back to indexDemis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity | Lex Fridman Podcast #299
link |
The following is a conversation with Demis Hassabis,
link |
CEO and co founder of DeepMind,
link |
a company that has published and built
link |
some of the most incredible artificial intelligence systems
link |
in the history of computing,
link |
including AlphaZero that learned all by itself
link |
to play the game of go better than any human in the world
link |
and AlphaFold2 that solved protein folding.
link |
Both tasks considered nearly impossible
link |
for a very long time.
link |
Demis is widely considered to be
link |
one of the most brilliant and impactful humans
link |
in the history of artificial intelligence
link |
and science and engineering in general.
link |
This was truly an honor and a pleasure for me
link |
to finally sit down with him for this conversation.
link |
And I'm sure we will talk many times again in the future.
link |
This is the Lux Readman podcast.
link |
To support it, please check out our sponsors
link |
in the description.
link |
And now, dear friends, here's Demis Hassabis.
link |
Let's start with a bit of a personal question.
link |
Am I an AI program you wrote to interview people
link |
until I get good enough to interview you?
link |
Well, I'd be impressed if you were.
link |
I'd be impressed by myself if you were.
link |
I don't think we're quite up to that yet,
link |
but maybe you're from the future, Lex.
link |
If you did, would you tell me?
link |
Is that a good thing to tell a language model
link |
that's tasked with interviewing
link |
that it is, in fact, AI?
link |
Maybe we're in a kind of meta Turing test.
link |
Probably it would be a good idea not to tell you,
link |
so it doesn't change your behavior, right?
link |
This is a kind of link.
link |
Heisenberg uncertainty principle situation.
link |
If I told you, you'd behave differently.
link |
Maybe that's what's happening with us, of course.
link |
This is a benchmark from the future
link |
where they replay 2022 as a year
link |
before AIs were good enough yet,
link |
and now we want to see, is it gonna pass?
link |
If I was such a program,
link |
would you be able to tell, do you think?
link |
So to the Turing test question,
link |
you've talked about the benchmark for solving intelligence.
link |
What would be the impressive thing?
link |
You've talked about winning a Nobel Prize
link |
and AIS system winning a Nobel Prize,
link |
but I still return to the Turing test as a compelling test,
link |
the spirit of the Turing test as a compelling test.
link |
Yeah, the Turing test, of course,
link |
it's been unbelievably influential,
link |
and Turing's one of my all time heroes,
link |
but I think if you look back at the 1950 paper,
link |
his original paper and read the original,
link |
you'll see, I don't think he meant it
link |
to be a rigorous formal test.
link |
I think it was more like a thought experiment,
link |
almost a bit of philosophy he was writing
link |
if you look at the style of the paper,
link |
and you can see he didn't specify it very rigorously.
link |
So for example, he didn't specify the knowledge
link |
that the expert or judge would have.
link |
How much time would they have to investigate this?
link |
So these are important parameters
link |
if you were gonna make it a true sort of formal test.
link |
And by some measures, people claim the Turing test passed
link |
several, a decade ago, I remember someone claiming that
link |
with a kind of very bog standard, normal logic model,
link |
because they pretended it was a kid.
link |
So the judges thought that the machine was a child.
link |
So that would be very different
link |
from an expert AI person interrogating a machine
link |
and knowing how it was built and so on.
link |
So I think we should probably move away from that
link |
as a formal test and move more towards a general test
link |
where we test the AI capabilities on a range of tasks
link |
and see if it reaches human level or above performance
link |
on maybe thousands, perhaps even millions of tasks
link |
eventually and cover the entire sort of cognitive space.
link |
So I think for its time,
link |
it was an amazing thought experiment.
link |
And also 1950s, obviously there's barely the dawn
link |
of the computer age.
link |
So of course he only thought about text
link |
and now we have a lot more different inputs.
link |
So yeah, maybe the better thing to test
link |
is the generalizability, so across multiple tasks.
link |
But I think it's also possible as systems like Gato show
link |
that eventually that might map right back to language.
link |
So you might be able to demonstrate your ability
link |
to generalize across tasks by then communicating
link |
your ability to generalize across tasks,
link |
which is kind of what we do through conversation anyway
link |
when we jump around.
link |
Ultimately what's in there in that conversation
link |
is not just you moving around knowledge,
link |
it's you moving around like these entirely different
link |
modalities of understanding that ultimately map
link |
to your ability to operate successfully
link |
in all of these domains, which you can think of as tasks.
link |
Yeah, I think certainly we as humans use language
link |
as our main generalization communication tool.
link |
So I think we end up thinking in language
link |
and expressing our solutions in language.
link |
So it's going to be a very powerful mode in which
link |
to explain the system, to explain what it's doing.
link |
But I don't think it's the only modality that matters.
link |
So I think there's going to be a lot of different ways
link |
to express capabilities other than just language.
link |
Yeah, visual, robotics, body language,
link |
yeah, actions, the interactive aspect of all that.
link |
That's all part of it.
link |
But what's interesting with Gato is that
link |
it's sort of pushing prediction to the maximum
link |
in terms of like mapping arbitrary sequences
link |
to other sequences and sort of just predicting
link |
what's going to happen next.
link |
So prediction seems to be fundamental to intelligence.
link |
And what you're predicting doesn't so much matter.
link |
Yeah, it seems like you can generalize that quite well.
link |
So obviously language models predict the next word,
link |
Gato predicts potentially any action or any token.
link |
And it's just the beginning really.
link |
It's our most general agent one could call it so far,
link |
but that itself can be scaled up massively more
link |
than we've done so far.
link |
And obviously we're in the middle of doing that.
link |
But the big part of solving AGI is creating benchmarks
link |
that help us get closer and closer,
link |
sort of creating benchmarks that test the generalizability.
link |
And it's just still interesting that this fella,
link |
Alan Turing, was one of the first
link |
and probably still one of the only people
link |
that was trying, maybe philosophically,
link |
but was trying to formulate a benchmark
link |
that could be followed.
link |
It is, even though it's fuzzy,
link |
it's still sufficiently rigorous
link |
to where you can run that test.
link |
And I still think something like the Turing test
link |
will, at the end of the day,
link |
be the thing that truly impresses other humans
link |
so that you can have a close friend who's an AI system.
link |
And for that friend to be a good friend,
link |
they're going to have to be able to play StarCraft
link |
and they're gonna have to do all of these tasks,
link |
get you a beer, so the robotics tasks,
link |
play games with you, use language,
link |
humor, all of those kinds of things.
link |
But that ultimately can boil down to language.
link |
It feels like, not in terms of the AI community,
link |
but in terms of the actual impact
link |
of general intelligence on the world,
link |
it feels like language will be the place
link |
where it truly shines.
link |
I think so, because it's such an important
link |
kind of input output for us.
link |
I think you're right.
link |
I think the Turing test,
link |
the kind of the philosophy behind it,
link |
which is the idea of can a machine mimic
link |
the behaviors of a human?
link |
And I would say wider than just language and text,
link |
then in terms of actions and everything else,
link |
creativity, all these things,
link |
then if it can sort of match or exceed
link |
human cognitive capabilities,
link |
then I think we have a true intelligence.
link |
So from that perspective, you're right.
link |
I think he did formulate the right kind of setup.
link |
I just, I think there'll be a kind of humor
link |
in the AI systems of the future looking back
link |
to this conversation and thinking about the Turing test
link |
and also thinking about by that time,
link |
they would know which year they were finally able
link |
to sort of cross the threshold of human level intelligence
link |
and think how funny it is that we humans
link |
were still confused about this whole problem
link |
that has been long ago been solved.
link |
Well, hello to the AI agents of the future.
link |
Anyway, so going back to your journey,
link |
when did you fall in love with programming first?
link |
Well, I was pretty young age actually.
link |
So, I started off, actually games was my first love.
link |
So starting to play chess when I was around four years old
link |
and then it was actually with winnings
link |
from a chess competition that I managed
link |
to buy my first chess computer
link |
when I was about eight years old.
link |
It was a ZX Spectrum, which was hugely popular
link |
in the UK at the time.
link |
And it was amazing machine because I think it trained
link |
a whole generation of programmers in the UK
link |
because it was so accessible.
link |
You know, you literally switched it on
link |
and there was the basic prompt
link |
and you could just get going.
link |
And my parents didn't really know anything about computers.
link |
So, but because it was my money from a chess competition,
link |
I could say I wanted to buy it.
link |
And then, you know, I just went to bookstores,
link |
got books on programming and started typing in,
link |
you know, the programming code.
link |
And then of course, once you start doing that,
link |
you start adjusting it and then making your own games.
link |
And that's when I fell in love with computers
link |
and realized that they were a very magical device.
link |
In a way, I kind of, I wouldn't have been able
link |
to explain this at the time,
link |
but I felt that they were sort of almost
link |
a magical extension of your mind.
link |
I always had this feeling and I've always loved this
link |
about computers that you can set them off doing something,
link |
some task for you, you can go to sleep,
link |
come back the next day and it's solved.
link |
You know, that feels magical to me.
link |
So, I mean, all machines do that to some extent.
link |
They all enhance our natural capabilities.
link |
Obviously cars make us, allow us to move faster
link |
than we can run, but this was a machine to extend the mind.
link |
And then of course, AI is the ultimate expression
link |
of what a machine may be able to do or learn.
link |
So very naturally for me, that thought extended
link |
into AI quite quickly.
link |
Do you remember the programming language
link |
that was first started and was it special to the machine?
link |
No, I think it was just basic on the ZX Spectrum.
link |
I don't know what specific form it was.
link |
And then later on I got a Commodore Amiga,
link |
which was a fantastic machine.
link |
Now you're just showing off.
link |
So yeah, well, lots of my friends had Atari STs
link |
and I managed to get Amigas, it was a bit more powerful
link |
and that was incredible and used to do programming
link |
in assembler and also Amos basic,
link |
this specific form of basic, it was incredible actually.
link |
So I learned all my coding skills.
link |
And when did you fall in love with AI?
link |
So when did you first start to gain an understanding
link |
that you can not just write programs
link |
that do some mathematical operations for you
link |
while you sleep, but something that's akin
link |
to bringing an entity to life,
link |
sort of a thing that can figure out something
link |
more complicated than a simple mathematical operation.
link |
Yeah, so there was a few stages for me
link |
all while I was very young.
link |
So first of all, as I was trying to improve
link |
at playing chess, I was captaining
link |
various England junior chess teams.
link |
And at the time when I was about maybe 10, 11 years old,
link |
I was gonna become a professional chess player.
link |
That was my first thought.
link |
So that dream was there to try to get
link |
to the highest levels of chess.
link |
Yeah, so when I was about 12 years old,
link |
I got to master standard and I was second highest rated
link |
player in the world to Judith Polgar,
link |
who obviously ended up being an amazing chess player
link |
and a world women's champion.
link |
And when I was trying to improve at chess,
link |
where what you do is you obviously, first of all,
link |
you're trying to improve your own thinking processes.
link |
So that leads you to thinking about thinking,
link |
how is your brain coming up with these ideas?
link |
Why is it making mistakes?
link |
How can you improve that thought process?
link |
But the second thing is that you,
link |
it was just the beginning, this was like in the early 80s,
link |
mid 80s of chess computers.
link |
If you remember, they were physical balls
link |
like the one we have in front of us.
link |
And you press down the squares.
link |
And I think Kasparov had a branded version of it
link |
And you used to, they're not as strong as they are today,
link |
but they were pretty strong and you used to practice
link |
against them to try and improve your openings
link |
And so I remember, I think I probably got my first one,
link |
I was around 11 or 12.
link |
And I remember thinking, this is amazing,
link |
how has someone programmed this chess board to play chess?
link |
And it was very formative book I bought,
link |
which was called The Chess Computer Handbook
link |
This thing came out in 1984 or something.
link |
So I must've got it when I was about 11, 12.
link |
And it explained fully how these chess programs were made.
link |
And I remember my first AI program
link |
being programming my Amiga.
link |
It couldn't, it wasn't powerful enough to play chess.
link |
I couldn't write a whole chess program,
link |
but I wrote a program for it to play Othello or reverse it,
link |
sometimes called I think in the US.
link |
And so a slightly simpler game than chess,
link |
but I used all of the principles that chess programs had,
link |
alpha, beta, search, all of that.
link |
And that was my first AI program.
link |
I remember that very well, I was around 12 years old.
link |
So that brought me into AI.
link |
And then the second part was later on,
link |
when I was around 16, 17,
link |
and I was writing games professionally, designing games,
link |
writing a game called Theme Park,
link |
which had AI as a core gameplay component
link |
as part of the simulation.
link |
And it sold millions of copies around the world.
link |
And people loved the way that the AI,
link |
even though it was relatively simple
link |
by today's AI standards,
link |
was reacting to the way you as the player played it.
link |
So it was called a sandbox game.
link |
So it was one of the first types of games like that,
link |
along with SimCity.
link |
And it meant that every game you played was unique.
link |
Is there something you could say just on a small tangent
link |
about really impressive AI
link |
from a game design, human enjoyment perspective,
link |
really impressive AI that you've seen in games
link |
and maybe what does it take to create an AI system?
link |
And how hard of a problem is that?
link |
So a million questions just as a brief tangent.
link |
Well, look, I think games have been significant in my life
link |
for three reasons.
link |
So first of all, I was playing them
link |
and training myself on games when I was a kid.
link |
Then I went through a phase of designing games
link |
and writing AI for games.
link |
So all the games I professionally wrote
link |
had AI as a core component.
link |
And that was mostly in the 90s.
link |
And the reason I was doing that in games industry
link |
was at the time the games industry,
link |
I think was the cutting edge of technology.
link |
So whether it was graphics with people like John Carmack
link |
and Quake and those kinds of things or AI,
link |
I think actually all the action was going on in games.
link |
And we're still reaping the benefits of that
link |
even with things like GPUs, which I find ironic
link |
was obviously invented for graphics, computer graphics,
link |
but then turns out to be amazingly useful for AI.
link |
It just turns out everything's a matrix multiplication
link |
it appears in the whole world.
link |
So I think games at the time had the most cutting edge AI.
link |
And a lot of the games, I was involved in writing.
link |
So there was a game called Black and White,
link |
which was one game I was involved with
link |
in the early stages of,
link |
which I still think is the most impressive example
link |
of reinforcement learning in a computer game.
link |
So in that game, you trained a little pet animal and...
link |
It's a brilliant game.
link |
And it sort of learned from how you were treating it.
link |
So if you treated it badly, then it became mean.
link |
And then it would be mean to your villagers
link |
and your population, the sort of the little tribe
link |
that you were running.
link |
But if you were kind to it, then it would be kind.
link |
And people were fascinated by how that works.
link |
And so was I to be honest with the way it kind of developed.
link |
Especially the mapping to good and evil.
link |
Made you realize, made me realize that you can sort of
link |
in the choices you make can define where you end up.
link |
And that means all of us are capable of the good, evil.
link |
It all matters in the different choices
link |
along the trajectory to those places that you make.
link |
I mean, games can do that philosophically to you.
link |
Well, games are, I think, a unique medium
link |
because you as the player,
link |
you're not just passively consuming the entertainment,
link |
You're actually actively involved as an agent.
