Kind: captions
Language: en
It's hard for us humans to make any kind
of clean predictions about highly
nonlinear dynamical systems. But again
to your point, we might be very
surprised what classical learning
systems might be able to do about even
fluid.
>> Yes, exactly. I mean fluid dynamics,
Navia Stokes equations, these are
traditionally thought of as very very
difficult intractable problems to do on
classical systems. They take enormous
amounts of compute, you know, weather
prediction systems, you know, these kind
of things all involve fluid dynamics
calculations. But again, if you look at
something like VO, our video generation
model, it can model liquids quite well,
surprisingly well, and materials,
specular lighting. I love the ones
where, you know, there's there's people
have generated videos where there's like
clear liquids going through hydraulic
presses and then being squeezed out. I I
used to write uh physics engines and
graphics engines and in my early days in
gaming. And I know it's just so
painstakingly hard to build programs
that can do that. And yet somehow these
systems are, you know, reverse
engineering from just watching YouTube
videos. So presumably what's happening
is it's extracting some underlying
structure around how these materials
behave. So perhaps there is some kind of
lower dimensional manifold that can be
learned if we actually fully understood
what's going on under the hood. That's
maybe, you know, maybe true of most of
reality.
The following is a conversation with
Demis Hassabis, his second time on the
podcast. He is the leader of Google Deep
Mind and is now a Nobel Prize winner.
Demis is one of the most brilliant and
fascinating minds in the world today,
working on understanding and building
intelligence and exploring the big
mysteries of our universe.
This was truly an honor and a pleasure
for me. This is the Lex Freedman
podcast. To support it, please check out
our sponsors in the description and
consider subscribing to this channel.
And now, dear friends, here's Deus
Hassavas.
In your Nobel Prize lecture, you propose
what I think is a super interesting
conjecture that quote any pattern that
can be generated or found in nature can
be efficiently discovered and modeled by
a classical learning algorithm. What
kind of patterns of systems might be
included in that? Biology, chemistry,
physics, maybe cosmology,
>> neuroscience. What what are we talking
about?
>> Sure. Well, look, I I felt that it's
sort of a tradition, I think, of Nobel
Prize lectures that you're supposed to
be a little bit provocative and I wanted
to follow that tradition. What I was
talking about there is if you take a
step back and you look at um all the
work that we've done especially with the
alpha x projects so I'm thinking alpho
of course alpha fold what they really
are is we're building models of very
combinatorily highdimensional spaces
that you know if you tried to brute
force a solution find the best move and
go or find the the exact shape of a
protein and if you enumerated all the
possibilities you there wouldn't be
enough time in the in the you know the
time of the universe. So you have to do
something much smarter and what we did
in both cases was build models of those
environments. Um and that guided the
search in a in a smart way and that
makes it tractable. So if you think
about protein folding which is obviously
a natural system you know why should
that be possible? How does physics do
that? You know proteins fold in
milliseconds in our bodies. So somehow
physics solves this problem that we've
now also solved computationally. And I
think the reason that's possible is that
in nature, natural systems have
structure because they were subject to
evolutionary processes that that shape
them. And if that's true, then you can
maybe learn uh uh what that structure
is. So this perspective, I think, is
really interesting one. You've hinted it
at it, which is almost like uh crudely
stated. Anything that can be evolved can
be efficiently modeled. Think there's
some truth to that. Yeah, I sometimes
call it survival of the stablest or
something like that because you know
it's it's of course there's evolution
for life uh living things but there's
also you know if you think about
geological time so the shape of
mountains that's been shaped by
weathering processes right over
thousands of years but then you can even
take it cosmological the orbits of
planets the um shapes of asteroids these
have all been survived kind of processes
that have acted on them many many times
so if that's true then there should some
sort of pattern um that you can kind of
reverse learn and uh a kind of manifold
really that helps you uh uh search to
the right solution to the right shape um
and actually allow you to predict things
about it uh in an efficient way because
it's not a random pattern right so um it
may not be possible for for man-made
things or abstract things like
factorizing large numbers because unless
there's patterns in the number space
which there might be but if there's not
and it's uniform then there's no pattern
to learn there's no model to learn that
will help you search. So you have to do
brute force. So in that case you you
know you maybe need a quantum computer
something like this. But in most things
in nature that we're interested in uh
are not like that. They have structure
um that evolved for a reason and
survived over time. And if that's true I
think that's potentially learnable by a
neural network.
>> It's like nature is doing a search
process and it's so fascinating that
it's in that search process is creating
systems that could be efficiently
modeled. That's right. Yeah.
>> So interesting.
>> So they can be efficiently rediscovered
or recovered um because nature is not
random, right? These everything that we
see around us, including like the
elements that are more stable, all of
those things, they're subject to um some
kind of selection process pressure. Do
you think because you're also a fan of
theoretical computer science and
complexity, do you think we can come up
with a kind of complexity class like a
complexity zoo type of class where maybe
it's the set of learnable systems, the
set of learnable natural systems, lns.
