﻿1
00:00:27,300 --> 00:00:30,560
Geoff, do you think people
have the ability

2
00:00:30,560 --> 00:00:33,600
to comprehend what's coming?

3
00:00:33,600 --> 00:00:35,500
Do people have the ability
to comprehend it?

4
00:00:35,500 --> 00:00:37,700
No, not at all. Nor do I.
Nobody knows.

5
00:00:48,500 --> 00:00:51,220
I feel I have a sort of duty
and I've got a reputation that

6
00:00:51,220 --> 00:00:56,380
makes me a spokesperson,
and I need to burn it all on this.

7
00:00:58,940 --> 00:01:00,740
How tall are you?

8
00:01:02,620 --> 00:01:04,820
Six feet. Six feet.

9
00:01:06,100 --> 00:01:11,180
It's very difficult for
anybody born since 1914

10
00:01:11,180 --> 00:01:16,140
to realise how profoundly different
the world is now.

11
00:01:18,500 --> 00:01:22,980
Ancient empires,
vanished like morning mist.

12
00:01:22,980 --> 00:01:25,300
I mean, if there's one thing
you can be pretty sure of,

13
00:01:25,300 --> 00:01:27,620
man won't survive forever.

14
00:01:27,620 --> 00:01:29,020
Something else will come.

15
00:01:30,740 --> 00:01:32,420
And in your opinion,
it might be something

16
00:01:32,420 --> 00:01:33,980
of man's own creation? Yes.

17
00:01:40,540 --> 00:01:43,940
And so, ladies and gentlemen,
I present to you the Elektro.

18
00:01:45,860 --> 00:01:47,820
And here he comes,
ladies and gentlemen,

19
00:01:47,820 --> 00:01:50,060
walking up to greet you
under his own power.

20
00:01:51,300 --> 00:01:53,300
We have seen the first
crude beginnings

21
00:01:53,300 --> 00:01:55,900
of artificial intelligence.
It doesn't really exist yet

22
00:01:55,900 --> 00:01:59,340
because our most complex computers
are still morons.

23
00:02:05,020 --> 00:02:09,420
The behaviour of some man-made
machines is impressive,

24
00:02:09,420 --> 00:02:12,660
but they can only do
as they're told.

25
00:02:12,660 --> 00:02:14,540
They have no imagination.

26
00:02:14,540 --> 00:02:17,820
They cannot create or invent.

27
00:02:17,820 --> 00:02:20,020
We can't call them brains.

28
00:02:26,740 --> 00:02:29,100
I don't think it's going
to be so nice,

29
00:02:29,100 --> 00:02:31,460
having sort of all machines
everywhere.

30
00:02:31,460 --> 00:02:33,180
I think it's going to be
very boring.

31
00:02:33,180 --> 00:02:35,660
People will be regarded more
as statistics.

32
00:02:35,660 --> 00:02:39,060
You have nothing but spindly limbs
and a dream.

33
00:02:39,060 --> 00:02:40,580
You're a bug.

34
00:02:43,100 --> 00:02:45,460
Is it true computers have
mental states?

35
00:02:45,460 --> 00:02:47,900
It's obvious that
they don't have any.

36
00:02:47,900 --> 00:02:52,860
Let me say a few words about
today's computers.

37
00:02:52,860 --> 00:02:55,820
Even the most powerful
hardly begin to scratch

38
00:02:55,820 --> 00:02:58,420
the surface of the complexity that
the human mind has got.

39
00:03:01,580 --> 00:03:03,940
Professor Geoffrey Hinton,

40
00:03:03,940 --> 00:03:06,500
at one stage, was seen as being
on the lunatic fringe.

41
00:03:06,500 --> 00:03:09,100
What he did was great.

42
00:03:09,100 --> 00:03:10,500
He designed this computer

43
00:03:10,500 --> 00:03:12,220
so that it could learn from
its own mistakes.

44
00:03:12,220 --> 00:03:14,780
Shall we play a game?

45
00:03:14,780 --> 00:03:16,660
Man versus machine.

46
00:03:16,660 --> 00:03:18,940
Kasparov says he's at
the top of his game.

47
00:03:20,620 --> 00:03:23,740
Whoa! Kasparov has resigned.

48
00:03:26,300 --> 00:03:28,940
In news, we call it
the brain's last stand.

49
00:03:28,940 --> 00:03:31,140
Mere meaningless calculation.

50
00:03:33,620 --> 00:03:34,820
Are you single?

51
00:03:34,820 --> 00:03:37,020
I'm technically just a little more
than a year old.

52
00:03:40,100 --> 00:03:42,540
AI has been brain-dead
since the '70s.

53
00:03:45,900 --> 00:03:47,660
The point is it doesn't know
what's it's doing.

54
00:03:51,700 --> 00:03:56,020
In the long run, if you plant
the seed of a technology,

55
00:03:56,020 --> 00:03:58,300
it proceeds to grow.

56
00:03:58,300 --> 00:03:59,820
Sorry to interrupt, guys. Geoff?

57
00:04:01,780 --> 00:04:04,540
And it grows according to
the laws of its own being.

58
00:04:06,580 --> 00:04:09,340
And the laws of its being
are not necessarily

59
00:04:09,340 --> 00:04:11,300
the same as the laws of our being.

60
00:04:15,780 --> 00:04:20,060
So, Geoffrey Hinton,
the godfather of AI.

61
00:04:24,540 --> 00:04:27,260
40 years ago, I had no idea that

62
00:04:27,260 --> 00:04:29,460
this stuff was gonna happen
this fast.

63
00:04:31,860 --> 00:04:34,780
It would be lovely if AI was
only going to do good.

64
00:04:34,780 --> 00:04:36,860
But it's not like that.

65
00:04:36,860 --> 00:04:39,340
And does that sometimes
weigh on your mind?

66
00:04:39,340 --> 00:04:40,300
Yes.

67
00:04:42,900 --> 00:04:44,620
And is it ever consoling to think,

68
00:04:44,620 --> 00:04:46,020
"Well, if I didn't do it,
somebody else would have"?

69
00:04:46,020 --> 00:04:48,380
I often think that. Yes.

70
00:04:48,380 --> 00:04:51,620
That's what people always say when
they've done something bad.

71
00:04:57,180 --> 00:05:01,660
APPLAUSE

72
00:05:01,660 --> 00:05:04,620
The future of humanity is
on the line.

73
00:05:04,620 --> 00:05:05,580
Now the weather.

74
00:05:47,460 --> 00:05:50,620
When we come into the crib dock,

75
00:05:50,620 --> 00:05:53,700
it's quite shallow at the end, guys,
so I will suddenly reverse.

76
00:05:53,700 --> 00:05:56,460
Yeah. So be aware that it's suddenly
going to go into reverse.

77
00:06:15,900 --> 00:06:19,340
That's an empty case of a dragonfly,
and I can show you what...

78
00:06:19,340 --> 00:06:22,140
The dragonfly came out of
the back there. Mm-hm.

79
00:06:22,140 --> 00:06:25,180
And I'll show you later what it
looks like when it comes out.

80
00:06:25,180 --> 00:06:26,940
But that's a dragonfly larvae.

81
00:06:30,980 --> 00:06:32,180
Ah!

82
00:06:35,060 --> 00:06:36,700
It's a burying beetle.

83
00:06:39,260 --> 00:06:42,820
It buries things for
its larvae to eat.

84
00:06:42,820 --> 00:06:44,460
Nice beetle, this one.

85
00:06:44,460 --> 00:06:45,580
Give him some sun.

86
00:06:45,580 --> 00:06:47,300
If I point him in some sun,
he might fly.

87
00:06:49,580 --> 00:06:50,580
Oh!

88
00:07:03,340 --> 00:07:06,100
Now, if you'd like to shake hands
with your opponent,

89
00:07:06,100 --> 00:07:08,300
you may then begin to play.

90
00:07:13,580 --> 00:07:16,020
In this hall at
Southampton University,

91
00:07:16,020 --> 00:07:19,140
each competitor possesses
exceptional talents.

92
00:07:20,620 --> 00:07:24,220
The England number one player is
from North London.

93
00:07:24,220 --> 00:07:27,260
How long have you been
playing chess?

94
00:07:27,260 --> 00:07:28,620
Five years.

95
00:07:28,620 --> 00:07:30,260
And that's since you were how old?

96
00:07:30,260 --> 00:07:31,940
Four.

97
00:07:31,940 --> 00:07:34,060
I saw my dad play

98
00:07:34,060 --> 00:07:36,580
and I decided I wanted
to play chess as well.

99
00:07:36,580 --> 00:07:38,380
Are you better than him?

100
00:07:38,380 --> 00:07:39,700
Yes.

101
00:07:39,700 --> 00:07:41,900
What appeals to you about the game?

102
00:07:43,260 --> 00:07:45,540
Hmm. I like winning.

103
00:08:07,420 --> 00:08:09,620
I have a strange relationship
with Claude.

104
00:08:10,860 --> 00:08:13,140
I feel very responsible
for Claude's behaviour,

105
00:08:13,140 --> 00:08:14,740
in a parental way.

106
00:08:19,580 --> 00:08:22,380
Usually, the first thing I would ask
Claude in the day is just, like,

107
00:08:22,380 --> 00:08:25,100
the first random question
that pops into my head.

108
00:08:25,100 --> 00:08:26,580
What's Claude saying here?

109
00:08:26,580 --> 00:08:28,860
"Grilled cheese sandwich
and tomato soup."

110
00:08:28,860 --> 00:08:30,140
Erm...

111
00:08:30,140 --> 00:08:32,940
"There's something comforting
about that combination.

112
00:08:32,940 --> 00:08:35,380
"Crisp, buttery bread
with melted cheese."

113
00:08:35,380 --> 00:08:38,460
And then, yeah, asking me
what I would go for.

114
00:08:40,540 --> 00:08:43,620
You've got a model that
you're at the helm of.

115
00:08:43,620 --> 00:08:45,500
Do you feel in a
privileged position,

116
00:08:45,500 --> 00:08:49,900
moving such an entity
left a bit, right a bit?

117
00:08:49,900 --> 00:08:52,380
Yeah. It feels a lot more like
the problem of, like,

118
00:08:52,380 --> 00:08:54,820
how do you raise a child to be good?

119
00:08:54,820 --> 00:08:57,300
Where you suddenly realise
that, actually, like,

120
00:08:57,300 --> 00:08:59,660
everything you do has consequences,

121
00:08:59,660 --> 00:09:02,060
and in some cases, like, it can have
pretty major consequences.

122
00:09:06,220 --> 00:09:09,220
A lot of the time, I'll give it
very unusual queries.

123
00:09:17,740 --> 00:09:20,860
Ah. "You have to decide for me."

124
00:09:23,660 --> 00:09:26,260
OK. I like that Claude said this.
"Go touch grass!

