AlphaGo (2017)

1
It's intensely contemplative.
It is almost hypnotic.
It's like putting your hand
on the third rail of the universe.
If you play Go seriously,
there is a chance
that you will get exposed
to this experience that...
is kind of like nothing else
on the planet.
Go is putting you in a place
where you're always
at the very farthest reaches
of your capacity.
There's a reason
that people have been playing Go
for thousands and thousands
of years, right?
It's not just that they want
to understand Go.
They want to understand
what understanding is.
And maybe that is truly
what it means to be human.
When I was a kid,
I loved playing games.
I started off with board games like chess.
And then I bought my first computer
when I was eight
with winnings from a chess tournament.
Ever since then, I felt that computers
were this sort of magical device
that could extend the power of your mind.
Virtual environments and games,
we think they're the perfect platform
for developing and testing AI algorithms.
Games are very convenient,
in that a lot of them have scores,
so it's very easy
to measure incremental progress.
So I'm going to show you a few videos
of the agent system, the AI.
So let's start off with Breakout.
So here you control the bat and ball,
and you're trying to break
through this rainbow-colored wall.
The agent system has to learn
everything for itself,
just from the raw pixels.
It doesn't know what it's controlling.
It doesn't even know
what the object of the game is.
Now, at the beginning, after 100 games,
you can see 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,
and pretty much gets
the ball back every time.
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.
The researchers working on this,
the amazing AI developers,
well, they're not so good at Breakout,
and they didn't know about that strategy.
So, they learned something
from their own system,
which is, uh, you know,
pretty funny and quite instructive,
I think, about the potential
for general AI.
So, for us, what's the next step now?
Go is the most complex game,
pretty much, ever devised by man.
Beating a professional player at Go
is a long-standing, grand challenge
of AI research.
Wow.
You really want to kill me, yeah?
No, I don't want to,
but I'm trying to survive.
Yes, that's it. That's it.
I am a Go professional player.
I am also a European Champion.
With that move,
you're going to lose for sure.
You were happy to defeat me
but then you panicked.
You're afraid that victory
will escape you.
In general, when you're afraid,
victory always escapes you.
I was born in China.
When I was 18, I wanted to change my life.
That was why I went to France.
I wanted to try to forget Go.
But it was impossible.
Because all the things I learned
in my life was through Go.
It looks like a mirror.
When I see Go, I also see myself.
For me, Go is real life.
Dear Mr. Fan.
My name is Demis Hassabis.
I run an artificial intelligence company
based in London called DeepMind.
As the strongest Go player in Europe,
we would like to invite you
to our offices in London,
both to meet you in person
and to share with you
an exciting Go project
that we are working on.
If you would be interested
in coming to visit us,
please let us know.
Many thanks. Kind regards, Demis.
When I saw this e-mail,
I didn't know if it was true or not.
I will accept this. Why not?
For me, everything is an adventure.
I want to go visit DeepMind
to know what this Go project is.
On my first visit,
I thought maybe they wanted me
to sit in a special room,
attach numerous wires to my head,
and also my body,
let me play to scan my brain.
I don't know,
to make some research.
We needed to get Fan Hui
to DeepMind
to see we were a serious operation,
and we were serious people...
doing proper research.
And the search time
is getting better and better
as we go from, for example,
one to eight, in any one of these,
it always goes up.
We think of DeepMind as...
kind of like an Apollo program effort
for AI.
Our mission is to fundamentally
understand intelligence,
and recreate it artificially.
And then, once we've done that,
we feel that we can use
that technology to help society
solve all sorts of other problems.
If you step through
the actual game,
we can see, kind of, how AlphaGo thinks,
what's the most likely variation
that it thinks will happen.
We've been working
on AlphaGo,
our program to play Go,
for just under two years now.
All the little patterns cascade together,
layer after layer, after layer,
after layer.
I started talking about the game of Go
with Demis more than 20 years ago.
And so, this has been
a really long journey.
The game of Go is the holy grail
of artificial intelligence.
For many years,
people have looked at this game
and they have thought,
"Wow, this is just too hard."
Everything we have 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.
If we can crack Go,
we know we've done something special.
So with Fan Hui,
we started talking around
what the real purpose of the visit was.
It wasn't just a Go project
we wanted him to help with,
that actually we wanted to play him,
and we had a very strong program.
It's okay, it's just a program.
It's so easy.
So, it will be easy to play.
A lot of people have thought
that it was decades away.
Some people thought it would be never.
Because they felt that to succeed at Go,
you needed human intuition.
Go is the world's oldest
continuously played board game.
And, in some sense,
it is one of the simplest
and also most abstract.
There's only one type of piece.
There's only one type of move.
You just place that piece on the board.
And then, your goal is to create
a linked group of your stones
that surrounds some empty territory.
And when you surround enemy stones,
you capture them,
and remove them from the board.
You earn points by surrounding territory.
And, at the end of the game,
the person with the most territory wins.
It seems really simple.
But, then you sit down to play,
and you realize right away,
it's like, "Well, I technically know
what I'm allowed to do,
but I have no clue what I should do."
Go is incredibly challenging
for computers to tackle
because compared to, say, chess,
the number of possible moves
in a position is much larger.
In chess, it's about 20.
In Go, it's about 200.
And the number of possible
configurations of the board
is more than the number
of atoms in the universe.
So even if you took all
the computers in the world,
and ran them for a million years,
that wouldn't be enough compute power
to calculate all the possible variations.
If you ask a great Go player
why they played a particular move,
sometimes, they will just tell you
it felt right.
So, we have to come up
with some kind of clever algorithm
to mimic what people do
with their intuition.
- Nice to meet you.
- Oh, hello. Nice to meet you, too.
So with Fan Hui, we agreed
a best-of-five match.
And we agreed it would be filmed.
You know, we would treat it
as a serious match.
I played with AlphaGo.
Aja Huang pushed the stones for AlphaGo.
And, of course, I thought I will win
the game with AlphaGo
because it's just a program.
Good game.
Have a nice game.
The first game,
I made some mistakes,
and I lost the game.
Oh. Strange. Very strange game.
In that first match,
I think something clicked for him
that this wasn't an ordinary Go program.
We weren't just doing
the same as everyone else,
that something new was happening.
The second game,
I tried to change my style.
But the problem
is I also lost the second game.
And also the third game,
the fourth game,
even the last game.
I lost, five-zero.
AlphaGo won all the games.
After losing, losing, losing,
you can feel his pressure
is getting heavier and heavier.
And several times after the game,
he said he wanted to go out for fresh air.
I said, "Oh, I can go with you
and have a chat."
He said, "No, I want to go by myself."
I felt something very strange.
I lost to a program.
And I didn't understand myself anymore.
I worried whether
he would even come back.
You know, he seemed very low,
and he spent about an hour away
from the office.
He came back completely changed.
I told Maddy,
"This is the first time in history
that a human professional Go player
lost to a program.
She told me, "Yes, we know.
But for you... Are you okay?"
I told her...
"No, and yes."
I am not happy to lose a game,
but I'm very happy to play for history.
Artificial intelligence researchers
have solved the game of Go
a decade earlier than expected.
The computer named AlphaGo,
was able to beat
the European human champion.
Artificial intelligence researchers
have made a significant breakthrough.
It really is a big leap forward.
There's a big difference between
the way the IBM computer beat Kasparov,
which was programmed
by expert chess players,
and the way the Go-playing computer,
more or less, learned itself.
