3
00:00:30,682 --> 00:00:31,814
Hi, Alpha.
4
00:00:32,684 --> 00:00:34,034
Hello.
5
00:00:34,077 --> 00:00:35,296
Can you
help me write code?
6
00:00:37,559 --> 00:00:39,082
I was trained
to answer questions,
7
00:00:40,562 --> 00:00:42,216
but I'm able to learn.
8
00:00:43,695 --> 00:00:45,958
That's very
open-minded of you.
9
00:00:46,002 --> 00:00:48,657
Thank you.
I'm glad you're happy with me.
10
00:00:50,311 --> 00:00:51,312
What's this guy doing?
11
00:00:54,532 --> 00:00:56,621
That's a developer.
12
00:00:56,665 --> 00:00:58,014
What do you think
he's working on?
13
00:00:59,537 --> 00:01:01,017
That's a tough question.
14
00:01:01,061 --> 00:01:02,888
He might be working
on a new feature,
15
00:01:02,932 --> 00:01:05,239
a bug fix or something else.
16
00:01:05,282 --> 00:01:06,327
It's quite possible.
17
00:01:06,370 --> 00:01:07,328
Yes.
18
00:01:10,548 --> 00:01:11,854
Do you see my backpack?
19
00:01:13,986 --> 00:01:15,292
That's a badminton racket.
20
00:01:15,336 --> 00:01:17,903
It's a squash racket,
but that's pretty close.
21
00:01:20,123 --> 00:01:21,690
That's a badminton racket.
22
00:01:21,733 --> 00:01:23,126
No, but you're not
the first person
23
00:01:23,170 --> 00:01:24,649
to make that mistake.
24
00:01:34,703 --> 00:01:35,965
AI, the technology
25
00:01:36,008 --> 00:01:38,402
that has been advancing
at breakneck speed.
26
00:01:38,446 --> 00:01:40,796
Artificial
intelligence is all the rage.
27
00:01:40,839 --> 00:01:42,319
Some are now
raising alarm about...
28
00:01:42,363 --> 00:01:43,886
It is definitely concerning.
29
00:01:43,929 --> 00:01:45,322
This is an AI arms race.
30
00:01:45,366 --> 00:01:46,454
We don't know
31
00:01:46,497 --> 00:01:47,585
how this is all
going to shake out,
32
00:01:47,629 --> 00:01:48,891
but it's clear
something is happening.
33
00:01:53,809 --> 00:01:55,115
I'm kind of restless.
34
00:01:57,073 --> 00:02:00,120
Trying to build AGI
is the most exciting journey,
35
00:02:00,163 --> 00:02:02,383
in my opinion, that humans
have ever embarked on.
36
00:02:04,254 --> 00:02:06,300
If you're really going
to take that seriously,
37
00:02:06,343 --> 00:02:08,345
there isn't a lot of time.
38
00:02:08,389 --> 00:02:09,825
Life's very short.
39
00:02:12,523 --> 00:02:14,482
My whole life goal is to solve
40
00:02:14,525 --> 00:02:16,875
artificial
general intelligence.
41
00:02:16,919 --> 00:02:20,357
And on the way,
use AI as the ultimate tool
42
00:02:20,401 --> 00:02:21,750
to solve all the world's
43
00:02:21,793 --> 00:02:23,230
most complex
scientific problems.
44
00:02:25,449 --> 00:02:27,103
I think that's bigger
than the Internet.
45
00:02:27,147 --> 00:02:28,452
I think that's bigger
than mobile.
46
00:02:29,671 --> 00:02:31,020
I think it's more like
47
00:02:31,063 --> 00:02:32,891
the advent
of electricity or fire.
48
00:02:45,687 --> 00:02:46,731
World leaders
49
00:02:46,775 --> 00:02:48,516
and artificial
intelligence experts
50
00:02:48,559 --> 00:02:49,908
are gathering
for the first ever
51
00:02:49,952 --> 00:02:52,781
global AI safety summit,
52
00:02:52,824 --> 00:02:54,348
set to look at the risks
53
00:02:54,391 --> 00:02:56,611
of the fast growing technology
and also...
54
00:02:56,654 --> 00:02:57,873
I think
this is a hugely
55
00:02:57,916 --> 00:03:00,136
critical moment
for all humanity.
56
00:03:01,093 --> 00:03:03,618
It feels like
we're on the cusp
57
00:03:03,661 --> 00:03:05,794
of some incredible things
happening.
58
00:03:05,837 --> 00:03:07,056
Let me take you through
59
00:03:07,099 --> 00:03:08,275
some of the reactions
in today's papers.
60
00:03:08,318 --> 00:03:10,015
AGI is pretty close,
I think.
61
00:03:10,059 --> 00:03:12,844
There's clearly huge interest
in what it is capable of,
62
00:03:12,888 --> 00:03:14,803
where it's taking us.
63
00:03:14,846 --> 00:03:15,978
This is the moment
64
00:03:16,021 --> 00:03:17,936
I've been living
my whole life for.
65
00:03:22,202 --> 00:03:24,987
I've always been fascinated
by the mind.
66
00:03:25,030 --> 00:03:27,990
So I set my heart
on studying neuroscience
67
00:03:28,033 --> 00:03:29,948
because I wanted
to get inspiration
68
00:03:29,992 --> 00:03:31,646
from the brain for AI.
69
00:03:31,689 --> 00:03:33,300
I remember asking Demis,
70
00:03:33,343 --> 00:03:34,649
"What's the end game?"
71
00:03:34,692 --> 00:03:36,172
You know?
So you're going to come here
72
00:03:36,216 --> 00:03:38,218
and you're going
to study neuroscience
73
00:03:38,261 --> 00:03:41,438
and you're going to maybe
get a Ph.D. if you work hard.
74
00:03:42,874 --> 00:03:44,136
And he said,
75
00:03:44,180 --> 00:03:46,965
"You know, I want
to be able to solve AI.
76
00:03:47,009 --> 00:03:49,359
"I want to be able
to solve intelligence."
77
00:03:49,403 --> 00:03:51,361
The human brain
is the only existent proof
78
00:03:51,405 --> 00:03:53,494
we have, perhaps
in the entire universe,
79
00:03:53,537 --> 00:03:56,279
that general intelligence
is possible at all.
80
00:03:56,323 --> 00:03:59,282
And I thought
someone in this building
81
00:03:59,326 --> 00:04:00,370
should be interested
82
00:04:00,414 --> 00:04:02,590
in general intelligence
like I am.
83
00:04:02,633 --> 00:04:04,809
And then Shane's name
popped up.
84
00:04:04,853 --> 00:04:07,334
Our next speaker today
is Shane Legg.
85
00:04:07,377 --> 00:04:08,596
He's from New Zealand,
86
00:04:08,639 --> 00:04:11,512
where he trained in math
and classical ballet.
87
00:04:11,555 --> 00:04:14,079
Are machines actually
becoming more intelligent?
88
00:04:14,123 --> 00:04:16,517
Some people say yes,
some people say no.
89
00:04:16,560 --> 00:04:17,648
It's not really clear.
90
00:04:17,692 --> 00:04:18,867
We know they're getting
a lot faster
91
00:04:18,910 --> 00:04:20,216
at doing computations.
92
00:04:20,260 --> 00:04:21,609
But are we actually
going forwards
93
00:04:21,652 --> 00:04:23,828
in terms
of general intelligence?
94
00:04:23,872 --> 00:04:25,700
We were both
obsessed with AGI,
95
00:04:25,743 --> 00:04:27,397
artificial
general intelligence.
96
00:04:27,441 --> 00:04:29,704
So today I'm going
to be talking about
97
00:04:29,747 --> 00:04:32,097
different approaches
to building AGI.
98
00:04:32,141 --> 00:04:33,882
With my colleague
Demis Hassabis,
99
00:04:33,925 --> 00:04:35,884
we're looking at ways
to bring in ideas
100
00:04:35,927 --> 00:04:37,451
from theoretical neuroscience.
101
00:04:37,494 --> 00:04:41,585
I felt like we were
the keepers of a secret
102
00:04:41,629 --> 00:04:43,152
that no one else knew.
103
00:04:43,195 --> 00:04:45,415
Shane and I knew
no one in academia
104
00:04:45,459 --> 00:04:47,809
would be supportive
of what we were doing.
105
00:04:47,852 --> 00:04:50,855
AI was almost
an embarrassing word
106
00:04:50,899 --> 00:04:52,857
to use in academic circles,
right?
107
00:04:52,901 --> 00:04:54,903
If you said
you were working on AI,
108
00:04:54,946 --> 00:04:57,993
then you clearly weren't
a serious scientist.
109
00:04:58,036 --> 00:05:00,082
So I convinced Shane
the right way to do it
110
00:05:00,125 --> 00:05:01,344
would be to start a company.
111
00:05:01,388 --> 00:05:03,215
Okay,
we're going to try to do
112
00:05:03,259 --> 00:05:04,913
artificial
general intelligence.
113
00:05:04,956 --> 00:05:06,915
It may not even be possible.
114
00:05:06,958 --> 00:05:08,482
We're not quite sure
how we're going to do it,
115
00:05:08,525 --> 00:05:11,528
but we have some ideas
or, kind of, approaches.
116
00:05:11,572 --> 00:05:14,966
Huge amounts of money,
huge amounts of risk,
117
00:05:15,010 --> 00:05:16,533
lots and lots of compute.
118
00:05:18,709 --> 00:05:20,320
And if we pull this off,
119
00:05:20,363 --> 00:05:23,540
it'll be the biggest thing
ever, right?
120
00:05:23,584 --> 00:05:26,108
That is a very hard thing
for a typical investor
121
00:05:26,151 --> 00:05:27,327
to put their money on.
122
00:05:27,370 --> 00:05:29,633
It's almost like
buying a lottery ticket.
123
00:05:29,677 --> 00:05:32,462
I'm going to be speaking about
the system of neuroscience
124
00:05:32,506 --> 00:05:36,727
and how it might be used
to help us build AGI.
125
00:05:36,771 --> 00:05:37,946
Finding initial funding
126
00:05:37,989 --> 00:05:39,208
for this was very hard.
127
00:05:39,251 --> 00:05:41,253
We're going to solve
all of intelligence.
128
00:05:41,297 --> 00:05:42,907
You can imagine
some of the looks I got
129
00:05:42,951 --> 00:05:44,779
when we were
pitching that around.
130
00:05:44,822 --> 00:05:47,912
So I'm a V.C.
and I look at about
131
00:05:47,956 --> 00:05:51,002
700 to 1,000 projects a year.
132
00:05:51,046 --> 00:05:54,789
And I fund
literally 1% of those.
133
00:05:54,832 --> 00:05:57,139
About eight projects a year.
134
00:05:57,182 --> 00:06:00,272
So that means 99% of the time,
you're in "No" mode.
135
00:06:00,316 --> 00:06:01,622
"Wait a minute.
I'm telling you,
136
00:06:01,665 --> 00:06:03,798
"this is the most important
thing of all time.
137
00:06:03,841 --> 00:06:05,103
"I'm giving you
all this build-up
138
00:06:05,147 --> 00:06:06,235
"about how... explain
139
00:06:06,278 --> 00:06:07,671
"how it connects
with the brain,
140
00:06:07,715 --> 00:06:09,412
"why the time's right now,
and then you're asking me,
141
00:06:09,456 --> 00:06:10,979
"'But what's your, how are you
going to make money?
142
00:06:11,022 --> 00:06:12,154
"'What's your product?'"
143
00:06:12,197 --> 00:06:15,984
It's like,
so prosaic a question.
144
00:06:16,506 --> 00:06:17,942
You know?
145
00:06:17,986 --> 00:06:19,117
"Have you not been listening
to what I've been saying?"
146
00:06:19,161 --> 00:06:21,076
We needed investors
147
00:06:21,119 --> 00:06:23,470
who aren't necessarily
going to invest
148
00:06:23,513 --> 00:06:25,123
because they think
it's the best
149
00:06:25,167 --> 00:06:26,690
investment decision.
150
00:06:26,734 --> 00:06:27,865
They're probably
going to invest
151
00:06:27,909 --> 00:06:29,476
because they just think
it's really cool.
152
00:06:29,519 --> 00:06:31,478
He's the Silicon Valley
153
00:06:31,521 --> 00:06:33,523
version of the man
behind the curtain
154
00:06:33,567 --> 00:06:34,872
inThe Wizard of Oz.
155
00:06:34,916 --> 00:06:36,265
He had a lot to do
with giving you
156
00:06:36,308 --> 00:06:39,007
PayPal, Facebook,
YouTube and Yelp.
157
00:06:39,050 --> 00:06:40,443
If everyone says "X,"
158
00:06:40,487 --> 00:06:43,185
Peter Thiel suspects
that the opposite of X
159
00:06:43,228 --> 00:06:44,708
is quite possibly true.
160
00:06:44,752 --> 00:06:47,276
So Peter Thiel
was our first big investor.
161
00:06:47,319 --> 00:06:50,105
But he insisted that
we come to Silicon Valley
162
00:06:50,148 --> 00:06:51,802
because that
was the only place we could...
163
00:06:51,846 --> 00:06:52,977
There would be the talent,
164
00:06:53,021 --> 00:06:55,066
and we could build
that kind of company.
165
00:06:55,110 --> 00:06:56,981
But I was pretty adamant
we should be in London
166
00:06:57,025 --> 00:06:59,027
because I think
London's an amazing city.
167
00:06:59,070 --> 00:07:01,464
Plus, I knew there were
really amazing people
168
00:07:01,508 --> 00:07:03,814
trained at Cambridge
and Oxford and UCL.
169
00:07:03,858 --> 00:07:05,033
In Silicon Valley,
170
00:07:05,076 --> 00:07:06,643
everybody's founding
a company every year,
171
00:07:06,687 --> 00:07:07,818
and then if it doesn't work,
172
00:07:07,862 --> 00:07:09,603
you chuck it
and you start something new.
173
00:07:09,646 --> 00:07:11,169
That is not conducive
174
00:07:11,213 --> 00:07:14,564
to a long-term
research challenge.
175
00:07:14,608 --> 00:07:17,349
So we were totally
an outlier for him.
176
00:07:17,393 --> 00:07:20,962
Hi, everyone.
Welcome to DeepMind.
177
00:07:21,005 --> 00:07:22,529
So, what is our mission?
178
00:07:23,747 --> 00:07:25,706
We summarize it as...
179
00:07:25,749 --> 00:07:28,012
DeepMind's mission is to build
the world's first
180
00:07:28,056 --> 00:07:29,579
general learning machine.
181
00:07:29,623 --> 00:07:31,712
So we always stress the word
"general" and "learning" here
182
00:07:31,755 --> 00:07:32,887
are the key things.
183
00:07:32,930 --> 00:07:35,106
Our mission
was to build an AGI,
184
00:07:35,150 --> 00:07:36,760
an artificial
general intelligence.
185
00:07:36,804 --> 00:07:40,372
And so that means that we need
a system which is general.
186
00:07:40,416 --> 00:07:42,723
It doesn't learn to do
one specific thing.
187
00:07:42,766 --> 00:07:45,073
That's a really key part
of human intelligence.
188
00:07:45,116 --> 00:07:46,683
We can learn to do
many, many things.
189
00:07:46,727 --> 00:07:48,685
It's going to, of course,
be a lot of hard work.
190
00:07:48,729 --> 00:07:51,079
But one of the things
that keeps me up at night
191
00:07:51,122 --> 00:07:53,168
is to not waste this
opportunity to, you know,
192
00:07:53,211 --> 00:07:54,430
to really make
a difference here,
193
00:07:54,474 --> 00:07:56,824
and have a big impact
on the world.
194
00:07:56,867 --> 00:07:58,303
The first people
that came
195
00:07:58,347 --> 00:08:00,262
and joined DeepMind
really believed in the dream.
196
00:08:00,305 --> 00:08:02,090
But this was, I think,
one of the first times
197
00:08:02,133 --> 00:08:04,440
they found a place
full of other dreamers.
198
00:08:04,484 --> 00:08:06,398
You know, we collected
this Manhattan Project,
199
00:08:06,442 --> 00:08:08,400
if you like,
together to solve AI.
200
00:08:08,444 --> 00:08:09,750
In the first two years,
201
00:08:09,793 --> 00:08:10,794
we were in total stealth mode.
202
00:08:10,838 --> 00:08:12,274
And so we couldn't
say to anyone
203
00:08:12,317 --> 00:08:14,755
what were we doing
or where we worked.
204
00:08:14,798 --> 00:08:16,104
It was all quite vague.
205
00:08:16,147 --> 00:08:17,584
It had
no public presence at all.
206
00:08:17,627 --> 00:08:18,933
You couldn't
look at a website.
207
00:08:18,976 --> 00:08:21,109
The office
was at a secret location.
208
00:08:21,152 --> 00:08:23,981
When we would interview people
in those early days,
209
00:08:24,025 --> 00:08:26,288
they would show up
very nervously.
210
00:08:27,463 --> 00:08:29,639
I had at least one candidate
who said,
211
00:08:29,683 --> 00:08:31,989
"I just messaged my wife
to tell her exactly
212
00:08:32,033 --> 00:08:33,338
"where I'm going just in case
213
00:08:33,382 --> 00:08:34,514
"this turns out to be some
kind of horrible scam
214
00:08:34,557 --> 00:08:35,819
"and I'm going
to get kidnapped."
215
00:08:35,863 --> 00:08:39,693
Well, my favorite new person
who's an investor,
216
00:08:39,736 --> 00:08:43,000
who I've been working
for a year, is Elon Musk.
217
00:08:43,044 --> 00:08:44,132
So for those of you
who don't know,
218
00:08:44,175 --> 00:08:45,350
this is what he looks like.
219
00:08:45,394 --> 00:08:47,483
And he hadn't really thought
much about AI
220
00:08:47,527 --> 00:08:49,137
until we chatted.
221
00:08:49,180 --> 00:08:51,269
His mission is to die on Mars
or something.
222
00:08:51,313 --> 00:08:52,923
But not on impact.
223
00:08:52,967 --> 00:08:54,185
So...
224
00:08:55,186 --> 00:08:57,145
We made some big decisions
225
00:08:57,188 --> 00:08:59,321
about how we were going
to approach building AI.
226
00:08:59,364 --> 00:09:01,105
This is a reinforcement
learning setup.
227
00:09:01,149 --> 00:09:02,803
This is the kind of setup
that we think about
228
00:09:02,846 --> 00:09:06,328
when we say we're building,
you know, an AI agent.
229
00:09:06,371 --> 00:09:08,504
It's basically the agent,
which is the AI,
230
00:09:08,548 --> 00:09:09,984
and then there's
the environment
231
00:09:10,027 --> 00:09:11,159
that it's interacting with.
232
00:09:11,202 --> 00:09:12,464
We decided that games,
233
00:09:12,508 --> 00:09:13,988
as long as
you're very disciplined
234
00:09:14,031 --> 00:09:15,293
about how you use them,
235
00:09:15,337 --> 00:09:17,252
are the perfect
training ground
236
00:09:17,295 --> 00:09:19,341
for AI development.
