Eng AlphaGo (2017)

1
00:00:18,601 --> 00:00:22,230
It's intensely contemplative.

2
00:00:26,025 --> 00:00:28,778
It is almost hypnotic.

3
00:00:32,657 --> 00:00:38,163
It's like putting your hand
on the third rail of the universe.

4
00:00:39,622 --> 00:00:42,459
If you play Go seriously,

5
00:00:42,542 --> 00:00:45,962
there is a chance
that you will get exposed

6
00:00:46,045 --> 00:00:50,008
to this experience that...

7
00:00:50,925 --> 00:00:53,762
is kind of like nothing else
on the planet.

8
00:00:54,637 --> 00:00:56,598
Go is putting you in a place

9
00:00:56,681 --> 00:01:01,061
where you're always
at the very farthest reaches

10
00:01:01,144 --> 00:01:02,437
of your capacity.

11
00:01:04,189 --> 00:01:06,274
There's a reason
that people have been playing Go

12
00:01:06,357 --> 00:01:08,077
for thousands and thousands
of years, right?

13
00:01:09,402 --> 00:01:11,529
It's not just that they want
to understand Go.

14
00:01:12,781 --> 00:01:16,242
They want to understand
what understanding is.

15
00:01:17,911 --> 00:01:21,498
And maybe that is truly
what it means to be human.

16
00:01:26,544 --> 00:01:29,130
When I was a kid,
I loved playing games.

17
00:01:30,215 --> 00:01:32,884
I started off with board games like chess.

18
00:01:33,802 --> 00:01:36,471
And then I bought my first computer
when I was eight

19
00:01:36,554 --> 00:01:38,264
with winnings from a chess tournament.

20
00:01:39,557 --> 00:01:42,185
Ever since then, I felt that computers

21
00:01:42,268 --> 00:01:46,481
were this sort of magical device
that could extend the power of your mind.

22
00:01:48,900 --> 00:01:50,735
Virtual environments and games,

23
00:01:51,361 --> 00:01:55,156
we think they're the perfect platform
for developing and testing AI algorithms.

24
00:01:55,949 --> 00:01:59,327
Games are very convenient,
in that a lot of them have scores,

25
00:01:59,410 --> 00:02:02,497
so it's very easy
to measure incremental progress.

26
00:02:02,872 --> 00:02:06,876
So I'm going to show you a few videos
of the agent system, the AI.

27
00:02:06,960 --> 00:02:08,837
So let's start off with Breakout.

28
00:02:08,920 --> 00:02:10,880
So here you control the bat and ball,

29
00:02:10,964 --> 00:02:13,364
and you're trying to break
through this rainbow-colored wall.

30
00:02:13,758 --> 00:02:16,511
The agent system has to learn
everything for itself,

31
00:02:16,594 --> 00:02:18,054
just from the raw pixels.

32
00:02:18,138 --> 00:02:19,806
It doesn't know what it's controlling.

33
00:02:19,889 --> 00:02:22,517
It doesn't even know
what the object of the game is.

34
00:02:22,600 --> 00:02:26,771
Now, at the beginning, after 100 games,
you can see the agent is not very good.

35
00:02:26,855 --> 00:02:28,481
It's missing the ball most of the time.

36
00:02:28,565 --> 00:02:30,441
But it's starting to get
the hang of the idea

37
00:02:30,525 --> 00:02:32,125
that the bat should go towards the ball.

38
00:02:33,111 --> 00:02:34,612
Now, after 300 games,

39
00:02:34,696 --> 00:02:36,990
it's about as good
as any human can play this,

40
00:02:37,073 --> 00:02:39,325
and pretty much gets
the ball back every time.

41
00:02:39,409 --> 00:02:40,845
We thought, "Well, that's pretty cool."

42
00:02:40,869 --> 00:02:43,496
But we left the system playing
for another 200 games,

43
00:02:43,580 --> 00:02:44,831
and it did this amazing thing.

44
00:02:44,914 --> 00:02:46,207
It found the optimal strategy

45
00:02:46,291 --> 00:02:48,251
was to dig a tunnel around the side.

46
00:02:48,334 --> 00:02:50,094
And put the ball
around the back of the wall.

47
00:02:50,879 --> 00:02:53,548
The researchers working on this,
the amazing AI developers,

48
00:02:53,631 --> 00:02:56,843
well, they're not so good at Breakout,
and they didn't know about that strategy.

49
00:02:56,926 --> 00:02:58,779
So, they learned something
from their own system,

50
00:02:58,803 --> 00:03:01,347
which is, uh, you know,
pretty funny and quite instructive,

51
00:03:01,431 --> 00:03:04,183
I think, about the potential
for general AI.

52
00:03:04,642 --> 00:03:06,895
So, for us, what's the next step now?

53
00:03:06,978 --> 00:03:10,481
Go is the most complex game,
pretty much, ever devised by man.

54
00:03:10,565 --> 00:03:12,400
Beating a professional player at Go

55
00:03:12,483 --> 00:03:15,904
is a long-standing, grand challenge
of AI research.

56
00:03:36,966 --> 00:03:38,009
Wow.

57
00:03:39,594 --> 00:03:41,387
You really want to kill me, yeah?

58
00:03:41,930 --> 00:03:44,599
No, I don't want to,
but I'm trying to survive.

59
00:03:44,682 --> 00:03:45,934
Yes, that's it. That's it.

60
00:03:46,017 --> 00:03:47,977
I am a Go professional player.

61
00:03:48,061 --> 00:03:49,646
I am also a European Champion.

62
00:03:50,230 --> 00:03:51,689
With that move,

63
00:03:51,773 --> 00:03:53,316
you're going to lose for sure.

64
00:03:53,399 --> 00:03:56,110
You were happy to defeat me
but then you panicked.

65
00:03:56,611 --> 00:03:59,822
You're afraid that victory
will escape you.

66
00:03:59,906 --> 00:04:03,534
In general, when you're afraid,
victory always escapes you.

67
00:04:05,703 --> 00:04:07,205
I was born in China.

68
00:04:07,664 --> 00:04:10,875
When I was 18, I wanted to change my life.

69
00:04:11,417 --> 00:04:13,169
That was why I went to France.

70
00:04:14,128 --> 00:04:16,506
I wanted to try to forget Go.

71
00:04:17,882 --> 00:04:19,300
But it was impossible.

72
00:04:19,884 --> 00:04:23,972
Because all the things I learned
in my life was through Go.

73
00:04:25,056 --> 00:04:26,432
It looks like a mirror.

74
00:04:27,225 --> 00:04:30,561
When I see Go, I also see myself.

75
00:04:32,272 --> 00:04:34,399
For me, Go is real life.

76
00:04:40,238 --> 00:04:41,447
Dear Mr. Fan.

77
00:04:41,823 --> 00:04:43,449
My name is Demis Hassabis.

78
00:04:44,659 --> 00:04:48,413
I run an artificial intelligence company
based in London called DeepMind.

79
00:04:49,289 --> 00:04:51,291
As the strongest Go player in Europe,

80
00:04:51,374 --> 00:04:53,918
we would like to invite you
to our offices in London,

81
00:04:54,002 --> 00:04:56,421
both to meet you in person
and to share with you

82
00:04:56,504 --> 00:04:58,840
an exciting Go project
that we are working on.

83
00:04:59,882 --> 00:05:01,802
If you would be interested
in coming to visit us,

84
00:05:01,884 --> 00:05:03,052
please let us know.

85
00:05:03,803 --> 00:05:06,806
Many thanks. Kind regards, Demis.

86
00:05:08,266 --> 00:05:10,184
When I saw this e-mail,

87
00:05:10,268 --> 00:05:12,562
I didn't know if it was true or not.

88
00:05:13,938 --> 00:05:16,065
I will accept this. Why not?

89
00:05:18,026 --> 00:05:19,944
For me, everything is an adventure.

90
00:05:21,612 --> 00:05:23,239
I want to go visit DeepMind

91
00:05:23,323 --> 00:05:27,702
to know what this Go project is.

92
00:05:31,581 --> 00:05:34,542
On my first visit,
I thought maybe they wanted me

93
00:05:34,625 --> 00:05:36,461
to sit in a special room,

94
00:05:36,544 --> 00:05:39,172
attach numerous wires to my head,

95
00:05:39,589 --> 00:05:41,049
and also my body,

96
00:05:41,132 --> 00:05:44,385
let me play to scan my brain.

97
00:05:45,011 --> 00:05:46,054
I don't know,

98
00:05:46,804 --> 00:05:48,222
to make some research.

99
00:05:48,890 --> 00:05:51,100
We needed to get Fan Hui
to DeepMind

100
00:05:51,184 --> 00:05:54,729
to see we were a serious operation,
and we were serious people...

101
00:05:55,563 --> 00:05:56,647
doing proper research.

102
00:05:58,107 --> 00:06:00,068
And the search time
is getting better and better

103
00:06:00,151 --> 00:06:02,987
as we go from, for example,
one to eight, in any one of these,

104
00:06:03,071 --> 00:06:04,781
it always goes up.

105
00:06:05,281 --> 00:06:07,158
We think of DeepMind as...

106
00:06:07,241 --> 00:06:10,036
kind of like an Apollo program effort
for AI.

107
00:06:10,995 --> 00:06:14,248
Our mission is to fundamentally
understand intelligence,

108
00:06:14,332 --> 00:06:16,000
and recreate it artificially.

109
00:06:16,334 --> 00:06:17,574
And then, once we've done that,

110
00:06:17,627 --> 00:06:21,089
we feel that we can use
that technology to help society

111
00:06:21,172 --> 00:06:23,049
solve all sorts of other problems.

112
00:06:23,132 --> 00:06:24,675
If you step through
the actual game,

113
00:06:24,759 --> 00:06:27,303
we can see, kind of, how AlphaGo thinks,

114
00:06:27,386 --> 00:06:30,473
what's the most likely variation
that it thinks will happen.

115
00:06:30,556 --> 00:06:32,391
We've been working
on AlphaGo,

116
00:06:32,475 --> 00:06:35,770
our program to play Go,
for just under two years now.

117
00:06:36,354 --> 00:06:38,397
All the little patterns cascade together,

118
00:06:38,481 --> 00:06:40,483
layer after layer, after layer,
after layer.

119
00:06:41,484 --> 00:06:45,822
I started talking about the game of Go
with Demis more than 20 years ago.

120
00:06:45,905 --> 00:06:48,324
And so, this has been
a really long journey.

121
00:06:48,407 --> 00:06:51,828
The game of Go is the holy grail
of artificial intelligence.

122
00:06:51,911 --> 00:06:54,163
For many years,
people have looked at this game

123
00:06:54,247 --> 00:06:56,874
and they have thought,
"Wow, this is just too hard."

124
00:06:56,958 --> 00:06:58,960
Everything we have ever tried in AI,

125
00:06:59,043 --> 00:07:01,212
it just falls over
when you try the game of Go.

126
00:07:01,295 --> 00:07:04,674
And so that's why it feels like
a real litmus test of progress.

127
00:07:04,757 --> 00:07:07,468
If we can crack Go,
we know we've done something special.

128
00:07:09,679 --> 00:07:11,639
So with Fan Hui,
we started talking around

129
00:07:11,722 --> 00:07:13,641
what the real purpose of the visit was.

130
00:07:13,724 --> 00:07:16,394
It wasn't just a Go project
we wanted him to help with,

131
00:07:16,477 --> 00:07:19,981
that actually we wanted to play him,
and we had a very strong program.

132
00:07:20,064 --> 00:07:21,691
It's okay, it's just a program.

133
00:07:21,774 --> 00:07:23,401
It's so easy.

134
00:07:23,484 --> 00:07:24,861
So, it will be easy to play.

135
00:07:25,278 --> 00:07:27,864
A lot of people have thought
that it was decades away.

136
00:07:27,947 --> 00:07:29,383
Some people thought it would be never.

137
00:07:29,407 --> 00:07:34,704
Because they felt that to succeed at Go,
you needed human intuition.

138
00:07:35,913 --> 00:07:40,168
Go is the world's oldest
continuously played board game.

139
00:07:40,251 --> 00:07:41,711
And, in some sense,

140
00:07:41,794 --> 00:07:45,882
it is one of the simplest
and also most abstract.

141
00:07:46,382 --> 00:07:47,758
There's only one type of piece.

142
00:07:47,842 --> 00:07:48,986
There's only one type of move.

143
00:07:49,010 --> 00:07:50,570
You just place that piece on the board.

144
00:07:51,053 --> 00:07:56,309
And then, your goal is to create
a linked group of your stones

145
00:07:56,392 --> 00:07:58,144
that surrounds some empty territory.

146
00:07:58,811 --> 00:08:01,522
And when you surround enemy stones,

147
00:08:01,606 --> 00:08:04,066
you capture them,
and remove them from the board.

148
00:08:05,067 --> 00:08:07,361
You earn points by surrounding territory.

149
00:08:07,445 --> 00:08:11,782
And, at the end of the game,
the person with the most territory wins.

150
00:08:12,825 --> 00:08:14,619
It seems really simple.

151
00:08:14,702 --> 00:08:17,580
But, then you sit down to play,
and you realize right away,

152
00:08:17,663 --> 00:08:20,917
it's like, "Well, I technically know
what I'm allowed to do,

153
00:08:21,000 --> 00:08:22,793
but I have no clue what I should do."

154
00:08:22,877 --> 00:08:25,505
Go is incredibly challenging
for computers to tackle

155
00:08:25,588 --> 00:08:27,632
because compared to, say, chess,

156
00:08:27,715 --> 00:08:30,801
the number of possible moves
in a position is much larger.

157
00:08:30,885 --> 00:08:32,303
In chess, it's about 20.

158
00:08:32,386 --> 00:08:34,222
In Go, it's about 200.

159
00:08:34,305 --> 00:08:37,183
And the number of possible
configurations of the board

160
00:08:37,266 --> 00:08:39,727
is more than the number
of atoms in the universe.

161
00:08:39,810 --> 00:08:42,313
So even if you took all
the computers in the world,

162
00:08:42,396 --> 00:08:43,814
and ran them for a million years,

163
00:08:43,898 --> 00:08:48,236
that wouldn't be enough compute power
to calculate all the possible variations.

164
00:08:48,569 --> 00:08:52,114
If you ask a great Go player
why they played a particular move,

165
00:08:52,198 --> 00:08:54,617
sometimes, they will just tell you
it felt right.

166
00:08:54,700 --> 00:08:57,495
So, we have to come up
with some kind of clever algorithm

167
00:08:57,578 --> 00:09:00,373
to mimic what people do
with their intuition.

168
00:09:01,415 --> 00:09:03,751
- Nice to meet you.
- Oh, hello. Nice to meet you, too.

169
00:09:03,834 --> 00:09:06,837
So with Fan Hui, we agreed
a best-of-five match.

170
00:09:06,921 --> 00:09:09,006
And we agreed it would be filmed.

171
00:09:09,090 --> 00:09:10,930
You know, we would treat it
as a serious match.

172
00:09:13,678 --> 00:09:15,346
I played with AlphaGo.

173
00:09:15,429 --> 00:09:18,391
Aja Huang pushed the stones for AlphaGo.

174
00:09:18,891 --> 00:09:21,978
And, of course, I thought I will win
the game with AlphaGo

175
00:09:22,061 --> 00:09:23,604
because it's just a program.

176
00:09:24,689 --> 00:09:25,689
Good game.

177
00:09:26,357 --> 00:09:27,357
Have a nice game.

178
00:09:30,903 --> 00:09:33,698
The first game,
I made some mistakes,

179
00:09:33,781 --> 00:09:35,616
and I lost the game.

180
00:09:36,158 --> 00:09:39,412
Oh. Strange. Very strange game.

181
00:09:39,829 --> 00:09:42,748
In that first match,
I think something clicked for him

182
00:09:42,832 --> 00:09:46,544
that this wasn't an ordinary Go program.

183
00:09:46,627 --> 00:09:49,630
We weren't just doing
the same as everyone else,

184
00:09:49,714 --> 00:09:51,215
that something new was happening.

185
00:09:51,299 --> 00:09:53,926
The second game,
I tried to change my style.

186
00:09:54,594 --> 00:09:57,763
But the problem
is I also lost the second game.

187
00:09:57,847 --> 00:09:59,682
And also the third game,

188
00:09:59,765 --> 00:10:00,891
the fourth game,

189
00:10:01,392 --> 00:10:02,518
even the last game.

190
00:10:03,936 --> 00:10:06,105
I lost, five-zero.

191
00:10:08,190 --> 00:10:09,650
AlphaGo won all the games.

192
00:10:10,318 --> 00:10:13,070
After losing, losing, losing,

193
00:10:13,154 --> 00:10:16,407
you can feel his pressure
is getting heavier and heavier.

194
00:10:16,949 --> 00:10:18,993
And several times after the game,

195
00:10:19,076 --> 00:10:23,289
he said he wanted to go out for fresh air.

196
00:10:23,372 --> 00:10:25,625
I said, "Oh, I can go with you
and have a chat."

197
00:10:25,708 --> 00:10:27,228
He said, "No, I want to go by myself."

198
00:10:30,004 --> 00:10:32,882
I felt something very strange.

199
00:10:33,549 --> 00:10:34,842
I lost to a program.

200
00:10:35,551 --> 00:10:39,096
And I didn't understand myself anymore.

201
00:10:43,225 --> 00:10:45,311
I worried whether
he would even come back.

202
00:10:45,394 --> 00:10:47,938
You know, he seemed very low,

203
00:10:48,022 --> 00:10:50,691
and he spent about an hour away
from the office.

204
00:10:50,775 --> 00:10:52,610
He came back completely changed.

205
00:10:52,693 --> 00:10:56,405
I told Maddy,
"This is the first time in history

206
00:10:56,489 --> 00:11:00,493
that a human professional Go player
lost to a program.

207
00:11:00,993 --> 00:11:02,828
She told me, "Yes, we know.

208
00:11:02,912 --> 00:11:05,873
But for you... Are you okay?"

209
00:11:06,582 --> 00:11:07,667
I told her...

210
00:11:09,418 --> 00:11:11,545
"No, and yes."

211
00:11:12,213 --> 00:11:14,924
I am not happy to lose a game,

212
00:11:15,383 --> 00:11:19,595
but I'm very happy to play for history.