The way we start off training AlphaGo
is by showing it 100,000 games
that strong amateurs have played,
that we have downloaded from the Internet.
And we first initially get AlphaGo
to mimic the human player.
And then, through self-playing,
reinforcement learning,
it plays against different versions
of itself many millions of times,
and learns from its errors.
The specific ideas
that are driving AlphaGo
are going to drive our future.
The technologies
at the heart of AlphaGo,
they are what are called
deep neural networks
which essentially mimic
the web of neurons in the human brain.
It's a very old idea,
but recently, due to increases
in computing power,
these neural networks
have become extremely powerful,
almost overnight.
Big neural networks
that operate on big data
can achieve surprising things.
AlphaGo found a way to learn
how to play Go.
Learning is the key thing here.
It is a machine learning.
The whole beauty
about these types of algorithms
is that because they are learning
for themselves,
they can go beyond
what we, as the programmers,
know how to do,
and allow us to make new breakthroughs
in areas of science and medicine.
So, AlphaGo is one significant step
towards that ultimate goal.
My wife called me.
She told me, "Don't check the Internet.
Don't connect to the Internet,
because people are saying terrible things
about your match with AlphaGo."
"Fan Hui has been living
for a long time in Europe,
he is not a real professional,
he's just an amateur player."
This is so hard for me, so hard.
The Go world was skeptical
about how strong really was AlphaGo,
and how much further did it need
to get to beat the top professionals.
So our program is improving over time,
and we want to push
the AI algorithm to the limit,
and see how far this kind
of self-improving process can go.
So, we needed to look
for an even greater challenge.
A match like no other
is about to get underway in South Korea.
Se-dol Lee,
the long-reigning global champ...
This guy is a genius...
...who will take on
artificial intelligence program AlphaGo
in the ultimate
human versus machine smackdown.
This is a huge moment for both
the world of artificial intelligence
and I think, the world of Go.
So far, AlphaGo has beaten
every challenge we have given it.
But we won't know
its true strength until we play somebody
who is at the top of the world,
like Se-dol Lee.
We chose Se-dol Lee
because we wanted
a legendary, historic player.
Somebody who has been acknowledged
as the greatest player of the last decade.
I don't want to sound too arrogant,
but, still, I think I have the advantage.
"I will not be too arrogant,
but I don't think
that it will be a very close match.
The level of the player that AlphaGo
went against in October
is not the same level as me.
So, given that a couple of months
has only passed,
I don't think that it is enough for it
to be able to catch up with me.
My hope is that it will be either
five-zero for me, or maybe four to one.
So the critical point for me
is to make sure I do not lose one."
High five, please.
We don't know how well
our system will play against someone
as creative as Se-dol Lee.
Also, he's very famous
- for very creative fighting play.
- Yeah, yeah. Yeah.
So, this could be difficult for us,
but we'll see.
Maybe, in some ways,
he's the most difficult opponent
we could pick.
Yeah.
There was still so much
of a question
about whether or not they could beat
someone like Se-dol Lee.
Fan Hui is a good player,
but he's nothing
like the very top players.
Se-dol Lee
is a 9-dan professional.
Nine-dan professional is the highest rank
that you can achieve in Go.
The ranking system.
Fan Hui, who is the European Champion,
is just a 2-dan professional.
On the other hand, at the very top,
Se-dol Lee is a 9-dan professional.
Se-dol Lee is to Go
what Roger Federer is to tennis.
He is playing in Wimbledon
to win the Grand Slam.
And it's not just this year.
He'll be there next year,
and the year after that,
and the year after that.
Of course, the Internet,
and all the Go community, everyone,
maybe not 100%, maybe 99.999%
think Se-dol Lee will win very easily.
Teacher, I got stuck.
How do I go from here?
Good.
Hold the stone properly.
Place the stone nicely, not like that.
And hold it nicely, like this.
Go is taken seriously in Korea.
It's so much part of the culture,
like breathing, or, you know,
like taking swimming lessons,
or something.
On the surface, it's a game,
but inside, it has a very deep philosophy.
The Go board reflects
the individual who's playing.
The truth is going to show itself
on the board.
You won't be able to hide it.
In ancient China, Japan, Korea,
Go is one of the four noble things,
like four accomplishments for literati,
with music, poetry, and painting.
So, people think the Go players
are very smart and very noble.
It's okay to lose,
as long as you played your best.
Bow to your opponent.
Master Kwon started this school,
so that he could produce
great players in Korea.
Se-dol Lee, when he was eight years old,
he came in here, attending class,
9:00 a.m. until 9:00 p.m.
for seven days.
And then, he stayed with Master Kwon.
I remember his round face
and dark brown eyes.
Since he came
from an island, he easily drew attention.
But unlike the other children,
his eyes shone brightly.
7TH WORLD GO CHAMPIONSHIP
Se-dol Lee was a little boy
that I remember when I was a student
studying Go.
He was very young and rural enough
to think that pizza grew on trees.
When I was young,
I started playing Go
because I thought it was fun
to beat guys who were older than me.
Since then, it has become
a type of creation of mine.
I want my style of Go
to be something different,
something new,
my own thing.
Something that no one
has thought of before.
Se-dol Lee plays things
that are interesting, where you felt like
he is beyond winning and losing.
He wants to do something that's innovative
or takes things to the next level.
So every Go player
studies his game for sure.
21ST ASIAN TV CUP CHAMPIONSHIP
Roughly for ten years, he dominated
the professional Go world.
He won 18 World Championships.
Se-dol Lee
is a genius of the century.
2012 GO CHAMPIONSHIP AWARD CEREMONY
AT GRAND CENTRAL HOTEL, CHINA
Even when I look back now,
I am proud of him.
And I'm proud of myself, too.
Definitely, Se-dol Lee
is going to win the game, five all.
IF YOU DESIRE TO WIN, YOU WILL LOSE
Yeah, all the games.
He's going to win all the games.
We had our evaluation match
last week.
We won a game, and we lost a game.
And we lost a game in a way
that would have made us look
extremely foolish
if that happened publicly.
It means that we still have work to do,
and we need to take this really seriously.
There's just too much risk that, actually,
we could lose, overall.
Not only that, but we could lose in a way
that makes us look rather silly.
- Yeah.
- So, I guess, um...
Yeah, Aja, did you want to say anything
about what you're trying?
I'm working hard.
So, we're working round the clock
at the moment,
training our algorithms further,
trying to incrementally keep on improving
right up to the moment
where we have the match.
We have collected together
people with different skill sets
onto the same team.
So we have researchers,
engineers, valuation guys.
But you mean what happens
if that creates something new
- which wasn't in heavy node?
- Yeah.
Then we don't see it.
You'll see it.
- Okay.
- The first thing to know...
I was thinking
we have the perfect solution,
- but it's not perfect.
- Yeah.
Aja is the lead programmer,
and built the original search engine.
So Aja's responsibility
is quite a big one.
He'll be the one sitting
opposite Se-dol Lee
and actually playing
the moves AlphaGo makes.
- I'm feeling excited.
- Yeah?
Se-dol Lee is a great player,
and I feel honored to play with him.
Many of my friends,
they are very excited about the match.
They keep telling me
that the whole world is watching.
"Just prepare AlphaGo."
Hello.
Hello, Demis. Can you hear us?
Yes, I can hear you, guys.
Can you hear me okay?
Yes. Perfect.
"I was curious to meet you,
such an amazing developer
who made AlphaGo.
