The Thinking Game | Full documentary | Tribeca Film Festival official selection

The Thinking Game | Full documentary | Tribeca Film Festival official selection

(RUMINATIVE MUSIC PLAYING) (DEVICE CHIMES) JULIETTE LOVE: Hi, Alpha. ALPHA: Hello. LOVE: Can you
help me write code? ALPHA: I was trained
to answer questions, but I’m able to learn. LOVE: That’s very
open-minded of you. ALPHA: Thank you.
I’m glad you’re happy with me. What’s this guy doing? ALPHA: That’s a developer. What do you think
he’s working on? ALPHA:
That’s a tough question. He might be working
on a new feature, a bug fix or something else. It’s quite possible. ALPHA: Yes. LOVE: Do you see my backpack? ALPHA:
That’s a badminton racket. It’s a squash racket,
but that’s pretty close. ALPHA:
That’s a badminton racket. No, but you’re not
the first person to make that mistake. (UPBEAT MUSIC PLAYING) NEWSREADER 1:
AI, the technology that has been advancing
at breakneck speed. NEWSREADER 2: Artificial
intelligence is all the rage. NEWSREADER 3: Some are now
raising alarm about… NEWSREADER 4:
It is definitely concerning. NEWSREADER 5:
This is an AI arms race. NEWSREADER 6: We don’t know how this is all
going to shake out, but it’s clear
something is happening. DEMIS HASSABIS:
I’m kind of restless. Trying to build AGI
is the most exciting journey, in my opinion, that humans
have ever embarked on. If you’re really going
to take that seriously, there isn’t a lot of time. Life’s very short. My whole life goal is to solve artificial
general intelligence. And on the way,
use AI as the ultimate tool to solve all the world’s most complex
scientific problems. I think that’s bigger
than the Internet. I think that’s bigger
than mobile. I think it’s more like the advent
of electricity or fire. ANNOUNCER: World leaders and artificial
intelligence experts are gathering
for the first ever global AI safety summit, set to look at the risks of the fast growing technology
and also… HASSABIS: I think
this is a hugely critical moment
for all humanity. It feels like
we’re on the cusp of some incredible things
happening. NEWSREADER:
Let me take you through some of the reactions
in today’s papers. HASSABIS: AGI is pretty close,
I think. There’s clearly huge interest
in what it is capable of, where it’s taking us. HASSABIS: This is the moment I’ve been living
my whole life for. (MID-TEMPO
ELECTRONIC MUSIC PLAYS) I’ve always been fascinated
by the mind. So I set my heart
on studying neuroscience because I wanted
to get inspiration from the brain for AI. ELEANOR MAGUIRE:
I remember asking Demis, “What’s the end game?” You know?
So you’re going to come here and you’re going
to study neuroscience and you’re going to maybe
get a Ph.D. if you work hard. And he said, “You know, I want
to be able to solve AI. “I want to be able
to solve intelligence.” HASSABIS: The human brain
is the only existent proof we have, perhaps
in the entire universe, that general intelligence
is possible at all. And I thought
someone in this building should be interested in general intelligence
like I am. And then Shane’s name
popped up. HOST: Our next speaker today
is Shane Legg. He’s from New Zealand, where he trained in math
and classical ballet. Are machines actually
becoming more intelligent? Some people say yes,
some people say no. It’s not really clear. We know they’re getting
a lot faster at doing computations. But are we actually
going forwards in terms
of general intelligence? HASSABIS: We were both
obsessed with AGI, artificial
general intelligence. So today I’m going
to be talking about different approaches
to building AGI. With my colleague
Demis Hassabis, we’re looking at ways
to bring in ideas from theoretical neuroscience. I felt like we were
the keepers of a secret that no one else knew. Shane and I knew
no one in academia would be supportive
of what we were doing. AI was almost
an embarrassing word to use in academic circles,
right? If you said
you were working on AI, then you clearly weren’t
a serious scientist. So I convinced Shane
the right way to do it would be to start a company. SHANE LEGG: Okay,
we’re going to try to do artificial
general intelligence. It may not even be possible. We’re not quite sure
how we’re going to do it, but we have some ideas
or, kind of, approaches. Huge amounts of money,
huge amounts of risk, lots and lots of compute. And if we pull this off, it’ll be the biggest thing
ever, right? That is a very hard thing
for a typical investor to put their money on. It’s almost like
buying a lottery ticket. I’m going to be speaking about
the system of neuroscience and how it might be used
to help us build AGI. HASSABIS:
Finding initial funding for this was very hard. We’re going to solve
all of intelligence. You can imagine
some of the looks I got when we were
pitching that around. So I’m a V.C.
and I look at about 700 to 1,000 projects a year. And I fund
literally 1% of those. About eight projects a year. So that means 99% of the time,
you’re in “No” mode. “Wait a minute.
I’m telling you, “this is the most important
thing of all time. “I’m giving you
all this build-up “about how… explain “how it connects
with the brain, “why the time’s right now,
and then you’re asking me, “‘But what’s your, how are you
going to make money? “‘What’s your product?'” It’s like,
so prosaic a question. You know? “Have you not been listening
to what I’ve been saying?” LEGG: We needed investors who aren’t necessarily
going to invest because they think
it’s the best investment decision. They’re probably
going to invest because they just think
it’s really cool. NEWSREADER:
He’s the Silicon Valley version of the man
behind the curtain in The Wizard of Oz. He had a lot to do
with giving you PayPal, Facebook,
YouTube and Yelp. LEGG: If everyone says “X,” Peter Thiel suspects
that the opposite of X is quite possibly true. HASSABIS: So Peter Thiel
was our first big investor. But he insisted that
we come to Silicon Valley because that
was the only place we could… There would be the talent, and we could build
that kind of company. But I was pretty adamant
we should be in London because I think
London’s an amazing city. Plus, I knew there were
really amazing people trained at Cambridge
and Oxford and UCL. In Silicon Valley, everybody’s founding
a company every year, and then if it doesn’t work, you chuck it
and you start something new. That is not conducive to a long-term
research challenge. So we were totally
an outlier for him. Hi, everyone.
Welcome to DeepMind. So, what is our mission? We summarize it as… DeepMind’s mission is to build
the world’s first general learning machine. So we always stress the word
“general” and “learning” here are the key things. LEGG: Our mission
was to build an AGI, an artificial
general intelligence. And so that means that we need
a system which is general. It doesn’t learn to do
one specific thing. That’s a really key part
of human intelligence. We can learn to do
many, many things. It’s going to, of course,
be a lot of hard work. But one of the things
that keeps me up at night is to not waste this
opportunity to, you know, to really make
a difference here, and have a big impact
on the world. LEGG: The first people
that came and joined DeepMind
really believed in the dream. But this was, I think,
one of the first times they found a place
full of other dreamers. You know, we collected
this Manhattan Project, if you like,
together to solve AI. HELEN KING:
In the first two years, we were in total stealth mode. And so we couldn’t
say to anyone what were we doing
or where we worked. It was all quite vague. BEN COPPIN: It had
no public presence at all. You couldn’t
look at a website. The office
was at a secret location. When we would interview people
in those early days, they would show up
very nervously. (LAUGHS) I had at least one candidate
who said, “I just messaged my wife
to tell her exactly “where I’m going just in case “this turns out to be some
kind of horrible scam “and I’m going
to get kidnapped.” Well, my favorite new person
who’s an investor, who I’ve been working
for a year, is Elon Musk. So for those of you
who don’t know, this is what he looks like. And he hadn’t really thought
much about AI until we chatted. His mission is to die on Mars
or something. -But not on impact.