213
00:11:21,764 --> 00:11:25,935
Artificial intelligence researchers
have solved the game of Go

214
00:11:26,018 --> 00:11:28,396
a decade earlier than expected.

215
00:11:28,479 --> 00:11:30,314
The computer named AlphaGo,

216
00:11:30,398 --> 00:11:32,942
was able to beat
the European human champion.

217
00:11:33,025 --> 00:11:34,902
Artificial intelligence researchers

218
00:11:34,985 --> 00:11:36,695
have made a significant breakthrough.

219
00:11:36,779 --> 00:11:38,864
It really is a big leap forward.

220
00:11:39,323 --> 00:11:44,120
There's a big difference between
the way the IBM computer beat Kasparov,

221
00:11:44,203 --> 00:11:46,747
which was programmed
by expert chess players,

222
00:11:46,831 --> 00:11:50,709
and the way the Go-playing computer,
more or less, learned itself.

223
00:11:50,793 --> 00:11:53,087
The way we start off training AlphaGo

224
00:11:53,170 --> 00:11:57,299
is by showing it 100,000 games
that strong amateurs have played,

225
00:11:57,383 --> 00:11:59,027
that we have downloaded from the Internet.

226
00:11:59,051 --> 00:12:02,680
And we first initially get AlphaGo
to mimic the human player.

227
00:12:02,763 --> 00:12:05,015
And then, through self-playing,
reinforcement learning,

228
00:12:05,099 --> 00:12:07,935
it plays against different versions
of itself many millions of times,

229
00:12:08,018 --> 00:12:09,687
and learns from its errors.

230
00:12:09,770 --> 00:12:13,732
The specific ideas
that are driving AlphaGo

231
00:12:13,816 --> 00:12:15,526
are going to drive our future.

232
00:12:15,860 --> 00:12:18,320
The technologies
at the heart of AlphaGo,

233
00:12:18,404 --> 00:12:20,698
they are what are called
deep neural networks

234
00:12:20,781 --> 00:12:22,700
which essentially mimic

235
00:12:22,783 --> 00:12:24,910
the web of neurons in the human brain.

236
00:12:24,994 --> 00:12:26,412
It's a very old idea,

237
00:12:26,495 --> 00:12:30,207
but recently, due to increases
in computing power,

238
00:12:30,291 --> 00:12:33,002
these neural networks
have become extremely powerful,

239
00:12:33,085 --> 00:12:33,961
almost overnight.

240
00:12:34,044 --> 00:12:37,214
Big neural networks
that operate on big data

241
00:12:37,298 --> 00:12:39,216
can achieve surprising things.

242
00:12:39,300 --> 00:12:41,802
AlphaGo found a way to learn
how to play Go.

243
00:12:41,886 --> 00:12:44,722
Learning is the key thing here.
It is a machine learning.

244
00:12:44,805 --> 00:12:46,849
The whole beauty
about these types of algorithms

245
00:12:46,932 --> 00:12:48,827
is that because they are learning
for themselves,

246
00:12:48,851 --> 00:12:49,727
they can go beyond

247
00:12:49,810 --> 00:12:52,730
what we, as the programmers,
know how to do,

248
00:12:52,813 --> 00:12:56,609
and allow us to make new breakthroughs
in areas of science and medicine.

249
00:12:56,692 --> 00:13:01,489
So, AlphaGo is one significant step
towards that ultimate goal.

250
00:13:11,165 --> 00:13:12,833
My wife called me.

251
00:13:13,417 --> 00:13:16,378
She told me, "Don't check the Internet.

252
00:13:16,462 --> 00:13:17,880
Don't connect to the Internet,

253
00:13:17,963 --> 00:13:20,424
because people are saying terrible things

254
00:13:20,508 --> 00:13:22,551
about your match with AlphaGo."

255
00:13:23,052 --> 00:13:26,013
"Fan Hui has been living
for a long time in Europe,

256
00:13:26,096 --> 00:13:28,641
he is not a real professional,
he's just an amateur player."

257
00:13:29,475 --> 00:13:32,186
This is so hard for me, so hard.

258
00:13:34,063 --> 00:13:35,856
The Go world was skeptical

259
00:13:35,940 --> 00:13:38,442
about how strong really was AlphaGo,

260
00:13:38,526 --> 00:13:40,027
and how much further did it need

261
00:13:40,110 --> 00:13:41,904
to get to beat the top professionals.

262
00:13:42,404 --> 00:13:45,074
So our program is improving over time,

263
00:13:45,157 --> 00:13:47,952
and we want to push
the AI algorithm to the limit,

264
00:13:48,035 --> 00:13:51,622
and see how far this kind
of self-improving process can go.

265
00:13:52,373 --> 00:13:55,251
So, we needed to look
for an even greater challenge.

266
00:13:56,502 --> 00:14:00,130
A match like no other
is about to get underway in South Korea.

267
00:14:00,214 --> 00:14:02,424
Se-dol Lee,
the long-reigning global champ...

268
00:14:02,508 --> 00:14:04,009
This guy is a genius...

269
00:14:04,093 --> 00:14:07,513
...who will take on
artificial intelligence program AlphaGo

270
00:14:07,596 --> 00:14:10,641
in the ultimate
human versus machine smackdown.

271
00:14:10,724 --> 00:14:14,687
This is a huge moment for both
the world of artificial intelligence

272
00:14:14,770 --> 00:14:16,855
and I think, the world of Go.

273
00:14:16,939 --> 00:14:20,609
So far, AlphaGo has beaten
every challenge we have given it.

274
00:14:20,693 --> 00:14:23,737
But we won't know
its true strength until we play somebody

275
00:14:23,821 --> 00:14:26,740
who is at the top of the world,
like Se-dol Lee.

276
00:14:26,824 --> 00:14:28,284
We chose Se-dol Lee

277
00:14:28,367 --> 00:14:31,912
because we wanted
a legendary, historic player.

278
00:14:31,996 --> 00:14:35,791
Somebody who has been acknowledged
as the greatest player of the last decade.

279
00:14:36,208 --> 00:14:40,170
I don't want to sound too arrogant,

280
00:14:40,254 --> 00:14:42,298
but, still, I think I have the advantage.

281
00:14:42,381 --> 00:14:44,508
"I will not be too arrogant,

282
00:14:44,592 --> 00:14:47,803
but I don't think
that it will be a very close match.

283
00:14:47,887 --> 00:14:52,308
The level of the player that AlphaGo
went against in October

284
00:14:52,391 --> 00:14:53,893
is not the same level as me.

285
00:14:53,976 --> 00:14:57,396
So, given that a couple of months
has only passed,

286
00:14:57,479 --> 00:15:01,233
I don't think that it is enough for it
to be able to catch up with me.

287
00:15:01,317 --> 00:15:06,322
My hope is that it will be either
five-zero for me, or maybe four to one.

288
00:15:06,405 --> 00:15:10,200
So the critical point for me
is to make sure I do not lose one."

289
00:15:10,784 --> 00:15:12,411
High five, please.

290
00:15:13,412 --> 00:15:15,497
We don't know how well

291
00:15:15,581 --> 00:15:17,291
our system will play against someone

292
00:15:17,374 --> 00:15:19,418
as creative as Se-dol Lee.

293
00:15:19,501 --> 00:15:20,586
Also, he's very famous

294
00:15:20,669 --> 00:15:24,048
- for very creative fighting play.
- Yeah, yeah. Yeah.

295
00:15:24,131 --> 00:15:27,301
So, this could be difficult for us,
but we'll see.

296
00:15:27,384 --> 00:15:28,218
Maybe, in some ways,

297
00:15:28,302 --> 00:15:30,142
he's the most difficult opponent
we could pick.

298
00:15:30,220 --> 00:15:31,055
Yeah.

299
00:15:31,138 --> 00:15:33,557
There was still so much
of a question

300
00:15:33,641 --> 00:15:35,643
about whether or not they could beat

301
00:15:35,726 --> 00:15:37,561
someone like Se-dol Lee.

302
00:15:38,062 --> 00:15:39,438
Fan Hui is a good player,

303
00:15:39,521 --> 00:15:42,650
but he's nothing
like the very top players.

304
00:15:42,733 --> 00:15:44,902
Se-dol Lee
is a 9-dan professional.

305
00:15:45,194 --> 00:15:48,739
Nine-dan professional is the highest rank
that you can achieve in Go.

306
00:15:48,822 --> 00:15:50,240
The ranking system.

307
00:15:50,324 --> 00:15:54,161
Fan Hui, who is the European Champion,
is just a 2-dan professional.

308
00:15:54,244 --> 00:15:58,791
On the other hand, at the very top,
Se-dol Lee is a 9-dan professional.

309
00:15:58,874 --> 00:16:01,627
Se-dol Lee is to Go
what Roger Federer is to tennis.

310
00:16:01,710 --> 00:16:03,879
He is playing in Wimbledon
to win the Grand Slam.

311
00:16:03,963 --> 00:16:06,024
And it's not just this year.
He'll be there next year,

312
00:16:06,048 --> 00:16:08,258
and the year after that,
and the year after that.

313
00:16:08,342 --> 00:16:10,302
Of course, the Internet,

314
00:16:10,386 --> 00:16:13,847
and all the Go community, everyone,

315
00:16:14,932 --> 00:16:19,311
maybe not 100%, maybe 99.999%

316
00:16:19,395 --> 00:16:22,940
think Se-dol Lee will win very easily.

317
00:16:23,565 --> 00:16:25,985
Teacher, I got stuck.

318
00:16:26,068 --> 00:16:27,361
How do I go from here?

319
00:16:29,863 --> 00:16:31,615
Good.
Hold the stone properly.

320
00:16:31,699 --> 00:16:34,326
Place the stone nicely, not like that.

321
00:16:34,410 --> 00:16:36,412
And hold it nicely, like this.

322
00:16:37,746 --> 00:16:39,665
Go is taken seriously in Korea.

323
00:16:39,957 --> 00:16:43,210
It's so much part of the culture,
like breathing, or, you know,

324
00:16:43,293 --> 00:16:45,796
like taking swimming lessons,
or something.

325
00:16:46,505 --> 00:16:48,132
On the surface, it's a game,

326
00:16:48,215 --> 00:16:51,760
but inside, it has a very deep philosophy.

327
00:16:51,844 --> 00:16:55,514
The Go board reflects
the individual who's playing.

328
00:16:55,597 --> 00:16:57,891
The truth is going to show itself
on the board.

329
00:16:57,975 --> 00:16:59,560
You won't be able to hide it.

330
00:17:00,436 --> 00:17:03,689
In ancient China, Japan, Korea,

331
00:17:03,772 --> 00:17:07,693
Go is one of the four noble things,

332
00:17:07,776 --> 00:17:10,863
like four accomplishments for literati,

333
00:17:11,280 --> 00:17:15,034
with music, poetry, and painting.

334
00:17:15,117 --> 00:17:18,746
So, people think the Go players

335
00:17:18,829 --> 00:17:21,415
are very smart and very noble.

336
00:17:21,498 --> 00:17:22,833
It's okay to lose,

337
00:17:22,916 --> 00:17:24,185
as long as you played your best.

338
00:17:24,209 --> 00:17:26,879
Bow to your opponent.

339
00:17:26,962 --> 00:17:29,339
Master Kwon started this school,

340
00:17:29,423 --> 00:17:31,258
so that he could produce

341
00:17:31,341 --> 00:17:32,885
great players in Korea.

342
00:17:33,427 --> 00:17:36,180
Se-dol Lee, when he was eight years old,

343
00:17:36,263 --> 00:17:38,557
he came in here, attending class,

344
00:17:38,640 --> 00:17:41,810
9:00 a.m. until 9:00 p.m.
for seven days.

345
00:17:41,894 --> 00:17:44,063
And then, he stayed with Master Kwon.

346
00:17:44,897 --> 00:17:47,024
I remember his round face

347
00:17:47,107 --> 00:17:49,359
and dark brown eyes.

348
00:17:49,985 --> 00:17:54,615
Since he came
from an island, he easily drew attention.

349
00:17:55,115 --> 00:17:58,911
But unlike the other children,

350
00:17:58,994 --> 00:18:01,663
his eyes shone brightly.

351
00:18:01,747 --> 00:18:02,747
7TH WORLD GO CHAMPIONSHIP

352
00:18:02,790 --> 00:18:05,292
Se-dol Lee was a little boy

353
00:18:05,375 --> 00:18:08,837
that I remember when I was a student
studying Go.

354
00:18:08,921 --> 00:18:14,343
He was very young and rural enough
to think that pizza grew on trees.

355
00:18:15,844 --> 00:18:17,346
When I was young,

356
00:18:17,429 --> 00:18:19,848
I started playing Go
because I thought it was fun

357
00:18:19,932 --> 00:18:21,600
to beat guys who were older than me.

358
00:18:21,683 --> 00:18:25,145
Since then, it has become
a type of creation of mine.

359
00:18:25,687 --> 00:18:27,856
I want my style of Go
to be something different,

360
00:18:27,940 --> 00:18:29,149
something new,

361
00:18:30,150 --> 00:18:31,819
my own thing.

362
00:18:33,654 --> 00:18:36,031
Something that no one
has thought of before.

363
00:18:36,615 --> 00:18:38,742
Se-dol Lee plays things

364
00:18:38,826 --> 00:18:41,495
that are interesting, where you felt like

365
00:18:41,578 --> 00:18:42,955
he is beyond winning and losing.

366
00:18:43,038 --> 00:18:47,000
He wants to do something that's innovative
or takes things to the next level.

367
00:18:47,084 --> 00:18:49,628
So every Go player
studies his game for sure.

368
00:18:49,711 --> 00:18:50,921
21ST ASIAN TV CUP CHAMPIONSHIP

369
00:18:51,004 --> 00:18:53,799
Roughly for ten years, he dominated
the professional Go world.

370
00:18:53,882 --> 00:18:55,968
He won 18 World Championships.

371
00:18:56,051 --> 00:18:57,302
Se-dol Lee

372
00:18:57,386 --> 00:18:59,054
is a genius of the century.

373
00:18:59,138 --> 00:19:01,449
2012 GO CHAMPIONSHIP AWARD CEREMONY
AT GRAND CENTRAL HOTEL, CHINA

374
00:19:01,473 --> 00:19:05,394
Even when I look back now,
I am proud of him.

375
00:19:05,477 --> 00:19:08,522
And I'm proud of myself, too.

376
00:19:09,148 --> 00:19:11,942
Definitely, Se-dol Lee
is going to win the game, five all.

377
00:19:12,025 --> 00:19:13,425
IF YOU DESIRE TO WIN, YOU WILL LOSE

378
00:19:14,153 --> 00:19:15,571
Yeah, all the games.

379
00:19:16,113 --> 00:19:17,531
He's going to win all the games.

380
00:19:18,699 --> 00:19:21,368
We had our evaluation match
last week.

381
00:19:21,660 --> 00:19:23,579
We won a game, and we lost a game.

382
00:19:23,662 --> 00:19:24,788
And we lost a game in a way

383
00:19:24,872 --> 00:19:26,516
that would have made us look
extremely foolish

384
00:19:26,540 --> 00:19:28,125
if that happened publicly.

385
00:19:28,917 --> 00:19:31,128
It means that we still have work to do,

386
00:19:31,211 --> 00:19:32,856
and we need to take this really seriously.

387
00:19:32,880 --> 00:19:35,215
There's just too much risk that, actually,

388
00:19:35,299 --> 00:19:37,926
we could lose, overall.

389
00:19:38,010 --> 00:19:41,638
Not only that, but we could lose in a way
that makes us look rather silly.

390
00:19:41,722 --> 00:19:43,307
- Yeah.
- So, I guess, um...

391
00:19:43,849 --> 00:19:46,244
Yeah, Aja, did you want to say anything
about what you're trying?

392
00:19:46,268 --> 00:19:47,644
I'm working hard.

393
00:19:51,356 --> 00:19:53,317
So, we're working round the clock
at the moment,

394
00:19:53,400 --> 00:19:55,027
training our algorithms further,

395
00:19:55,110 --> 00:19:57,112
trying to incrementally keep on improving

396
00:19:57,196 --> 00:19:59,448
right up to the moment
where we have the match.

397
00:19:59,531 --> 00:20:00,866
We have collected together

398
00:20:00,949 --> 00:20:03,452
people with different skill sets
onto the same team.

399
00:20:03,535 --> 00:20:07,456
So we have researchers,
engineers, valuation guys.

400
00:20:07,831 --> 00:20:09,809
But you mean what happens
if that creates something new

401
00:20:09,833 --> 00:20:11,501
- which wasn't in heavy node?
- Yeah.

402
00:20:12,044 --> 00:20:13,128
Then we don't see it.

403
00:20:13,212 --> 00:20:14,254
You'll see it.

404
00:20:14,338 --> 00:20:15,797
- Okay.
- The first thing to know...

405
00:20:15,881 --> 00:20:17,483
I was thinking
we have the perfect solution,

406
00:20:17,507 --> 00:20:19,218
- but it's not perfect.
- Yeah.

407
00:20:19,760 --> 00:20:24,431
Aja is the lead programmer,
and built the original search engine.

408
00:20:24,514 --> 00:20:26,516
So Aja's responsibility
is quite a big one.

409
00:20:26,600 --> 00:20:29,144
He'll be the one sitting
opposite Se-dol Lee

410
00:20:29,228 --> 00:20:32,606
and actually playing
the moves AlphaGo makes.

411
00:20:32,689 --> 00:20:34,691
- I'm feeling excited.
- Yeah?

412
00:20:36,401 --> 00:20:38,695
Se-dol Lee is a great player,

413
00:20:38,779 --> 00:20:42,199
and I feel honored to play with him.

414
00:20:42,866 --> 00:20:45,535
Many of my friends,
they are very excited about the match.

415
00:20:45,619 --> 00:20:48,914
They keep telling me
that the whole world is watching.

416
00:20:48,997 --> 00:20:51,083
"Just prepare AlphaGo."

417
00:20:51,541 --> 00:20:52,542
Hello.

418
00:20:52,626 --> 00:20:54,127
Hello, Demis. Can you hear us?

419
00:20:54,211 --> 00:20:55,504
Yes, I can hear you, guys.