It's very nice to meet you."
Thank you, thank you. Likewise.
It's a real pleasure to meet you.
I love the game of Go. I'm not very good,
but I don't know if you know,
but we actually have something in common.
I actually also trained at a game
at a young age.
I used to be, when I was young,
a professional chess player.
He said that he was once
a professional chess player.
I used to play for the England team,
and then I stopped playing
when I was about 14. I was...
When I was 13, I was the second
highest-rated player in the world.
"After the match,
I would like to propose
that I give you one teaching game of Go,
and you give me
one teaching game of chess."
- Yeah.
- "How does that sound?"
That sounds very good.
I would be very honored to do that.
"I'm not sure if it's okay
to ask you this question,
but I saw the games against Fan Hui,
and I didn't think it was quite
at the level to play with me,
but I heard
that it is getting really stronger.
Can I ask how much
it got stronger since then?"
Yeah. I can't say too much, of course,
but it's definitely got
significantly stronger.
With more time,
I think, a logical approach
would be to follow up
on Aja's local search,
- run that fast in AlphaGo,
- Right.
- generate a whole new data set,
- Exactly.
- and iterate more time.
- Yeah.
But we're just out of time for that,
so we have to be realistic.
We realized that
if we wanted to prepare
for a mammoth task
like taking on Se-dol Lee,
there would be nothing better
than talking to a professional.
And we couldn't have picked
a better person than Fan Hui.
We invited him as an advisor
because during the match,
we realized he is a man of good spirit.
- Welcome back.
- Thank you very much.
It's very precious
that we have Fan Hui with us.
Let's go down this way.
He was crushed by AlphaGo,
but then, he became very positive
and was a big help to us.
When Demis told me,
"Can you come back to help us
make AlphaGo stronger?"
I felt a sense of respect.
Of course, I accepted it.
I played with AlphaGo
to understand
what AlphaGo's strong points are
and what may be its weakness.
I played in the morning,
in the afternoon,
all the time.
And I found something.
I found AlphaGo's big weakness.
It was a big one.
The superheroes always have
a hidden vulnerability, right?
And the same is true for AlphaGo.
It's unbelievably superhuman, in general,
but it has some particular weaknesses
in some situations.
We can think of there
being this space
of all the things it knows about,
and it knows about most of it
extremely well.
Then there will be these tricky lumps
of knowledge,
that it just understands very poorly.
And it's really hard for us
to characterize
when it is going to enter
into one of these lumps.
But if it does,
it can be completely delusional,
thinking that it is alive
on one part of the board
when, in fact, it's dead or vice versa.
So, there is a real risk
that we could lose the match.
Everybody in the team
tried to work more and more
to fix the problem,
but I thought it was difficult
for it to be fixed very quickly.
We've got version 18
in the pipeline, haven't we?
Well, we will only go
for version 18 if it makes
- a significant improvement which...
- If it... Yes.
We're having to rerun all the tests
because they went wrong, unfortunately,
so we're basically a day behind
in our evaluation because of that.
Well, maybe we just need
to be realistic.
That we've tried a bunch of things,
and we've come to the point
where we said we would actually
start freezing the code, and saying,
"Look, it's...
not... actually, not panning out."
These delusions are still
a realistic possibility for the match.
We have some weaknesses
that we... I don't think we are going
to fix fully before the match,
so that's causing us
a little bit of anxiety.
Se-dol Lee is getting ready to rumble.
On Wednesday, live across the Internet,
this professional South Korean Go player
will take on artificial intelligence
program AlphaGo.
I am confident
that I will win five to zero.
Tomorrow, Julian and George,
they pack up version 18,
they stick it on a laptop,
and they fly out to Seoul.
- See you all in Korea.
- Safe trip.
- Have a safe trip.
- See you, too.
Is everything all set
for the victory against Se-dol Lee?
Everything is set, I think.
But when I get to the hotel,
I'm going to catch up with the team.
What we should do
in the time remaining
is list things that could go wrong
in the solidity of the system.
We need somewhere
to put AlphaGo under here.
All right, rehearsal.
Let's go, please, folks!
Have you arranged the place
for Se-dol Lee to smoke?
When he is in a difficult situation,
he likes to smoke.
Yes, there is a terrace,
and we will have security,
so he will be able to go up there
and be by himself.
Which one do you support,
Se-dol Lee or AlphaGo?
I am in an awkward position.
- As you can see, I am a referee.
- I see.
I am a referee.
When I got here,
I didn't expect
the attention on the match.
It was literally front-page news.
About eight million Koreans
play the game of Go,
and, even those who don't,
you know, recognize Se-dol Lee.
He is a national figure.
And so, there's that. Right?
There's some national pride involved.
But, it's more than that.
Just the very thought of a machine
playing a human in something like this,
I think, is inherently intriguing
to people.
We can bring him in?
Let's go in? Yes.
I am confident about the match.
I believe that human intuition
is still too advanced
for AI to have caught up.
I'm going to do my best
to protect human intelligence.
The whole world put pressure
to Se-dol Lee.
Before this, he played the tournament
for his country, for himself.
But this time,
he's playing for the human race.
I just really hope we win
this first game.
If you lose the first game,
you literally have to win
three out of four.
Yeah, that's right.
- Which is hard work.
- Yeah.
Where are you going to be?
- I'll start in the match room.
- Okay.
Then, I'm gonna come in here,
make sure everything's fine.
- Yeah, you should probably be in here.
- Or something like that.
- I know you're nervous.
- Yeah, I'm nervous.
I'm nervous.
How are you? You're not nervous?
- A bit.
- Little bit, yeah?
It's good to have a little bit of nerves.
It should be fine.
Be fine. Be fine.
Hello, and welcome
to the DeepMind Challenge,
Game One, Round One,
live from the Four Seasons
here in Seoul, Korea.
I'm Chris Garlock
of The American Go E-Journal.
I'm here with Michael Redmond,
9-dan professional.
Welcome, Michael.
Want to give a shout-out to all the folks
watching around the world?
Well, the excitement
is pretty palpable here in the hotel.
I've never seen a crush of interest
of reporters from around the world.
- The cameras...
- The media is overwhelming.
I am not sure how much AlphaGo
has changed in the last five months.
All right, folks. You're here,
you're gonna see history made.
Stay with us.
Five minutes, guys. Five minutes.
Please don't push.
I just thought we should
take a moment together and just think
about what is about to happen.
I'm extremely excited to be here.
Extremely proud of every one of you
and what we have done.
And, win or loss, I think
it is just amazing that we're here.
Isn't it strange your dad
is fighting against a machine?
I would like it if the machine
did not beat a human in Go yet.
One minute left.
Five, four, three, two, one.
The match will be
in Chinese rules, 7.5 points komi.
The time limit is two hours,
and 60 seconds times three.
Please choose the stone.
Se-dol Lee, 9 dan, what have you chosen?
Black.
Please start the game.
We had worked so hard
to make sure that this would go
technically smoothly.
We tested it and tested it and tested it.
And still, there comes that moment
when you're live,
all the TV cameras
are broadcasting everything,
and now it has to do
the thing you built it to do.
AlphaGo is even considering
what to do from the first step.
If you look at the time closely,
- it has been over 30 seconds.
- Yes.
AlphaGo is thinking too long
at the second move.
Yes.
I was a bit nervous.
It's the first time that I sat
in front of a world-class Go player.
And I actually felt
the spirit and courtesy
of a great Go player like Se-dol Lee.