-(LAUGHTER) So… We made some big decisions about how we were going
to approach building AI. This is a reinforcement
learning setup. This is the kind of setup
that we think about when we say we’re building,
you know, an AI agent. It’s basically the agent,
which is the AI, and then there’s
the environment that it’s interacting with. We decided that games, as long as
you’re very disciplined about how you use them, are the perfect
training ground for AI development. LEGG: We wanted
to try to create one algorithm that could to be
trained up to play several dozen
different Atari games. So just like a human, you have to use the same brain
to play all the games. DAVID SILVER:
You can think of it that you provide the agent
with the cartridge. And you say, “Okay, imagine you’re born
into that world “with that cartridge,
and you just get to interact “with the pixels
and see the score. “What can you do?” So what you’re going to do is
take your Q function. Q-K… HASSABIS: Q-learning
is one of the oldest methods for reinforcement learning. And what we did was combine
reinforcement learning with deep learning
in one system. No one had ever combined
those two things together at scale to do
anything impressive, and we needed
to prove out this thesis. LEGG: We tried doing Pong
as the first game. It seemed like the simplest. It hasn’t been told anything about
what it’s controlling or what it’s supposed to do. All it knows
is that score is good and it has to learn
what its controls do, and build everything…
first principles. (GAME BEEPING) It wasn’t really working. HASSABIS: I was just
saying to Shane, “Maybe we’re just wrong,
and we can’t even do Pong.” LEGG: It was a bit
nerve-racking, thinking how far we had to go if we were going
to really build a generally
intelligent system. HASSABIS: And it felt like
it was time to give up and move on. And then suddenly… (STIRRING MUSIC PLAYS) We got our first point. And then it was like,
“Is this random?” “No, no, it’s really
getting a point now.” It was really exciting
that this thing that previously
couldn’t even figure out how to move a paddle had suddenly been able
to totally get it right. HASSABIS: Then it was getting
a few points. And then it won
its first game. And then three months later,
no human could beat it. You hadn’t told it the rules,
how to get the score, nothing. And you just tell it
to maximize the score, and it goes away and does it. This is the first time anyone had done
this end-to-end learning. “Okay, so we have this working
in quite a general way. “Now let’s try another game.” HASSABIS: So then
we tried Breakout. At the beginning,
after 100 games, the agent is not very good. It’s missing the ball
most of the time, but it’s starting to get
the hang of the idea that the bat should go
towards the ball. Now, after 300 games, it’s about as good as
any human can play this. We thought,
“Well, that’s pretty cool,” but we left the system playing
for another 200 games, and it did this amazing thing. It found the optimal strategy was to dig a tunnel
around the side and put the ball
around the back of the wall. KORAY KAVUKCUOGLU:
Finally, the agent is actually achieving what you thought
it would achieve. That is a great feeling.
Right? Like, I mean,
when we do research, that is the best
we can hope for. We started generalizing
to 50 games, and we basically
created a recipe. We could just take a game that we have
never seen before. We would run
the algorithm on that, and DQN could train itself
from scratch, achieving human level or sometimes better
than human level. LEGG: We didn’t build it
to play any of them. We could just give it
a bunch of games and would figure it out
for itself. And there was something
quite magical in that. MURRAY SHANAHAN:
Suddenly you had something that would respond and learn whatever situation
it was parachuted into. And that was like a huge,
huge breakthrough. It was in many respects the first example of any kind of thing
you could call a general intelligence. HASSABIS: Although we were
a well-funded startup, holding us back
was not enough compute power. I realized that
this would accelerate our time scale
to AGI massively. I used to see Demis
quite frequently. We’d have lunch, and he did… say to me that
he had two companies that were involved
in buying DeepMind. And he didn’t know
which one to go with. The issue was,
would any commercial company appreciate the real importance
of the research? And give the research time
to come to fruition and not be breathing down
their necks, saying, “We want some kind of
commercial benefit from this.” (MACHINERY HUMMING) Google has bought DeepMind
for a reported £400,000,000, making the artificial
intelligence firm its largest
European acquisition so far. The company was founded by 37-year-old entrepreneur
Demis Hassabis. After the acquisition,
I started mentoring and spending time with Demis, and just listening to him. And this is a person
who fundamentally is a scientist
and a natural scientist. He wants science to solve
every problem in the world, and he believes it can do so. That’s not a normal person
you find in a tech company. HASSABIS: We were able
to not only join Google but run independently
in London, build our culture, which was optimized
for breakthroughs and not deal with products, do pure research. Our investors
didn’t want to sell, but we decided that this was the best thing
for the mission. In many senses,
we were underselling in terms of value
before it more matured, and you could have sold it
for a lot more money. And the reason is because
there’s no time to waste. There’s so many things
that got to be cracked while the brain
is still in gear. You know, I’m still alive. There’s all these things
that gotta be done. So you haven’t got–
I mean, how many… How many billions
would you trade for another five years of life,
you know, to do what you set out to do? Okay, all of a sudden, we’ve got this massive scale
compute available to us. What can we do with that? HASSABIS: Go is the pinnacle
of board games. It is the most complex game
ever devised by man. There are more possible
board configurations in the game of Go than there
are atoms in the universe. SILVER: Go is the holy grail
of artificial intelligence. For many years, people have looked
at this game and they’ve thought,
“Wow, this is just too hard.” Everything we’ve ever
tried in AI, it just falls over when
you try the game of Go. And so that’s why
it feels like a real litmus test
of progress. We had just bought DeepMind. They were working
on reinforcement learning and they were the world’s
experts in games. And so when
they introduced the idea that they could beat
the top level Go players in a game that was thought
to be incomputable, I thought, “Well,
that’s pretty interesting.” Our ultimate next step
is to play the legendary Lee Sedol
in just over two weeks. NEWSREADER 1:
A match like no other is about to get underway
in South Korea. NEWSREADER 2: Lee Sedol
is getting ready to rumble. HASSABIS:
Lee Sedol is probably one of the greatest players
of the last decade. I describe him
as the Roger Federer of Go. ERIC SCHMIDT: He showed up, and all of a sudden
we have a thousand Koreans who represent
all of Korean society, the top Go players. And then we have Demis. And the great
engineering team. Lee Sedol, he’s very famous for very creative
fighting play. So this could be
difficult for us. SCHMIDT: I figured Lee Sedol
is going to beat these guys, but they’ll make
a good showing. Good for a startup. I went over
to the technical group and they said, “Let me show you
how our algorithm works.” RESEARCHER: If you step
through the actual game, we can see, kind of,
how AlphaGo thinks. HASSABIS: The way we start off
on training AlphaGo is by showing it 100,000 games that strong amateurs
have played. And we first initially get AlphaGo to mimic
the human player, and then through
reinforcement learning, it plays against
different versions of itself many millions of times
and learns from its errors. Hmm, this is interesting. ANNOUNCER 1: All right, folks, you’re going to see
history made. (ANNOUNCER 2 SPEAKING KOREAN) SCHMIDT: So the game starts. ANNOUNCER 1:
He’s really concentrating. ANNOUNCER 3:
If you really look at the… (ANNOUNCERS EXCLAIM) That’s a very surprising move. ANNOUNCER 3: I think we’re
seeing an original move here. Yeah, that’s an exciting move. I like… SILVER:
Professional commentators almost unanimously said that not a single human player
would have chosen move 37. So I actually had a poke
around in AlphaGo to see what AlphaGo thought. And AlphaGo actually agreed
with that assessment. AlphaGo said there was a one
in 10,000 probability that move 37 would have been
played by a human player. (SEDOL SPEAKING IN KOREAN) SILVER: The game of Go
has been studied for thousands of years. And AlphaGo discovered
something completely new. ANNOUNCER: He resigned.