420
00:20:55,587 --> 00:20:57,089
Can you hear me okay?

421
00:20:57,172 --> 00:20:58,507
Yes. Perfect.

422
00:20:58,590 --> 00:21:00,550
"I was curious to meet you,

423
00:21:00,634 --> 00:21:02,803
such an amazing developer

424
00:21:02,886 --> 00:21:04,846
who made AlphaGo.

425
00:21:04,930 --> 00:21:06,306
It's very nice to meet you."

426
00:21:06,390 --> 00:21:08,809
Thank you, thank you. Likewise.

427
00:21:08,892 --> 00:21:11,395
It's a real pleasure to meet you.

428
00:21:11,478 --> 00:21:14,147
I love the game of Go. I'm not very good,

429
00:21:14,231 --> 00:21:16,775
but I don't know if you know,

430
00:21:16,858 --> 00:21:19,111
but we actually have something in common.

431
00:21:19,194 --> 00:21:22,572
I actually also trained at a game
at a young age.

432
00:21:22,656 --> 00:21:25,742
I used to be, when I was young,
a professional chess player.

433
00:21:25,826 --> 00:21:27,744
He said that he was once

434
00:21:27,828 --> 00:21:28,888
a professional chess player.

435
00:21:28,912 --> 00:21:32,082
I used to play for the England team,

436
00:21:32,165 --> 00:21:36,086
and then I stopped playing
when I was about 14. I was...

437
00:21:36,169 --> 00:21:39,506
When I was 13, I was the second
highest-rated player in the world.

438
00:21:39,589 --> 00:21:40,924
"After the match,

439
00:21:41,008 --> 00:21:44,761
I would like to propose
that I give you one teaching game of Go,

440
00:21:44,845 --> 00:21:47,347
and you give me
one teaching game of chess."

441
00:21:47,431 --> 00:21:48,974
- Yeah.
- "How does that sound?"

442
00:21:49,057 --> 00:21:52,561
That sounds very good.
I would be very honored to do that.

443
00:21:52,644 --> 00:21:56,523
"I'm not sure if it's okay
to ask you this question,

444
00:21:56,606 --> 00:21:59,192
but I saw the games against Fan Hui,

445
00:21:59,276 --> 00:22:03,447
and I didn't think it was quite
at the level to play with me,

446
00:22:03,530 --> 00:22:06,700
but I heard
that it is getting really stronger.

447
00:22:06,783 --> 00:22:10,287
Can I ask how much
it got stronger since then?"

448
00:22:13,373 --> 00:22:16,585
Yeah. I can't say too much, of course,

449
00:22:16,668 --> 00:22:20,172
but it's definitely got
significantly stronger.

450
00:22:21,757 --> 00:22:24,718
With more time,
I think, a logical approach

451
00:22:24,801 --> 00:22:27,929
would be to follow up
on Aja's local search,

452
00:22:28,013 --> 00:22:29,765
- run that fast in AlphaGo,
- Right.

453
00:22:29,848 --> 00:22:31,325
- generate a whole new data set,
- Exactly.

454
00:22:31,349 --> 00:22:33,018
- and iterate more time.
- Yeah.

455
00:22:33,101 --> 00:22:35,812
But we're just out of time for that,
so we have to be realistic.

456
00:22:35,896 --> 00:22:38,648
We realized that
if we wanted to prepare

457
00:22:38,732 --> 00:22:42,069
for a mammoth task
like taking on Se-dol Lee,

458
00:22:42,152 --> 00:22:45,405
there would be nothing better
than talking to a professional.

459
00:22:45,489 --> 00:22:48,200
And we couldn't have picked
a better person than Fan Hui.

460
00:22:51,661 --> 00:22:54,331
We invited him as an advisor

461
00:22:54,414 --> 00:22:58,126
because during the match,
we realized he is a man of good spirit.

462
00:22:58,210 --> 00:22:59,961
- Welcome back.
- Thank you very much.

463
00:23:00,045 --> 00:23:02,506
It's very precious
that we have Fan Hui with us.

464
00:23:02,589 --> 00:23:03,757
Let's go down this way.

465
00:23:05,383 --> 00:23:07,302
He was crushed by AlphaGo,

466
00:23:07,385 --> 00:23:09,054
but then, he became very positive

467
00:23:09,137 --> 00:23:10,847
and was a big help to us.

468
00:23:11,973 --> 00:23:13,266
When Demis told me,

469
00:23:13,350 --> 00:23:16,561
"Can you come back to help us
make AlphaGo stronger?"

470
00:23:17,646 --> 00:23:19,189
I felt a sense of respect.

471
00:23:20,023 --> 00:23:22,025
Of course, I accepted it.

472
00:23:24,152 --> 00:23:26,029
I played with AlphaGo

473
00:23:26,113 --> 00:23:30,200
to understand
what AlphaGo's strong points are

474
00:23:30,283 --> 00:23:33,411
and what may be its weakness.

475
00:23:34,329 --> 00:23:35,539
I played in the morning,

476
00:23:36,123 --> 00:23:37,207
in the afternoon,

477
00:23:38,333 --> 00:23:39,333
all the time.

478
00:23:40,710 --> 00:23:42,295
And I found something.

479
00:23:43,380 --> 00:23:46,258
I found AlphaGo's big weakness.

480
00:23:46,925 --> 00:23:48,135
It was a big one.

481
00:23:48,927 --> 00:23:50,345
The superheroes always have

482
00:23:50,428 --> 00:23:52,389
a hidden vulnerability, right?

483
00:23:52,472 --> 00:23:54,766
And the same is true for AlphaGo.

484
00:23:54,850 --> 00:23:59,146
It's unbelievably superhuman, in general,

485
00:23:59,229 --> 00:24:03,108
but it has some particular weaknesses
in some situations.

486
00:24:03,191 --> 00:24:04,818
We can think of there
being this space

487
00:24:04,901 --> 00:24:06,278
of all the things it knows about,

488
00:24:06,361 --> 00:24:08,572
and it knows about most of it
extremely well.

489
00:24:08,655 --> 00:24:11,116
Then there will be these tricky lumps
of knowledge,

490
00:24:11,199 --> 00:24:13,910
that it just understands very poorly.

491
00:24:13,994 --> 00:24:16,079
And it's really hard for us
to characterize

492
00:24:16,163 --> 00:24:18,206
when it is going to enter
into one of these lumps.

493
00:24:18,290 --> 00:24:20,542
But if it does,
it can be completely delusional,

494
00:24:20,625 --> 00:24:22,520
thinking that it is alive
on one part of the board

495
00:24:22,544 --> 00:24:24,546
when, in fact, it's dead or vice versa.

496
00:24:25,005 --> 00:24:27,716
So, there is a real risk
that we could lose the match.

497
00:24:28,884 --> 00:24:30,552
Everybody in the team

498
00:24:30,635 --> 00:24:33,513
tried to work more and more
to fix the problem,

499
00:24:33,597 --> 00:24:35,974
but I thought it was difficult

500
00:24:36,057 --> 00:24:37,934
for it to be fixed very quickly.

501
00:24:38,018 --> 00:24:40,520
We've got version 18
in the pipeline, haven't we?

502
00:24:41,062 --> 00:24:42,942
Well, we will only go
for version 18 if it makes

503
00:24:43,023 --> 00:24:45,108
- a significant improvement which...
- If it... Yes.

504
00:24:45,650 --> 00:24:49,112
We're having to rerun all the tests
because they went wrong, unfortunately,

505
00:24:49,196 --> 00:24:52,782
so we're basically a day behind
in our evaluation because of that.

506
00:24:53,325 --> 00:24:55,076
Well, maybe we just need
to be realistic.

507
00:24:55,160 --> 00:24:56,912
That we've tried a bunch of things,

508
00:24:56,995 --> 00:24:59,315
and we've come to the point
where we said we would actually

509
00:24:59,372 --> 00:25:02,167
start freezing the code, and saying,

510
00:25:02,250 --> 00:25:03,293
"Look, it's...

511
00:25:04,419 --> 00:25:06,630
not... actually, not panning out."

512
00:25:07,047 --> 00:25:10,842
These delusions are still
a realistic possibility for the match.

513
00:25:10,926 --> 00:25:12,510
We have some weaknesses

514
00:25:12,594 --> 00:25:17,265
that we... I don't think we are going
to fix fully before the match,

515
00:25:17,349 --> 00:25:20,518
so that's causing us
a little bit of anxiety.

516
00:25:26,441 --> 00:25:29,194
Se-dol Lee is getting ready to rumble.

517
00:25:29,277 --> 00:25:31,446
On Wednesday, live across the Internet,

518
00:25:31,529 --> 00:25:33,865
this professional South Korean Go player

519
00:25:33,949 --> 00:25:37,202
will take on artificial intelligence
program AlphaGo.

520
00:25:37,285 --> 00:25:42,249
I am confident
that I will win five to zero.

521
00:25:43,458 --> 00:25:46,503
Tomorrow, Julian and George,
they pack up version 18,

522
00:25:46,586 --> 00:25:49,047
they stick it on a laptop,
and they fly out to Seoul.

523
00:25:53,301 --> 00:25:54,803
- See you all in Korea.
- Safe trip.

524
00:25:54,886 --> 00:25:56,246
- Have a safe trip.
- See you, too.

525
00:26:07,983 --> 00:26:11,444
Is everything all set
for the victory against Se-dol Lee?

526
00:26:11,528 --> 00:26:12,612
Everything is set, I think.

527
00:26:12,696 --> 00:26:15,615
But when I get to the hotel,
I'm going to catch up with the team.

528
00:26:15,699 --> 00:26:17,367
What we should do
in the time remaining

529
00:26:17,450 --> 00:26:21,579
is list things that could go wrong
in the solidity of the system.

530
00:26:21,663 --> 00:26:25,041
We need somewhere
to put AlphaGo under here.

531
00:26:25,125 --> 00:26:27,377
All right, rehearsal.
Let's go, please, folks!

532
00:26:28,003 --> 00:26:29,379
Have you arranged the place

533
00:26:29,462 --> 00:26:30,922
for Se-dol Lee to smoke?

534
00:26:31,006 --> 00:26:33,166
When he is in a difficult situation,
he likes to smoke.

535
00:26:33,216 --> 00:26:35,844
Yes, there is a terrace,
and we will have security,

536
00:26:35,927 --> 00:26:38,013
so he will be able to go up there
and be by himself.

537
00:26:38,096 --> 00:26:40,348
Which one do you support,

538
00:26:40,432 --> 00:26:42,142
Se-dol Lee or AlphaGo?

539
00:26:43,268 --> 00:26:44,853
I am in an awkward position.

540
00:26:44,936 --> 00:26:46,855
- As you can see, I am a referee.
- I see.

541
00:26:46,938 --> 00:26:48,189
I am a referee.

542
00:26:49,441 --> 00:26:50,692
When I got here,

543
00:26:50,775 --> 00:26:52,485
I didn't expect
the attention on the match.

544
00:26:52,569 --> 00:26:54,487
It was literally front-page news.

545
00:26:54,571 --> 00:26:57,866
About eight million Koreans
play the game of Go,

546
00:26:57,949 --> 00:27:00,452
and, even those who don't,
you know, recognize Se-dol Lee.

547
00:27:00,535 --> 00:27:02,037
He is a national figure.

548
00:27:02,704 --> 00:27:03,955
And so, there's that. Right?

549
00:27:04,039 --> 00:27:05,582
There's some national pride involved.

550
00:27:05,665 --> 00:27:07,167
But, it's more than that.

551
00:27:07,250 --> 00:27:11,713
Just the very thought of a machine
playing a human in something like this,

552
00:27:11,796 --> 00:27:13,882
I think, is inherently intriguing
to people.

553
00:27:14,883 --> 00:27:16,134
We can bring him in?

554
00:27:16,968 --> 00:27:18,219
Let's go in? Yes.

555
00:27:29,022 --> 00:27:31,483
I am confident about the match.

556
00:27:31,566 --> 00:27:35,987
I believe that human intuition
is still too advanced

557
00:27:36,863 --> 00:27:40,825
for AI to have caught up.

558
00:27:41,326 --> 00:27:45,747
I'm going to do my best
to protect human intelligence.

559
00:27:47,499 --> 00:27:49,834
The whole world put pressure

560
00:27:49,918 --> 00:27:51,086
to Se-dol Lee.

561
00:27:51,920 --> 00:27:56,091
Before this, he played the tournament
for his country, for himself.

562
00:27:56,674 --> 00:27:58,927
But this time,
he's playing for the human race.

563
00:28:07,936 --> 00:28:09,896
I just really hope we win
this first game.

564
00:28:09,979 --> 00:28:11,040
If you lose the first game,

565
00:28:11,064 --> 00:28:12,708
you literally have to win
three out of four.

566
00:28:12,732 --> 00:28:13,732
Yeah, that's right.

567
00:28:13,775 --> 00:28:15,276
- Which is hard work.
- Yeah.

568
00:28:15,360 --> 00:28:16,400
Where are you going to be?

569
00:28:16,444 --> 00:28:17,838
- I'll start in the match room.
- Okay.

570
00:28:17,862 --> 00:28:19,924
Then, I'm gonna come in here,
make sure everything's fine.

571
00:28:19,948 --> 00:28:22,508
- Yeah, you should probably be in here.
- Or something like that.

572
00:28:25,745 --> 00:28:28,039
- I know you're nervous.
- Yeah, I'm nervous.

573
00:28:28,790 --> 00:28:29,874
I'm nervous.

574
00:28:31,334 --> 00:28:32,603
How are you? You're not nervous?

575
00:28:32,627 --> 00:28:34,212
- A bit.
- Little bit, yeah?

576
00:28:34,295 --> 00:28:36,089
It's good to have a little bit of nerves.

577
00:28:38,133 --> 00:28:40,135
It should be fine.

578
00:28:41,219 --> 00:28:42,470
Be fine. Be fine.

579
00:28:42,971 --> 00:28:46,433
Hello, and welcome
to the DeepMind Challenge,

580
00:28:46,516 --> 00:28:48,685
Game One, Round One,

581
00:28:48,768 --> 00:28:52,272
live from the Four Seasons
here in Seoul, Korea.

582
00:28:52,355 --> 00:28:55,233
I'm Chris Garlock
of The American Go E-Journal.

583
00:28:55,316 --> 00:28:58,445
I'm here with Michael Redmond,
9-dan professional.

584
00:28:58,528 --> 00:28:59,528
Welcome, Michael.

585
00:28:59,571 --> 00:29:03,700
Want to give a shout-out to all the folks
watching around the world?

586
00:29:03,783 --> 00:29:06,453
Well, the excitement
is pretty palpable here in the hotel.

587
00:29:06,536 --> 00:29:10,206
I've never seen a crush of interest
of reporters from around the world.

588
00:29:10,290 --> 00:29:12,050
- The cameras...
- The media is overwhelming.

589
00:29:12,125 --> 00:29:13,918
I am not sure how much AlphaGo

590
00:29:14,002 --> 00:29:17,714
has changed in the last five months.

591
00:29:17,797 --> 00:29:20,216
All right, folks. You're here,
you're gonna see history made.

592
00:29:20,300 --> 00:29:21,300
Stay with us.

593
00:29:21,468 --> 00:29:24,262
Five minutes, guys. Five minutes.

594
00:29:29,851 --> 00:29:32,645
Please don't push.

595
00:29:32,729 --> 00:29:34,272
I just thought we should

596
00:29:34,355 --> 00:29:39,819
take a moment together and just think
about what is about to happen.

597
00:29:39,903 --> 00:29:41,488
I'm extremely excited to be here.

598
00:29:41,571 --> 00:29:44,574
Extremely proud of every one of you
and what we have done.

599
00:29:44,657 --> 00:29:47,160
And, win or loss, I think
it is just amazing that we're here.

600
00:29:47,243 --> 00:29:49,621
Isn't it strange your dad

601
00:29:49,704 --> 00:29:50,904
is fighting against a machine?

602
00:29:52,373 --> 00:29:54,876
I would like it if the machine

603
00:29:54,959 --> 00:29:56,795
did not beat a human in Go yet.

604
00:29:59,005 --> 00:30:00,173
One minute left.

605
00:30:45,176 --> 00:30:48,304
Five, four, three, two, one.

606
00:30:53,977 --> 00:30:55,436
The match will be

607
00:30:55,520 --> 00:30:57,689
in Chinese rules, 7.5 points komi.

608
00:30:57,981 --> 00:31:01,192
The time limit is two hours,
and 60 seconds times three.

609
00:31:24,132 --> 00:31:25,842
Please choose the stone.

610
00:31:27,176 --> 00:31:29,262
Se-dol Lee, 9 dan, what have you chosen?

611
00:31:30,555 --> 00:31:31,555
Black.

612
00:31:32,348 --> 00:31:34,142
Please start the game.

613
00:31:59,042 --> 00:32:00,543
We had worked so hard

614
00:32:00,627 --> 00:32:04,380
to make sure that this would go
technically smoothly.

615
00:32:04,756 --> 00:32:07,342
We tested it and tested it and tested it.

616
00:32:08,009 --> 00:32:12,639
And still, there comes that moment
when you're live,

617
00:32:12,722 --> 00:32:15,850
all the TV cameras
are broadcasting everything,

618
00:32:15,934 --> 00:32:19,062
and now it has to do
the thing you built it to do.

619
00:32:20,521 --> 00:32:22,732
AlphaGo is even considering

620
00:32:22,815 --> 00:32:24,317
what to do from the first step.

621
00:32:24,400 --> 00:32:26,486
If you look at the time closely,

622
00:32:26,569 --> 00:32:27,963
- it has been over 30 seconds.
- Yes.

623
00:32:27,987 --> 00:32:30,156
AlphaGo is thinking too long
at the second move.

624
00:32:30,239 --> 00:32:31,239
Yes.

625
00:33:08,695 --> 00:33:10,363
I was a bit nervous.

626
00:33:10,446 --> 00:33:15,535
It's the first time that I sat
in front of a world-class Go player.

627
00:33:16,369 --> 00:33:19,580
And I actually felt
the spirit and courtesy

628
00:33:19,664 --> 00:33:22,250
of a great Go player like Se-dol Lee.