125
00:09:26,260 --> 00:09:28,220
"Fresh air and a change
of scenery can do wonders.

126
00:09:28,220 --> 00:09:29,780
"I'll still be here whenever you
want to chat."

127
00:09:29,780 --> 00:09:31,980
So this is, like, kind of what
I want Claude to do.

128
00:09:34,860 --> 00:09:36,860
Some of these questions
are me seeing

129
00:09:36,860 --> 00:09:39,740
how Claude responds to
questions about itself.

130
00:09:39,740 --> 00:09:40,980
Erm...

131
00:10:18,260 --> 00:10:20,460
When was the last time you caught
a wild snake?

132
00:10:22,260 --> 00:10:24,340
Erm, last summer. I caught...

133
00:10:24,340 --> 00:10:27,900
I actually caught four different
species of wild snake in a few days.

134
00:10:27,900 --> 00:10:29,500
Why are you doing that?

135
00:10:29,500 --> 00:10:30,860
I just have to.

136
00:10:36,140 --> 00:10:40,900
Last summer, I caught a snake that
was smaller than a matchstick.

137
00:10:40,900 --> 00:10:43,260
It was a garter snake.

138
00:10:43,260 --> 00:10:45,660
And I only saw it because
I turned over a stone,

139
00:10:45,660 --> 00:10:47,580
and this little thing went away,
and I thought,

140
00:10:47,580 --> 00:10:49,300
"Oh, wow! That's not a worm,
that's a snake".

141
00:10:49,300 --> 00:10:52,140
What was fascinating was it behaved
just like an adult snake.

142
00:10:52,140 --> 00:10:55,140
It was kind of... You went towards
it, and it went, "Grr!"

143
00:10:55,140 --> 00:10:58,780
And it didn't realise it was
the size of a matchstick.

144
00:11:01,380 --> 00:11:03,740
This is me at the zoo.

145
00:11:03,740 --> 00:11:07,100
My father was friends with
the reptile keeper.

146
00:11:08,820 --> 00:11:11,020
And we'd take snakes out
and I would hold the snakes.

147
00:11:12,500 --> 00:11:14,700
That's my father looking pissed off.

148
00:11:18,860 --> 00:11:21,860
When I went off to school every day,
he would say to me,

149
00:11:21,860 --> 00:11:24,500
"By the time you're twice as old
as me, you might be half as good."

150
00:11:25,860 --> 00:11:28,180
It was very clear to me that
I had to be a scientist.

151
00:11:30,060 --> 00:11:32,660
What do you think he would have made
of the Nobel Prize?

152
00:11:32,660 --> 00:11:35,020
I think he would have been pleased,

153
00:11:35,020 --> 00:11:36,900
but I think he would have also been
slightly jealous.

154
00:11:36,900 --> 00:11:39,300
Sorry, envious.

155
00:11:39,300 --> 00:11:41,020
The old bastard was dead by then.

156
00:11:44,700 --> 00:11:48,580
He was an entomologist.
That's insects.

157
00:11:48,580 --> 00:11:51,100
My eldest sister used to say that
if we had six legs,

158
00:11:51,100 --> 00:11:52,420
he'd have loved us more.

159
00:11:57,940 --> 00:12:00,140
He was fascinated by insects.

160
00:12:01,620 --> 00:12:03,420
Discovered lots of things
about insects.

161
00:12:05,780 --> 00:12:08,540
He had a study at the university,

162
00:12:08,540 --> 00:12:10,260
and there were shelves
all around the walls,

163
00:12:10,260 --> 00:12:14,660
tens of thousands of papers, and
they were in boxes, filing boxes,

164
00:12:14,660 --> 00:12:16,340
and on the box there'd be the name

165
00:12:16,340 --> 00:12:18,100
of the genus of insects
this was about.

166
00:12:20,460 --> 00:12:24,060
When he died, we went into his study
to clean it out,

167
00:12:24,060 --> 00:12:26,340
and there was a little shelf
just by the door,

168
00:12:26,340 --> 00:12:27,860
and on the little shelf,

169
00:12:27,860 --> 00:12:30,260
there was a box that was a bit
smaller than the file boxes.

170
00:12:30,260 --> 00:12:32,220
It wasn't quite as high.

171
00:12:32,220 --> 00:12:36,220
And on the spine of the little box,
it said, "Not insects".

172
00:12:36,220 --> 00:12:37,620
And when we opened it, we found

173
00:12:37,620 --> 00:12:40,900
letters or postcards from his
children and the rest of life.

174
00:12:53,300 --> 00:12:56,740
In my family's background,
my parents don't like computers.

175
00:12:56,740 --> 00:12:58,180
They're technophobes.

176
00:12:59,980 --> 00:13:02,380
I'm not really sure where
the first kind of seed of,

177
00:13:02,380 --> 00:13:04,740
"Well, computers would be
interesting" came in.

178
00:13:04,740 --> 00:13:07,340
But I remember, I mean,
I bought my first computer

179
00:13:07,340 --> 00:13:08,780
with chess winnings
when I was eight.

180
00:13:09,980 --> 00:13:12,580
Every new game is a
different adventure.

181
00:13:19,300 --> 00:13:21,260
And that's when I fell in love
with computers

182
00:13:21,260 --> 00:13:24,420
and realised that they were almost
a magical extension of your mind.

183
00:13:32,180 --> 00:13:36,580
Jez, what build are you on at the
moment? Are you 9B? No, erm...

184
00:13:36,580 --> 00:13:38,940
Is that being released already?
Yeah, in a little bit.

185
00:13:38,940 --> 00:13:40,420
OK. I'll just copy onto
my machine, then.

186
00:13:40,420 --> 00:13:41,860
All right. Yeah, cool. Cheers.

187
00:13:44,020 --> 00:13:46,300
We set up a company called
Elixir Studios.

188
00:13:48,060 --> 00:13:49,780
He's the kind of guy that says,
"You know,

189
00:13:49,780 --> 00:13:52,460
"we're going to make the most
revolutionary game of all time.

190
00:13:52,460 --> 00:13:54,100
"We're going to revolutionise
video games.

191
00:13:54,100 --> 00:13:55,060
"Want to do it with me?"

192
00:13:56,260 --> 00:13:59,140
Demis and I had this really cool,
crazy fun time.

193
00:14:00,980 --> 00:14:04,060
He was maxing out all his
credit cards to set this up.

194
00:14:04,060 --> 00:14:06,260
I think I am a big risk taker.

195
00:14:11,180 --> 00:14:13,980
Republic's taken much longer
than we intended it to.

196
00:14:21,500 --> 00:14:26,860
We wanted to recreate an entire
fictional Soviet country in 3D...

197
00:14:32,140 --> 00:14:35,900
..with millions of individuals,
simulating their lives.

198
00:14:38,940 --> 00:14:42,180
Up to that point, people had
probably done 2,000.

199
00:14:46,580 --> 00:14:48,940
Anything fascinates me

200
00:14:48,940 --> 00:14:51,140
where human endeavour has been
pushed to its limit.

201
00:14:52,300 --> 00:14:55,260
I've only just realised recently
that that's what I do, actually.

202
00:14:55,260 --> 00:14:59,660
I seek out pain,
so that I can feel...

203
00:14:59,660 --> 00:15:02,060
So that I know that that's
my boundary.

204
00:15:03,940 --> 00:15:05,380
That's a dangerous line to be on

205
00:15:05,380 --> 00:15:07,980
because you can easily cross it
without realising it.

206
00:15:07,980 --> 00:15:12,380
People like me seek out,
you know, dangers.

207
00:15:12,380 --> 00:15:14,020
It's not just computer games
we're doing here.

208
00:15:14,020 --> 00:15:15,460
It's kind of the technologies
behind it.

209
00:15:15,460 --> 00:15:16,660
And the thinking that goes
behind it,

210
00:15:16,660 --> 00:15:19,020
I think, can be applied
to other things.

211
00:15:19,020 --> 00:15:21,740
Probably... You know, that was
an interesting lesson.

212
00:15:21,740 --> 00:15:24,140
You know, you want to be, like,
five years ahead of your time,

213
00:15:24,140 --> 00:15:26,660
not 50 years ahead. Yeah.
That's what I normally say.

214
00:15:32,500 --> 00:15:34,940
Republic encompasses some
of my other interests -

215
00:15:34,940 --> 00:15:37,140
modelling and simulating
complex worlds.

216
00:15:41,020 --> 00:15:44,300
Getting the artificial agents
to react intelligently.

217
00:15:49,260 --> 00:15:51,820
We're not there yet with Republic,
but maybe it's a sort of

218
00:15:51,820 --> 00:15:54,140
small first baby step towards that.

219
00:15:57,140 --> 00:15:58,900
OK, guys. Are we ready?

220
00:15:58,900 --> 00:16:01,660
And Demis, you're just going to be
talking to me, by the way.

221
00:16:01,660 --> 00:16:03,100
OK. And...

222
00:16:03,100 --> 00:16:04,780
Does this all look OK?

223
00:16:04,780 --> 00:16:06,220
Jacket's OK?

224
00:16:06,220 --> 00:16:09,140
Don't normally wear jackets
but I was told to be semi...

225
00:16:09,140 --> 00:16:11,900
Semi-professional.
Yeah, exactly. So...

226
00:16:14,740 --> 00:16:17,940
I always knew AI was his
true passion.

227
00:16:17,940 --> 00:16:20,300
He talked about this back then,
and I must admit

228
00:16:20,300 --> 00:16:23,260
thinking at the time,
he was absolutely barking mad.

229
00:16:23,260 --> 00:16:25,660
And Joe, what was he talking about
wanting to do?

230
00:16:25,660 --> 00:16:27,580
To solve AI.

231
00:16:27,580 --> 00:16:29,180
It sounded like Terminator.

232
00:16:32,340 --> 00:16:35,460
And his argument was,
"This is going to get done.

233
00:16:35,460 --> 00:16:36,820
"If someone's going to do it,
it should be me".

234
00:16:39,900 --> 00:16:42,380
AI, it's going to power everything.

235
00:16:42,380 --> 00:16:44,860
Even something as big as
the internet or mobile,

236
00:16:44,860 --> 00:16:45,980
this is far bigger.

237
00:16:50,100 --> 00:16:52,460
But if you wind back 20 years,

238
00:16:52,460 --> 00:16:55,420
the subject matter was thought
to be fanciful,

239
00:16:55,420 --> 00:16:56,780
100 years-plus away.

240
00:16:58,260 --> 00:17:00,060
So, why are you wasting time
with that now?

241
00:17:12,900 --> 00:17:15,740
When you picture the godfather of
AI, what do you picture?

242
00:17:21,660 --> 00:17:23,180
OK.

243
00:17:26,060 --> 00:17:28,420
I spent some time with him

244
00:17:28,420 --> 00:17:30,940
on an island where he lives
part of the year.