>> Yeah,
>> this is a deis new class of systems that
could be actually learnable by classical
systems in this kind of way. Natural
systems that can be uh modeled
efficiently. Yeah, I mean I' I've always
been fascinated by the P= MP question
and what is modelable by classical
systems I non-quantum systems you know
cheuring machines in effect and that's
exactly what I'm working on actually in
kind of my few moments of spare time
with a few colleagues about is should
there be you know maybe a new class of
problem that is solvable by this type of
neural network process and kind of
mapped on to these natural systems so
you know the things that exist in
physics
and have structure. So I think that
could be a very interesting uh new way
of thinking about it. And it sort of
fits with the way I think about physics
in general which is that you know I
think information is primary.
Information is the most sort of
fundamental unit of the universe more
fundamental than energy and matter. I
think they can all be converted into
each other but I think of the universe
as a kind of informationational system.
>> So when you think of the universe as
anformational system then the P= NP
question is a is a physics question.
>> That's right. And it's a question that
can help us actually solve the entirety
of this whole thing going on.
>> Yeah, I think it's one of the most uh
fundamental questions actually if you
think of physics asformational
uh and and the answer to that I think is
going to be you know very enlightening.
more specific to the PNNP question.
This again, some of the stuff we're
saying is kind of crazy right now. Just
like the Christian Edinson Nobel Prize
speech controversial thing that he said
sounded crazy and then you went and got
a Nobel Prize for this with John Jumper
solved the problem. So, let me let me
just stick to the P equals MP. Do you
think there's something
in this thing we're talking about that
could be shown if you can do something
like uh polomial time or constant time
compute ahead of time and construct this
gigantic model then you can solve some
of these extremely difficult problems in
a theoretical computer science kind of
way.
>> Yeah, I think that there are actually a
huge class of problems that could be
couched in this way. the way we did
alpha go and the way we did alpha fold
where you know you you model what the
dynamics of the system is the the the
the properties of that system the
environment that you're trying to
understand and then that makes the
search for the solution or the
prediction of the next step efficient
basically polomial time so tractable by
a uh classical system uh which a neural
network is it runs on normal computers
right classical computers uh chewing
machines in effect and um I think it's
one of the most interesting questions
there is is how far can that paradigm
go? You know, I think we've proven uh
and the AI community in general that
classical systems, cheuring machines can
go a lot further than we previously
thought. You know, they can do things
like model the structures of proteins
and play go to better than world
champion level. And uh you know a lot of
people would have thought maybe 10 20
years ago that was decades away or maybe
you would need some sort of quantum
machines to to quantum systems to be
able to do things like protein folding.
And so I think we haven't really uh even
sort of scratched the surface yet of
what uh classical systems socalled uh uh
could do. And of course AGI being built
on a on a neural network system on top
of a neural network system on top of a
classical computer would be the ultimate
expression of that. And I think the
limit you know the the what what the
bounds of that kind of system what it
can do it's a very interesting question
and and and directly speaks to the P
equals MP question.
What do you think again hypothetical
might be outside of this maybe
emergent phenomena like if you look at
cellular automa some of the you have
extremely simple systems and then some
complexity emerges yes
>> maybe that would be outside or even
would you guess even that might be
amendable
>> to efficient modeling by a classical
machine
>> yeah I think those systems would be
right on the boundary right so um I
think most emergent systems cellular
automter things that could be modelable
by a classical system. You just sort of
do a forward simulation of it and it
probably be efficient enough. Um, of
course there's the question of things
like chaotic systems where the initial
conditions really matter and then you
get to some, you know, uncorrelated end
state those could be difficult to model.
So I think these are kind of the open
questions. But I think when you step
back and look at what we've done with
the systems and the and the problems
that we've solved and then you look at
things like V3 on like video generation
sort of rendering physics and lighting
and things like that, you know, really
core fundamental things in physics. Um
it's pretty interesting. I think it's
telling us something quite fundamental
about how the universe is structured in
my opinion. Um so you know in a way
that's what I want to build AGI for is
to help uh us uh as scientists answer
these questions uh like p=mp.
>> Yeah I think we might be continuously
surprised about what is modelable by
classical computers. I mean alpha fold 3
on the interaction side is surprising
that you can make any kind of progress
on that direction. Alpha genome is
surprising that you can map the genetic
code to the function kind of playing
with the emergent kind of phenomena. You
think there's so many combinatorial
options that and then here you go. You
can find the kernel that is efficiently
modeled.