629
00:33:22,750 --> 00:33:25,670
Because I think
it was the first time he faced

630
00:33:25,753 --> 00:33:27,463
a strange opponent, I think.

631
00:33:27,547 --> 00:33:30,967
It is non-human,
has no emotion, it's cold.

632
00:33:31,509 --> 00:33:33,511
But he stayed very calm.

633
00:33:34,095 --> 00:33:37,432
And I can feel his mental strength.

634
00:33:50,903 --> 00:33:51,946
Oh.

635
00:33:52,071 --> 00:33:53,573
AlphaGo peeped.

636
00:33:54,615 --> 00:33:56,701
It's unbelievable.

637
00:33:57,493 --> 00:33:59,829
This is the first time

638
00:33:59,912 --> 00:34:01,748
I am seeing AlphaGo
making a move like this.

639
00:34:01,831 --> 00:34:04,417
It feels like AlphaGo is playing
like a human.

640
00:34:04,500 --> 00:34:05,626
I'm Andrew Jackson.

641
00:34:05,710 --> 00:34:09,297
With me here is Myung-wan Kim, a 9-dan pro
from the Korean Baduk Association.

642
00:34:09,380 --> 00:34:11,799
We are here live at the Four Seasons Hotel

643
00:34:11,883 --> 00:34:12,967
on the 21st floor.

644
00:34:13,176 --> 00:34:15,136
What are we looking at?
How's the game going?

645
00:34:16,387 --> 00:34:18,347
Oh, it's fighting from the beginning.

646
00:34:18,431 --> 00:34:20,516
- From the very beginning?
- Yeah, yeah.

647
00:34:20,600 --> 00:34:23,019
- AlphaGo is playing very well.
- Yeah?

648
00:34:23,102 --> 00:34:26,481
- It's just like a top professional.
- Just like a top professional?

649
00:34:26,564 --> 00:34:28,316
Yeah, it's very aggressive.

650
00:34:30,359 --> 00:34:31,736
- It blocked!
- Did it block?

651
00:34:32,737 --> 00:34:34,781
How dare he disconnect it?

652
00:34:35,448 --> 00:34:37,909
This can't be right.

653
00:34:37,992 --> 00:34:40,203
Now, the fight is getting
really complicated.

654
00:34:40,286 --> 00:34:43,372
Um, this is actually the first time
I have seen AlphaGo

655
00:34:43,998 --> 00:34:47,335
playing a game
that has this difficult fight.

656
00:34:52,048 --> 00:34:53,466
This could make

657
00:34:53,549 --> 00:34:54,884
Se-dol Lee nervous.

658
00:34:55,343 --> 00:34:58,221
Because AlphaGo uses one minute
to 1.5 minutes in any situation.

659
00:34:59,013 --> 00:35:02,558
This action is not like a human at all.

660
00:35:03,559 --> 00:35:05,520
No matter how complex you make the game,

661
00:35:05,603 --> 00:35:07,483
AlphaGo plays as if it knows
everything already.

662
00:35:09,899 --> 00:35:12,235
I just saw him looking
at his opponent's face.

663
00:35:12,485 --> 00:35:15,363
- And that's just kind of a habit.
- It's just an instinct,

664
00:35:15,571 --> 00:35:17,549
as a player, to look at the person
across from you.

665
00:35:17,573 --> 00:35:21,661
Yeah, it's sort of something
that Se-dol Lee would do

666
00:35:21,744 --> 00:35:23,871
when he was wondering
how his opponent was feeling.

667
00:35:23,955 --> 00:35:25,957
- Right.
- It's just a habit.

668
00:35:26,040 --> 00:35:29,544
So, it's not as if Aja Huang
is going to do any giveaway

669
00:35:29,627 --> 00:35:32,505
- because Aja Huang isn't AlphaGo, right?
- Right. Right.

670
00:35:41,472 --> 00:35:42,682
He almost made a move.

671
00:35:42,765 --> 00:35:44,433
Lee smiled for a second.

672
00:35:44,517 --> 00:35:46,727
- No. That's...
- I guess Lee is mistaken?

673
00:35:46,811 --> 00:35:48,980
He was about to make a move
that makes no sense.

674
00:35:49,063 --> 00:35:51,065
- Here is the...
- What is he doing?

675
00:35:52,275 --> 00:35:54,586
If he had put his stone there,
he would have lost the game today.

676
00:35:54,610 --> 00:35:57,405
- He was trying to place it here,
- His hand...

677
00:35:57,488 --> 00:35:59,365
but he moved his hand there
without thinking.

678
00:35:59,448 --> 00:36:01,701
- He has so many things to ponder.
- Well, I think...

679
00:36:03,119 --> 00:36:04,745
He is exhausted now.

680
00:36:05,079 --> 00:36:06,414
- I think...
- Well...

681
00:36:07,331 --> 00:36:08,583
He looks a little panicked.

682
00:36:10,418 --> 00:36:13,588
This is the thing
we were most afraid of.

683
00:36:13,671 --> 00:36:14,671
Yes.

684
00:36:14,714 --> 00:36:16,299
- Self-doubt.
- Self-doubt.

685
00:36:28,519 --> 00:36:30,146
With a human, when you play,

686
00:36:30,229 --> 00:36:32,523
you can have an exchange, by feeling.

687
00:36:33,149 --> 00:36:36,986
I look at you. I know that, okay,
maybe you want to talk with me.

688
00:36:37,403 --> 00:36:38,988
Maybe you fear me.

689
00:36:39,071 --> 00:36:40,781
I can feel many, many things.

690
00:36:40,865 --> 00:36:43,659
But with AlphaGo, you can feel nothing.

691
00:36:45,077 --> 00:36:47,914
So, when you feel nothing when you play,

692
00:36:47,997 --> 00:36:50,791
you have more questions to yourself.

693
00:36:51,417 --> 00:36:53,586
In the beginning, you think,
"Okay, my move was good.

694
00:36:53,669 --> 00:36:56,589
Is it really good? Really good?

695
00:36:56,923 --> 00:36:57,924
Oh, maybe it's bad.

696
00:36:58,007 --> 00:36:59,258
Oh, terrible!

697
00:36:59,342 --> 00:37:00,676
Why did I put it here?"

698
00:37:00,760 --> 00:37:01,886
More and more.

699
00:37:02,929 --> 00:37:04,263
It's hard to know where to be.

700
00:37:04,347 --> 00:37:06,617
All the different rooms
are like exciting in different ways.

701
00:37:06,641 --> 00:37:08,309
Mm-hm. Right.

702
00:37:08,392 --> 00:37:11,145
It's quite nice to be here
at the heart of the operation.

703
00:37:11,229 --> 00:37:12,813
Yeah, I feel safe here.

704
00:37:14,899 --> 00:37:16,943
White is thinking of doing
this huge invasion here

705
00:37:17,026 --> 00:37:18,069
from its thick wall.

706
00:37:19,070 --> 00:37:20,154
Yes, go for it.

707
00:37:20,905 --> 00:37:23,074
Stone in. He's going in.

708
00:37:23,491 --> 00:37:25,251
- Look at his face. Look at his face.
- Yeah.

709
00:37:26,953 --> 00:37:28,454
That is not a confident face.

710
00:37:28,829 --> 00:37:31,082
He's pretty horrified by that.

711
00:37:31,624 --> 00:37:34,210
I can't believe
what I'm seeing right now.

712
00:37:34,293 --> 00:37:35,503
Really...

713
00:37:35,962 --> 00:37:39,298
- Se-dol, he's a little bit behind.
- Mm-hm.

714
00:37:39,382 --> 00:37:41,592
- And he made a very aggressive move here.
- Okay.

715
00:37:41,676 --> 00:37:43,052
And made it very complicated.

716
00:37:43,135 --> 00:37:44,804
- Okay.
- I think Se-dol Lee,

717
00:37:44,887 --> 00:37:47,306
- if he doesn't respond correctly,
- Uh-huh.

718
00:37:47,390 --> 00:37:49,267
then he can collapse.

719
00:37:50,977 --> 00:37:55,106
I mean, this is kind of unthinkable
as a human Go player.

720
00:37:55,564 --> 00:37:57,441
What if it knows everything

721
00:37:57,525 --> 00:37:59,193
about what's going to happen next?

722
00:37:59,277 --> 00:38:00,194
Where are we?

723
00:38:00,278 --> 00:38:04,532
Here's our search depth,
we're searching, 50 or 60 moves ahead.

724
00:38:05,074 --> 00:38:07,368
That's the maximum number
of moves ahead

725
00:38:07,451 --> 00:38:10,871
that AlphaGo is looking
from the current game position.

726
00:38:10,955 --> 00:38:15,209
It's typically over 50.
It's often over 60.

727
00:38:16,043 --> 00:38:19,630
In the games we have seen,
often around move 150,

728
00:38:19,714 --> 00:38:21,215
AlphaGo goes for the kill.

729
00:38:21,299 --> 00:38:24,719
We are at move 150 now,
so we are getting to that tactical point.

730
00:38:28,347 --> 00:38:30,057
We are all astonished,

731
00:38:30,725 --> 00:38:35,354
just in the middle of the game
because AlphaGo is, uh...

732
00:38:35,438 --> 00:38:38,941
It seems to be doing a much better job
than we all thought it would.

733
00:38:39,442 --> 00:38:43,946
I thought Se-dol Lee would be leading
the game comfortably,

734
00:38:44,030 --> 00:38:46,949
but it turned out that he's struggling
at the moment.

735
00:38:47,283 --> 00:38:50,578
But, I think, eventually,
he will prevail. I hope.

736
00:38:51,120 --> 00:38:54,123
I was, you know, more five to zero,

737
00:38:55,041 --> 00:38:57,543
but now, I'm not sure about that.

738
00:39:01,881 --> 00:39:03,215
What is this?

739
00:39:03,674 --> 00:39:04,967
AlphaGo made a mistake.

740
00:39:06,052 --> 00:39:07,172
AlphaGo made a mistake here.

741
00:39:07,470 --> 00:39:09,096
What a mistake.

742
00:39:09,180 --> 00:39:11,265
It might be the first mistake,

743
00:39:11,640 --> 00:39:15,227
kind of a clear mistake that white made.

744
00:39:15,603 --> 00:39:19,523
AlphaGo is making mistakes
at the endgame from time to time.

745
00:39:20,358 --> 00:39:23,402
If AlphaGo wins
by a small difference in the end,

746
00:39:23,861 --> 00:39:26,572
it is possible that it played
just enough to win.

747
00:39:27,239 --> 00:39:29,367
The reason I stopped talking a minute ago

748
00:39:29,450 --> 00:39:32,370
was that according to my count,
AlphaGo may have more points.

749
00:39:33,371 --> 00:39:35,539
I didn't tell you
because I couldn't believe it.

750
00:39:35,623 --> 00:39:37,208
It seems like a mistake,

751
00:39:37,291 --> 00:39:39,728
but because it's based on calculation,
it's actually not a mistake.

752
00:39:39,752 --> 00:39:40,878
Yes, yes.

753
00:39:41,253 --> 00:39:43,589
So if it's actually based
on calculation...

754
00:39:44,590 --> 00:39:46,717
it is so scary.

755
00:39:46,801 --> 00:39:50,137
It means it's just playing
with its opponent.

756
00:39:50,221 --> 00:39:52,061
Myung-wan, it looks like
you have got to count.

757
00:39:52,390 --> 00:39:53,474
Yeah, right.

758
00:39:53,557 --> 00:39:55,935
- What do you think? Yeah.
- It's done. Yes.

759
00:39:56,018 --> 00:39:58,396
White won a lot at this time.
It looks like it.

760
00:39:58,479 --> 00:40:00,272
If it's like this, white has won by a lot.

761
00:40:00,356 --> 00:40:01,190
Yeah.

762
00:40:01,273 --> 00:40:02,900
Two, three. I can't believe it.

763
00:40:03,317 --> 00:40:06,112
Thirty-six, thirty-seven...

764
00:40:06,195 --> 00:40:07,196
Seventy-two.

765
00:40:08,781 --> 00:40:09,949
He lost.

766
00:40:11,367 --> 00:40:12,368
He lost.

767
00:40:13,160 --> 00:40:14,829
Whoa, it's so shocking.

768
00:40:14,912 --> 00:40:18,916
I... I expected AlphaGo
to win only one game.

769
00:40:24,296 --> 00:40:25,297
Oh, gosh.

770
00:40:32,596 --> 00:40:36,267
I can feel his pain. Like he was...

771
00:40:36,350 --> 00:40:39,103
He couldn't believe, you know.
He couldn't accept it.

772
00:40:39,186 --> 00:40:42,648
It took him some time
to accept the outcome.

773
00:40:44,733 --> 00:40:47,194
Maybe AlphaGo is very strong now,

774
00:40:47,278 --> 00:40:49,989
but he did not want
to believe he will lose.

775
00:40:50,573 --> 00:40:52,199
And as Go professional players,

776
00:40:52,283 --> 00:40:54,160
we can't believe this,

777
00:40:54,243 --> 00:40:59,206
because, for us,
it's something unimaginable.

778
00:40:59,290 --> 00:41:02,543
It's not something that can happen now.
It's impossible.

779
00:41:03,085 --> 00:41:04,879
But, in reality...

780
00:41:05,671 --> 00:41:06,714
it is happening now.

781
00:41:08,257 --> 00:41:10,759
I think he resigned
in a very polite way.

782
00:41:11,677 --> 00:41:14,305
Se-dol Lee plays black,
but he put a white stone.

783
00:41:22,855 --> 00:41:24,190
I think he resigned.

784
00:41:26,484 --> 00:41:27,526
Oh, my gosh.

785
00:41:29,820 --> 00:41:31,322
- Look at the time. It's slow.
- Yeah.

786
00:41:31,405 --> 00:41:32,448
Wow.

787
00:41:40,581 --> 00:41:41,581
You got this.

788
00:41:42,166 --> 00:41:43,834
- Congratulations.
- Hey.

789
00:41:44,126 --> 00:41:45,126
What's up?

790
00:41:46,504 --> 00:41:51,008
I feel really good.
I feel like I really believe in AlphaGo.

791
00:41:51,717 --> 00:41:54,136
Of course, it's natural that humans
want humans to win.

792
00:41:54,220 --> 00:41:56,180
I mean, I think that's a natural response.

793
00:41:56,263 --> 00:41:58,015
But AlphaGo is human-created,

794
00:41:58,098 --> 00:42:01,602
and I think that's the ultimate sign
of human ingenuity and cleverness.

795
00:42:01,977 --> 00:42:03,812
Everything that AlphaGo does,

796
00:42:03,896 --> 00:42:07,983
it does, because a human has either
created the data that it learns from,

797
00:42:08,067 --> 00:42:10,903
created the learning algorithm
that learns from that data,

798
00:42:10,986 --> 00:42:12,446
created the search algorithm.

799
00:42:12,530 --> 00:42:15,574
All of these things have come from humans.

800
00:42:15,658 --> 00:42:18,536
So, really, this is a human endeavor.

801
00:42:33,092 --> 00:42:35,261
In the battle
between man versus machine,

802
00:42:35,344 --> 00:42:36,929
a computer just came out the victor.

803
00:42:37,012 --> 00:42:39,765
DeepMind put
its computer program to the test

804
00:42:39,848 --> 00:42:42,643
against one of the brightest minds
in the world and won.

805
00:42:42,726 --> 00:42:45,020
AlphaGo beat
a professional player

806
00:42:45,104 --> 00:42:48,482
who has 18 Go World Championships
under his belt.

807
00:42:48,566 --> 00:42:52,194
The victory is considered a breakthrough
in artificial intelligence.

808
00:42:54,196 --> 00:42:56,323
First of all,

809
00:42:56,407 --> 00:42:58,093
I would have to say
that I was very surprised,

810
00:42:58,117 --> 00:43:01,620
because I didn't think
that I would lose the game.

811
00:43:02,705 --> 00:43:05,457
I think the mistakes I made
in the beginning

812
00:43:05,541 --> 00:43:07,376
lasted until the very end.

813
00:43:07,459 --> 00:43:09,420
That's why I lost this game.

814
00:43:09,795 --> 00:43:12,006
I wasn't able to foresee.

815
00:43:12,089 --> 00:43:15,593
I didn't think that AlphaGo would
play the game in such a perfect manner.

816
00:43:15,676 --> 00:43:21,974
I have won world championship titles
and have a lot of experience,

817
00:43:22,057 --> 00:43:27,229
so losing one game won't affect me
in playing games in the future.

818
00:43:27,313 --> 00:43:31,483
I think now it's 50/50.

819
00:43:31,984 --> 00:43:37,197
I would like to express
my respect to the team

820
00:43:37,281 --> 00:43:41,744
for developing such an amazing program

821
00:43:41,827 --> 00:43:43,454
like AlphaGo.

822
00:43:45,664 --> 00:43:46,790
In research,

823
00:43:46,874 --> 00:43:49,710
we normally work
to produce an academic paper.

824
00:43:49,793 --> 00:43:51,813
It gets published,
and maybe we get to talk about it

825
00:43:51,837 --> 00:43:53,297
in a conference if we're lucky.

826
00:43:55,215 --> 00:43:57,092
This is not normal for a research.

827
00:43:57,176 --> 00:44:01,764
In fact, I've never experienced any
media attention remotely close to this.

828
00:44:01,847 --> 00:44:06,101
So, it's a special moment for us all,
and we're just enjoying it while it lasts.

829
00:44:11,690 --> 00:44:14,818
In a matchup between
man and machine, who wins?

830
00:44:14,902 --> 00:44:16,612
So far, it's the machine.

831
00:44:16,695 --> 00:44:20,324
In Seoul, South Korea,
the artificially intelligent computer

832
00:44:20,407 --> 00:44:25,162
defeated the global champion
in the ancient Chinese board game Go.

833
00:44:25,245 --> 00:44:26,514
AlphaGo Shock
AI Shocked the World

834
00:44:26,538 --> 00:44:30,000
Se-dol Lee lost the first matchup,
but he's got four more chances.

835
00:44:30,376 --> 00:44:31,669
This is the team

836
00:44:31,752 --> 00:44:35,255
that played Go with Se-dol Lee.

837
00:44:35,339 --> 00:44:36,840
We got our newspapers.

838
00:44:36,924 --> 00:44:39,284
But we are not responsible.
It's just the man in the beanie.

