things. And then there are other people who argue,
well, the sort of safari view of like, well, we're a primitive species still because we're not
space faring yet. And, and, and we're, you know, there's some kind of globe, like universal rule
not to interfere. Yeah, Star Trek rule. But like, look, look, we can't even coordinate humans
to deal with climate change. And we're one species. What, what is the chance that of all
of these different human civilization, you know, alien civilizations, they would have the same
priorities and, and, and agree or cross the, you know, these kind of matters. And even if that was
true, and we were in some sort of safari for our own good, to me, that's not much different from
the simulation hypothesis. Because what does it mean, the simulation hypothesis? I think in
its most fundamental level, it means what we're seeing is not quite reality, right? It's something,
there's something more deeper underlying it, maybe computational. Now, if we were in a,
if we were in a sort of safari park, and everything we were seeing was a hologram and
it was projected by the aliens or whatever, that to me is not much different than thinking
we're inside of another universe, because we still can't see true reality, right?
I mean, there's, there's other explanations. It could be that the way they're communicating is
just fundamentally different, that we're too dumb to understand the much better methods of
communication they have. It could be, I mean, I mean, it's silly to say, but our own thoughts
could be the methods by which they're communicating, like the place from which our ideas, writers talk
about this, like the muse. It sounds like very kind of wild, but it could be thoughts,
it could be some interactions with our mind that we think are originating from us is actually
something that is coming from other life forms elsewhere. Consciousness itself might be that.
It could be, but I don't see any sensible argument to the why, why would all of the alien species
behave this way? Some of them will be more primitive, they will be close to our level.
There should be a whole sort of norm distribution of these things, right? Some would be aggressive,
some would be, you know, curious, others would be very stoical and philosophical,
because maybe they're a million years older than us, but it's not, it shouldn't be like,
I mean, one alien civilization might be like that and communicating thoughts and others,
but I don't see why, you know, potentially the hundreds there should be would be uniform in this
way, right? It could be a violent dictatorship that the people, the alien civilizations that
become successful, gain the ability to be destructive, an order of magnitude more
destructive. But of course, the sad thought, well, either humans are very special. We took a lot of
leaps that arrived at what it means to be human. There's a question there, which was the hardest,
which was the most special, but also if others have reached this level, and maybe many others
have reached this level, the great filter that prevented them from going farther to becoming
a multiplayer species or reaching out into the stars. And those are really important questions
for us, whether there's other alien civilizations that there are not, this is very useful for us
to think about. If we destroy ourselves, how will we do it, and how easy is it to do?
Yeah. Well, these are big questions, and I've thought about these a lot, but the interesting
thing is that if we're alone, that's somewhat comforting from the great filter perspective,
because it probably means the great filters were past us, and I'm pretty sure they are. So going
back to your origin of life question, there are some incredible things that no one knows how
happened. Like obviously, the first life form from chemical soup, that seems pretty hard.
But I would guess the multicellular, I wouldn't be that surprised if we saw single cell sort of
life forms elsewhere, bacteria type things. But multicellular life seems incredibly hard,
that step of capturing mitochondria and then sort of using that as part of yourself when
you've just eaten it. Would you say that's the biggest, the most, if you had to choose one,
sort of Hitchhiker's got this galaxy, one set in summary of like, oh, those clever
creatures did this, there would be the multicellular. I think that's probably the one that's the
biggest. I mean, there's a great book called The Ten Great Inventions of Evolution by Nic Laine,
and he speculates on 10 of these, what could be great filters. I think that's one. I think the
advent of intelligence and conscious intelligence and in order to us to be able to do science and
things like that is huge as well. I mean, it's only evolved once as far as in earth history.