Because I think
it was the first time he faced
a strange opponent, I think.
It is non-human,
has no emotion, it's cold.
But he stayed very calm.
And I can feel his mental strength.
Oh.
AlphaGo peeped.
It's unbelievable.
This is the first time
I am seeing AlphaGo
making a move like this.
It feels like AlphaGo is playing
like a human.
I'm Andrew Jackson.
With me here is Myung-wan Kim, a 9-dan pro
from the Korean Baduk Association.
We are here live at the Four Seasons Hotel
on the 21st floor.
What are we looking at?
How's the game going?
Oh, it's fighting from the beginning.
- From the very beginning?
- Yeah, yeah.
- AlphaGo is playing very well.
- Yeah?
- It's just like a top professional.
- Just like a top professional?
Yeah, it's very aggressive.
- It blocked!
- Did it block?
How dare he disconnect it?
This can't be right.
Now, the fight is getting
really complicated.
Um, this is actually the first time
I have seen AlphaGo
playing a game
that has this difficult fight.
This could make
Se-dol Lee nervous.
Because AlphaGo uses one minute
to 1.5 minutes in any situation.
This action is not like a human at all.
No matter how complex you make the game,
AlphaGo plays as if it knows
everything already.
I just saw him looking
at his opponent's face.
- And that's just kind of a habit.
- It's just an instinct,
as a player, to look at the person
across from you.
Yeah, it's sort of something
that Se-dol Lee would do
when he was wondering
how his opponent was feeling.
- Right.
- It's just a habit.
So, it's not as if Aja Huang
is going to do any giveaway
- because Aja Huang isn't AlphaGo, right?
- Right. Right.
He almost made a move.
Lee smiled for a second.
- No. That's...
- I guess Lee is mistaken?
He was about to make a move
that makes no sense.
- Here is the...
- What is he doing?
If he had put his stone there,
he would have lost the game today.
- He was trying to place it here,
- His hand...
but he moved his hand there
without thinking.
- He has so many things to ponder.
- Well, I think...
He is exhausted now.
- I think...
- Well...
He looks a little panicked.
This is the thing
we were most afraid of.
Yes.
- Self-doubt.
- Self-doubt.
With a human, when you play,
you can have an exchange, by feeling.
I look at you. I know that, okay,
maybe you want to talk with me.
Maybe you fear me.
I can feel many, many things.
But with AlphaGo, you can feel nothing.
So, when you feel nothing when you play,
you have more questions to yourself.
In the beginning, you think,
"Okay, my move was good.
Is it really good? Really good?
Oh, maybe it's bad.
Oh, terrible!
Why did I put it here?"
More and more.
It's hard to know where to be.
All the different rooms
are like exciting in different ways.
Mm-hm. Right.
It's quite nice to be here
at the heart of the operation.
Yeah, I feel safe here.
White is thinking of doing
this huge invasion here
from its thick wall.
Yes, go for it.
Stone in. He's going in.
- Look at his face. Look at his face.
- Yeah.
That is not a confident face.
He's pretty horrified by that.
I can't believe
what I'm seeing right now.
Really...
- Se-dol, he's a little bit behind.
- Mm-hm.
- And he made a very aggressive move here.
- Okay.
And made it very complicated.
- Okay.
- I think Se-dol Lee,
- if he doesn't respond correctly,
- Uh-huh.
then he can collapse.
I mean, this is kind of unthinkable
as a human Go player.
What if it knows everything
about what's going to happen next?
Where are we?
Here's our search depth,
we're searching, 50 or 60 moves ahead.
That's the maximum number
of moves ahead
that AlphaGo is looking
from the current game position.
It's typically over 50.
It's often over 60.
In the games we have seen,
often around move 150,
AlphaGo goes for the kill.
We are at move 150 now,
so we are getting to that tactical point.
We are all astonished,
just in the middle of the game
because AlphaGo is, uh...
It seems to be doing a much better job
than we all thought it would.
I thought Se-dol Lee would be leading
the game comfortably,
but it turned out that he's struggling
at the moment.
But, I think, eventually,
he will prevail. I hope.
I was, you know, more five to zero,
but now, I'm not sure about that.
What is this?
AlphaGo made a mistake.
AlphaGo made a mistake here.
What a mistake.
It might be the first mistake,
kind of a clear mistake that white made.
AlphaGo is making mistakes
at the endgame from time to time.
If AlphaGo wins
by a small difference in the end,
it is possible that it played
just enough to win.
The reason I stopped talking a minute ago
was that according to my count,
AlphaGo may have more points.
I didn't tell you
because I couldn't believe it.
It seems like a mistake,
but because it's based on calculation,
it's actually not a mistake.
Yes, yes.
So if it's actually based
on calculation...
it is so scary.
It means it's just playing
with its opponent.
Myung-wan, it looks like
you have got to count.
Yeah, right.
- What do you think? Yeah.
- It's done. Yes.
White won a lot at this time.
It looks like it.
If it's like this, white has won by a lot.
Yeah.
Two, three. I can't believe it.
Thirty-six, thirty-seven...
Seventy-two.
He lost.
He lost.
Whoa, it's so shocking.
I... I expected AlphaGo
to win only one game.
Oh, gosh.
I can feel his pain. Like he was...
He couldn't believe, you know.
He couldn't accept it.
It took him some time
to accept the outcome.
Maybe AlphaGo is very strong now,
but he did not want
to believe he will lose.
And as Go professional players,
we can't believe this,
because, for us,
it's something unimaginable.
It's not something that can happen now.
It's impossible.
But, in reality...
it is happening now.
I think he resigned
in a very polite way.
Se-dol Lee plays black,
but he put a white stone.
I think he resigned.
Oh, my gosh.
- Look at the time. It's slow.
- Yeah.
Wow.
You got this.
- Congratulations.
- Hey.
What's up?
I feel really good.
I feel like I really believe in AlphaGo.
Of course, it's natural that humans
want humans to win.
I mean, I think that's a natural response.
But AlphaGo is human-created,
and I think that's the ultimate sign
of human ingenuity and cleverness.
Everything that AlphaGo does,
it does, because a human has either
created the data that it learns from,
created the learning algorithm
that learns from that data,
created the search algorithm.
All of these things have come from humans.
So, really, this is a human endeavor.
In the battle
between man versus machine,
a computer just came out the victor.
DeepMind put
its computer program to the test
against one of the brightest minds
in the world and won.
AlphaGo beat
a professional player
who has 18 Go World Championships
under his belt.
The victory is considered a breakthrough
in artificial intelligence.
First of all,
I would have to say
that I was very surprised,
because I didn't think
that I would lose the game.
I think the mistakes I made
in the beginning
lasted until the very end.
That's why I lost this game.
I wasn't able to foresee.
I didn't think that AlphaGo would
play the game in such a perfect manner.
I have won world championship titles
and have a lot of experience,
so losing one game won't affect me
in playing games in the future.
I think now it's 50/50.
I would like to express
my respect to the team
for developing such an amazing program
like AlphaGo.
In research,
we normally work
to produce an academic paper.
It gets published,
and maybe we get to talk about it
in a conference if we're lucky.
This is not normal for a research.
In fact, I've never experienced any
media attention remotely close to this.
So, it's a special moment for us all,
and we're just enjoying it while it lasts.
In a matchup between
man and machine, who wins?
So far, it's the machine.
In Seoul, South Korea,
the artificially intelligent computer
defeated the global champion
in the ancient Chinese board game Go.