839
00:44:42,221 --> 00:44:43,055
Sorry.

840
00:44:43,138 --> 00:44:45,182
AI Defeats Human Brain

841
00:44:45,265 --> 00:44:47,309
It is a bit strange
being on the front cover

842
00:44:47,393 --> 00:44:49,478
and everything as a computer scientist.

843
00:44:49,561 --> 00:44:51,939
Normally, you sit there
in your corner and you code.

844
00:44:52,022 --> 00:44:53,649
Nobody really knows about it.

845
00:44:53,732 --> 00:44:54,918
Perhaps you've heard the joke,

846
00:44:54,942 --> 00:44:57,069
"How can you tell
that a computer scientist

847
00:44:57,152 --> 00:44:59,405
is an extrovert and not an introvert?"

848
00:45:00,114 --> 00:45:02,241
If he's an extrovert,
he looks at your shoes

849
00:45:02,324 --> 00:45:04,535
when he's talking to you,
instead of at his own.

850
00:45:06,412 --> 00:45:10,916
If you look at Aja,
he avoids all the cameras like crazy.

851
00:45:11,250 --> 00:45:14,253
He's like, the game is finished,
and he's, like, out the back,

852
00:45:14,336 --> 00:45:15,921
and back in his room.

853
00:45:16,672 --> 00:45:20,175
And I think a lot of computer scientists
would be like that.

854
00:45:20,676 --> 00:45:24,179
We are more about doing our work
than standing in the spotlight.

855
00:45:25,848 --> 00:45:29,810
Hello, and welcome
to Game Two, Round Two,

856
00:45:29,893 --> 00:45:32,438
in the Google DeepMind Challenge

857
00:45:32,521 --> 00:45:35,315
throwdown between man and machine.

858
00:45:35,607 --> 00:45:39,111
Game One, the machine takes down man.

859
00:45:39,445 --> 00:45:42,573
Huge shock. Headlines around the world.

860
00:45:42,656 --> 00:45:46,702
The reactions on the ground here
from the folks in Korea were just stunned.

861
00:45:46,785 --> 00:45:50,289
They estimate 60 million people
watched the game in China alone.

862
00:45:50,372 --> 00:45:52,708
Probably bringing it up
to maybe 80 million people

863
00:45:52,791 --> 00:45:54,126
who watched this game worldwide.

864
00:45:54,209 --> 00:45:55,711
It was just incredible.

865
00:45:55,794 --> 00:45:59,381
And if anything, today is probably
even more of a madhouse.

866
00:46:00,048 --> 00:46:01,550
So, Se-dol Lee knows today.

867
00:46:01,633 --> 00:46:04,261
He knows that this is just
a really important game. Right?

868
00:46:04,344 --> 00:46:05,344
He has got to win.

869
00:46:09,725 --> 00:46:11,643
Go, Se-dol Lee!

870
00:46:11,727 --> 00:46:12,895
Fighting!

871
00:46:14,104 --> 00:46:15,606
Can you tell what Lee

872
00:46:15,689 --> 00:46:17,000
is thinking by looking at his face?

873
00:46:17,024 --> 00:46:17,858
Does he seem nervous?

874
00:46:17,941 --> 00:46:20,360
Yes.
Lee seems to be very tense.

875
00:46:21,320 --> 00:46:24,531
My guess is that
he didn't sleep well last night.

876
00:46:24,615 --> 00:46:26,783
Se-dol Lee on white,

877
00:46:26,867 --> 00:46:30,037
I think probably looking
for a little payback.

878
00:46:30,454 --> 00:46:33,707
My impression is that maybe
he underestimated AlphaGo,

879
00:46:33,790 --> 00:46:35,459
and he is going to change his tactics.

880
00:46:37,711 --> 00:46:40,339
Se-dol Lee is playing
much, much slower today.

881
00:46:40,422 --> 00:46:41,298
Yeah.

882
00:46:41,381 --> 00:46:42,925
- Which is going to affect us.
- Yes.

883
00:46:43,342 --> 00:46:44,982
He's playing at half the speed, actually.

884
00:46:45,719 --> 00:46:47,846
It is hard to say who is ahead

885
00:46:47,930 --> 00:46:49,097
or better now.

886
00:46:49,181 --> 00:46:53,310
The last two moves
make me doubt AlphaGo's ability.

887
00:46:53,393 --> 00:46:56,104
- But we have to stay alert.
- Yes.

888
00:46:56,188 --> 00:46:58,469
- Yes. This is difficult.
- AlphaGo is hard to understand.

889
00:46:58,524 --> 00:47:00,400
If I see anything about AlphaGo

890
00:47:00,484 --> 00:47:03,445
that is not normal,
is maybe the way it handles a game

891
00:47:03,529 --> 00:47:05,822
- when it thinks it is ahead.
- Yes.

892
00:47:05,906 --> 00:47:08,325
We're actually going to have a visit
from one of the team,

893
00:47:08,408 --> 00:47:10,994
and we will talk about exactly
that point from the inside.

894
00:47:11,078 --> 00:47:12,538
Thanks so much for coming by, Thore.

895
00:47:12,621 --> 00:47:13,705
I really appreciate it.

896
00:47:13,914 --> 00:47:18,418
Can you sort of share a bit
of what is going on in AlphaGo?

897
00:47:18,502 --> 00:47:21,797
So, AlphaGo has these three
main components.

898
00:47:21,880 --> 00:47:23,757
There's the policy network,

899
00:47:23,840 --> 00:47:28,720
which was trained on high-level games
to imitate those players.

900
00:47:29,221 --> 00:47:33,058
And then, we have a second component.
We call this the value net.

901
00:47:33,141 --> 00:47:35,686
And it can evaluate the board position

902
00:47:35,769 --> 00:47:39,773
and say what is the probability of winning
in this particular position.

903
00:47:40,107 --> 00:47:42,985
And the third component
is the tree search,

904
00:47:43,068 --> 00:47:46,780
where it would look through
different variations of the game

905
00:47:46,863 --> 00:47:49,449
and try to figure out
what will happen in the future.

906
00:47:52,744 --> 00:47:54,913
So if we now take
a position like this,

907
00:47:55,455 --> 00:47:58,875
first, the policy network
would scan the position

908
00:47:59,209 --> 00:48:03,839
and come up with what would be
the interesting spots to play,

909
00:48:04,339 --> 00:48:07,259
and it builds up a tree of variations.

910
00:48:07,342 --> 00:48:10,053
And then employs this value net

911
00:48:10,137 --> 00:48:14,975
that tells it how promising is the outcome
of this particular variation.

912
00:48:17,603 --> 00:48:22,274
So, AlphaGo tries to maximize
its probability of winning,

913
00:48:22,691 --> 00:48:26,445
but it doesn't care at all about
the margin by which it wins.

914
00:48:26,528 --> 00:48:28,655
Okay, so when you see
a slow-looking move,

915
00:48:28,739 --> 00:48:32,409
that's maybe an indication that AlphaGo
thinks it has a good chance to win.

916
00:48:32,492 --> 00:48:35,120
- Yeah, that is a little giveaway.
- Yeah.

917
00:48:35,203 --> 00:48:37,323
- A little tell. We're looking for a tell.
- Oh, yeah.

918
00:48:37,372 --> 00:48:39,333
Se-dol Lee is playing

919
00:48:39,416 --> 00:48:41,877
in a completely different style
from his usual style.

920
00:48:41,960 --> 00:48:47,049
This is a historic moment
and Se-dol Lee is the center of attention.

921
00:48:47,132 --> 00:48:49,468
He must be feeling immense pressure.

922
00:48:49,551 --> 00:48:54,264
I hope he gets over
this pressure and enjoys the match.

923
00:49:07,152 --> 00:49:08,904
Ooh, it looks like Lee

924
00:49:08,987 --> 00:49:10,347
is taking a little bit of a break.

925
00:49:28,632 --> 00:49:30,342
Se-dol Lee goes to smoke,

926
00:49:30,842 --> 00:49:32,427
and AlphaGo just plays.

927
00:49:32,511 --> 00:49:35,472
It does not think about
whether the opponent will be there or not.

928
00:49:35,889 --> 00:49:39,685
So, Aja sees AlphaGo plays move 37,

929
00:49:40,477 --> 00:49:42,604
and Aja puts the stone on the board.

930
00:49:44,648 --> 00:49:46,149
- Oh.
- Oh, wow.

931
00:49:48,193 --> 00:49:50,278
- Oh, it's totally an unthinkable move.
- Yes.

932
00:49:50,987 --> 00:49:52,114
The value...

933
00:49:53,990 --> 00:49:56,702
That's a very...
That's a very surprising move.

934
00:49:56,785 --> 00:49:59,996
I thought it was a mistake.

935
00:50:01,331 --> 00:50:04,710
When I saw this move,
for me, it was just a big shock.

936
00:50:04,793 --> 00:50:05,793
What?

937
00:50:06,628 --> 00:50:08,422
Normally, humans will never play this one

938
00:50:08,505 --> 00:50:10,048
because it's bad.

939
00:50:11,299 --> 00:50:13,019
It is just bad.
We don't know why, it's bad.

940
00:50:13,593 --> 00:50:15,220
It is a little bit high.

941
00:50:15,512 --> 00:50:16,388
Yeah?

942
00:50:16,471 --> 00:50:17,764
It's the fifth line.

943
00:50:17,848 --> 00:50:20,267
Normally, you don't make
a shoulder hit on the fifth line.

944
00:50:20,851 --> 00:50:24,146
So, coming on top of a fourth line zone
is really unusual.

945
00:50:24,646 --> 00:50:26,273
Yeah, that's an exciting move.

946
00:50:26,356 --> 00:50:29,067
I think we are seeing
an original move here.

947
00:50:29,735 --> 00:50:34,114
That is the kind of move
that you play Go for.

948
00:50:35,657 --> 00:50:36,657
Hey.

949
00:50:38,785 --> 00:50:40,203
Interesting stuff.

950
00:50:40,287 --> 00:50:42,831
- This fifth line shoulder hit was good.
- Yeah.

951
00:50:42,914 --> 00:50:44,249
And I wasn't expecting that.

952
00:50:45,250 --> 00:50:48,128
I don't really know if it's a good
or bad move at this point.

953
00:50:48,211 --> 00:50:51,673
The professional commentators
almost unanimously said

954
00:50:51,757 --> 00:50:54,926
that not a single human player
would have chosen move 37.

955
00:50:55,510 --> 00:50:57,554
So, I actually had a poke around
in AlphaGo

956
00:50:57,637 --> 00:50:59,389
to see what AlphaGo thought.

957
00:50:59,890 --> 00:51:02,684
And AlphaGo actually agreed
with that assessment.

958
00:51:03,226 --> 00:51:06,521
AlphaGo said there was
a 1-in-10,000 probability

959
00:51:06,605 --> 00:51:10,108
that move 37 would have been played
by a human player.

960
00:51:10,192 --> 00:51:12,611
So it knew that this was
an extremely unlikely move.

961
00:51:12,694 --> 00:51:15,739
It went beyond its human guide,

962
00:51:15,822 --> 00:51:19,034
and it came up with something new
and creative and different.

963
00:51:19,993 --> 00:51:24,748
I am very much watching the game
through these commentators.

964
00:51:24,831 --> 00:51:25,957
That is the way it works.

965
00:51:26,041 --> 00:51:28,418
So when they are confused,
I'm certainly confused.

966
00:51:28,502 --> 00:51:30,262
At the same time,
I'm latching on to the fact

967
00:51:30,337 --> 00:51:31,922
that they are confused, right?

968
00:51:32,005 --> 00:51:33,924
That is an interesting moment.

969
00:51:34,007 --> 00:51:36,676
When everyone else is confused,
who is not confused, right,

970
00:51:36,760 --> 00:51:37,761
besides the machine?

971
00:51:42,182 --> 00:51:45,769
I want to see Se-dol Lee
when he sees this move.

972
00:51:50,982 --> 00:51:52,651
He is back. Lee is back.

973
00:52:14,881 --> 00:52:16,383
I thought AlphaGo

974
00:52:16,466 --> 00:52:19,177
was based on probability calculation

975
00:52:19,761 --> 00:52:23,682
and that it was merely a machine.

976
00:52:24,349 --> 00:52:26,184
But when I saw this move,
I changed my mind.

977
00:52:27,352 --> 00:52:29,104
Surely, AlphaGo is creative.

978
00:52:30,480 --> 00:52:36,528
This move was really creative
and beautiful.

979
00:52:39,781 --> 00:52:41,116
Normally, he thinks

980
00:52:41,199 --> 00:52:43,243
for about one or two minutes
not more than that.

981
00:52:43,326 --> 00:52:46,371
But this time, he thought
for more than 12 minutes.

982
00:52:55,130 --> 00:52:59,050
The more I look at this move,
I feel something changed.

983
00:53:00,677 --> 00:53:03,471
Maybe, for humans, we think it is bad,

984
00:53:03,555 --> 00:53:05,682
but for AlphaGo, why not?

985
00:53:06,683 --> 00:53:08,184
Go is like geopolitics,

986
00:53:08,268 --> 00:53:10,061
like something small that happens here,

987
00:53:10,145 --> 00:53:11,396
it can have a ripple effect,

988
00:53:11,938 --> 00:53:15,483
you know, hours down the road
in a different part of the board.

989
00:53:15,859 --> 00:53:18,945
The game kind of turned on its axis
at that moment.

990
00:53:19,654 --> 00:53:21,448
This move was very special,

991
00:53:21,531 --> 00:53:22,824
because with this move,

992
00:53:23,199 --> 00:53:25,243
all the stones played before

993
00:53:25,327 --> 00:53:26,411
worked together.

994
00:53:26,494 --> 00:53:27,704
It was connected.

995
00:53:27,787 --> 00:53:30,123
It looked like a network,
linked everywhere.

996
00:53:30,582 --> 00:53:31,791
It was very special.

997
00:53:32,250 --> 00:53:33,335
Very special.

998
00:53:34,794 --> 00:53:36,504
This move made me think

999
00:53:36,588 --> 00:53:39,799
about Go in a new light.

1000
00:53:40,675 --> 00:53:44,763
What does creativity mean in Go?

1001
00:53:45,555 --> 00:53:47,766
It was a really meaningful move.

1002
00:53:50,769 --> 00:53:52,270
This is a tough game

1003
00:53:52,354 --> 00:53:53,229
for Se-dol Lee.

1004
00:53:53,313 --> 00:53:55,690
AlphaGo is just not letting Se-dol Lee

1005
00:53:55,774 --> 00:53:57,025
- do what he wants.
- Right.

1006
00:53:57,400 --> 00:54:00,070
Black has almost 60 points. That's a lot.

1007
00:54:00,612 --> 00:54:01,988
That's not a good sign.

1008
00:54:02,614 --> 00:54:04,032
Oh. Oh.

1009
00:54:04,115 --> 00:54:06,701
Se-dol Lee just slapped himself
on the side of the head.

1010
00:54:07,827 --> 00:54:08,954
Oh, wow.

1011
00:54:10,080 --> 00:54:11,665
I think black is ahead at this point.

1012
00:54:12,999 --> 00:54:15,210
It's looking good, isn't it?
We're on that steady...

1013
00:54:16,086 --> 00:54:17,420
steady path now.

1014
00:54:19,172 --> 00:54:21,132
I just saw Se-dol Lee...

1015
00:54:22,092 --> 00:54:23,635
lose so much.

1016
00:54:23,718 --> 00:54:26,846
Normally, we would have resigned
a long time ago, but he wanted to try.

1017
00:54:27,222 --> 00:54:28,682
He continued to play.

1018
00:54:28,765 --> 00:54:31,059
He just didn't want to resign.

1019
00:54:46,700 --> 00:54:48,576
Because then white...
Oh, he resigned.

1020
00:54:49,327 --> 00:54:52,038
It looks like Se-dol Lee
has just resigned.

1021
00:54:55,291 --> 00:54:57,419
It doesn't make any sense.
I didn't foresee that one.

1022
00:54:57,502 --> 00:54:59,879
- Move 37, very beautiful.
- Yes.

1023
00:54:59,963 --> 00:55:01,589
Beautiful, yeah. Beautiful.

1024
00:55:05,010 --> 00:55:06,052
Is he okay?

1025
00:55:07,345 --> 00:55:10,390
I brought his friend,
because yesterday I noticed

1026
00:55:10,473 --> 00:55:12,726
that he really wanted to analyze the game.

1027
00:55:12,809 --> 00:55:16,104
I didn't realize how bad it was.

1028
00:55:20,608 --> 00:55:22,402
There was this heavy sadness

1029
00:55:22,485 --> 00:55:24,362
over that whole floor,

1030
00:55:24,446 --> 00:55:25,989
and you could feel it during the game.

1031
00:55:26,072 --> 00:55:27,532
I felt it during the game.

1032
00:55:27,615 --> 00:55:31,453
And I'm leaving the commentary room
to go to the press conference,

1033
00:55:31,536 --> 00:55:32,954
and I was stopped by someone,

1034
00:55:33,038 --> 00:55:34,330
another technology reporter.

1035
00:55:34,414 --> 00:55:36,207
At first, all he wanted to talk about

1036
00:55:36,291 --> 00:55:38,084
was the technology and how great this was,

1037
00:55:38,168 --> 00:55:42,297
but then, even he kind of slipped
into this moment of melancholy

1038
00:55:42,380 --> 00:55:43,882
where he was upset as well.

1039
00:55:45,675 --> 00:55:51,139
Yesterday, I was surprised,
but today, I am quite speechless.

1040
00:55:51,848 --> 00:55:53,683
"I am quite speechless.

1041
00:55:53,767 --> 00:55:56,853
I admit that it was a very clear loss
on my part.

1042
00:55:57,270 --> 00:55:59,105
From the very beginning of the game,

1043
00:55:59,189 --> 00:56:02,901
there was not a moment in time
that I felt that I was leading the game."

1044
00:56:03,610 --> 00:56:07,155
You feel elated,
and you feel a little bit scared.

1045
00:56:07,238 --> 00:56:10,492
There is something, I think,
frightening to people

1046
00:56:10,575 --> 00:56:12,786
about a machine that learns on its own.