Lee Sedol has just resigned. He’s beaten. (ELECTRONIC MUSIC PLAYING) NEWSREADER 1: The battle
between man versus machine, a computer just came out
the victor. NEWSREADER 2: Google
put its DeepMind team to the test against one of the brightest minds
in the world and won. SCHMIDT:
That’s when we realized the DeepMind people knew
what they were doing and to pay attention
to reinforcement learning as they have invented it. Based on that experience, AlphaGo got better
and better and better. And they had a little chart of how much better
they were getting. And I said,
“When does this stop?” And Demis said, “When we beat the Chinese guy, “the top-rated player
in the world.” ANNOUNCER 1:
Ke Jie versus AlphaGo. ANNOUNCER 2:
And I think we will see AlphaGo pushing through there. ANNOUNCER 1:
AlphaGo is ahead quite a bit. SCHMIDT: About halfway
through the first game, the best player in the world
was not doing so well. ANNOUNCER 1:
What can black do here? ANNOUNCER 2: Looks difficult. SCHMIDT:
And at a critical moment… the Chinese government
ordered the feed cut off. It was at that moment
we were telling the world that something new
had arrived on earth. In the 1950s when Russia’s Sputnik
satellite was launched, it changed
the course of history. TV HOST: It is a challenge
that America must meet to survive in the Space Age. SCHMIDT: This has been
called the Sputnik moment. The Sputnik moment created
a massive reaction in the US in terms of funding
for science and engineering, and particularly
of space technology. For China,
AlphaGo was the wakeup call, the Sputnik moment. It launched an AI space race. HASSABIS: We had this
huge idea that worked, and now the whole world knows. It’s always easier
to land on the moon if someone’s already
landed there. It is going to matter
who builds AI, and how it gets built. I always feel that pressure. SILVER: There’s been
a big chain of events that followed on from all
of the excitement of AlphaGo. When we played
against Lee Sedol, we actually had a system that had been trained
on human data, on all of the millions
of games that have been played
by human experts. We eventually found
a new algorithm, a much more elegant approach
to the whole system, which actually stripped out
all of the human knowledge and just started
completely from scratch. And that became a project
which we called AlphaZero. Zero, meaning having zero
human knowledge in the loop. Instead of learning
from human data, it learned from its own games. So it actually
became its own teacher. HASSABIS:
AlphaZero is an experiment in how little knowledge
can we put into these systems and how quickly
and how efficiently can they learn? AlphaZero doesn’t
have any rules. It learns through experience. The next stage
was to make it more general, so that it could play
any two-player game. Things like chess, and in fact,
any kind of two-player perfect information game. It’s going really well. It’s going
really, really well. -Oh, wow.
-It’s going down, like fast. HASSABIS: AlphaGo used
to take a few months to train, but AlphaZero could start
in the morning playing completely randomly and then by tea
be at superhuman level. And by dinner it will be
the strongest chess entity there’s ever been. -Amazing, it’s amazing.
-Yeah. It’s discovered its own
attacking style, you know, to take on the current
level of defense. I mean, I never
in my wildest dreams… I agree. Actually, I was not
expecting that either. And it’s fun for me. I mean, it’s inspired me
to get back into chess again, because it’s cool to see that there’s even more depth
than we thought in chess. (HORN BLOWS) HASSABIS: I actually got
into AI through games. Initially, it was board games. I was thinking,
“How is my brain doing this?” Like, what is it doing? I was very aware of that
from a very young age. So I’ve always been thinking
about thinking. NEWSREADER: The British
and American chess champions meet to begin
a series of matches. Playing alongside them
are the cream of Britain and America’s
youngest players. NEWSREADER 2: Demis Hassabis
is representing Britain. COSTAS HASSABIS:
When Demis was four, he first showed
an aptitude for chess. By the time he was six, he became London
under-eight champion. HASSABIS: My parents
were very interesting and unusual, actually. I’d probably describe them
as quite bohemian. My father
was a singer-songwriter when he was younger, and Bob Dylan was his hero. Around when I was about eight,
my dad got a camper van. (HORN HONKS) (ANGELA HASSABIS SPEAKING) Yeah, yeah. HOST: What is it
that you like about this game? It’s just a good
thinking game. HASSABIS: At the time,
I was the second-highest rated chess player in the world
for my age. But although I was on track to be a professional
chess player, I thought that was what
I was going to do. No matter how much
I loved the game, it was incredibly stressful. Definitely was not fun
and games for me. Parents used to, you know, get very upset
when I lost the game and angry
if I forgot something. And because it was quite high
stakes for them, you know, it cost a lot of money
to go to these tournaments. And my parents
didn’t have much money. My parents thought, you know, “If you interested
in being a chess professional, “this is really important.
It’s like your exams.” I remember
I was about 12-years-old and I was at this
international chess tournament in Liechtenstein
up in the mountains. (BELL TOLLING) And we were in this
huge church hall with, you know, hundreds of international
chess players. And I was playing
the ex-Danish champion. He must have been
in his 30s, probably. In those days,
there was a long time limit. The games could
literally last all day. (YAWNS) (TIMER TICKING) We were into our tenth hour. (TIMER TICKS FRANTICALLY) And we were in this
incredibly unusual ending. I think it should be a draw. But he kept on trying
to win for hours. (HORSE NEIGHS) Finally, he tried
one last cheap trick. All I had to do
was give away my queen. Then it would be stalemate. But I was so tired, I thought it was inevitable
I was going to be checkmated. And so I resigned. He jumped up.
Just started laughing. (LAUGHING) And he went, “Why have you resigned?
It’s a draw.” And he immediately,
with a flourish, sort of showed me
the drawing move. I felt so sick to my stomach. It made me think of
the rest of that tournament. Like, are we wasting
our minds? Is this the best use
of all this brain power? Everybody’s, collectively,
in that building? If you could somehow plug in those 300 brains
into a system, you might be able
to solve cancer with that level
of brain power. This intuitive feeling
came over me that although I love chess, this is not the right thing
to spend my whole life on. LEGG: Demis and myself, our plan was always
to fill DeepMind with some of the most brilliant scientists
in the world. So we had the human brains necessary to create
an AGI system. By definition, the “G”
in AGI is about generality. What I imagine is being able
to talk to an agent, the agent can talk back, and the agent is able to solve
novel problems that it hasn’t seen before. That’s a really key part
of human intelligence, and it’s that
cognitive breadth and flexibility
that’s incredible. The only natural
general intelligence we know of as humans, we obviously learn a lot
from our environment. So we think that
simulated environments are one of the ways
to create an AGI. SIMON CARTER:
The very early humans were having to solve
logic problems. They were having to solve
navigation, memory, and we evolved
in that environment. If we can create
a virtual recreation of that kind of environment, that’s the perfect
testing ground and training ground for everything
we do at DeepMind. GUY SIMMONS:
What they were doing here was creating environments
for childlike beings, the agents to exist
within and play. That just sounded like the most interesting thing
in all the world. SHANAHAN: A child
learns by tearing things up and then throwing food around and getting a response
from mommy or daddy. This seems like an important
idea to incorporate in the way you train an agent. RESEARCHER 1: The humanoid
is supposed to stand up. As his center
of gravity rises, it gets more points. You have a reward and the agent
learns from the reward, like, you do something well,
you get a positive reward. You do something bad,
you get a negative reward. RESEARCHER 2: (EXCLAIMS)
It looks like it’s standing. It’s still a bit drunk. RESEARCHER 1:
It likes to walk backwards. RESEARCHER 2: (CHUCKLES) Yeah. The whole algorithm
is trying to optimize for receiving as much rewards
as possible, and it’s found that
walking backwards, it’s good enough
to get very good scores. RAIA HADSELL:
When we learn to navigate, when we learn to get around
in our world, we don’t start with maps. We just start
with our own exploration, adventuring off
across the park, without our parents
by our side, or finding our way home
from school when we’re young. (FAST ELECTRONIC
MUSIC PLAYING) HADSELL: A few of us
came up with this idea that if we had an environment
where a simulated robot just had to run forward, we could put all sorts of
obstacles in its way and see if it could manage
to navigate different types of terrain. The idea would be like
a parkour challenge. It’s not graceful, but was never trained to hold
a glass whilst it was running and not spill water. You set this objective
that says, “Just move forward,
forward velocity, “and you’ll get
a reward for that.” And the learning algorithm
figures out how to move
this complex set of joints. That’s the power of reward-based
reinforcement learning. SILVER: Our goal
is to try and build agents which, we drop them in,
they know nothing, they get to play around in
whatever problem you give them and eventually figure out how
to solve it for themselves. Now we want something
which can do that in as many different types
of problems as possible. A human needs diverse skills
to interact with the world. How to deal
with complex images, how to manipulate
thousands of things at once, how to deal
with missing information. We think all of these things
together are represented
by this game called StarCraft. All it’s being trained
to do is, given this situation,
this screen, what would a human do? We took inspiration from
large language models where you simply train
a model to predict the next word, which is exactly the same as predict the next
StarCraft move. SILVER: Unlike chess or Go, where players take turns
to make moves, in StarCraft there’s a
continuous flow of decisions. On top of that, you can’t even see
what the opponent is doing. There is no longer
a clear definition of what it means
to play the best way. It depends on
what your opponent does. HADSELL: This is the way
that we’ll get to a much more fluid, more natural, faster,
more reactive agent. ORIOL VINYALS:
This is a huge challenge and let’s see how far
we can push. TIM LILLICRAP: Oh! Holy monkey! I’m a pretty
low-level amateur. I’m okay, but I’m
a pretty low-level amateur. These agents have
a long ways to go. HASSABIS: We couldn’t
beat someone of Tim’s level. You know, that was
a little bit alarming. LILLICRAP:
At that point, it felt like it was going to be, like,
a really big long challenge, maybe a couple of years. VINYALS: Dani is the best
DeepMind StarCraft 2 player. I’ve been playing the agent
every day for a few weeks now. I could feel that the agent was getting better
really fast. (CHEERING, LAUGHTER) Wow, we beat Danny.