link |
So I think that's what makes it in some ways
link |
can be more visceral than other mediums
link |
like films and books.
link |
So the second, so that was designing AI in games.
link |
And then the third use we've used of AI
link |
is in DeepMind from the beginning,
link |
which is using games as a testing ground
link |
for proving out AI algorithms and developing AI algorithms.
link |
And that was a sort of a core component
link |
of our vision at the start of DeepMind
link |
was that we would use games very heavily
link |
as our main testing ground, certainly to begin with,
link |
because it's super efficient to use games.
link |
And also, it's very easy to have metrics
link |
to see how well your systems are improving
link |
and what direction your ideas are going in
link |
and whether you're making incremental improvements.
link |
And because those games are often rooted
link |
in something that humans did for a long time beforehand,
link |
there's already a strong set of rules.
link |
Like it's already a damn good benchmark.
link |
Yes, it's really good for so many reasons
link |
because you've got clear measures
link |
of how good humans can be at these things.
link |
And in some cases like Go,
link |
we've been playing it for thousands of years
link |
and often they have scores or at least win conditions.
link |
So it's very easy for reward learning systems
link |
It's very easy to specify what that reward is.
link |
And also at the end, it's easy to test externally
link |
at how strong is your system by of course,
link |
playing against the world's strongest players at those games.
link |
So it's so good for so many reasons
link |
and it's also very efficient to run potentially millions
link |
of simulations in parallel on the cloud.
link |
So I think there's a huge reason why we were so successful
link |
back in starting out 2010,
link |
how come we were able to progress so quickly
link |
because we've utilized games.
link |
And at the beginning of DeepMind,
link |
we also hired some amazing game engineers
link |
who I knew from my previous lives in the games industry.
link |
And that helped to bootstrap us very quickly.
link |
And plus it's somehow super compelling
link |
almost at a philosophical level of man versus machine
link |
over a chess board or a Go board.
link |
And especially given that the entire history of AI
link |
is defined by people saying it's gonna be impossible
link |
to make a machine that beats a human being in chess.
link |
And then once that happened,
link |
people were certain when I was coming up in AI
link |
that Go is not a game that can be solved
link |
because of the combinatorial complexity is just too,
link |
it's no matter how much Moore's law you have,
link |
compute is just never going to be able
link |
to crack the game of Go.
link |
And so then there's something compelling about facing,
link |
sort of taking on the impossibility of that task
link |
from the AI researcher perspective,
link |
engineer perspective, and then as a human being,
link |
just observing this whole thing.
link |
Your beliefs about what you thought was impossible
link |
being broken apart,
link |
it's humbling to realize we're not as smart as we thought.
link |
It's humbling to realize that the things we think
link |
are impossible now perhaps will be done in the future.
link |
There's something really powerful about a game,
link |
AI system beating human being in a game
link |
that drives that message home
link |
for like millions, billions of people,
link |
especially in the case of Go.
link |
Well, look, I think it's,
link |
I mean, it has been a fascinating journey
link |
and especially as I think about it from,
link |
I can understand it from both sides,
link |
both as the AI, creators of the AI,
link |
but also as a games player originally.
link |
So, it was a really interesting,
link |
I mean, it was a fantastic, but also somewhat
link |
bittersweet moment, the AlphaGo match for me,
link |
seeing that and being obviously heavily involved in that.
link |
But as you say, chess has been the,
link |
I mean, Kasparov, I think rightly called it
link |
the Drosophila of intelligence, right?
link |
So, it's sort of, I love that phrase
link |
and I think he's right because chess has been
link |
hand in hand with AI from the beginning
link |
of the whole field, right?
link |
So, I think every AI practitioner,
link |
starting with Turing and Claude Shannon and all those,
link |
the sort of forefathers of the field,
link |
tried their hand at writing a chess program.
link |
I've got original edition of Claude Shannon's
link |
first chess program, I think it was 1949,
link |
the original sort of paper.
link |
And they all did that and Turing famously wrote
link |
a chess program, but all the computers around them
link |
were obviously too slow to run it.
link |
So, he had to run, he had to be the computer, right?
link |
So, he literally, I think spent two or three days
link |
running his own program by hand with pencil and paper
link |
and playing a friend of his with his chess program.
link |
So, of course, Deep Blue was a huge moment,
link |
beating Kasparov, but actually when that happened,
link |
I remember that very vividly, of course,
link |
because it was chess and computers and AI,
link |
all the things I loved and I was at college at the time.
link |
But I remember coming away from that,
link |
being more impressed by Kasparov's mind
link |
than I was by Deep Blue.
link |
Because here was Kasparov with his human mind,
link |
not only could he play chess more or less
link |
to the same level as this brute of a calculation machine,
link |
but of course, Kasparov can do everything else
link |
humans can do, ride a bike, talk many languages,
link |
do politics, all the rest of the amazing things
link |
that Kasparov does.
link |
And so, with the same brain.
link |
And yet Deep Blue, brilliant as it was at chess,
link |
it'd been hand coded for chess and actually had distilled
link |
the knowledge of chess grandmasters into a cool program,
link |
but it couldn't do anything else.
link |
It couldn't even play a strictly simpler game
link |
So, something to me was missing from intelligence
link |
from that system that we would regard as intelligence.
link |
And I think it was this idea of generality
link |
and also learning.
link |
So, and that's obviously what we tried to do with AlphaGo.
link |
Yeah, with AlphaGo and AlphaZero, MuZero,
link |
and then God and all the things that we'll get into
link |
some parts of, there's just a fascinating trajectory here.
link |
But let's just stick on chess briefly.
link |
On the human side of chess, you've proposed that
link |
from a game design perspective,
link |
the thing that makes chess compelling as a game
link |
is that there's a creative tension between a bishop
link |
Can you explain this?
link |
First of all, it's really interesting to think about
link |
what makes a game compelling,
link |
makes it stick across centuries.
link |
Yeah, I was sort of thinking about this,
link |
and actually a lot of even amazing chess players
link |
don't think about it necessarily
link |
from a game's designer point of view.
link |
So, it's with my game design hat on
link |
that I was thinking about this, why is chess so compelling?
link |
And I think a critical reason is the dynamicness
link |
of the different kind of chess positions you can have,
link |
whether they're closed or open and other things,
link |
comes from the bishop and the knight.
link |
So, if you think about how different
link |
the capabilities of the bishop and knight are
link |
in terms of the way they move,
link |
and then somehow chess has evolved
link |
to balance those two capabilities more or less equally.
link |
So, they're both roughly worth three points each.
link |
So, you think that dynamics is always there
link |
and then the rest of the rules
link |
are kind of trying to stabilize the game.
link |
Well, maybe, I mean, it's sort of,
link |
I don't know, it's chicken and egg situation,
link |
probably both came together.
link |
But the fact that it's got to this beautiful equilibrium
link |
where you can have the bishop and knight
link |
that are so different in power,
link |
but so equal in value across the set
link |
of the universe of all positions, right?
link |
Somehow they've been balanced by humanity
link |
over hundreds of years,
link |
I think gives the game the creative tension
link |
that you can swap the bishop and knights
link |
for a bishop for a knight,
link |
and they're more or less worth the same,
link |
but now you aim for a different type of position.
link |
If you have the knight, you want a closed position.
link |
If you have the bishop, you want an open position.
link |
So, I think that creates
link |
a lot of the creative tension in chess.
link |
So, some kind of controlled creative tension.
link |
From an AI perspective,
link |
do you think AI systems could eventually design games
link |
that are optimally compelling to humans?
link |
Well, that's an interesting question.
link |
Sometimes I get asked about AI and creativity,
link |
and the way I answered that is relevant to that question,
link |
which is that I think there are different levels
link |
of creativity, one could say.
link |
So, I think if we define creativity
link |
as coming up with something original, right,
link |
that's useful for a purpose,
link |
then I think the kind of lowest level of creativity
link |
is like an interpolation.
link |
So, an averaging of all the examples you see.
link |
So, maybe a very basic AI system could say
link |
you could have that.
link |
So, you show it millions of pictures of cats,
link |
and then you say, give me an average looking cat, right?
link |
Generate me an average looking cat.
link |
I would call that interpolation.
link |
Then there's extrapolation,
link |
which something like AlphaGo showed.
link |
So, AlphaGo played millions of games of Go against itself,
link |
and then it came up with brilliant new ideas
link |
like Move 37 in game two, brilliant motif strategies in Go
link |
that no humans had ever thought of,
link |
even though we've played it for thousands of years
link |
and professionally for hundreds of years.
link |
So, that I call that extrapolation,
link |
but then there's still a level above that,
link |
which is, you could call out of the box thinking
link |
or true innovation, which is, could you invent Go, right?
link |
Could you invent chess and not just come up
link |
with a brilliant chess move or brilliant Go move,
link |
but can you actually invent chess
link |
or something as good as chess or Go?
link |
And I think one day AI could, but then what's missing
link |
is how would you even specify that task
link |
to a program right now?
link |
And the way I would do it if I was telling a human to do it
link |
or a human games designer to do it is I would say,
link |
something like Go, I would say, come up with a game
link |
that only takes five minutes to learn,
link |
which Go does because it's got simple rules,
link |
but many lifetimes to master, right?
link |
Or impossible to master in one lifetime
link |
because it's so deep and so complex.
link |
And then it's aesthetically beautiful.
link |
And also it can be completed in three or four hours
link |
of gameplay time, which is useful for us in a human day.
link |
And so you might specify these sort of high level concepts
link |
like that, and then with that
link |
and then maybe a few other things,
link |
one could imagine that Go satisfies those constraints.
link |
But the problem is that we're not able
link |
to specify abstract notions like that,
link |
high level abstract notions like that yet to our AI systems.
link |
And I think there's still something missing there
link |
in terms of high level concepts or abstractions
link |
that they truly understand
link |
and they're combinable and compositional.
link |
So for the moment, I think AI is capable
link |
of doing interpolation and extrapolation,
link |
but not true invention.
link |
So coming up with rule sets and optimizing
link |
with complicated objectives around those rule sets,
link |
we can't currently do.
link |
But you could take a specific rule set
link |
and then run a kind of self play experiment
link |
to see how long, just observe how an AI system
link |
from scratch learns, how long is that journey of learning?
link |
And maybe if it satisfies some of those other things
link |
you mentioned in terms of quickness to learn and so on,
link |
and you could see a long journey to master
link |
for even an AI system, then you could say
link |
that this is a promising game.
link |
But it would be nice to do almost like AlphaCode
link |
so programming rules.
link |
So generating rules that automate even that part
link |
of the generation of rules.
link |
So I have thought about systems actually
link |
that I think would be amazing for a games designer.
link |
If you could have a system that takes your game,
link |
plays it tens of millions of times, maybe overnight,
link |
and then self balances the rules better.
link |
So it tweaks the rules and maybe the equations
link |
and the parameters so that the game is more balanced,
link |
the units in the game or some of the rules could be tweaked.
link |
So it's a bit of like giving a base set
link |
and then allowing Monte Carlo Tree Search
link |
or something like that to sort of explore it.
link |
And I think that would be super powerful tool actually
link |
for balancing, auto balancing a game,
link |
which usually takes thousands of hours
link |
from hundreds of human games testers normally
link |
to balance a game like StarCraft,
link |
which is Blizzard are amazing at balancing their games,
link |
but it takes them years and years and years.
link |
So one could imagine at some point
link |
when this stuff becomes efficient enough
link |
to you might be able to do that like overnight.
link |
Do you think a game that is optimal designed by an AI system
link |
would look very much like a planet earth?
link |
Maybe, maybe it's only the sort of game
link |
I would love to make is, and I've tried in my games career,
link |
the games design career, my first big game
link |
was designing a theme park, an amusement park.
link |
Then with games like Republic, I tried to have games
link |
where we designed whole cities and allowed you to play in.
link |
So, and of course people like Will Wright
link |
have written games like SimEarth,
link |
trying to simulate the whole of earth, pretty tricky,
link |
SimEarth, I haven't actually played that one.
link |
Does it incorporate of evolution or?
link |
Yeah, it has evolution and it sort of tries to,
link |
it sort of treats it as an entire biosphere,
link |
but from quite high level.
link |
It'd be nice to be able to sort of zoom in,
link |
zoom out and zoom in.
link |
So obviously it couldn't do, that was in the 90s.
link |
I think he wrote that in the 90s.
link |
So it couldn't, it wasn't able to do that,
link |
but that would be obviously the ultimate sandbox game.
link |
On that topic, do you think we're living in a simulation?
link |
Yes, well, so, okay.
link |
We're gonna jump around from the absurdly philosophical
link |
Sure, sure, very, very happy to.
link |
So I think my answer to that question
link |
is a little bit complex because there is simulation theory,
link |
which obviously Nick Bostrom,
link |
I think famously first proposed.
link |
And I don't quite believe it in that sense.
link |
So in the sense that are we in some sort of computer game
link |
or have our descendants somehow recreated earth
link |
in the 21st century and some,
link |
for some kind of experimental reason.
link |
I think that, but I do think that we,
link |
that we might be, that the best way to understand physics
link |
and the universe is from a computational perspective.
link |
So understanding it as an information universe
link |
and actually information being the most fundamental unit
link |
of reality rather than matter or energy.
link |
So a physicist would say, you know, matter or energy,
link |
you know, E equals MC squared.
link |
These are the things that are the fundamentals
link |
I'd actually say information,
link |
which of course itself can be,
link |
can specify energy or matter, right?
link |
Matter is actually just, you know,
link |
we're just out the way our bodies
link |
and the molecules in our body are arranged as information.
link |
So I think information may be the most fundamental way
link |
to describe the universe.
link |
And therefore you could say we're in some sort of simulation
link |
But I don't, I do, I'm not,
link |
I'm not really a subscriber to the idea that, you know,
link |
these are sort of throw away billions of simulations around.
link |
I think this is actually very critical and possibly unique,
link |
This particular one.
link |
And you just mean treating the universe as a computer
link |
that's processing and modifying information
link |
is a good way to solve the problems of physics,
link |
of chemistry, of biology,
link |
and perhaps of humanity and so on.
link |
Yes, I think understanding physics
link |
in terms of information theory
link |
might be the best way to really understand
link |
what's going on here.
link |
From our understanding of a universal Turing machine,
link |
from our understanding of a computer,
link |
do you think there's something outside
link |
of the capabilities of a computer
link |
that is present in our universe?
link |
You have a disagreement with Roger Penrose
link |
about the nature of consciousness.
link |
He thinks that consciousness is more
link |
than just a computation.
link |
Do you think all of it, the whole shebangs,
link |
can be a computation?
link |
Yeah, I've had many fascinating debates
link |
with Sir Roger Penrose,
link |
and obviously he's famously,
link |
and I read, you know, Emperors of the New Mind
link |
and his books, his classical books,
link |
and they were pretty influential in the 90s.
link |
And he believes that there's something more,
link |
something quantum that is needed
link |
to explain consciousness in the brain.
link |
I think about what we're doing actually at DeepMind
link |
and what my career is being,
link |
we're almost like Turing's champion.
link |
So we are pushing Turing machines or classical computation
link |
What are the limits of what classical computing can do?
link |
Now, and at the same time,
link |
I've also studied neuroscience to see,
link |
and that's why I did my PhD in,
link |
was to see, also to look at, you know,
link |
is there anything quantum in the brain
link |
from a neuroscience or biological perspective?