So that would be a later candidate. But certainly for the early candidates,
I think multicellular life forms is huge. By the way, what it's interesting to ask you,
if you can hypothesize about what is the origin of intelligence, is it that we started
cooking meat over fire? Is it that we somehow figured out that we could be very powerful when
we started collaborating? So cooperation between our ancestors, so that we can overthrow the alpha
male. What is it, Richard? I talked to Richard Randham, who thinks we're all just beta males who
figured out how to collaborate to defeat the one, the dictator, the authoritarian alpha male
that controlled the tribe. Is there other explanation? Was there a 2001 space obviously
type of monolith that came down to earth? Well, I think all of those things you suggested
are good candidates, fire and cooking, right? So that's clearly important for energy efficiency,
cooking our meat, and then being able to be more efficient about eating it and consuming
the energy. I think that's huge and then utilizing fire and tools. I think you're right about the
tribal cooperation aspects and probably language as part of that, because probably that's what
allowed us to outcompete Neanderthals and perhaps less cooperative species. So that may be the case,
toolmaking, spears, axes. I think it's pretty clear now that humans were responsible for
a lot of the extinctions of megafauna, especially in the Americas when humans arrived.
So you can imagine once you discover a tool usage, how powerful that would have been and how
scary for animals. So I think all of those could have been explanations for it. The interesting
thing is that it's a bit like general intelligence too. It's very costly to begin with, to have a
brain and especially a general purpose brain rather than a special purpose one, because you
might have energy our brains use. I think it's like 20% of the body's energy and it's massive.
And when you're thinking chess, one of the funny things that we used to say is that as much as
a racing driver uses for a whole Formula One race, just playing a game of serious high-level
chess, which we know you wouldn't think just sitting there, because the brain's using so much
energy. So in order for an animal and organism to justify that, there has to be a huge payoff.
And the problem with half a brain or half intelligence, say an IQ of a monkey brain,
it's not clear you can justify that evolutionary until you get to the human level brain.
But how do you do that jump? It's very difficult, which is why I think it's only been done once,
from the sort of specialized brains that you see in animals to this sort of general purpose,
chewing powerful brains that humans have, which allows us to invent the modern world.
And it takes a lot to cross that barrier. And I think we've seen the same with AI systems,
which is that maybe until very recently, it's always been easier to craft a specific solution
to a problem like chess than it has been to build a general learning system that could
potentially do many things. Because initially, that system will be way worse than less efficient
than the specialized system. So one of the interesting quirks of the human mind of this
evolved system is that it appears to be conscious. This thing that we don't quite understand,
but it seems very special, its ability to have a subjective experience that it feels like something
to eat a cookie, the deliciousness of it or see a color and that kind of stuff.
Do you think in order to solve intelligence, we also need to solve consciousness along the way?
Do you think AGI systems need to have consciousness in order to be truly intelligent?
Yeah, we thought about this a lot actually. And I think that my guess is that consciousness
and intelligence are double dissociable. So you can have one without the other both ways.
And I think you can see that with consciousness in that I think some animals and pets, if you
have a pet dog or something like that, you can see some of the higher animals and dolphins,
things like that, have self-awareness and a very sociable, seem to dream.
You know, those kinds of, a lot of the traits one would regard as being kind of conscious
and self-aware. But yet they're not that smart, right? So they're not that intelligent by,
say IQ standards or something like that. Yeah, it's also possible that our understanding
of intelligence is flawed, like putting an IQ to it. Maybe the thing that a dog can do
is actually gone very far along the path of intelligence and we humans are just able to
play chess and maybe write poems. Right. But if we go back to the idea of AGI and general
intelligence, you know, dogs are very specialized, right? Most animals are pretty specialized.
They can be amazing at what they do, but they're like kind of elite sports, sports people or
something, right? So they have one thing extremely well because their entire brain is,
is optimized. They have somehow convinced the entirety of the human population to feed them
and service them. So in some way, they're controlling. Yes, exactly. Well, we co-evolved
to some crazy degree, right? Including the way the dogs, you know, even, even wag their tails
and twitch their noses, right? We find, we find inexorably cute. But I think you can also see
intelligence on the other side. So systems like artificial systems that are amazingly
smart at certain things like maybe playing go in chess and other things, but they don't feel
at all in any shape or form conscious in the way that, you know, you do to me or I do to you.
And, and I think actually building AI is these intelligent constructs is one of the best ways
to explore the mystery of consciousness to break it down because we're going to have devices that
are pretty smart at certain things or capable at certain things, but potentially won't have any
semblance of self-awareness or other things. And in fact, I would advocate, if there's a choice,
building systems in the first place, AI systems that are not conscious to begin with are just tools
until we understand them better and the capabilities better.