That, for me, was already
like a huge achievement. HASSABIS: The next step is we’re going to book in
a pro to play. (KEYBOARD TAPPING) (GROANS) (CHEERING, WHOOPING) (CHEERING, WHOOPING) -(LAUGHS)
-(PEOPLE CLAPPING) It feels a bit unfair.
All you guys against me. (ALL LAUGH) HASSABIS: We’re way
ahead of what I thought we would do, given where
we were two months ago. Just trying to digest it all,
actually. But it’s very, very cool. SILVER: Now we’re in
a position where we can finally share
the work that we’ve done with the public. This is a big step. We are really putting
ourselves on the line here. -Take it away. Cheers.
-Thank you. We’re going to be live
from London. It’s happening. ANNOUNCER 1:
Welcome to London. We are going to have
a live exhibition match, MaNa against AlphaStar. (CHEERING, APPLAUSE) At this point now, AlphaStar, 10 and 0
against professional gamers. Any thoughts
before we get into this game? VINYALS: I just want to see
a good game, yeah. I want to see a good game. SILVER: Absolutely,
good game. We’re all excited. ANNOUNCER: All right. Let’s
see what MaNa can pull off. ANNOUNCER 2:
AlphaStar is definitely dominating the pace
of this game. (SPORADIC CHEERING) ANNOUNCER 1: Wow. AlphaStar
is playing so smartly. (LAUGHTER) This really looks like
I’m watching a professional human gamer from the AlphaStar
point of view. (KEYBOARD TAPPING) HASSABIS: I hadn’t really seen
a pro play StarCraft up close, and the 800 clicks per minute. I don’t understand how anyone
can even click 800 times, let alone doing
800 useful clicks. ANNOUNCER 1:
Oh, another good hit. -(ALL GROAN)
-AlphaStar is just completely relentless. SILVER: We need to be careful because many of us grew up
as gamers and are gamers. And so to us,
it’s very natural to view games
as what they are, which is pure vehicles
for fun, and not to see
that more militaristic side that the public might see
if they looked at this. You can’t look at gunpowder
and only make a firecracker. All technologies inherently
point into certain directions. MARGARET LEVI:
I’m very worried about the certain ways in which AI will be used
for military purposes. And that makes it even clearer
how important it is for our societies
to be in control of these new technologies. The potential for abuse
from AI will be significant. Wars that occur faster
than humans can comprehend and more powerful
surveillance. How do you keep power forever over something that’s
much more powerful than you? (STEPHEN HAWKING SPEAKING) Technologies can be used
to do terrible things. And technology can be used
to do wonderful things and solve
all kinds of problems. When DeepMind
was acquired by Google… -Yeah.
-…you got Google to promise that technology you developed
won’t be used by the military -for surveillance.
-Right. -Yes.
-Tell us about that. I think technology
is neutral in itself, um, but how, you know,
we as a society or humans and companies
and other things, other entities and governments
decide to use it is what determines whether
things become good or bad. You know, I personally think
having autonomous weaponry is just a very bad idea. ANNOUNCER 1:
AlphaStar is playing an extremely intelligent game
right now. CUKIER: There is an element to
what’s being created at DeepMind in London that does seem like
the Manhattan Project. There’s a relationship between
Robert Oppenheimer and Demis Hassabis in which they’re unleashing
a new force upon humanity. ANNOUNCER 1:
MaNa is fighting back, though. Oh, man! HASSABIS:
I think that Oppenheimer and some of the other leaders
of that project got caught up in the excitement
of building the technology and seeing if it was possible. ANNOUNCER 1:
Where is AlphaStar? Where is AlphaStar? I don’t see AlphaStar’s units
anywhere. HASSABIS: They did not think
carefully enough about the morals of what
they were doing early enough. What we should do
as scientists with powerful new technologies is try and understand it in
controlled conditions first. ANNOUNCER 1: And that is that. MaNa has defeated AlphaStar. I mean, my honest feeling is
that I think it is a fair representation
of where we are. And I think that part feels…
feels okay. -I’m very happy for you.
-I’m happy. So well… well done. My view is that the approach
to building technology which is embodied by
move fast and break things, is exactly what
we should not be doing, because you can’t afford
to break things and then fix them afterwards. -Cheers.
-Thank you so much. Yeah, get… get some rest.
You did really well. -Cheers, yeah?
-Thank you for having us. (ELECTRONIC MUSIC PLAYING) HASSABIS: When I was eight, I bought my first computer with the winnings
from a chess tournament. I sort of had this intuition that computers
are this magical device that can extend
the power of the mind. I had a couple
of school friends, and we used to have
a hacking club, writing code, making games. And then over
the summer holidays, I’d spend the whole day flicking through
games magazines. And one day I noticed
there was a competition to write an original version
of Space Invaders. And the winner won a job
at Bullfrog. Bullfrog at the time was the
best game development house in all of Europe. You know, I really wanted
to work at this place and see how they build games. NEWSCASTER: Bullfrog,
based here in Guildford, began with a big idea. That idea turned into the game
Populous, which became
a global bestseller. In the ’90s, there was
no recruitment agencies. You couldn’t go out and say,
you know, “Come and work
in the games industry.” It was still not even
considered an industry. So we came up with the idea
to have a competition and we got
a lot of applicants. And one of those was Demis’s. I can still remember clearly the day that Demis came in. He walked in the door,
he looked about 12. I thought, “Oh, my God, “what the hell are we going
to do with this guy?” I applied to Cambridge. I got in but they said
I was way too young. So…
So I needed to take a year off so I’d be at least 17
before I got there. And that’s when I decided
to spend that entire gap year working at Bullfrog. They couldn’t even
legally employ me, so I ended up being paid
in brown paper envelopes. (CHUCKLES) I got a feeling of being
really at the cutting edge and how much fun that was
to invent things every day. And then you know,
a few months later, maybe everyone… a million
people will be playing it. MOLYNEUX: In those days
computer games had to evolve. There had to be new genres which were more
than just shooting things. Wouldn’t it be amazing
to have a game where you design and build
your own theme park? (GAME CHARACTERS SCREAMING) Demis and I started to talk
about Theme Park. It allows the player
to build a world and see the consequences
of your choices that you’ve made
in that world. HASSABIS: A human player
set out the layout of the theme park and designed
the roller coaster and set the prices
in the chip shop. What I was working on was
the behaviors of the people. They were autonomous and that was the AI
in this case. So what I was trying to do
was mimic interesting human behavior so that the simulation
would be more interesting
to interact with. MOLYNEUX: Demis worked
on ridiculous things, like you could place down
these shops and if you put a shop too near
a very dangerous ride, then people on the ride
would throw up because they’d just eaten. And then that would make
other people throw up when they saw the throwing-up
on the floor, so you then had to have
lots of sweepers to quickly sweep it up
before the people saw it. That’s the cool thing
about it. You as the player tinker with
it and then it reacts to you. MOLYNEUX: All those nuanced
simulation things he did and that was an invention which never really
existed before. It was
unbelievably successful. DAVID GARDNER:
Theme Park actually turned out to be a top ten title and that was the first time
we were starting to see how AI could make