>> Yes. Because there's some structure
there's some landscape you know in the
energy landscape or whatever it is that
you can follow some gradient you can
follow. And of course what neural
networks are very good at is following
gradients. And so if there's one to
follow and object and you can specify
the objective function correctly you
know you don't have to deal with all
that complexity which I think is how we
maybe have naively thought about it for
decades those problems if you just
enumerate all the possibilities it looks
totally intractable and there's many
many problems like that and then you
think well it's like 10^ the 300
possible protein structures uh it's 10^
theund you know 70 possible go positions
all of these are way more than atoms in
the universe so how could one possibly
find the the right solution or predict
the next step and and it but it turns
out that it is possible and of course
reality nature does do it right proteins
do fold so that that gives you
confidence that there must be if we
understood how physics was doing that uh
in a sense uh then and we could mimic
that process I model that process uh it
should be possible on our classical
systems is is is basically what the
conjecture is about
>> and of course there's nonlinear
dynamical systems, highly nonlinear
dynamical systems, everything involving
fluid.
>> Yes.
>> Right.
>> You know, I recently had a conversation
with Terrence Ta who mathematically
uh it contends with a very difficult
aspect of systems that have some
singularities in them that break the
mathematics and it's just hard for us
humans to make any kind of clean
predictions about highly nonlinear
dynamical systems. But again to your
point we might be very surprised what
classical learning systems might be able
to do about even fluid. Yes, exactly. I
mean fluid dynamics, Navia Stokes
equations, these are traditionally
thought of as very very difficult
intractable kind of problems to do on
classical systems. They take enormous
amounts of compute, you know, weather
prediction systems, you know, these kind
of things all involve fluid dynamics
calculations. And um but again, if you
look at something like VO, our video
generation model, it can model liquids
quite well, surprisingly well, and
materials, specular lighting. I love the
ones where, you know, there's there's
people have generated videos where
there's like clear liquids going through
hydraulic presses and then being
squeezed out. I I used to write uh
physics engines and graphics engines and
in my early days in gaming and I know
it's just so painstakingly hard to build
programs that can do that and yet
somehow these systems are, you know,
reverse engineering from just watching
YouTube videos. So presumably what's
happening is it's extracting some
underlying structure around how these
materials behave. So perhaps there is
some kind of lower dimensional manifold
that can be learned if we actually fully
understood what's going on under the
hood. That's maybe you know maybe true
of most of reality.
>> Yeah. I've been continuously precisely
by this aspect of V3. I think a lot of
people highlight different aspects,
including the comedic and the meme and
all that kind of stuff. And then the
ultra realistic ability to capture
humans in a really nice way that's
compelling and feels close to reality
and then combine that with native audio.
All of those are marvelous things about
V3. But the exactly the thing you're
mentioning, which is the physics.
>> Yeah,
>> it's not perfect, but it's pretty damn
good. And then the really interesting
scientific question is what is it
understanding about our world
>> in order to be able to do that because
of the cynical take with diffusion
models there's no way it understands
anything
>> but it seem I mean I don't think you can
generate that kind of video without
understanding and then our own
philosophical notion what it means to
understand then is like brought to the
surface like do to what degree do you
think V3 understands our world? I think
to the extent that it can predict the
next frames you know in a coherent way
that's some that is a form you know of
understanding right not in the
anthropomorphic version of you know it's
not some kind of deep philosophical
understanding of what's going on I don't
think these systems have that but they
they certainly have uh modeled enough of
the dynamics you know put it that way
that they can pretty accurately generate
whatever it is 8 seconds of consistent
video that by eye at least you know at a
glance is quite hard to distinguish what
the issues are and imagine that in two
or three more years time. That's the
thing I'm thinking about and how
incredible that there will look uh given
where we've come from, you know, the
early versions of that uh one or two
years ago. And so, um the rate of
progress is incredible. And I think um
I'm like you is like a lot of people
love all of the the the the standup
comedians and the the that actually
captures a lot of human dynamics very
well and and body language, but actually
the thing I'm most impressed with and
fascinated by is the physics behavior,
the lighting and materials and liquids.
And it's pretty amazing that it can do
that. And I think that shows that it has
some notion of at least intuitive
physics, right? um how things are
supposed to work uh intuitively maybe
the way that uh a human child would
understand physics right as opposed to a
you know a PhD student really uh being
able to unpack all the equations it's
more of an intuitive physics
understanding
>> well that intuitive
physics understanding that's the base
layer that's the thing people sometimes
call like common sense like it it really
understands something I think that
really surprised a lot of people it
blows my mind that
>> I just didn't think it would be possible
to generate that level of realism
without understanding.
>> You there's this notion that you can
only understand the physical world by
having an embodied AI system, a robot
that interacts with that world. That's
the only way to construct an
understanding of that world. But V3 is
directly challenging that it feels like
>> yes and it's very interesting you know
even if we if you were to ask me 5 10
years ago I would have said even though
I was immersed in all of this I would
have said well yeah you probably need to
understand intuitive physics you know
like if I push this off the table this
glass it will maybe shatter you know um
and and the liquid will spill out right
so we know all of these things but I
thought that you know and there's a lot
of theories in neuroscience it's called
action in perception where you know you
you need to act in the world to really
truly perceive it in a deep way. And
there was a lot of theories about you
need embodied intelligence or robotics
or something or maybe at least simulated
action uh so that you would understand
things like intuitive physics. But it
seems like um you can understand it
through passive observation which is
pretty surprising to me and and again I
think hints at something underlying
about the nature of uh reality in in my
opinion beyond um just the you know the
cool videos that it generates. Um and
and of course there's next stages is
maybe even making those videos
interactive. So uh one can actually step
into them and move around them. Um which
would be really mind-blowing especially
given my games background. So you can
imagine and then and then I think you
know you're we're starting to get
towards what I would call a world model
a model of how the world works the
mechanics of the world the physics of
the world and the things in that world.