237
00:09:19,384 --> 00:09:21,691
We wanted
to try to create one algorithm
238
00:09:21,735 --> 00:09:23,650
that could to be
trained up to play
239
00:09:23,693 --> 00:09:26,043
several dozen
different Atari games.
240
00:09:26,087 --> 00:09:27,479
So just like a human,
241
00:09:27,523 --> 00:09:29,525
you have to use the same brain
to play all the games.
242
00:09:29,569 --> 00:09:30,874
You can think of it
243
00:09:30,918 --> 00:09:33,007
that you provide the agent
with the cartridge.
244
00:09:33,050 --> 00:09:34,312
And you say,
245
00:09:34,356 --> 00:09:35,705
"Okay, imagine you're born
into that world
246
00:09:35,749 --> 00:09:37,533
"with that cartridge,
and you just get to interact
247
00:09:37,577 --> 00:09:39,404
"with the pixels
and see the score.
248
00:09:40,362 --> 00:09:41,668
"What can you do?"
249
00:09:43,974 --> 00:09:47,369
So what you're going to do is
take your Q function. Q-K...
250
00:09:47,412 --> 00:09:49,197
Q-learning
is one of the oldest methods
251
00:09:49,240 --> 00:09:50,894
for reinforcement learning.
252
00:09:50,938 --> 00:09:53,680
And what we did was combine
reinforcement learning
253
00:09:53,723 --> 00:09:56,770
with deep learning
in one system.
254
00:09:56,813 --> 00:09:59,337
No one had ever combined
those two things together
255
00:09:59,381 --> 00:10:01,426
at scale to do
anything impressive,
256
00:10:01,470 --> 00:10:03,515
and we needed
to prove out this thesis.
257
00:10:03,559 --> 00:10:06,780
We tried doingPong
as the first game.
258
00:10:06,823 --> 00:10:08,129
It seemed like the simplest.
259
00:10:08,172 --> 00:10:10,218
It hasn't been told
260
00:10:10,261 --> 00:10:11,828
anything about
what it's controlling
261
00:10:11,872 --> 00:10:12,873
or what it's supposed to do.
262
00:10:12,916 --> 00:10:14,657
All it knows
is that score is good
263
00:10:14,701 --> 00:10:18,052
and it has to learn
what its controls do,
264
00:10:18,095 --> 00:10:20,620
and build everything...
first principles.
265
00:10:29,193 --> 00:10:30,412
It wasn't really working.
266
00:10:32,588 --> 00:10:34,068
I was just
saying to Shane,
267
00:10:34,111 --> 00:10:37,071
"Maybe we're just wrong,
and we can't even doPong."
268
00:10:37,114 --> 00:10:38,812
It was a bit
nerve-racking,
269
00:10:38,855 --> 00:10:40,422
thinking how far we had to go
270
00:10:40,465 --> 00:10:42,337
if we were going
to really build
271
00:10:42,380 --> 00:10:44,426
a generally
intelligent system.
272
00:10:44,469 --> 00:10:45,732
And it felt like
it was time
273
00:10:45,775 --> 00:10:47,255
to give up and move on.
274
00:10:48,169 --> 00:10:49,387
And then suddenly...
275
00:10:51,389 --> 00:10:53,609
We got our first point.
276
00:10:53,653 --> 00:10:56,612
And then it was like,
"Is this random?"
277
00:10:56,656 --> 00:10:59,180
"No, no, it's really
getting a point now."
278
00:10:59,223 --> 00:11:00,703
It was really exciting
that this thing
279
00:11:00,747 --> 00:11:02,226
that previously
couldn't even figure out
280
00:11:02,270 --> 00:11:03,532
how to move a paddle
281
00:11:03,575 --> 00:11:05,926
had suddenly been able
to totally get it right.
282
00:11:05,969 --> 00:11:07,144
Then it was getting
a few points.
283
00:11:07,188 --> 00:11:08,624
And then it won
its first game.
284
00:11:08,668 --> 00:11:10,974
And then three months later,
no human could beat it.
285
00:11:11,018 --> 00:11:14,238
You hadn't told it the rules,
how to get the score, nothing.
286
00:11:14,282 --> 00:11:16,110
And you just tell it
to maximize the score,
287
00:11:16,153 --> 00:11:17,372
and it goes away and does it.
288
00:11:17,415 --> 00:11:18,678
This is the first time
289
00:11:18,721 --> 00:11:20,549
anyone had done
this end-to-end learning.
290
00:11:20,592 --> 00:11:24,292
"Okay, so we have this working
in quite a general way.
291
00:11:24,335 --> 00:11:25,772
"Now let's try another game."
292
00:11:25,815 --> 00:11:27,512
So then
we triedBreakout.
293
00:11:27,556 --> 00:11:29,123
At the beginning,
after 100 games,
294
00:11:29,166 --> 00:11:30,777
the agent is not very good.
295
00:11:30,820 --> 00:11:32,474
It's missing the ball
most of the time,
296
00:11:32,517 --> 00:11:34,389
but it's starting to get
the hang of the idea
297
00:11:34,432 --> 00:11:35,999
that the bat should go
towards the ball.
298
00:11:36,043 --> 00:11:37,566
Now, after 300 games,
299
00:11:37,609 --> 00:11:40,656
it's about as good as
any human can play this.
300
00:11:40,700 --> 00:11:42,049
We thought,
"Well, that's pretty cool,"
301
00:11:42,092 --> 00:11:44,312
but we left the system playing
for another 200 games,
302
00:11:44,355 --> 00:11:46,053
and it did this amazing thing.
303
00:11:46,096 --> 00:11:47,358
It found the optimal strategy
304
00:11:47,402 --> 00:11:49,404
was to dig a tunnel
around the side
305
00:11:49,447 --> 00:11:51,667
and put the ball
around the back of the wall.
306
00:11:51,711 --> 00:11:53,234
Finally, the agent
307
00:11:53,277 --> 00:11:54,365
is actually achieving
308
00:11:54,409 --> 00:11:55,627
what you thought
it would achieve.
309
00:11:55,671 --> 00:11:57,325
That is a great feeling.
Right?
310
00:11:57,368 --> 00:11:59,283
Like, I mean,
when we do research,
311
00:11:59,327 --> 00:12:00,676
that is the best
we can hope for.
312
00:12:00,720 --> 00:12:03,200
We started generalizing
to 50 games,
313
00:12:03,244 --> 00:12:05,463
and we basically
created a recipe.
314
00:12:05,507 --> 00:12:06,813
We could just take a game
315
00:12:06,856 --> 00:12:08,379
that we have
never seen before.
316
00:12:08,423 --> 00:12:09,903
We would run
the algorithm on that,
317
00:12:09,946 --> 00:12:13,036
and DQN could train itself
from scratch,
318
00:12:13,080 --> 00:12:14,429
achieving human level
319
00:12:14,472 --> 00:12:15,996
or sometimes better
than human level.
320
00:12:16,039 --> 00:12:18,433
We didn't build it
to play any of them.
321
00:12:18,476 --> 00:12:20,522
We could just give it
a bunch of games
322
00:12:20,565 --> 00:12:22,654
and would figure it out
for itself.
323
00:12:22,698 --> 00:12:25,179
And there was something
quite magical in that.
324
00:12:25,222 --> 00:12:26,528
Suddenly you had something
325
00:12:26,571 --> 00:12:27,921
that would respond and learn
326
00:12:27,964 --> 00:12:30,358
whatever situation
it was parachuted into.
327
00:12:30,401 --> 00:12:33,013
And that was like a huge,
huge breakthrough.
328
00:12:33,056 --> 00:12:35,276
It was in many respects
329
00:12:35,319 --> 00:12:36,625
the first example
330
00:12:36,668 --> 00:12:39,062
of any kind of thing
you could call
331
00:12:39,106 --> 00:12:40,542
a general intelligence.
332
00:12:42,109 --> 00:12:43,893
Although we were
a well-funded startup,
333
00:12:43,937 --> 00:12:47,375
holding us back
was not enough compute power.
334
00:12:47,418 --> 00:12:49,203
I realized that
this would accelerate
335
00:12:49,246 --> 00:12:51,466
our time scale
to AGI massively.
336
00:12:51,509 --> 00:12:53,163
I used to see Demis
quite frequently.
337
00:12:53,207 --> 00:12:55,035
We'd have lunch, and he did...
338
00:12:56,210 --> 00:12:58,865
say to me that
he had two companies
339
00:12:58,908 --> 00:13:02,564
that were involved
in buying DeepMind.
340
00:13:02,607 --> 00:13:04,609
And he didn't know
which one to go with.
341
00:13:04,653 --> 00:13:08,526
The issue was,
would any commercial company
342
00:13:08,570 --> 00:13:12,661
appreciate the real importance
of the research?
343
00:13:12,704 --> 00:13:15,838
And give the research time
to come to fruition
344
00:13:15,882 --> 00:13:17,797
and not be breathing down
their necks,
345
00:13:17,840 --> 00:13:21,539
saying, "We want some kind of
commercial benefit from this."
346
00:13:27,284 --> 00:13:32,463
Google has bought DeepMind
for a reported £400,000,000,
347
00:13:32,507 --> 00:13:34,726
making the artificial
intelligence firm
348
00:13:34,770 --> 00:13:37,947
its largest
European acquisition so far.
349
00:13:37,991 --> 00:13:39,644
The company was founded
350
00:13:39,688 --> 00:13:43,344
by 37-year-old entrepreneur
Demis Hassabis.
351
00:13:43,387 --> 00:13:45,563
After the acquisition,
I started mentoring
352
00:13:45,607 --> 00:13:47,261
and spending time with Demis,
353
00:13:47,304 --> 00:13:48,915
and just listening to him.
354
00:13:48,958 --> 00:13:52,396
And this is a person
who fundamentally
355
00:13:52,440 --> 00:13:55,573
is a scientist
and a natural scientist.
356
00:13:55,617 --> 00:13:58,489
He wants science to solve
every problem in the world,
357
00:13:58,533 --> 00:14:00,622
and he believes it can do so.
358
00:14:00,665 --> 00:14:03,625
That's not a normal person
you find in a tech company.
359
00:14:05,496 --> 00:14:07,455
We were able
to not only join Google
360
00:14:07,498 --> 00:14:10,240
but run independently
in London,
361
00:14:10,284 --> 00:14:11,328
build our culture,
362
00:14:11,372 --> 00:14:13,461
which was optimized
for breakthroughs
363
00:14:13,504 --> 00:14:15,419
and not deal with products,
364
00:14:15,463 --> 00:14:17,726
do pure research.
365
00:14:17,769 --> 00:14:19,293
Our investors
didn't want to sell,
366
00:14:19,336 --> 00:14:20,685
but we decided
367
00:14:20,729 --> 00:14:22,731
that this was the best thing
for the mission.
368
00:14:22,774 --> 00:14:24,646
In many senses,
we were underselling
369
00:14:24,689 --> 00:14:26,169
in terms of value
before it more matured,
370
00:14:26,213 --> 00:14:28,128
and you could have sold it
for a lot more money.
371
00:14:28,171 --> 00:14:32,828
And the reason is because
there's no time to waste.
372
00:14:32,872 --> 00:14:35,178
There's so many things
that got to be cracked
373
00:14:35,222 --> 00:14:37,659
while the brain
is still in gear.
374
00:14:37,702 --> 00:14:39,008
You know, I'm still alive.
375
00:14:39,052 --> 00:14:40,880
There's all these things
that gotta be done.
376
00:14:40,923 --> 00:14:42,925
So you haven't got--
I mean, how many...
377
00:14:42,969 --> 00:14:44,492
How many billions
would you trade for
378
00:14:44,535 --> 00:14:45,797
another five years of life,
you know,
379
00:14:45,841 --> 00:14:48,365
to do what you set out to do?
380
00:14:48,409 --> 00:14:49,758
Okay, all of a sudden,
381
00:14:49,801 --> 00:14:52,717
we've got this massive scale
compute available to us.
382
00:14:52,761 --> 00:14:53,936
What can we do with that?
383
00:14:56,591 --> 00:14:59,942
Go is the pinnacle
of board games.
384
00:14:59,986 --> 00:15:04,642
It is the most complex game
ever devised by man.
385
00:15:04,686 --> 00:15:06,688
There are more possible
board configurations
386
00:15:06,731 --> 00:15:09,821
in the game of Go than there
are atoms in the universe.
387
00:15:09,865 --> 00:15:13,390
Go is the holy grail
of artificial intelligence.
388
00:15:13,434 --> 00:15:14,522
For many years,
389
00:15:14,565 --> 00:15:16,002
people have looked
at this game
390
00:15:16,045 --> 00:15:17,917
and they've thought,
"Wow, this is just too hard."
391
00:15:17,960 --> 00:15:20,180
Everything we've ever
tried in AI,
392
00:15:20,223 --> 00:15:22,704
it just falls over when
you try the game of Go.
393
00:15:22,747 --> 00:15:23,966
And so that's why
it feels like
394
00:15:24,010 --> 00:15:26,099
a real litmus test
of progress.
395
00:15:26,142 --> 00:15:28,579
We had just bought DeepMind.
396
00:15:28,623 --> 00:15:30,712
They were working
on reinforcement learning
397
00:15:30,755 --> 00:15:32,975
and they were the world's
experts in games.
398
00:15:33,019 --> 00:15:34,846
And so when
they introduced the idea
399
00:15:34,890 --> 00:15:37,197
that they could beat
the top level Go players
400
00:15:37,240 --> 00:15:40,113
in a game that was thought
to be incomputable,
401
00:15:40,156 --> 00:15:42,724
I thought, "Well,
that's pretty interesting."
402
00:15:42,767 --> 00:15:46,467
Our ultimate next step
is to play the legendary
403
00:15:46,510 --> 00:15:49,209
Lee Sedol
in just over two weeks.
404
00:15:50,514 --> 00:15:52,038
A match like no other
405
00:15:52,081 --> 00:15:54,257
is about to get underway
in South Korea.
406
00:15:54,301 --> 00:15:57,826
Lee Sedol
is getting ready to rumble.
407
00:15:57,869 --> 00:15:59,262
Lee Sedol is probably
408
00:15:59,306 --> 00:16:01,699
one of the greatest players
of the last decade.
409
00:16:01,743 --> 00:16:04,224
I describe him
as the Roger Federer of Go.
410
00:16:05,573 --> 00:16:06,922
He showed up,
411
00:16:06,966 --> 00:16:09,838
and all of a sudden
we have a thousand Koreans
412
00:16:09,881 --> 00:16:13,059
who represent
all of Korean society,
413
00:16:13,102 --> 00:16:14,190
the top Go players.
414
00:16:15,583 --> 00:16:17,802
And then we have Demis.
415
00:16:17,846 --> 00:16:19,848
And the great
engineering team.
416
00:16:20,588 --> 00:16:22,285
He's very famous
417
00:16:22,329 --> 00:16:25,506
for very creative
fighting play.
418
00:16:25,549 --> 00:16:28,770
So this could be
difficult for us.
419
00:16:28,813 --> 00:16:31,991
I figured Lee Sedol
is going to beat these guys,
420
00:16:32,034 --> 00:16:34,297
but they'll make
a good showing.
421
00:16:34,341 --> 00:16:35,733
Good for a startup.
422
00:16:38,040 --> 00:16:39,563
I went over
to the technical group
423
00:16:39,607 --> 00:16:40,912
and they said,
424
00:16:40,956 --> 00:16:42,305
"Let me show you
how our algorithm works."
425
00:16:43,654 --> 00:16:45,091
If you step
through the actual game,
426
00:16:45,134 --> 00:16:47,789
we can see, kind of,
how AlphaGo thinks.
427
00:16:47,832 --> 00:16:50,226
The way we start off
on training AlphaGo
428
00:16:50,270 --> 00:16:52,968
is by showing it 100,000 games
429
00:16:53,012 --> 00:16:54,491
that strong amateurs
have played.
430
00:16:54,535 --> 00:16:55,710
And we first initially
431
00:16:55,753 --> 00:16:58,887
get AlphaGo to mimic
the human player,
432
00:16:58,930 --> 00:17:00,802
and then through
reinforcement learning,
433
00:17:00,845 --> 00:17:02,369
it plays against
different versions of itself
434
00:17:02,412 --> 00:17:05,720
many millions of times
and learns from its errors.
435
00:17:05,763 --> 00:17:07,591
Hmm, this is interesting.
436
00:17:07,635 --> 00:17:08,679
All right, folks,
437
00:17:08,723 --> 00:17:10,594
you're going to see
history made.
438
00:17:12,770 --> 00:17:14,772
So the game starts.
439
00:17:14,816 --> 00:17:15,991
He's really concentrating.
440
00:17:16,035 --> 00:17:17,645
If you really look at the...
441
00:17:20,909 --> 00:17:25,348
That's a very surprising move.
442
00:17:25,392 --> 00:17:27,916
I think we're
seeing an original move here.
443
00:17:34,401 --> 00:17:35,793
Yeah, that's an exciting move.
444
00:17:36,185 --> 00:17:37,230
I like...
445
00:17:37,273 --> 00:17:38,579
Professional commentators
446
00:17:38,622 --> 00:17:40,102
almost unanimously said
447
00:17:40,146 --> 00:17:43,323
that not a single human player
would have chosen move 37.
448
00:17:43,366 --> 00:17:45,803
So I actually had a poke
around in AlphaGo
449
00:17:45,847 --> 00:17:47,414
to see what AlphaGo thought.
450
00:17:47,457 --> 00:17:50,330
And AlphaGo actually agreed
with that assessment.
451
00:17:50,373 --> 00:17:53,811
AlphaGo said there was a one
in 10,000 probability
452
00:17:53,855 --> 00:17:57,772
that move 37 would have been
played by a human player.
453
00:18:08,826 --> 00:18:10,176
The game of Go
has been studied
454
00:18:10,219 --> 00:18:11,568
for thousands of years.
455
00:18:11,612 --> 00:18:15,137
And AlphaGo discovered
something completely new.
456
00:18:16,878 --> 00:18:19,707
He resigned.
Lee Sedol has just resigned.
457
00:18:19,750 --> 00:18:21,187
He's beaten.
458
00:18:22,710 --> 00:18:24,668
The battle
between man versus machine,
459
00:18:24,712 --> 00:18:26,235
a computer just came out
the victor.
460
00:18:26,279 --> 00:18:28,237
Google
put its DeepMind team
461
00:18:28,281 --> 00:18:29,673
to the test against
462
00:18:29,717 --> 00:18:32,023
one of the brightest minds
in the world and won.
463
00:18:32,067 --> 00:18:33,721
That's when we realized
464
00:18:33,764 --> 00:18:35,244
the DeepMind people knew
what they were doing
465
00:18:35,288 --> 00:18:37,551
and to pay attention
to reinforcement learning
466
00:18:37,594 --> 00:18:38,900
as they have invented it.
467
00:18:40,075 --> 00:18:41,816
Based on that experience,
468
00:18:41,859 --> 00:18:44,819
AlphaGo got better
and better and better.
469
00:18:44,862 --> 00:18:45,950
And they had a little chart
470
00:18:45,994 --> 00:18:47,517
of how much better
they were getting.