AlphaGo Shock
AI Shocked the World
Se-dol Lee lost the first matchup,
but he's got four more chances.
This is the team
that played Go with Se-dol Lee.
We got our newspapers.
But we are not responsible.
It's just the man in the beanie.
Sorry.
AI Defeats Human Brain
It is a bit strange
being on the front cover
and everything as a computer scientist.
Normally, you sit there
in your corner and you code.
Nobody really knows about it.
Perhaps you've heard the joke,
"How can you tell
that a computer scientist
is an extrovert and not an introvert?"
If he's an extrovert,
he looks at your shoes
when he's talking to you,
instead of at his own.
If you look at Aja,
he avoids all the cameras like crazy.
He's like, the game is finished,
and he's, like, out the back,
and back in his room.
And I think a lot of computer scientists
would be like that.
We are more about doing our work
than standing in the spotlight.
Hello, and welcome
to Game Two, Round Two,
in the Google DeepMind Challenge
throwdown between man and machine.
Game One, the machine takes down man.
Huge shock. Headlines around the world.
The reactions on the ground here
from the folks in Korea were just stunned.
They estimate 60 million people
watched the game in China alone.
Probably bringing it up
to maybe 80 million people
who watched this game worldwide.
It was just incredible.
And if anything, today is probably
even more of a madhouse.
So, Se-dol Lee knows today.
He knows that this is just
a really important game. Right?
He has got to win.
Go, Se-dol Lee!
Fighting!
Can you tell what Lee
is thinking by looking at his face?
Does he seem nervous?
Yes.
Lee seems to be very tense.
My guess is that
he didn't sleep well last night.
Se-dol Lee on white,
I think probably looking
for a little payback.
My impression is that maybe
he underestimated AlphaGo,
and he is going to change his tactics.
Se-dol Lee is playing
much, much slower today.
Yeah.
- Which is going to affect us.
- Yes.
He's playing at half the speed, actually.
It is hard to say who is ahead
or better now.
The last two moves
make me doubt AlphaGo's ability.
- But we have to stay alert.
- Yes.
- Yes. This is difficult.
- AlphaGo is hard to understand.
If I see anything about AlphaGo
that is not normal,
is maybe the way it handles a game
- when it thinks it is ahead.
- Yes.
We're actually going to have a visit
from one of the team,
and we will talk about exactly
that point from the inside.
Thanks so much for coming by, Thore.
I really appreciate it.
Can you sort of share a bit
of what is going on in AlphaGo?
So, AlphaGo has these three
main components.
There's the policy network,
which was trained on high-level games
to imitate those players.
And then, we have a second component.
We call this the value net.
And it can evaluate the board position
and say what is the probability of winning
in this particular position.
And the third component
is the tree search,
where it would look through
different variations of the game
and try to figure out
what will happen in the future.
So if we now take
a position like this,
first, the policy network
would scan the position
and come up with what would be
the interesting spots to play,
and it builds up a tree of variations.
And then employs this value net
that tells it how promising is the outcome
of this particular variation.
So, AlphaGo tries to maximize
its probability of winning,
but it doesn't care at all about
the margin by which it wins.
Okay, so when you see
a slow-looking move,
that's maybe an indication that AlphaGo
thinks it has a good chance to win.
- Yeah, that is a little giveaway.
- Yeah.
- A little tell. We're looking for a tell.
- Oh, yeah.
Se-dol Lee is playing
in a completely different style
from his usual style.
This is a historic moment
and Se-dol Lee is the center of attention.
He must be feeling immense pressure.
I hope he gets over
this pressure and enjoys the match.
Ooh, it looks like Lee
is taking a little bit of a break.
Se-dol Lee goes to smoke,
and AlphaGo just plays.
It does not think about
whether the opponent will be there or not.
So, Aja sees AlphaGo plays move 37,
and Aja puts the stone on the board.
- Oh.
- Oh, wow.
- Oh, it's totally an unthinkable move.
- Yes.
The value...
That's a very...
That's a very surprising move.
I thought it was a mistake.
When I saw this move,
for me, it was just a big shock.
What?
Normally, humans will never play this one
because it's bad.
It is just bad.
We don't know why, it's bad.
It is a little bit high.
Yeah?
It's the fifth line.
Normally, you don't make
a shoulder hit on the fifth line.
So, coming on top of a fourth line zone
is really unusual.
Yeah, that's an exciting move.
I think we are seeing
an original move here.
That is the kind of move
that you play Go for.
Hey.
Interesting stuff.
- This fifth line shoulder hit was good.
- Yeah.
And I wasn't expecting that.
I don't really know if it's a good
or bad move at this point.
The 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 1-in-10,000 probability
that move 37 would have been played
by a human player.
So it knew that this was
an extremely unlikely move.
It went beyond its human guide,
and it came up with something new
and creative and different.
I am very much watching the game
through these commentators.
That is the way it works.
So when they are confused,
I'm certainly confused.
At the same time,
I'm latching on to the fact
that they are confused, right?
That is an interesting moment.
When everyone else is confused,
who is not confused, right,
besides the machine?
I want to see Se-dol Lee
when he sees this move.
He is back. Lee is back.
I thought AlphaGo
was based on probability calculation
and that it was merely a machine.
But when I saw this move,
I changed my mind.
Surely, AlphaGo is creative.
This move was really creative
and beautiful.
Normally, he thinks
for about one or two minutes
not more than that.
But this time, he thought
for more than 12 minutes.
The more I look at this move,
I feel something changed.
Maybe, for humans, we think it is bad,
but for AlphaGo, why not?
Go is like geopolitics,
like something small that happens here,
it can have a ripple effect,
you know, hours down the road
in a different part of the board.
The game kind of turned on its axis
at that moment.
This move was very special,
because with this move,
all the stones played before
worked together.
It was connected.
It looked like a network,
linked everywhere.
It was very special.
Very special.
This move made me think
about Go in a new light.
What does creativity mean in Go?
It was a really meaningful move.
This is a tough game
for Se-dol Lee.
AlphaGo is just not letting Se-dol Lee
- do what he wants.
- Right.
Black has almost 60 points. That's a lot.
That's not a good sign.
Oh. Oh.
Se-dol Lee just slapped himself
on the side of the head.
Oh, wow.
I think black is ahead at this point.
It's looking good, isn't it?
We're on that steady...
steady path now.
I just saw Se-dol Lee...
lose so much.
Normally, we would have resigned
a long time ago, but he wanted to try.
He continued to play.
He just didn't want to resign.
Because then white...
Oh, he resigned.
It looks like Se-dol Lee
has just resigned.
It doesn't make any sense.
I didn't foresee that one.
- Move 37, very beautiful.
- Yes.
Beautiful, yeah. Beautiful.
Is he okay?
I brought his friend,
because yesterday I noticed
that he really wanted to analyze the game.
I didn't realize how bad it was.
There was this heavy sadness
over that whole floor,
and you could feel it during the game.
I felt it during the game.
And I'm leaving the commentary room
to go to the press conference,
and I was stopped by someone,
another technology reporter.
At first, all he wanted to talk about
was the technology and how great this was,
but then, even he kind of slipped
into this moment of melancholy
where he was upset as well.
Yesterday, I was surprised,
but today, I am quite speechless.
"I am quite speechless.
I admit that it was a very clear loss
on my part.
From the very beginning of the game,
there was not a moment in time
that I felt that I was leading the game."