So on that topic, just not as the CEO of DeepMind, just as a human being, let me ask you about this
one particular anecdotal evidence of the Google engineer who made a comment or believed that
there's some aspect of a language model, the Lambda language model that exhibited sentience.
So you said you believe there might be a responsibility to build systems that are not
sentient. And this experience of a particular engineer, I think, I'd love to get your general
opinion on this kind of thing, but I think it will happen more and more and more, which not
one engineers, but when people out there that don't have an engineer background start interacting
with increasingly intelligent systems, we anthropomorphize them, they start to have
deep impactful interactions with us in a way that we miss them when they're gone.
And we sure as heck feel like they're living entities, self-aware entities, and maybe even
we project sentience onto them. So what's your thought about this particular
system? Have you ever met a language model that's sentient on record?
What do you make of the case of when you feel that there's some elements of sentience to this
system? Yeah, so this is an interesting question and obviously a very fundamental one. So the
first thing to say is I think that none of the systems we have today, I would say, even have
one iota of semblance of consciousness or sentience, that's my personal feeling interacting with them
every day. So I think this way premature to be discussing what that engineer talked about,
I think at the moment it's more of a projection of other way our own minds work, which is to see
sort of purpose and direction in almost anything that we, our brains are trained to interpret
agency, basically in things, even inanimate things sometimes. And of course with a language
system, because language is so fundamental to intelligence, that's going to be easy for us
to anthropomorphize that. I mean, back in the day, even the first, the dumbest sort of template
chatbots ever, Eliza and the ilk of the original chatbots back in the 60s fooled some people under
certain circumstances, right, it pretended to be a psychologist. So we just basically wrap it back
to you, the same question you asked it back to you. And some people believe that. So I don't
think we can, this is why I think the truing test is a little bit flawed as a formal test,
because it depends on the sophistication of the, of the judge, whether or not they are
qualified to make that distinction. So I think we should talk to, you know, the top philosophers
about this people like Daniel Dennett and David Chalmers and others who've obviously thought
deeply about consciousness. Of course, consciousness itself hasn't been, well, there's no agreed
definition. If I was to, you know, speculate about that, you know, I kind of the definite,
the working definition I like is, it's the way information feels when, you know, it gets processed.
I think maybe Max Tegmark came up with that. I like that idea. I don't know if it helps us get
towards any more operational thing. But it's, it's, it's, I think it's a nice way of viewing it.
I think we can obviously see from neuroscience certain prerequisites that are required,
like self-awareness, I think is necessary, but not sufficient component, this idea of a self and
other and set of coherent preferences that are coherent over time. You know, these things are
maybe memory. These things are probably needed for a sentient or conscious being.
But the reason, the difficult thing I think for us when we get, and I think this is a really
interesting philosophical debate, is when we get closer to AGI and, you know, and much more powerful
systems than we have today, how are we going to make this judgment? And one way, which is the
Turing test is sort of a behavioral judgment, is, is the system exhibiting all the behaviors
that a human sentient or sentient being would, would, would exhibit? Is it answering the right
questions? Is it saying the right things? Is it indistinguishable from a human? And so on.
But I think there's a second thing that makes us as humans regard each other as sentient, right?
Why do we, why do we think this? And I debated this with Daniel Dennett. And I think there's a
second reason that's over often overlooked, which is that we're running on the same substrate, right?
So if we're exhibiting the same behavior, more or less as humans, and we're running on the same,
you know, carbon based biological substrate, the squishy, you know, few pounds of flesh in our skulls,
then the most parsimonious, I think, explanation is that you're feeling the same thing as I'm
feeling, right? But we will never have that second part, the substrate equivalence with a machine,
right? So we will have to only judge based on the behavior. And I think the substrate equivalence
is a critical part of why we make assumptions that we're conscious. And in fact, even with,
with animals, high level animals, why we think they might be, because they're exhibiting some
of the behaviors we would expect from a sentient animal. And we know they're made of the same
things, biological neurons. So we're going to have to come up with explanations or models of the gap
between substrate differences between machines and humans to get anywhere beyond the behavioral.