1047
00:56:13,369 --> 00:56:16,664
For us, AlphaGo is obviously
just some computer program.

1048
00:56:17,123 --> 00:56:19,667
But looking at the commentary
on the Internet,

1049
00:56:19,751 --> 00:56:22,045
I already saw the commentators
call AlphaGo

1050
00:56:22,128 --> 00:56:24,255
like "he" and "she" during the games.

1051
00:56:25,006 --> 00:56:26,591
Completely unconsciously.

1052
00:56:27,175 --> 00:56:30,512
Which... AlphaGo is really
a very, very simple program.

1053
00:56:30,595 --> 00:56:34,099
It's not anywhere close to full AI,
and we already see that happening.

1054
00:56:34,182 --> 00:56:35,558
So I find that very interesting.

1055
00:56:40,647 --> 00:56:43,149
The tendency
to anthropomorphize AI systems

1056
00:56:43,233 --> 00:56:45,401
is one of the big obstacles

1057
00:56:45,485 --> 00:56:47,403
in the way of actually trying
to understand

1058
00:56:47,487 --> 00:56:50,115
how AI might impact the world
in the future.

1059
00:56:50,198 --> 00:56:51,634
For example, the conversation is about

1060
00:56:51,658 --> 00:56:53,618
what could go wrong, what the risks are.

1061
00:56:53,993 --> 00:56:56,204
And invariably, you see
this Terminator picture.

1062
00:56:56,287 --> 00:56:58,957
Every single time,
there are these red glowing eyes, right?

1063
00:56:59,040 --> 00:57:01,543
We're really closer
to a smart washing machine

1064
00:57:01,626 --> 00:57:03,211
than Terminator.

1065
00:57:03,294 --> 00:57:06,422
If you look at today's AI,

1066
00:57:06,506 --> 00:57:09,050
we are really very nascent.

1067
00:57:09,134 --> 00:57:11,136
I'm extremely excited

1068
00:57:11,219 --> 00:57:14,472
and passionate about AI's potential.

1069
00:57:14,556 --> 00:57:17,892
But AI is still very limited in its power.

1070
00:57:18,393 --> 00:57:20,520
I think that people
are right to think

1071
00:57:21,146 --> 00:57:25,024
that there is a danger that as we continue
to improve these systems,

1072
00:57:25,108 --> 00:57:30,155
that we might miss that threshold
where we do cross over into danger.

1073
00:57:30,238 --> 00:57:31,239
But the good news is

1074
00:57:31,322 --> 00:57:33,467
that there are already people
thinking about those dangers.

1075
00:57:33,491 --> 00:57:36,303
You know, there's a lot of talk now,
and we are leading the discussion on this,

1076
00:57:36,327 --> 00:57:37,763
that maybe there should be a kind of,

1077
00:57:37,787 --> 00:57:40,540
cross-industry best practices
working group or something.

1078
00:57:40,623 --> 00:57:41,458
Right.

1079
00:57:41,541 --> 00:57:45,295
Where the leaders of the research teams
in those organizations,

1080
00:57:45,378 --> 00:57:48,548
you know, the big ones that are working
on AI, IBM, Microsoft, so on,

1081
00:57:48,631 --> 00:57:52,552
come together and make sure that AI
is used ethically and responsibly.

1082
00:57:53,511 --> 00:57:57,390
I think what is important
is that there is this community of people

1083
00:57:57,474 --> 00:58:01,644
who are leading the cutting edge of AI,

1084
00:58:01,728 --> 00:58:03,813
who are interacting with academics,

1085
00:58:03,897 --> 00:58:06,900
and already are thinking
about the long term,

1086
00:58:06,983 --> 00:58:10,653
and how we can ensure
that innovation is responsible

1087
00:58:10,737 --> 00:58:12,780
as the power of these machines
gets even greater.

1088
00:58:12,864 --> 00:58:15,241
Se-dol Lee Lost by Resignation

1089
00:58:15,325 --> 00:58:16,409
This is it, folks.

1090
00:58:17,118 --> 00:58:21,664
Day Three, Game Three.
Se-dol Lee, Go Master.

1091
00:58:22,207 --> 00:58:25,627
Back to the wall. He's down two-zero.

1092
00:58:25,710 --> 00:58:29,088
He's got to win today to keep hope alive.

1093
00:58:29,172 --> 00:58:32,592
Se-dol Lee had a day off

1094
00:58:32,675 --> 00:58:34,302
after losing two games,

1095
00:58:34,719 --> 00:58:39,474
and he gathered with Go professionals

1096
00:58:39,557 --> 00:58:43,186
and analyzed the game all night.

1097
00:58:43,269 --> 00:58:46,606
I heard that four pros, yeah,
went to visit him.

1098
00:58:46,689 --> 00:58:49,776
- Hm.
- To console him and to review AlphaGo's...

1099
00:58:49,859 --> 00:58:52,570
Console him?
Do you think he's upset, or...

1100
00:58:52,654 --> 00:58:54,822
I mean, Se-dol Lee
is upset. Yeah.

1101
00:58:55,323 --> 00:58:58,117
In the beginning,
there was a fierce fight,

1102
00:58:58,201 --> 00:59:00,453
and AlphaGo played very well.

1103
00:59:00,537 --> 00:59:02,580
So he secured a very early lead.

1104
00:59:02,997 --> 00:59:06,334
From move 50, the win rate
was very high already.

1105
00:59:06,417 --> 00:59:08,795
It was climbing toward 100%.

1106
00:59:09,128 --> 00:59:10,296
Oh, that's a pretty move.

1107
00:59:10,380 --> 00:59:12,465
What's your probability rating?

1108
00:59:12,549 --> 00:59:14,551
- Like, 91?
- No.

1109
00:59:14,634 --> 00:59:17,637
- It seems 70, you think?
- No.

1110
00:59:17,720 --> 00:59:19,931
He must want to win desperately right now.

1111
00:59:20,014 --> 00:59:21,683
THE CHALLENGE OF AI
0 VS. 2

1112
00:59:21,766 --> 00:59:24,060
But I think the pressure

1113
00:59:24,143 --> 00:59:26,688
and the psychological burden is adding up.

1114
00:59:26,771 --> 00:59:29,816
Today must be the worst.

1115
00:59:29,899 --> 00:59:31,609
I personally believe,

1116
00:59:31,693 --> 00:59:35,947
when you try too hard to win,
you will lose.

1117
00:59:36,656 --> 00:59:37,907
He tried to fight

1118
00:59:37,991 --> 00:59:39,200
directly in the game

1119
00:59:39,284 --> 00:59:40,493
but it's not his style.

1120
00:59:41,160 --> 00:59:44,581
When we change our style
to play against an opponent,

1121
00:59:44,664 --> 00:59:46,666
normally, it's very, very bad.

1122
00:59:49,877 --> 00:59:53,506
So it's an easier game for AlphaGo.

1123
00:59:54,632 --> 00:59:56,676
- It's looking good for us.
- Hopefully, I think...

1124
00:59:56,759 --> 00:59:59,721
You know, that black group is huge,
and it has got nowhere to go,

1125
00:59:59,804 --> 01:00:01,764
and it is going to be running around.

1126
01:00:11,190 --> 01:00:14,152
I don't know how
to describe the situation.

1127
01:00:14,235 --> 01:00:16,404
If I were black, I will resign.

1128
01:00:16,863 --> 01:00:21,618
We should admit that we are facing
the strongest existence ever,

1129
01:00:21,701 --> 01:00:23,244
ever in the Go history.

1130
01:00:23,703 --> 01:00:25,663
There's no point
in playing out the endgame,

1131
01:00:25,747 --> 01:00:28,166
and you're going to lose, right? So...

1132
01:00:31,127 --> 01:00:32,670
Even if black can live there...

1133
01:00:32,754 --> 01:00:34,964
- It's done. It's done.
- Oh, he resigned. Okay.

1134
01:00:35,048 --> 01:00:36,799
Wow. Wow.

1135
01:00:38,468 --> 01:00:41,095
You saw history made here tonight.

1136
01:00:41,429 --> 01:00:43,306
AlphaGo has won again.
Three straight wins.

1137
01:00:43,389 --> 01:00:45,308
Three straight wins.
Has won the match.

1138
01:00:46,768 --> 01:00:50,480
When Se-dol Lee resigned
from the game, he looked unhappy.

1139
01:00:50,563 --> 01:00:53,024
It was not just he lost the tournament,

1140
01:00:53,107 --> 01:00:55,401
it was especially about this game,

1141
01:00:55,485 --> 01:00:58,029
because he did not play his game.

1142
01:00:58,988 --> 01:01:02,241
I am very hurt about this, very hurt.

1143
01:01:03,743 --> 01:01:05,870
But, I can't do anything.

1144
01:01:08,956 --> 01:01:11,250
You know, suddenly, I feel
a bit ambivalent about it,

1145
01:01:11,334 --> 01:01:12,752
given I'm a games player,

1146
01:01:12,835 --> 01:01:15,546
and, you know, Go is the pinnacle
of board games.

1147
01:01:15,630 --> 01:01:16,815
But I really liked the statement

1148
01:01:16,839 --> 01:01:19,801
of one of the top Chinese professionals.
He said, you know,

1149
01:01:19,884 --> 01:01:25,139
"If AlphaGo wins, maybe we'll really start
to get to see what this game is about."

1150
01:01:26,808 --> 01:01:28,434
I couldn't celebrate.

1151
01:01:28,768 --> 01:01:31,270
It was fantastic that we had won,

1152
01:01:31,354 --> 01:01:33,231
but there was such a big part of me

1153
01:01:33,314 --> 01:01:37,527
that saw this man trying so hard

1154
01:01:37,610 --> 01:01:41,864
and being so disappointed.

1155
01:01:58,005 --> 01:02:01,217
I see on the Internet, many people
are talking about Se-dol Lee.

1156
01:02:01,300 --> 01:02:02,927
Maybe he didn't play his best.

1157
01:02:03,219 --> 01:02:05,388
You know, we are Go players.

1158
01:02:05,471 --> 01:02:08,182
Okay, sometimes in China,
in Korea, in Japan

1159
01:02:08,266 --> 01:02:09,892
we see Go like art.

1160
01:02:10,309 --> 01:02:11,686
We are artists, you know.

1161
01:02:11,769 --> 01:02:14,897
- We play our best for Go.
- Right. Right.

1162
01:02:14,981 --> 01:02:17,692
- So, please be gentle with Se-dol Lee.
- Right.

1163
01:02:17,775 --> 01:02:20,862
He is a very, very good player.
He is a great player.

1164
01:02:21,404 --> 01:02:24,073
I was in the room. I saw Se-dol Lee.
He wanted to win.

1165
01:02:24,157 --> 01:02:26,284
He tried everything.

1166
01:02:27,118 --> 01:02:30,496
It's just we can't. It's just... So...

1167
01:02:32,749 --> 01:02:35,752
I think I have to express
my apologies first.

1168
01:02:37,962 --> 01:02:40,923
If I had been able to play better
or smarter,

1169
01:02:41,007 --> 01:02:43,050
the results might have been different.

1170
01:02:43,134 --> 01:02:46,262
I think I disappointed too many of you
this time.

1171
01:02:46,345 --> 01:02:49,474
I want to apologize
for being so powerless.

1172
01:02:50,475 --> 01:02:52,477
I've never felt...

1173
01:02:52,560 --> 01:02:55,146
this much pressure,

1174
01:02:55,772 --> 01:02:58,065
this much weight.

1175
01:02:58,149 --> 01:03:02,195
I think I was too weak to overcome it.

1176
01:03:03,821 --> 01:03:07,074
I can't believe
this is happening.

1177
01:03:07,158 --> 01:03:10,036
Regardless of your opponent's level,

1178
01:03:10,119 --> 01:03:12,914
to be defeated not by three-zero,
but five-zero?

1179
01:03:13,623 --> 01:03:17,835
Losing to AlphaGo by five-zero

1180
01:03:17,919 --> 01:03:20,755
would really hurt my pride.

1181
01:03:20,838 --> 01:03:26,052
I also feel so bad for the people
who have supported me.

1182
01:03:26,761 --> 01:03:28,179
Se-dol Lee, 9 dan,

1183
01:03:28,262 --> 01:03:32,141
has the strongest heart of anyone I know.

1184
01:03:32,517 --> 01:03:38,773
He is fighting a lonely fight.

1185
01:03:38,856 --> 01:03:42,318
His opponent does not exist
in physical form.

1186
01:03:42,401 --> 01:03:44,779
I really feel for him.

1187
01:03:44,862 --> 01:03:47,114
We still have Games Four and Five,

1188
01:03:47,198 --> 01:03:53,162
and if Se-dol Lee, 9 dan,
plays like himself,

1189
01:03:53,246 --> 01:03:55,414
I believe we can beat the machine.

1190
01:04:06,926 --> 01:04:08,886
You could see he was more relaxed

1191
01:04:08,970 --> 01:04:11,055
after he had lost three games in a row.

1192
01:04:11,556 --> 01:04:13,474
That said, the stakes were still high.

1193
01:04:13,808 --> 01:04:16,811
You know, in the end,
it isn't about pride.

1194
01:04:18,229 --> 01:04:20,189
He did not feel confident,

1195
01:04:20,606 --> 01:04:22,024
but he felt light.

1196
01:04:22,900 --> 01:04:26,779
Can Se-dol Lee
find AlphaGo's weakness?

1197
01:04:26,863 --> 01:04:28,823
Is there, in fact, a weakness?

1198
01:04:28,906 --> 01:04:30,586
I was thinking I would just pull the plug.

1199
01:04:30,616 --> 01:04:32,052
- You would just pull the plug?
- Yeah.

1200
01:04:32,076 --> 01:04:34,412
Everyone is still
cheering for Se-dol Lee.

1201
01:04:34,495 --> 01:04:35,621
And...

1202
01:04:36,497 --> 01:04:38,207
Yeah, it's not going very well.

1203
01:04:38,708 --> 01:04:40,293
Se-dol Lee has chosen

1204
01:04:40,376 --> 01:04:42,003
to play conservatively.

1205
01:04:42,086 --> 01:04:45,172
But I am not sure
it will benefit him.

1206
01:04:45,256 --> 01:04:46,256
I agree.

1207
01:04:46,299 --> 01:04:48,459
Yes, it doesn't look like
a very good strategy for him.

1208
01:04:49,010 --> 01:04:50,136
It feels pretty good

1209
01:04:50,219 --> 01:04:51,929
for black at this point.

1210
01:04:52,013 --> 01:04:54,616
- It feels pretty good for black, yes.
- It's pretty good for black. Yeah.

1211
01:04:54,640 --> 01:04:56,976
I want Se-dol Lee to play his game.

1212
01:04:57,518 --> 01:05:00,187
Because, at that moment,
he tried various things

1213
01:05:00,271 --> 01:05:02,940
to play with AlphaGo,
to understand AlphaGo,

1214
01:05:03,024 --> 01:05:05,234
but he never tried to play himself.

1215
01:05:06,319 --> 01:05:07,546
I've said this many, many times,

1216
01:05:07,570 --> 01:05:10,448
AlphaGo looks like a real mirror.

1217
01:05:10,990 --> 01:05:12,617
When you play with AlphaGo,

1218
01:05:13,200 --> 01:05:15,369
you feel very strange.

1219
01:05:15,745 --> 01:05:18,456
It feels like you are naked all the time.

1220
01:05:19,081 --> 01:05:21,560
The first time you see this,
you wouldn't want to see because...

1221
01:05:21,584 --> 01:05:24,086
"Oh, is this me? The real me?"

1222
01:05:24,921 --> 01:05:27,298
And the more you see,
you learn to accept it.

1223
01:05:27,673 --> 01:05:29,175
"Oh, this is the real me.

1224
01:05:29,258 --> 01:05:32,053
So, how... Now, what shall I do?"

1225
01:05:36,682 --> 01:05:39,226
It's developing into a very,
very dangerous fight.

1226
01:05:40,019 --> 01:05:44,190
This is really Se-dol Lee's type of game.
He likes this kind of fight.

1227
01:05:44,273 --> 01:05:46,525
White has to find something
inside black's territory.

1228
01:05:46,609 --> 01:05:49,862
I think that he
is already planning on trying something.

1229
01:05:49,946 --> 01:05:52,782
So you believe in Se-dol Lee's ability

1230
01:05:52,865 --> 01:05:55,117
to live in that very small area, right?

1231
01:05:55,201 --> 01:05:56,243
Yes. Yes.

1232
01:05:56,327 --> 01:05:58,204
There is a little potential there.

1233
01:05:58,621 --> 01:06:00,873
And I think, maybe he is going to try
to do something.

1234
01:06:00,957 --> 01:06:02,541
- Se-dol Lee magic!
- Nice.

1235
01:06:02,959 --> 01:06:05,711
Oh, what would be
the magic move?

1236
01:06:05,795 --> 01:06:07,588
Se-dol Lee is running short on time,

1237
01:06:07,672 --> 01:06:09,590
but he's going to have to use up
all his time.

1238
01:06:09,674 --> 01:06:12,426
Yeah, he has just burned, like,
seven or eight minutes,

1239
01:06:12,510 --> 01:06:13,886
just on this move already.

1240
01:06:14,512 --> 01:06:15,930
I don't see anything.

1241
01:06:16,973 --> 01:06:18,849
We can't find anything.

1242
01:06:18,933 --> 01:06:20,226
Show us something.

1243
01:06:20,685 --> 01:06:22,144
I don't see any possible move.

1244
01:06:22,228 --> 01:06:24,063
What is Se-dol Lee up to here?

1245
01:06:24,146 --> 01:06:25,940
Yeah, he is really concentrating.

1246
01:06:26,023 --> 01:06:27,483
He really is. Look at that.

1247
01:06:28,734 --> 01:06:30,236
Se-dol Lee is very patient.

1248
01:06:30,319 --> 01:06:32,863
He waits. He waits for his moment.

1249
01:06:33,656 --> 01:06:36,158
I feel, sometimes, he is like a wolf,

1250
01:06:36,659 --> 01:06:40,121
waiting in the forest, in the winter.

1251
01:06:41,414 --> 01:06:43,541
He is cold. He's very, very cold.

1252
01:06:44,333 --> 01:06:45,835
But he needs patience.