link |
And so far, I think most neuroscientists
link |
and most mainstream biologists and neuroscientists
link |
would say there's no evidence of any quantum systems
link |
or effects in the brain.
link |
As far as we can see, it can be mostly explained
link |
by classical theories.
link |
So, and then, so there's sort of the search
link |
from the biology side.
link |
And then at the same time,
link |
there's the raising of the water, the bar,
link |
from what classical Turing machines can do.
link |
And, you know, including our new AI systems.
link |
And as you alluded to earlier, you know,
link |
I think AI, especially in the last decade plus,
link |
has been a continual story now of surprising events
link |
and surprising successes,
link |
knocking over one theory after another
link |
of what was thought to be impossible, you know,
link |
from Go to protein folding and so on.
link |
And so I think I would be very hesitant
link |
to bet against how far the universal Turing machine
link |
and classical computation paradigm can go.
link |
And my betting would be that all of,
link |
certainly what's going on in our brain,
link |
can probably be mimicked or approximated
link |
on a classical machine,
link |
not requiring something metaphysical or quantum.
link |
And we'll get there with some of the work with AlphaFold,
link |
which I think begins the journey of modeling
link |
this beautiful and complex world of biology.
link |
So you think all the magic of the human mind
link |
comes from this, just a few pounds of mush,
link |
of biological computational mush,
link |
that's akin to some of the neural networks,
link |
not directly, but in spirit
link |
that DeepMind has been working with.
link |
Well, look, I think it's, you say it's a few, you know,
link |
of course it's, this is the,
link |
I think the biggest miracle of the universe
link |
is that it is just a few pounds of mush in our skulls.
link |
And yet it's also our brains are the most complex objects
link |
that we know of in the universe.
link |
So there's something profoundly beautiful
link |
and amazing about our brains.
link |
And I think that it's an incredibly,
link |
incredible efficient machine.
link |
And it's, you know, phenomenon basically.
link |
And I think that building AI,
link |
one of the reasons I wanna build AI,
link |
and I've always wanted to is,
link |
I think by building an intelligent artifact like AI,
link |
and then comparing it to the human mind,
link |
that will help us unlock the uniqueness
link |
and the true secrets of the mind
link |
that we've always wondered about since the dawn of history,
link |
like consciousness, dreaming, creativity, emotions,
link |
what are all these things, right?
link |
We've wondered about them since the dawn of humanity.
link |
And I think one of the reasons,
link |
and, you know, I love philosophy and philosophy of mind is,
link |
we found it difficult is there haven't been the tools
link |
for us to really, other than introspection,
link |
from very clever people in history,
link |
very clever philosophers,
link |
to really investigate this scientifically.
link |
But now suddenly we have a plethora of tools.
link |
Firstly, we have all of the neuroscience tools,
link |
fMRI machines, single cell recording, all of this stuff,
link |
but we also have the ability, computers and AI,
link |
to build intelligent systems.
link |
So I think that, you know,
link |
I think it is amazing what the human mind does.
link |
And I'm kind of in awe of it really.
link |
And I think it's amazing that with our human minds,
link |
we're able to build things like computers
link |
and actually even, you know,
link |
think and investigate about these questions.
link |
I think that's also a testament to the human mind.
link |
The universe built the human mind
link |
that now is building computers that help us understand
link |
both the universe and our own human mind.
link |
This is actually it.
link |
I mean, I think that's one, you know,
link |
one could say we are,
link |
maybe we're the mechanism by which the universe
link |
is going to try and understand itself.
link |
So let's go to the basic building blocks of biology
link |
that I think is another angle at which you can start
link |
to understand the human mind, the human body,
link |
which is quite fascinating,
link |
which is from the basic building blocks,
link |
start to simulate, start to model
link |
how from those building blocks,
link |
you can construct bigger and bigger, more complex systems,
link |
maybe one day the entirety of the human biology.
link |
So here's another problem that thought
link |
to be impossible to solve, which is protein folding.
link |
And Alpha Fold or specifically Alpha Fold 2 did just that.
link |
It solved protein folding.
link |
I think it's one of the biggest breakthroughs,
link |
certainly in the history of structural biology,
link |
but in general in science,
link |
maybe from a high level, what is it and how does it work?
link |
And then we can ask some fascinating questions after.
link |
So maybe to explain it to people not familiar
link |
with protein folding is, you know,
link |
first of all, explain proteins, which is, you know,
link |
proteins are essential to all life.
link |
Every function in your body depends on proteins.
link |
Sometimes they're called the workhorses of biology.
link |
And if you look into them and I've, you know,
link |
obviously as part of Alpha Fold,
link |
I've been researching proteins and structural biology
link |
for the last few years, you know,
link |
they're amazing little bio nano machines proteins.
link |
They're incredible if you actually watch little videos
link |
of how they work, animations of how they work.
link |
And proteins are specified by their genetic sequence
link |
called the amino acid sequence.
link |
So you can think of it as their genetic makeup.
link |
And then in the body in nature,
link |
they fold up into a 3D structure.
link |
So you can think of it as a string of beads
link |
and then they fold up into a ball.
link |
Now, the key thing is you want to know
link |
what that 3D structure is because the structure,
link |
the 3D structure of a protein is what helps to determine
link |
what does it do, the function it does in your body.
link |
And also if you're interested in drugs or disease,
link |
you need to understand that 3D structure
link |
because if you want to target something
link |
with a drug compound about to block something
link |
the protein's doing, you need to understand
link |
where it's gonna bind on the surface of the protein.
link |
So obviously in order to do that,
link |
you need to understand the 3D structure.
link |
So the structure is mapped to the function.
link |
The structure is mapped to the function
link |
and the structure is obviously somehow specified
link |
by the amino acid sequence.
link |
And that's the, in essence, the protein folding problem is,
link |
can you just from the amino acid sequence,
link |
the one dimensional string of letters,
link |
can you immediately computationally predict
link |
And this has been a grand challenge in biology
link |
for over 50 years.
link |
So I think it was first articulated by Christian Anfinsen,
link |
a Nobel prize winner in 1972,
link |
as part of his Nobel prize winning lecture.
link |
And he just speculated this should be possible
link |
to go from the amino acid sequence to the 3D structure,
link |
but he didn't say how.
link |
So it's been described to me as equivalent
link |
to Fermat's last theorem, but for biology.
link |
You should, as somebody that very well might win
link |
the Nobel prize in the future.
link |
But outside of that, you should do more
link |
of that kind of thing.
link |
In the margin, just put random things
link |
that will take like 200 years to solve.
link |
Set people off for 200 years.
link |
It should be possible.
link |
And just don't give any details.
link |
I think everyone exactly should be,
link |
I'll have to remember that for future.
link |
So yeah, so he set off, you know,
link |
with this one throwaway remark, just like Fermat,
link |
you know, he set off this whole 50 year field really
link |
of computational biology.
link |
And they had, you know, they got stuck.
link |
They hadn't really got very far with doing this.
link |
And until now, until AlphaFold came along,
link |
this is done experimentally, right?
link |
Very painstakingly.
link |
So the rule of thumb is, and you have to like
link |
crystallize the protein, which is really difficult.
link |
Some proteins can't be crystallized like membrane proteins.
link |
And then you have to use very expensive electron microscopes
link |
or X ray crystallography machines.
link |
Really painstaking work to get the 3D structure
link |
and visualize the 3D structure.
link |
So the rule of thumb in experimental biology
link |
is that it takes one PhD student,
link |
their entire PhD to do one protein.
link |
And with AlphaFold 2, we were able to predict
link |
the 3D structure in a matter of seconds.
link |
And so we were, you know, over Christmas,
link |
we did the whole human proteome
link |
or every protein in the human body or 20,000 proteins.
link |
So the human proteomes like the equivalent
link |
of the human genome, but on protein space.
link |
And sort of revolutionized really
link |
what a structural biologist can do.
link |
Because now they don't have to worry
link |
about these painstaking experimental,
link |
should they put all of that effort in or not?
link |
They can almost just look up the structure
link |
of their proteins like a Google search.
link |
And so there's a data set on which it's trained
link |
and how to map this amino acid sequence.
link |
First of all, it's incredible that a protein,
link |
this little chemical computer is able to do
link |
that computation itself in some kind of distributed way
link |
and do it very quickly.
link |
That's a weird thing.
link |
And they evolve that way because, you know,
link |
in the beginning, I mean, that's a great invention,
link |
just the protein itself.
link |
And then there's, I think, probably a history
link |
of like they evolved to have many of these proteins
link |
and those proteins figure out how to be computers themselves
link |
in such a way that you can create structures
link |
that can interact in complexes with each other
link |
in order to form high level functions.
link |
I mean, it's a weird system that they figured it out.
link |
I mean, you know, maybe we should talk
link |
about the origins of life too,
link |
but proteins themselves, I think are magical
link |
and incredible, as I said, little bio nano machines.
link |
And actually Leventhal, who was another scientist,
link |
a contemporary of Amphinson, he coined this Leventhal,
link |
what became known as Leventhal's paradox,
link |
which is exactly what you're saying.
link |
He calculated roughly an average protein,
link |
which is maybe 2000 amino acids base as long,
link |
is can fold in maybe 10 to the power 300
link |
different confirmations.
link |
So there's 10 to the power 300 different ways
link |
that protein could fold up.
link |
And yet somehow in nature, physics solves this,
link |
solves this in a matter of milliseconds.
link |
So proteins fold up in your body in, you know,
link |
sometimes in fractions of a second.
link |
So physics is somehow solving that search problem.
link |
And just to be clear, in many of these cases,
link |
maybe you can correct me if I'm wrong,
link |
there's often a unique way for that sequence to form itself.
link |
So among that huge number of possibilities,
link |
it figures out a way how to stably,
link |
in some cases there might be a misfunction, so on,
link |
which leads to a lot of the disorders and stuff like that.
link |
But most of the time it's a unique mapping
link |
and that unique mapping is not obvious.
link |
Which is what the problem is.
link |
Exactly, so there's a unique mapping usually in a healthy,
link |
if it's healthy, and as you say in disease,
link |
so for example, Alzheimer's,
link |
one conjecture is that it's because of misfolded protein,
link |
a protein that folds in the wrong way, amyloid beta protein.
link |
So, and then because it folds in the wrong way,
link |
it gets tangled up, right, in your neurons.
link |
So it's super important to understand
link |
both healthy functioning and also disease
link |
is to understand, you know, what these things are doing
link |
and how they're structuring.
link |
Of course, the next step is sometimes proteins change shape
link |
when they interact with something.
link |
So they're not just static necessarily in biology.
link |
Maybe you can give some interesting,
link |
so beautiful things to you about these early days
link |
of AlphaFold, of solving this problem,
link |
because unlike games, this is real physical systems
link |
that are less amenable to self play type of mechanisms.
link |
The size of the data set is smaller
link |
than you might otherwise like,
link |
so you have to be very clever about certain things.
link |
Is there something you could speak to
link |
what was very hard to solve
link |
and what are some beautiful aspects about the solution?
link |
Yeah, I would say AlphaFold is the most complex
link |
and also probably most meaningful system
link |
we've built so far.
link |
So it's been an amazing time actually in the last,
link |
you know, two, three years to see that come through
link |
because as we talked about earlier, you know,
link |
games is what we started on
link |
building things like AlphaGo and AlphaZero,
link |
but really the ultimate goal was to,
link |
not just to crack games,
link |
it was just to build,
link |
use them to bootstrap general learning systems
link |
we could then apply to real world challenges.
link |
Specifically, my passion is scientific challenges
link |
like protein folding.
link |
And then AlphaFold of course
link |
is our first big proof point of that.
link |
And so, you know, in terms of the data
link |
and the amount of innovations that had to go into it,
link |
we, you know, it was like
link |
more than 30 different component algorithms
link |
needed to be put together to crack the protein folding.
link |
I think some of the big innovations were that
link |
kind of building in some hard coded constraints
link |
around physics and evolutionary biology
link |
to constrain sort of things like the bond angles
link |
in the protein and things like that,
link |
a lot, but not to impact the learning system.
link |
So still allowing the system to be able to learn
link |
the physics itself from the examples that we had.
link |
And the examples, as you say,
link |
there are only about 150,000 proteins,
link |
even after 40 years of experimental biology,
link |
only around 150,000 proteins have been,
link |
the structures have been found out about.
link |
So that was our training set,
link |
which is much less than normally we would like to use,
link |
but using various tricks, things like self distillation.
link |
So actually using AlphaFold predictions,
link |
some of the best predictions
link |
that it thought was highly confident in,
link |
we put them back into the training set, right?
link |
To make the training set bigger,
link |
that was critical to AlphaFold working.
link |
So there was actually a huge number
link |
of different innovations like that,
link |
that were required to ultimately crack the problem.
link |
AlphaFold one, what it produced was a distrogram.
link |
So a kind of a matrix of the pairwise distances
link |
between all of the molecules in the protein.
link |
And then there had to be a separate optimization process
link |
to create the 3D structure.
link |
And what we did for AlphaFold two
link |
is make it truly end to end.
link |
So we went straight from the amino acid sequence of bases
link |
to the 3D structure directly
link |
without going through this intermediate step.
link |
And in machine learning, what we've always found is
link |
that the more end to end you can make it,
link |
the better the system.
link |
And it's probably because in the end,
link |
the system's better at learning what the constraints are
link |
than we are as the human designers of specifying it.
link |
So anytime you can let it flow end to end
link |
and actually just generate what it is
link |
you're really looking for, in this case, the 3D structure,
link |
you're better off than having this intermediate step,
link |
which you then have to handcraft the next step for.
link |
So it's better to let the gradients and the learning
link |
flow all the way through the system from the end point,
link |
the end output you want to the inputs.
link |
So that's a good way to start on a new problem.
link |
Handcraft a bunch of stuff,
link |
add a bunch of manual constraints
link |
with a small end to end learning piece
link |
or a small learning piece and grow that learning piece
link |
until it consumes the whole thing.
link |
And so you can also see,
link |
this is a bit of a method we've developed
link |
over doing many sort of successful alpha,
link |
we call them alpha X projects, right?
link |
And the easiest way to see that is the evolution
link |
of alpha go to alpha zero.
link |
So alpha go was a learning system,
link |
but it was specifically trained to only play go, right?
link |
So, and what we wanted to do with first version of alpha go
link |
is just get to world champion performance
link |
no matter how we did it, right?
link |
And then of course, alpha go zero,
link |
we remove the need to use human games as a starting point,
link |
So it could just play against itself
link |
from random starting point from the beginning.
link |
So that removed the need for human knowledge about go.
link |
And then finally alpha zero then generalized it
link |
so that any things we had in there, the system,
link |
including things like symmetry of the go board were removed.
link |
So the alpha zero could play from scratch
link |
any two player game and then mu zero,
link |
which is the final, our latest version
link |
of that set of things was then extending it
link |
so that you didn't even have to give it
link |
the rules of the game.
link |
It would learn that for itself.
link |
So it could also deal with computer games
link |
as well as board games.
link |
So that line of alpha go, alpha go zero, alpha zero,
link |
mu zero, that's the full trajectory
link |
of what you can take from imitation learning
link |
to full self supervised learning.