But to me, sort of the practical question is very interesting and very important.
When you have millions, perhaps billions of people believing that you have a sentient AI,
believing what that Google engineer believed, which I just see as an obvious, very near term
future thing, certainly on the path to AGI, how does that change the world? What's the responsibility
of the AI system to help those millions of people? And also, what's the ethical thing? Because
you can, you can make a lot of people happy by creating a meaningful, deep experience with a
system that's faking it before it makes it. And I don't, are we the right, who is to say,
what's the right thing to do? Should AI always be tools? Like why? Why are we constraining AI
to always be tools as opposed to friends? Yeah. I think, well, I mean, these are,
you know, fantastic questions and also critical ones. And we've been thinking about this
since the start of DeepMind and before that, because we plan for success and, you know,
however remote that looked like back in 2010. And we've always had sort of these ethical
considerations as fundamental at DeepMind. And my current thinking on the language models is,
and large models, is they're not ready. We don't understand them well enough yet.
And, you know, in terms of analysis tools and guardrails, what they can and can't do and so on
to deploy them at scale. Because I think, you know, there are big still ethical questions like
should an AI system always announce that it is an AI system to begin with? Probably yes.
What do you do about answering those philosophical questions about the feelings people may have
about AI systems, perhaps incorrectly attributed? So I think there's a whole bunch of research
that needs to be done first. To responsibly, before, you know, you can responsibly deploy
these systems at scale, that would be at least be my current position. Over time, I'm very confident
we'll have those tools, like interpretability questions and analysis questions. And then with
the ethical quandary, you know, I think there, it's important to look beyond just science.
That's why I think philosophy, social sciences, even theology, other things like that come into it,
where, you know, arts and humanities, what does it mean to be human and the spirit of
being human and to enhance that and the human condition, right? And allow us to experience
things we could never experience before and improve the overall human condition and humanity
overall, you know, get radical abundance, solve many scientific problems, solve disease.
So this is the era I think, this is the amazing era I think we're heading into if we do it right.
But we've got to be careful. We've already seen with things like social media how dual
use technologies can be misused by firstly by bad actors or naive actors or crazy actors,
right? So there's that set of just the common or garden misuse of existing dual use technology.
And then of course, there's an additional thing that has to be overcome with AI that
eventually it may have its own agency. So it could be good or bad in itself. So I think
these questions have to be approached very carefully using the scientific method, I would say,
in terms of hypothesis generation, careful control testing, not live AB testing out in the
world. Because with powerful dual technologies like AI, if something goes wrong, it may cause,
you know, a lot of harm before you can fix it. It's not like a, you know, an imaging app or game
app where, you know, if something goes wrong, it's relatively easy to fix and the harm is
relatively small. So I think it comes with, you know, the usual cliche of like with a lot of power
comes a lot of responsibility. And I think that's the case here with things like AI, given the
enormous opportunity in front of us. And I think we need a lot of voices and as many inputs into
things like the design of the systems and the values they should have and what goals should
they be put to, I think as wide a group of voices as possible beyond just the technologist is needed
to input into that and to have a say in that, especially when it comes to deployment of these
systems, which is when the rubber really hits the road, it really affects the general person
in the street rather than fundamental research. And that's why I say, I think as a first step,
it would be better if we have the choice to build these systems as tools to give and I'm not saying
that it should never, they should never go beyond tools because of course the potential is there
for it to go way beyond just tools. But I think that would be a good first step in order for us
to allow us to carefully experiment, understand what these things can do. So the leap between tool,
the sentient entity being as well should take very careful. Let me ask a dark personal question.
So you're one of the most brilliant people in the AI community, also one of the most kind.
And if I may say sort of love people in the community, that said,
creation of a super intelligent AI system would be one of the most powerful things in the world,
tools or otherwise. And again, as the old saying goes, power corrupts and absolute power corrupts,
absolutely. You are likely to be one of the people, I would say probably the most likely
person to be in the control of such a system. Do you think about the corrupting nature of power
when you talk about these kinds of systems that as all dictators and people have caused atrocities
in the past always think they're doing good, but they don't do good because the powers polluted
their mind about what is good and what is evil. Do you think about this stuff or we just focus
on language model? No, I think about them all the time. And I think what are the defences against
that? I think one thing is to remain very grounded and sort of humble no matter what you do or achieve.