471
00:18:47,561 --> 00:18:49,302
And I said,
"When does this stop?"
472
00:18:50,085 --> 00:18:50,999
And Demis said,
473
00:18:51,042 --> 00:18:52,696
"When we beat the Chinese guy,
474
00:18:52,740 --> 00:18:55,743
"the top-rated player
in the world."
475
00:18:56,961 --> 00:18:59,225
Ke Jie versus AlphaGo.
476
00:19:03,620 --> 00:19:04,795
And I think we will see
477
00:19:04,839 --> 00:19:06,145
AlphaGo pushing through there.
478
00:19:06,188 --> 00:19:08,190
AlphaGo is ahead quite a bit.
479
00:19:08,234 --> 00:19:11,454
About halfway
through the first game,
480
00:19:11,498 --> 00:19:14,675
the best player in the world
was not doing so well.
481
00:19:14,718 --> 00:19:17,808
What can black do here?
482
00:19:19,114 --> 00:19:21,247
Looks difficult.
483
00:19:21,290 --> 00:19:23,336
And at a critical moment...
484
00:19:32,997 --> 00:19:35,913
the Chinese government
ordered the feed cut off.
485
00:19:38,307 --> 00:19:41,745
It was at that moment
we were telling the world
486
00:19:41,789 --> 00:19:44,966
that something new
had arrived on earth.
487
00:19:47,621 --> 00:19:48,970
In the 1950s
488
00:19:49,013 --> 00:19:51,929
when Russia'sSputnik
satellite was launched,
489
00:19:53,279 --> 00:19:55,194
it changed
the course of history.
490
00:19:55,237 --> 00:19:57,544
It is a challenge
that America must meet
491
00:19:57,587 --> 00:19:59,633
to survive in the Space Age.
492
00:19:59,676 --> 00:20:02,375
This has been
called theSputnik moment.
493
00:20:02,418 --> 00:20:06,335
The Sputnikmoment created
a massive reaction in the US
494
00:20:06,379 --> 00:20:10,034
in terms of funding
for science and engineering,
495
00:20:10,078 --> 00:20:12,254
and particularly
of space technology.
496
00:20:12,298 --> 00:20:15,823
For China,
AlphaGo was the wakeup call,
497
00:20:15,866 --> 00:20:17,128
the Sputnikmoment.
498
00:20:17,172 --> 00:20:19,870
It launched an AI space race.
499
00:20:21,220 --> 00:20:23,047
We had this
huge idea that worked,
500
00:20:23,091 --> 00:20:26,355
and now the whole world knows.
501
00:20:26,399 --> 00:20:28,879
It's always easier
to land on the moon
502
00:20:28,923 --> 00:20:30,838
if someone's already
landed there.
503
00:20:32,056 --> 00:20:34,450
It is going to matter
who builds AI,
504
00:20:34,494 --> 00:20:36,626
and how it gets built.
505
00:20:36,670 --> 00:20:38,324
I always feel that pressure.
506
00:20:42,066 --> 00:20:43,764
There's been
a big chain of events
507
00:20:43,807 --> 00:20:46,680
that followed on from all
of the excitement of AlphaGo.
508
00:20:46,723 --> 00:20:48,072
When we played
against Lee Sedol,
509
00:20:48,116 --> 00:20:49,248
we actually had a system
510
00:20:49,291 --> 00:20:50,684
that had been trained
on human data,
511
00:20:50,727 --> 00:20:52,251
on all of the millions
of games
512
00:20:52,294 --> 00:20:55,036
that have been played
by human experts.
513
00:20:55,079 --> 00:20:56,994
We eventually found
a new algorithm,
514
00:20:57,038 --> 00:20:59,170
a much more elegant approach
to the whole system,
515
00:20:59,214 --> 00:21:01,172
which actually stripped out
all of the human knowledge
516
00:21:01,216 --> 00:21:03,697
and just started
completely from scratch.
517
00:21:03,740 --> 00:21:06,700
And that became a project
which we called AlphaZero.
518
00:21:06,743 --> 00:21:09,529
Zero, meaning having zero
human knowledge in the loop.
519
00:21:11,879 --> 00:21:13,054
Instead of learning
from human data,
520
00:21:13,097 --> 00:21:15,796
it learned from its own games.
521
00:21:15,839 --> 00:21:17,841
So it actually
became its own teacher.
522
00:21:21,280 --> 00:21:23,499
AlphaZero is an experiment
523
00:21:23,543 --> 00:21:26,720
in how little knowledge
can we put into these systems
524
00:21:26,763 --> 00:21:28,243
and how quickly
and how efficiently
525
00:21:28,287 --> 00:21:29,723
can they learn?
526
00:21:29,766 --> 00:21:32,552
But the other thing is AlphaZero
doesn't have any rules.
527
00:21:32,595 --> 00:21:33,553
It learns through experience.
528
00:21:36,120 --> 00:21:38,862
The next stage
was to make it more general,
529
00:21:38,906 --> 00:21:40,995
so that it could play
any two-player game.
530
00:21:41,038 --> 00:21:42,344
Things like chess,
531
00:21:42,388 --> 00:21:44,085
and in fact,
any kind of two-player
532
00:21:44,128 --> 00:21:45,391
perfect information game.
533
00:21:45,434 --> 00:21:46,653
It's going really well.
534
00:21:46,696 --> 00:21:47,828
It's going
really, really well.
535
00:21:47,871 --> 00:21:50,091
- Oh, wow.
- It's going down, like fast.
536
00:21:50,134 --> 00:21:53,050
AlphaGo used
to take a few months to train,
537
00:21:53,094 --> 00:21:55,662
but AlphaZero could start
in the morning
538
00:21:55,705 --> 00:21:57,794
playing completely randomly
539
00:21:57,838 --> 00:22:01,015
and then by tea
be at superhuman level.
540
00:22:01,058 --> 00:22:03,365
And by dinner it will be
the strongest chess entity
541
00:22:03,409 --> 00:22:04,758
there's ever been.
542
00:22:04,801 --> 00:22:06,629
- Amazing, it's amazing.
- Yeah.
543
00:22:06,673 --> 00:22:09,371
It's discovered its own
attacking style, you know,
544
00:22:09,415 --> 00:22:11,417
to take on the current
level of defense.
545
00:22:11,460 --> 00:22:12,766
I mean, I never
in my wildest dreams...
546
00:22:12,809 --> 00:22:14,942
I agree. Actually, I was not
expecting that either.
547
00:22:14,985 --> 00:22:16,422
And it's fun for me.
548
00:22:16,465 --> 00:22:18,598
I mean, it's inspired me
to get back into chess again,
549
00:22:18,641 --> 00:22:20,164
because it's cool to see
550
00:22:20,208 --> 00:22:22,384
that there's even more depth
than we thought in chess.
551
00:22:31,611 --> 00:22:34,440
I actually got
into AI through games.
552
00:22:35,658 --> 00:22:37,660
Initially, it was board games.
553
00:22:37,704 --> 00:22:40,054
I was thinking,
"How is my brain doing this?"
554
00:22:40,097 --> 00:22:41,969
Like, what is it doing?
555
00:22:43,362 --> 00:22:47,148
I was very aware of that
from a very young age.
556
00:22:47,191 --> 00:22:50,064
So I've always been thinking
about thinking.
557
00:22:50,107 --> 00:22:52,762
The British
and American chess champions
558
00:22:52,806 --> 00:22:55,025
meet to begin
a series of matches.
559
00:22:55,069 --> 00:22:56,592
Playing alongside them
are the cream
560
00:22:56,636 --> 00:22:59,073
of Britain and America's
youngest players.
561
00:22:59,116 --> 00:23:01,467
Demis Hassabis
is representing Britain.
562
00:23:06,210 --> 00:23:07,864
When Demis was four,
563
00:23:07,908 --> 00:23:11,259
he first showed
an aptitude for chess.
564
00:23:12,739 --> 00:23:14,044
By the time he was six,
565
00:23:14,088 --> 00:23:18,048
he became London
under-eight champion.
566
00:23:18,092 --> 00:23:19,485
My parents
were very interesting
567
00:23:19,528 --> 00:23:20,834
and unusual, actually.
568
00:23:20,877 --> 00:23:23,750
I'd probably describe them
as quite bohemian.
569
00:23:23,793 --> 00:23:25,491
My father
was a singer-songwriter
570
00:23:25,534 --> 00:23:26,709
when he was younger,
571
00:23:26,753 --> 00:23:28,363
and Bob Dylan was his hero.
572
00:23:38,068 --> 00:23:39,287
Yeah, yeah.
573
00:23:41,637 --> 00:23:44,031
What is it
that you like about this game?
574
00:23:45,075 --> 00:23:47,121
It's just a good
thinking game.
575
00:23:49,253 --> 00:23:51,038
At the time,
I was the second-highest rated
576
00:23:51,081 --> 00:23:52,692
chess player in the world
for my age.
577
00:23:52,735 --> 00:23:54,345
But although I was on track
578
00:23:54,389 --> 00:23:55,956
to be a professional
chess player,
579
00:23:55,999 --> 00:23:57,523
I thought that was what
I was going to do.
580
00:23:57,566 --> 00:23:59,220
No matter how much
I loved the game,
581
00:23:59,263 --> 00:24:01,222
it was incredibly stressful.
582
00:24:01,265 --> 00:24:03,354
Definitely was not fun
and games for me.
583
00:24:03,398 --> 00:24:05,226
My parents used to, you know,
584
00:24:05,269 --> 00:24:06,923
get very upset
when I lost the game
585
00:24:06,967 --> 00:24:10,405
and angry
if I forgot something.
586
00:24:10,449 --> 00:24:12,407
And because it was quite high
stakes for them, you know,
587
00:24:12,451 --> 00:24:14,061
it cost a lot of money
to go to these tournaments.
588
00:24:14,104 --> 00:24:15,715
And my parents
didn't have much money.
589
00:24:18,413 --> 00:24:19,719
My parents thought, you know,
590
00:24:19,762 --> 00:24:22,069
"If you interested
in being a chess professional,
591
00:24:22,112 --> 00:24:25,376
"this is really important.
It's like your exams."
592
00:24:27,291 --> 00:24:30,077
I remember
I was about 12-years-old
593
00:24:30,120 --> 00:24:32,079
and I was at this
international chess tournament
594
00:24:32,122 --> 00:24:34,255
in Liechtenstein
up in the mountains.
595
00:24:43,307 --> 00:24:45,571
And we were in this
huge church hall
596
00:24:47,181 --> 00:24:48,399
with, you know,
597
00:24:48,443 --> 00:24:50,184
hundreds of international
chess players.
598
00:24:52,403 --> 00:24:56,016
And I was playing
the ex-Danish champion.
599
00:24:56,059 --> 00:24:58,801
He must have been
in his 30s, probably.
600
00:25:00,411 --> 00:25:02,979
In those days,
there was a long time limit.
601
00:25:03,023 --> 00:25:05,068
The games could
literally last all day.
602
00:25:08,332 --> 00:25:10,813
- We were into our tenth hour.
603
00:25:20,257 --> 00:25:23,086
And we were in this
incredibly unusual ending.
604
00:25:23,130 --> 00:25:24,566
I think it should be a draw.
605
00:25:26,307 --> 00:25:28,657
But he kept on trying
to win for hours.
606
00:25:35,359 --> 00:25:38,319
Finally, he tried
one last cheap trick.
607
00:25:42,541 --> 00:25:44,455
All I had to do
was give away my queen.
608
00:25:44,499 --> 00:25:45,674
Then it would be stalemate.
609
00:25:47,023 --> 00:25:48,634
But I was so tired,
610
00:25:48,677 --> 00:25:50,157
I thought it was inevitable
I was going to be checkmated.
611
00:25:52,289 --> 00:25:53,508
And so I resigned.
612
00:25:57,294 --> 00:25:59,558
He jumped up.
Just started laughing.
613
00:26:01,777 --> 00:26:02,909
And he went,
614
00:26:02,952 --> 00:26:04,171
"Why have you resigned?
It's a draw."
615
00:26:04,214 --> 00:26:05,346
And he immediately,
with a flourish,
616
00:26:05,389 --> 00:26:06,652
sort of showed me
the drawing move.
617
00:26:09,045 --> 00:26:12,396
I felt so sick to my stomach.
618
00:26:12,440 --> 00:26:14,137
It made me think of
the rest of that tournament.
619
00:26:14,181 --> 00:26:16,662
Like, are we wasting
our minds?
620
00:26:16,705 --> 00:26:19,708
Is this the best use
of all this brain power?
621
00:26:19,752 --> 00:26:22,363
Everybody's, collectively,
in that building?
622
00:26:22,406 --> 00:26:24,060
If you could somehow plug in
623
00:26:24,104 --> 00:26:27,542
those 300 brains
into a system,
624
00:26:27,586 --> 00:26:29,022
you might be able
to solve cancer
625
00:26:29,065 --> 00:26:30,501
with that level
of brain power.
626
00:26:31,633 --> 00:26:33,504
This intuitive feeling
came over me
627
00:26:33,548 --> 00:26:35,071
that although I love chess,
628
00:26:35,115 --> 00:26:38,335
this is not the right thing
to spend my whole life on.
629
00:26:51,479 --> 00:26:53,089
Demis and myself,
630
00:26:53,133 --> 00:26:56,092
our plan was always
to fill DeepMind
631
00:26:56,136 --> 00:26:57,224
with some of the most
632
00:26:57,267 --> 00:26:59,226
brilliant scientists
in the world.
633
00:26:59,269 --> 00:27:01,054
So we had the human brains
634
00:27:01,097 --> 00:27:04,927
necessary to create
an AGI system.
635
00:27:04,971 --> 00:27:09,279
By definition, the "G"
in AGI is about generality.
636
00:27:09,323 --> 00:27:13,109
What I imagine is being able
to talk to an agent,
637
00:27:13,153 --> 00:27:15,198
the agent can talk back,
638
00:27:15,242 --> 00:27:18,898
and the agent is able to solve
novel problems
639
00:27:18,941 --> 00:27:20,595
that it hasn't seen before.
640
00:27:20,639 --> 00:27:22,728
That's a really key part
of human intelligence,
641
00:27:22,771 --> 00:27:24,468
and it's that
cognitive breadth
642
00:27:24,512 --> 00:27:27,733
and flexibility
that's incredible.
643
00:27:27,776 --> 00:27:29,473
The only natural
general intelligence
644
00:27:29,517 --> 00:27:30,910
we know of as humans,
645
00:27:30,953 --> 00:27:33,434
we obviously learn a lot
from our environment.
646
00:27:33,477 --> 00:27:35,871
So we think that
simulated environments
647
00:27:35,915 --> 00:27:38,874
are one of the ways
to create an AGI.
648
00:27:40,528 --> 00:27:42,356
The very early humans
649
00:27:42,399 --> 00:27:44,314
were having to solve
logic problems.
650
00:27:44,358 --> 00:27:46,752
They were having to solve
navigation, memory,
651
00:27:46,795 --> 00:27:48,971
and we evolved
in that environment.
652
00:27:50,364 --> 00:27:52,496
If we can create
a virtual recreation
653
00:27:52,540 --> 00:27:54,629
of that kind of environment,
654
00:27:54,673 --> 00:27:56,326
that's the perfect
testing ground
655
00:27:56,370 --> 00:27:57,458
and training ground
656
00:27:57,501 --> 00:27:59,329
for everything
we do at DeepMind.
657
00:28:04,247 --> 00:28:05,727
What they were doing here
658
00:28:05,771 --> 00:28:09,252
was creating environments
for childlike beings,
659
00:28:09,296 --> 00:28:11,690
the agents to exist
within and play.
660
00:28:12,429 --> 00:28:13,648
That just sounded like
661
00:28:13,692 --> 00:28:16,782
the most interesting thing
in all the world.
662
00:28:16,825 --> 00:28:19,132
A child
learns by tearing things up
663
00:28:19,175 --> 00:28:20,655
and then throwing food around
664
00:28:20,699 --> 00:28:23,353
and getting a response
from mommy or daddy.
665
00:28:23,397 --> 00:28:25,616
This seems like an important
idea to incorporate
666
00:28:25,660 --> 00:28:27,880
in the way you train an agent.
667
00:28:27,923 --> 00:28:30,970
The humanoid
is supposed to stand up.
668
00:28:31,013 --> 00:28:32,972
As his center
of gravity rises,
669
00:28:33,015 --> 00:28:34,408
it gets more points.
670
00:28:37,454 --> 00:28:38,804
You have a reward
671
00:28:38,847 --> 00:28:40,849
and the agent
learns from the reward,
672
00:28:40,893 --> 00:28:43,373
like, you do something well,
you get a positive reward.
673
00:28:43,417 --> 00:28:47,508
You do something bad,
you get a negative reward.
674
00:28:47,551 --> 00:28:49,379
It looks like it's standing.
675
00:28:50,859 --> 00:28:52,165
It's still a bit drunk.
676
00:28:52,208 --> 00:28:53,644
It likes to walk backwards.
677
00:28:53,688 --> 00:28:55,342
Yeah.
678
00:28:55,385 --> 00:28:57,518
The whole algorithm
is trying to optimize
679
00:28:57,561 --> 00:28:59,694
for receiving as much rewards
as possible,
680
00:28:59,738 --> 00:29:02,305
and it's found that
walking backwards,
681
00:29:02,349 --> 00:29:05,352
it's good enough
to get very good scores.
682
00:29:07,746 --> 00:29:09,573
When we learn to navigate,
683
00:29:09,617 --> 00:29:11,271
when we learn to get around
in our world,
684
00:29:11,314 --> 00:29:13,490
we don't start with maps.
685
00:29:13,534 --> 00:29:16,319
We just start
with our own exploration,
686
00:29:16,363 --> 00:29:18,017
adventuring off
across the park,
687
00:29:18,060 --> 00:29:21,890
without our parents
by our side,
688
00:29:21,934 --> 00:29:24,327
or finding our way home
from school when we're young.
689
00:29:26,939 --> 00:29:28,723
A few of us
came up with this idea
690
00:29:28,767 --> 00:29:31,987
that if we had an environment
where a simulated robot
691
00:29:32,031 --> 00:29:33,728
just had to run forward,
692
00:29:33,772 --> 00:29:36,296
we could put all sorts of
obstacles in its way
693
00:29:36,339 --> 00:29:38,211
and see if it could manage
to navigate
694
00:29:38,254 --> 00:29:40,474
different types of terrain.
695
00:29:40,517 --> 00:29:43,042
The idea would be like
a parkour challenge.
696
00:29:46,393 --> 00:29:48,743
It's not graceful,
697
00:29:48,787 --> 00:29:51,833
but was never trained to hold
a glass whilst it was running
698
00:29:51,877 --> 00:29:53,052
and not spill water.
699
00:29:54,183 --> 00:29:55,706
You set this objective
that says,
700
00:29:55,750 --> 00:29:58,231
"Just move forward,
forward velocity,
701
00:29:58,274 --> 00:30:00,581
"and you'll get
a reward for that."
702
00:30:00,624 --> 00:30:02,670
And the learning algorithm
figures out
703
00:30:02,713 --> 00:30:05,412
how to move
this complex set of joints.