link |
And learning the entire structure
link |
of the environment you're put in from scratch, right?
link |
And bootstrapping it through self play yourself.
link |
But the thing is it would have been impossible, I think,
link |
or very hard for us to build alpha zero
link |
or mu zero first out of the box.
link |
Even psychologically, because you have to believe
link |
in yourself for a very long time.
link |
You're constantly dealing with doubt
link |
because a lot of people say that it's impossible.
link |
Exactly, so it's hard enough just to do go.
link |
As you were saying, everyone thought that was impossible
link |
or at least a decade away from when we did it
link |
back in 2015, 2016.
link |
And so yes, it would have been psychologically
link |
probably very difficult as well as the fact
link |
that of course we learn a lot by building alpha go first.
link |
Right, so I think this is why I call AI
link |
an engineering science.
link |
It's one of the most fascinating science disciplines,
link |
but it's also an engineering science in the sense
link |
that unlike natural sciences, the phenomenon you're studying
link |
doesn't exist out in nature.
link |
You have to build it first.
link |
So you have to build the artifact first,
link |
and then you can study and pull it apart and how it works.
link |
This is tough to ask you this question
link |
because you probably will say it's everything,
link |
but let's try to think through this
link |
because you're in a very interesting position
link |
where DeepMind is a place of some of the most brilliant
link |
ideas in the history of AI,
link |
but it's also a place of brilliant engineering.
link |
So how much of solving intelligence,
link |
this big goal for DeepMind,
link |
how much of it is science?
link |
How much is engineering?
link |
So how much is the algorithms?
link |
How much is the data?
link |
How much is the hardware compute infrastructure?
link |
How much is it the software compute infrastructure?
link |
What else is there?
link |
How much is the human infrastructure?
link |
And like just the humans interacting certain kinds of ways
link |
in all the space of all those ideas.
link |
And how much is maybe like philosophy?
link |
How much, what's the key?
link |
If you were to sort of look back,
link |
like if we go forward 200 years and look back,
link |
what was the key thing that solved intelligence?
link |
Is it the ideas or the engineering?
link |
I think it's a combination.
link |
First of all, of course,
link |
it's a combination of all those things,
link |
but the ratios of them changed over time.
link |
So even in the last 12 years,
link |
so we started DeepMind in 2010,
link |
which is hard to imagine now because 2010,
link |
it's only 12 short years ago,
link |
but nobody was talking about AI.
link |
I don't know if you remember back to your MIT days,
link |
no one was talking about it.
link |
I did a postdoc at MIT back around then.
link |
And it was sort of thought of as a,
link |
well, look, we know AI doesn't work.
link |
We tried this hard in the 90s at places like MIT,
link |
mostly using logic systems and old fashioned,
link |
sort of good old fashioned AI, we would call it now.
link |
People like Minsky and Patrick Winston,
link |
and you know all these characters, right?
link |
And used to debate a few of them.
link |
And they used to think I was mad thinking about
link |
that some new advance could be done with learning systems.
link |
And I was actually pleased to hear that
link |
because at least you know you're on a unique track
link |
at that point, right?
link |
Even if all of your professors are telling you you're mad.
link |
And of course in industry,
link |
we couldn't get, it was difficult to get two cents together,
link |
which is hard to imagine now as well,
link |
given that it's the biggest sort of buzzword in VCs
link |
and fundraisings easy and all these kinds of things today.
link |
So back in 2010, it was very difficult.
link |
And the reason we started then,
link |
and Shane and I used to discuss
link |
what were the sort of founding tenants of DeepMind.
link |
And it was various things.
link |
One was algorithmic advances.
link |
So deep learning, you know,
link |
Jeff Hinton and Co had just sort of invented that
link |
in academia, but no one in industry knew about it.
link |
We love reinforcement learning.
link |
We thought that could be scaled up.
link |
But also understanding about the human brain
link |
had advanced quite a lot in the decade prior
link |
with fMRI machines and other things.
link |
So we could get some good hints about architectures
link |
and algorithms and sort of representations maybe
link |
that the brain uses.
link |
So at a systems level, not at a implementation level.
link |
And then the other big things were compute and GPUs, right?
link |
So we could see a compute was going to be really useful
link |
and had got to a place where it become commoditized
link |
mostly through the games industry
link |
and that could be taken advantage of.
link |
And then the final thing was also mathematical
link |
and theoretical definitions of intelligence.
link |
So things like AIXI, AIXE,
link |
which Shane worked on with his supervisor, Marcus Hutter,
link |
which is this sort of theoretical proof really
link |
of universal intelligence,
link |
which is actually a reinforcement learning system
link |
I mean, it assumes infinite compute and infinite memory
link |
in the way, you know, like a Turing machine proves.
link |
But I was also waiting to see something like that too,
link |
to, you know, like Turing machines and computation theory
link |
that people like Turing and Shannon came up with
link |
underpins modern computer science.
link |
You know, I was waiting for a theory like that
link |
to sort of underpin AGI research.
link |
So when I, you know, met Shane
link |
and saw he was working on something like that,
link |
you know, that to me was a sort of final piece
link |
So in the early days, I would say that ideas
link |
were the most important.
link |
You know, for us, it was deep reinforcement learning,
link |
scaling up deep learning.
link |
Of course, we've seen transformers.
link |
So huge leaps, I would say, you know, three or four
link |
from, if you think from 2010 till now,
link |
huge evolutions, things like AlphaGo.
link |
And maybe there's a few more still needed.
link |
But as we get closer to AI, AGI,
link |
I think engineering becomes more and more important
link |
and data because scale and of course the recent,
link |
you know, results of GPT3 and all the big language models
link |
and large models, including our ones,
link |
has shown that scale and large models
link |
are clearly gonna be a necessary,
link |
but perhaps not sufficient part of an AGI solution.
link |
And throughout that, like you said,
link |
and I'd like to give you a big thank you.
link |
You're one of the pioneers in this is sticking by ideas
link |
like reinforcement learning, that this can actually work
link |
given actually limited success in the past.
link |
And also, which we still don't know,
link |
but proudly having the best researchers in the world
link |
and talking about solving intelligence.
link |
So talking about whatever you call it,
link |
AGI or something like this, speaking of MIT,
link |
that's just something you wouldn't bring up.
link |
Not maybe you did in like 40, 50 years ago,
link |
but that was, AI was a place where you do tinkering,
link |
very small scale, not very ambitious projects.
link |
And maybe the biggest ambitious projects
link |
were in the space of robotics
link |
and doing like the DARPA challenge.
link |
But the task of solving intelligence and believing you can,
link |
that's really, really powerful.
link |
So in order for engineering to do its work,
link |
to have great engineers, build great systems,
link |
you have to have that belief,
link |
that threads throughout the whole thing
link |
that you can actually solve
link |
some of these impossible challenges.
link |
Yeah, that's right.
link |
And back in 2010, our mission statement and still is today,
link |
it was used to be solving step one, solve intelligence,
link |
step two, use it to solve everything else.
link |
So if you can imagine pitching that to a VC in 2010,
link |
the kind of looks we got,
link |
we managed to find a few kooky people to back us,
link |
but it was tricky.
link |
And it got to the point where we wouldn't mention it
link |
to any of our professors because they would just eye roll
link |
and think we committed career suicide.
link |
And so it was, there's a lot of things that we had to do,
link |
but we always believed it.
link |
And one reason, by the way,
link |
one reason I've always believed in reinforcement learning
link |
is that if you look at neuroscience,
link |
that is the way that the primate brain learns.
link |
One of the main mechanisms is the dopamine system
link |
implements some form of TD learning.
link |
It was a very famous result in the late 90s
link |
where they saw this in monkeys
link |
and as a propagating prediction error.
link |
So again, in the limit,
link |
this is what I think you can use neuroscience for is,
link |
at mathematics, when you're doing something as ambitious
link |
as trying to solve intelligence
link |
and it's blue sky research, no one knows how to do it,
link |
you need to use any evidence
link |
or any source of information you can
link |
to help guide you in the right direction
link |
or give you confidence you're going in the right direction.
link |
So that was one reason we pushed so hard on that.
link |
And just going back to your earlier question
link |
about organization, the other big thing
link |
that I think we innovated with at DeepMind
link |
to encourage invention and innovation
link |
was the multidisciplinary organization we built
link |
and we still have today.
link |
So DeepMind originally was a confluence
link |
of the most cutting edge knowledge in neuroscience
link |
with machine learning, engineering and mathematics, right?
link |
And then since then we've built that out even further.
link |
So we have philosophers here and ethicists,
link |
but also other types of scientists, physicists and so on.
link |
And that's what brings together,
link |
I tried to build a sort of new type of Bell Labs,
link |
but in its golden era, right?
link |
And a new expression of that to try and foster
link |
this incredible sort of innovation machine.
link |
So talking about the humans in the machine,
link |
DeepMind itself is a learning machine
link |
with lots of amazing human minds in it
link |
coming together to try and build these learning systems.
link |
If we return to the big ambitious dream of AlphaFold,
link |
that may be the early steps on a very long journey
link |
in biology, do you think the same kind of approach
link |
can use to predict the structure and function
link |
of more complex biological systems?
link |
So multi protein interaction,
link |
and then, I mean, you can go out from there,
link |
just simulating bigger and bigger systems
link |
that eventually simulate something like the human brain
link |
or the human body, just the big mush,
link |
the mess of the beautiful, resilient mess of biology.
link |
Do you see that as a long term vision?
link |
I do, and I think, if you think about
link |
what are the top things I wanted to apply AI to
link |
once we had powerful enough systems,
link |
biology and curing diseases and understanding biology
link |
was right up there, top of my list.
link |
That's one of the reasons I personally pushed that myself
link |
and with AlphaFold, but I think AlphaFold,
link |
amazing as it is, is just the beginning.
link |
And I hope it's evidence of what could be done
link |
with computational methods.
link |
So AlphaFold solved this huge problem
link |
of the structure of proteins, but biology is dynamic.
link |
So really what I imagine from here,
link |
and we're working on all these things now,
link |
is protein, protein interaction, protein ligand binding,
link |
so reacting with molecules,
link |
then you wanna build up to pathways,
link |
and then eventually a virtual cell.
link |
That's my dream, maybe in the next 10 years.
link |
And I've been talking actually
link |
to a lot of biologists, friends of mine,
link |
Paul Nurse, who runs the Crick Institute,
link |
amazing biologists, Nobel Prize winning biologists.
link |
We've been discussing for 20 years now, virtual cells.
link |
Could you build a virtual simulation of a cell?
link |
And if you could, that would be incredible
link |
for biology and disease discovery,
link |
because you could do loads of experiments
link |
on the virtual cell, and then only at the last stage,
link |
validate it in the wet lab.
link |
So you could, in terms of the search space
link |
of discovering new drugs, it takes 10 years roughly
link |
to go from identifying a target,
link |
to having a drug candidate.
link |
Maybe that could be shortened by an order of magnitude,
link |
if you could do most of that work in silico.
link |
So in order to get to a virtual cell,
link |
we have to build up understanding
link |
of different parts of biology and the interactions.
link |
And so every few years we talk about this,
link |
I talked about this with Paul.
link |
And then finally, last year after AlphaFold,
link |
I said, now's the time we can finally go for it.
link |
And AlphaFold is the first proof point
link |
that this might be possible.
link |
And he's very excited, and we have some collaborations
link |
with his lab, they're just across the road actually
link |
from us, it's wonderful being here in King's Cross
link |
with the Crick Institute across the road.
link |
And I think the next steps,
link |
I think there's gonna be some amazing advances
link |
in biology built on top of things like AlphaFold.
link |
We're already seeing that with the community doing that
link |
after we've open sourced it and released it.
link |
And I often say that I think if you think of mathematics
link |
is the perfect description language for physics,
link |
I think AI might be end up being
link |
the perfect description language for biology
link |
because biology is so messy, it's so emergent,
link |
so dynamic and complex.
link |
I think I find it very hard to believe
link |
we'll ever get to something as elegant
link |
as Newton's laws of motions to describe a cell, right?
link |
It's just too complicated.
link |
So I think AI is the right tool for that.
link |
So you have to start at the basic building blocks
link |
and use AI to run the simulation
link |
for all those building blocks.
link |
So have a very strong way to do prediction
link |
of what given these building blocks,
link |
what kind of biology, how the function
link |
and the evolution of that biological system.
link |
It's almost like a cellular automata,
link |
you have to run it, you can't analyze it from a high level.
link |
You have to take the basic ingredients,
link |
figure out the rules and let it run.
link |
But in this case, the rules are very difficult
link |
to figure out, you have to learn them.
link |
That's exactly it.
link |
So the biology is too complicated to figure out the rules.
link |
It's too emergent, too dynamic,
link |
say compared to a physics system,
link |
like the motion of a planet, right?
link |
And so you have to learn the rules
link |
and that's exactly the type of systems that we're building.
link |
So you mentioned you've open sourced AlphaFold
link |
and even the data involved.
link |
To me personally, also really happy
link |
and a big thank you for open sourcing Mojoko,
link |
the physics simulation engine that's often used
link |
for robotics research and so on.
link |
So I think that's a pretty gangster move.
link |
So what's the, I mean, very few companies
link |
or people do that kind of thing.
link |
What's the philosophy behind that?
link |
You know, it's a case by case basis.
link |
And in both of those cases,
link |
we felt that was the maximum benefit to humanity to do that.
link |
And the scientific community, in one case,
link |
the robotics physics community with Mojoko, so.
link |
We purchased it for, yes,
link |
we purchased it for the express principle to open source it.
link |
So, you know, I hope people appreciate that.
link |
It's great to hear that you do.
link |
And then the second thing was,
link |
and mostly we did it because the person building it
link |
was not able to cope with supporting it anymore
link |
because it got too big for him.
link |
He's an amazing professor who built it in the first place.
link |
So we helped him out with that.
link |
And then with AlphaFold is even bigger, I would say.
link |
And I think in that case,
link |
we decided that there were so many downstream applications
link |
of AlphaFold that we couldn't possibly even imagine
link |
what they all were.
link |
So the best way to accelerate drug discovery
link |
and also fundamental research would be to give all
link |
that data away and the system itself.
link |
You know, it's been so gratifying to see
link |
what people have done that within just one year,
link |
which is a short amount of time in science.
link |
And it's been used by over 500,000 researchers have used it.
link |
We think that's almost every biologist in the world.
link |
I think there's roughly 500,000 biologists in the world,
link |
professional biologists,
link |
have used it to look at their proteins of interest.
link |
We've seen amazing fundamental research done.
link |
So a couple of weeks ago, front cover,
link |
there was a whole special issue of science,
link |
including the front cover,
link |
which had the nuclear pore complex on it,
link |
which is one of the biggest proteins in the body.
link |
The nuclear pore complex is a protein that governs
link |
all the nutrients going in and out of your cell nucleus.
link |
So they're like little gateways that open and close
link |
to let things go in and out of your cell nucleus.
link |
So they're really important, but they're huge
link |
because they're massive donut ring shaped things.
link |
And they've been looking to try and figure out
link |
that structure for decades.
link |
And they have lots of experimental data,
link |
but it's too low resolution, there's bits missing.