1253
01:06:45,918 --> 01:06:48,045
But, when the moment comes,

1254
01:06:48,129 --> 01:06:50,464
he goes out to attack.

1255
01:06:51,841 --> 01:06:54,010
This is the hinge of the game.

1256
01:06:54,093 --> 01:06:55,093
Oh.

1257
01:06:55,177 --> 01:06:56,929
Look at that move.
That's an exciting move.

1258
01:06:57,096 --> 01:06:59,640
- Basically, I believe that this time...
- He found a wedge.

1259
01:06:59,724 --> 01:07:00,725
Whoa.

1260
01:07:00,808 --> 01:07:03,686
It's going to change
the equation

1261
01:07:03,769 --> 01:07:05,438
because now black cannot escape.

1262
01:07:05,521 --> 01:07:07,189
That would be so cool if that works.

1263
01:07:11,193 --> 01:07:12,193
Oh!

1264
01:07:14,447 --> 01:07:17,783
AlphaGo has just played
something maybe unusual.

1265
01:07:18,242 --> 01:07:21,662
You know, I'm not actually sure
what AlphaGo is trying to do here.

1266
01:07:21,746 --> 01:07:22,788
What's that about?

1267
01:07:22,872 --> 01:07:24,498
I don't really understand it.

1268
01:07:24,582 --> 01:07:25,458
Well, well.

1269
01:07:25,541 --> 01:07:28,294
That was a sharp drop in win rate.

1270
01:07:28,377 --> 01:07:30,880
That is the sharpest drop
in win rate we have seen.

1271
01:07:31,380 --> 01:07:33,799
She dropped at eight percent.

1272
01:07:33,883 --> 01:07:35,092
- Wow.
- Wow.

1273
01:07:36,052 --> 01:07:39,180
This could be that it actually
can't find a way through.

1274
01:07:39,263 --> 01:07:40,365
- I think this is...
- Uh-huh.

1275
01:07:40,389 --> 01:07:43,100
It has looked far enough ahead
to see that it doesn't work,

1276
01:07:43,184 --> 01:07:44,935
and now maybe it's on tilt. I don't know.

1277
01:07:45,728 --> 01:07:47,146
There it comes.

1278
01:07:47,229 --> 01:07:49,190
- That was one time.
- There it comes, though.

1279
01:07:49,273 --> 01:07:51,734
It looks like it has fallen off the cliff.

1280
01:07:52,568 --> 01:07:54,695
- Yeah, it has made a mistake.
- Yeah.

1281
01:07:54,779 --> 01:07:56,572
Did anything strange happen in the...

1282
01:07:56,989 --> 01:07:58,449
- No, it all looked normal.
- Yeah.

1283
01:07:58,532 --> 01:08:01,077
Well, we can definitely say
there is a weakness.

1284
01:08:01,702 --> 01:08:03,382
Well, we definitely say there's a mistake.

1285
01:08:04,789 --> 01:08:09,168
I felt this mixture
of this sinking feeling in my stomach,

1286
01:08:09,251 --> 01:08:12,588
where I was wondering if AlphaGo
was becoming delusional in this situation.

1287
01:08:12,671 --> 01:08:14,965
Where I could see that it was starting
to play strangely,

1288
01:08:15,341 --> 01:08:17,259
and at the same time,
relief, that Se-dol Lee,

1289
01:08:17,343 --> 01:08:19,970
that he was actually in with a chance now.

1290
01:08:21,597 --> 01:08:24,350
Aja is like trying not to look horrified.

1291
01:08:26,143 --> 01:08:30,439
I knew after move 78,
after, like, 10 or 20 moves,

1292
01:08:30,815 --> 01:08:33,859
I saw AlphaGo's strange moves.

1293
01:08:33,943 --> 01:08:38,405
And, I knew AlphaGo somehow
became crazy, but I didn't realize why.

1294
01:08:38,489 --> 01:08:40,157
What?

1295
01:08:40,241 --> 01:08:43,202
We searched to 95 ahead at that point?

1296
01:08:43,285 --> 01:08:44,805
At the part where it made the mistake?

1297
01:08:45,246 --> 01:08:47,206
- Yes.
- I think that something went wrong.

1298
01:08:47,289 --> 01:08:49,559
That's the longest it has searched
the entire game, isn't it?

1299
01:08:49,583 --> 01:08:50,417
Yeah.

1300
01:08:50,501 --> 01:08:52,169
I think it's like it searched so deeply,

1301
01:08:52,253 --> 01:08:54,213
- it has lost itself.
- It's tired.

1302
01:08:55,256 --> 01:08:56,465
I do get the impression

1303
01:08:56,549 --> 01:08:58,884
that AlphaGo has sort of,
gone off on a tangent.

1304
01:08:58,968 --> 01:09:02,012
- What is it doing now?
- Well, maybe it has a master plan.

1305
01:09:02,096 --> 01:09:04,557
No, it doesn't even think it has, does it?

1306
01:09:04,640 --> 01:09:08,269
So it knows it has made a mistake,
and it starts evaluating it the other way.

1307
01:09:08,769 --> 01:09:11,147
Look, look, Lee is confused.
He's like, "What is he doing?"

1308
01:09:11,230 --> 01:09:14,108
That's not a "I'm scared" confused,
that's a "What is it doing?"

1309
01:09:14,191 --> 01:09:15,442
What's going on?

1310
01:09:15,818 --> 01:09:17,319
You know, you asked if it was a bug.

1311
01:09:17,820 --> 01:09:20,656
I've said before that if...

1312
01:09:20,739 --> 01:09:23,492
- Yeah, right. That's what I thought.
- If DeepMind has figured out

1313
01:09:23,576 --> 01:09:26,162
how to write code that doesn't have bugs,

1314
01:09:26,245 --> 01:09:28,539
that is a bigger news story than AlphaGo.

1315
01:09:28,622 --> 01:09:30,457
- Are you kidding me?
- Hm?

1316
01:09:30,541 --> 01:09:32,710
This... Literally, this next move
we are going to play,

1317
01:09:34,837 --> 01:09:37,715
I think they are going to laugh.
I think Lee is going to laugh.

1318
01:09:46,182 --> 01:09:47,516
Oh! Oh!

1319
01:09:47,600 --> 01:09:48,601
What is this?

1320
01:09:49,977 --> 01:09:51,812
- Oh, that's ridiculous!
- What's going on?

1321
01:09:51,896 --> 01:09:53,272
What is it thinking?

1322
01:09:53,606 --> 01:09:55,858
I don't really know what AlphaGo
is trying to do here.

1323
01:09:57,151 --> 01:09:58,819
That's the understatement of the year.

1324
01:09:59,320 --> 01:10:00,571
Is it a mouse misclick

1325
01:10:00,654 --> 01:10:02,364
- from Aja Huang?
- Nope, that's the move.

1326
01:10:02,448 --> 01:10:03,908
Aja Huang makes no misclicks.

1327
01:10:04,909 --> 01:10:06,535
Se-dol Lee is very confused.

1328
01:10:06,994 --> 01:10:08,621
These are not human moves.

1329
01:10:08,704 --> 01:10:12,208
This move is also sort of inexplicable.

1330
01:10:12,291 --> 01:10:15,628
I mean, those, you can clearly
call those mistakes.

1331
01:10:15,711 --> 01:10:16,629
Yes, of course.

1332
01:10:16,712 --> 01:10:18,881
But it's the first time
in the four matches

1333
01:10:18,964 --> 01:10:20,724
- that we have seen moves like that.
- Right.

1334
01:10:20,799 --> 01:10:22,593
Oh, the value dropped even more.

1335
01:10:23,385 --> 01:10:24,386
That's weird.

1336
01:10:25,221 --> 01:10:26,680
They're like 45%.

1337
01:10:28,724 --> 01:10:29,724
White.

1338
01:10:29,975 --> 01:10:32,228
- White.
- I think white is winning.

1339
01:10:32,311 --> 01:10:34,063
Come on, Se-dol Lee.

1340
01:10:34,146 --> 01:10:37,024
- It's unbelievable.
- Yeah, all right!

1341
01:10:37,107 --> 01:10:41,695
I was so confident that this black area
would just be consolidated by black,

1342
01:10:41,779 --> 01:10:42,905
and there was nothing there.

1343
01:10:42,988 --> 01:10:46,283
And somehow, he has just erased it all.
Like, it's gone.

1344
01:10:46,367 --> 01:10:47,743
So, he has found his weakness.

1345
01:10:47,826 --> 01:10:50,329
That wedge move probably surprised it.

1346
01:10:50,412 --> 01:10:51,747
- The wedge, yeah.
- Yeah.

1347
01:10:51,830 --> 01:10:54,375
AlphaGo seemed to have
such a good game, and then...

1348
01:10:54,458 --> 01:10:57,878
this whole sequence in the center
sort of changed that.

1349
01:10:57,962 --> 01:10:59,797
He pulls off a miracle.

1350
01:11:00,130 --> 01:11:01,757
He managed to make it so complicated

1351
01:11:01,840 --> 01:11:05,344
that the artificial intelligence
doesn't evaluate it correctly.

1352
01:11:05,427 --> 01:11:08,055
I think it looks like Aja Huang
maybe, sitting there,

1353
01:11:08,138 --> 01:11:10,015
also knows that the game is over.

1354
01:11:10,432 --> 01:11:13,060
I can't... I can't stop smiling.

1355
01:11:13,602 --> 01:11:14,895
Did he smile?

1356
01:11:14,979 --> 01:11:16,647
No, I haven't seen him smile yet.

1357
01:11:17,106 --> 01:11:19,191
- I think he is still checking.
- He's so serious.

1358
01:11:19,275 --> 01:11:20,526
He's so serious.

1359
01:11:20,609 --> 01:11:22,861
- He wants to be careful.
- Yeah.

1360
01:11:22,945 --> 01:11:25,447
He's concerned a lot.

1361
01:11:25,531 --> 01:11:26,991
He really does his best.

1362
01:11:27,783 --> 01:11:30,244
I think he feels like he has a sense of...

1363
01:11:30,953 --> 01:11:34,290
uh, responsibility, or burden.

1364
01:11:34,873 --> 01:11:36,542
I think it's almost over.

1365
01:11:37,209 --> 01:11:38,419
So, what's the percentage?

1366
01:11:38,502 --> 01:11:40,713
It shows as 18.2%.

1367
01:11:40,796 --> 01:11:42,715
The next move should be resign.

1368
01:11:46,885 --> 01:11:49,471
There you go. AlphaGo resigned.

1369
01:11:51,974 --> 01:11:54,226
- It looks like AlphaGo has resigned.
- Wow.

1370
01:11:55,352 --> 01:11:57,271
The most amazing game.

1371
01:11:58,981 --> 01:12:03,068
I'm almost going to tear up,
was, uh, Game Four...

1372
01:12:03,152 --> 01:12:05,654
um, where he comes back and wins, right?

1373
01:12:05,738 --> 01:12:07,072
It resigned!

1374
01:12:09,867 --> 01:12:11,452
I'm so happy!

1375
01:12:17,583 --> 01:12:19,918
I heard people shouting in joy

1376
01:12:20,002 --> 01:12:22,546
when it was clear that AlphaGo
had lost the game.

1377
01:12:23,047 --> 01:12:24,715
I think it is clear why.

1378
01:12:26,258 --> 01:12:28,427
People felt helplessness and fear.

1379
01:12:29,219 --> 01:12:32,222
It seemed like we humans
are so weak and fragile.

1380
01:12:33,265 --> 01:12:35,976
And this victory meant...

1381
01:12:36,810 --> 01:12:40,898
we could still hold our own.

1382
01:12:40,981 --> 01:12:43,901
As time goes on, it will probably be
very difficult to beat AI.

1383
01:12:44,610 --> 01:12:50,282
But winning this one time,
it felt like it was enough.

1384
01:12:51,367 --> 01:12:52,743
One time was enough.

1385
01:12:53,369 --> 01:12:55,746
I heard from so many people saying

1386
01:12:55,829 --> 01:12:58,374
they were running out in the street.

1387
01:12:58,457 --> 01:13:00,000
They were so happy.

1388
01:13:00,084 --> 01:13:02,211
They were chanting, they were celebrating.

1389
01:13:02,294 --> 01:13:05,172
Especially after zero-three down,

1390
01:13:05,547 --> 01:13:10,469
at the time, it seemed to be hopeless,
that the end of the world was coming,

1391
01:13:10,552 --> 01:13:11,720
but we saw the light.

1392
01:13:11,970 --> 01:13:13,514
Usually...

1393
01:13:13,597 --> 01:13:17,643
I'm happy when I win,
not when my colleague wins.

1394
01:13:18,227 --> 01:13:21,897
But, this time, it felt like my win.

1395
01:13:21,980 --> 01:13:24,400
I didn't want to make a move
which AlphaGo could predict.

1396
01:13:24,483 --> 01:13:25,526
This part...

1397
01:13:26,485 --> 01:13:29,196
It didn't anticipate this wedge move.

1398
01:13:30,155 --> 01:13:31,698
I won one game at least.

1399
01:13:32,741 --> 01:13:33,826
Good job today.

1400
01:14:02,271 --> 01:14:04,690
But I didn't expect it

1401
01:14:04,773 --> 01:14:06,150
to be like this.

1402
01:14:06,692 --> 01:14:08,819
I couldn't believe I won one game.

1403
01:14:08,902 --> 01:14:11,321
It was unbelievable.

1404
01:14:11,905 --> 01:14:12,990
Thank you very much.

1405
01:14:14,324 --> 01:14:17,995
I have never been congratulated so much
for winning one game.

1406
01:14:18,745 --> 01:14:22,499
After losing three games in a row,
I couldn't be happier.

1407
01:14:25,377 --> 01:14:30,090
This victory is so valuable
that I wouldn't exchange it

1408
01:14:30,174 --> 01:14:32,885
for anything in the world.

1409
01:14:33,343 --> 01:14:36,597
My question is about move number 78.

1410
01:14:36,680 --> 01:14:39,933
Chinese top Go player
Gu Li said that it was a god's play.

1411
01:14:40,017 --> 01:14:42,519
What were you thinking
when you made that play? Thank you.

1412
01:14:42,895 --> 01:14:45,189
At that point in the game,

1413
01:14:45,272 --> 01:14:48,150
move 78 was the only move I could see.

1414
01:14:48,233 --> 01:14:50,611
There was no other placement.

1415
01:14:50,694 --> 01:14:53,614
It was the only option for me,
so I put it there.

1416
01:14:53,697 --> 01:14:57,367
I am quite humbled by all the praise
I am getting for it.

1417
01:14:58,243 --> 01:15:01,663
It does leave me a little bit
in awe of the human brain's power,

1418
01:15:01,747 --> 01:15:05,584
in particular, Lee's amazing ability
to cause AlphaGo problems

1419
01:15:05,667 --> 01:15:08,337
and find something
seemingly out of nothing.

1420
01:15:08,420 --> 01:15:10,589
And so, we really want
to understand what had happened.

1421
01:15:11,298 --> 01:15:13,091
So, it really was Lee's move?

1422
01:15:13,175 --> 01:15:14,176
The key is the center.

1423
01:15:14,259 --> 01:15:15,945
After the center got out,
there was no chance.

1424
01:15:15,969 --> 01:15:16,887
Come, come.

1425
01:15:16,970 --> 01:15:18,388
We were winning before this.

1426
01:15:19,097 --> 01:15:20,682
But if AlphaGo thought it was winning

1427
01:15:20,766 --> 01:15:24,019
and it couldn't convert that win,
that means it wasn't actually winning.

1428
01:15:24,102 --> 01:15:25,955
Otherwise, it would have had
the right responses.

1429
01:15:25,979 --> 01:15:27,105
Would we have played it?

1430
01:15:27,189 --> 01:15:28,357
Nerves of steel in there.

1431
01:15:28,440 --> 01:15:30,192
What probability does it give it as white?

1432
01:15:33,445 --> 01:15:36,532
It's 0.007% of...

1433
01:15:36,615 --> 01:15:38,283
Is that in the position that he played?

1434
01:15:38,367 --> 01:15:39,636
- That's when he played it.
- I see.

1435
01:15:39,660 --> 01:15:41,971
- We thought this was a 1-in-10,000 move.
- I know it was like...

1436
01:15:41,995 --> 01:15:43,635
- Yeah.
- The values we didn't know about.

1437
01:15:45,499 --> 01:15:48,585
- So the god move was literally a god move,
- Yeah.

1438
01:15:48,669 --> 01:15:51,380
because we believe
that only one in 10,000 humans

1439
01:15:51,463 --> 01:15:53,066
- would have found that move.
- That's right.

1440
01:15:53,090 --> 01:15:55,384
It doesn't come to mind as the top five...

1441
01:15:55,467 --> 01:15:57,803
in the top five moves
that you would consider.

1442
01:15:57,886 --> 01:15:59,906
Unless you're Se-dol Lee,
he said it was the only move.

1443
01:15:59,930 --> 01:16:02,891
He thought it was the only move.
Yeah, remarkable, isn't it?

1444
01:16:06,645 --> 01:16:08,814
Welcome back to the Four Seasons.

1445
01:16:08,897 --> 01:16:10,732
World Champion Se-dol Lee

1446
01:16:10,816 --> 01:16:16,113
went looking for AlphaGo's weakness
in Game Four, and he found it.

1447
01:16:16,196 --> 01:16:20,742
Today, last round, he is looking to see
if he can repeat that.

1448
01:16:20,826 --> 01:16:21,994
We will see what happens.

1449
01:16:22,077 --> 01:16:23,704
The place
is a madhouse downstairs.

1450
01:16:23,787 --> 01:16:26,582
Maybe as many, if not more,
journalists here today

1451
01:16:26,665 --> 01:16:28,584
than there were for Game One.

1452
01:16:29,793 --> 01:16:32,254
We were really excited
for the fifth match,

1453
01:16:32,337 --> 01:16:33,505
to see what would happen.

1454
01:16:33,589 --> 01:16:36,842
Was this going to be a three-two thing
or was it going to be a four-one thing?

1455
01:16:36,925 --> 01:16:39,803
And, you know, there's a different message
with both of those.

1456
01:16:41,013 --> 01:16:44,433
For me, I don't think
AlphaGo will make the mistake again.