And of course that's what you would need
for a true AGI system.
>> I have to talk to you about video games.
So, you were being a bit trolly. I I
think you're you're having more and more
fun on Twitter on X, which is great to
see. So, guy named Jimmy Apples tweeted,
"Let me play a video game of my V3
videos already. Uh, Google cooked so
good playable world models when spelled
we n question mark." Uh, and then you
quote tweeted that with, "Now, wouldn't
that be something?" So, how how hard is
it to build game worlds with AI? Maybe
can you look out into the future
uh of video games 5 10 years out? What
do you think that looks like?
>> Well, games were my first love really
and doing AI for games was the first
thing I did professionally in my teenage
years and and was the first major AI
systems that I built and uh I always
want to have I want to scratch that itch
one day and come back to that. though,
you know, and I will do, I think, and um
I think I'd sort of dream about, you
know, what would I have done back in the
'90s if I'd had access to the kind of AI
systems we have today? And I think you
could build absolutely mind-blowing
games. Um, and I think the next stage is
I always used to love making all the
games I've made are openw world games.
So, they're games where there's a
simulation and then there's AI
characters and then the player uh
interacts with that simulation and the
simulation adapts to the way the player
plays. And I always thought they were
the coolest games because uh so games
like theme park that I worked on where
everybody's game experience would be
unique to them, right? Because you're
kind of co-creating the game, right? Uh
we set up the parameters, we set up
initial conditions, and then you as the
player immersed in it, and then you are
co-creating it with the with the
simulation. But of course, it's very
hard to program open world games. you
know, you've got to be able to create uh
content whichever direction the player
goes in and you want it to be compelling
no matter what the player chooses. Um,
and so it was always quite difficult to
build uh things like cellular automter
actually type of those kind of classical
systems which created some emergent
behavior. Um, but they're always a
little bit fragile, a little bit
limited. Now we're maybe on the cusp in
the next few years, 5 10 years of having
AI systems that can truly create around
your imagination. um can nar sort of
dynamically change the story and
storytell the narrative around uh and
make it dramatic no matter what you end
up choosing. So it's like the ultimate
choose your own adventure sort of game.
And uh you know I think maybe we're
within reach if you think of a kind of
interactive version of VO uh and then
wind that forward 5 to 10 years and you
know imagine how good it's going to be.
>> Yeah. So you said a lot of super
interesting stuff there. So one the open
world
built into that is a deep
personalization the way you've described
it.
>> So it's not just that it's open world
like you can open any door and there'll
be something there. It's that the choice
of which door you open
>> in an unconstrained way defines the
worlds you see. So some games try to do
that to give you choice. Yes. But it's
really just an illusion of choice
because
>> the only uh like like Stanley Parable
game I recently played. It's it's it's
really there's a couple of doors and it
really just takes you down a narrative.
Stanley Parable is a great video game I
recommend people play that kind of uh in
a meta way uh mocks the illusion of
choice and there's philosophical notions
of free will and so on. But uh I do like
one of my favorite games of Elder
Scrolls is Daggerfall. I believe that
they really played with a like random
generation of the dungeons.
>> Yeah.
>> Of you can step in and they give you
this feeling of an open world and there
you mentioned interactivity. You don't
need to interact. That's a first step
cuz you don't need to interact that
much. You just when you open the door,
whatever you see is randomly generated
for you.
>> Yeah. And that's already an incredible
experience because you might be the only
person to ever see that.
>> Yeah. Exactly. And and so but what you'd
like is a little bit better than just
sort of a random generation, right? So
you'd like uh and and also better than a
simple AB hardcoded choice, right?
That's not really uh open world, right?
As you say, it's just giving you the
illusion of choice. What you want to be
able to do is is potentially anything in
that game environment. Um, and I think
the only way you can do that is to have
uh generated systems, systems that uh
will generate that on the fly. Of
course, you can't create infinite
amounts of game assets, right? It's
expensive enough already how AAA games
are made today. And that was obvious to
to us back in the '9s when I was working
on all these games. I think maybe Black
and White uh was the game that I worked
on, early stages of that that had the
still probably the best AI learning AI
in it. It was an early reinforcement
learning system that you, you know, you
were, you were looking after this
mythical creature and growing it and
nurturing it and depending how you
treated it, it would treat the villagers
in that world in the same way. So if you
were mean to it, it would be mean. If
you were good, it would be protective.