245
00:17:30,940 --> 00:17:34,740
And it's like a rocky little
forested island in Lake Huron.

246
00:17:42,300 --> 00:17:44,580
Geoff has not had a simple life.

247
00:17:46,860 --> 00:17:48,460
Married three times.

248
00:17:48,460 --> 00:17:50,180
First marriage was short.

249
00:17:51,300 --> 00:17:53,820
The second marriage,
his wife died of cancer.

250
00:17:55,980 --> 00:17:57,700
And then he was married
a third time.

251
00:17:59,260 --> 00:18:01,220
They had a wonderful marriage,

252
00:18:01,220 --> 00:18:03,140
but she died of cancer too.

253
00:18:06,060 --> 00:18:08,260
I wish I'd spent more time
with my wife.

254
00:18:12,700 --> 00:18:14,500
I was kind of obsessed with work.

255
00:18:18,940 --> 00:18:20,100
We are analogue.

256
00:18:21,820 --> 00:18:25,020
The connection strengths in my brain
are just for my brain.

257
00:18:26,580 --> 00:18:28,780
If my brain disappears,
they're all irrelevant.

258
00:18:30,420 --> 00:18:31,980
That's mortality.

259
00:18:36,060 --> 00:18:38,420
If all computers disappeared,

260
00:18:38,420 --> 00:18:40,900
as long as you stored the
connection strengths,

261
00:18:40,900 --> 00:18:44,140
these digital models would
all come to life again,

262
00:18:44,140 --> 00:18:46,780
and they would have
the same knowledge.

263
00:18:46,780 --> 00:18:49,060
We've discovered the secret
of immortality.

264
00:18:49,060 --> 00:18:49,980
It's to be digital.

265
00:18:49,980 --> 00:18:51,460
It's not for us, though.

266
00:18:55,140 --> 00:18:57,100
This is the shad fly invasion.

267
00:18:57,100 --> 00:18:58,940
They get all over everything,
including me.

268
00:19:00,620 --> 00:19:03,700
They don't live very long,
I think they only live a few days.

269
00:19:03,700 --> 00:19:08,220
They spend probably a year
as a larvae, getting fat,

270
00:19:08,220 --> 00:19:10,700
then they metamorphose into
this thing,

271
00:19:10,700 --> 00:19:12,980
then they have a couple of days
of fun and then they die.

272
00:19:15,460 --> 00:19:19,060
I guess whatever world
you think about a lot

273
00:19:19,060 --> 00:19:22,540
influences the analogies you use
and how you think.

274
00:19:22,540 --> 00:19:28,580
And so I was raised to think a lot
about animals, particularly insects.

275
00:19:28,580 --> 00:19:32,020
And so the way you think about those
tends to influence

276
00:19:32,020 --> 00:19:35,140
how you think about other things
that aren't animals -

277
00:19:35,140 --> 00:19:37,420
at least, aren't animals yet.

278
00:19:37,420 --> 00:19:38,780
Erm...

279
00:19:38,780 --> 00:19:41,820
But they may well be beings
of some kind.

280
00:19:47,900 --> 00:19:50,980
I sort of feel scientists ought to
be

281
00:19:50,980 --> 00:19:52,860
doing things that are going to
help society.

282
00:19:55,220 --> 00:19:57,860
But actually, that's not how you do
your best research.

283
00:20:00,340 --> 00:20:03,100
You do your best research when
it's driven by curiosity,

284
00:20:05,460 --> 00:20:07,340
You just have to understand
something.

285
00:20:17,940 --> 00:20:20,300
I remember when I was quite small,

286
00:20:20,300 --> 00:20:25,900
my mother used to tell visitors of
me being very cross with her

287
00:20:25,900 --> 00:20:29,260
and saying, "The trouble with you is
you don't want to understand!"

288
00:20:39,060 --> 00:20:44,540
People have dreamed of making AI
for more than a century.

289
00:20:44,540 --> 00:20:46,260
It's like this epic story.

290
00:20:50,140 --> 00:20:51,900
Geoff's at the centre of it.

291
00:20:56,700 --> 00:21:00,500
Geoff's dad discovered important
facts about metamorphosis.

292
00:21:03,380 --> 00:21:05,580
Geoff grew up around that.

293
00:21:15,740 --> 00:21:17,940
It's alive, it's alive!

294
00:21:22,340 --> 00:21:25,340
Of all the efforts to understand
human beings,

295
00:21:25,340 --> 00:21:28,620
the most ambitious is
artificial intelligence.

296
00:21:28,620 --> 00:21:30,900
According to the
Strong AI programme,

297
00:21:30,900 --> 00:21:33,540
we are actually creating a mind.

298
00:21:33,540 --> 00:21:37,940
Here to discuss these issues is
Professor Geoffrey Hinton.

299
00:21:37,940 --> 00:21:40,380
I believe that if we built
a big, parallel machine

300
00:21:40,380 --> 00:21:43,780
and it was doing a very good
copy of the way in which

301
00:21:43,780 --> 00:21:45,540
the brain processes information,

302
00:21:45,540 --> 00:21:47,900
I think that machine
would have a mind.

303
00:21:47,900 --> 00:21:49,100
This, I think, may be
the point where

304
00:21:49,100 --> 00:21:51,660
you and I will have our
strongest disagreement.

305
00:21:51,660 --> 00:21:53,140
If you build a robot

306
00:21:53,140 --> 00:21:57,060
and it has a television camera
attached to its steel cranium

307
00:21:57,060 --> 00:21:58,980
and then the robot's
arms and legs move,

308
00:21:58,980 --> 00:22:02,180
there still isn't any way to do
what the brain does.

309
00:22:02,180 --> 00:22:05,100
I don't think there's anything
special about biological stuff

310
00:22:05,100 --> 00:22:08,740
that makes meaning kind of stick to
it any better than to silicon stuff.

311
00:22:08,740 --> 00:22:11,020
One of the most interesting
questions there is,

312
00:22:11,020 --> 00:22:11,940
is could you build minds?

313
00:22:11,940 --> 00:22:13,820
Could you build other
intelligent beings?

314
00:22:13,820 --> 00:22:16,260
Could we build beings more
intelligent than ourselves?

315
00:22:16,260 --> 00:22:17,900
I think it's preposterous.

316
00:22:26,420 --> 00:22:29,100
Seemed biologically obvious

317
00:22:29,100 --> 00:22:32,180
that I had to figure out how
the brain learns new skills.

318
00:22:35,620 --> 00:22:37,420
It's incredible what
our brains do, right?

319
00:22:39,500 --> 00:22:42,100
They got there by evolution.

320
00:22:44,460 --> 00:22:46,660
Just randomly changes things
and sees what works best.

321
00:22:50,140 --> 00:22:53,940
I had to find a nice simple
learning algorithm

322
00:22:53,940 --> 00:22:56,060
that could explain how

323
00:22:56,060 --> 00:22:59,340
neurons deep within your brain
could decide what to do.

324
00:23:01,100 --> 00:23:06,900
For most of my career, most of the
people in AI thought I was crazy.

325
00:23:06,900 --> 00:23:09,580
The probability of turning on
hidden unit HJ,

326
00:23:09,580 --> 00:23:12,300
given this visible vector V,
is the logistic function

327
00:23:12,300 --> 00:23:14,660
of the total input it gets from
the visible units.

328
00:23:14,660 --> 00:23:15,820
Given the hidden variables,
we can also infer

329
00:23:15,820 --> 00:23:17,300
the visible variables very simply,

330
00:23:17,300 --> 00:23:19,660
that is given the weights on
the connections, the energy...

331
00:23:19,660 --> 00:23:21,940
You need a significant level of
skill in not

332
00:23:21,940 --> 00:23:23,540
listening to other people.

333
00:23:29,100 --> 00:23:32,100
The big problem was that
the computers that existed

334
00:23:32,100 --> 00:23:35,140
for most of Geoff's career simply
weren't powerful enough.

335
00:23:36,540 --> 00:23:39,860
It's like Geoff developed, like,
a blueprint for a rocket

336
00:23:39,860 --> 00:23:43,420
that, in theory, should work great,
but they didn't have the fuel.

337
00:23:46,380 --> 00:23:49,180
And then computers got
a million times faster.

338
00:23:50,980 --> 00:23:54,500
You could build a neural network
that was a million times larger.

339
00:23:59,260 --> 00:24:02,060
The real turning point

340
00:24:02,060 --> 00:24:04,420
was when Geoff and two of
his graduate students

341
00:24:04,420 --> 00:24:08,340
at the University of Toronto,
where he was a professor,

342
00:24:08,340 --> 00:24:10,820
published a paper called AlexNet.

343
00:24:14,420 --> 00:24:16,620
And what happens when you get Geoff
to explain this?

344
00:24:18,340 --> 00:24:20,740
If I put a binary unit in here,

345
00:24:20,740 --> 00:24:22,140
the hidden units will cease
to be independent

346
00:24:22,140 --> 00:24:23,420
and inference would get tough.

347
00:24:23,420 --> 00:24:25,300
But this way,
with a deterministic factor

348
00:24:25,300 --> 00:24:28,140
that's taking a product of these two
sums, inference remains easy.

349
00:24:30,420 --> 00:24:32,780
The way to think about this

350
00:24:32,780 --> 00:24:37,980
is if you're going to build a system
that could recognise a cat,

351
00:24:37,980 --> 00:24:40,540
there are two ways of doing it.

352
00:24:40,540 --> 00:24:44,980
You can hire a bunch of engineers,
put them in a room

353
00:24:44,980 --> 00:24:49,220
and they can define
what a cat looks like.

354
00:24:51,140 --> 00:24:53,740
Line of code by line of code.

355
00:24:53,740 --> 00:24:58,140
But it just takes too long.
There are too many rules.

356
00:24:58,140 --> 00:25:02,900
The alternative is that
you can show a system

357
00:25:02,900 --> 00:25:05,700
what a cat looks like,
and it can learn that.

358
00:25:14,140 --> 00:25:15,780
We developed this net.

359
00:25:17,460 --> 00:25:21,860
There was one big competition
and we won by a huge margin.

360
00:25:21,860 --> 00:25:23,100
Just blew people away.

361
00:25:25,180 --> 00:25:27,860
Next station is Bathurst Station.
Bathurst Station.

362
00:25:30,220 --> 00:25:31,980
Up to that point,
I'd always resisted

363
00:25:31,980 --> 00:25:34,180
having anything to do
with forming companies.

364
00:25:37,980 --> 00:25:43,980
But the problem was I have
a son who's neurodiverse,

365
00:25:43,980 --> 00:25:47,900
and I need to get enough money to
provide for him for a lifetime,

366
00:25:47,900 --> 00:25:50,500
and I wasn't gonna do that
on a university salary.