704
00:30:06,152 --> 00:30:07,370
That's the power of
705
00:30:07,414 --> 00:30:10,069
reward-based
reinforcement learning.
706
00:30:10,112 --> 00:30:12,767
Our goal
is to try and build agents
707
00:30:12,811 --> 00:30:15,770
which, we drop them in,
they know nothing,
708
00:30:15,814 --> 00:30:18,381
they get to play around in
whatever problem you give them
709
00:30:18,425 --> 00:30:22,342
and eventually figure out how
to solve it for themselves.
710
00:30:22,385 --> 00:30:24,823
Now we want something
which can do that
711
00:30:24,866 --> 00:30:27,390
in as many different types
of problems as possible.
712
00:30:29,262 --> 00:30:32,918
A human needs diverse skills
to interact with the world.
713
00:30:32,961 --> 00:30:35,050
How to deal
with complex images,
714
00:30:35,094 --> 00:30:37,879
how to manipulate
thousands of things at once,
715
00:30:37,923 --> 00:30:40,447
how to deal
with missing information.
716
00:30:40,490 --> 00:30:42,014
We think all of these things
together
717
00:30:42,057 --> 00:30:45,278
are represented
by this game calledStarCraft.
718
00:30:45,321 --> 00:30:47,280
All it's being trained
to do is,
719
00:30:47,323 --> 00:30:50,457
given this situation,
this screen,
720
00:30:50,500 --> 00:30:51,850
what would a human do?
721
00:30:51,893 --> 00:30:55,244
We took inspiration from
large language models
722
00:30:55,288 --> 00:30:57,638
where you simply train
a model
723
00:30:57,681 --> 00:30:59,553
to predict the next word,
724
00:31:03,731 --> 00:31:05,472
which is exactly the same as
725
00:31:05,515 --> 00:31:07,735
predict the next
StarCraft move.
726
00:31:07,778 --> 00:31:08,954
Unlike chess or Go,
727
00:31:08,997 --> 00:31:11,217
where players take turns
to make moves,
728
00:31:11,260 --> 00:31:14,133
inStarCraft there's a
continuous flow of decisions.
729
00:31:15,134 --> 00:31:16,483
On top of that,
730
00:31:16,526 --> 00:31:18,659
you can't even see
what the opponent is doing.
731
00:31:18,702 --> 00:31:20,879
There is no longer
a clear definition
732
00:31:20,922 --> 00:31:22,315
of what it means
to play the best way.
733
00:31:22,358 --> 00:31:23,925
It depends on
what your opponent does.
734
00:31:23,969 --> 00:31:25,492
This is the way
that we'll get to
735
00:31:25,535 --> 00:31:27,494
a much more fluid,
736
00:31:27,537 --> 00:31:31,672
more natural, faster,
more reactive agent.
737
00:31:31,715 --> 00:31:33,021
This is a huge challenge
738
00:31:33,065 --> 00:31:35,415
and let's see how far
we can push.
739
00:31:35,458 --> 00:31:36,459
Oh!
740
00:31:36,503 --> 00:31:38,200
Holy monkey!
741
00:31:38,244 --> 00:31:40,376
I'm a pretty
low-level amateur.
742
00:31:40,420 --> 00:31:42,857
I'm okay, but I'm
a pretty low-level amateur.
743
00:31:42,901 --> 00:31:45,991
These agents have
a long ways to go.
744
00:31:46,034 --> 00:31:48,515
We couldn't
beat someone of Tim's level.
745
00:31:48,558 --> 00:31:50,430
You know, that was
a little bit alarming.
746
00:31:50,473 --> 00:31:51,997
At that point, it felt like
747
00:31:52,040 --> 00:31:53,520
it was going to be, like,
a really big long challenge,
748
00:31:53,563 --> 00:31:55,000
maybe a couple of years.
749
00:31:58,177 --> 00:32:01,832
Dani is the best
DeepMind StarCraft 2 player.
750
00:32:01,876 --> 00:32:05,097
I've been playing the agent
every day for a few weeks now.
751
00:32:07,186 --> 00:32:08,883
I could feel that the agent
752
00:32:08,927 --> 00:32:11,103
was getting better
really fast.
753
00:32:13,409 --> 00:32:15,020
Wow, we beat Danny.
That, for me,
754
00:32:15,063 --> 00:32:16,935
was already
like a huge achievement.
755
00:32:18,284 --> 00:32:19,372
The next step is
756
00:32:19,415 --> 00:32:21,722
we're going to book in
a pro to play.
757
00:32:42,003 --> 00:32:44,049
It feels a bit unfair.
All you guys against me.
758
00:32:45,615 --> 00:32:46,965
We're way
ahead of what I thought
759
00:32:47,008 --> 00:32:49,402
we would do, given where
we were two months ago.
760
00:32:49,445 --> 00:32:50,794
Just trying to digest it all,
actually.
761
00:32:50,838 --> 00:32:52,753
But it's very, very cool.
762
00:32:52,796 --> 00:32:54,146
Now we're in
a position where
763
00:32:54,189 --> 00:32:56,061
we can finally share
the work that we've done
764
00:32:56,104 --> 00:32:57,192
with the public.
765
00:32:57,236 --> 00:32:58,585
This is a big step.
766
00:32:58,628 --> 00:33:00,674
We are really putting
ourselves on the line here.
767
00:33:00,717 --> 00:33:02,589
- Take it away. Cheers.
- Thank you.
768
00:33:02,632 --> 00:33:04,460
We're going to be live
from London.
769
00:33:04,504 --> 00:33:05,722
It's happening.
770
00:33:08,638 --> 00:33:10,597
Welcome to London.
771
00:33:10,640 --> 00:33:13,252
We are going to have
a live exhibition match,
772
00:33:13,295 --> 00:33:15,341
MaNa against AlphaStar.
773
00:33:18,344 --> 00:33:19,998
At this point now,
774
00:33:20,041 --> 00:33:23,827
AlphaStar, 10 and 0
against professional gamers.
775
00:33:23,871 --> 00:33:25,873
Any thoughts
before we get into this game?
776
00:33:25,916 --> 00:33:27,483
I just want to see
a good game, yeah.
777
00:33:27,527 --> 00:33:28,963
I want to see a good game.
778
00:33:29,007 --> 00:33:30,486
Absolutely,
good game. We're all excited.
779
00:33:30,530 --> 00:33:33,011
All right. Let's
see what MaNa can pull off.
780
00:33:34,969 --> 00:33:36,405
AlphaStar is definitely
781
00:33:36,449 --> 00:33:38,364
dominating the pace
of this game.
782
00:33:41,149 --> 00:33:44,152
Wow. AlphaStar
is playing so smartly.
783
00:33:46,850 --> 00:33:48,461
This really looks like
I'm watching
784
00:33:48,504 --> 00:33:49,940
a professional human gamer
785
00:33:49,984 --> 00:33:51,290
from the AlphaStar
point of view.
786
00:33:57,252 --> 00:34:01,952
I hadn't really seen
a pro playStarCraft up close,
787
00:34:01,996 --> 00:34:03,563
and the 800 clicks per minute.
788
00:34:03,606 --> 00:34:06,000
I don't understand how anyone
can even click 800 times,
789
00:34:06,044 --> 00:34:09,482
let alone doing
800 useful clicks.
790
00:34:09,525 --> 00:34:11,092
Oh, another good hit.
791
00:34:11,136 --> 00:34:13,094
- AlphaStar is just
792
00:34:13,138 --> 00:34:14,617
completely relentless.
793
00:34:14,661 --> 00:34:16,141
We need to be careful
794
00:34:16,184 --> 00:34:19,361
because many of us grew up
as gamers and are gamers.
795
00:34:19,405 --> 00:34:21,363
And so to us,
it's very natural
796
00:34:21,407 --> 00:34:23,800
to view games
as what they are,
797
00:34:23,844 --> 00:34:26,977
which is pure vehicles
for fun,
798
00:34:27,021 --> 00:34:29,719
and not to see
that more militaristic side
799
00:34:29,763 --> 00:34:32,853
that the public might see
if they looked at this.
800
00:34:32,896 --> 00:34:37,553
You can't look at gunpowder
and only make a firecracker.
801
00:34:37,597 --> 00:34:41,209
All technologies inherently
point into certain directions.
802
00:34:43,124 --> 00:34:44,691
I'm very worried about
803
00:34:44,734 --> 00:34:46,606
the certain ways in which AI
804
00:34:46,649 --> 00:34:49,652
will be used
for military purposes.
805
00:34:51,306 --> 00:34:55,049
And that makes it even clearer
how important it is
806
00:34:55,093 --> 00:34:58,357
for our societies
to be in control
807
00:34:58,400 --> 00:35:01,055
of these new technologies.
808
00:35:01,099 --> 00:35:05,190
The potential for abuse
from AI will be significant.
809
00:35:05,233 --> 00:35:08,758
Wars that occur faster
than humans can comprehend
810
00:35:08,802 --> 00:35:11,065
and more powerful
surveillance.
811
00:35:12,458 --> 00:35:15,765
How do you keep power forever
812
00:35:15,809 --> 00:35:19,421
over something that's
much more powerful than you?
813
00:35:43,053 --> 00:35:45,752
Technologies can be used
to do terrible things.
814
00:35:47,493 --> 00:35:50,452
And technology can be used
to do wonderful things
815
00:35:50,496 --> 00:35:52,150
and solve
all kinds of problems.
816
00:35:53,586 --> 00:35:54,978
When DeepMind
was acquired by Google...
817
00:35:55,022 --> 00:35:56,589
- Yeah.
- ...you got Google to promise
818
00:35:56,632 --> 00:35:58,025
that technology you developed
won't be used by the military
819
00:35:58,068 --> 00:35:59,418
- for surveillance.
- Right.
820
00:35:59,461 --> 00:36:00,593
- Yes.
- Tell us about that.
821
00:36:00,636 --> 00:36:03,161
I think technology
is neutral in itself,
822
00:36:03,204 --> 00:36:05,598
um, but how, you know,
we as a society
823
00:36:05,641 --> 00:36:07,382
or humans and companies
and other things,
824
00:36:07,426 --> 00:36:09,515
other entities and governments
decide to use it
825
00:36:09,558 --> 00:36:12,561
is what determines whether
things become good or bad.
826
00:36:12,605 --> 00:36:16,261
You know, I personally think
having autonomous weaponry
827
00:36:16,304 --> 00:36:17,479
is just a very bad idea.
828
00:36:19,177 --> 00:36:21,266
AlphaStar is playing
829
00:36:21,309 --> 00:36:24,094
an extremely intelligent game
right now.
830
00:36:24,138 --> 00:36:27,359
There is an element to
what's being created
831
00:36:27,402 --> 00:36:28,882
at DeepMind in London
832
00:36:28,925 --> 00:36:34,148
that does seem like
the Manhattan Project.
833
00:36:34,192 --> 00:36:37,586
There's a relationship between
Robert Oppenheimer
834
00:36:37,630 --> 00:36:39,675
and Demis Hassabis
835
00:36:39,719 --> 00:36:44,202
in which they're unleashing
a new force upon humanity.
836
00:36:44,245 --> 00:36:46,204
MaNa is fighting back, though.
837
00:36:46,247 --> 00:36:48,162
Oh, man!
838
00:36:48,206 --> 00:36:50,208
I think that Oppenheimer
839
00:36:50,251 --> 00:36:52,471
and some of the other leaders
of that project got caught up
840
00:36:52,514 --> 00:36:54,908
in the excitement
of building the technology
841
00:36:54,951 --> 00:36:56,170
and seeing if it was possible.
842
00:36:56,214 --> 00:36:58,520
Where is AlphaStar?
843
00:36:58,564 --> 00:36:59,782
Where is AlphaStar?
844
00:36:59,826 --> 00:37:01,958
I don't see AlphaStar's units
anywhere.
845
00:37:02,002 --> 00:37:03,525
They did not think
carefully enough
846
00:37:03,569 --> 00:37:07,312
about the morals of what
they were doing early enough.
847
00:37:07,355 --> 00:37:08,965
What we should do
as scientists
848
00:37:09,009 --> 00:37:11,011
with powerful new technologies
849
00:37:11,054 --> 00:37:13,883
is try and understand it in
controlled conditions first.
850
00:37:14,928 --> 00:37:16,799
And that is that.
851
00:37:16,843 --> 00:37:19,411
MaNa has defeated AlphaStar.
852
00:37:29,551 --> 00:37:31,336
I mean, my honest feeling is
that I think it is
853
00:37:31,379 --> 00:37:33,207
a fair representation
of where we are.
854
00:37:33,251 --> 00:37:35,949
And I think that part feels...
feels okay.
855
00:37:35,992 --> 00:37:37,429
- I'm very happy for you.
- I'm happy.
856
00:37:37,472 --> 00:37:38,865
So well... well done.
857
00:37:38,908 --> 00:37:40,867
My view is that the approach
to building technology
858
00:37:40,910 --> 00:37:43,348
which is embodied by
move fast and break things,
859
00:37:43,391 --> 00:37:46,220
is exactly what
we should not be doing,
860
00:37:46,264 --> 00:37:47,961
because you can't afford
to break things
861
00:37:48,004 --> 00:37:49,049
and then fix them afterwards.
862
00:37:49,092 --> 00:37:50,398
- Cheers.
- Thank you so much.
863
00:37:50,442 --> 00:37:52,008
Yeah, get... get some rest.
You did really well.
864
00:37:52,052 --> 00:37:53,923
- Cheers, yeah?
- Thank you for having us.
865
00:38:04,238 --> 00:38:05,500
When I was eight,
866
00:38:05,544 --> 00:38:06,849
I bought my first computer
867
00:38:06,893 --> 00:38:09,548
with the winnings
from a chess tournament.
868
00:38:09,591 --> 00:38:11,158
I sort of had this intuition
869
00:38:11,201 --> 00:38:13,726
that computers
are this magical device
870
00:38:13,769 --> 00:38:15,902
that can extend
the power of the mind.
871
00:38:15,945 --> 00:38:17,382
I had a couple
of school friends,
872
00:38:17,425 --> 00:38:19,166
and we used to have
a hacking club,
873
00:38:19,209 --> 00:38:21,908
writing code, making games.
874
00:38:26,260 --> 00:38:27,827
And then over
the summer holidays,
875
00:38:27,870 --> 00:38:29,219
I'd spend the whole day
876
00:38:29,263 --> 00:38:31,526
flicking through
games magazines.
877
00:38:31,570 --> 00:38:33,441
And one day I noticed
there was a competition
878
00:38:33,485 --> 00:38:35,878
to write an original version
of Space Invaders.
879
00:38:35,922 --> 00:38:39,621
And the winner won a job
at Bullfrog.
880
00:38:39,665 --> 00:38:42,320
Bullfrog at the time was the
best game development house
881
00:38:42,363 --> 00:38:43,756
in all of Europe.
882
00:38:43,799 --> 00:38:45,279
You know, I really wanted
to work at this place
883
00:38:45,323 --> 00:38:48,587
and see how they build games.
884
00:38:48,630 --> 00:38:50,415
Bullfrog,
based here in Guildford,
885
00:38:50,458 --> 00:38:52,286
began with a big idea.
886
00:38:52,330 --> 00:38:54,680
That idea turned into the game
Populous,
887
00:38:54,723 --> 00:38:56,551
which became
a global bestseller.
888
00:38:56,595 --> 00:38:59,859
In the '90s, there was
no recruitment agencies.
889
00:38:59,902 --> 00:39:02,122
You couldn't go out and say,
you know,
890
00:39:02,165 --> 00:39:04,951
"Come and work
in the games industry."
891
00:39:04,994 --> 00:39:08,171
It was still not even
considered an industry.
892
00:39:08,215 --> 00:39:11,218
So we came up with the idea
to have a competition
893
00:39:11,261 --> 00:39:13,655
and we got
a lot of applicants.
894
00:39:14,700 --> 00:39:17,616
And one of those was Demis's.
895
00:39:17,659 --> 00:39:20,706
I can still remember clearly
896
00:39:20,749 --> 00:39:23,970
the day that Demis came in.
897
00:39:24,013 --> 00:39:27,147
He walked in the door,
he looked about 12.
898
00:39:28,670 --> 00:39:30,019
I thought, "Oh, my God,
899
00:39:30,063 --> 00:39:31,586
"what the hell are we going
to do with this guy?"
900
00:39:31,630 --> 00:39:32,979
I applied to Cambridge.
901
00:39:33,022 --> 00:39:35,373
I got in but they said
I was way too young.
902
00:39:35,416 --> 00:39:37,853
So...
So I needed to take a year off
903
00:39:37,897 --> 00:39:39,899
so I'd be at least 17
before I got there.
904
00:39:39,942 --> 00:39:42,771
And that's when I decided
to spend that entire gap year
905
00:39:42,815 --> 00:39:44,469
working at Bullfrog.
906
00:39:44,512 --> 00:39:46,166
They couldn't even
legally employ me,
907
00:39:46,209 --> 00:39:48,298
so I ended up being paid
in brown paper envelopes.
908
00:39:50,823 --> 00:39:54,304
I got a feeling of being
really at the cutting edge
909
00:39:54,348 --> 00:39:58,047
and how much fun that was
to invent things every day.
910
00:39:58,091 --> 00:40:00,572
And then you know,
a few months later,
911
00:40:00,615 --> 00:40:03,662
maybe everyone... a million
people will be playing it.
912
00:40:03,705 --> 00:40:06,665
In those days
computer games had to evolve.
913
00:40:06,708 --> 00:40:08,536
There had to be new genres
914
00:40:08,580 --> 00:40:11,757
which were more
than just shooting things.
915
00:40:11,800 --> 00:40:14,063
Wouldn't it be amazing
to have a game
916
00:40:14,107 --> 00:40:18,807
where you design and build
your own theme park?
917
00:40:22,594 --> 00:40:25,945
Demis and I started to talk
aboutTheme Park.
918
00:40:25,988 --> 00:40:28,904
It allows the player
to build a world
919
00:40:28,948 --> 00:40:31,864
and see the consequences
of your choices
920
00:40:31,907 --> 00:40:34,127
that you've made
in that world.
921
00:40:34,170 --> 00:40:36,085
A human player
set out the layout
922
00:40:36,129 --> 00:40:38,566
of the theme park and designed
the roller coaster
923
00:40:38,610 --> 00:40:41,351
and set the prices
in the chip shop.
924
00:40:41,395 --> 00:40:43,615
What I was working on was
the behaviors of the people.
925
00:40:43,658 --> 00:40:45,138
They were autonomous
926
00:40:45,181 --> 00:40:47,314
and that was the AI
in this case.
927
00:40:47,357 --> 00:40:48,881
So what I was trying to do
was mimic
928
00:40:48,924 --> 00:40:51,013
interesting human behavior
929
00:40:51,057 --> 00:40:52,319
so that the simulation
would be
930
00:40:52,362 --> 00:40:54,582
more interesting
to interact with.