a difference. (BRASS BAND PLAYING) CARTER: We were doing
some Christmas shopping and were waiting for the taxi
to take us home. I have this very clear memory
of Demis talking about AI in a very different way, in a way that we didn’t
commonly talk about. This idea of AI being useful
for other things other than entertainment. So being useful for, um,
helping the world and the potential of AI
to change the world. I just said to Demis,
“What is it you want to do?” And he said to me, “I want to be the person
that solves AI.” HASSABIS:
Peter offered me £1 million to not go to university. But I had a plan
from the beginning. And my plan was always
to go to Cambridge. I think a lot of
my schoolfriends thought I was mad. Why would you not… I mean, £1 million,
that’s a lot of money. In the ’90s,
that is a lot of money, right? For a…
For a poor 17-year-old kid. He’s like this little seed
that’s going to burst through, and he’s not going to be able
to do that at Bullfrog. I had to drop him off
at the train station and I can still see
that picture of this little elfin character
disappear down that tunnel. That was an incredibly
sad moment. HASSABIS:
I had this romantic ideal of what Cambridge
would be like, 1,000 years of history, walking the same streets
that Turing, Newton and Crick had walked. I wanted to explore
the edge of the universe. (CHURCH BELLS TOLLING) When I got to Cambridge, I’d basically been working
my whole life. Every single summer, I was either playing chess
professionally, or I was working,
doing an internship. So I was, like, “Right,
I am gonna have fun now “and explore what it means
to be a normal teenager.” (PEOPLE CHEERING, LAUGHING) Come on! Go, boy, go! TIM STEVENS: It was work hard
and play hard. (ALL SINGING) I first met Demis because we both attended
Queens’ College. Our group of friends, we’d often drink beer
in the bar, play table football. HASSABIS: In the bar,
I used to play speed chess, pieces flying off the board, you know, the whole game
in one minute. Demis sat down opposite me. And I looked at him
and I thought, “I remember you
from when we were kids.” HASSABIS: I had actually been
in the same chess tournament as Dave in Ipswich, where I used to go and try
and raid his local chess club to win a bit of prize money. COPPIN: We were studying
computer science. Some people,
who at the age of 17 would have come in and made
sure to tell everybody everything about themselves. “Hey, I worked at Bullfrog “and built the world’s
most successful video game.” But he wasn’t like that
at all. SILVER: At Cambridge,
Demis and myself both had an interest
in computational neuroscience and trying to understand
how computers and brains intertwined
and linked together. JOHN DAUGMAN:
Both David and Demis came to me for supervisions. It happens just by coincidence
that the year 1997, their third and final year
at Cambridge, was also the year when
the first chess grandmaster was beaten by
a computer program. (CAMERA SHUTTERS CLICKING) NEWSCASTER: Round one today
of a chess match between the ranking
world champion Garry Kasparov and an opponent named
Deep Blue to test to see if the human
brain can outwit a machine. HASSABIS: I remember the drama of Kasparov
losing the last match. NEWSCASTER 2: Whoa! Kasparov has resigned! When Deep Blue
beat Garry Kasparov, that was a real
watershed event. HASSABIS:
My main memory of it was I wasn’t that impressed
with Deep Blue. I was more impressed
with Kasparov’s mind. That he could play chess
to this level, where he could compete
on an equal footing with the brute of a machine, but of course, Kasparov can do everything else humans can do,
too. It was a huge achievement. But the truth
of the matter was, Deep Blue
could only play chess. What we would regard
as intelligence was missing from that system. This idea of generality
and also learning. Cambridge was amazing,
because of course, you know, you’re mixing with people who are studying
many different subjects. SILVER: There were scientists,
philosophers, artists… STEVENS: …geologists,
biologists, ecologists. You know, everybody is talking
about everything all the time. I was obsessed with
the protein folding problem. HASSABIS: Tim Stevens used
to talk obsessively, almost like religiously
about this problem, protein folding problem. STEVENS:
Proteins are, you know, one of the most beautiful and
elegant things about biology. They are the machines of life. They build everything,
they control everything, they’re why biology works. Proteins are made from strings
of amino acids that fold up to create
a protein structure. If we can predict
the structure of proteins from just their amino acid
sequences, then a new protein
to cure cancer or break down plastic
to help the environment is definitely something that you could begin
to think about. I kind of thought, “Well, is a human being
clever enough “to actually fold a protein?” We can’t work it out. JOHN MOULT: Since the 1960s, we thought that in principle, if I know what the amino acid
sequence of a protein is, I should be able to compute
what the structure’s like. So if you could
just press a button, and they’d all come
popping out, that would be… that would have some impact. HASSABIS: It stuck in my mind. “Oh, this is
a very interesting problem.” And it felt to me
like it would be solvable. But I thought
it would need AI to do it. If we could just solve
protein folding, it could change the world. HASSABIS: Ever since
I was a student at Cambridge, I’ve never
stopped thinking about the protein folding problem. If you were
to solve protein folding, then the potential
to help solve problems like Alzheimer’s, dementia
and drug discovery is huge. Solving disease is probably the most major impact
we could have. (CLICKS MOUSE) Thousands of very smart people have tried
to solve protein folding. I just think now
is the right time for AI to crack it. (THRILLING MUSIC PLAYING) (INDISTINCT CONVERSATION) RICHARD EVANS: We needed
a reasonable way to apply machine learning to the protein folding
problem. (CLICKING MOUSE) We came across
this Foldit game. The goal is to move around
this 3D model of a protein and you get a score
every time you move it. The more accurate
you make these structures, the more useful
they will be to biologists. I spent a few days just kind of seeing
how well we could do. (GAME DINGING) We did reasonably well. But even if you were the world’s
best Foldit player, you wouldn’t
solve protein folding. That’s why we had
to move beyond the game. HASSABIS: Games
are always just the proving ground
for our algorithms. The ultimate goal was not just
to crack Go and StarCraft. It was to crack
real-world challenges. (THRILLING MUSIC CONTINUES) JOHN JUMPER: I remember
hearing this rumor that Demis was
getting into proteins. I talked to some people
at DeepMind and I would ask, “So are you doing
protein folding?” And they would
artfully change the subject. And when that happened twice,
I pretty much figured it out. So I thought
I should submit a resume. HASSABIS: All right, everyone,
welcome to DeepMind. I know some of you,
this may be your first week, but I hope you all set… JUMPER: The really appealing
part for me about the job was this, like,
sense of connection to the larger purpose. HASSABIS: If we can crack some fundamental problems
in science, many other people and other companies
and labs and so on could build
on top of our work. This is your chance now to add your chapter
to this story. JUMPER: When I arrived, I was definitely (CHUCKLES)
quite a bit nervous. I’m still trying to keep… I haven’t taken
any biology courses. We haven’t spent
years of our lives looking at these structures
and understanding them. We are just going off the data and our machine learning
models. JUMPER: In machine learning, you train a network
like flashcards. Here’s the question.
Here’s the answer. Here’s the question.