And I try to do that. My best friends are still my set of friends from my undergraduate Cambridge
days. My families and friends are very important. I think trying to be a multidisciplinary person,
it helps to keep you humble because no matter how good you are at one topic, someone will be better
than you at that. And always relearning a new topic again from scratch or a new field is very
humbling. So for me, that's been biology over the last five years. Huge area topic and I just love
doing that, but it helps to keep you grounded and keeps you open-minded. And then the other
important thing is to have a really group, amazing set of people around you at your company or your
organization who are also very ethical and grounded themselves and help to keep you that way.
And then ultimately, just to answer your question, I hope we're going to be a big part of
birthing AI and that being the greatest benefit to humanity of any tool or technology ever and
getting us into a world of radical abundance and curing diseases and solving many of the big
challenges we have in front of us and then ultimately help the ultimate flourishing of
humanity to travel the stars and find those aliens if they are there. And if they're not there,
find out why they're not there, what is going on here in the universe.
This is all to come and that's what I've always dreamed about. But I think AI is too big an idea.
It's not going to be, there'll be a certain set of pioneers who get there first. I hope we're in
the vanguard so we can influence how that goes. And I think it matters which cultures they come
from and what values they have, the builders of AI systems. Because I think even though the AI
system is going to learn for itself, most of its knowledge, there'll be a residue in the system
of the culture and the values of the creators of that system. And there's interesting questions
to discuss about that geopolitically, different cultures as we're in a more fragmented world than
ever. Unfortunately, I think in terms of global cooperation, we see that in things like climate
where we can't seem to get our act together globally to cooperate on these pressing matters.
I hope that will change over time. Perhaps if we get to an era of radical abundance,
we don't have to be so competitive anymore. Maybe we can be more co-operative if resources aren't
so scarce. It's true that in terms of power corrupting and leading to destructive things,
it seems that some of the atrocities of the past happen when there's a significant
constraint on resources. I think that's the first thing. I don't think that's enough. I think
scarcity is one thing that's led to competition, zero-sum game thinking. I would like us to all
be in a positive sum world. And I think for that, you have to remove scarcity. I don't think that's
enough, unfortunately, to get world peace because there's also other corrupting things like wanting
power over people and this kind of stuff, which is not necessarily satisfied by just abundance.
But I think it will help. Ultimately, AI is not going to be run by any one person or one
organization. I think it should belong to the world, belong to humanity. I think there'll be
many ways this will happen. Ultimately, everybody should have a say in that.
Do you have advice for young people in high school and college? Maybe
if they're interested in AI or interested in having a big impact on the world, what they should do
to have a career they can be proud of or to have a life they can be proud of?
I love giving talks to the next generation. What I say to them is actually two things.
I think the most important things to learn about and to find out about when you're young
is what are your true passions? First of all, there's two things. One is find your true passions.
I think the way to do that is to explore as many things as possible when you're young and you have
the time and you can take those risks. I would also encourage people to look at finding the
connections between things in a unique way. I think that's a really great way to find a passion.
Second thing I would say advice is know yourself. Spend a lot of time understanding how you work
best. What are the optimal times to work? What are the optimal ways that you study?
How do you deal with pressure? Test yourself in various scenarios and try and improve your
weaknesses, but also find out what your unique skills and strengths are and then hone those.
That's what will be your super value in the world later on. If you can then combine those
two things and find passions that you're genuinely excited about, that intersect with what your unique
strong skills are, then you're onto something incredible and I think you can make a huge
difference in the world. Let me ask about know yourself. This is fun. This is fun. Quick questions
about day in the life, the perfect day, the perfect productive day in the life of Demesus
Hub. Maybe these days there's a lot involved. Maybe a slightly younger Demesus Hub where
you could focus on a single project maybe. How early do you wake up? Are you night out? Do you
wake up early in the morning? What are some interesting habits? How many dozens of cups
of coffees do you drink a day? What's the computer that you use? What's the setup? How many screens
will keyboard? Are we talking Emax Vim? Are we talking something more modern? There's a bunch
of those questions. Maybe day in the life, what's the perfect day involved?