931
00:40:54,626 --> 00:40:56,541
Demis worked
on ridiculous things,
932
00:40:56,584 --> 00:40:59,413
like you could place down
these shops
933
00:40:59,457 --> 00:41:03,591
and if you put a shop too near
a very dangerous ride,
934
00:41:03,635 --> 00:41:05,375
then people on the ride
would throw up
935
00:41:05,419 --> 00:41:08,030
because they'd just eaten.
936
00:41:08,074 --> 00:41:09,641
And then that would make
other people throw up
937
00:41:09,684 --> 00:41:12,121
when they saw the throwing-up
on the floor,
938
00:41:12,165 --> 00:41:14,559
so you then had to have
lots of sweepers
939
00:41:14,602 --> 00:41:17,823
to quickly sweep it up
before the people saw it.
940
00:41:17,866 --> 00:41:19,520
That's the cool thing
about it.
941
00:41:19,564 --> 00:41:22,784
You as the player tinker with
it and then it reacts to you.
942
00:41:22,828 --> 00:41:25,874
All those nuanced
simulation things he did
943
00:41:25,918 --> 00:41:28,094
and that was an invention
944
00:41:28,137 --> 00:41:31,227
which never really
existed before.
945
00:41:31,271 --> 00:41:34,230
It was
unbelievably successful.
946
00:41:34,274 --> 00:41:35,710
Theme Park actually turned out
947
00:41:35,754 --> 00:41:37,190
to be a top ten title
948
00:41:37,233 --> 00:41:39,932
and that was the first time
we were starting to see
949
00:41:39,975 --> 00:41:43,022
how AI could make
a difference.
950
00:41:46,155 --> 00:41:47,592
We were doing
some Christmas shopping
951
00:41:47,635 --> 00:41:51,247
and were waiting for the taxi
to take us home.
952
00:41:51,291 --> 00:41:54,947
I have this very clear memory
of Demis talking about AI
953
00:41:54,990 --> 00:41:56,209
in a very different way,
954
00:41:56,252 --> 00:41:58,428
in a way that we didn't
commonly talk about.
955
00:41:58,472 --> 00:42:02,345
This idea of AI being useful
for other things
956
00:42:02,389 --> 00:42:04,086
other than entertainment.
957
00:42:04,130 --> 00:42:07,437
So being useful for, um,
helping the world
958
00:42:07,481 --> 00:42:10,310
and the potential of AI
to change the world.
959
00:42:10,353 --> 00:42:13,226
I just said to Demis,
"What is it you want to do?"
960
00:42:13,269 --> 00:42:14,532
And he said to me,
961
00:42:14,575 --> 00:42:16,795
"I want to be the person
that solves AI."
962
00:42:22,670 --> 00:42:25,760
Peter offered me £1 million
963
00:42:25,804 --> 00:42:27,675
to not go to university.
964
00:42:30,199 --> 00:42:32,593
But I had a plan
from the beginning.
965
00:42:32,637 --> 00:42:35,814
And my plan was always
to go to Cambridge.
966
00:42:35,857 --> 00:42:36,902
I think a lot of
my schoolfriends
967
00:42:36,945 --> 00:42:38,033
thought I was mad.
968
00:42:38,077 --> 00:42:39,252
Why would you not...
969
00:42:39,295 --> 00:42:40,688
I mean, £1 million,
that's a lot of money.
970
00:42:40,732 --> 00:42:43,517
In the '90s,
that is a lot of money, right?
971
00:42:43,561 --> 00:42:46,346
For a...
For a poor 17-year-old kid.
972
00:42:46,389 --> 00:42:50,219
He's like this little seed
that's going to burst through,
973
00:42:50,263 --> 00:42:53,658
and he's not going to be able
to do that at Bullfrog.
974
00:42:56,443 --> 00:42:59,098
I had to drop him off
at the train station
975
00:42:59,141 --> 00:43:02,580
and I can still see
that picture
976
00:43:02,623 --> 00:43:07,019
of this little elven character
disappear down that tunnel.
977
00:43:07,062 --> 00:43:09,804
That was an incredibly
sad moment.
978
00:43:13,242 --> 00:43:14,635
I had this romantic ideal
979
00:43:14,679 --> 00:43:16,942
of what Cambridge
would be like,
980
00:43:16,985 --> 00:43:18,639
1,000 years of history,
981
00:43:18,683 --> 00:43:21,033
walking the same streets
that Turing,
982
00:43:21,076 --> 00:43:23,601
Newton and Crick had walked.
983
00:43:23,644 --> 00:43:26,647
I wanted to explore
the edge of the universe.
984
00:43:29,084 --> 00:43:30,346
When I got to Cambridge,
985
00:43:30,390 --> 00:43:32,653
I'd basically been working
my whole life.
986
00:43:33,741 --> 00:43:35,090
Every single summer,
987
00:43:35,134 --> 00:43:37,136
I was either playing chess
professionally,
988
00:43:37,179 --> 00:43:39,704
or I was working,
doing an internship.
989
00:43:39,747 --> 00:43:43,708
So I was, like, "Right,
I am gonna have fun now
990
00:43:43,751 --> 00:43:46,711
"and explore what it means
to be a normal teenager."
991
00:43:50,236 --> 00:43:52,238
Come on! Go, boy, go!
992
00:43:52,281 --> 00:43:54,022
It was work hard
and play hard.
993
00:43:55,850 --> 00:43:57,025
I first met Demis
994
00:43:57,069 --> 00:43:59,201
because we both attended
Queens' College.
995
00:44:00,115 --> 00:44:01,203
Our group of friends,
996
00:44:01,247 --> 00:44:03,205
we'd often drink beer
in the bar,
997
00:44:03,249 --> 00:44:04,946
play table football.
998
00:44:04,990 --> 00:44:07,340
In the bar,
I used to play speed chess,
999
00:44:07,383 --> 00:44:09,255
pieces flying off the board,
1000
00:44:09,298 --> 00:44:11,083
you know, the whole game
in one minute.
1001
00:44:11,126 --> 00:44:12,301
Demis sat down opposite me.
1002
00:44:12,345 --> 00:44:13,563
And I looked at him
and I thought,
1003
00:44:13,607 --> 00:44:15,217
"I remember you
from when we were kids."
1004
00:44:15,261 --> 00:44:17,176
I had actually been
in the same chess tournament
1005
00:44:17,219 --> 00:44:18,786
as Dave in Ipswich,
1006
00:44:18,830 --> 00:44:20,440
where I used to go and try
and raid his local chess club
1007
00:44:20,483 --> 00:44:22,703
to win a bit of prize money.
1008
00:44:22,747 --> 00:44:24,618
We were studying
computer science.
1009
00:44:24,662 --> 00:44:26,794
Some people,
who at the age of 17
1010
00:44:26,838 --> 00:44:28,404
would have come in and made
sure to tell everybody
1011
00:44:28,448 --> 00:44:29,492
everything about themselves.
1012
00:44:29,536 --> 00:44:30,972
"Hey, I worked at Bullfrog
1013
00:44:31,016 --> 00:44:33,018
"and built the world's
most successful video game."
1014
00:44:33,061 --> 00:44:34,715
But he wasn't like that
at all.
1015
00:44:34,759 --> 00:44:36,412
At Cambridge,
Demis and myself
1016
00:44:36,456 --> 00:44:38,414
both had an interest
in computational neuroscience
1017
00:44:38,458 --> 00:44:40,242
and trying to understand
how computers and brains
1018
00:44:40,286 --> 00:44:42,636
intertwined
and linked together.
1019
00:44:42,680 --> 00:44:44,290
Both David and Demis
1020
00:44:44,333 --> 00:44:46,422
came to me for supervisions.
1021
00:44:46,466 --> 00:44:49,774
It happens just by coincidence
that the year 1997,
1022
00:44:49,817 --> 00:44:51,645
their third and final year
at Cambridge,
1023
00:44:51,689 --> 00:44:55,301
was also the year when
the first chess grandmaster
1024
00:44:55,344 --> 00:44:56,781
was beaten by
a computer program.
1025
00:44:58,304 --> 00:45:00,088
Round one today
of a chess match
1026
00:45:00,132 --> 00:45:03,701
between the ranking
world champion Garry Kasparov
1027
00:45:03,744 --> 00:45:06,007
and an opponent named
Deep Blue
1028
00:45:06,051 --> 00:45:10,490
to test to see if the human
brain can outwit a machine.
1029
00:45:10,533 --> 00:45:11,621
I remember the drama
1030
00:45:11,665 --> 00:45:13,798
of Kasparov
losing the last match.
1031
00:45:13,841 --> 00:45:15,234
Whoa!
1032
00:45:15,277 --> 00:45:17,192
Kasparov has resigned!
1033
00:45:17,236 --> 00:45:19,586
When Deep Blue
beat Garry Kasparov,
1034
00:45:19,629 --> 00:45:21,457
that was a real
watershed event.
1035
00:45:21,501 --> 00:45:23,155
My main memory of it was
1036
00:45:23,198 --> 00:45:25,331
I wasn't that impressed
with Deep Blue.
1037
00:45:25,374 --> 00:45:27,246
I was more impressed
with Kasparov's mind.
1038
00:45:27,289 --> 00:45:29,509
That he could play chess
to this level,
1039
00:45:29,552 --> 00:45:31,641
where he could compete
on an equal footing
1040
00:45:31,685 --> 00:45:33,252
with the brute of a machine,
1041
00:45:33,295 --> 00:45:35,167
but of course, Kasparov can do
1042
00:45:35,210 --> 00:45:36,951
everything else humans can do,
too.
1043
00:45:36,995 --> 00:45:38,257
It was a huge achievement.
1044
00:45:38,300 --> 00:45:39,388
But the truth
of the matter was,
1045
00:45:39,432 --> 00:45:40,868
Deep Blue
could only play chess.
1046
00:45:42,435 --> 00:45:44,524
What we would regard
as intelligence
1047
00:45:44,567 --> 00:45:46,874
was missing from that system.
1048
00:45:46,918 --> 00:45:49,834
This idea of generality
and also learning.
1049
00:45:53,751 --> 00:45:55,404
Cambridge was amazing,
because of course, you know,
1050
00:45:55,448 --> 00:45:56,666
you're mixing with people
1051
00:45:56,710 --> 00:45:58,233
who are studying
many different subjects.
1052
00:45:58,277 --> 00:46:01,410
There were scientists,
philosophers, artists...
1053
00:46:01,454 --> 00:46:04,457
...geologists,
biologists, ecologists.
1054
00:46:04,500 --> 00:46:07,416
You know, everybody is talking
about everything all the time.
1055
00:46:07,460 --> 00:46:10,768
I was obsessed with
the protein folding problem.
1056
00:46:10,811 --> 00:46:13,248
Tim Stevens used
to talk obsessively,
1057
00:46:13,292 --> 00:46:15,381
almost like religiously
about this problem,
1058
00:46:15,424 --> 00:46:17,165
protein folding problem.
1059
00:46:17,209 --> 00:46:18,863
Proteins are, you know,
1060
00:46:18,906 --> 00:46:22,083
one of the most beautiful and
elegant things about biology.
1061
00:46:22,127 --> 00:46:24,738
They are the machines of life.
1062
00:46:24,782 --> 00:46:27,001
They build everything,
they control everything,
1063
00:46:27,045 --> 00:46:29,569
they're why biology works.
1064
00:46:29,612 --> 00:46:32,659
Proteins are made from strings
of amino acids
1065
00:46:32,702 --> 00:46:37,055
that fold up to create
a protein structure.
1066
00:46:37,098 --> 00:46:39,884
If we can predict
the structure of proteins
1067
00:46:39,927 --> 00:46:43,104
from just their amino acid
sequences,
1068
00:46:43,148 --> 00:46:46,107
then a new protein
to cure cancer
1069
00:46:46,151 --> 00:46:49,415
or break down plastic
to help the environment
1070
00:46:49,458 --> 00:46:50,808
is definitely something
1071
00:46:50,851 --> 00:46:52,940
that you could begin
to think about.
1072
00:46:53,941 --> 00:46:55,029
I kind of thought,
1073
00:46:55,073 --> 00:46:58,293
"Well, is a human being
clever enough
1074
00:46:58,337 --> 00:46:59,991
"to actually fold a protein?"
1075
00:47:00,034 --> 00:47:02,080
We can't work it out.
1076
00:47:02,123 --> 00:47:04,082
Since the 1960s,
1077
00:47:04,125 --> 00:47:05,953
we thought that in principle,
1078
00:47:05,997 --> 00:47:08,913
if I know what the amino acid
sequence of a protein is,
1079
00:47:08,956 --> 00:47:11,437
I should be able to compute
what the structure's like.
1080
00:47:11,480 --> 00:47:13,700
So, if you could
just press a button,
1081
00:47:13,743 --> 00:47:16,311
and they'd all come
popping out, that would be...
1082
00:47:16,355 --> 00:47:18,009
that would have some impact.
1083
00:47:20,272 --> 00:47:21,577
It stuck in my mind.
1084
00:47:21,621 --> 00:47:23,405
"Oh, this is
a very interesting problem."
1085
00:47:23,449 --> 00:47:26,844
And it felt to me
like it would be solvable.
1086
00:47:26,887 --> 00:47:29,934
But I thought
it would need AI to do it.
1087
00:47:31,500 --> 00:47:34,025
If we could just solve
protein folding,
1088
00:47:34,068 --> 00:47:35,635
it could change the world.
1089
00:47:50,868 --> 00:47:52,826
Ever since
I was a student at Cambridge,
1090
00:47:54,001 --> 00:47:55,611
I've never
stopped thinking about
1091
00:47:55,655 --> 00:47:57,135
the protein folding problem.
1092
00:47:59,789 --> 00:48:02,792
If you were
to solve protein folding,
1093
00:48:02,836 --> 00:48:05,665
then the potential
to help solve problems like
1094
00:48:05,708 --> 00:48:09,669
Alzheimer's, dementia
and drug discovery is huge.
1095
00:48:09,712 --> 00:48:11,671
Solving disease is probably
1096
00:48:11,714 --> 00:48:13,325
the most major impact
we could have.
1097
00:48:15,457 --> 00:48:16,676
Thousands of very smart people
1098
00:48:16,719 --> 00:48:18,765
have tried
to solve protein folding.
1099
00:48:18,808 --> 00:48:20,985
I just think now
is the right time
1100
00:48:21,028 --> 00:48:22,508
for AI to crack it.
1101
00:48:26,555 --> 00:48:28,383
We needed
a reasonable way
1102
00:48:28,427 --> 00:48:29,515
to apply machine learning
1103
00:48:29,558 --> 00:48:30,516
to the protein folding
problem.
1104
00:48:32,431 --> 00:48:35,173
We came across
this Foldit game.
1105
00:48:35,216 --> 00:48:38,785
The goal is to move around
this 3D model of a protein
1106
00:48:38,828 --> 00:48:41,701
and you get a score
every time you move it.
1107
00:48:41,744 --> 00:48:43,050
The more accurate
you make these structures,
1108
00:48:43,094 --> 00:48:45,531
the more useful
they will be to biologists.
1109
00:48:46,227 --> 00:48:47,489
I spent a few days
1110
00:48:47,533 --> 00:48:48,838
just kind of seeing
how well we could do.
1111
00:48:50,666 --> 00:48:52,407
We did reasonably well.
1112
00:48:52,451 --> 00:48:53,843
But even if you were
1113
00:48:53,887 --> 00:48:55,280
the world's
best Foldit player,
1114
00:48:55,323 --> 00:48:57,499
you wouldn't
solve protein folding.
1115
00:48:57,543 --> 00:48:59,501
That's why we had
to move beyond the game.
1116
00:48:59,545 --> 00:49:00,720
Games
are always just
1117
00:49:00,763 --> 00:49:03,723
the proving ground
for our algorithms.
1118
00:49:03,766 --> 00:49:07,727
The ultimate goal was not just
to crack Go and StarCraft.
1119
00:49:07,770 --> 00:49:10,034
It was to crack
real-world challenges.
1120
00:49:16,127 --> 00:49:18,520
I remember
hearing this rumor
1121
00:49:18,564 --> 00:49:21,175
that Demis was
getting into proteins.
1122
00:49:21,219 --> 00:49:23,743
I talked to some people
at DeepMind and I would ask,
1123
00:49:23,786 --> 00:49:25,049
"So are you doing
protein folding?"
1124
00:49:25,092 --> 00:49:26,920
And they would
artfully change the subject.
1125
00:49:26,964 --> 00:49:30,097
And when that happened twice,
I pretty much figured it out.
1126
00:49:30,141 --> 00:49:32,839
So I thought
I should submit a resume.
1127
00:49:32,882 --> 00:49:35,668
All right, everyone,
welcome to DeepMind.
1128
00:49:35,711 --> 00:49:37,626
I know some of you,
this may be your first week,
1129
00:49:37,670 --> 00:49:39,193
but I hope you all set...
1130
00:49:39,237 --> 00:49:40,890
The really appealing
part for me about the job
1131
00:49:40,934 --> 00:49:42,675
was this, like,
sense of connection
1132
00:49:42,718 --> 00:49:44,503
to the larger purpose.
1133
00:49:44,546 --> 00:49:45,591
If we can crack
1134
00:49:45,634 --> 00:49:48,028
some fundamental problems
in science,
1135
00:49:48,072 --> 00:49:49,160
many other people
1136
00:49:49,203 --> 00:49:50,900
and other companies
and labs and so on
1137
00:49:50,944 --> 00:49:52,467
could build
on top of our work.
1138
00:49:52,511 --> 00:49:53,816
This is your chance now
1139
00:49:53,860 --> 00:49:55,818
to add your chapter
to this story.
1140
00:49:55,862 --> 00:49:57,298
When I arrived,
1141
00:49:57,342 --> 00:49:59,605
I was definitely
quite a bit nervous.
1142
00:49:59,648 --> 00:50:00,736
I'm still trying to keep...
1143
00:50:00,780 --> 00:50:02,912
I haven't taken
any biology courses.
1144
00:50:02,956 --> 00:50:05,393
We haven't spent
years of our lives
1145
00:50:05,437 --> 00:50:07,917
looking at these structures
and understanding them.
1146
00:50:07,961 --> 00:50:09,963
We are just going off the data
1147
00:50:10,007 --> 00:50:11,269
and our machine learning
models.
1148
00:50:12,705 --> 00:50:13,836
In machine learning,
1149
00:50:13,880 --> 00:50:15,795
you train a network
like flashcards.
1150
00:50:15,838 --> 00:50:18,798
Here's the question.
Here's the answer.
1151
00:50:18,841 --> 00:50:20,756
Here's the question.
Here's the answer.
1152
00:50:20,800 --> 00:50:22,323
But in protein folding,
1153
00:50:22,367 --> 00:50:25,457
we're not doing the kind
of standard task at DeepMind
1154
00:50:25,500 --> 00:50:28,155
where you have unlimited data.
1155
00:50:28,199 --> 00:50:30,766
Your job is to get better
at chess or Go
1156
00:50:30,810 --> 00:50:32,986
and you can play
as many games of chess or Go
1157
00:50:33,030 --> 00:50:34,814
as your computers will allow.