You feel elated,
and you feel a little bit scared.
There is something, I think,
frightening to people
about a machine that learns on its own.
For us, AlphaGo is obviously
just some computer program.
But looking at the commentary
on the Internet,
I already saw the commentators
call AlphaGo
like "he" and "she" during the games.
Completely unconsciously.
Which... AlphaGo is really
a very, very simple program.
It's not anywhere close to full AI,
and we already see that happening.
So I find that very interesting.
The tendency
to anthropomorphize AI systems
is one of the big obstacles
in the way of actually trying
to understand
how AI might impact the world
in the future.
For example, the conversation is about
what could go wrong, what the risks are.
And invariably, you see
this Terminator picture.
Every single time,
there are these red glowing eyes, right?
We're really closer
to a smart washing machine
than Terminator.
If you look at today's AI,
we are really very nascent.
I'm extremely excited
and passionate about AI's potential.
But AI is still very limited in its power.
I think that people
are right to think
that there is a danger that as we continue
to improve these systems,
that we might miss that threshold
where we do cross over into danger.
But the good news is
that there are already people
thinking about those dangers.
You know, there's a lot of talk now,
and we are leading the discussion on this,
that maybe there should be a kind of,
cross-industry best practices
working group or something.
Right.
Where the leaders of the research teams
in those organizations,
you know, the big ones that are working
on AI, IBM, Microsoft, so on,
come together and make sure that AI
is used ethically and responsibly.
I think what is important
is that there is this community of people
who are leading the cutting edge of AI,
who are interacting with academics,
and already are thinking
about the long term,
and how we can ensure
that innovation is responsible
as the power of these machines
gets even greater.
Se-dol Lee Lost by Resignation
This is it, folks.
Day Three, Game Three.
Se-dol Lee, Go Master.
Back to the wall. He's down two-zero.
He's got to win today to keep hope alive.
Se-dol Lee had a day off
after losing two games,
and he gathered with Go professionals
and analyzed the game all night.
I heard that four pros, yeah,
went to visit him.
- Hm.
- To console him and to review AlphaGo's...
Console him?
Do you think he's upset, or...
I mean, Se-dol Lee
is upset. Yeah.
In the beginning,
there was a fierce fight,
and AlphaGo played very well.
So he secured a very early lead.
From move 50, the win rate
was very high already.
It was climbing toward 100%.
Oh, that's a pretty move.
What's your probability rating?
- Like, 91?
- No.
- It seems 70, you think?
- No.
He must want to win desperately right now.
THE CHALLENGE OF AI
0 VS. 2
But I think the pressure
and the psychological burden is adding up.
Today must be the worst.
I personally believe,
when you try too hard to win,
you will lose.
He tried to fight
directly in the game
but it's not his style.
When we change our style
to play against an opponent,
normally, it's very, very bad.
So it's an easier game for AlphaGo.
- It's looking good for us.
- Hopefully, I think...
You know, that black group is huge,
and it has got nowhere to go,
and it is going to be running around.
I don't know how
to describe the situation.
If I were black, I will resign.
We should admit that we are facing
the strongest existence ever,
ever in the Go history.
There's no point
in playing out the endgame,
and you're going to lose, right? So...
Even if black can live there...
- It's done. It's done.
- Oh, he resigned. Okay.
Wow. Wow.
You saw history made here tonight.
AlphaGo has won again.
Three straight wins.
Three straight wins.
Has won the match.
When Se-dol Lee resigned
from the game, he looked unhappy.
It was not just he lost the tournament,
it was especially about this game,
because he did not play his game.
I am very hurt about this, very hurt.
But, I can't do anything.
You know, suddenly, I feel
a bit ambivalent about it,
given I'm a games player,
and, you know, Go is the pinnacle
of board games.
But I really liked the statement
of one of the top Chinese professionals.
He said, you know,
"If AlphaGo wins, maybe we'll really start
to get to see what this game is about."
I couldn't celebrate.
It was fantastic that we had won,
but there was such a big part of me
that saw this man trying so hard
and being so disappointed.
I see on the Internet, many people
are talking about Se-dol Lee.
Maybe he didn't play his best.
You know, we are Go players.
Okay, sometimes in China,
in Korea, in Japan
we see Go like art.
We are artists, you know.
- We play our best for Go.
- Right. Right.
- So, please be gentle with Se-dol Lee.
- Right.
He is a very, very good player.
He is a great player.
I was in the room. I saw Se-dol Lee.
He wanted to win.
He tried everything.
It's just we can't. It's just... So...
I think I have to express
my apologies first.
If I had been able to play better
or smarter,
the results might have been different.
I think I disappointed too many of you
this time.
I want to apologize
for being so powerless.
I've never felt...
this much pressure,
this much weight.
I think I was too weak to overcome it.
I can't believe
this is happening.
Regardless of your opponent's level,
to be defeated not by three-zero,
but five-zero?
Losing to AlphaGo by five-zero
would really hurt my pride.
I also feel so bad for the people
who have supported me.
Se-dol Lee, 9 dan,
has the strongest heart of anyone I know.
He is fighting a lonely fight.
His opponent does not exist
in physical form.
I really feel for him.
We still have Games Four and Five,
and if Se-dol Lee, 9 dan,
plays like himself,
I believe we can beat the machine.
You could see he was more relaxed
after he had lost three games in a row.
That said, the stakes were still high.
You know, in the end,
it isn't about pride.
He did not feel confident,
but he felt light.
Can Se-dol Lee
find AlphaGo's weakness?
Is there, in fact, a weakness?
I was thinking I would just pull the plug.
- You would just pull the plug?
- Yeah.
Everyone is still
cheering for Se-dol Lee.
And...
Yeah, it's not going very well.
Se-dol Lee has chosen
to play conservatively.
But I am not sure
it will benefit him.
I agree.
Yes, it doesn't look like
a very good strategy for him.
It feels pretty good
for black at this point.
- It feels pretty good for black, yes.
- It's pretty good for black. Yeah.
I want Se-dol Lee to play his game.
Because, at that moment,
he tried various things
to play with AlphaGo,
to understand AlphaGo,
but he never tried to play himself.
I've said this many, many times,
AlphaGo looks like a real mirror.
When you play with AlphaGo,
you feel very strange.
It feels like you are naked all the time.
The first time you see this,
you wouldn't want to see because...
"Oh, is this me? The real me?"
And the more you see,
you learn to accept it.
"Oh, this is the real me.
So, how... Now, what shall I do?"
It's developing into a very,
very dangerous fight.
This is really Se-dol Lee's type of game.
He likes this kind of fight.
White has to find something
inside black's territory.
I think that he
is already planning on trying something.
So you believe in Se-dol Lee's ability
to live in that very small area, right?
Yes. Yes.
There is a little potential there.
And I think, maybe he is going to try
to do something.
- Se-dol Lee magic!
- Nice.
Oh, what would be
the magic move?
Se-dol Lee is running short on time,
but he's going to have to use up
all his time.
Yeah, he has just burned, like,
seven or eight minutes,
just on this move already.
I don't see anything.
We can't find anything.
Show us something.
I don't see any possible move.
What is Se-dol Lee up to here?
Yeah, he is really concentrating.
He really is. Look at that.
Se-dol Lee is very patient.
He waits. He waits for his moment.
I feel, sometimes, he is like a wolf,
waiting in the forest, in the winter.
He is cold. He's very, very cold.
But he needs patience.
But, when the moment comes,
he goes out to attack.