link |
And they were able to, like a giant Lego jigsaw puzzle,
link |
use alpha fold predictions plus experimental data
link |
and combined those two independent sources of information,
link |
actually four different groups around the world
link |
were able to put it together more or less simultaneously
link |
using alpha fold predictions.
link |
So that's been amazing to see.
link |
And pretty much every pharma company,
link |
every drug company executive I've spoken to
link |
has said that their teams are using alpha fold
link |
to accelerate whatever drugs they're trying to discover.
link |
So I think the knock on effect has been enormous
link |
in terms of the impact that alpha fold has made.
link |
And it's probably bringing in, it's creating biologists,
link |
it's bringing more people into the field,
link |
both on the excitement and both on the technical skills
link |
involved in, it's almost like a gateway drug to biology.
link |
And to get more computational people involved too, hopefully.
link |
And I think for us, the next stage, as I said,
link |
in future we have to have other considerations too.
link |
We're building on top of alpha fold
link |
and these other ideas I discussed with you
link |
about protein interactions and genomics and other things.
link |
And not everything will be open source.
link |
Some of it we'll do commercially
link |
because that will be the best way
link |
to actually get the most resources and impact behind it.
link |
In other ways, some other projects
link |
we'll do nonprofit style.
link |
And also we have to consider for future things as well,
link |
safety and ethics as well.
link |
Like synthetic biology, there is dual use.
link |
And we have to think about that as well.
link |
With alpha fold, we consulted with 30 different bioethicists
link |
and other people expert in this field
link |
to make sure it was safe before we released it.
link |
So there'll be other considerations in future.
link |
But for right now, I think alpha fold
link |
is a kind of a gift from us to the scientific community.
link |
So I'm pretty sure that something like alpha fold
link |
will be part of Nobel prizes in the future.
link |
But us humans, of course,
link |
are horrible with credit assignment.
link |
So we'll of course give it to the humans.
link |
Do you think there will be a day
link |
when AI system can't be denied
link |
that it earned that Nobel prize?
link |
Do you think we will see that in 21st century?
link |
It depends what type of AIs we end up building, right?
link |
Whether they're goal seeking agents
link |
who specifies the goals, who comes up with the hypotheses,
link |
who determines which problems to tackle, right?
link |
And tweets about it, announcement of the results.
link |
Yes, and tweets about results exactly as part of it.
link |
So I think right now, of course,
link |
it's amazing human ingenuity that's behind these systems.
link |
And then the system, in my opinion, is just a tool.
link |
Be a bit like saying with Galileo and his telescope,
link |
the ingenuity that the credit should go to the telescope.
link |
I mean, it's clearly Galileo building the tool
link |
which he then uses.
link |
So I still see that in the same way today,
link |
even though these tools learn for themselves.
link |
There, I think of things like alpha fold
link |
and the things we're building as the ultimate tools
link |
for science and for acquiring new knowledge
link |
to help us as scientists acquire new knowledge.
link |
I think one day there will come a point
link |
where an AI system may solve
link |
or come up with something like general relativity
link |
of its own bat, not just by averaging everything
link |
on the internet or averaging everything on PubMed,
link |
although that would be interesting to see
link |
what that would come up with.
link |
So that to me is a bit like our earlier debate
link |
about creativity, you know, inventing go
link |
rather than just coming up with a good go move.
link |
And so I think solving, I think to, you know,
link |
if we wanted to give it the credit
link |
of like a Nobel type of thing,
link |
then it would need to invent go
link |
and sort of invent that new conjecture out of the blue
link |
rather than being specified by the human scientists
link |
or the human creators.
link |
So I think right now it's definitely just a tool.
link |
Although it is interesting how far you get
link |
by averaging everything on the internet, like you said,
link |
because, you know, a lot of people do see science
link |
as you're always standing on the shoulders of giants.
link |
And the question is how much are you really reaching
link |
up above the shoulders of giants?
link |
Maybe it's just simulating different kinds
link |
of results of the past with ultimately this new perspective
link |
that gives you this breakthrough idea.
link |
But that idea may not be novel in the way
link |
that it can't be already discovered on the internet.
link |
Maybe the Nobel prizes of the next 100 years
link |
are already all there on the internet to be discovered.
link |
They could be, they could be.
link |
I mean, I think this is one of the big mysteries,
link |
I think is that I, first of all,
link |
I believe a lot of the big new breakthroughs
link |
that are gonna come in the next few decades
link |
and even in the last decade are gonna come
link |
at the intersection between different subject areas
link |
where there'll be some new connection that's found
link |
between what seemingly were disparate areas.
link |
And one can even think of DeepMind, as I said earlier,
link |
as a sort of interdisciplinary between neuroscience ideas
link |
and AI engineering ideas originally.
link |
And so I think there's that.
link |
And then one of the things we can't imagine today is,
link |
and one of the reasons I think people,
link |
we were so surprised by how well large models worked
link |
is that actually it's very hard for our human minds,
link |
our limited human minds to understand
link |
what it would be like to read the whole internet, right?
link |
I think we can do a thought experiment
link |
and I used to do this of like,
link |
well, what if I read the whole of Wikipedia?
link |
What would I know?
link |
And I think our minds can just about comprehend
link |
maybe what that would be like,
link |
but the whole internet is beyond comprehension.
link |
So I think we just don't understand what it would be like
link |
to be able to hold all of that in mind potentially, right?
link |
And then active at once,
link |
and then maybe what are the connections
link |
that are available there?
link |
So I think no doubt there are huge things
link |
to be discovered just like that.
link |
But I do think there is this other type of creativity
link |
of true spark of new knowledge, new idea,
link |
never thought before about,
link |
can't be averaged from things that are known,
link |
that really, of course, everything come,
link |
nobody creates in a vacuum,
link |
so there must be clues somewhere,
link |
but just a unique way of putting those things together.
link |
I think some of the greatest scientists in history
link |
have displayed that I would say,
link |
although it's very hard to know going back to their time,
link |
what was exactly known when they came up with those things.
link |
Although you're making me really think because just a thought
link |
experiment of deeply knowing a hundred Wikipedia pages.
link |
I don't think I can,
link |
I've been really impressed by Wikipedia for technical topics.
link |
So if you know a hundred pages or a thousand pages,
link |
I don't think we can truly comprehend
link |
what kind of intelligence that is.
link |
That's a pretty powerful intelligence.
link |
If you know how to use that
link |
and integrate that information correctly,
link |
I think you can go really far.
link |
You can probably construct thought experiments
link |
based on that, like simulate different ideas.
link |
So if this is true, let me run this thought experiment
link |
that maybe this is true.
link |
It's not really invention.
link |
It's like just taking literally the knowledge
link |
and using it to construct the very basic simulation
link |
I mean, some argue it's romantic in part,
link |
but Einstein would do the same kind of things
link |
with a thought experiment.
link |
Yeah, one could imagine doing that systematically
link |
across millions of Wikipedia pages,
link |
plus PubMed, all these things.
link |
I think there are many, many things to be discovered
link |
like that that are hugely useful.
link |
You could imagine,
link |
and I want us to do some of these things in material science
link |
like room temperature superconductors
link |
is something on my list one day.
link |
I'd like to have an AI system to help build
link |
better optimized batteries,
link |
all of these sort of mechanical things.
link |
I think a systematic sort of search
link |
could be guided by a model,
link |
could be extremely powerful.
link |
So speaking of which,
link |
you have a paper on nuclear fusion,
link |
magnetic control of tachymic plasmas
link |
through deep reinforcement learning.
link |
So you're seeking to solve nuclear fusion with deep RL.
link |
So it's doing control of high temperature plasmas.
link |
Can you explain this work
link |
and can AI eventually solve nuclear fusion?
link |
It's been very fun last year or two and very productive
link |
because we've been taking off a lot of my dream projects,
link |
if you like, of things that I've collected
link |
over the years of areas of science
link |
that I would like to,
link |
I think could be very transformative if we helped accelerate
link |
and really interesting problems,
link |
scientific challenges in of themselves.
link |
So this is energy.
link |
So energy, yes, exactly.
link |
So energy and climate.
link |
So we talked about disease and biology
link |
as being one of the biggest places I think AI can help with.
link |
I think energy and climate is another one.
link |
So maybe they would be my top two.
link |
And fusion is one area I think AI can help with.
link |
Now, fusion has many challenges,
link |
mostly physics and material science
link |
and engineering challenges as well
link |
to build these massive fusion reactors
link |
and contain the plasma.
link |
And what we try to do,
link |
and whenever we go into a new field to apply our systems,
link |
is we look for, we talk to domain experts.
link |
We try and find the best people in the world
link |
to collaborate with.
link |
In this case, in fusion,
link |
we collaborated with EPFL in Switzerland,
link |
the Swiss Technical Institute, who are amazing.
link |
They have a test reactor.
link |
They were willing to let us use,
link |
which I double checked with the team
link |
we were gonna use carefully and safely.
link |
I was impressed they managed to persuade them
link |
And it's an amazing test reactor they have there.
link |
And they try all sorts of pretty crazy experiments on it.
link |
And what we tend to look at is,
link |
if we go into a new domain like fusion,
link |
what are all the bottleneck problems?
link |
Like thinking from first principles,
link |
what are all the bottleneck problems
link |
that are still stopping fusion working today?
link |
And then we look at, we get a fusion expert to tell us,
link |
and then we look at those bottlenecks
link |
and we look at the ones,
link |
which ones are amenable to our AI methods today, right?
link |
And would be interesting from a research perspective,
link |
from our point of view, from an AI point of view,
link |
and that would address one of their bottlenecks.
link |
And in this case, plasma control was perfect.
link |
So, the plasma, it's a million degrees Celsius,
link |
something like that, it's hotter than the sun.
link |
And there's obviously no material that can contain it.
link |
So, they have to be containing these magnetic,
link |
very powerful and superconducting magnetic fields.
link |
But the problem is plasma,
link |
it's pretty unstable as you imagine,
link |
you're kind of holding a mini sun, mini star in a reactor.
link |
So, you kind of want to predict ahead of time,
link |
what the plasma is gonna do.
link |
So, you can move the magnetic field
link |
within a few milliseconds,
link |
to basically contain what it's gonna do next.
link |
So, it seems like a perfect problem if you think of it
link |
for like a reinforcement learning prediction problem.
link |
So, you got controller, you're gonna move the magnetic field.
link |
And until we came along, they were doing it
link |
with traditional operational research type of controllers,
link |
which are kind of handcrafted.
link |
And the problem is, of course,
link |
they can't react in the moment
link |
to something the plasma is doing,
link |
they have to be hard coded.
link |
And again, knowing that that's normally our go to solution
link |
is we would like to learn that instead.
link |
And they also had a simulator of these plasma.
link |
So, there were lots of criteria
link |
that matched what we like to use.
link |
So, can AI eventually solve nuclear fusion?
link |
Well, so with this problem,
link |
and we published it in a nature paper last year,
link |
we held the fusion, we held the plasma in a specific shapes.
link |
So, actually, it's almost like carving the plasma
link |
into different shapes and hold it there
link |
for a record amount of time.
link |
So, that's one of the problems of fusion sort of solved.
link |
So, have a controller that's able to,
link |
no matter the shape.
link |
Contain it. Contain it.
link |
Yeah, contain it and hold it in structure.
link |
And there's different shapes that are better
link |
for the energy productions called droplets and so on.
link |
So, that was huge.
link |
And now we're looking,
link |
we're talking to lots of fusion startups
link |
to see what's the next problem we can tackle
link |
in the fusion area.
link |
So, another fascinating place in a paper titled,
link |
Pushing the Frontiers of Density Functionals
link |
by Solving the Fractional Electron Problem.
link |
So, you're taking on modeling and simulating
link |
the quantum mechanical behavior of electrons.
link |
Can you explain this work and can AI model
link |
and simulate arbitrary quantum mechanical systems
link |
Yeah, so this is another problem I've had my eye on
link |
for a decade or more,
link |
which is sort of simulating the properties of electrons.
link |
If you can do that, you can basically describe
link |
how elements and materials and substances work.
link |
So, it's kind of like fundamental
link |
if you want to advance material science.
link |
And we have Schrodinger's equation
link |
and then we have approximations
link |
to that density functional theory.
link |
These things are famous.
link |
And people try and write approximations
link |
to these functionals and kind of come up
link |
with descriptions of the electron clouds,
link |
where they're going to go,
link |
how they're going to interact
link |
when you put two elements together.
link |
And what we try to do is learn a simulation,
link |
learn a functional that will describe more chemistry,
link |
types of chemistry.
link |
So, until now, you can run expensive simulations,
link |
but then you can only simulate very small molecules,
link |
very simple molecules.
link |
We would like to simulate large materials.
link |
And so, today there's no way of doing that.
link |
And we're building up towards building functionals
link |
that approximate Schrodinger's equation
link |
and then allow you to describe what the electrons are doing.
link |
And all material sort of science
link |
and material properties are governed by the electrons
link |
and how they interact.
link |
So, have a good summarization of the simulation
link |
through the functional,
link |
but one that is still close
link |
to what the actual simulation would come out with.
link |
So, how difficult is that task?
link |
What's involved in that task?
link |
Is it running those complicated simulations
link |
and learning the task of mapping
link |
from the initial conditions
link |
and the parameters of the simulation,
link |
learning what the functional would be?
link |
So, it's pretty tricky.
link |
And we've done it with,
link |
the nice thing is we can run a lot of the simulations,
link |
the molecular dynamic simulations on our compute clusters.
link |
And so, that generates a lot of data.
link |
So, in this case, the data is generated.
link |
So, we like those sort of systems and that's why we use games.
link |
It's simulated, generated data.
link |
And we can kind of create as much of it as we want, really.
link |
And just let's leave some,
link |
if any computers are free in the cloud,
link |
we just run, we run some of these calculations, right?
link |
Compute cluster calculation.
link |
I like how the free compute time
link |
is used up on quantum mechanics.
link |
Yeah, quantum mechanics, exactly.
link |
Simulations and protein simulations and other things.
link |
And so, when you're not searching on YouTube
link |
for free video, cat videos,
link |
we're using those computers usefully in quantum chemistry.
link |
And putting them to good use.
link |
And then, yeah, and then all of that computational data
link |
we can then try and learn the functionals from that,
link |
which of course are way more efficient
link |
once we learn the functional
link |
than running those simulations would be.
link |
Do you think one day AI may allow us
link |
to do something like basically crack open physics?
link |
So, do something like travel faster than the speed of light?
link |
My ultimate aim is always being with AI
link |
is the reason I am personally working on AI
link |
for my whole life, it was to build a tool
link |
to help us understand the universe.
link |
So, I wanted to, and that means physics, really,
link |
and the nature of reality.
link |
So, I don't think we have systems
link |
that are capable of doing that yet,
link |
but when we get towards AGI,
link |
I think that's one of the first things
link |
I think we should apply AGI to.
link |
I would like to test the limits of physics
link |
and our knowledge of physics.
link |
There's so many things we don't know.
link |
This is one thing I find fascinating about science.
link |
And as a huge proponent of the scientific method
link |
as being one of the greatest ideas humanity has ever had
link |
and allowed us to progress with our knowledge,
link |
but I think as a true scientist,
link |
I think what you find is the more you find out,
link |
the more you realize we don't know.