And so it was really a reflection of the
way you played it. So actually all of
the uh I've been working on sort of
simulations and AI through the medium of
games at the beginning of my career and
and really the whole of what I do today
is still a follow on from uh those early
more hardcoded ways of doing the AI to
now you know fully general learning
systems that that are trying to achieve
the same thing.
>> Yeah, it's been uh interesting,
hilarious, and uh fun to watch you and
Elon obviously itching to create games
because you're both gamers. And one of
the sad aspects of your uh incredible
success in so many domains of science
like serious adult stuff.
>> Yeah.
>> That you might not have time to really
create a game. You might end up creating
the tooling that others would create the
game. You have to watch
>> other others create the thing you've
always dreamed of. Do you think it's
possible you can somehow in your
extremely busy schedule actually find
time to create something like black and
white? some some an actual video game
where like you could make the childhood
dream come become reality.
>> You know, there's two things way to
think about that is maybe with vibe
coding as it gets better and there's a
possibility that I could, you know, one
could do that actually in in your spare
time. So, I'm quite excited about that
as a as that would be my project if if I
got the time to do some vibe coding. Um
I'm actually itching to do that. And
then the other thing is, you know, maybe
it's a sbatical after agi has been
safely stewarded into the world and
delivered into the world. You know, that
and then working on my physics theory as
we talked about at the beginning. Those
would be the two my my two post AGI
projects. Let's call it that way.
>> I I would love to see which game post
AGI which you choose. Solving uh the the
problem that some of the smartest people
in human history contended with, you
know, P equals MP
or creating a cool video. Yeah. Well,
but they might but in my world they'd be
related because it would be an openw
world simulated game uh as realistic as
possible. So, you know what what is what
is the universe? That's that's that's
speaking to the same question, right?
NPL MP. I think all these things are
related, at least in my mind. I mean in
a really serious way like video games
sometimes are looked down upon as just
this fun side activity but especially as
AI does more and more of the difficult
uh boring tasks something we in in
modern world call work.
You know video games is the thing in
which we may find meaning in which we
may find like what to do with our time.
You could create incredibly rich,
meaningful experiences. Like that's what
human life is. And then in video games,
you can create more sophisticated,
more diverse
ways of living,
>> right?
>> I think so. I mean, those of us who love
games, and I still do, is is is
um you know, it's almost can let your
imagination run wild, right? Like I I
used to love games um and working on
games. so much because it's the fusion
especially in the '9s and early 2000s
the sort of golden era maybe the 80s of
of of game of the games industry and it
was all being discovered new genres were
being discovered we weren't just making
games we felt we were we were creating a
new entertainment medium that never
existed before especially with these
open world games and simulation games
where you were co-create you as the
player were co-creating the story
there's no other media uh entertainment
media where you do that where you as the
audience actually co-create the the
story and of course Now with multiplayer
games as well, it can be a very social
activity and can explore all kinds of
interesting worlds in that. But on the
other hand, you know, it's very
important to um also enjoy and
experience uh the physical world. But
the question is then, you know, I think
we're going to have to kind of confront
the question again of what is the
fundamental nature of reality? uh what
is the going to be the difference
between these increasingly realistic
simulations and uh multiplayer ones and
emergent um and what we do in the real
world.
>> Yeah, there's clearly a huge amount of
value to experiencing the real world
nature. There's also a huge amount of
value in experiencing other humans
directly in person the way we're sitting
here today.
>> But we need to really scientifically
rigorously answer the question why.
>> Yeah. And which aspect of that can be
mapped into the virtual world.
>> Exactly.
>> It's not it's not enough to say, "Yeah,
you should go touch grass and hang out
in nature." It's like, why exactly is
that valuable?
>> Yes. And I guess that's maybe the thing
that's been uh haunting me, obsessing me
from the beginning of my career. If you
think about all the different things
I've done, that's they're all related in
that way. This simulation, nature of
reality, and what is the bounds of, you
know, what can be modeled. Sorry for the
ridiculous question, but so far, what is
the greatest video game of all time?
What's up there?
>> Well, my favorite one of all time is
Civilization. I have to say that that
was the the the Civilization 1 and
Civilization 2. My favorite games of all
time. Um
>> I can only assume you've avoided the
most recent one because it would
probably you would that would be your
sobatical that you would disappear.
>> Yes, exactly. They take a lot of time
these Civilization games. So, I got to
be careful with them.
>> Fun question. You and Elon seem to be
somehow solid gamers. Uh is there a
connection between being great at gaming
and and uh being great leaders of AI
companies?
>> I don't know. I It's an interesting one.