This is the hinge of the game.
Oh.
Look at that move.
That's an exciting move.
- Basically, I believe that this time...
- He found a wedge.
Whoa.
It's going to change
the equation
because now black cannot escape.
That would be so cool if that works.
Oh!
AlphaGo has just played
something maybe unusual.
You know, I'm not actually sure
what AlphaGo is trying to do here.
What's that about?
I don't really understand it.
Well, well.
That was a sharp drop in win rate.
That is the sharpest drop
in win rate we have seen.
She dropped at eight percent.
- Wow.
- Wow.
This could be that it actually
can't find a way through.
- I think this is...
- Uh-huh.
It has looked far enough ahead
to see that it doesn't work,
and now maybe it's on tilt. I don't know.
There it comes.
- That was one time.
- There it comes, though.
It looks like it has fallen off the cliff.
- Yeah, it has made a mistake.
- Yeah.
Did anything strange happen in the...
- No, it all looked normal.
- Yeah.
Well, we can definitely say
there is a weakness.
Well, we definitely say there's a mistake.
I felt this mixture
of this sinking feeling in my stomach,
where I was wondering if AlphaGo
was becoming delusional in this situation.
Where I could see that it was starting
to play strangely,
and at the same time,
relief, that Se-dol Lee,
that he was actually in with a chance now.
Aja is like trying not to look horrified.
I knew after move 78,
after, like, 10 or 20 moves,
I saw AlphaGo's strange moves.
And, I knew AlphaGo somehow
became crazy, but I didn't realize why.
What?
We searched to 95 ahead at that point?
At the part where it made the mistake?
- Yes.
- I think that something went wrong.
That's the longest it has searched
the entire game, isn't it?
Yeah.
I think it's like it searched so deeply,
- it has lost itself.
- It's tired.
I do get the impression
that AlphaGo has sort of,
gone off on a tangent.
- What is it doing now?
- Well, maybe it has a master plan.
No, it doesn't even think it has, does it?
So it knows it has made a mistake,
and it starts evaluating it the other way.
Look, look, Lee is confused.
He's like, "What is he doing?"
That's not a "I'm scared" confused,
that's a "What is it doing?"
What's going on?
You know, you asked if it was a bug.
I've said before that if...
- Yeah, right. That's what I thought.
- If DeepMind has figured out
how to write code that doesn't have bugs,
that is a bigger news story than AlphaGo.
- Are you kidding me?
- Hm?
This... Literally, this next move
we are going to play,
I think they are going to laugh.
I think Lee is going to laugh.
Oh! Oh!
What is this?
- Oh, that's ridiculous!
- What's going on?
What is it thinking?
I don't really know what AlphaGo
is trying to do here.
That's the understatement of the year.
Is it a mouse misclick
- from Aja Huang?
- Nope, that's the move.
Aja Huang makes no misclicks.
Se-dol Lee is very confused.
These are not human moves.
This move is also sort of inexplicable.
I mean, those, you can clearly
call those mistakes.
Yes, of course.
But it's the first time
in the four matches
- that we have seen moves like that.
- Right.
Oh, the value dropped even more.
That's weird.
They're like 45%.
White.
- White.
- I think white is winning.
Come on, Se-dol Lee.
- It's unbelievable.
- Yeah, all right!
I was so confident that this black area
would just be consolidated by black,
and there was nothing there.
And somehow, he has just erased it all.
Like, it's gone.
So, he has found his weakness.
That wedge move probably surprised it.
- The wedge, yeah.
- Yeah.
AlphaGo seemed to have
such a good game, and then...
this whole sequence in the center
sort of changed that.
He pulls off a miracle.
He managed to make it so complicated
that the artificial intelligence
doesn't evaluate it correctly.
I think it looks like Aja Huang
maybe, sitting there,
also knows that the game is over.
I can't... I can't stop smiling.
Did he smile?
No, I haven't seen him smile yet.
- I think he is still checking.
- He's so serious.
He's so serious.
- He wants to be careful.
- Yeah.
He's concerned a lot.
He really does his best.
I think he feels like he has a sense of...
uh, responsibility, or burden.
I think it's almost over.
So, what's the percentage?
It shows as 18.2%.
The next move should be resign.
There you go. AlphaGo resigned.
- It looks like AlphaGo has resigned.
- Wow.
The most amazing game.
I'm almost going to tear up,
was, uh, Game Four...
um, where he comes back and wins, right?
It resigned!
I'm so happy!
I heard people shouting in joy
when it was clear that AlphaGo
had lost the game.
I think it is clear why.
People felt helplessness and fear.
It seemed like we humans
are so weak and fragile.
And this victory meant...
we could still hold our own.
As time goes on, it will probably be
very difficult to beat AI.
But winning this one time,
it felt like it was enough.
One time was enough.
I heard from so many people saying
they were running out in the street.
They were so happy.
They were chanting, they were celebrating.
Especially after zero-three down,
at the time, it seemed to be hopeless,
that the end of the world was coming,
but we saw the light.
Usually...
I'm happy when I win,
not when my colleague wins.
But, this time, it felt like my win.
I didn't want to make a move
which AlphaGo could predict.
This part...
It didn't anticipate this wedge move.
I won one game at least.
Good job today.
But I didn't expect it
to be like this.
I couldn't believe I won one game.
It was unbelievable.
Thank you very much.
I have never been congratulated so much
for winning one game.
After losing three games in a row,
I couldn't be happier.
This victory is so valuable
that I wouldn't exchange it
for anything in the world.
My question is about move number 78.
Chinese top Go player
Gu Li said that it was a god's play.
What were you thinking
when you made that play? Thank you.
At that point in the game,
move 78 was the only move I could see.
There was no other placement.
It was the only option for me,
so I put it there.
I am quite humbled by all the praise
I am getting for it.
It does leave me a little bit
in awe of the human brain's power,
in particular, Lee's amazing ability
to cause AlphaGo problems
and find something
seemingly out of nothing.
And so, we really want
to understand what had happened.
So, it really was Lee's move?
The key is the center.
After the center got out,
there was no chance.
Come, come.
We were winning before this.
But if AlphaGo thought it was winning
and it couldn't convert that win,
that means it wasn't actually winning.
Otherwise, it would have had
the right responses.
Would we have played it?
Nerves of steel in there.
What probability does it give it as white?
It's 0.007% of...
Is that in the position that he played?
- That's when he played it.
- I see.
- We thought this was a 1-in-10,000 move.
- I know it was like...
- Yeah.
- The values we didn't know about.
- So the god move was literally a god move,
- Yeah.
because we believe
that only one in 10,000 humans
- would have found that move.
- That's right.
It doesn't come to mind as the top five...
in the top five moves
that you would consider.
Unless you're Se-dol Lee,
he said it was the only move.
He thought it was the only move.
Yeah, remarkable, isn't it?
Welcome back to the Four Seasons.
World Champion Se-dol Lee
went looking for AlphaGo's weakness
in Game Four, and he found it.
Today, last round, he is looking to see
if he can repeat that.
We will see what happens.
The place
is a madhouse downstairs.
Maybe as many, if not more,
journalists here today
than there were for Game One.
We were really excited
for the fifth match,
to see what would happen.
Was this going to be a three-two thing
or was it going to be a four-one thing?
And, you know, there's a different message
with both of those.
For me, I don't think
AlphaGo will make the mistake again.
But who knows.
Maybe it will happen again.
And Se-dol Lee now have more confidence
about this.