1158
00:50:35,510 --> 00:50:36,772
With proteins,
1159
00:50:36,816 --> 00:50:39,732
we're sitting on
a very thick size of data
1160
00:50:39,775 --> 00:50:41,995
that's been determined
by a half century
1161
00:50:42,039 --> 00:50:46,478
of time-consuming experimental
methods in laboratories.
1162
00:50:46,521 --> 00:50:49,698
These painstaking methods
can take months or years
1163
00:50:49,742 --> 00:50:52,353
to determine
a single protein structure,
1164
00:50:52,397 --> 00:50:55,791
and sometimes, a structure
can never be determined.
1165
00:50:57,315 --> 00:51:00,274
That's why we're working
with such small datasets
1166
00:51:00,318 --> 00:51:02,233
to train our algorithms.
1167
00:51:02,276 --> 00:51:04,365
When DeepMind
started to explore
1168
00:51:04,409 --> 00:51:05,975
the folding problem,
1169
00:51:06,019 --> 00:51:07,977
they were talking to us about
which datasets they were using
1170
00:51:08,021 --> 00:51:09,892
and what would be
the possibilities
1171
00:51:09,936 --> 00:51:11,633
if they did
solve this problem.
1172
00:51:12,373 --> 00:51:13,940
Many people have tried,
1173
00:51:13,983 --> 00:51:16,638
and yet no one on the planet
has solved protein folding.
1174
00:51:16,682 --> 00:51:18,205
I did think to myself,
1175
00:51:18,249 --> 00:51:19,946
"Well, you know, good luck."
1176
00:51:19,989 --> 00:51:22,731
If we can solve
the protein folding problem,
1177
00:51:22,775 --> 00:51:25,691
it would have an incredible
kind of medical relevance.
1178
00:51:25,734 --> 00:51:27,736
This is the cycle of science.
1179
00:51:27,780 --> 00:51:30,043
You do a huge amount
of exploration,
1180
00:51:30,087 --> 00:51:31,914
and then you go
into exploitation mode,
1181
00:51:31,958 --> 00:51:33,568
and you focus and you see
1182
00:51:33,612 --> 00:51:35,353
how good
are those ideas, really?
1183
00:51:35,396 --> 00:51:36,528
And there's nothing better
1184
00:51:36,571 --> 00:51:38,095
than external competition
for that.
1185
00:51:39,835 --> 00:51:43,056
So we decided
to enter CASP competition.
1186
00:51:43,100 --> 00:51:47,234
CASP, we started
to try and speed up
1187
00:51:47,278 --> 00:51:49,802
the solution to
the protein folding problem.
1188
00:51:49,845 --> 00:51:52,065
CASP is when we say,
1189
00:51:52,109 --> 00:51:54,372
"Look, DeepMind
is doing protein folding,
1190
00:51:54,415 --> 00:51:55,590
"this is how good we are,
1191
00:51:55,634 --> 00:51:57,592
"and maybe it's better
than everybody else.
1192
00:51:57,636 --> 00:51:58,680
"Maybe it isn't."
1193
00:51:58,724 --> 00:52:00,073
CASP is a bit like
1194
00:52:00,117 --> 00:52:02,075
the Olympic Games
of protein folding.
1195
00:52:03,642 --> 00:52:06,035
CASP is
a community-wide assessment
1196
00:52:06,079 --> 00:52:08,125
that's held every two years.
1197
00:52:09,561 --> 00:52:10,866
Teams are given
1198
00:52:10,910 --> 00:52:14,261
the amino acid sequences
of about 100 proteins,
1199
00:52:14,305 --> 00:52:17,525
and then they try
to solve this folding problem
1200
00:52:17,569 --> 00:52:20,833
using computational methods.
1201
00:52:20,876 --> 00:52:23,401
These proteins have
already been determined
1202
00:52:23,444 --> 00:52:25,968
by experiments
in a laboratory,
1203
00:52:26,012 --> 00:52:29,276
but have not yet
been revealed publicly.
1204
00:52:29,320 --> 00:52:30,756
And these known structures
1205
00:52:30,799 --> 00:52:33,280
represent the gold standard
against which
1206
00:52:33,324 --> 00:52:36,936
all the computational
predictions will be compared.
1207
00:52:37,937 --> 00:52:39,330
We've got a score
1208
00:52:39,373 --> 00:52:42,159
that measures the accuracy
of the predictions.
1209
00:52:42,202 --> 00:52:44,726
And you would expect
a score of over 90
1210
00:52:44,770 --> 00:52:47,425
to be a solution to
the protein folding problem.
1211
00:52:48,556 --> 00:52:50,036
Welcome, everyone,
1212
00:52:50,079 --> 00:52:52,038
to our first, uh, semifinals
in the winners' bracket.
1213
00:52:52,081 --> 00:52:54,954
Nick and John
versus Demis and Frank.
1214
00:52:54,997 --> 00:52:57,217
Please join us, come around.
This will be an intense match.
1215
00:52:57,261 --> 00:52:59,306
When I learned that Demis was
1216
00:52:59,350 --> 00:53:02,048
going to tackle
the protein folding issue,
1217
00:53:02,091 --> 00:53:04,790
um, I wasn't at all surprised.
1218
00:53:04,833 --> 00:53:06,792
It's very typical of Demis.
1219
00:53:06,835 --> 00:53:08,924
You know,
he loves competition.
1220
00:53:08,968 --> 00:53:10,143
And that's the end
1221
00:53:10,187 --> 00:53:12,972
of the first game, 10-7.
1222
00:53:13,015 --> 00:53:14,103
The aim for CASP would be
1223
00:53:14,147 --> 00:53:15,975
to not just
win the competition,
1224
00:53:16,018 --> 00:53:19,892
but sort of, um,
retire the need for it.
1225
00:53:19,935 --> 00:53:23,417
So, 20 targets total
have been released by CASP.
1226
00:53:23,461 --> 00:53:24,810
We were thinking maybe
1227
00:53:24,853 --> 00:53:26,899
throw in the standard
kind of machine learning
1228
00:53:26,942 --> 00:53:28,814
and see how far
that could take us.
1229
00:53:28,857 --> 00:53:30,729
Instead of having a couple
of days on an experiment,
1230
00:53:30,772 --> 00:53:33,558
we can turn around
five experiments a day.
1231
00:53:33,601 --> 00:53:35,342
Great. Well done, everyone.
1232
00:53:36,778 --> 00:53:38,693
Can you show me the real one
instead of ours?
1233
00:53:38,737 --> 00:53:39,781
The true answer is
1234
00:53:39,825 --> 00:53:42,044
supposed to look
something like that.
1235
00:53:42,088 --> 00:53:45,047
It's a lot more
cylindrical than I thought.
1236
00:53:45,091 --> 00:53:47,398
The results
were not very good.
1237
00:53:47,441 --> 00:53:48,703
Okay.
1238
00:53:48,747 --> 00:53:49,835
We throw
all the obvious ideas to it
1239
00:53:49,878 --> 00:53:51,793
and the problem laughs at you.
1240
00:53:52,881 --> 00:53:54,361
This makes no sense.
1241
00:53:54,405 --> 00:53:56,015
We thought
we could just throw
1242
00:53:56,058 --> 00:53:58,322
some of our best algorithms
at the problem.
1243
00:53:59,366 --> 00:54:00,976
We were slightly naive.
1244
00:54:01,020 --> 00:54:02,282
We should be learning this,
1245
00:54:02,326 --> 00:54:04,284
you know,
in the blink of an eye.
1246
00:54:05,372 --> 00:54:06,982
The thing
I'm worried about is,
1247
00:54:07,026 --> 00:54:08,375
we take the field from
1248
00:54:08,419 --> 00:54:10,899
really bad answers
to moderately bad answers.
1249
00:54:10,943 --> 00:54:13,946
I feel like we need
some sort of new technology
1250
00:54:13,989 --> 00:54:15,164
for moving around
these things.
1251
00:54:20,431 --> 00:54:22,215
With only
a week left of CASP,
1252
00:54:22,259 --> 00:54:24,348
it's now a sprint
to get it deployed.
1253
00:54:26,654 --> 00:54:28,090
You've done your best.
1254
00:54:28,134 --> 00:54:29,875
Then there's
nothing more you can do
1255
00:54:29,918 --> 00:54:32,399
but wait for CASP
to deliver the results.
1256
00:54:52,593 --> 00:54:53,725
This famous thing of Einstein,
1257
00:54:53,768 --> 00:54:55,030
the last couple of years
of his life,
1258
00:54:55,074 --> 00:54:57,381
when he was here,
he overlapped with Kurt Godel
1259
00:54:57,424 --> 00:54:59,774
and he said one of the reasons
he still comes in to work
1260
00:54:59,818 --> 00:55:01,646
is so that
he gets to walk home
1261
00:55:01,689 --> 00:55:03,517
and discuss things with Godel.
1262
00:55:03,561 --> 00:55:05,911
It's a pretty big compliment
for Kurt Godel,
1263
00:55:05,954 --> 00:55:07,478
shows you how amazing he was.
1264
00:55:09,131 --> 00:55:10,568
The Institute
for Advanced Study
1265
00:55:10,611 --> 00:55:12,918
was formed in 1933.
1266
00:55:12,961 --> 00:55:14,223
In the early years,
1267
00:55:14,267 --> 00:55:16,487
the intense scientific
atmosphere attracted
1268
00:55:16,530 --> 00:55:19,359
some of the most brilliant
mathematicians and physicists
1269
00:55:19,403 --> 00:55:22,536
ever concentrated
in a single place and time.
1270
00:55:22,580 --> 00:55:24,582
The founding
principle of this place,
1271
00:55:24,625 --> 00:55:28,020
it's the idea of unfettered
intellectual pursuits,
1272
00:55:28,063 --> 00:55:30,152
even if you don't know
what you're exploring.
1273
00:55:30,196 --> 00:55:32,154
Will result
in some cool things,
1274
00:55:32,198 --> 00:55:34,896
and sometimes that then
ends up being useful,
1275
00:55:34,940 --> 00:55:36,376
which, of course,
1276
00:55:36,420 --> 00:55:37,986
is partially what I've been
trying to do at DeepMind.
1277
00:55:38,030 --> 00:55:39,988
How many big breakthroughs
do you think are required
1278
00:55:40,032 --> 00:55:41,686
to get all the way to AGI?
1279
00:55:41,729 --> 00:55:43,078
And, you know,
I estimate maybe
1280
00:55:43,122 --> 00:55:44,341
there's about
a dozen of those.
1281
00:55:44,384 --> 00:55:46,125
You know, I hope
it's within my lifetime.
1282
00:55:46,168 --> 00:55:47,605
- Yes, okay.
-But then,
1283
00:55:47,648 --> 00:55:49,171
all scientists
hope that, right?
1284
00:55:49,215 --> 00:55:51,130
Demis has
many accolades.
1285
00:55:51,173 --> 00:55:54,002
He was elected Fellow to
the Royal Society last year.
1286
00:55:54,046 --> 00:55:55,961
He is also a Fellow
of Royal Society of Arts.
1287
00:55:56,004 --> 00:55:57,528
A big hand for Demis Hassabis.
1288
00:56:04,230 --> 00:56:05,927
My dream
has always been to try
1289
00:56:05,971 --> 00:56:08,103
and make
AI-assisted science possible.
1290
00:56:08,147 --> 00:56:09,235
And what I think is
1291
00:56:09,278 --> 00:56:11,150
our most exciting project,
last year,
1292
00:56:11,193 --> 00:56:13,152
which is our work
in protein folding.
1293
00:56:13,195 --> 00:56:15,328
Uh, and we call this system
AlphaFold.
1294
00:56:15,372 --> 00:56:18,331
We entered it into CASP
and our system, uh,
1295
00:56:18,375 --> 00:56:20,507
was the most accurate,
uh, predicting structures
1296
00:56:20,551 --> 00:56:24,946
for 25 out of the 43 proteins
in the hardest category.
1297
00:56:24,990 --> 00:56:26,208
So we're state of the art,
1298
00:56:26,252 --> 00:56:27,514
but we still...
I have to make... Be clear,
1299
00:56:27,558 --> 00:56:28,559
we're still a long way from
1300
00:56:28,602 --> 00:56:30,474
solving the protein
folding problem.
1301
00:56:30,517 --> 00:56:31,866
We're working hard
on this, though,
1302
00:56:31,910 --> 00:56:33,738
and we're exploring
many other techniques.
1303
00:56:49,188 --> 00:56:50,232
Let's get started.
1304
00:56:50,276 --> 00:56:53,148
So kind of
a rapid debrief,
1305
00:56:53,192 --> 00:56:55,455
these are
our final rankings for CASP.
1306
00:56:56,500 --> 00:56:57,544
We beat the second team
1307
00:56:57,588 --> 00:57:00,155
in this competition
by nearly 50%,
1308
00:57:00,199 --> 00:57:01,592
but we've still got
a long way to go
1309
00:57:01,635 --> 00:57:04,333
before we've solved
the protein folding problem
1310
00:57:04,377 --> 00:57:07,032
in a sense that
a biologist could use it.
1311
00:57:07,075 --> 00:57:08,990
It is area of concern.
1312
00:57:11,602 --> 00:57:14,213
The quality
of predictions varied
1313
00:57:14,256 --> 00:57:16,737
and they were no more useful
than the previous methods.
1314
00:57:16,781 --> 00:57:19,914
AlphaFold didn't
produce good enough data
1315
00:57:19,958 --> 00:57:22,526
for it to be useful
in a practical way
1316
00:57:22,569 --> 00:57:24,005
to, say, somebody like me
1317
00:57:24,049 --> 00:57:28,227
investigating
my own biological problems.
1318
00:57:28,270 --> 00:57:30,316
That was kind of
a humbling moment
1319
00:57:30,359 --> 00:57:32,753
'cause we thought we'd worked
very hard and succeeded.
1320
00:57:32,797 --> 00:57:34,886
And what we'd found is
we were the best in the world
1321
00:57:34,929 --> 00:57:36,453
at a problem
the world's not good at.
1322
00:57:37,671 --> 00:57:38,933
We knew we sucked.
1323
00:57:40,457 --> 00:57:42,328
It doesn't help
if you have the tallest ladder
1324
00:57:42,371 --> 00:57:44,635
when you're going to the moon.
1325
00:57:44,678 --> 00:57:47,115
The opinion of quite
a few people on the team,
1326
00:57:47,159 --> 00:57:51,468
that this is sort of
a fool's errand in some ways.
1327
00:57:51,511 --> 00:57:54,079
And I might have been wrong
with protein folding.
1328
00:57:54,122 --> 00:57:55,559
Maybe it's too hard still
1329
00:57:55,602 --> 00:57:58,431
for where we're at
generally with AI.
1330
00:57:58,475 --> 00:58:01,173
If you want to do
biological research,
1331
00:58:01,216 --> 00:58:03,044
you have to be
prepared to fail
1332
00:58:03,088 --> 00:58:06,570
because biology
is very complicated.
1333
00:58:06,613 --> 00:58:09,790
I've run a laboratory
for nearly 50 years,
1334
00:58:09,834 --> 00:58:11,096
and half my time,
1335
00:58:11,139 --> 00:58:12,619
I'm just
an amateur psychiatrist
1336
00:58:12,663 --> 00:58:18,103
to keep, um, my colleagues
cheerful when nothing works.
1337
00:58:18,146 --> 00:58:22,542
And quite a lot of the time
and I mean, 80, 90%,
1338
00:58:22,586 --> 00:58:24,413
it does not work.
1339
00:58:24,457 --> 00:58:26,720
If you are
at the forefront of science,
1340
00:58:26,764 --> 00:58:30,115
I can tell you,
you will fail a great deal.
1341
00:58:35,163 --> 00:58:37,165
I just felt disappointed.
1342
00:58:38,689 --> 00:58:41,605
Lesson I learned is that
ambition is a good thing,
1343
00:58:41,648 --> 00:58:43,694
but you need
to get the timing right.
1344
00:58:43,737 --> 00:58:46,784
There's no point being
50 years ahead of your time.
1345
00:58:46,827 --> 00:58:48,133
You will never survive
1346
00:58:48,176 --> 00:58:49,917
fifty years of
that kind of endeavor
1347
00:58:49,961 --> 00:58:51,963
before it yields something.
1348
00:58:52,006 --> 00:58:53,268
You'll literally die trying.
1349
00:59:08,936 --> 00:59:11,286
When we talk about AGI,
1350
00:59:11,330 --> 00:59:14,376
the holy grail
of artificial intelligence,
1351
00:59:14,420 --> 00:59:15,508
it becomes really difficult
1352
00:59:15,552 --> 00:59:17,815
to know what
we're even talking about.
1353
00:59:17,858 --> 00:59:19,643
Which bits
are we gonna see today?
1354
00:59:19,686 --> 00:59:21,645
We're going
to start in the garden.
1355
00:59:23,124 --> 00:59:25,649
This is the garden looking
from the observation area.
1356
00:59:25,692 --> 00:59:27,433
Research scientists
and engineers
1357
00:59:27,476 --> 00:59:30,871
can analyze and collaborate
and evaluate
1358
00:59:30,915 --> 00:59:33,004
what's going on in real time.
1359
00:59:33,047 --> 00:59:34,614
So in the 1800s,
1360
00:59:34,658 --> 00:59:37,008
we'd think of things like
television and the submarine
1361
00:59:37,051 --> 00:59:38,139
or a rocket ship to the moon
1362
00:59:38,183 --> 00:59:40,228
and say these things
are impossible.
1363
00:59:40,272 --> 00:59:41,490
Yet Jules Verne
wrote about them and,
1364
00:59:41,534 --> 00:59:44,406
a century and a half later,
they happened.
1365
00:59:44,450 --> 00:59:45,451
We'll be
experimenting
1366
00:59:45,494 --> 00:59:47,888
on civilizations really,
1367
00:59:47,932 --> 00:59:50,587
civilizations of AI agents.
1368
00:59:50,630 --> 00:59:52,719
Once the experiments
start going,
1369
00:59:52,763 --> 00:59:54,242
it's going to be
the most exciting thing ever.
1370
00:59:54,286 --> 00:59:56,984
So how will we get sleep?
1371
00:59:57,028 --> 00:59:58,682
I won't be able to sleep.
1372
00:59:58,725 --> 01:00:00,684
Full AGI
will be able to do
1373
01:00:00,727 --> 01:00:03,861
any cognitive task
a person can do.
1374
01:00:03,904 --> 01:00:08,387
It will be at a scale,
potentially, far beyond that.
1375
01:00:08,430 --> 01:00:10,302
It's really impossible for us
1376
01:00:10,345 --> 01:00:14,828
to imagine the outputs
of a superintelligent entity.
1377
01:00:14,872 --> 01:00:18,963
It's like asking a gorilla
to imagine, you know,
1378
01:00:19,006 --> 01:00:20,181
what Einstein does
1379
01:00:20,225 --> 01:00:23,402
when he produces
the theory of relativity.
1380
01:00:23,445 --> 01:00:25,491
People often ask me
these questions like,
1381
01:00:25,534 --> 01:00:29,495
"What happens if you're wrong,
and AGI is quite far away?"