I mean uh we both love games and uh it's
interesting he wrote games as well to
start off with. It's probably especially
in the era I grew up in where home
computers were just became a thing, you
know, in the late ' 80s and '9s,
especially in the UK. I had a Spectrum
and then a Commodore Omega 500 which is
my my favorite computer ever and that's
why I learned all my programming and of
course it's a very fun thing uh to
program is to program games. So I think
it's a great way to learn programming
probably still is and um and then of
course I immediately took it in
directions of AI and simulations which
so I may was able to express my interest
in in games and my sort of wider
scientific interests alto together. And
then the final thing I think that's
great about games is it fuses um
artistic design, you know, art with the
the the most cutting edge programming.
Um so again, in the '90s, all of the
most interesting uh technical advances
were happening in gaming, whether that
was AI, graphics, physics engines, uh
hardware, even GPUs of course were
designed for gaming originally. Um so
everything that was pushing computing
forward in the in the '9s was due to
gaming. So interestingly that was where
the forefront of research was going on
and it was this incredible fusion with
with art um you know graphics but also
music and just the whole new media of
storytelling and I love that. For me
it's this sort of multi-disiplinary kind
of effort is again something I've
enjoyed my whole my whole life. I have
to ask you, I almost forgot about one of
the many and I would say one of the most
incredible things recently uh that
somehow didn't yet get enough attention
is alpha evolve.
>> We talked about evolution a little bit
but it's the Google deep mind system
that evolves algorithms.
>> Yeah.
>> Are these kinds of evolution-like
techniques promising as a component of
future super intelligence system? So for
people who don't know, it's kind of um I
don't know if it's fair to say it's LLM
guided
evolution search.
>> Yeah.
>> So evolutionary algorithms are doing the
search and LLMs are telling you where.
>> Yes. Exactly. So LLMs are kind of
proposing some possible solutions and
then you do you use evolutionary
computing on top to to to find some
novel part of the of the search space.
So actually I think it's an example of
very promising directions where you
combine LLMs or foundation models with
other computational techniques.
Evolutionary methods is one but you
could also imagine Monte Carlo research
basically many types of search
algorithms or reasoning algorithms sort
of on top of or using the foundation
models as a basis. So, I actually think
there's quite a lot of interesting uh
things to be discovered probably with
these sort of hybrid systems, let's call
them.
>> But not to romanticize evolution. Yeah,
>> I'm only human. But you think there's
some value in whatever that mechanism is
because we already talked about natural
systems. Do you think where there's a
lot of lowhanging fruit of us
understanding being being able to model
uh being able to simulate evolution and
then using that whatever we understand
about that nature inspired mechanism to
to then do surge better and better and
better.
>> Yes. So if you think about uh again
breaking down the sort of systems we've
built uh to their really fundamental
core, you've got like the model of the
of the underlying dynamics of the
system. Uh and then if you want to
discover something new, something novel
that hasn't been seen before, um then
you need some kind of search process on
top to take you to a novel region of the
of the of the search space. And um you
can do that in a number of ways.
Evolutionary computing is one. um with
Alph Go we just use Monte Carlo research
right and that's what found move 37 the
new kind of never seen before strategy
in go and so that's how you can go
beyond potentially what is already known
so the model can model everything that
you currently know about right all the
data that you currently have but then
how do you go beyond that so that starts
to speak about the ideas of creativity
how can these systems create something
new discover something new obviously
this is super relevant for scientific
discovery or pushing met science and
medicine forward, which we want to do
with these systems. And you can actually
bolt on some uh fairly simple search
systems on top of these models and get
you into a new region of space. Of
course, you also have to um make sure
that you're not searching that space
totally randomly. It would be too big.
So, you have to have some objective
function that you're trying to optimize
and hill climb towards and that guides
that search. But there's some mechanism
of evolution that are interesting maybe
in the space of programs. But then the
space of programs is an extremely
important space because you can probably
generalize to to everything you know for
example mutation.
So it's not just Monte Carlo tree search
where it's like a search.
>> You could every once in a while
>> combine things. Yeah.
>> Combine things alter like sub like a
components of a thing. Yes. So then you
know what evolution is really good at is
not just the natural selection.
It's combining things and building
increasingly complex hierarchical
systems.
>> So that component is super interesting
especially like with alpha evolve in the
space of programs.
>> Yeah. Exactly. So there's a you can get
a bit of an extra property out of
evolutionary systems which is some new
emergent capability may come about but
of course like happened with life.
Interestingly, with naive uh sort of
traditional evolutionary computing
methods without LLMs and the modern AI,
the problem with them, there was they
were very well studied in the 90s and
and and and early 2000s and some
promising results, but the problem was
they could never work out how to evolve
new properties, new emergent properties.
You always had a sort of subset of the
properties that you put into the system.
But maybe if we combine them with these
foundation models, perhaps we can
overcome that limitation. Obviously uh
natural evolution clearly did because it
it did evolve new capabilities right so
bacteria to where we are now. So clearly
that it must be possible with
evolutionary systems to generate uh new
patterns you know going back to the
first thing we talked about and uh new
capabilities and emergent properties and
maybe we're on the cusp of discovering
how to do that.