Well, these days it's quite different from say 10, 20 years ago. Back 10, 20 years ago, it would have
been a whole day of research, individual research or programming, doing some experiment, neuroscience,
computer science experiment, reading lots of research papers and then perhaps at night
time reading science fiction books or playing some games.
But lots of focus, deep focused work on whether it's programming or reading research papers.
Yes. That would be lots of deep focused work. These days for the last five to 10 years,
I've actually got quite a structure that works very well for me now, which is that
I'm a complete night owl, always have been. I optimize for that. I basically do a normal
day's work, get into work about 11 o'clock and do work to about seven in the office.
I will arrange back-to-back meetings for the entire time of that. With as many as many people as
possible. That's my collaboration management part of the day. Then I go home, spend time with the
family and friends, have dinner, relax a little bit. Then I start a second day of work. I call it
my second day of work around 10 p.m., 11 p.m. That's the time till about the small hours of the
morning, 4, 5 in the morning, where I will do my thinking and reading research, writing research
papers. Sadly, I don't have time to code anymore, but it's not efficient to do that these days,
given the amount of time I have. That's when I do maybe do the long stretches of thinking
and planning. Then probably using email or other things, I would fire off a lot of things to my
team to deal with the next morning. Actually, thinking about this overnight, we should go
for this project or arrange this meeting the next day. When you think it through a problem,
are you talking about sheet of paper? Is there some structured process?
I still like pencil and paper best for working out things, but these days it's just so efficient
to read research papers just on the screen. I still often print them out, actually. I still
prefer to mark out things. I find it goes into the brain better and sticks in the brain better
when you're still using physical pen and pencil and paper.
So you take notes? I have lots of notes, electronic ones and also whole stacks of notebooks
that I use at home. On some of these most challenging next steps, for example,
none of us know about that you're working on, you're thinking there's some deep thinking
required there. What is the right problem? What is the right approach? Because you're going to
have to invest a huge amount of time for the whole team. They're going to have to pursue this thing.
What's the right way to do it? Is RL going to work here or not? Yes.
Yes. What's the right thing to try? What's the right benchmark to you? Do we need to
construct a benchmark from scratch? All those kinds of things.
Yes. So I think of all those kind of things in the nighttime phase, but also much more.
I find I've always found the quiet hours of the morning when everyone's asleep. It's super
quiet outside. I love that time. It's the golden hours between like one and three in the morning.
Yes. Put some music on, some inspiring music on, and then think these deep thoughts. So that's
when I would read my philosophy books and Spinoza's, my recent Freire Cant, all these things. I read
about a great scientist of history, how they did things, how they thought things. So that's when
you do all your creative thinking. That's when I do all my creative thinking. It's good. I think
people recommend you do your sort of creative thinking in one block. The way I organize the day,
that way I don't get interrupted because obviously no one else is up at those times.
I can sort of get super deep and super into flow. The other nice thing about doing it
nighttime wise is if I'm really onto something or I've got really deep into something, I can
choose to extend it and I'll go into six in the morning, whatever, and then I'll just pay for
it the next day because I'll be a bit tired and I won't be my best. But that's fine. I can decide,
looking at my schedule the next day, and given where I'm at with this particular thought or
creative idea, that I'm going to pay that cost the next day. So I think that's more flexible
than morning people who do that, they get up at four in the morning. They can also do those
golden hours then. But then their start of their scheduled day starts at breakfast,
AAM, whatever they have their first meeting. And then it's hard. You have to reschedule
day if you and flow. Yeah, that could be a truly special thread of thoughts that
you're too passionate about. This is where some of the greatest ideas could potentially come is
when you just lose yourself late into the night. And for the meetings, I mean, you're loading in
really hard problems in a very short amount of time. So you have to do some kind of first
principles thinking here. It's like, what's the problem? What's the state of things? What's the
right next step? Yes, you have to get really good at context switching, which is one of the hardest
things because especially as we do so many things, if you include all the scientific things we do,
scientific fields we're working in, these are complex fields in themselves. And you have to
sort of keep up to a rest of that. But I enjoy it. I've always been a sort of generalist in a way.