link |
And I always think that it's surprising
link |
that more people aren't troubled.
link |
Every night I think about all these things
link |
we interact with all the time,
link |
that we have no idea how they work.
link |
Time, consciousness, gravity, life, we can't,
link |
I mean, these are all the fundamental things of nature.
link |
I think the way we don't really know what they are.
link |
To live life, we pin certain assumptions on them
link |
and kind of treat our assumptions as if they're a fact.
link |
That allows us to sort of box them off somehow.
link |
Yeah, box them off somehow.
link |
But the reality is when you think of time,
link |
you should remind yourself,
link |
you should take it off the shelf
link |
and realize like, no, we have a bunch of assumptions.
link |
There's still a lot of, there's even now a lot of debate.
link |
There's a lot of uncertainty about exactly what is time.
link |
Is there an error of time?
link |
You know, there's a lot of fundamental questions
link |
that you can't just make assumptions about.
link |
And maybe AI allows you to not put anything on the shelf.
link |
Not make any hard assumptions
link |
and really open it up and see what's.
link |
Exactly, I think we should be truly open minded about that.
link |
And exactly that, not be dogmatic to a particular theory.
link |
It'll also allow us to build better tools,
link |
experimental tools eventually,
link |
that can then test certain theories
link |
that may not be testable today.
link |
Things about like what we spoke about at the beginning
link |
about the computational nature of the universe.
link |
How one might, if that was true,
link |
how one might go about testing that, right?
link |
And how much, you know, there are people
link |
who've conjectured people like Scott Aaronson and others
link |
about, you know, how much information
link |
can a specific plank unit of space and time
link |
So one might be able to think about testing those ideas
link |
if you had AI helping you build
link |
some new exquisite experimental tools.
link |
This is what I imagine that, you know,
link |
many decades from now we'll be able to do.
link |
And what kind of questions can be answered
link |
through running a simulation of them?
link |
So there's a bunch of physics simulations
link |
you can imagine that could be run
link |
in some kind of efficient way,
link |
much like you're doing in the quantum simulation work.
link |
And perhaps even the origin of life.
link |
So figuring out how going even back
link |
before the work of AlphaFold begins
link |
of how this whole thing emerges from a rock.
link |
From a static thing.
link |
What do you think AI will allow us to,
link |
is that something you have your eye on?
link |
It's trying to understand the origin of life.
link |
First of all, yourself, what do you think,
link |
how the heck did life originate on Earth?
link |
Yeah, well, maybe I'll come to that in a second,
link |
but I think the ultimate use of AI
link |
is to kind of use it to accelerate science to the maximum.
link |
So I think of it a little bit
link |
like the tree of all knowledge.
link |
If you imagine that's all the knowledge there is
link |
in the universe to attain.
link |
And we sort of barely scratched the surface of that so far.
link |
And even though we've done pretty well
link |
since the enlightenment, right, as humanity.
link |
And I think AI will turbocharge all of that,
link |
like we've seen with AlphaFold.
link |
And I want to explore as much of that tree of knowledge
link |
as is possible to do.
link |
And I think that involves AI helping us
link |
with understanding or finding patterns,
link |
but also potentially designing and building new tools,
link |
experimental tools.
link |
So I think that's all,
link |
and also running simulations and learning simulations,
link |
all of that we're sort of doing at a baby steps level here.
link |
But I can imagine that in the decades to come
link |
as what's the full flourishing of that line of thinking.
link |
It's gonna be truly incredible, I would say.
link |
If I visualized this tree of knowledge,
link |
something tells me that that tree of knowledge for humans
link |
is much smaller in the set of all possible trees
link |
of knowledge, it's actually quite small
link |
given our cognitive limitations,
link |
limited cognitive capabilities,
link |
that even with the tools we build,
link |
we still won't be able to understand a lot of things.
link |
And that's perhaps what nonhuman systems
link |
might be able to reach farther, not just as tools,
link |
but in themselves understanding something
link |
that they can bring back.
link |
Yeah, it could well be.
link |
So, I mean, there's so many things
link |
that are sort of encapsulated in what you just said there.
link |
I think first of all, there's two different things.
link |
There's like, what do we understand today?
link |
What could the human mind understand?
link |
And what is the totality of what is there to be understood?
link |
And so there's three concentric,
link |
you can think of them as three larger and larger trees
link |
or exploring more branches of that tree.
link |
And I think with AI, we're gonna explore that whole lot.
link |
Now, the question is, if you think about
link |
what is the totality of what could be understood,
link |
there may be some fundamental physics reasons
link |
why certain things can't be understood,
link |
like what's outside a simulation or outside the universe.
link |
Maybe it's not understandable from within the universe.
link |
So there may be some hard constraints like that.
link |
It could be smaller constraints,
link |
like we think of space time as fundamental.
link |
Our human brains are really used to this idea
link |
of a three dimensional world with time, maybe.
link |
But our tools could go beyond that.
link |
They wouldn't have that limitation necessarily.
link |
They could think in 11 dimensions, 12 dimensions,
link |
whatever is needed.
link |
But we could still maybe understand that
link |
in several different ways.
link |
The example I always give is,
link |
when I play Garry Kasparov for speed chess,
link |
or we've talked about chess and these kinds of things,
link |
you know, if you're reasonably good at chess,
link |
you can't come up with the move Garry comes up with
link |
in his move, but he can explain it to you.
link |
And you can understand.
link |
And you can understand post hoc the reasoning.
link |
So I think there's an even further level of like,
link |
well, maybe you couldn't have invented that thing,
link |
but going back to using language again,
link |
perhaps you can understand and appreciate that.
link |
Same way that you can appreciate, you know,
link |
Vivaldi or Mozart or something without,
link |
you can appreciate the beauty of that
link |
without being able to construct it yourself, right?
link |
Invent the music yourself.
link |
So I think we see this in all forms of life.
link |
So it will be that times, you know, a million,
link |
but you can imagine also one sign of intelligence
link |
is the ability to explain things clearly and simply, right?
link |
You know, people like Richard Feynman,
link |
another one of my old time heroes used to say that, right?
link |
If you can't, you know, if you can explain it
link |
something simply, then that's the best sign,
link |
a complex topic simply,
link |
then that's one of the best signs of you understanding it.
link |
I can see myself talking trash in the AI system in that way.
link |
It gets frustrated how dumb I am
link |
and trying to explain something to me.
link |
I was like, well, that means you're not intelligent
link |
because if you were intelligent,
link |
you'd be able to explain it simply.
link |
Yeah, of course, you know, there's also the other option.
link |
Of course, we could enhance ourselves and with our devices,
link |
we are already sort of symbiotic with our compute devices,
link |
right, with our phones and other things.
link |
And, you know, there's stuff like Neuralink and Xceptra
link |
that could advance that further.
link |
So I think there's lots of really amazing possibilities
link |
that I could foresee from here.
link |
Well, let me ask you some wild questions.
link |
So out there looking for friends,
link |
do you think there's a lot of alien civilizations out there?
link |
So I guess this also goes back
link |
to your origin of life question too,
link |
because I think that that's key.
link |
My personal opinion, looking at all this,
link |
and, you know, it's one of my hobbies, physics, I guess.
link |
So, you know, it's something I think about a lot
link |
and talk to a lot of experts on and read a lot of books on.
link |
And I think my feeling currently is that we are alone.
link |
I think that's the most likely scenario
link |
given what evidence we have.
link |
So, and the reasoning is I think that, you know,
link |
we've tried since things like SETI program
link |
and I guess since the dawning of the space age,
link |
we've, you know, had telescopes,
link |
open radio telescopes and other things.
link |
And if you think about and try to detect signals,
link |
now, if you think about the evolution of humans on earth,
link |
we could have easily been a million years ahead
link |
of our time now or million years behind,
link |
right, easily with just some slightly different quirk
link |
thing happening hundreds of thousands of years ago.
link |
You know, things could have been slightly different
link |
if the meteor would hit the dinosaurs a million years earlier,
link |
maybe things would have evolved.
link |
We'd be a million years ahead of where we are now.
link |
So what that means is if you imagine where humanity will be
link |
in a few hundred years, let alone a million years,
link |
especially if we hopefully, you know,
link |
solve things like climate change and other things,
link |
and we continue to flourish and we build things like AI
link |
and we do space traveling and all of the stuff
link |
that humans have dreamed of forever, right?
link |
And sci fi is talked about forever.
link |
We will be spreading across the stars, right?
link |
And von Neumann famously calculated, you know,
link |
it would only take about a million years
link |
if you sent out von Neumann probes to the nearest,
link |
you know, the nearest other solar systems.
link |
And then all they did was build two more versions
link |
of themselves and sent those two out
link |
to the next nearest systems.
link |
You know, within a million years,
link |
I think you would have one of these probes
link |
in every system in the galaxy.
link |
So it's not actually in cosmological time.
link |
That's actually a very short amount of time.
link |
So, and you know, people like Dyson have thought
link |
about constructing Dyson spheres around stars
link |
to collect all the energy coming out of the star.
link |
You know, there would be constructions like that
link |
would be visible across space,
link |
probably even across a galaxy.
link |
So, and then, you know, if you think about
link |
all of our radio, television emissions
link |
that have gone out since the, you know, 30s and 40s,
link |
imagine a million years of that.
link |
And now hundreds of civilizations doing that.
link |
When we opened our ears at the point
link |
we got technologically sophisticated enough
link |
we should have heard a cacophony of voices.
link |
We should have joined that cacophony of voices.
link |
And what we did, we opened our ears and we heard nothing.
link |
And many people who argue that there are aliens
link |
would say, well, we haven't really done
link |
exhaustive search yet.
link |
And maybe we're looking in the wrong bands
link |
and we've got the wrong devices
link |
and we wouldn't notice what an alien form was like
link |
because it'd be so different to what we're used to.
link |
But, you know, I don't really buy that,
link |
that it shouldn't be as difficult as that.
link |
Like, I think we've searched enough.
link |
There should be everywhere.
link |
If it was, yeah, it should be everywhere.
link |
We should see Dyson spheres being put up,
link |
sun's blinking in and out.
link |
You know, there should be a lot of evidence
link |
And then there are other people who argue,
link |
well, the sort of safari view of like,
link |
well, we're a primitive species still
link |
because we're not space faring yet.
link |
And we're, you know, there's some kind of global,
link |
like universal rule not to interfere,
link |
you know, Star Trek rule.
link |
But like, look, we can't even coordinate humans
link |
to deal with climate change and we're one species.
link |
What is the chance that of all of these different
link |
human civilization, you know, alien civilizations,
link |
they would have the same priorities
link |
and agree across these kinds of matters.
link |
And even if that was true
link |
and we were in some sort of safari for our own good,
link |
to me, that's not much different
link |
from the simulation hypothesis
link |
because what does it mean, the simulation hypothesis?
link |
I think in its most fundamental level,
link |
it means what we're seeing is not quite reality, right?
link |
It's something, there's something more deeper underlying it,
link |
maybe computational.
link |
Now, if we were in a sort of safari park
link |
and everything we were seeing was a hologram
link |
and it was projected by the aliens or whatever,
link |
that to me is not much different
link |
than thinking we're inside of another universe
link |
because we still can't see true reality, right?
link |
I mean, there's other explanations.
link |
It could be that the way they're communicating
link |
is just fundamentally different,
link |
that we're too dumb to understand the much better methods
link |
of communication they have.
link |
It could be, I mean, it's silly to say,
link |
but our own thoughts could be the methods
link |
by which they're communicating.
link |
Like the place from which our ideas,
link |
writers talk about this, like the muse.
link |
I mean, it sounds like very kind of wild,
link |
but it could be thoughts.
link |
It could be some interactions with our mind
link |
that we think are originating from us
link |
is actually something that is coming
link |
from other life forms elsewhere.
link |
Consciousness itself might be that.
link |
It could be, but I don't see any sensible argument
link |
to the why would all of the alien species
link |
behave in this way?
link |
Yeah, some of them will be more primitive.
link |
They will be close to our level.
link |
There should be a whole sort of normal distribution
link |
of these things, right?
link |
Some would be aggressive.
link |
Some would be curious.
link |
Others would be very historical and philosophical
link |
because maybe they're a million years older than us,
link |
but it's not, it shouldn't be like,
link |
I mean, one alien civilization might be like that,
link |
communicating thoughts and others,
link |
but I don't see why potentially the hundreds there should be
link |
would be uniform in this way, right?
link |
It could be a violent dictatorship that the people,
link |
the alien civilizations that become successful
link |
gain the ability to be destructive,
link |
an order of magnitude more destructive,
link |
but of course the sad thought,
link |
well, either humans are very special.
link |
We took a lot of leaps that arrived
link |
at what it means to be human.
link |
There's a question there, which was the hardest,
link |
which was the most special,
link |
but also if others have reached this level
link |
and maybe many others have reached this level,
link |
the great filter that prevented them from going farther
link |
to becoming a multi planetary species
link |
or reaching out into the stars.
link |
And those are really important questions for us,
link |
whether there's other alien civilizations out there or not,
link |
this is very useful for us to think about.
link |
If we destroy ourselves, how will we do it?
link |
And how easy is it to do?
link |
Yeah, well, these are big questions
link |
and I've thought about these a lot,
link |
but the interesting thing is that if we're alone,
link |
that's somewhat comforting from the great filter perspective
link |
because it probably means the great filters were passed us.
link |
And I'm pretty sure they are.
link |
So going back to your origin of life question,
link |
there are some incredible things
link |
that no one knows how happened,
link |
like obviously the first life form from chemical soup,
link |
that seems pretty hard,
link |
but I would guess the multicellular,
link |
I wouldn't be that surprised if we saw single cell
link |
sort of life forms elsewhere, bacteria type things,
link |
but multicellular life seems incredibly hard,
link |
that step of capturing mitochondria
link |
and then sort of using that as part of yourself,
link |
you know, when you've just eaten it.
link |
Would you say that's the biggest, the most,
link |
like if you had to choose one sort of,
link |
Hitchhiker's Galaxy, one sentence summary of like,
link |
oh, those clever creatures did this,
link |
that would be the multicellular.
link |
I think that was probably the one that's the biggest.
link |
I mean, there's a great book
link |
called The 10 Great Inventions of Evolution by Nick Lane,
link |
and he speculates on 10 of these, you know,
link |
what could be great filters.
link |
I think that's one.
link |
I think the advent of intelligence
link |
and conscious intelligence and in order, you know,
link |
to us to be able to do science and things like that
link |
I mean, it's only evolved once as far as, you know,
link |
So that would be a later candidate,
link |
but there's certainly for the early candidates,
link |
I think multicellular life forms is huge.
link |
By the way, what it's interesting to ask you,
link |
if you can hypothesize about
link |
what is the origin of intelligence?
link |
Is it that we started cooking meat over fire?
link |
Is it that we somehow figured out
link |
that we could be very powerful when we started collaborating?
link |
So cooperation between our ancestors
link |
so that we can overthrow the alpha male.
link |
What is it, Richard?
link |
I talked to Richard Ranham,
link |
who thinks we're all just beta males
link |
who figured out how to collaborate to defeat the one,
link |
the dictator, the authoritarian alpha male
link |
that controlled the tribe.