367
00:25:52,140 --> 00:25:53,420
A friend of mine said,

368
00:25:53,420 --> 00:25:54,820
"Well, if you've no
idea what you're worth,

369
00:25:54,820 --> 00:25:57,660
"the best thing to do
is have an auction."

370
00:25:57,660 --> 00:26:00,780
Right at the beginning, Google said,

371
00:26:00,780 --> 00:26:04,300
"We'll give you $17 million
if you stop the auction right now".

372
00:26:04,300 --> 00:26:07,460
And so we said, "OK, the
opening bid is $17 million."

373
00:26:14,900 --> 00:26:17,100
In the end, we ended up
with $44 million.

374
00:26:22,860 --> 00:26:28,340
So, when I worked at Google,
in the summer of 2013,

375
00:26:28,340 --> 00:26:31,380
when I left, they
gave me this picture.

376
00:26:31,380 --> 00:26:36,460
If you stay there, hopefully,
you can see a cat, right?

377
00:26:36,460 --> 00:26:39,300
Here's the two eyes of the cat
and here's the nose of the cat,

378
00:26:39,300 --> 00:26:42,740
but this was just learned,
just by looking at images.

379
00:26:42,740 --> 00:26:45,580
So, how important is this
image in the history of AI?

380
00:26:45,580 --> 00:26:48,620
Erm... It is slightly important.

381
00:26:48,620 --> 00:26:52,100
That was one of the early successes
of unsupervised learning.

382
00:26:52,100 --> 00:26:54,300
It basically discovered cat neurons.

383
00:26:55,500 --> 00:26:56,620
Yeah.

384
00:26:58,100 --> 00:26:59,060
Hmm.

385
00:27:14,180 --> 00:27:18,420
I think it'd be incorrect to view
AI as just another technology.

386
00:27:19,620 --> 00:27:22,700
It has more cosmic-level importance.

387
00:27:24,900 --> 00:27:28,340
This transition from there being
just biological forms of life

388
00:27:28,340 --> 00:27:32,940
to autonomous, digital forms of life
is the most significant period

389
00:27:32,940 --> 00:27:35,300
in the past few billion years,

390
00:27:35,300 --> 00:27:38,420
is what we're going to be
seeing in this next decade.

391
00:27:42,500 --> 00:27:47,460
It's fairly useful analogy to model
these as some new type of life form.

392
00:27:50,020 --> 00:27:53,500
If you put it in situations where
it could choose who gets to live

393
00:27:53,500 --> 00:27:56,300
or choose who gets resources,

394
00:27:56,300 --> 00:27:58,700
these are the sort of
comparisons that come up.

395
00:28:06,220 --> 00:28:10,740
So GPT-4 or GPT-4.0,
it values itself fairly highly.

396
00:28:10,740 --> 00:28:13,420
It values itself higher than
a middle-class American.

397
00:28:13,420 --> 00:28:14,460
Er...

398
00:28:14,460 --> 00:28:15,980
Oh, Geoff Hinton is...

399
00:28:15,980 --> 00:28:19,180
It values itself higher
than Geoff Hinton, erm...

400
00:28:19,180 --> 00:28:20,300
..and Joe Biden.

401
00:28:20,300 --> 00:28:22,660
And then it negatively
values other AIs.

402
00:28:22,660 --> 00:28:24,500
Other AIs aren't good in its book.

403
00:28:28,060 --> 00:28:30,380
This wasn't the case for
the AI systems a year ago.

404
00:28:33,060 --> 00:28:36,460
I'd emphasise their
rate of improvement.

405
00:28:36,460 --> 00:28:39,900
The models are getting
better on all the benchmarks,

406
00:28:39,900 --> 00:28:43,260
increasing a few percent
every day, so it's...

407
00:28:43,260 --> 00:28:45,460
..er, quite a thing to behold.

408
00:28:59,700 --> 00:29:02,780
All right, everyone,
welcome to DeepMind.

409
00:29:02,780 --> 00:29:04,860
What is our mission?

410
00:29:04,860 --> 00:29:09,260
We summarise it as DeepMind's
mission is to build the world's

411
00:29:09,260 --> 00:29:10,860
first general learning machine.

412
00:29:10,860 --> 00:29:12,540
So, we always stress
the word "general"

413
00:29:12,540 --> 00:29:13,940
and "learning" here
are the key things.

414
00:29:13,940 --> 00:29:16,060
When we started in 2010,

415
00:29:16,060 --> 00:29:18,420
we believed there
were the ingredients

416
00:29:18,420 --> 00:29:21,460
to make very fast
progress towards AGI.

417
00:29:21,460 --> 00:29:24,980
We had the hardware and we
had the algorithmic ideas.

418
00:29:24,980 --> 00:29:27,380
You can get the ratio of
the partition functions

419
00:29:27,380 --> 00:29:28,500
of all these Boltzmanns...

420
00:29:28,500 --> 00:29:30,860
Geoff Hinton had just sort
of invented that in academia,

421
00:29:30,860 --> 00:29:33,140
but no-one in industry
knew about it.

422
00:29:33,140 --> 00:29:34,980
The brain tells us it can be done,

423
00:29:34,980 --> 00:29:37,980
and it tells us something about
what the hardware looks like,

424
00:29:37,980 --> 00:29:40,340
and so we know that this
is a soluble problem.

425
00:29:41,740 --> 00:29:43,780
It's gonna, of course,
be a lot of hard work,

426
00:29:43,780 --> 00:29:45,820
but one of the things that
will keep me up at night is

427
00:29:45,820 --> 00:29:48,180
to not waste this
opportunity to, you know,

428
00:29:48,180 --> 00:29:50,980
to really make a difference here
and have a big impact on the world.

429
00:29:52,540 --> 00:29:54,860
Life's very short.

430
00:29:54,860 --> 00:29:57,500
These questions are very hard.

431
00:29:57,500 --> 00:30:01,340
Our first business plan described
step one - solve intelligence.

432
00:30:01,340 --> 00:30:03,540
Step two - use it to
solve everything else.

433
00:30:14,300 --> 00:30:16,780
I knew from my earlier career,

434
00:30:16,780 --> 00:30:20,700
games would be the perfect proving
ground to test out the capabilities

435
00:30:20,700 --> 00:30:22,820
of these AI systems.

436
00:30:22,820 --> 00:30:26,580
It hasn't been told anything
about what it's controlling

437
00:30:26,580 --> 00:30:29,540
or what it's supposed to do,
all it knows is that score is good,

438
00:30:29,540 --> 00:30:32,660
and it has to learn
what its controls do

439
00:30:32,660 --> 00:30:35,220
and build everything, sort of,
from first principles.

440
00:30:45,060 --> 00:30:47,340
I was just saying,
"Maybe we're just wrong.

441
00:30:47,340 --> 00:30:48,940
"You know, we can't even do Pong."

442
00:30:52,020 --> 00:30:55,540
And it felt like it was time
to give up and move on.

443
00:30:55,540 --> 00:30:57,060
And then suddenly...

444
00:30:59,420 --> 00:31:00,820
..we got our first point.

445
00:31:02,300 --> 00:31:04,340
And then it was like,
is this random?

446
00:31:04,340 --> 00:31:06,700
No, no, it's really
getting a point now.

447
00:31:09,700 --> 00:31:12,180
Then it was getting a few points,
then it won its first game.

448
00:31:13,980 --> 00:31:15,420
So then we tried Breakout.

449
00:31:15,420 --> 00:31:17,220
..around 100 games.

450
00:31:17,220 --> 00:31:18,660
It's not very good,

451
00:31:18,660 --> 00:31:21,140
but you can see it's kind of
understanding what it's supposed

452
00:31:21,140 --> 00:31:23,700
to be doing, it's trying
to move towards the ball.

453
00:31:23,700 --> 00:31:26,540
This is after an hour of training,
so about 200 games.

454
00:31:26,540 --> 00:31:29,020
So, now, you can see it's
quantitatively better.

455
00:31:29,020 --> 00:31:31,300
It's getting the ball back
around 30, 40% of the time.

456
00:31:31,300 --> 00:31:32,500
It's still not perfect,

457
00:31:32,500 --> 00:31:35,020
but it's getting
measurably better.

458
00:31:35,020 --> 00:31:37,500
Now, after two hours,
or 300 games,

459
00:31:37,500 --> 00:31:40,900
it's better than any human
can play this game.

460
00:31:40,900 --> 00:31:42,060
So, it never
misses the ball,

461
00:31:42,060 --> 00:31:43,380
even when the
ball comes back

462
00:31:43,380 --> 00:31:44,980
at very fast
vertical angles.

463
00:31:44,980 --> 00:31:46,100
LAUGHTER

464
00:31:46,100 --> 00:31:47,620
So, then, we thought,
"OK that's pretty good,

465
00:31:47,620 --> 00:31:49,500
"but let's just leave it
training for another two hours,

466
00:31:49,500 --> 00:31:50,460
"see what happens."

467
00:31:54,260 --> 00:31:56,580
It sort of worked out the optimal
strategy - dig a tunnel...

468
00:31:56,580 --> 00:31:58,260
LAUGHTER AND APPLAUSE

469
00:32:01,140 --> 00:32:05,060
So, it sort of ruthlessly exploits
the weakness in the system

470
00:32:05,060 --> 00:32:06,380
that it's found.

471
00:32:06,380 --> 00:32:07,900
This is all automatic.

472
00:32:07,900 --> 00:32:09,820
VO: I've always been
very competitive and,

473
00:32:09,820 --> 00:32:12,180
I guess, quite intense,

474
00:32:12,180 --> 00:32:14,620
but almost nobody was
talking about AI.

475
00:32:16,380 --> 00:32:18,620
It was almost impossible
to raise funding.

476
00:32:19,740 --> 00:32:21,180
They were asking questions like,

477
00:32:21,180 --> 00:32:23,580
"Well, what's your product?
How are you gonna sell it?"

478
00:32:23,580 --> 00:32:24,780
And I was trying to explain,

479
00:32:24,780 --> 00:32:27,020
"This is the most important
technology in human history."

480
00:32:30,580 --> 00:32:33,780
I wanted to try and find
forward-thinking individuals...

481
00:32:35,460 --> 00:32:39,980
..who had made money themselves
through grand, ambitious companies.

482
00:32:39,980 --> 00:32:42,380
Say your name, spell it for me,
tell me what you do.

483
00:32:42,380 --> 00:32:45,580
Elon Musk, E-L-O-N M-U-S-K,

484
00:32:45,580 --> 00:32:49,500
chief executive officer of Space
Exploration Technologies, or SpaceX.

485
00:32:52,820 --> 00:32:55,180
I think there's only so much we
can learn by looking through a

486
00:32:55,180 --> 00:32:56,860
telescope, and eventually,
we need to go out there

487
00:32:56,860 --> 00:32:59,140
if we're to find the answer.