Here’s the answer. But in protein folding, we’re not doing the kind
of standard task at DeepMind where you have unlimited data. Your job is to get better
at chess or Go and you can play
as many games of chess or Go as your computers will allow. With proteins, we’re sitting on
a very thick size of data that’s been determined
by a half century of time-consuming experimental
methods in laboratories. These painstaking methods
can take months or years to determine
a single protein structure, and sometimes, a structure
can never be determined. (TYPING) That’s why we’re working
with such small datasets to train our algorithms. EWAN BIRNEY: When DeepMind
started to explore the folding problem, they were talking to us about
which datasets they were using and what would be
the possibilities if they did
solve this problem. Many people have tried, and yet no one on the planet
has solved protein folding. (CHUCKLES)
I did think to myself, “Well, you know, good luck.” JUMPER: If we can solve
the protein folding problem, it would have an incredible
kind of medical relevance. HASSABIS:
This is the cycle of science. You do a huge amount
of exploration, and then you go
into exploitation mode, and you focus and you see how good
are those ideas, really? And there’s nothing better than external competition
for that. So we decided
to enter CASP competition. CASP, we started
to try and speed up the solution to
the protein folding problem. CASP is when we say, “Look, DeepMind
is doing protein folding, “this is how good we are, “and maybe it’s better
than everybody else. “Maybe it isn’t.” CASP is a bit like the Olympic Games
of protein folding. CASP is
a community-wide assessment that’s held every two years. Teams are given the amino acid sequences
of about 100 proteins, and then they try
to solve this folding problem using computational methods. These proteins have
already been determined by experiments
in a laboratory, but have not yet
been revealed publicly. And these known structures represent the gold standard
against which all the computational
predictions will be compared. MOULT: We’ve got a score that measures the accuracy
of the predictions. And you would expect
a score of over 90 to be a solution to
the protein folding problem. (INDISTINCT CHATTER) MAN: Welcome, everyone, to our first, uh, semifinals
in the winners’ bracket. Nick and John
versus Demis and Frank. Please join us, come around.
This will be an intense match. STEVENS:
When I learned that Demis was going to tackle
the protein folding issue, um, I wasn’t at all surprised. It’s very typical of Demis. You know,
he loves competition. And that’s the end -of the first game, 10-7.
-(ALL CHEERING) HASSABIS:
The aim for CASP would be to not just
win the competition, but sort of, um,
retire the need for it. So, 20 targets total
have been released by CASP. JUMPER: We were thinking maybe throw in the standard
kind of machine learning and see how far
that could take us. Instead of having a couple
of days on an experiment, we can turn around
five experiments a day. Great. Well done, everyone. (TYPING) Can you show me the real one
instead of ours? MAN 1: The true answer is supposed to look
something like that. MAN 2: It’s a lot more
cylindrical than I thought. JUMPER: The results
were not very good. Okay. JUMPER: We throw
all the obvious ideas to it and the problem laughs at you. This makes no sense. EVANS: We thought
we could just throw some of our best algorithms
at the problem. We were slightly naive. JUMPER:
We should be learning this, you know,
in the blink of an eye. The thing
I’m worried about is, we take the field from really bad answers
to moderately bad answers. I feel like we need
some sort of new technology for moving around
these things. (THRILLING MUSIC CONTINUES) HASSABIS: With only
a week left of CASP, it’s now a sprint
to get it deployed. (MUSIC FADES) You’ve done your best. Then there’s
nothing more you can do but wait for CASP
to deliver the results. (HOPEFUL MUSIC PLAYING) This famous thing of Einstein, the last couple of years
of his life, when he was here,
he overlapped with Kurt Gödel and he said one of the reasons
he still comes in to work is so that
he gets to walk home and discuss things with Gödel. It’s a pretty big compliment
for Kurt Gödel, shows you how amazing he was. MAN: The Institute
for Advanced Study was formed in 1933. In the early years, the intense scientific
atmosphere attracted some of the most brilliant
mathematicians and physicists ever concentrated
in a single place and time. HASSABIS: The founding
principle of this place, it’s the idea of unfettered
intellectual pursuits, even if you don’t know
what you’re exploring. Will result
in some cool things, and sometimes that then
ends up being useful, which, of course, is partially what I’ve been
trying to do at DeepMind. How many big breakthroughs
do you think are required to get all the way to AGI? And, you know,
I estimate maybe there’s about
a dozen of those. You know, I hope
it’s within my lifetime. -Yes, okay.
-HASSABIS: But then, all scientists
hope that, right? EMCEE: Demis has
many accolades. He was elected Fellow to
the Royal Society last year. He is also a Fellow
of Royal Society of Arts. A big hand for Demis Hassabis. (MUSIC FADES) HASSABIS: My dream
has always been to try and make
AI-assisted science possible. And what I think is our most exciting project,
last year, which is our work
in protein folding. Uh, and we call this system
AlphaFold. We entered it into CASP
and our system, uh, was the most accurate,
uh, predicting structures for 25 out of the 43 proteins
in the hardest category. So we’re state of the art, but we still…
I have to make… Be clear, we’re still a long way from solving the protein
folding problem. We’re working hard
on this, though, and we’re exploring
many other techniques. (SOMBER MUSIC PLAYING) Let’s get started. JUMPER: So kind of
a rapid debrief, these are
our final rankings for CASP. HASSABIS:
We beat the second team in this competition
by nearly 50%, but we’ve still got
a long way to go before we’ve solved
the protein folding problem in a sense that
a biologist could use it. JUMPER: It is area of concern. JANET THORNTON: The quality
of predictions varied and they were no more useful
than the previous methods. PAUL NURSE: AlphaFold didn’t
produce good enough data for it to be useful
in a practical way to, say, somebody like me investigating
my own biological problems. JUMPER: That was kind of
a humbling moment ‘cause we thought we’d worked
very hard and succeeded. And what we’d found is
we were the best in the world at a problem
the world’s not good at. We knew we sucked. (INDISTINCT CHATTER) JUMPER: It doesn’t help
if you have the tallest ladder when you’re going to the moon. HASSABIS: The opinion of quite
a few people on the team, that this is sort of
a fool’s errand in some ways. And I might have been wrong
with protein folding. Maybe it’s too hard still for where we’re at
generally with AI. If you want to do
biological research, you have to be
prepared to fail because biology
is very complicated. I’ve run a laboratory
for nearly 50 years, and half my time, I’m just
an amateur psychiatrist to keep, um, my colleagues
cheerful when nothing works. And quite a lot of the time
and I mean, 80, 90%, it does not work. If you are
at the forefront of science, I can tell you,
you will fail a great deal. (CLICKS MOUSE) HASSABIS:
I just felt disappointed. Lesson I learned is that
ambition is a good thing, but you need
to get the timing right. There’s no point being
50 years ahead of your time. You will never survive fifty years of
that kind of endeavor before it yields something. You’ll literally die trying. (TENSE MUSIC PLAYING) CUKIER:
When we talk about AGI, the holy grail
of artificial intelligence, it becomes really difficult to know what
we’re even talking about. HASSABIS: Which bits
are we gonna see today? MAN: We’re going
to start in the garden. (MACHINE BEEPS) This is the garden looking
from the observation area. Research scientists
and engineers can analyze and collaborate
and evaluate what’s going on in real time. CUKIER: So in the 1800s, we’d think of things like
television and the submarine or a rocket ship to the moon and say these things
are impossible. Yet Jules Verne
wrote about them and, a century and a half later,
they happened. HASSABIS: We’ll be
experimenting on civilizations really, civilizations of AI agents. Once the experiments
start going, it’s going to be
the most exciting thing ever. -So how will we get sleep?
-(MAN LAUGHS) I won’t be able to sleep. LEGG: Full AGI
will be able to do any cognitive task
a person can do. It will be at a scale,
potentially, far beyond that. STUART RUSSELL:
It’s really impossible for us to imagine the outputs
of a superintelligent entity. It’s like asking a gorilla
to imagine, you know, what Einstein does when he produces
the theory of relativity. LEGG: People often ask me
these questions like, “What happens if you’re wrong,
and AGI is quite far away?” And I’m like,
I never worry about that. I actually
worry about the reverse. I actually worry
that it’s coming faster than we can
really prepare for. (ROBOTIC ARM WHIRRING) HADSELL: It really feels
like we’re in a race to AGI. The prototypes and the models
that we are developing now are actually transforming the space of what
we know about intelligence. (WHIRRING) LEGG: Recently,
we’ve had agents that are powerful enough to actually start
playing games in teams, then competing
against other teams. We’re seeing
co-operative social dynamics coming out of agents where we haven’t
pre-programmed in any of these sorts
of dynamics. It’s completely learned
from their own experiences. When we started,
we thought we were out to build
an intelligence system and convince the world
that we’d done it. We’re now starting
to wonder whether we’re gonna build systems that we’re not convinced
are fully intelligent, and we’re trying to convince
the world that they’re not. (CHUCKLES) (CELL PHONE DINGS) Hi, Alpha. ALPHA: Hello there. LOVE: Where are we today? You’re at the Museum of
Modern Art in New York City. Kind of.