link |
Is there other explanation?
link |
Was there 2001 Space Odyssey type of monolith
link |
that came down to Earth?
link |
Well, I think all of those things
link |
you suggested are good candidates,
link |
fire and cooking, right?
link |
So that's clearly important for energy efficiency,
link |
cooking our meat and then being able to be more efficient
link |
about eating it and consuming the energy.
link |
I think that's huge and then utilizing fire and tools
link |
I think you're right about the tribal cooperation aspects
link |
and probably language is part of that
link |
because probably that's what allowed us
link |
to outcompete Neanderthals
link |
and perhaps less cooperative species.
link |
So that may be the case.
link |
Tool making, spears, axes, I think that let us,
link |
I mean, I think it's pretty clear now
link |
that humans were responsible
link |
for a lot of the extinctions of megafauna,
link |
especially in the Americas when humans arrived.
link |
So you can imagine once you discover tool usage
link |
how powerful that would have been
link |
and how scary for animals.
link |
So I think all of those could have been explanations for it.
link |
The interesting thing is that it's a bit
link |
like general intelligence too,
link |
is it's very costly to begin with to have a brain
link |
and especially a general purpose brain
link |
rather than a special purpose one
link |
because the amount of energy our brains use,
link |
I think it's like 20% of the body's energy
link |
and it's massive and even your thinking chest,
link |
one of the funny things that we used to say
link |
is it's as much as a racing driver uses
link |
for a whole Formula One race,
link |
just playing a game of serious high level chess,
link |
which you wouldn't think just sitting there
link |
because the brain's using so much energy.
link |
So in order for an animal, an organism to justify that,
link |
there has to be a huge payoff.
link |
And the problem with half a brain
link |
or half intelligence, say an IQs of like a monkey brain,
link |
it's not clear you can justify that evolutionary
link |
until you get to the human level brain.
link |
And so, but how do you do that jump?
link |
It's very difficult,
link |
which is why I think it has only been done once
link |
from the sort of specialized brains that you see in animals
link |
to this sort of general purpose,
link |
chewing powerful brains that humans have
link |
and which allows us to invent the modern world.
link |
And it takes a lot to cross that barrier.
link |
And I think we've seen the same with AI systems,
link |
which is that maybe until very recently,
link |
it's always been easier to craft a specific solution
link |
to a problem like chess than it has been
link |
to build a general learning system
link |
that could potentially do many things.
link |
Cause initially that system will be way worse
link |
than less efficient than the specialized system.
link |
So one of the interesting quirks of the human mind
link |
of this evolved system is that it appears to be conscious.
link |
This thing that we don't quite understand,
link |
but it seems very special is ability
link |
to have a subjective experience
link |
that it feels like something to eat a cookie,
link |
the deliciousness of it or see a color
link |
and that kind of stuff.
link |
Do you think in order to solve intelligence,
link |
we also need to solve consciousness along the way?
link |
Do you think AGI systems need to have consciousness
link |
in order to be truly intelligent?
link |
Yeah, we thought about this a lot actually.
link |
And I think that my guess is that consciousness
link |
and intelligence are double dissociable.
link |
So you can have one without the other both ways.
link |
And I think you can see that with consciousness
link |
in that I think some animals and pets,
link |
if you have a pet dog or something like that,
link |
you can see some of the higher animals and dolphins,
link |
things like that have self awareness
link |
and are very sociable, seem to dream.
link |
A lot of the traits one would regard
link |
as being kind of conscious and self aware,
link |
but yet they're not that smart, right?
link |
So they're not that intelligent
link |
by say IQ standards or something like that.
link |
Yeah, it's also possible that our understanding
link |
of intelligence is flawed, like putting an IQ to it.
link |
Maybe the thing that a dog can do
link |
is actually gone very far along the path of intelligence
link |
and we humans are just able to play chess
link |
and maybe write poems.
link |
Right, but if we go back to the idea of AGI
link |
and general intelligence, dogs are very specialized, right?
link |
Most animals are pretty specialized.
link |
They can be amazing at what they do,
link |
but they're like kind of elite sports people or something,
link |
right, so they do one thing extremely well
link |
because their entire brain is optimized.
link |
They have somehow convinced the entirety
link |
of the human population to feed them and service them.
link |
So in some way they're controlling.
link |
Well, we co evolved to some crazy degree, right?
link |
Including the way the dogs even wag their tails
link |
and twitch their noses, right?
link |
We find inextricably cute.
link |
But I think you can also see intelligence on the other side.
link |
So systems like artificial systems
link |
that are amazingly smart at certain things
link |
like maybe playing go and chess and other things,
link |
but they don't feel at all in any shape or form conscious
link |
in the way that you do to me or I do to you.
link |
And I think actually building AI
link |
is these intelligent constructs
link |
is one of the best ways to explore
link |
the mystery of consciousness, to break it down
link |
because we're gonna have devices
link |
that are pretty smart at certain things
link |
or capable at certain things,
link |
but potentially won't have any semblance
link |
of self awareness or other things.
link |
And in fact, I would advocate if there's a choice,
link |
building systems in the first place,
link |
AI systems that are not conscious to begin with
link |
are just tools until we understand them better
link |
and the capabilities better.
link |
So on that topic, just not as the CEO of DeepMind,
link |
just as a human being, let me ask you
link |
about this one particular anecdotal evidence
link |
of the Google engineer who made a comment
link |
or believed that there's some aspect of a language model,
link |
the Lambda language model that exhibited sentience.
link |
So you said you believe there might be a responsibility
link |
to build systems that are not sentient.
link |
And this experience of a particular engineer,
link |
I think I'd love to get your general opinion
link |
on this kind of thing, but I think it will happen
link |
more and more and more, which not when engineers,
link |
but when people out there that don't have
link |
an engineering background start interacting
link |
with increasingly intelligent systems,
link |
we anthropomorphize them.
link |
They start to have deep, impactful interactions with us
link |
in a way that we miss them when they're gone.
link |
And we sure as heck feel like they're living entities,
link |
self aware entities, and maybe even
link |
we project sentience onto them.
link |
So what's your thought about this particular system?
link |
Have you ever met a language model that's sentient?
link |
What do you make of the case of when you kind of feel
link |
that there's some elements of sentience to the system?
link |
Yeah, so this is an interesting question
link |
and obviously a very fundamental one.
link |
So the first thing to say is I think that none
link |
of the systems we have today, I would say,
link |
even have one iota of semblance
link |
of consciousness or sentience.
link |
That's my personal feeling interacting with them every day.
link |
So I think this way premature to be discussing
link |
what that engineer talked about.
link |
I think at the moment it's more of a projection
link |
of the way our own minds work,
link |
which is to see sort of purpose and direction
link |
in almost anything that we, you know,
link |
our brains are trained to interpret agency,
link |
basically in things, even inanimate things sometimes.
link |
And of course with a language system,
link |
because language is so fundamental to intelligence,
link |
that's going to be easy for us to anthropomorphize that.
link |
I mean, back in the day, even the first, you know,
link |
the dumbest sort of template chatbots ever,
link |
Eliza and the ilk of the original chatbots
link |
back in the sixties fooled some people
link |
under certain circumstances, right?
link |
It pretended to be a psychologist.
link |
So just basically rabbit back to you
link |
the same question you asked it back to you.
link |
And some people believe that.
link |
So I don't think we can, this is why I think
link |
the Turing test is a little bit flawed as a formal test
link |
because it depends on the sophistication of the judge,
link |
whether or not they are qualified to make that distinction.
link |
So I think we should talk to, you know,
link |
the top philosophers about this,
link |
people like Daniel Dennett and David Chalmers and others
link |
who've obviously thought deeply about consciousness.
link |
Of course, consciousness itself hasn't been well,
link |
there's no agreed definition.
link |
If I was to, you know, speculate about that, you know,
link |
I kind of, the working definition I like is
link |
it's the way information feels when it gets processed.
link |
I think maybe Max Tegmark came up with that.
link |
I don't know if it helps us get towards
link |
any more operational thing,
link |
but I think it's a nice way of viewing it.
link |
I think we can obviously see from neuroscience
link |
certain prerequisites that are required,
link |
like self awareness, I think is necessary,
link |
but not sufficient component.
link |
This idea of a self and other
link |
and set of coherent preferences
link |
that are coherent over time.
link |
You know, these things are maybe memory.
link |
These things are probably needed
link |
for a sentient or conscious being.
link |
But the reason, the difficult thing,
link |
I think for us when we get,
link |
and I think this is a really interesting
link |
philosophical debate is when we get closer to AGI
link |
and, you know, and much more powerful systems
link |
than we have today,
link |
how are we going to make this judgment?
link |
And one way, which is the Turing test
link |
is sort of a behavioral judgment,
link |
is the system exhibiting all the behaviors
link |
that a human sentient or a sentient being would exhibit?
link |
Is it answering the right questions?
link |
Is it saying the right things?
link |
Is it indistinguishable from a human?
link |
But I think there's a second thing
link |
that makes us as humans regard each other as sentient,
link |
Why do we think this?
link |
And I debated this with Daniel Dennett.
link |
And I think there's a second reason
link |
that's often overlooked,
link |
which is that we're running on the same substrate, right?
link |
So if we're exhibiting the same behavior,
link |
more or less as humans,
link |
and we're running on the same, you know,
link |
carbon based biological substrate,
link |
the squishy, you know, few pounds of flesh in our skulls,
link |
then the most parsimonious, I think, explanation
link |
is that you're feeling the same thing as I'm feeling, right?
link |
But we will never have that second part,
link |
the substrate equivalence with a machine, right?
link |
So we will have to only judge based on the behavior.
link |
And I think the substrate equivalence
link |
is a critical part of why we make assumptions
link |
that we're conscious.
link |
And in fact, even with animals, high level animals,
link |
why we think they might be,
link |
because they're exhibiting some of the behaviors
link |
we would expect from a sentient animal.
link |
And we know they're made of the same things,
link |
biological neurons.
link |
So we're gonna have to come up with explanations
link |
or models of the gap between substrate differences,
link |
between machines and humans
link |
to get anywhere beyond the behavioral.
link |
But to me, sort of the practical question
link |
is very interesting and very important.
link |
When you have millions, perhaps billions of people
link |
believing that you have a sentient AI,
link |
believing what that Google engineer believed,
link |
which I just see as an obvious, very near term future thing,
link |
certainly on the path to AGI,
link |
how does that change the world?
link |
What's the responsibility of the AI system
link |
to help those millions of people?
link |
And also what's the ethical thing?
link |
Because you can make a lot of people happy
link |
by creating a meaningful, deep experience
link |
with a system that's faking it before it makes it.
link |
And I don't, are we the right,
link |
who is to say what's the right thing to do?
link |
Should AI always be tools?
link |
Why are we constraining AI to always be tools
link |
as opposed to friends?
link |
Yeah, I think, well, I mean, these are fantastic questions
link |
and also critical ones.
link |
And we've been thinking about this
link |
since the start of DeepMind and before that,
link |
because we plan for success
link |
and however remote that looked like back in 2010.
link |
And we've always had sort of these ethical considerations
link |
as fundamental at DeepMind.
link |
And my current thinking on the language models
link |
and large models is they're not ready,
link |
we don't understand them well enough yet.
link |
And in terms of analysis tools and guard rails,
link |
what they can and can't do and so on,
link |
to deploy them at scale, because I think,
link |
there are big, still ethical questions
link |
like should an AI system always announce
link |
that it is an AI system to begin with?
link |
What do you do about answering those philosophical questions
link |
about the feelings people may have about AI systems,
link |
perhaps incorrectly attributed?
link |
So I think there's a whole bunch of research
link |
that needs to be done first to responsibly,
link |
before you can responsibly deploy these systems at scale.
link |
That will be at least be my current position.
link |
Over time, I'm very confident we'll have those tools
link |
like interpretability questions and analysis questions.
link |
And then with the ethical quandary,
link |
I think there it's important to look beyond just science.
link |
That's why I think philosophy, social sciences,
link |
even theology, other things like that come into it,
link |
where arts and humanities,
link |
what does it mean to be human and the spirit of being human
link |
and to enhance that and the human condition, right?
link |
And allow us to experience things
link |
we could never experience before
link |
and improve the overall human condition
link |
and humanity overall, get radical abundance,
link |
solve many scientific problems, solve disease.
link |
So this is the era I think, this is the amazing era
link |
I think we're heading into if we do it right.
link |
But we've got to be careful.
link |
We've already seen with things like social media,
link |
how dual use technologies can be misused by,
link |
firstly, by bad actors or naive actors or crazy actors,
link |
right, so there's that set of just the common
link |
or garden misuse of existing dual use technology.
link |
And then of course, there's an additional thing
link |
that has to be overcome with AI
link |
that eventually it may have its own agency.
link |
So it could be good or bad in and of itself.
link |
So I think these questions have to be approached
link |
very carefully using the scientific method, I would say,
link |
in terms of hypothesis generation, careful control testing,
link |
not live A, B testing out in the world,
link |
because with powerful technologies like AI,
link |
if something goes wrong, it may cause a lot of harm
link |
before you can fix it.
link |
It's not like an imaging app or game app
link |
where if something goes wrong, it's relatively easy to fix
link |
and the harm is relatively small.
link |
So I think it comes with the usual cliche of,
link |
like with a lot of power comes a lot of responsibility.
link |
And I think that's the case here with things like AI,
link |
given the enormous opportunity in front of us.
link |
And I think we need a lot of voices
link |
and as many inputs into things like the design
link |
of the systems and the values they should have
link |
and what goals should they be put to.
link |
I think as wide a group of voices as possible
link |
beyond just the technologists is needed to input into that
link |
and to have a say in that,
link |
especially when it comes to deployment of these systems,
link |
which is when the rubber really hits the road,
link |
it really affects the general person in the street
link |
rather than fundamental research.
link |
And that's why I say, I think as a first step,
link |
it would be better if we have the choice
link |
to build these systems as tools to give,
link |
and I'm not saying that they should never go beyond tools
link |
because of course the potential is there
link |
for it to go way beyond just tools.
link |
But I think that would be a good first step
link |
in order for us to allow us to carefully experiment
link |
and understand what these things can do.
link |
So the leap between tool, the sentient entity being
link |
is one we should take very careful of.
link |
Let me ask a dark personal question.
link |
So you're one of the most brilliant people
link |
in the AI community, you're also one of the most kind
link |
and if I may say sort of loved people in the community.
link |
That said, creation of a super intelligent AI system
link |
would be one of the most powerful things in the world,
link |
tools or otherwise.
link |
And again, as the old saying goes, power corrupts
link |
and absolute power corrupts absolutely.
link |
You are likely to be one of the people,
link |
I would say probably the most likely person
link |
to be in the control of such a system.
link |
Do you think about the corrupting nature of power
link |
when you talk about these kinds of systems
link |
that as all dictators and people have caused atrocities
link |
in the past, always think they're doing good,
link |
but they don't do good because the power
link |
has polluted their mind about what is good
link |
Do you think about this stuff
link |
or are we just focused on language model?
link |
No, I think about them all the time
link |
and I think what are the defenses against that?
link |
I think one thing is to remain very grounded
link |
and sort of humble, no matter what you do or achieve.