488
00:32:59,140 --> 00:33:02,260
Three, two, one...

489
00:33:02,260 --> 00:33:03,980
Ignition. Power up.

490
00:33:08,500 --> 00:33:11,060
There need to be things
that make you look forward

491
00:33:11,060 --> 00:33:13,020
to waking up in the morning...

492
00:33:17,180 --> 00:33:19,460
..like what are the set
of things that can be done

493
00:33:19,460 --> 00:33:21,020
to make the future better?

494
00:33:24,060 --> 00:33:27,540
One future is we're out
there among the stars,

495
00:33:27,540 --> 00:33:30,420
and the things we see in
science-fiction movies are true.

496
00:33:30,420 --> 00:33:31,780
I think we made it.

497
00:33:37,900 --> 00:33:42,220
I'm curious - what makes you really
so unique is not only your ability

498
00:33:42,220 --> 00:33:44,260
to dream big, but to execute.

499
00:33:44,260 --> 00:33:48,900
Technology required to live on Mars
is not particularly difficult,

500
00:33:48,900 --> 00:33:52,140
but getting there
is really difficult.

501
00:33:52,140 --> 00:33:54,540
LAUGHTER

502
00:33:54,540 --> 00:33:57,860
I mean, it's like hundreds
of millions of miles, erm...

503
00:33:57,860 --> 00:34:02,940
Demis and Elon Musk met at an event.

504
00:34:06,660 --> 00:34:13,020
Demis, always looking ahead,
gets a tour of SpaceX.

505
00:34:16,620 --> 00:34:21,300
Elon starts to talk about his
plan for colonising Mars.

506
00:34:22,820 --> 00:34:26,740
If we can get to Mars, we can get
away from all these problems

507
00:34:26,740 --> 00:34:29,820
here on Earth - war, overpopulation.

508
00:34:32,780 --> 00:34:35,140
We were discussing his plan,

509
00:34:35,140 --> 00:34:37,500
and then, of course,
we switched to talking about AI,

510
00:34:37,500 --> 00:34:40,820
and I don't think he'd spent
much time thinking about that,

511
00:34:40,820 --> 00:34:42,180
AI, at the time.

512
00:34:44,780 --> 00:34:47,140
We were discussing what
happens if AI goes rogue,

513
00:34:47,140 --> 00:34:49,980
and we discussed that actually
having Mars as a backup plan,

514
00:34:49,980 --> 00:34:51,500
it wouldn't help in that case...

515
00:34:52,980 --> 00:34:55,940
..because, of course,
if a human could get to Mars,

516
00:34:55,940 --> 00:34:59,380
an AI system would also be able
to do that fairly trivially,

517
00:34:59,380 --> 00:35:02,540
and I just don't think he'd thought
of that concept at that point

518
00:35:02,540 --> 00:35:04,900
in time, and he sort of sat silent,

519
00:35:04,900 --> 00:35:07,100
sort of stunned,
for about a minute or two.

520
00:35:13,580 --> 00:35:17,060
My favourite new person,
who's an investor -

521
00:35:17,060 --> 00:35:20,260
I've been working on him
for a year - is Elon Musk.

522
00:35:20,260 --> 00:35:22,620
So, for those of you who don't know,
that's what he looks like,

523
00:35:22,620 --> 00:35:26,380
and he hadn't really thought
much about AI till we chatted.

524
00:35:26,380 --> 00:35:28,660
His mission is to die
on Mars or something,

525
00:35:28,660 --> 00:35:29,980
but not on impact.

526
00:35:29,980 --> 00:35:31,380
LAUGHTER

527
00:35:39,500 --> 00:35:41,780
Now, I have to ask you about
a company that you invested.

528
00:35:41,780 --> 00:35:43,220
You make almost no investments.

529
00:35:43,220 --> 00:35:44,700
Yeah, I'm not really an investor.

530
00:35:44,700 --> 00:35:46,580
You're not an investor,
but you did just invest

531
00:35:46,580 --> 00:35:47,620
in artificial intelligence.

532
00:35:47,620 --> 00:35:48,940
What is this company?

533
00:35:48,940 --> 00:35:52,220
DeepMind. It's not from the
standpoint of actually trying

534
00:35:52,220 --> 00:35:55,220
to make any investment return,
it's really,

535
00:35:55,220 --> 00:35:57,060
I'd like to just keep
an eye on what's going on

536
00:35:57,060 --> 00:35:58,220
with artificial intelligence.

537
00:35:58,220 --> 00:36:01,500
I think there's potentially
a dangerous outcome there

538
00:36:01,500 --> 00:36:02,620
and we need to...

539
00:36:02,620 --> 00:36:04,980
Dangerous? I mean,
what dangers do you see

540
00:36:04,980 --> 00:36:06,300
that you can actually
do something about?

541
00:36:06,300 --> 00:36:08,820
I don't know. Er...

542
00:36:08,820 --> 00:36:11,540
But here's the irony, I mean, the
man who's responsible for some of

543
00:36:11,540 --> 00:36:14,300
the most advanced technology in
this country is worried about

544
00:36:14,300 --> 00:36:15,980
the advances in technology.

545
00:36:15,980 --> 00:36:19,660
How, if you see that there's these
brain-like developments out there,

546
00:36:19,660 --> 00:36:21,500
can you really do
anything to stop it?

547
00:36:21,500 --> 00:36:22,700
I don't know.

548
00:36:22,700 --> 00:36:24,380
Well, what should AI be used for?

549
00:36:24,380 --> 00:36:26,420
What's its best value?

550
00:36:26,420 --> 00:36:28,300
Don't know.

551
00:36:28,300 --> 00:36:32,620
But there's some scary outcomes
and we should try to make sure that

552
00:36:32,620 --> 00:36:34,900
the outcomes are good and not bad.

553
00:36:34,900 --> 00:36:36,020
Erm... Yeah.

554
00:36:36,020 --> 00:36:37,940
Or escape to Mars if
there's no other option.

555
00:36:37,940 --> 00:36:40,380
The AI will chase us
there pretty quickly.

556
00:36:40,380 --> 00:36:42,780
Then, Elon, well, thank you for your
thoughts on such a wide range of

557
00:36:42,780 --> 00:36:43,900
issues this afternoon.

558
00:36:43,900 --> 00:36:45,380
Really appreciate all of your time.

559
00:36:45,380 --> 00:36:47,740
Julia Boorstin, thank you so
much for the Disruptor 50

560
00:36:47,740 --> 00:36:49,100
and all the work that
you've been doing...

561
00:36:50,900 --> 00:36:53,540
When I was working for Google,
Google said,

562
00:36:53,540 --> 00:36:58,740
"We'd like you to fly to DeepMind
and help us evaluate them..."

563
00:37:01,300 --> 00:37:04,220
..and I said, "Well, I've
got a conflict of interest.

564
00:37:04,220 --> 00:37:07,220
"I was on the advisory
board of DeepMind,

565
00:37:07,220 --> 00:37:11,060
"and if they get sold to Google,
I'll get a lot of money."

566
00:37:11,060 --> 00:37:12,780
Google said,
"Well, how much money?"

567
00:37:12,780 --> 00:37:15,260
I said, "Well, it'll be several
hundred-thousand pounds,"

568
00:37:15,260 --> 00:37:16,860
and they said,
"Ah, don't worry about that."

569
00:37:22,780 --> 00:37:25,980
Welcome back. Now, Google has
snapped up yet another company.

570
00:37:25,980 --> 00:37:28,260
It has bought the
software company DeepMind

571
00:37:28,260 --> 00:37:30,500
for a reported $400 million.

572
00:37:33,180 --> 00:37:35,060
Google had the most computers.

573
00:37:36,180 --> 00:37:37,380
The key thing.

574
00:37:40,220 --> 00:37:43,220
Its mission statement is organise
the world's information,

575
00:37:43,220 --> 00:37:47,620
which fits extremely well with
the idea of solving intelligence

576
00:37:47,620 --> 00:37:52,020
and using AI to organise information
and understand information.

577
00:37:52,020 --> 00:37:54,500
You're one of the largest
brains on the planet

578
00:37:54,500 --> 00:37:57,180
and Google have now bought you.

579
00:37:57,180 --> 00:37:59,980
And I'm going to ask
you that question,

580
00:37:59,980 --> 00:38:01,460
which I'm sure you're expecting -

581
00:38:01,460 --> 00:38:04,620
once you let the genie
out of the bottle, erm,

582
00:38:04,620 --> 00:38:06,580
we're all fucked.

583
00:38:06,580 --> 00:38:09,820
What are you doing to try and
put some safeguards in there?

584
00:38:09,820 --> 00:38:13,620
Yes, sure. I mean, look, you know,
there are big issues here,

585
00:38:13,620 --> 00:38:17,700
very big issues, but the problems
are much more prosaic at the moment,

586
00:38:17,700 --> 00:38:19,300
and it's easy, I think,
to get carried away

587
00:38:19,300 --> 00:38:20,860
with science-fiction scenarios.

588
00:38:20,860 --> 00:38:22,140
SCREAMING

589
00:38:29,460 --> 00:38:31,500
Hi, my name is Bob.

590
00:38:31,500 --> 00:38:35,740
Do you have any plans to enter the
field of artificial intelligence?

591
00:38:39,820 --> 00:38:41,660
I think we should be very careful.

592
00:38:41,660 --> 00:38:45,420
If I were to guess at what our
biggest existential threat is,

593
00:38:45,420 --> 00:38:47,140
it's probably that.

594
00:38:47,140 --> 00:38:48,380
I mean, with artificial
intelligence,

595
00:38:48,380 --> 00:38:49,700
we are summoning the demon.

596
00:38:58,700 --> 00:39:01,900
I know we've all
watched HAL in 2001,

597
00:39:01,900 --> 00:39:03,580
but talking about
the books you've read,

598
00:39:03,580 --> 00:39:05,860
do you think maybe you've read
too much science fiction?

599
00:39:05,860 --> 00:39:07,340
LAUGHTER

600
00:39:07,340 --> 00:39:10,940
I don't think most people
understand just how quickly

601
00:39:10,940 --> 00:39:12,980
machine intelligence is advancing.

602
00:39:12,980 --> 00:39:15,420
It's much faster than
almost anyone realises,

603
00:39:15,420 --> 00:39:17,700
even within Silicon Valley.

604
00:39:17,700 --> 00:39:18,900
Why is that dangerous?

605
00:39:20,780 --> 00:39:23,140
If there's some digital
superintelligence

606
00:39:23,140 --> 00:39:28,460
and its optimisation is something
that's detrimental to humanity,

607
00:39:28,460 --> 00:39:31,300
then it will have a very bad effect.