And that's actually what happened with my games career after chess. One of the reasons I stopped
playing chess was because I got into computers, but also I started realizing there were many other
great games out there to play too. So I've always been that way inclined, multidisciplinary. And
there's too many interesting things in the world to spend all your time just on one thing.
So you mentioned Spinoza. Gotta ask the big, ridiculously big question about life. What do
you think is the meaning of this whole thing? Why are we humans here? You've already mentioned that
perhaps the universe created us. Is that why you think we're here? To understand how the
universe works? Yeah, I think my answer to that would be, and at least the life I'm living, is to
gain knowledge and understand the universe. That's what I think. I can't see any higher purpose
than that if you think back to the classical Greeks, the virtue of gaining knowledge.
I think it's one of the few true virtues is to understand the world around us and the context
and humanity better. And I think if you do that, you become more compassionate and more understanding
yourself and more tolerant. And all these other things may flow from that. And to me,
understanding the nature of reality, that is the biggest question. What is going on here is sometimes
the colloquial way I say. What is really going on here? It's so mysterious. I feel like we're in
some huge puzzle. But the world also seems to be, the universe seems to be structured in a way.
Why is it structured in a way that science is even possible? The scientific method works,
things are repeatable. It feels like it's almost structured in a way to be conducive to gaining
knowledge. And why should computers be even possible? Isn't that amazing that
computational or electronic devices can be possible? And they're made of sand, our most
common element that we have, silicon on the Earth's crust, that could be made of diamond or
something. Then we would have only had one computer. So a lot of things are kind of slightly
suspicious to me. It sure as heck sounds, this puzzle sure as heck sounds like something we
talked about earlier, what it takes to design a game that's really fun to play for prolonged
periods of time. And it does seem like this puzzle, like you mentioned, the more you learn about it,
the more you realize how little you know. So it humbles you, but excites you by the possibility
of learning more. It's one heck of a puzzle we got going on here. So like I mentioned,
of all the people in the world, you're very likely to be the one who creates the AGI system
that achieves human level intelligence and goes beyond it. So if you got a chance,
and very well you could be the person that goes into the room with the system and have a conversation,
maybe you only get to ask one question. If you do, what question would you ask her?
I would probably ask, what is the true nature of reality? I think that's the question. I don't
know if I'd understand the answer because maybe it would be 42 or something like that. But that's
the question I would ask. And then there'll be a deep sigh from the systems like, all right,
how do I explain to this human? All right, let me, I don't have time to explain. Maybe I'll
draw you a picture. I mean, how do you even begin to answer that question?
Well, I think it would... What would you think the answer could possibly look like?
I think it could start looking like more fundamental explanations of physics would
be the beginning. More careful specification of that taking walking us through by the hand as to
what one would do to maybe prove those things out. Maybe giving you glimpses of what things you
totally missed in the physics of today. Exactly. Here's glimpses of, no, like there's a much
more elaborate world or a much simpler world or something. A much deeper, maybe simpler explanation
of things, right? Then the standard model of physics, which we know doesn't work, but we still
keep adding to. And that's how I think the beginning of an explanation would look. And it
would start encompassing many of the mysteries that we have wandered about for thousands of years,
like consciousness, dreaming, life, and gravity, all of these things.
Yeah, giving us a glimpses of explanations for those things. Well,
Demis, you're one of the special human beings in this giant puzzle of ours. And it's a huge honor
that you would take a pause from the bigger puzzle to solve this small puzzle of a conversation with
me today. It's truly an honor and a pleasure. Thank you so much. I really enjoyed it. Thanks,
Lex. Thanks for listening to this conversation with Demis and Sabis. To support this podcast,
please check out our sponsors in the description. And now let me leave you with some words from
Ezker Dijkstra. Computer science is no more about computers than astronomy is about telescopes.
Thank you for listening and hope to see you next time.