>> Yeah listen uh alpha evolve is one of
the coolest things I've ever seen. I've
I've on my desk at home, you know, most
of my time is spent behind that
computers just programming. And next to
the the three screens is a skull of a
tectalic, which is one of the early
organisms that crawled out of the water
onto land. And I just kind of watch that
little guy.
It's like you whatever the computation
mechanism of evolution is is quite
incredible. It's truly truly incredible.
Now whether that's exactly the thing we
need to do to do our search but never
dismiss the power of nature what it did
here.
>> Yeah. And it's amazing um which is a
relatively simple algorithm right
effectively and it can generate all of
this immense complexity emerges
obviously running over you know 4
billion years of time but but it's it's
it's you know you can think about that
as again a pro a search process that ran
over the physics substrate of the
universe for a long amount of
computational time but then it generated
all this incredible uh rich diversity.
>> So uh so many questions I want to ask
you. But one, you do have a dream. One
of the natural systems you want to uh
try to model is a is a cell.
>> Yes,
>> that's a beautiful dream. Uh I could ask
you about that. I also just for that
purpose on the AI scientist front just
broadly. So there's a essay uh from
Daniel Cocatalio, Scott Alexander, and
others that outlines steps along the way
to get to ASI and has a lot of
interesting ideas in it. one of which is
uh including a superhuman coder and a
superhuman AI researcher
and in that there's a term of research
taste that's really interesting. So in
everything you've seen, do you think
it's possible for AI systems to have
research taste to help you in the way
that AI co-scientist does to help steer
human um human brilliant scientists and
then potentially by itself to figure out
what are the directions
where you want to generate truly novel
ideas because that seems to be like a
really important component how to do
great science. Yeah, I think that's
going to be one of the hardest things to
to uh mimic or model is is this this
idea of taste or or judgment. I think
that's what separates the you know the
the great scientists from the good
scientists like all all professional
scientists are good technically right
otherwise they wouldn't have made it
that far in in academia and things like
that but then do you have the taste to
sort of sniff out what the right
direction is what the right experiment
is what the right question is. So the
it's the it's picking the right question
is is the hardest part of science. Um
and and making the right hypothesis and
um that's what you know today's systems
definitely they can't do. So you know I
often say it's harder to come up with a
conjecture a really good conjecture than
it is to solve it. So we may have
systems soon that can solve pretty hard
conjectures. um you know I I um mass
Olympiad problems where we we you know
alpha proof last year our system got you
know silver medal in that really hard
problems maybe eventually we'll be able
to solve a millennium prize kind of
problem but could a system have come up
with a conjecture worthy of study that
someone like Terren Tower would have
gone you know what that's a really deep
question about the nature of maths or
the nature of numbers or the nature of
physics and that is far harder type of
creativity and we don't really Oh,
systems clearly can't do that and we're
not quite sure what that mechanism would
be. This kind of leap of imagination
like like Einstein had when he came up
with, you know, special relativity and
then general relativity with the
knowledge he had at the time.
>> As for conjecture,
the you want to come up with a thing
that's interesting and amenable to
proof.
>> Yes.
>> So like it's easy to come up with a
thing that's extremely difficult.
>> Yeah.
>> It's easy to come up with a thing that's
extremely easy. at that at that very
edge,
>> that sweet spot, right, of of basically
advancing the science and splitting the
hypothesis space into two ideally,
right? Whether if it's true or not true,
you you've learned something really
useful and um and and that's hard and
and and and making something that's also
uh you know falsifiable and within sort
of the technologies that you have you
currently have available. So it's a very
creative process actually highly
creative process that um I think just a
kind of naive search on top of a model
won't be enough for that.
>> Okay. The idea of splitting the
hypothesis space in two is super
interesting. So uh I've heard you say
that there's basically no failure in or
failure is extremely valuable if it's
done if you construct the questions
right if you construct the experiments
right if you design them right that
failure success are both useful. So
perhaps because it splits the hypothesis
basically two, it's like a binary
search.
>> That's right. So when you do like, you
know, real blue sky research, there's no
such thing as failure really as long as
you're picking experiments and
hypotheses that that that that
meaningfully spit the hypothesis space.
So you know, and you learn something,
you can learn something kind of equally
valuable from an experiment that doesn't
work. That should tell you, if you've
designed the experiment well and your
hypothesis are interesting, it should
tell you a lot about where to go next.
and um and then it's you're effectively
doing a search process um and using that
information in in you know very helpful
ways. So to go to your dream
of uh modeling a cell uh what are the
big challenges that lay ahead for us to
make that happen? We should maybe
highlight that alpha I mean there's just
so many leaps.
>> Yeah.
>> So AlphaFold solved if it's fair to say
protein folding and there's so many
incredible things we could talk about
there including the open sourcing uh the
everything you've released. Alpha Fold 3
is doing protein, RNA, DNA interactions,
>> which is super complicated and and
fascinating. That's amendable to
modeling. Alpha genome uh predicts uh
how small genetic changes like if we
think about single mutations, how they
link to actual uh function. So um those
are it seems like it's creeping along to
sophistic to to much more complicated u
things like a cell but a cell has a lot
of really complicated components.