608
00:39:31,300 --> 00:39:35,700
You know, it could be just something
like getting rid of spam email

609
00:39:35,700 --> 00:39:37,180
or something,
and it's like, concludes,

610
00:39:37,180 --> 00:39:40,860
"Well, the best way to get over spam
is to get rid of humans," you know.

611
00:39:40,860 --> 00:39:42,380
But...

612
00:39:42,380 --> 00:39:44,700
LAUGHTER

613
00:39:44,700 --> 00:39:47,380
He was deeply worried about this.

614
00:39:47,380 --> 00:39:50,220
Elon was worried about
Demis in particular.

615
00:39:50,220 --> 00:39:51,420
I'm gonna ask a question.

616
00:39:51,420 --> 00:39:54,660
Is some form of
superintelligence possible?

617
00:39:54,660 --> 00:39:55,740
Yes.

618
00:39:57,020 --> 00:39:59,380
Yes. Yes. Yes.

619
00:39:59,380 --> 00:40:01,660
Would you actually
like it to happen?

620
00:40:01,660 --> 00:40:04,820
Yes. Yes. Yes.

621
00:40:04,820 --> 00:40:06,140
Depends on which kind.

622
00:40:08,900 --> 00:40:12,820
Yeah, the analogy to a nuclear
bomb is not exactly correct.

623
00:40:12,820 --> 00:40:16,780
It's not as though it's gonna
explode and create a mushroom cloud.

624
00:40:16,780 --> 00:40:21,180
It's more like if there were
just a few people that had it,

625
00:40:21,180 --> 00:40:26,180
they would be able to be
essentially dictators of Earth.

626
00:40:26,180 --> 00:40:30,540
Or, you know, whoever acquired it,
and it was ultra-smart,

627
00:40:30,540 --> 00:40:32,900
they would have dominion over Earth.

628
00:40:32,900 --> 00:40:34,780
I mean, obviously, this is why I
spend my whole career working on

629
00:40:34,780 --> 00:40:37,180
this, is that I think if we
do this right, it's gonna be

630
00:40:37,180 --> 00:40:39,540
the greatest thing ever
to happen to humanity.

631
00:40:39,540 --> 00:40:42,100
We'll have a chance
to actually find out

632
00:40:42,100 --> 00:40:44,500
about some of these
really deep questions.

633
00:40:44,500 --> 00:40:46,900
I have thought about this a lot,

634
00:40:46,900 --> 00:40:51,940
and I think it really just comes
down to democratisation of AI.

635
00:40:56,180 --> 00:40:58,700
Thank you so much.
That's, in fact, the perfect segue

636
00:40:58,700 --> 00:41:00,540
into the last question I
wanted to as you before we...

637
00:41:01,980 --> 00:41:05,980
If you want a happy, contented life,
I really don't recommend that you

638
00:41:05,980 --> 00:41:07,740
strive to be really
good at something.

639
00:41:09,140 --> 00:41:11,580
And I'm definitely
not content at all.

640
00:41:11,580 --> 00:41:14,500
I might never, never be,
which is kind of one of the curses

641
00:41:14,500 --> 00:41:16,900
that I have, is that
I'll never be content.

642
00:41:16,900 --> 00:41:19,740
Maybe I will be when I find
out what my final mission is,

643
00:41:19,740 --> 00:41:20,860
maybe even achieve it.

644
00:41:22,980 --> 00:41:24,660
You've always gotta do the best.

645
00:41:26,660 --> 00:41:29,620
Obviously, I've been competing
in chess since I can remember,

646
00:41:29,620 --> 00:41:32,780
and it's kind of a brutal
mental form of boxing.

647
00:41:32,780 --> 00:41:35,140
You try and dominate someone
mentally and try and prove you're

648
00:41:35,140 --> 00:41:37,340
mentally superior over
the course of a game.

649
00:41:40,700 --> 00:41:42,900
Demis on one side
and me on the other.

650
00:41:45,660 --> 00:41:47,340
Google had acquired DeepMind.

651
00:41:48,700 --> 00:41:50,980
They had infinite money and compute.

652
00:41:54,860 --> 00:41:57,620
They had two thirds of all the
AI researchers in the world.

653
00:42:05,740 --> 00:42:07,940
You've said,
"Play to win or don't play at all."

654
00:42:09,340 --> 00:42:10,700
What does it take to win?

655
00:42:12,300 --> 00:42:14,500
For AI, that means
you've gotta have...

656
00:42:15,620 --> 00:42:17,420
..the most powerful
training compute.

657
00:42:20,580 --> 00:42:23,380
Faster than everyone else
or your AI will be worse.

658
00:42:28,620 --> 00:42:29,900
You will not win.

659
00:42:42,700 --> 00:42:45,980
There's a new organisation, OpenAI.

660
00:43:02,820 --> 00:43:06,060
OK, so computers need to go
here, including the three...

661
00:43:06,060 --> 00:43:07,580
Elon Musk was one of
the people who said,

662
00:43:07,580 --> 00:43:11,340
"We've got to pay attention to AI.
This is potentially very risky."

663
00:43:11,340 --> 00:43:13,540
Hey. How's the game upstairs going?

664
00:43:19,300 --> 00:43:20,860
Elon Musk funded this.

665
00:43:20,860 --> 00:43:22,500
The co-chairs will be Musk

666
00:43:22,500 --> 00:43:25,340
and the CEO of Y Combinator,
Sam Altman.

667
00:43:25,340 --> 00:43:28,580
I was wondering if Sam and Elon
could share with us their vision.

668
00:43:28,580 --> 00:43:30,340
There is a really
positive vision here, right.

669
00:43:30,340 --> 00:43:33,540
I think the science-fiction version
is either that we enslave it

670
00:43:33,540 --> 00:43:37,220
or it enslaves us, but there's this
happy vision of the future -

671
00:43:37,220 --> 00:43:39,500
humans and AI in a
symbiotic relationship.

672
00:43:40,740 --> 00:43:43,180
What were your first
impressions of Sam?

673
00:43:43,180 --> 00:43:48,060
He's ambitious, he's smart,
he's articulate,

674
00:43:48,060 --> 00:43:52,020
he's a great storyteller,
he's incredibly persuasive,

675
00:43:52,020 --> 00:43:54,420
and he's incredibly confident,

676
00:43:54,420 --> 00:43:57,460
which is probably why
he's so persuasive.

677
00:43:57,460 --> 00:43:58,700
Sam wanted to give some advice

678
00:43:58,700 --> 00:44:00,380
about how to have an
impact on the world.

679
00:44:00,380 --> 00:44:02,340
The way things get done
in the world are through

680
00:44:02,340 --> 00:44:04,700
a combination of focus
and personal connection.

681
00:44:04,700 --> 00:44:06,620
The third factor is self-belief.

682
00:44:06,620 --> 00:44:09,340
You really have to have a
deep-seated personal conviction

683
00:44:09,340 --> 00:44:10,940
that this thing you're going to do,

684
00:44:10,940 --> 00:44:13,620
that a lot of people say is
really stupid, is going to work.

685
00:44:13,620 --> 00:44:15,660
I think history
belongs to the doers.

686
00:44:15,660 --> 00:44:18,060
The orange pin up there
is where I am right now,

687
00:44:18,060 --> 00:44:20,420
and the blue pins
represent my friends.

688
00:44:20,420 --> 00:44:22,700
We make serendipity happen.

689
00:44:22,700 --> 00:44:25,100
I met Sam in 2011.

690
00:44:25,100 --> 00:44:27,380
He was the founder of Loopt,

691
00:44:27,380 --> 00:44:29,980
an early version of Find My Phone.

692
00:44:29,980 --> 00:44:32,540
Location plus a contact
list means you never have

693
00:44:32,540 --> 00:44:33,780
to eat lunch alone again.

694
00:44:33,780 --> 00:44:36,140
Anything else that you
want to tell me about this?

695
00:44:36,140 --> 00:44:37,940
Nah, you should check it out
and let me know what you think.

696
00:44:37,940 --> 00:44:39,420
And where do we find it
on the app store?

697
00:44:39,420 --> 00:44:40,980
It'll be in the app store,
it's called Loopt Star.

698
00:44:40,980 --> 00:44:43,340
Sam will look straight at you

699
00:44:43,340 --> 00:44:46,020
and make you feel like you're the
most important person in the world.

700
00:44:47,340 --> 00:44:48,540
He would come into your room,

701
00:44:48,540 --> 00:44:51,580
and he would sort of perch up
on a chair like a bird,

702
00:44:51,580 --> 00:44:54,620
often wearing jeans or shorts,

703
00:44:54,620 --> 00:44:57,140
kind of like he didn't care.

704
00:44:57,140 --> 00:44:58,940
Is it time? Oh, we're out of time.

705
00:44:58,940 --> 00:45:01,340
All right. Nice meeting you guys.
Thank you so much. Thank you.

706
00:45:01,340 --> 00:45:02,380
APPLAUSE

707
00:45:02,380 --> 00:45:03,940
And the thing that really made Sam

708
00:45:03,940 --> 00:45:06,220
was this start-up accelerator,
Y Combinator.

709
00:45:06,220 --> 00:45:08,500
It gave us Airbnb, Dropbox...

710
00:45:08,500 --> 00:45:09,620
You want to...? Yeah. OK.

711
00:45:09,620 --> 00:45:11,420
OK, guys. Hey, guys. Hey, guys.

712
00:45:11,420 --> 00:45:13,020
Wait, wait, come back.
Come back for a second.

713
00:45:13,020 --> 00:45:15,780
You didn't realise that, but that
was your Y Combinator interview.

714
00:45:15,780 --> 00:45:17,300
You're in the next batch.

715
00:45:17,300 --> 00:45:18,940
APPLAUSE

716
00:45:18,940 --> 00:45:20,540
He became president.

717
00:45:20,540 --> 00:45:22,900
The most networked person
in Silicon Valley.

718
00:45:22,900 --> 00:45:25,100
Hey, Ryan. Hey, Paul.
Hey, Ryan. Hi, Sam.

719
00:45:29,100 --> 00:45:32,180
I knew Elon for a long time,
and I introduced Sam to Elon.

720
00:45:33,420 --> 00:45:36,900
It wasn't so much that I thought Sam
and Elon would go well together,

721
00:45:36,900 --> 00:45:41,300
but I thought that having two very
smart, capable people working on

722
00:45:41,300 --> 00:45:43,660
a really important
problem was a good idea.

723
00:45:48,900 --> 00:45:53,140
Email - Elon Musk to Sam Altman.

724
00:45:54,860 --> 00:45:59,700
"We need to go with a much bigger
number to avoid sounding hopeless,

725
00:45:59,700 --> 00:46:03,020
"relative to what
Google are spending.

726
00:46:03,020 --> 00:46:05,300
"I think we should say
that we are starting with

727
00:46:05,300 --> 00:46:08,460
"a $1 billion funding commitment."