1382
01:00:29,538 --> 01:00:31,453
And I'm like,
I never worry about that.
1383
01:00:31,497 --> 01:00:33,847
I actually
worry about the reverse.
1384
01:00:33,891 --> 01:00:37,242
I actually worry
that it's coming faster
1385
01:00:37,285 --> 01:00:39,723
than we can
really prepare for.
1386
01:00:42,073 --> 01:00:45,859
It really feels
like we're in a race to AGI.
1387
01:00:45,903 --> 01:00:49,907
The prototypes and the models
that we are developing now
1388
01:00:49,950 --> 01:00:51,822
are actually transforming
1389
01:00:51,865 --> 01:00:54,215
the space of what
we know about intelligence.
1390
01:00:57,349 --> 01:00:58,785
Recently,
we've had agents
1391
01:00:58,829 --> 01:01:00,047
that are powerful enough
1392
01:01:00,091 --> 01:01:03,442
to actually start
playing games in teams,
1393
01:01:03,485 --> 01:01:06,140
then competing
against other teams.
1394
01:01:06,184 --> 01:01:08,795
We're seeing
co-operative social dynamics
1395
01:01:08,839 --> 01:01:10,492
coming out of agents
1396
01:01:10,536 --> 01:01:13,321
where we haven't
pre-programmed in
1397
01:01:13,365 --> 01:01:15,584
any of these sorts
of dynamics.
1398
01:01:15,628 --> 01:01:19,240
It's completely learned
from their own experiences.
1399
01:01:20,807 --> 01:01:23,288
When we started,
we thought we were
1400
01:01:23,331 --> 01:01:25,725
out to build
an intelligence system
1401
01:01:25,769 --> 01:01:28,336
and convince the world
that we'd done it.
1402
01:01:28,380 --> 01:01:29,947
We're now starting
to wonder whether
1403
01:01:29,990 --> 01:01:31,296
we're gonna build systems
1404
01:01:31,339 --> 01:01:32,906
that we're not convinced
are fully intelligent,
1405
01:01:32,950 --> 01:01:34,691
and we're trying to convince
the world that they're not.
1406
01:01:38,651 --> 01:01:40,000
Hi, Alpha.
1407
01:01:40,044 --> 01:01:41,523
Hello there.
1408
01:01:41,567 --> 01:01:43,917
Where are we today?
1409
01:01:43,961 --> 01:01:46,659
You're at the Museum of
Modern Art in New York City.
1410
01:01:48,400 --> 01:01:53,013
Kind of.
Um, what painting is this?
1411
01:01:53,057 --> 01:01:55,494
This isThe Creation of Adam
by Michelangelo.
1412
01:01:55,537 --> 01:01:58,410
I don't think that painting
is in New York City.
1413
01:01:58,453 --> 01:02:01,543
You are right.
It's in the Vatican City.
1414
01:02:01,587 --> 01:02:02,806
Do you think
that's where we are?
1415
01:02:04,677 --> 01:02:05,809
No.
1416
01:02:05,852 --> 01:02:07,593
Then how
am I looking at this?
1417
01:02:10,509 --> 01:02:11,640
You're looking at
1418
01:02:11,684 --> 01:02:12,946
a digital copy
of the painting.
1419
01:02:12,990 --> 01:02:13,991
Yes.
1420
01:02:15,035 --> 01:02:16,341
Who lives
in this house?
1421
01:02:19,213 --> 01:02:20,345
An astronaut.
1422
01:02:26,046 --> 01:02:28,005
What's this astronaut's name?
1423
01:02:28,048 --> 01:02:29,397
His name is Mark.
1424
01:02:29,441 --> 01:02:32,052
This is a famous equation
by Albert Einstein.
1425
01:02:32,096 --> 01:02:35,577
It shows that energy and mass
are equivalent.
1426
01:02:35,621 --> 01:02:37,666
How would you explain it
to a five year old?
1427
01:02:40,931 --> 01:02:43,542
Well, imagine
you have a big rock,
1428
01:02:43,585 --> 01:02:45,326
and you want to know
how much it weighs.
1429
01:02:45,370 --> 01:02:47,589
Is Mark in danger?
1430
01:02:47,633 --> 01:02:49,722
No.
Mark is not in danger.
1431
01:02:49,766 --> 01:02:50,767
Why is that?
1432
01:02:54,422 --> 01:02:56,163
Because the dinosaur
is not real.
1433
01:02:57,338 --> 01:02:58,470
All right.
1434
01:02:58,513 --> 01:03:00,515
Why could you
not answer my question?
1435
01:03:00,559 --> 01:03:01,908
I think I couldn't
answer your question
1436
01:03:01,952 --> 01:03:04,041
because I don't know
enough about the subject.
1437
01:03:04,911 --> 01:03:06,130
Okay.
1438
01:03:06,173 --> 01:03:08,654
How could you learn
more about the subject?
1439
01:03:08,697 --> 01:03:09,960
I can learn more
about the subject
1440
01:03:10,003 --> 01:03:12,876
by reading books,
articles, and websites.
1441
01:03:12,919 --> 01:03:14,268
You can teach me.
1442
01:03:14,312 --> 01:03:15,835
My conclusion is,
1443
01:03:15,879 --> 01:03:17,619
if you now think about
what we're doing,
1444
01:03:17,663 --> 01:03:20,057
is learning from all humans,
1445
01:03:20,100 --> 01:03:22,581
all their knowledge at once
put on the Internet,
1446
01:03:22,624 --> 01:03:24,931
you would actually
know a lot about the world.
1447
01:03:24,975 --> 01:03:27,716
Like a significant portion
of everything humans can do.
1448
01:03:27,760 --> 01:03:29,631
And now,
I think it's more like,
1449
01:03:29,675 --> 01:03:30,937
"Well, it might just work."
1450
01:03:30,981 --> 01:03:33,113
This is a big moment.
1451
01:03:33,157 --> 01:03:34,636
Who is this?
1452
01:03:34,680 --> 01:03:36,551
This is God
reaching out to Adam.
1453
01:03:36,595 --> 01:03:37,988
The advent of AGI
1454
01:03:38,031 --> 01:03:40,860
will divide human history
into two parts.
1455
01:03:40,904 --> 01:03:43,950
The part up to that point
and the part after that point.
1456
01:03:43,994 --> 01:03:47,258
Why is he reaching out
to touch Adam?
1457
01:03:47,301 --> 01:03:48,912
God is
reaching out to touch Adam
1458
01:03:48,955 --> 01:03:50,217
to give him life.
1459
01:03:50,261 --> 01:03:52,219
It will give us a tool
1460
01:03:52,263 --> 01:03:57,529
that can completely reinvent
our entire civilization.
1461
01:03:57,572 --> 01:03:59,400
What does
this painting mean to you?
1462
01:04:01,750 --> 01:04:03,752
The painting
means a lot to me.
1463
01:04:03,796 --> 01:04:04,797
Okay. Like what?
1464
01:04:10,324 --> 01:04:11,456
I think
the painting is a reminder
1465
01:04:11,499 --> 01:04:12,674
that we are all
connected to each other
1466
01:04:12,718 --> 01:04:13,937
and that we are
all part of something
1467
01:04:13,980 --> 01:04:15,112
bigger than ourselves.
1468
01:04:16,461 --> 01:04:17,766
That's pretty nice.
1469
01:04:19,029 --> 01:04:21,379
When you cross
that barrier of
1470
01:04:21,422 --> 01:04:23,947
"AGI might happen
one day in the future"
1471
01:04:23,990 --> 01:04:26,645
to "No, actually, this could
really happen in a time frame
1472
01:04:26,688 --> 01:04:28,690
"that is sort of, like,
on my watch, you know,"
1473
01:04:28,734 --> 01:04:30,475
something changes
in your thinking.
1474
01:04:30,518 --> 01:04:32,694
...learned to orient
itself by looking...
1475
01:04:32,738 --> 01:04:35,045
We have to be
careful with how we use it
1476
01:04:35,088 --> 01:04:37,177
and thoughtful about
how we deploy it.
1477
01:04:39,832 --> 01:04:41,138
You'd have to consider
1478
01:04:41,181 --> 01:04:42,487
what's its top level goal.
1479
01:04:42,530 --> 01:04:45,011
If it's to keep humans happy,
1480
01:04:45,055 --> 01:04:48,928
which set of humans?
What does happiness mean?
1481
01:04:48,972 --> 01:04:52,018
A lot of our collective goals
are very tricky,
1482
01:04:52,062 --> 01:04:54,891
even for humans to figure out.
1483
01:04:54,934 --> 01:04:58,503
Technology always
embeds our values.
1484
01:04:58,546 --> 01:05:01,680
It's not just technical,
it's ethical as well.
1485
01:05:01,723 --> 01:05:02,899
So we've got
to be really cautious
1486
01:05:02,942 --> 01:05:04,291
about what
we're building into it.
1487
01:05:04,335 --> 01:05:06,076
We're trying to find
a single algorithm which...
1488
01:05:06,119 --> 01:05:07,816
The reality is
that this is an algorithm
1489
01:05:07,860 --> 01:05:11,037
that has been created
by people, by us.
1490
01:05:11,081 --> 01:05:13,213
You know, what does it mean
to endow our agents
1491
01:05:13,257 --> 01:05:15,607
with the same kind of values
that we hold dear?
1492
01:05:15,650 --> 01:05:17,652
What is the purpose
of making these AI systems
1493
01:05:17,696 --> 01:05:19,045
appear so humanlike
1494
01:05:19,089 --> 01:05:20,742
so that they do
capture hearts and minds
1495
01:05:20,786 --> 01:05:21,961
because they're kind of
1496
01:05:22,005 --> 01:05:24,703
exploiting a human
vulnerability also?
1497
01:05:24,746 --> 01:05:26,531
The heart and mind
of these systems
1498
01:05:26,574 --> 01:05:28,054
are very much
human-generated data...
1499
01:05:28,098 --> 01:05:29,055
Mmm-hmm.
1500
01:05:29,099 --> 01:05:30,491
...for all the good
and the bad.
1501
01:05:30,535 --> 01:05:32,015
There is a parallel
1502
01:05:32,058 --> 01:05:34,017
between
the Industrial Revolution,
1503
01:05:34,060 --> 01:05:36,758
which was an incredible
moment of displacement
1504
01:05:36,802 --> 01:05:42,373
and the current technological
change created by AI.
1505
01:05:42,416 --> 01:05:43,722
Pause AI!
1506
01:05:43,765 --> 01:05:45,724
We have to think
about who's displaced
1507
01:05:45,767 --> 01:05:48,596
and how we're going
to support them.
1508
01:05:48,640 --> 01:05:50,076
This technology
is coming a lot sooner,
1509
01:05:50,120 --> 01:05:52,426
uh, than really
the world knows or kind of
1510
01:05:52,470 --> 01:05:55,908
even we 18, 24 months
ago thought.
1511
01:05:55,952 --> 01:05:57,257
So there's
a tremendous opportunity,
1512
01:05:57,301 --> 01:05:58,389
tremendous excitement,
1513
01:05:58,432 --> 01:06:00,391
but also
tremendous responsibility.
1514
01:06:00,434 --> 01:06:01,740
It's happening so fast.
1515
01:06:02,654 --> 01:06:04,003
How will we govern it?
1516
01:06:05,135 --> 01:06:06,223
How will we decide
1517
01:06:06,266 --> 01:06:08,181
what is okay
and what is not okay?
1518
01:06:08,225 --> 01:06:10,923
AI-generated images are
getting more sophisticated.
1519
01:06:10,967 --> 01:06:14,535
The use of AI
for generating disinformation
1520
01:06:14,579 --> 01:06:17,016
and manipulating
human psychology
1521
01:06:17,060 --> 01:06:20,237
is only going to get
much, much worse.
1522
01:06:21,194 --> 01:06:22,587
AGI is coming,
1523
01:06:22,630 --> 01:06:24,632
whether we do it here
at DeepMind or not.
1524
01:06:25,459 --> 01:06:26,765
It's gonna happen,
1525
01:06:26,808 --> 01:06:29,028
so we better create
institutions to protect us.
1526
01:06:29,072 --> 01:06:30,595
It's gonna require
global coordination.
1527
01:06:30,638 --> 01:06:32,727
And I worry that humanity is
1528
01:06:32,771 --> 01:06:35,382
increasingly getting worse
at that rather than better.
1529
01:06:35,426 --> 01:06:37,123
We need
a lot more people
1530
01:06:37,167 --> 01:06:40,039
really taking this seriously
and thinking about this.
1531
01:06:40,083 --> 01:06:42,999
It's, yeah, it's serious.
It worries me.
1532
01:06:44,043 --> 01:06:45,871
It worries me. Yeah.
1533
01:06:45,914 --> 01:06:48,613
If you received
an email saying
1534
01:06:48,656 --> 01:06:50,832
this superior
alien civilization
1535
01:06:50,876 --> 01:06:52,791
is going to arrive on Earth,
1536
01:06:52,834 --> 01:06:54,575
there would be
emergency meetings
1537
01:06:54,619 --> 01:06:56,273
of all the governments.
1538
01:06:56,316 --> 01:06:58,144
We would go into overdrive
1539
01:06:58,188 --> 01:07:00,103
trying to figure out
how to prepare.
1540
01:07:01,669 --> 01:07:03,976
The arrival of AGI will be
1541
01:07:04,020 --> 01:07:06,935
the most important moment
that we have ever faced.
1542
01:07:14,378 --> 01:07:17,555
My dream
was that on the way to AGI,
1543
01:07:17,598 --> 01:07:20,688
we would create
revolutionary technologies
1544
01:07:20,732 --> 01:07:23,082
that would be
of use to humanity.
1545
01:07:23,126 --> 01:07:25,171
That's what I wanted
with AlphaFold.
1546
01:07:26,694 --> 01:07:28,653
I think
it's more important than ever
1547
01:07:28,696 --> 01:07:31,047
that we should solve
the protein folding problem.
1548
01:07:32,004 --> 01:07:34,224
This is gonna be really hard,
1549
01:07:34,267 --> 01:07:36,791
but I won't give up
until it's done.
1550
01:07:36,835 --> 01:07:37,879
You know,
we need to double down
1551
01:07:37,923 --> 01:07:40,317
and go as fast as possible
from here.
1552
01:07:40,360 --> 01:07:41,796
I think we've got
no time to lose.
1553
01:07:41,840 --> 01:07:45,757
So we are going to make
a protein folding strike team.
1554
01:07:45,800 --> 01:07:47,541
Team lead for the strike team
will be John.
1555
01:07:47,585 --> 01:07:48,673
Yeah, we've seen Alpha...
1556
01:07:48,716 --> 01:07:50,283
You know,
we're gonna try everything,
1557
01:07:50,327 --> 01:07:51,328
kitchen sink, the whole lot.
1558
01:07:52,198 --> 01:07:53,330
CASP14 is about
1559
01:07:53,373 --> 01:07:55,158
proving we can
solve the whole problem.
1560
01:07:56,333 --> 01:07:57,725
And I felt that to do that,
1561
01:07:57,769 --> 01:08:00,337
we would need to incorporate
some domain knowledge.
1562
01:08:01,903 --> 01:08:03,731
We had some
fantastic engineers on it,
1563
01:08:03,775 --> 01:08:05,733
but they were
not trained in biology.
1564
01:08:08,475 --> 01:08:10,260
As a computational biologist,
1565
01:08:10,303 --> 01:08:12,131
when I initially joined
the AlphaFold team,
1566
01:08:12,175 --> 01:08:14,220
I didn't immediately feel
confident about anything.
1567
01:08:14,264 --> 01:08:15,352
You know,
1568
01:08:15,395 --> 01:08:17,223
whether we were
gonna be successful.
1569
01:08:17,267 --> 01:08:21,097
Biology is so
ridiculously complicated.
1570
01:08:21,140 --> 01:08:25,101
It just felt like this very
far-off mountain to climb.
1571
01:08:25,144 --> 01:08:26,754
I'm starting to play with
the underlying temperatures
1572
01:08:26,798 --> 01:08:27,973
to see if we can get...
1573
01:08:28,016 --> 01:08:29,148
As one of the few people
on the team
1574
01:08:29,192 --> 01:08:31,846
who's done work
in biology before,
1575
01:08:31,890 --> 01:08:34,849
you feel this huge sense
of responsibility.
1576
01:08:34,893 --> 01:08:36,112
"We're expecting you to do
1577
01:08:36,155 --> 01:08:37,678
"great things
on this strike team."
1578
01:08:37,722 --> 01:08:38,897
That's terrifying.
1579
01:08:40,464 --> 01:08:42,727
But one of the reasons
why I wanted to come here
1580
01:08:42,770 --> 01:08:45,556
was to do
something that matters.
1581
01:08:45,599 --> 01:08:48,472
This is the number
of missing things.
1582
01:08:48,515 --> 01:08:49,951
What about making use
1583
01:08:49,995 --> 01:08:52,563
of whatever understanding
you have of physics?
1584
01:08:52,606 --> 01:08:54,391
Using that
as a source of data?
1585
01:08:54,434 --> 01:08:55,479
But if it's systematic...
1586
01:08:55,522 --> 01:08:56,784
Then, that can't be
right, though.
1587
01:08:56,828 --> 01:08:58,308
If it's systematically wrong
in some weird way,
1588
01:08:58,351 --> 01:09:01,224
you might be learning that
systematically wrong physics.
1589
01:09:01,267 --> 01:09:02,355
The team is already
1590
01:09:02,399 --> 01:09:04,749
trying to think
of multiple ways that...
1591
01:09:04,792 --> 01:09:06,229
Biological relevance
1592
01:09:06,272 --> 01:09:07,795
is what we're going for.
1593
01:09:09,057 --> 01:09:11,364
So we rewrote
the whole data pipeline
1594
01:09:11,408 --> 01:09:13,279
that AlphaFold uses to learn.
1595
01:09:13,323 --> 01:09:15,586
You can't
force the creative phase.
1596
01:09:15,629 --> 01:09:18,241
You have to give it space
for those flowers to bloom.
1597
01:09:19,242 --> 01:09:20,286
We won CASP.
1598
01:09:20,330 --> 01:09:22,070
Then it was
back to the drawing board
1599
01:09:22,114 --> 01:09:24,116
and like,
what are our new ideas?
1600
01:09:24,160 --> 01:09:26,945
Um, and then it's taken
a little while, I would say,
1601
01:09:26,988 --> 01:09:28,686
for them to get back
to where they were,
1602
01:09:28,729 --> 01:09:30,340
but with the new ideas.
1603
01:09:30,383 --> 01:09:31,515
And then now I think
1604
01:09:31,558 --> 01:09:33,952
we're seeing the benefits
of the new ideas.
1605
01:09:33,995 --> 01:09:35,736
They can go further, right?
1606
01:09:35,780 --> 01:09:38,130
So, um, that's a really
important moment.
1607
01:09:38,174 --> 01:09:40,959
I've seen that moment
so many times now,
1608
01:09:41,002 --> 01:09:42,613
but I know
what that means now.
1609
01:09:42,656 --> 01:09:44,484
And I know
this is the time now to press.