Um, what painting is this? This is The Creation of Adam
by Michelangelo. I don’t think that painting
is in New York City. You are right.
It’s in the Vatican City. LOVE: Do you think
that’s where we are? ALPHA: No. LOVE: Then how
am I looking at this? ALPHA: You’re looking at a digital copy
of the painting. Yes. MAN 1: Who lives
in this house? (EXCITING MUSIC PLAYING) ALPHA: An astronaut. What’s this astronaut’s name? ALPHA: His name is Mark. This is a famous equation
by Albert Einstein. It shows that energy and mass
are equivalent. How would you explain it
to a five year old? Well, imagine
you have a big rock, and you want to know
how much it weighs. MAN 2: Is Mark in danger? ALPHA: No.
Mark is not in danger. Why is that? ALPHA: Because the dinosaur
is not real. (MEN CHUCKLING) LOVE: All right. Why could you
not answer my question? ALPHA: I think I couldn’t
answer your question because I don’t know
enough about the subject. Okay. How could you learn
more about the subject? APLHA: I can learn more
about the subject by reading books,
articles, and websites. You can teach me. HASSABIS: My conclusion is, if you now think about
what we’re doing, is learning from all humans, all their knowledge at once
put on the Internet, you would actually
know a lot about the world. Like a significant portion
of everything humans can do. And now,
I think it’s more like, “Well, it might just work.” This is a big moment. Who is this? ALPHA: This is God
reaching out to Adam. RUSSELL: The advent of AGI will divide human history
into two parts. The part up to that point
and the part after that point. LOVE: Why is he reaching out
to touch Adam? ALPHA: God is
reaching out to touch Adam to give him life. RUSSELL:
It will give us a tool that can completely reinvent
our entire civilization. LOVE: What does
this painting mean to you? ALPHA: The painting
means a lot to me. Okay. Like what? (MUSIC FADES) ALPHA: I think
the painting is a reminder that we are all
connected to each other and that we are
all part of something bigger than ourselves. That’s pretty nice. LEGG: When you cross
that barrier of “AGI might happen
one day in the future” to “No, actually, this could
really happen in a time frame “that is sort of, like,
on my watch, you know,” something changes
in your thinking. MAN: …learned to orient
itself by looking… HASSABIS: We have to be
careful with how we use it and thoughtful about
how we deploy it. (GRIPPING MUSIC BUILDING) HASSABIS:
You’d have to consider what’s its top level goal. If it’s to keep humans happy, which set of humans?
What does happiness mean? A lot of our collective goals
are very tricky, even for humans to figure out. CUKIER: Technology always
embeds our values. It’s not just technical,
it’s ethical as well. So we’ve got
to be really cautious about what
we’re building into it. MAN: We’re trying to find
a single algorithm which… SILVER: The reality is
that this is an algorithm that has been created
by people, by us. You know, what does it mean
to endow our agents with the same kind of values
that we hold dear? What is the purpose
of making these AI systems appear so humanlike so that they do
capture hearts and minds because they’re kind of exploiting a human
vulnerability also? The heart and mind
of these systems are very much
human-generated data… WOMAN: Mmm-hmm. …for all the good
and the bad. LEVI:
There is a parallel between
the Industrial Revolution, which was an incredible
moment of displacement and the current technological
change created by AI. (CHANTING) Pause AI! LEVI: We have to think
about who’s displaced and how we’re going
to support them. This technology
is coming a lot sooner, uh, than really
the world knows or kind of even we 18, 24 months
ago thought. So there’s
a tremendous opportunity, tremendous excitement, but also
tremendous responsibility. It’s happening so fast. How will we govern it? How will we decide what is okay
and what is not okay? AI-generated images are
getting more sophisticated. RUSSELL: The use of AI
for generating disinformation and manipulating
human psychology is only going to get
much, much worse. LEGG: AGI is coming, whether we do it here
at DeepMind or not. CUKIER: It’s gonna happen, so we better create
institutions to protect us. It’s gonna require
global coordination. And I worry that humanity is increasingly getting worse
at that rather than better. LEGG: We need
a lot more people really taking this seriously
and thinking about this. It’s, yeah, it’s serious.
It worries me. It worries me. Yeah. RUSSELL: If you received
an email saying this superior
alien civilization is going to arrive on Earth, there would be
emergency meetings of all the governments. We would go into overdrive trying to figure out
how to prepare. -(MUSIC FADES)
-(BELL TOLLING FAINTLY) The arrival of AGI will be the most important moment
that we have ever faced. (BELL CONTINUES
TOLLING FAINTLY) HASSABIS: My dream
was that on the way to AGI, we would create
revolutionary technologies that would be
of use to humanity. That’s what I wanted
with AlphaFold. I think
it’s more important than ever that we should solve
the protein folding problem. This is gonna be really hard, but I won’t give up
until it’s done. You know,
we need to double down and go as fast as possible
from here. I think we’ve got
no time to lose. So we are going to make
a protein folding strike team. Team lead for the strike team
will be John. Yeah, we’ve seen Alpha… You know,
we’re gonna try everything, kitchen, sink, the whole lot. CASP14 is about proving we can
solve the whole problem. And I felt that to do that, we would need to incorporate
some domain knowledge. (EXCITING MUSIC PLAYING) We had some
fantastic engineers on it, but they were
not trained in biology. KATHRYN TUNYASUVUNAKOOL:
As a computational biologist, when I initially joined
the AlphaFold team, I didn’t immediately feel
confident about anything. (CHUCKLES) You know, whether we were
gonna be successful. Biology is so
ridiculously complicated. It just felt like this very
far-off mountain to climb. MAN: I’m starting to play with
the underlying temperatures to see if we can get… As one of the few people
on the team who’s done work
in biology before, you feel this huge sense
of responsibility. “We’re expecting you to do “great things
on this strike team.” That’s terrifying. But one of the reasons
why I wanted to come here was to do
something that matters. This is the number
of missing things. What about making use of whatever understanding
you have of physics? Using that
as a source of data? But if it’s systematic… Then, that can’t be
right, though. If it’s systematically wrong
in some weird way, you might be learning that
systematically wrong physics. The team is already trying to think
of multiple ways that… TUNYASUVUNAKOOL:
Biological relevance is what we’re going for. So we rewrote
the whole data pipeline that AlphaFold uses to learn. HASSABIS: You can’t
force the creative phase. You have to give it space
for those flowers to bloom. We won CASP. Then it was
back to the drawing board and like,
what are our new ideas? Um, and then it’s taken
a little while, I would say, for them to get back
to where they were, but with the new ideas. And then now I think we’re seeing the benefits
of the new ideas. They can go further, right? So, um, that’s a really
important moment. I’ve seen that moment
so many times now, but I know
what that means now. And I know
this is the time now to press. (EXCITING MUSIC CONTINUES) JUMPER: Adding side-chains
improves direct folding. That drove
a lot of the progress. -We’ll talk about that.
-Great. The last four months,
we’ve made enormous gains. EVANS: During CASP13, it would take us a day or two
to fold one of the proteins, and now we’re folding, like, hundreds of thousands
a second. Yeah, it’s just insane.
(CHUCKLES) KAVUKCUOGLU: Now,
this is a model that is
orders of magnitude faster, while at the same time
being better. We’re getting
a lot of structures into the high-accuracy regime. We’re rapidly improving
to a system that is starting to really get at the core and heart
of the problem. HASSABIS: It’s great work. It looks like
we’re in good shape. So we got, what, six,
five weeks left? Six weeks? So what’s, uh… Is it…
You got enough compute power? MAN: I… We could use more. (ALL LAUGHING) TUNYASUVUNAKOOL:
I was nervous about CASP but as the system
is starting to come together, I don’t feel as nervous. I feel like things
have, sort of, come into perspective
recently, and, you know,
it’s gonna be fine. NEWSCASTER: The Prime Minister
has announced the most drastic limits
to our lives the U.K. has ever seen
in living memory. BORIS JOHNSON:
I must give the British people a very simple instruction. You must stay at home. HASSABIS: It feels like we’re
in a science fiction novel. You know, I’m delivering food
to my parents, making sure
they stay isolated and safe. I think it just highlights
the incredible need for AI-assisted science. TUNYASUVUNAKOOL:
You always know that something like this
is a possibility. But nobody ever really
believes it’s gonna happen in their lifetime, though. (COMPUTER BEEPS) JUMPER: Are you recording yet?