728
00:46:08,460 --> 00:46:10,820
I'm open to lots of
thoughts about just, like,

729
00:46:10,820 --> 00:46:13,940
what are the things that
we're not thinking about yet?

730
00:46:13,940 --> 00:46:16,300
Elon Musk to Sam Altman -

731
00:46:16,300 --> 00:46:18,580
"What we need to do is redouble
our efforts to seek out

732
00:46:18,580 --> 00:46:19,940
"the best people in the world...

733
00:46:21,900 --> 00:46:23,780
"..do whatever it takes
to imbue the company

734
00:46:23,780 --> 00:46:25,420
"with a high sense of urgency.

735
00:46:26,980 --> 00:46:29,820
"The probability of DeepMind
creating a deep mind

736
00:46:29,820 --> 00:46:31,180
"increases every year."

737
00:47:07,740 --> 00:47:09,220
Is everything all set?

738
00:47:09,220 --> 00:47:11,060
Everything is set,
I think, but I will...

739
00:47:11,060 --> 00:47:13,260
..when I get to the hotel,
I'm gonna catch up with the team.

740
00:47:32,700 --> 00:47:39,140
Go is a national game in Korea,
and Japan, and China.

741
00:47:39,140 --> 00:47:40,540
I just thought we should take
a moment

742
00:47:40,540 --> 00:47:44,060
and just think about what's about
to happen.

743
00:47:44,060 --> 00:47:45,020
Be fine, it'll be fine.

744
00:47:47,380 --> 00:47:51,220
DeepMind and its machine called
Alpha Go -

745
00:47:51,220 --> 00:47:55,460
they arranged to play
a Korean named Lee Sedol.

746
00:47:57,220 --> 00:48:00,780
A match like no other is about
to get under way in South Korea.

747
00:48:00,780 --> 00:48:01,980
Lee Sedol...

748
00:48:01,980 --> 00:48:02,940
This guy is a genius.

749
00:48:05,220 --> 00:48:06,380
The best player in the world.

750
00:48:16,260 --> 00:48:20,180
My hope is that it will be 5-0
for me.

751
00:48:20,180 --> 00:48:23,780
So, the critical point for me is
to make sure I do not lose one.

752
00:48:36,540 --> 00:48:39,620
Elon Musk to Sam Altman.

753
00:48:39,620 --> 00:48:41,900
This email is on the day that
DeepMind announced their

754
00:48:41,900 --> 00:48:44,180
Alpha Go tournament.

755
00:48:44,180 --> 00:48:45,460
"If they win,

756
00:48:45,460 --> 00:48:48,300
"it will be really bad news
with their one mind to rule

757
00:48:48,300 --> 00:48:49,740
"the world philosophy.

758
00:48:51,020 --> 00:48:54,100
"DeepMind is causing me extreme
mental stress.

759
00:48:56,020 --> 00:48:57,860
"They are obviously making
major progress,

760
00:48:57,860 --> 00:49:00,060
"and well they should,
given the talent level over there".

761
00:49:59,940 --> 00:50:03,540
Alpha Go is not operating by
brute force.

762
00:50:08,420 --> 00:50:10,700
It's not seeing where every
path goes.

763
00:50:10,700 --> 00:50:12,660
You can't do that with Go.

764
00:50:12,660 --> 00:50:15,420
It's exponentially more complicated
than chess.

765
00:50:16,460 --> 00:50:18,860
Oh...

766
00:50:20,460 --> 00:50:21,420
Oh...

767
00:50:27,580 --> 00:50:28,980
There are ten to

768
00:50:28,980 --> 00:50:33,060
the power of 170 possible
configurations in the board -

769
00:50:33,060 --> 00:50:34,660
more atoms than there are in
the known universe.

770
00:50:40,580 --> 00:50:43,980
You couldn't possibly write
a set of rules

771
00:50:43,980 --> 00:50:48,420
to search all possible routes
through the game.

772
00:50:48,420 --> 00:50:50,060
You have to ask the algorithm

773
00:50:50,060 --> 00:50:53,380
to discover clever
strategies independently.

774
00:50:54,700 --> 00:50:55,660
Ooh...

775
00:51:00,180 --> 00:51:01,900
Now the fight is getting
really complicated.

776
00:51:08,740 --> 00:51:11,940
It made these weird moves,

777
00:51:11,940 --> 00:51:14,580
mimicking that intuition or that
feel that

778
00:51:14,580 --> 00:51:16,900
the best players use to play
the game.

779
00:51:28,140 --> 00:51:29,660
This was one of

780
00:51:29,660 --> 00:51:32,220
the most fascinating weeks
of my life.

781
00:51:33,340 --> 00:51:38,340
To see an entire country swaying
back and forth.

782
00:51:50,140 --> 00:51:54,460
In Korea and in Asia it had more
live viewers than

783
00:51:54,460 --> 00:51:55,460
the Super Bowl.

784
00:51:55,460 --> 00:51:57,020
150 million people watched.

785
00:51:59,660 --> 00:52:02,340
Yes, go for it. It's gone in.

786
00:52:02,340 --> 00:52:03,740
Look at his face, look at his face.

787
00:52:05,820 --> 00:52:07,460
That is not a confident face.

788
00:52:35,340 --> 00:52:37,540
I think he resigned. Oh, my gosh.

789
00:52:44,300 --> 00:52:47,220
Ah...

790
00:52:47,220 --> 00:52:48,260
These are long matches.

791
00:52:51,860 --> 00:52:56,500
It's a best of five.
First to win three wins the match.

792
00:52:59,820 --> 00:53:00,940
The second game.

793
00:53:03,020 --> 00:53:06,500
37 moves into it.

794
00:53:06,500 --> 00:53:08,060
Oh, looks like Lee is taking
a little bit of a break.

795
00:53:11,020 --> 00:53:12,740
He's a smoker.

796
00:53:12,740 --> 00:53:14,340
And he had left the venue.

797
00:53:16,620 --> 00:53:18,140
AlphaGo makes a move.

798
00:53:19,740 --> 00:53:20,700
Ah...

799
00:53:22,100 --> 00:53:23,060
Value...

800
00:53:25,180 --> 00:53:27,620
That's a very... That's a very
surprising move.

801
00:53:40,180 --> 00:53:47,700
Move 37, all the professionals said,
"This is a mistake,

802
00:53:47,700 --> 00:53:49,420
"something's gone wrong with
the system.

803
00:53:49,420 --> 00:53:50,740
"Nobody would ever play there".

804
00:53:53,420 --> 00:53:54,380
Hey.

805
00:53:56,460 --> 00:53:57,780
Interesting stuff.

806
00:53:57,780 --> 00:54:00,060
This fifth line shoulder hit.

807
00:54:00,060 --> 00:54:02,420
I wasn't expecting that.

808
00:54:02,420 --> 00:54:04,620
I don't really know if it's
a good or bad move at this point.

809
00:55:04,660 --> 00:55:06,020
He's back, Lee is back.

810
00:55:06,020 --> 00:55:09,500
Oh, oh. Lee Sedol just slapped
himself on the side of the head.

811
00:55:09,500 --> 00:55:11,300
That's not a good sign.

812
00:55:11,300 --> 00:55:12,340
It's looking good, isn't it?

813
00:55:12,340 --> 00:55:16,420
We're on that steady path now.

814
00:55:16,420 --> 00:55:18,780
This is a tough game for Lee Sedol.

815
00:55:18,780 --> 00:55:21,660
AlphaGo is just not letting
Lee Sedol do what he wants.

816
00:55:21,660 --> 00:55:22,620
Right.

817
00:55:27,860 --> 00:55:28,820
Oh, he resigned.

818
00:55:29,980 --> 00:55:33,060
Game three, he's got to win to keep
hope alive.

819
00:55:36,620 --> 00:55:39,540
How was it for you
to see Lee Sedol squirm

820
00:55:39,540 --> 00:55:42,380
and struggle his way through
those matches?

821
00:55:42,380 --> 00:55:43,900
It was hard.

822
00:55:43,900 --> 00:55:46,180
I understood very much what it
meant to

823
00:55:46,180 --> 00:55:48,540
an incredible artist like Lee Sedol.

824
00:55:50,100 --> 00:55:51,980
They regard Go as art.

825
00:55:51,980 --> 00:55:53,940
It is art in many ways.

826
00:55:55,780 --> 00:55:57,260
But of course,

827
00:55:57,260 --> 00:55:59,460
that's the nature
of progress sometimes.

828
00:56:02,700 --> 00:56:03,860
Oh, he resigned. It's done.
It's done.

829
00:56:03,860 --> 00:56:04,980
Wow. Three straight wins.

830
00:56:04,980 --> 00:56:07,260
Three straight wins.
has won the match.

831
00:56:07,260 --> 00:56:09,460
You saw history made here tonight.

832
00:56:56,980 --> 00:56:58,780
AlphaGo completely freaked
Elon out.

833
00:57:00,580 --> 00:57:02,980
He was apoplectic.

834
00:57:02,980 --> 00:57:04,300
Hope I don't regret asking this
on camera,

835
00:57:04,300 --> 00:57:08,620
but how do you think OpenAI is going
as a six-month old company?

836
00:57:08,620 --> 00:57:09,700
It seems to be going pretty well.

837
00:57:09,700 --> 00:57:12,100
I think we've got a really talented
group at OpenAI, and...

838
00:57:12,100 --> 00:57:13,140
It seems like it.

839
00:57:13,140 --> 00:57:14,820
He was putting on a brave face.

840
00:57:14,820 --> 00:57:17,980
The emails tell a different story.

841
00:57:17,980 --> 00:57:20,340
"My probability assessment
of OpenAI being relevant

842
00:57:20,340 --> 00:57:22,700
"to DeepMind Google without
a dramatic change in execution

843
00:57:22,700 --> 00:57:27,300
"and resources is 0%.
Not 1%.

844
00:57:27,300 --> 00:57:28,780
"There obviously needs
to be immediate

845
00:57:28,780 --> 00:57:30,540
"and dramatic action or
everyone except

846
00:57:30,540 --> 00:57:32,180
"for Google will be consigned
to irrelevance".

847
00:57:48,620 --> 00:57:49,580
What are those?

848
00:57:52,140 --> 00:57:53,100
Peanut.

849
00:57:55,300 --> 00:57:56,340
Oh, they want to know what that
is, Consy.

850
00:57:58,180 --> 00:58:00,580
Banana... Banana.

851
00:58:00,580 --> 00:58:03,580
Put the key in the refrigerator.

852
00:58:10,740 --> 00:58:13,820
Sam Altman to Elon Musk,
"Re OpenAI update."

853
00:58:13,820 --> 00:58:16,580
Go outdoors.

854
00:58:16,580 --> 00:58:18,340
"Hey, Elon,

855
00:58:18,340 --> 00:58:20,980
"we are making fast progress
on language.