It looks like, maybe, Se-dol Lee has
found the key in solving AlphaGo.
Oh, this works!
Is it fair to say
that AlphaGo made a mistake?
We might have another victory today.
You weren't crazy
about the timing on this move.
- Yeah.
- And now, you don't approve of this.
Yeah, I'm sort of thinking
that maybe AlphaGo
hasn't recovered from Game Four yet. Yeah.
Still deluded?
We don't know.
Oh, no, I hope he doesn't...
No, no, he needs to play the...
He played some moves.
It was a bad move.
I felt something.
Oh, maybe his weakness
has come back again.
Are we seeing
another short circuit?
Or is there something, what's...
I think it could be
a kind of a misreading.
And we are...
You're pretty comfortable
with saying that...
That it is looking good
for Se-dol Lee?
It's looking good for Se-dol Lee.
What is he playing in there for?
There's no reason for white
to be playing that move.
It's a bad move.
And in some cases,
it's going to lose a point, too.
It is exactly the same thing.
It's 91% certain now.
Someone was telling me
that maybe it had...
Yeah, because it's incorrect again.
The whole game,
we thought that AlphaGo
was wrong about the board position.
We were super worried that,
"Oh, it's going to play garbage.
It's going to be like lose
in a very embarrassing way."
And this continued for the whole game.
This is a bit weird.
- When did it play that?
- Just now.
- So I don't know. I mean, it doesn't...
- Well, that we don't know.
But we're not good enough
Go players to know that, right?
Yeah.
As it turns out,
none of us know Go well enough
to accurately judge what AlphaGo is doing.
The white is winning now.
This one looks like white is winning.
Ah!
We all say some of AlphaGo's moves
are so weird and strange,
and may be mistakes.
But after a game is finished,
we have to doubt ourselves, our judgment.
AlphaGo making another
kind of nonsensical throw-in.
I'm not really sure what that's about.
This is what 10
or maybe 11-dan play looks like.
It looks weird,
and we don't quite understand it.
I think it's important to study more
about AlphaGo's mistake-like moves.
Then maybe we can adjust
our knowledge about Go.
To me, the most amazing thing
to come out of my understanding of Go
as a result of watching AlphaGo play
are the infamous "slack moves."
Well, there's something strange
- about the way it is playing,
- Yeah.
because it's playing some moves
that are not really necessary.
Right.
A "slack move" is a move
that looks lazy.
You can see these other better moves,
and AlphaGo is rejecting them.
But, what I think AlphaGo is teaching us
is that we have been using score
as a proxy for chance of winning.
So the bigger my margin of territory,
the more confident I am
that I'm going to win.
And AlphaGo is saying, "No, no, no.
It shouldn't matter how much you win by,
you only need to win by a single point.
Why should I be seizing all this
extra territory when I don't need it?"
The lessons that AlphaGo is teaching us
are going to influence how Go is played
for the next thousand years.
By my counting, white won this game.
How... By how much?
Are we talking two points?
- One-and-a-half.
- One-and-a-half points.
One-and-a-half. Maybe half,
one-and-a-half.
Of course, Go is just a game,
but we can learn important lessons
from a computer being so successful at Go.
Machines will have the capability
not only to crunch
through a huge amount of data,
but also to analyze it intelligently.
Just as in the case of the Go games,
the machine made moves
that surprised even the experts.
And, eventually, the machines
will gain our confidence
because we will see that very, very often
they make a better guess
than we could have made as humans.
AlphaGo has strong foundation.
What surprised me the most was...
that AlphaGo showed us
that moves humans may have thought
are creative, were actually conventional.
I think this will bring
a new paradigm to Go.
AlphaGo won this area as well? Oh, god.
- That makes me very uncomfortable.
- Oh!
Unfortunately, for Se-dol Lee,
I think white might have
a slight advantage here.
White lights up in places,
but if we count it like that, it's 1.5.
They keep saying
one-and-a-half points.
How do you know the Korean room
has called it?
Someone, um...
- Who told us the Koreans called it?
- I don't know.
The Koreans told you?
Koreans called it.
The Korean room
has called it for AlphaGo.
We would just like to know
if someone Korean in there can find out.
The Korean room
think AlphaGo is winning.
"Winning," okay, not... then not won.
- That's what she said.
- "Winning."
She said they called it right there.
AlphaGo is saying it's going to resign.
- What...
- No. I'm only joking.
Jesus, I almost had a heart attack.
That is ridiculous.
That was very good.
- Well played.
- Yeah.
- Well played.
- No, not really.
I couldn't do it for too long.
I just couldn't think like that.
David, you've been waiting
two years to do that.
He needs to do that.
We don't know the score.
We have no idea what's going on, right?
Two-and-a-half points.
Lee is two-and-a-half points behind,
at least.
The parting in the center,
it's very hard to... It's very hard to...
Lee is touching the bowl.
It looks like
he's considering resigning.
Yes.
I think he is thinking of resigning.
- Did he resign?
- He resigned.
Hm? He has what?
I think he resigned.
- He resigned?
- I think he resigned. Look.
- He's moved everything in there.
- What?
He played the white.
Aja is looking relaxed.
- Yeah, he has resigned.
- He has resigned.
Yes, he is giving up.
- Okay. That's it. He has given up.
- Yeah.
- Yeah!
- Yes!
I mean, I think it just really
is a once-in-a-lifetime thing.
I would say it is the most amazing thing,
you know, I have experienced.
You know, for us, it is the culmination
of a 20-year dream.
It started as a pure research endeavor.
We just wanted to understand.
Can neural networks play the game of Go?
And from there, it went on
to a level that I never expected.
And, I'm unbelievably proud
of the team.
When I started doing
artificial intelligence,
it was four or five years ago,
and I was really interested in that,
but, like, many people
would discourage me,
and they would say,
okay, there is no future.
So, actually seeing that within five years
we are at this stage right now,
it's amazing by itself.
Everybody say, "Nine dan."
- Nine dan!
- Nine dan!
Kimchi!
There are so many
possible application domains
where creativity, in a different dimension
to what humans could do,
could be immensely valuable to us.
And I just love to have more
of those moments
where we look back and say,
"Yeah, that was just like move 37.
Something beautiful occurred there."
At least in a broad sense,
move 37 begat move 78,
begat a new attitude in Se-dol Lee,
a new way of seeing the game.
He improved through this machine.
His humanness was expanded
after playing this inanimate creation.
And the hope is that, that machine,
and in particular,
the technology behind it,
can have the same effect with all of us.
I remember hearing a talk
by Kasparov, who says that, uh,
a good human plus a machine
is the best combination.
This is a unique experience.
Nobody can have this experience
playing with AlphaGo,
five games like this.
So, I hope Se-dol Lee can find
something in these five games.
Maybe some change in his game.
I saw today how he continuously fought.
SE-DOL LEE
2016-03-15
He's very good,
a master, a really great master.
I have grown
through this experience.
I will make something out of it
with the lessons I have learned.
I feel thankful and feel like
I have found the reason I play Go.
I realize it was really a good choice,
learning to play Go.
It has been an unforgettable experience.
This event is done,
but for the story,
maybe, it's just beginning.
We don't know.
It's just when I played with AlphaGo,
he showed me something.
I felt something beautiful. That was it.
I saw the world differently,
before everything began.
What is really behind the game of Go?
What that is, it can change my game.
Maybe it can just show humans
something we have never discovered.
Maybe, there is beauty.