1457
01:16:45,350 --> 01:16:46,476
But who knows.

1458
01:16:47,311 --> 01:16:48,437
Maybe it will happen again.

1459
01:16:49,146 --> 01:16:53,066
And Se-dol Lee now have more confidence
about this.

1460
01:16:53,150 --> 01:16:58,739
It looks like, maybe, Se-dol Lee has
found the key in solving AlphaGo.

1461
01:16:59,364 --> 01:17:00,407
Oh, this works!

1462
01:17:03,035 --> 01:17:06,246
Is it fair to say
that AlphaGo made a mistake?

1463
01:17:06,330 --> 01:17:07,770
We might have another victory today.

1464
01:17:07,831 --> 01:17:11,084
You weren't crazy
about the timing on this move.

1465
01:17:11,168 --> 01:17:14,171
- Yeah.
- And now, you don't approve of this.

1466
01:17:14,838 --> 01:17:17,299
Yeah, I'm sort of thinking
that maybe AlphaGo

1467
01:17:17,382 --> 01:17:19,509
hasn't recovered from Game Four yet. Yeah.

1468
01:17:20,052 --> 01:17:21,136
Still deluded?

1469
01:17:21,845 --> 01:17:22,971
We don't know.

1470
01:17:23,513 --> 01:17:24,640
Oh, no, I hope he doesn't...

1471
01:17:24,723 --> 01:17:26,892
No, no, he needs to play the...

1472
01:17:29,436 --> 01:17:31,688
He played some moves.

1473
01:17:31,772 --> 01:17:32,773
It was a bad move.

1474
01:17:32,856 --> 01:17:33,857
I felt something.

1475
01:17:33,940 --> 01:17:36,777
Oh, maybe his weakness
has come back again.

1476
01:17:36,902 --> 01:17:39,863
Are we seeing
another short circuit?

1477
01:17:39,946 --> 01:17:42,074
Or is there something, what's...

1478
01:17:42,157 --> 01:17:44,117
I think it could be
a kind of a misreading.

1479
01:17:44,910 --> 01:17:46,244
And we are...

1480
01:17:46,328 --> 01:17:48,088
You're pretty comfortable
with saying that...

1481
01:17:48,121 --> 01:17:49,831
That it is looking good
for Se-dol Lee?

1482
01:17:49,915 --> 01:17:51,416
It's looking good for Se-dol Lee.

1483
01:17:54,044 --> 01:17:55,420
What is he playing in there for?

1484
01:17:55,504 --> 01:17:58,006
There's no reason for white
to be playing that move.

1485
01:17:58,090 --> 01:17:59,466
It's a bad move.

1486
01:17:59,549 --> 01:18:02,386
And in some cases,
it's going to lose a point, too.

1487
01:18:02,469 --> 01:18:04,596
It is exactly the same thing.

1488
01:18:04,680 --> 01:18:06,139
It's 91% certain now.

1489
01:18:06,223 --> 01:18:08,225
Someone was telling me
that maybe it had...

1490
01:18:08,308 --> 01:18:10,268
Yeah, because it's incorrect again.

1491
01:18:11,978 --> 01:18:13,647
The whole game,
we thought that AlphaGo

1492
01:18:13,730 --> 01:18:15,130
was wrong about the board position.

1493
01:18:15,190 --> 01:18:18,193
We were super worried that,
"Oh, it's going to play garbage.

1494
01:18:18,276 --> 01:18:20,445
It's going to be like lose
in a very embarrassing way."

1495
01:18:20,529 --> 01:18:22,489
And this continued for the whole game.

1496
01:18:23,323 --> 01:18:24,324
This is a bit weird.

1497
01:18:25,158 --> 01:18:26,576
- When did it play that?
- Just now.

1498
01:18:26,660 --> 01:18:29,097
- So I don't know. I mean, it doesn't...
- Well, that we don't know.

1499
01:18:29,121 --> 01:18:31,182
But we're not good enough
Go players to know that, right?

1500
01:18:31,206 --> 01:18:32,040
Yeah.

1501
01:18:32,124 --> 01:18:34,626
As it turns out,
none of us know Go well enough

1502
01:18:34,710 --> 01:18:36,920
to accurately judge what AlphaGo is doing.

1503
01:18:37,003 --> 01:18:38,380
The white is winning now.

1504
01:18:38,463 --> 01:18:39,943
This one looks like white is winning.

1505
01:18:41,133 --> 01:18:42,008
Ah!

1506
01:18:42,092 --> 01:18:46,888
We all say some of AlphaGo's moves
are so weird and strange,

1507
01:18:46,972 --> 01:18:48,390
and may be mistakes.

1508
01:18:48,682 --> 01:18:51,977
But after a game is finished,

1509
01:18:52,060 --> 01:18:56,148
we have to doubt ourselves, our judgment.

1510
01:18:56,231 --> 01:18:59,401
AlphaGo making another
kind of nonsensical throw-in.

1511
01:18:59,985 --> 01:19:01,379
I'm not really sure what that's about.

1512
01:19:01,403 --> 01:19:03,947
This is what 10
or maybe 11-dan play looks like.

1513
01:19:04,030 --> 01:19:06,700
It looks weird,
and we don't quite understand it.

1514
01:19:06,783 --> 01:19:10,245
I think it's important to study more

1515
01:19:10,328 --> 01:19:13,832
about AlphaGo's mistake-like moves.

1516
01:19:13,915 --> 01:19:17,878
Then maybe we can adjust
our knowledge about Go.

1517
01:19:17,961 --> 01:19:21,673
To me, the most amazing thing
to come out of my understanding of Go

1518
01:19:21,757 --> 01:19:23,383
as a result of watching AlphaGo play

1519
01:19:23,467 --> 01:19:26,052
are the infamous "slack moves."

1520
01:19:26,136 --> 01:19:27,471
Well, there's something strange

1521
01:19:27,554 --> 01:19:29,097
- about the way it is playing,
- Yeah.

1522
01:19:29,181 --> 01:19:32,058
because it's playing some moves
that are not really necessary.

1523
01:19:32,142 --> 01:19:32,976
Right.

1524
01:19:33,059 --> 01:19:36,062
A "slack move" is a move
that looks lazy.

1525
01:19:36,146 --> 01:19:39,900
You can see these other better moves,
and AlphaGo is rejecting them.

1526
01:19:39,983 --> 01:19:43,320
But, what I think AlphaGo is teaching us

1527
01:19:43,403 --> 01:19:49,534
is that we have been using score
as a proxy for chance of winning.

1528
01:19:49,618 --> 01:19:52,621
So the bigger my margin of territory,

1529
01:19:52,704 --> 01:19:54,706
the more confident I am
that I'm going to win.

1530
01:19:54,790 --> 01:19:56,416
And AlphaGo is saying, "No, no, no.

1531
01:19:56,500 --> 01:19:58,251
It shouldn't matter how much you win by,

1532
01:19:58,335 --> 01:20:00,462
you only need to win by a single point.

1533
01:20:00,545 --> 01:20:03,465
Why should I be seizing all this
extra territory when I don't need it?"

1534
01:20:03,548 --> 01:20:05,926
The lessons that AlphaGo is teaching us

1535
01:20:06,009 --> 01:20:11,598
are going to influence how Go is played
for the next thousand years.

1536
01:20:12,098 --> 01:20:14,518
By my counting, white won this game.

1537
01:20:14,601 --> 01:20:17,562
How... By how much?
Are we talking two points?

1538
01:20:17,646 --> 01:20:19,564
- One-and-a-half.
- One-and-a-half points.

1539
01:20:19,981 --> 01:20:22,192
One-and-a-half. Maybe half,
one-and-a-half.

1540
01:20:22,275 --> 01:20:25,737
Of course, Go is just a game,
but we can learn important lessons

1541
01:20:25,821 --> 01:20:28,573
from a computer being so successful at Go.

1542
01:20:28,657 --> 01:20:31,910
Machines will have the capability
not only to crunch

1543
01:20:31,993 --> 01:20:36,331
through a huge amount of data,
but also to analyze it intelligently.

1544
01:20:36,665 --> 01:20:39,251
Just as in the case of the Go games,

1545
01:20:39,334 --> 01:20:42,712
the machine made moves
that surprised even the experts.

1546
01:20:42,796 --> 01:20:45,465
And, eventually, the machines
will gain our confidence

1547
01:20:45,549 --> 01:20:47,759
because we will see that very, very often

1548
01:20:47,843 --> 01:20:50,846
they make a better guess
than we could have made as humans.

1549
01:20:53,849 --> 01:20:55,684
AlphaGo has strong foundation.

1550
01:21:01,648 --> 01:21:03,149
What surprised me the most was...

1551
01:21:03,608 --> 01:21:05,360
that AlphaGo showed us

1552
01:21:05,443 --> 01:21:09,072
that moves humans may have thought
are creative, were actually conventional.

1553
01:21:09,781 --> 01:21:14,828
I think this will bring
a new paradigm to Go.

1554
01:21:14,911 --> 01:21:18,248
AlphaGo won this area as well? Oh, god.

1555
01:21:18,331 --> 01:21:22,544
- That makes me very uncomfortable.
- Oh!

1556
01:21:22,627 --> 01:21:25,088
Unfortunately, for Se-dol Lee,

1557
01:21:25,171 --> 01:21:27,424
I think white might have
a slight advantage here.

1558
01:21:27,507 --> 01:21:30,176
White lights up in places,
but if we count it like that, it's 1.5.

1559
01:21:30,260 --> 01:21:31,887
They keep saying
one-and-a-half points.

1560
01:21:31,970 --> 01:21:33,805
How do you know the Korean room
has called it?

1561
01:21:33,889 --> 01:21:35,432
Someone, um...

1562
01:21:36,016 --> 01:21:38,101
- Who told us the Koreans called it?
- I don't know.

1563
01:21:38,184 --> 01:21:39,936
The Koreans told you?

1564
01:21:40,020 --> 01:21:41,020
Koreans called it.

1565
01:21:41,062 --> 01:21:44,024
The Korean room
has called it for AlphaGo.

1566
01:21:44,107 --> 01:21:47,360
We would just like to know
if someone Korean in there can find out.

1567
01:21:49,070 --> 01:21:50,906
The Korean room
think AlphaGo is winning.

1568
01:21:51,615 --> 01:21:53,241
"Winning," okay, not... then not won.

1569
01:21:53,325 --> 01:21:54,552
- That's what she said.
- "Winning."

1570
01:21:54,576 --> 01:21:56,016
She said they called it right there.

1571
01:22:01,499 --> 01:22:03,168
AlphaGo is saying it's going to resign.

1572
01:22:03,251 --> 01:22:04,669
- What...
- No. I'm only joking.

1573
01:22:06,338 --> 01:22:08,214
Jesus, I almost had a heart attack.

1574
01:22:08,798 --> 01:22:10,467
That is ridiculous.

1575
01:22:11,217 --> 01:22:13,345
That was very good.

1576
01:22:13,428 --> 01:22:14,763
- Well played.
- Yeah.

1577
01:22:14,846 --> 01:22:16,139
- Well played.
- No, not really.

1578
01:22:16,222 --> 01:22:19,309
I couldn't do it for too long.
I just couldn't think like that.

1579
01:22:19,392 --> 01:22:21,436
David, you've been waiting
two years to do that.

1580
01:22:21,519 --> 01:22:23,688
He needs to do that.

1581
01:22:23,772 --> 01:22:26,316
We don't know the score.
We have no idea what's going on, right?

1582
01:22:27,150 --> 01:22:28,360
Two-and-a-half points.

1583
01:22:28,443 --> 01:22:30,243
Lee is two-and-a-half points behind,
at least.

1584
01:22:31,488 --> 01:22:33,198
The parting in the center,

1585
01:22:33,281 --> 01:22:35,116
it's very hard to... It's very hard to...

1586
01:22:35,200 --> 01:22:38,119
Lee is touching the bowl.

1587
01:22:38,203 --> 01:22:39,996
It looks like
he's considering resigning.

1588
01:22:40,080 --> 01:22:41,998
Yes.
I think he is thinking of resigning.

1589
01:22:53,176 --> 01:22:54,511
- Did he resign?
- He resigned.

1590
01:22:54,594 --> 01:22:56,054
Hm? He has what?

1591
01:22:56,137 --> 01:22:57,555
I think he resigned.

1592
01:22:57,639 --> 01:23:00,517
- He resigned?
- I think he resigned. Look.

1593
01:23:01,851 --> 01:23:03,812
- He's moved everything in there.
- What?

1594
01:23:03,895 --> 01:23:04,896
He played the white.

1595
01:23:04,980 --> 01:23:06,147
Aja is looking relaxed.

1596
01:23:06,231 --> 01:23:08,066
- Yeah, he has resigned.
- He has resigned.

1597
01:23:08,149 --> 01:23:09,275
Yes, he is giving up.

1598
01:23:09,359 --> 01:23:11,861
- Okay. That's it. He has given up.
- Yeah.

1599
01:23:13,071 --> 01:23:14,906
- Yeah!
- Yes!

1600
01:23:23,623 --> 01:23:27,460
I mean, I think it just really
is a once-in-a-lifetime thing.

1601
01:23:27,877 --> 01:23:31,589
I would say it is the most amazing thing,
you know, I have experienced.

1602
01:23:32,298 --> 01:23:34,759
You know, for us, it is the culmination
of a 20-year dream.

1603
01:23:35,343 --> 01:23:39,055
It started as a pure research endeavor.

1604
01:23:39,139 --> 01:23:42,350
We just wanted to understand.
Can neural networks play the game of Go?

1605
01:23:42,434 --> 01:23:47,981
And from there, it went on
to a level that I never expected.

1606
01:23:48,481 --> 01:23:51,568
And, I'm unbelievably proud
of the team.

1607
01:23:52,944 --> 01:23:55,155
When I started doing
artificial intelligence,

1608
01:23:55,238 --> 01:23:57,699
it was four or five years ago,

1609
01:23:57,782 --> 01:23:59,617
and I was really interested in that,

1610
01:23:59,701 --> 01:24:01,578
but, like, many people
would discourage me,

1611
01:24:01,661 --> 01:24:03,663
and they would say,
okay, there is no future.

1612
01:24:03,747 --> 01:24:07,125
So, actually seeing that within five years

1613
01:24:07,208 --> 01:24:09,169
we are at this stage right now,

1614
01:24:09,294 --> 01:24:10,795
it's amazing by itself.

1615
01:24:10,879 --> 01:24:12,756
Everybody say, "Nine dan."

1616
01:24:12,839 --> 01:24:14,090
- Nine dan!
- Nine dan!

1617
01:24:14,174 --> 01:24:15,300
Kimchi!

1618
01:24:16,051 --> 01:24:18,428
There are so many
possible application domains

1619
01:24:18,511 --> 01:24:22,766
where creativity, in a different dimension
to what humans could do,

1620
01:24:22,849 --> 01:24:24,476
could be immensely valuable to us.

1621
01:24:24,893 --> 01:24:26,653
And I just love to have more
of those moments

1622
01:24:26,728 --> 01:24:29,814
where we look back and say,
"Yeah, that was just like move 37.

1623
01:24:29,898 --> 01:24:31,900
Something beautiful occurred there."

1624
01:24:32,192 --> 01:24:35,862
At least in a broad sense,
move 37 begat move 78,

1625
01:24:35,945 --> 01:24:40,784
begat a new attitude in Se-dol Lee,
a new way of seeing the game.

1626
01:24:41,785 --> 01:24:45,163
He improved through this machine.

1627
01:24:45,663 --> 01:24:51,211
His humanness was expanded
after playing this inanimate creation.

1628
01:24:51,294 --> 01:24:53,755
And the hope is that, that machine,

1629
01:24:53,838 --> 01:24:55,924
and in particular,
the technology behind it,

1630
01:24:56,007 --> 01:24:58,635
can have the same effect with all of us.

1631
01:25:00,178 --> 01:25:03,723
I remember hearing a talk
by Kasparov, who says that, uh,

1632
01:25:03,807 --> 01:25:07,811
a good human plus a machine
is the best combination.

1633
01:25:09,479 --> 01:25:11,856
This is a unique experience.

1634
01:25:12,982 --> 01:25:14,859
Nobody can have this experience

1635
01:25:14,943 --> 01:25:16,986
playing with AlphaGo,
five games like this.

1636
01:25:17,403 --> 01:25:22,951
So, I hope Se-dol Lee can find
something in these five games.

1637
01:25:23,034 --> 01:25:25,286
Maybe some change in his game.

1638
01:25:26,246 --> 01:25:29,040
I saw today how he continuously fought.

1639
01:25:29,582 --> 01:25:30,583
SE-DOL LEE
2016-03-15

1640
01:25:30,667 --> 01:25:31,835
He's very good,

1641
01:25:31,918 --> 01:25:34,170
a master, a really great master.

1642
01:25:36,089 --> 01:25:37,757
I have grown

1643
01:25:37,841 --> 01:25:39,634
through this experience.

1644
01:25:39,717 --> 01:25:44,889
I will make something out of it
with the lessons I have learned.

1645
01:25:45,390 --> 01:25:49,894
I feel thankful and feel like

1646
01:25:50,478 --> 01:25:54,607
I have found the reason I play Go.

1647
01:25:54,691 --> 01:25:59,070
I realize it was really a good choice,
learning to play Go.

1648
01:26:00,238 --> 01:26:06,494
It has been an unforgettable experience.

1649
01:26:09,581 --> 01:26:11,416
This event is done,

1650
01:26:11,499 --> 01:26:14,836
but for the story,
maybe, it's just beginning.

1651
01:26:14,919 --> 01:26:15,919
We don't know.

1652
01:26:19,966 --> 01:26:23,011
It's just when I played with AlphaGo,
he showed me something.

1653
01:26:23,261 --> 01:26:25,471
I felt something beautiful. That was it.

1654
01:26:26,723 --> 01:26:30,977
I saw the world differently,
before everything began.

1655
01:26:32,437 --> 01:26:35,732
What is really behind the game of Go?

1656
01:26:36,524 --> 01:26:39,611
What that is, it can change my game.

1657
01:26:43,198 --> 01:26:47,076
Maybe it can just show humans
something we have never discovered.

1658
01:26:48,828 --> 01:26:50,705
Maybe, there is beauty.