1610
01:09:45,920 --> 01:09:48,009
Adding side-chains
improves direct folding.
1611
01:09:48,053 --> 01:09:49,663
That drove
a lot of the progress.
1612
01:09:49,707 --> 01:09:51,012
- We'll talk about that.
- Great.
1613
01:09:51,056 --> 01:09:54,799
The last four months,
we've made enormous gains.
1614
01:09:54,842 --> 01:09:56,453
During CASP13,
1615
01:09:56,496 --> 01:09:59,499
it would take us a day or two
to fold one of the proteins,
1616
01:09:59,543 --> 01:10:01,762
and now we're folding, like,
1617
01:10:01,806 --> 01:10:03,938
hundreds of thousands
a second.
1618
01:10:03,982 --> 01:10:05,636
Yeah, it's just insane.
1619
01:10:05,679 --> 01:10:06,985
Now,
this is a model
1620
01:10:07,028 --> 01:10:09,901
that is
orders of magnitude faster,
1621
01:10:09,944 --> 01:10:12,251
while at the same time
being better.
1622
01:10:12,295 --> 01:10:13,644
We're getting
a lot of structures
1623
01:10:13,687 --> 01:10:15,254
into the high-accuracy regime.
1624
01:10:15,298 --> 01:10:17,517
We're rapidly improving
to a system
1625
01:10:17,561 --> 01:10:18,823
that is starting to really
1626
01:10:18,866 --> 01:10:20,477
get at the core and heart
of the problem.
1627
01:10:20,520 --> 01:10:21,695
It's great work.
1628
01:10:21,739 --> 01:10:23,088
It looks like
we're in good shape.
1629
01:10:23,131 --> 01:10:26,222
So we got, what, six,
five weeks left? Six weeks?
1630
01:10:26,265 --> 01:10:29,616
So what's, uh... Is it...
You got enough compute power?
1631
01:10:29,660 --> 01:10:31,531
I... We could use more.
1632
01:10:32,967 --> 01:10:34,360
I was nervous about CASP
1633
01:10:34,404 --> 01:10:36,580
but as the system
is starting to come together,
1634
01:10:36,623 --> 01:10:37,972
I don't feel as nervous.
1635
01:10:38,016 --> 01:10:39,496
I feel like things
have, sort of,
1636
01:10:39,539 --> 01:10:41,193
come into perspective
recently,
1637
01:10:41,237 --> 01:10:44,240
and, you know,
it's gonna be fine.
1638
01:10:47,330 --> 01:10:48,853
The Prime Minister
has announced
1639
01:10:48,896 --> 01:10:51,290
the most drastic limits
to our lives
1640
01:10:51,334 --> 01:10:53,858
the U.K. has ever seen
in living memory.
1641
01:10:53,901 --> 01:10:55,033
I must give the British people
1642
01:10:55,076 --> 01:10:56,904
a very simple instruction.
1643
01:10:56,948 --> 01:10:59,037
You must stay at home.
1644
01:10:59,080 --> 01:11:02,519
It feels like we're
in a science fiction novel.
1645
01:11:02,562 --> 01:11:04,869
You know, I'm delivering food
to my parents,
1646
01:11:04,912 --> 01:11:08,220
making sure
they stay isolated and safe.
1647
01:11:08,264 --> 01:11:10,570
I think it just highlights
the incredible need
1648
01:11:10,614 --> 01:11:12,877
for AI-assisted science.
1649
01:11:17,098 --> 01:11:18,361
You always know that
1650
01:11:18,404 --> 01:11:21,015
something like this
is a possibility.
1651
01:11:21,059 --> 01:11:23,888
But nobody ever really
believes it's gonna happen
1652
01:11:23,931 --> 01:11:25,585
in their lifetime, though.
1653
01:11:26,978 --> 01:11:29,154
-Are you recording yet?
-Yes.
1654
01:11:29,197 --> 01:11:31,025
- Okay, morning, all.
- Hey.
1655
01:11:31,069 --> 01:11:32,679
Good. CASP has started.
1656
01:11:32,723 --> 01:11:36,074
It's nice I get to sit around
in my pajama bottoms all day.
1657
01:11:36,117 --> 01:11:37,597
I never
thought I'd live in a house
1658
01:11:37,641 --> 01:11:39,164
where so much was going on.
1659
01:11:39,207 --> 01:11:41,427
I would be trying to solve
protein folding in one room,
1660
01:11:41,471 --> 01:11:42,559
and my husband would be trying
1661
01:11:42,602 --> 01:11:43,908
to make robots walk
in the other.
1662
01:11:46,954 --> 01:11:49,392
One of the hardest proteins
we've gotten in CASP thus far
1663
01:11:49,435 --> 01:11:51,219
is the SARS-CoV-2 protein
1664
01:11:51,263 --> 01:11:52,220
called ORF8.
1665
01:11:52,264 --> 01:11:54,919
ORF8 is
a coronavirus protein.
1666
01:11:54,962 --> 01:11:56,964
It's one of the main proteins,
um,
1667
01:11:57,008 --> 01:11:58,749
that dampens
the immune system.
1668
01:11:58,792 --> 01:12:00,054
We tried really hard
1669
01:12:00,098 --> 01:12:01,752
to improve our prediction.
1670
01:12:01,795 --> 01:12:03,493
Like, really, really hard.
1671
01:12:03,536 --> 01:12:05,582
Probably the most time
that we have ever spent
1672
01:12:05,625 --> 01:12:07,105
on a single target.
1673
01:12:07,148 --> 01:12:08,933
To the point where
my husband is, like,
1674
01:12:08,976 --> 01:12:12,197
"It's midnight.
You need to go to bed."
1675
01:12:12,240 --> 01:12:16,419
So I think we're at
Day 102 since lockdown.
1676
01:12:16,462 --> 01:12:19,944
My daughter
is keeping a journal.
1677
01:12:19,987 --> 01:12:22,120
Now you can go out
as much as you want.
1678
01:12:25,036 --> 01:12:27,212
We have received
the last target.
1679
01:12:27,255 --> 01:12:29,649
They've said they will be
sending out no more targets
1680
01:12:29,693 --> 01:12:31,347
in our category of CASP.
1681
01:12:32,652 --> 01:12:33,653
So we're just making sure
1682
01:12:33,697 --> 01:12:35,481
we get
the best possible answer.
1683
01:12:40,530 --> 01:12:43,315
As soon as we started
to get the results,
1684
01:12:43,359 --> 01:12:48,233
I'd sit down and start looking
at how close did anybody come
1685
01:12:48,276 --> 01:12:50,583
to getting the protein
structures correct.
1686
01:13:00,245 --> 01:13:01,551
- Oh, hi there.
-Hello.
1687
01:13:03,814 --> 01:13:07,078
It is an unbelievable thing,
CASP has finally ended.
1688
01:13:07,121 --> 01:13:09,472
I think it's at least time
to raise a glass.
1689
01:13:09,515 --> 01:13:11,212
Um, I don't know
if everyone has a glass
1690
01:13:11,256 --> 01:13:12,823
of something
that they can raise.
1691
01:13:12,866 --> 01:13:14,955
If not, raise,
I don't know, your laptops.
1692
01:13:14,999 --> 01:13:17,088
Um...
1693
01:13:17,131 --> 01:13:18,611
I'll probably make a speech
in a minute.
1694
01:13:18,655 --> 01:13:20,483
I feel like I should but I
just have no idea what to say.
1695
01:13:21,005 --> 01:13:24,269
So... let's see.
1696
01:13:24,312 --> 01:13:27,054
I feel like
a reading of email...
1697
01:13:27,098 --> 01:13:28,534
is the right thing to do.
1698
01:13:29,927 --> 01:13:31,232
When John said,
1699
01:13:31,276 --> 01:13:33,191
"I'm gonna read an email,"
at a team social,
1700
01:13:33,234 --> 01:13:35,498
I thought, "Wow, John,
you know how to have fun."
1701
01:13:35,541 --> 01:13:38,370
We're gonna read an email now.
1702
01:13:38,414 --> 01:13:41,634
Uh, I got this
about four o'clock today.
1703
01:13:42,722 --> 01:13:44,724
Um, it is from John Moult.
1704
01:13:45,725 --> 01:13:47,031
And I'll just read it.
1705
01:13:47,074 --> 01:13:49,381
It says,
"As I expect you know,
1706
01:13:49,425 --> 01:13:53,603
"your group has performed
amazingly well in CASP 14,
1707
01:13:53,646 --> 01:13:55,387
"both relative to other groups
1708
01:13:55,431 --> 01:13:57,911
"and in absolute
model accuracy."
1709
01:13:59,826 --> 01:14:01,219
"Congratulations on this work.
1710
01:14:01,262 --> 01:14:03,047
"It is really outstanding."
1711
01:14:03,090 --> 01:14:05,266
The structures were so good,
1712
01:14:05,310 --> 01:14:07,443
it was... it was just amazing.
1713
01:14:09,140 --> 01:14:10,750
After half a century,
1714
01:14:10,794 --> 01:14:12,230
we finally have a solution
1715
01:14:12,273 --> 01:14:14,928
to the protein folding
problem.
1716
01:14:14,972 --> 01:14:17,409
When I saw this email,
I read it,
1717
01:14:17,453 --> 01:14:19,585
I go, "Oh, shit!"
1718
01:14:19,629 --> 01:14:21,587
And my wife goes,
"Is everything okay?"
1719
01:14:21,631 --> 01:14:24,242
I call my parents, and just,
like, "Hey, Mum.
1720
01:14:24,285 --> 01:14:26,244
"Um, got something
to tell you.
1721
01:14:26,287 --> 01:14:27,550
"We've done this thing
1722
01:14:27,593 --> 01:14:29,813
"and it might be kind of
a big deal."
1723
01:14:29,856 --> 01:14:31,641
When I learned of
the CASP 14 results,
1724
01:14:32,642 --> 01:14:34,034
I was gobsmacked.
1725
01:14:34,078 --> 01:14:35,819
I was just excited.
1726
01:14:35,862 --> 01:14:38,909
This is a problem
that I was beginning to think
1727
01:14:38,952 --> 01:14:42,086
would not get solved
in my lifetime.
1728
01:14:42,129 --> 01:14:44,741
Now we have a tool
that can be used
1729
01:14:44,784 --> 01:14:46,612
practically by scientists.
1730
01:14:46,656 --> 01:14:48,440
These people
are asking us, you know,
1731
01:14:48,484 --> 01:14:50,224
"I've got this protein
involved in malaria,"
1732
01:14:50,268 --> 01:14:52,139
or, you know,
some infectious disease.
1733
01:14:52,183 --> 01:14:53,227
"We don't know the structure.
1734
01:14:53,271 --> 01:14:55,186
"Can we use AlphaFold
to solve it?"
1735
01:14:55,229 --> 01:14:56,970
We can easily predict
all known sequences
1736
01:14:57,014 --> 01:14:58,276
in a month.
1737
01:14:58,319 --> 01:14:59,973
All known sequences
in a month?
1738
01:15:00,017 --> 01:15:01,279
- Yeah, easily.
- Mmm-hmm?
1739
01:15:01,322 --> 01:15:02,585
A billion, two billion.
1740
01:15:02,628 --> 01:15:03,673
Um, and they're...
1741
01:15:03,716 --> 01:15:05,196
So why don't we just do that?
Yeah.
1742
01:15:05,239 --> 01:15:07,111
- We should just do that a lot.
- Well, I mean...
1743
01:15:07,154 --> 01:15:09,243
That's way better.
Why don't we just do that?
1744
01:15:09,287 --> 01:15:11,115
So that's
one of the options.
1745
01:15:11,158 --> 01:15:12,638
-Right.
- There's this...
1746
01:15:12,682 --> 01:15:15,119
We should just...
Right, that's a great idea.
1747
01:15:15,162 --> 01:15:17,513
We should just run
every protein in existence.
1748
01:15:18,296 --> 01:15:19,471
And then release that.
1749
01:15:19,515 --> 01:15:20,994
Why didn't someone
suggest this before?
1750
01:15:21,038 --> 01:15:22,126
Of course that's
what we should do.
1751
01:15:22,169 --> 01:15:23,954
Why are we thinking about
making a service
1752
01:15:23,997 --> 01:15:25,651
and then people submit
their protein?
1753
01:15:25,695 --> 01:15:26,913
We just fold everything.
1754
01:15:26,957 --> 01:15:28,654
And then give it to
everyone in the world.
1755
01:15:28,698 --> 01:15:31,483
Who knows how many discoveries
will be made from that?
1756
01:15:31,527 --> 01:15:33,790
Demis called us up
and said,
1757
01:15:33,833 --> 01:15:35,618
"We want to make this open.
1758
01:15:35,661 --> 01:15:37,837
"Not just make sure
the code is open,
1759
01:15:37,881 --> 01:15:39,578
"but we're gonna make it
really easy
1760
01:15:39,622 --> 01:15:42,668
"for everybody to get access
to the predictions."
1761
01:15:45,062 --> 01:15:47,238
That is fantastic.
1762
01:15:47,281 --> 01:15:49,327
It's like drawing back
the curtain
1763
01:15:49,370 --> 01:15:52,852
and seeing the whole world
of protein structures.
1764
01:15:55,246 --> 01:15:56,987
They released the structures
1765
01:15:57,030 --> 01:15:59,772
of 200 million proteins.
1766
01:15:59,816 --> 01:16:01,818
These are gifts to humanity.
1767
01:16:07,650 --> 01:16:10,914
The moment AlphaFold
is live to the world,
1768
01:16:10,957 --> 01:16:13,873
we will no longer be
the most important people
1769
01:16:13,917 --> 01:16:15,222
in AlphaFold's story.
1770
01:16:15,266 --> 01:16:16,833
Can't quite believe
it's all out.
1771
01:16:16,876 --> 01:16:18,356
Aw!
1772
01:16:18,399 --> 01:16:20,314
A hundred
and sixty-four users.
1773
01:16:20,358 --> 01:16:22,578
Loads of activity in Japan.
1774
01:16:22,621 --> 01:16:24,928
We have 655 users currently.
1775
01:16:24,971 --> 01:16:26,930
We currently
have 100,000 concurrent users.
1776
01:16:26,973 --> 01:16:28,192
Wow!
1777
01:16:31,108 --> 01:16:33,893
Today is just crazy.
1778
01:16:33,937 --> 01:16:36,504
What an absolutely
unbelievable effort
1779
01:16:36,548 --> 01:16:37,723
from everyone.
1780
01:16:37,767 --> 01:16:38,550
We're gonna all remember
these moments
1781
01:16:38,594 --> 01:16:40,030
for the rest of our lives.
1782
01:16:40,073 --> 01:16:41,727
I'm excited about AlphaFold.
1783
01:16:41,771 --> 01:16:45,601
For my research, it's already
propelling lots of progress.
1784
01:16:45,644 --> 01:16:47,385
And this is
just the beginning.
1785
01:16:47,428 --> 01:16:48,908
My guess is,
1786
01:16:48,952 --> 01:16:53,043
every single biological
and chemistry achievement
1787
01:16:53,086 --> 01:16:55,698
will be related to AlphaFold
in some way.
1788
01:17:13,367 --> 01:17:15,413
AlphaFold is an index moment.
1789
01:17:15,456 --> 01:17:18,068
It's a moment
that people will not forget
1790
01:17:18,111 --> 01:17:20,244
because the world changed.
1791
01:17:39,655 --> 01:17:41,482
Everybody's realized now
1792
01:17:41,526 --> 01:17:43,746
what Shane and I have known
for more than 20 years,
1793
01:17:43,789 --> 01:17:46,618
that AI is going to be
the most important thing
1794
01:17:46,662 --> 01:17:48,446
humanity's ever gonna invent.
1795
01:17:48,489 --> 01:17:50,230
We will shortly be arriving
1796
01:17:50,274 --> 01:17:52,058
at our final destination.
1797
01:18:02,068 --> 01:18:04,767
The pace of
innovation and capabilities
1798
01:18:04,810 --> 01:18:06,507
is accelerating,
1799
01:18:06,551 --> 01:18:09,293
like a boulder rolling down
a hill that we've kicked off
1800
01:18:09,336 --> 01:18:12,644
and now it's continuing
to gather speed.
1801
01:18:12,688 --> 01:18:15,299
We are at
a crossroads in human history.
1802
01:18:15,342 --> 01:18:16,735
AI has the potential
1803
01:18:16,779 --> 01:18:19,172
to transform our lives
in every aspect.
1804
01:18:19,216 --> 01:18:23,786
It's no less important than
the discovery of electricity.
1805
01:18:23,829 --> 01:18:26,484
We should be looking
at the scientific method
1806
01:18:26,527 --> 01:18:28,834
and trying to understand
each step of the way
1807
01:18:28,878 --> 01:18:30,096
in a rigorous way.
1808
01:18:30,140 --> 01:18:32,664
This is a moment
of profound opportunity.
1809
01:18:32,708 --> 01:18:34,753
Harnessing this technology
1810
01:18:34,797 --> 01:18:37,713
could eclipse anything
we have ever known.
1811
01:18:42,195 --> 01:18:43,675
Hi, Alpha.
1812
01:18:44,676 --> 01:18:45,764
Hi.
1813
01:18:47,157 --> 01:18:48,419
What is this?
1814
01:18:50,682 --> 01:18:53,729
This is a chessboard.
1815
01:18:53,772 --> 01:18:56,514
If I was to play white, what
move would you recommend?
1816
01:18:59,865 --> 01:19:00,953
I would recommend
1817
01:19:00,997 --> 01:19:02,781
that you move your pawn
from E2 to E4.
1818
01:19:05,871 --> 01:19:08,787
And now if you were black,
what would you play now?
1819
01:19:11,572 --> 01:19:13,618
I would play
the Sicilian Defense.
1820
01:19:15,838 --> 01:19:16,882
That's a good choice.
1821
01:19:19,406 --> 01:19:21,452
Thanks.
1822
01:19:23,715 --> 01:19:25,891
So what do you see?
What is this object?
1823
01:19:28,546 --> 01:19:30,504
This is a pencil sculpture.
1824
01:19:32,811 --> 01:19:35,031
What happens if I move
one of the pencils?
1825
01:19:37,990 --> 01:19:39,470
If you move
one of the pencils,
1826
01:19:39,513 --> 01:19:42,081
the sculpture will fall apart.
1827
01:19:42,125 --> 01:19:44,301
I'd better leave it alone,
then.
1828
01:19:44,344 --> 01:19:45,868
That's probably
a good idea.
1829
01:19:50,568 --> 01:19:52,744
AGI is on the horizon now.
1830
01:19:54,833 --> 01:19:56,661
Very clearly
the next generation
1831
01:19:56,704 --> 01:19:58,141
is going to live
in a future world
1832
01:19:58,184 --> 01:20:01,057
where things will be radically
different because of AI.
1833
01:20:02,493 --> 01:20:05,496
And if you want to steward
that responsibly,
1834
01:20:05,539 --> 01:20:09,239
every moment is vital.
1835
01:20:09,282 --> 01:20:12,677
This is the moment I've been
living my whole life for.
1836
01:20:19,162 --> 01:20:21,120
It's just
a good thinking game.