>> Yeah. So what I've tried to do
throughout my career is I have these
really grand dreams and then I try to as
you've noticed and then I try to break
but I try to break them down any you
know it's easy to have a kind of a crazy
ambitious dream but the the the trick is
how do you break it down into manageable
achievable uh interim steps that are
meaningful and useful in their own right
and so virtual cell which is what I call
the project of modeling a cell I've had
this idea you know of wanting to do that
for maybe more like 25 is and I used to
talk with Paul Nurse who is a bit of a
mentor of mine in biology. He runs the
the you know founded the Craig Institute
and and won the Nobel Prize in in 2001.
uh is is we've been talking about it
since you know before the you know in
the '90s and um and I come used to come
back to every 5 years is like what would
you need to model the full internals of
a cell so that you could do experiments
on the virtual cell and what those
experiment you know in silicone and
those predictions would be useful for
you to save you a lot of time in the wet
lab right that would be the dream maybe
you could 100x speed up experiments by
doing most of it in silicone the search
in silicico and then you do the
validation step in the wet lab. That
would be that's the that's the dream.
And so u but maybe now finally uh so I
was trying to build these components
alpha fold being one that that would
allow you eventually to model the full
interaction a full simulation of a cell
and I'd probably start with a yeast cell
and partly that's what Paul nurse
studied because a yeast cell is like a
full organism that's a single cell right
so it's the kind of simplest single cell
organism and so it's not just a cell
it's a full organism and um and yeast is
very well understood And so that would
be a good candidate for uh a a kind of
full simulated model. Now alpha fold is
the is the solution to the kind of
static picture of what does a what does
a protein look 3D structure protein look
like a static picture of it. But we know
that biology all the interesting things
happen with the dynamics the
interactions and that's what alpha 3 is
is the first step towards is modeling
those interactions. So first of all
pairwise you know proteins with proteins
proteins with RNA and DNA but then um
the next step after that would be
modeling maybe a whole pathway maybe
like the to pathway that's involved in
cancer or something like this and then
eventually you might be able to model
you know a whole cell
>> also there's another complexity here
that stuff in a cell happens at
different time scales is that tricky
like there you know protein uh folding
is you know super fast
>> yes
>> um I don't know all the bi ological
mechanisms, but some of them take a long
time. And so is that that's an level. So
the levels of interaction has a
different temporal scale that you have
to be able to model.
>> So that would be hard. So you'd probably
need several simulated systems that can
interact at these different temporal
dynamics or at least maybe it's like a
hierarchical system. So um you can jump
up and down the the different temporal
stages. So can you avoid I mean one of
the challenges here is
not avoid simulating for example the the
the quantum mechanical aspects of any of
this right you want to not overm model
you can skip ahead to just model the
really highlevel things that get you a
really good estimate of what's going to
happen
>> so you you got to make a decision when
you're modeling any natural system what
is the cutoff level of the granularity
that you're going to model it to that
then captures the dynamics that you're
interested in. So probably for a cell I
would hope that would be the protein
level uh and that one wouldn't have to
go down to the atomic level. Um so you
know of course that's where alpha volt
stock kicks in. So that would be kind of
the basis and then you'd build these um
uh higher level simulations that um take
those as building blocks and then you
get the emergent behavior. Apologize for
the pthead questions ahead of time, but
uh will do you think uh we'll be able to
simulate and model the origin of life.
So being able to simulate the first from
from non-living organisms the the birth
of a living organism.
>> I think that's a one of the of course
one of the deepest and most fascinating
questions. Um I love that area of
biology. you know, uh, people like
there's a great book by Nick Lane, one
of the top top experts in this area
called the the 10 great inventions of of
of evolution. I think it's fantastic and
it also speaks to what the great filters
might be, you know, prior or are they
ahead of us. I think I think they're
most likely in the past if you read that
book of how unlikely to go, you know,
have any life at all and then single
cell to multisell seems an unbelievably
big jump that took like a billion years,
I think, on Earth to do, right? So it
shows you how hard it was, right?
>> Bacteria were super happy for a very
long time,
>> a very long time before they captured
mitochondria somehow, right? I don't see
why not why AI couldn't help with that
some kind of simulation. Again, it's
again, it's a bit of a search process
through a combinatorial space. Here's
like all the chem, you know, the
chemical soup that that you start with,
the primordial soup that, you know,
maybe was on Earth near these hot vents.
Here's some initial conditions. Can you
uh generate something that looks like a
cell? So perhaps that would be a next
stage after the virtual cell project is
well how how could you actually um
something like that emerge from the
chemical soup?
>> Well, I would love it if there was a
move 37 for the origin of life. Yeah,
>> I think that's one of the sort of great
mysteries. I think ultimately what we
will figure out is their continuum.
There's no such thing as a li