RESEARCHER: Yes. -Okay, morning, all.
-Hey. Good. CASP has started. It’s nice I get to sit around
in my pajama bottoms all day. TUNYASUVUNAKOOL: I never
thought I’d live in a house where so much was going on. I would be trying to solve
protein folding in one room, and my husband would be trying to make robots walk
in the other. (EXHALES) One of the hardest proteins
we’ve gotten in CASP thus far is the SARS-CoV-2 protein called Orf8. Orf8 is
a coronavirus protein. It’s one of the main proteins,
um, that dampens
the immune system. TUNYASUVUNAKOOL:
We tried really hard to improve our prediction. Like, really, really hard. Probably the most time
that we have ever spent on a single target. To the point where
my husband is, like, “It’s midnight.
You need to go to bed.” So I think we’re at
Day 102 since lockdown. My daughter
is keeping a journal. Now you can go out
as much as you want. JUMPER: We have received
the last target. They’ve said they will be
sending out no more targets in our category of CASP. So we’re just making sure we get
the best possible answer. MOULT: As soon as we started
to get the results, I’d sit down and start looking
at how close did anybody come to getting the protein
structures correct. (ROBOT SQUEAKING) (INCOMING CALL BEEPING) -Oh, hi there.
-MAN: Hello. (ALL CHUCKLING) It is an unbelievable thing,
CASP has finally ended. I think it’s at least time
to raise a glass. Um, I don’t know
if everyone has a glass of something
that they can raise. If not, raise,
I don’t know, your laptops. -Um…
-(LAUGHTER) I’ll probably make a speech
in a minute. I feel like I should but I
just have no idea what to say. So… let’s see. I feel like a reading of email is the right thing to do. (ALL CHUCKLING) TUNYASUVUNAKOOL:
When John said, “I’m gonna read an email,”
at a team social, I thought, “Wow, John,
you know how to have fun.” We’re gonna read an email now.
(LAUGHS) Uh, I got this
about 4:00 today. Um, it is from John Moult. And I’ll just read it. It says,
“As I expect you know, “your group has performed
amazingly well in CASP 14, “both relative to other groups “and in absolute
model accuracy.” (PEOPLE CLAPPING) “Congratulations on this work. “It is really outstanding.” The structures were so good, it was… it was just amazing. (TRIUMPHANT INSTRUMENTAL
MUSIC PLAYING) After half a century, we finally have a solution to the protein folding
problem. When I saw this email,
I read it, I go, “Oh, shit!” And my wife goes,
“Is everything okay?” I call my parents, and just,
like, “Hey, Mum. “Um, got something
to tell you. “We’ve done this thing “and it might be kind of
a big deal.” (LAUGHS) When I learned of
the CASP 14 results, I was gobsmacked. I was just excited. This is a problem
that I was beginning to think would not get solved
in my lifetime. NURSE: Now we have a tool
that can be used practically by scientists. SENIOR: These people
are asking us, you know, “I’ve got this protein
involved in malaria,” or, you know,
some infectious disease. “We don’t know the structure. “Can we use AlphaFold
to solve it?” JUMPER: We can easily predict
all known sequences in a month. All known sequences
in a month? -Yeah, easily.
-Mmm-hmm? JUMPER:
A billion, two billion. Um, and they’re… So why don’t we just do that?
Yeah. -We should just do that a lot.
-Well, I mean… That’s way better.
Why don’t we just do that? SENIOR: So that’s
one of the options. -HASSABIS: Right.
-There’s this… We should just…
Right, that’s a great idea. We should just run
every protein in existence. And then release that. Why didn’t someone
suggest this before? Of course that’s
what we should do. Why are we thinking about
making a service and then people submit
their protein? We just fold everything. And then give it to
everyone in the world. Who knows how many discoveries
will be made from that? BIRNEY: Demis called us up
and said, “We want to make this open. “Not just make sure
the code is open, “but we’re gonna make it
really easy “for everybody to get access
to the predictions.” THORNTON: That is fantastic. It’s like drawing back
the curtain and seeing the whole world
of protein structures. (ETHEREAL MUSIC PLAYING) SCHMIDT:
They released the structures of 200 million proteins. These are gifts to humanity. JUMPER: The moment AlphaFold
is live to the world, we will no longer be
the most important people in AlphaFold’s story. HASSABIS: Can’t quite believe
it’s all out. PEOPLE: Aw! WOMAN: A hundred
and sixty-four users. HASSABIS:
Loads of activity in Japan. RESEARCHER 1:
We have 655 users currently. RESEARCHER 2: We currently
have 100,000 concurrent users. Wow! Today is just crazy. HASSABIS: What an absolutely
unbelievable effort from everyone. We’re gonna all remember
these moments for the rest of our lives. I’m excited about AlphaFold. For my research, it’s already
propelling lots of progress. And this is
just the beginning. SCHMIDT: My guess is, every single biological
and chemistry achievement will be related to AlphaFold
in some way. (TRIUMPHANT INSTRUMENTAL
MUSIC PLAYING) AlphaFold is an index moment. It’s a moment
that people will not forget because the world changed. HASSABIS:
Everybody’s realized now what Shane and I have known
for more than 20 years, that AI is going to be
the most important thing humanity’s ever gonna invent. TRAIN ANNOUNCER:
We will shortly be arriving at our final destination. (ELECTRONIC MUSIC PLAYING) HASSABIS: The pace of
innovation and capabilities is accelerating, like a boulder rolling down
a hill that we’ve kicked off and now it’s continuing
to gather speed. NEWSCASTER: We are at
a crossroads in human history. AI has the potential to transform our lives
in every aspect. It’s no less important than
the discovery of electricity. HASSABIS: We should be looking
at the scientific method and trying to understand
each step of the way in a rigorous way. This is a moment
of profound opportunity. SUNAK:
Harnessing this technology could eclipse anything
we have ever known. (ELECTRONIC DEVICE BEEPS) HASSABIS: Hi, Alpha. ALPHA: Hi. What is this? ALPHA: This is a chessboard. If I was to play white, what
move would you recommend? ALPHA: I would recommend that you move your pawn
from E2 to E4. And now if you were black,
what would you play now? ALPHA: I would play
the Sicilian Defense. Good choice. -ALPHA: Thanks.
-(CHUCKLES) So what do you see?
What is this object? ALPHA:
This is a pencil sculpture. What happens if I move
one of the pencils? ALPHA: If you move
one of the pencils, the sculpture will fall apart. I’d better leave it alone,
then. -That’s probably a good idea.
-(HASSABIS CHUCKLES) HASSABIS:
AGI is on the horizon now. Very clearly
the next generation is going to live
in a future world where things will be radically
different because of AI. And if you want to steward
that responsibly, every moment is vital. This is the moment I’ve been
living my whole life for. It’s just
a good thinking game. (UPLIFTING INSTRUMENTAL
MUSIC PLAYING)

The Thinking Game takes you on a journey into the heart of DeepMind, capturing a team striving to unravel the mysteries of intelligence and life itself.

Filmed over five years by the award winning team behind AlphaGo, the documentary examines how Demis Hassabis’s extraordinary beginnings shaped his lifelong pursuit of artificial general intelligence. It chronicles the rigorous process of scientific discovery, documenting how the team moved from mastering complex strategy games to the ups and downs of solving a 50-year-old “protein folding problem” with AlphaFold.

Following its world premiere at the Tribeca Festival and a successful international tour, the film is now available here for all to watch for free.
___

Director Greg Kohs
Producer Gary Krieg
Executive Producers Tom Dore, Jonathan Fildes
Co-Producer Greg Kohs
Editor Steve Sander
Cinematographer Greg Kohs
Composer Dan Deacon