link |
And I try to do that, my best friends
link |
are still my set of friends
link |
from my undergraduate Cambridge days,
link |
my family's and friends are very important.
link |
I've always, I think trying to be a multidisciplinary person,
link |
it helps to keep you humble
link |
because no matter how good you are at one topic,
link |
someone will be better than you at that.
link |
And always relearning a new topic again from scratch
link |
is a new field is very humbling, right?
link |
So for me, that's been biology over the last five years,
link |
huge area topic and I just love doing that,
link |
but it helps to keep you grounded
link |
like it keeps you open minded.
link |
And then the other important thing
link |
is to have a really good, amazing set of people around you
link |
at your company or your organization
link |
who are also very ethical and grounded themselves
link |
and help to keep you that way.
link |
And then ultimately just to answer your question,
link |
I hope we're gonna be a big part of birthing AI
link |
and that being the greatest benefit to humanity
link |
of any tool or technology ever,
link |
and getting us into a world of radical abundance
link |
and curing diseases and solving many of the big challenges
link |
we have in front of us.
link |
And then ultimately help the ultimate flourishing
link |
of humanity to travel the stars
link |
and find those aliens if they are there.
link |
And if they're not there, find out why they're not there,
link |
what is going on here in the universe.
link |
This is all to come.
link |
And that's what I've always dreamed about.
link |
But I think AI is too big an idea.
link |
It's not going to be,
link |
there'll be a certain set of pioneers who get there first.
link |
I hope we're in the vanguard
link |
so we can influence how that goes.
link |
And I think it matters who builds,
link |
which cultures they come from and what values they have,
link |
the builders of AI systems.
link |
Cause I think even though the AI system
link |
is gonna learn for itself most of its knowledge,
link |
there'll be a residue in the system of the culture
link |
and the values of the creators of that system.
link |
And there's interesting questions
link |
to discuss about that geopolitically.
link |
Different cultures,
link |
we're in a more fragmented world than ever, unfortunately.
link |
I think in terms of global cooperation,
link |
we see that in things like climate
link |
where we can't seem to get our act together globally
link |
to cooperate on these pressing matters.
link |
I hope that will change over time.
link |
Perhaps if we get to an era of radical abundance,
link |
we don't have to be so competitive anymore.
link |
Maybe we can be more cooperative
link |
if resources aren't so scarce.
link |
It's true that in terms of power corrupting
link |
and leading to destructive things,
link |
it seems that some of the atrocities of the past happen
link |
when there's a significant constraint on resources.
link |
I think that's the first thing.
link |
I don't think that's enough.
link |
I think scarcity is one thing that's led to competition,
link |
sort of zero sum game thinking.
link |
I would like us to all be in a positive sum world.
link |
And I think for that, you have to remove scarcity.
link |
I don't think that's enough, unfortunately,
link |
to get world peace
link |
because there's also other corrupting things
link |
like wanting power over people and this kind of stuff,
link |
which is not necessarily satisfied by just abundance.
link |
But I think it will help.
link |
But I think ultimately, AI is not gonna be run
link |
by any one person or one organization.
link |
I think it should belong to the world, belong to humanity.
link |
And I think there'll be many ways this will happen.
link |
And ultimately, everybody should have a say in that.
link |
Do you have advice for young people in high school,
link |
in college, maybe if they're interested in AI
link |
or interested in having a big impact on the world,
link |
what they should do to have a career they can be proud of
link |
or to have a life they can be proud of?
link |
I love giving talks to the next generation.
link |
What I say to them is actually two things.
link |
I think the most important things to learn about
link |
and to find out about when you're young
link |
is what are your true passions is first of all,
link |
One is find your true passions.
link |
And I think you can do that by,
link |
the way to do that is to explore as many things as possible
link |
when you're young and you have the time
link |
and you can take those risks.
link |
I would also encourage people to look at
link |
finding the connections between things in a unique way.
link |
I think that's a really great way to find a passion.
link |
Second thing I would say, advise is know yourself.
link |
So spend a lot of time understanding
link |
how you work best.
link |
Like what are the optimal times to work?
link |
What are the optimal ways that you study?
link |
What are your, how do you deal with pressure?
link |
Sort of test yourself in various scenarios
link |
and try and improve your weaknesses,
link |
but also find out what your unique skills and strengths are
link |
and then hone those.
link |
So then that's what will be your super value
link |
in the world later on.
link |
And if you can then combine those two things
link |
and find passions that you're genuinely excited about
link |
that intersect with what your unique strong skills are,
link |
then you're onto something incredible
link |
and I think you can make a huge difference in the world.
link |
So let me ask about know yourself.
link |
Quick questions about day in the life, the perfect day,
link |
the perfect productive day in the life of Demis's Hub.
link |
Maybe these days you're, there's a lot involved.
link |
So maybe a slightly younger Demis's Hub
link |
where you could focus on a single project maybe.
link |
How early do you wake up?
link |
Are you a night owl?
link |
Do you wake up early in the morning?
link |
What are some interesting habits?
link |
How many dozens of cups of coffees do you drink a day?
link |
What's the computer that you use?
link |
What kind of keyboard?
link |
Are we talking Emacs Vim
link |
or are we talking something more modern?
link |
So there's a bunch of those questions.
link |
So maybe day in the life, what's the perfect day involved?
link |
Well, these days it's quite different
link |
from say 10, 20 years ago.
link |
Back 10, 20 years ago, it would have been
link |
a whole day of research, individual research or programming,
link |
doing some experiment, neuroscience,
link |
computer science experiment,
link |
reading lots of research papers.
link |
And then perhaps at nighttime,
link |
reading science fiction books or playing some games.
link |
But lots of focus, so like deep focused work
link |
on whether it's programming or reading research papers.
link |
Yes, so that would be lots of deep focus work.
link |
These days for the last sort of, I guess, five to 10 years,
link |
I've actually got quite a structure
link |
that works very well for me now,
link |
which is that I'm a complete night owl, always have been.
link |
So I optimize for that.
link |
So I'll basically do a normal day's work,
link |
get into work about 11 o clock
link |
and sort of do work to about seven in the office.
link |
And I will arrange back to back meetings
link |
for the entire time of that.
link |
And with as many, meet as many people as possible.
link |
So that's my collaboration management part of the day.
link |
Then I go home, spend time with the family and friends,
link |
have dinner, relax a little bit.
link |
And then I start a second day of work.
link |
I call it my second day of work around 10 p.m., 11 p.m.
link |
And that's the time to about the small hours of the morning,
link |
four or five in the morning, where I will do my thinking
link |
and reading and research, writing research papers.
link |
Sadly, I don't have time to code anymore,
link |
but it's not efficient to do that these days,
link |
given the amount of time I have.
link |
But that's when I do, you know,
link |
maybe do the long kind of stretches
link |
of thinking and planning.
link |
And then probably, you know, using email, other things,
link |
I would set, I would fire off a lot of things to my team
link |
to deal with the next morning.
link |
But actually thinking about this overnight,
link |
we should go for this project
link |
or arrange this meeting the next day.
link |
When you're thinking through a problem,
link |
are you talking about a sheet of paper with a pen?
link |
Is there some structured process?
link |
I still like pencil and paper best for working out things,
link |
but these days it's just so efficient
link |
to read research papers just on the screen.
link |
I still often print them out, actually.
link |
I still prefer to mark out things.
link |
And I find it goes into the brain better
link |
and sticks in the brain better
link |
when you're still using physical pen and pencil and paper.
link |
So you take notes with the...
link |
I have lots of notes, electronic ones,
link |
and also whole stacks of notebooks that I use at home, yeah.
link |
On some of these most challenging next steps, for example,
link |
stuff none of us know about that you're working on,
link |
there's some deep thinking required there, right?
link |
Like what is the right problem?
link |
What is the right approach?
link |
Because you're gonna have to invest a huge amount of time
link |
for the whole team.
link |
They're going to have to pursue this thing.
link |
What's the right way to do it?
link |
Is RL gonna work here or not?
link |
What's the right thing to try?
link |
What's the right benchmark to use?
link |
Do we need to construct a benchmark from scratch?
link |
All those kinds of things.
link |
So I think of all those kinds of things
link |
in the nighttime phase, but also much more,
link |
I find I've always found the quiet hours of the morning
link |
when everyone's asleep, it's super quiet outside.
link |
It's the golden hours,
link |
like between one and three in the morning.
link |
Put some music on, some inspiring music on,
link |
and then think these deep thoughts.
link |
So that's when I would read my philosophy books
link |
and Spinoza's, my recent favorite can, all these things.
link |
And I read about a great scientist of history,
link |
how they did things, how they thought things.
link |
So that's when you do all your creative,
link |
that's when I do all my creative thinking.
link |
And it's good, I think people recommend
link |
you do your sort of creative thinking in one block.
link |
And the way I organize the day,
link |
that way I don't get interrupted.
link |
There's obviously no one else is up at those times.
link |
So I can go, I can sort of get super deep
link |
and super into flow.
link |
The other nice thing about doing it nighttime wise
link |
is if I'm really onto something
link |
or I've got really deep into something,
link |
I can choose to extend it
link |
and I'll go into six in the morning, whatever.
link |
And then I'll just pay for it the next day.
link |
So I'll be a bit tired and I won't be my best,
link |
I can decide looking at my schedule the next day
link |
and given where I'm at with this particular thought
link |
or creative idea that I'm gonna pay that cost the next day.
link |
So I think that's more flexible than morning people
link |
who do that, they get up at four in the morning.
link |
They can also do those golden hours then,
link |
but then their start of their scheduled day
link |
starts at breakfast, 8 a.m.,
link |
whatever they have their first meeting.
link |
And then it's hard, you have to reschedule a day
link |
if you're in flow.
link |
So I don't have to do that.
link |
So that could be a true special thread of thoughts
link |
that you're too passionate about.
link |
This is where some of the greatest ideas
link |
could potentially come is when you just lose yourself
link |
late into the night.
link |
And for the meetings, I mean, you're loading in
link |
really hard problems in a very short amount of time.
link |
So you have to do some kind of first principles thinking
link |
here, it's like, what's the problem?
link |
What's the state of things?
link |
What's the right next steps?
link |
You have to get really good at context switching,
link |
which is one of the hardest things,
link |
because especially as we do so many things,
link |
if you include all the scientific things we do,
link |
scientific fields we're working in,
link |
these are complex fields in themselves.
link |
And you have to sort of keep abreast of that.
link |
I've always been a sort of generalist in a way.
link |
And that's actually what happened in my games career
link |
One of the reasons I stopped playing chess
link |
was because I got into computers,
link |
but also I started realizing there were many other
link |
great games out there to play too.
link |
So I've always been that way inclined, multidisciplinary.
link |
And there's too many interesting things in the world
link |
to spend all your time just on one thing.
link |
So you mentioned Spinoza, gotta ask the big, ridiculously
link |
big question about life.
link |
What do you think is the meaning of this whole thing?
link |
Why are we humans here?
link |
You've already mentioned that perhaps the universe
link |
Is that why you think we're here?
link |
To understand how the universe works?
link |
Yeah, I think my answer to that would be,
link |
and at least the life I'm living,
link |
is to gain and understand the knowledge,
link |
to gain knowledge and understand the universe.
link |
That's what I think, I can't see any higher purpose
link |
than that if you think back to the classical Greeks,
link |
the virtue of gaining knowledge.
link |
It's, I think it's one of the few true virtues
link |
is to understand the world around us
link |
and the context and humanity better.
link |
And I think if you do that, you become more compassionate
link |
and more understanding yourself and more tolerant
link |
and all these, I think all these other things
link |
may flow from that.
link |
And to me, understanding the nature of reality,
link |
that is the biggest question.
link |
What is going on here is sometimes the colloquial way I say.
link |
What is really going on here?
link |
It's so mysterious.
link |
I feel like we're in some huge puzzle.
link |
And it's, but the world is also seems to be,
link |
the universe seems to be structured in a way.
link |
You know, why is it structured in a way
link |
that science is even possible?
link |
That, you know, methods, the scientific method works,
link |
things are repeatable.
link |
It feels like it's almost structured in a way
link |
to be conducive to gaining knowledge.
link |
So I feel like, and you know,
link |
why should computers be even possible?
link |
Wasn't that amazing that computational electronic devices
link |
can be possible, and they're made of sand,
link |
our most common element that we have,
link |
you know, silicon on the Earth's crust.
link |
It could have been made of diamond or something,
link |
then we would have only had one computer.
link |
So a lot of things are kind of slightly suspicious to me.
link |
It sure as heck sounds, this puzzle sure as heck sounds
link |
like something we talked about earlier,
link |
what it takes to design a game that's really fun to play
link |
for prolonged periods of time.
link |
And it does seem like this puzzle, like you mentioned,
link |
the more you learn about it,
link |
the more you realize how little you know.
link |
So it humbles you, but excites you
link |
by the possibility of learning more.
link |
It's one heck of a puzzle we got going on here.
link |
So like I mentioned, of all the people in the world,
link |
you're very likely to be the one who creates the AGI system
link |
that achieves human level intelligence and goes beyond it.
link |
So if you got a chance and very well,
link |
you could be the person that goes into the room
link |
with the system and have a conversation.
link |
Maybe you only get to ask one question.
link |
If you do, what question would you ask her?
link |
I would probably ask, what is the true nature of reality?
link |
I think that's the question.
link |
I don't know if I'd understand the answer
link |
because maybe it would be 42 or something like that,
link |
but that's the question I would ask.
link |
And then there'll be a deep sigh from the systems,
link |
like, all right, how do I explain to this human?
link |
All right, let me, I don't have time to explain.
link |
Maybe I'll draw you a picture that it is.
link |
I mean, how do you even begin to answer that question?
link |
Well, I think it would.
link |
What would you think the answer could possibly look like?
link |
I think it could start looking like
link |
more fundamental explanations of physics
link |
would be the beginning.
link |
More careful specification of that,
link |
taking you, walking us through by the hand
link |
as to what one would do to maybe prove those things out.
link |
Maybe giving you glimpses of what things
link |
you totally miss in the physics of today.
link |
Just here's glimpses of, no, like there's a much,
link |
a much more elaborate world
link |
or a much simpler world or something.
link |
A much deeper, maybe simpler explanation of things,
link |
right, than the standard model of physics,
link |
which we know doesn't work, but we still keep adding to.
link |
So, and that's how I think the beginning
link |
of an explanation would look.
link |
And it would start encompassing many of the mysteries
link |
that we have wondered about for thousands of years,
link |
like consciousness, dreaming, life, and gravity,
link |
all of these things.
link |
Yeah, giving us glimpses of explanations for those things.
link |
Well, Damasir, one of the special human beings
link |
in this giant puzzle of ours,
link |
and it's a huge honor that you would take a pause
link |
from the bigger puzzle to solve this small puzzle
link |
of a conversation with me today.
link |
It's truly an honor and a pleasure.
link |
Thank you so much.
link |
Thank you, I really enjoyed it.
link |
Thanks for listening to this conversation
link |
with Damas Ashabis.
link |
To support this podcast,
link |
please check out our sponsors in the description.
link |
And now, let me leave you with some words
link |
from Edgar Dykstra.
link |
Computer science is no more about computers
link |
than astronomy is about telescopes.
link |
Thank you for listening, and hope to see you next time.