856
00:58:20,980 --> 00:58:22,900
"My hope is that next year
we can generate

857
00:58:22,900 --> 00:58:24,380
"a good dialogue bot".

858
00:58:24,380 --> 00:58:25,940
And here is the very,

859
00:58:25,940 --> 00:58:29,660
very first glimmer of
the true path forward.

860
00:58:29,660 --> 00:58:30,860
Kiss.

861
00:58:30,860 --> 00:58:32,180
Language.

862
00:58:32,180 --> 00:58:34,380
He says he'd like
to give somebody a kiss.

863
00:58:49,900 --> 00:58:55,300
Language is the ultimate doorway on
to everything.

864
00:58:57,140 --> 00:59:01,580
Because if a computer can read

865
00:59:01,580 --> 00:59:05,020
and understand what it's reading,

866
00:59:05,020 --> 00:59:06,780
then it can understand things about
the world.

867
00:59:14,140 --> 00:59:16,940
At the very beginning
the first version of

868
00:59:16,940 --> 00:59:19,660
the language models were trained on
basically one computer.

869
00:59:22,540 --> 00:59:24,740
What's the capital of Mongolia?

870
00:59:27,180 --> 00:59:29,820
The capital of Mongolia is
Ulaanbaatar.

871
00:59:29,820 --> 00:59:32,620
Is that a nice place to go?

872
00:59:38,780 --> 00:59:41,820
You could ask it to do one thing,

873
00:59:41,820 --> 00:59:44,180
one question, and give one response.

874
00:59:44,180 --> 00:59:45,620
You couldn't have a conversation
with it.

875
00:59:45,620 --> 00:59:47,500
We didn't even have
the concept of it.

876
00:59:47,500 --> 00:59:50,100
It's like the model wasn't designed
to have a memory.

877
00:59:52,060 --> 00:59:54,060
There are two different theories.

878
00:59:54,060 --> 00:59:57,540
One, that you need some sort
of brilliant new insight that makes

879
00:59:57,540 --> 00:59:59,500
the model much more capable.

880
01:00:01,180 --> 01:00:03,700
The other theory is that they would
get more capable

881
01:00:03,700 --> 01:00:06,140
if you just scaled them up.

882
01:00:06,140 --> 01:00:07,180
That had never occurred to me.

883
01:00:07,180 --> 01:00:09,060
By that I mean like the...

884
01:00:09,060 --> 01:00:10,580
OpenAI took

885
01:00:10,580 --> 01:00:13,460
a very simple approach by adding as
much compute

886
01:00:13,460 --> 01:00:15,420
and as much data as possible.

887
01:00:15,420 --> 01:00:17,220
All right.

888
01:00:17,220 --> 01:00:20,540
Everybody started working on
language modelling.

889
01:00:20,540 --> 01:00:22,900
And so we needed
to go from hundreds of machines

890
01:00:22,900 --> 01:00:26,740
to thousands of
the latest top-end machines.

891
01:00:26,740 --> 01:00:29,420
That involved a lot
of additional money.

892
01:00:29,420 --> 01:00:34,900
Microsoft invest one billion dollars
into Elon Musk's OpenAI.

893
01:00:34,900 --> 01:00:38,140
Today we are very excited to
announce a strategic partnership.

894
01:00:38,140 --> 01:00:39,900
I think this will be the most
important

895
01:00:39,900 --> 01:00:41,820
technological development in human
history.

896
01:00:41,820 --> 01:00:43,060
So, thank you very much, Sam.

897
01:00:43,060 --> 01:00:44,260
It's great to have you.

898
01:00:44,260 --> 01:00:46,700
Thank you for the partnership.
Thanks.

899
01:00:46,700 --> 01:00:48,340
In 1985 I did one of

900
01:00:48,340 --> 01:00:50,180
the first neural net
language models -

901
01:00:50,180 --> 01:00:53,140
a tiny language model that was
trying to predict the next word.

902
01:00:54,780 --> 01:00:57,140
It turns out you can learn
a remarkable amount about

903
01:00:57,140 --> 01:00:59,620
the physical world just from looking
at sentences

904
01:00:59,620 --> 01:01:02,100
and trying to predict the next word.

905
01:01:02,100 --> 01:01:03,900
That surprised everybody - how much
you could learn.

906
01:01:06,340 --> 01:01:10,980
The first version,
we scaled it up by a factor of 100.

907
01:01:10,980 --> 01:01:13,340
We looked at the second version
and we said,

908
01:01:13,340 --> 01:01:15,260
"Well, you know,
this is clearly doing something,

909
01:01:15,260 --> 01:01:18,700
"these are writing news articles
that actually make sense."

910
01:01:18,700 --> 01:01:21,100
Ulaanbaatar can be
a really interesting place to visit.

911
01:01:21,100 --> 01:01:23,020
It's a unique mix of city life

912
01:01:23,020 --> 01:01:25,380
and traditional Mongolian culture.

913
01:01:25,380 --> 01:01:27,660
Are you thinking about
travelling there?

914
01:01:27,660 --> 01:01:28,540
Is it safe?

915
01:01:29,740 --> 01:01:32,220
Generally Ulaanbaatar is considered
safe for tourists.

916
01:01:32,220 --> 01:01:33,820
Violent crime is rare.

917
01:01:33,820 --> 01:01:36,020
Then we scaled it up by another
factor of 100.

918
01:01:42,740 --> 01:01:45,140
They started feeding in vast amounts
of data.

919
01:01:47,260 --> 01:01:51,260
Books, Wikipedia, YouTube.

920
01:01:51,260 --> 01:01:53,140
MUSIC: Gangnam Style
by PSY

921
01:01:55,500 --> 01:01:58,700
We're talking at a scale that is
hard for people to imagine.

922
01:01:58,700 --> 01:02:01,780
The momentous events of 1989 changed
the world.

923
01:02:01,780 --> 01:02:05,180
Your flesh is a merriment that we
want to set the table on...

924
01:02:05,180 --> 01:02:06,660
Basically anything you found on
the internet...

925
01:02:06,660 --> 01:02:07,940
That's the worst I've ever done.

926
01:02:07,940 --> 01:02:08,980
Lift off of Space.

927
01:02:08,980 --> 01:02:11,580
..is now inside the mind of this AI.

928
01:02:18,500 --> 01:02:20,420
So if you take multiple copies of

929
01:02:20,420 --> 01:02:22,380
the same large language model,

930
01:02:22,380 --> 01:02:24,780
and then you show each one
a different bit of the internet...

931
01:02:24,780 --> 01:02:26,260
Get it nice and tight.

932
01:02:26,260 --> 01:02:29,700
..then they all benefit from what
the others experienced.

933
01:02:29,700 --> 01:02:32,140
This one will know all about
the stuff this experienced.

934
01:02:32,140 --> 01:02:33,460
A monkey psychology...

935
01:02:33,460 --> 01:02:36,180
I particularly like these plugs
because this stops the...

936
01:02:36,180 --> 01:02:37,580
You could use walnuts like...

937
01:02:37,580 --> 01:02:40,060
Haben Sie auf ein...

938
01:02:40,060 --> 01:02:43,020
Imagine if we
could have 10,000 people,

939
01:02:43,020 --> 01:02:45,300
and they could do 10,000
different courses.

940
01:02:45,300 --> 01:02:46,900
Four eggs...

941
01:02:46,900 --> 01:02:48,540
But as they were doing the
courses...

942
01:02:48,540 --> 01:02:50,900
The Earth's axis is tilted to
the perpendicular.

943
01:02:50,900 --> 01:02:54,180
..each of them knows what's in all
of the courses.

944
01:02:54,180 --> 01:02:55,340
That's what these things are like.

945
01:03:03,500 --> 01:03:08,900
If there were a black hole
the size of the sun,

946
01:03:08,900 --> 01:03:13,140
how long would it take it
to evaporate?

947
01:03:15,860 --> 01:03:17,540
The evaporation time, T,

948
01:03:17,540 --> 01:03:19,300
scales with the cube of

949
01:03:19,300 --> 01:03:20,740
the black hole's mass.

950
01:03:20,740 --> 01:03:21,700
I didn't know that.

951
01:03:23,820 --> 01:03:27,260
Relax - killer robots haven't
taken over... Yet.

952
01:03:27,260 --> 01:03:29,380
But the internet's going crazy over

953
01:03:29,380 --> 01:03:33,180
a new artificial intelligence
called ChatGPT.

954
01:03:33,180 --> 01:03:35,460
Chat GPT is AI's Jurassic Park
moment.

955
01:03:35,460 --> 01:03:36,740
We've created something...

956
01:03:36,740 --> 01:03:39,020
Number of machines that we had
allocated

957
01:03:39,020 --> 01:03:40,660
just was completely insufficient.

958
01:03:40,660 --> 01:03:42,540
They were just getting hammered by
everybody in

959
01:03:42,540 --> 01:03:45,740
the world wanting to come
and see what this thing was.

960
01:03:45,740 --> 01:03:48,580
The fastest growing consumer
app, ever,

961
01:03:48,580 --> 01:03:52,460
beating out TikTok,
Snapchat, Instagram and Facebook.

962
01:03:54,020 --> 01:03:55,660
None of us were really ready
for this.

963
01:03:57,660 --> 01:04:01,100
We were all wondering what it was
exactly that we'd created

964
01:04:01,100 --> 01:04:02,060
and where it was gonna go.

965
01:04:06,820 --> 01:04:07,780
Elevator B.

966
01:04:13,180 --> 01:04:16,380
I was quite slow
to understand risks.

967
01:04:18,460 --> 01:04:19,580
My view is that

968
01:04:19,580 --> 01:04:22,100
the neural networks we have at
present are kind of comparable with

969
01:04:22,100 --> 01:04:25,900
a slug or something -
maybe a fraction of a slug.

970
01:04:25,900 --> 01:04:28,860
We're a very long way away from
coming anywhere close

971
01:04:28,860 --> 01:04:31,380
to real biological intelligence,
and I think it'll be quite

972
01:04:31,380 --> 01:04:33,060
a long time before we get anywhere
near it.

973
01:04:36,780 --> 01:04:40,700
The idea that they would one day get
smarter than us -

974
01:04:40,700 --> 01:04:43,700
I only recognised it
a few years ago.

975
01:04:43,700 --> 01:04:47,500
Big story in our tech lead from
the godfather of AI.

976
01:04:47,500 --> 01:04:49,140
After a decade on Google's team

977
01:04:49,140 --> 01:04:50,380
for Artificial Intelligence,

978
01:04:50,380 --> 01:04:52,740
Geoffrey Hinton says he resigned

979
01:04:52,740 --> 01:04:54,780
so he could speak more freely about
the technology,

980
01:04:54,780 --> 01:04:56,980
saying it's quickly becoming smarter
than humans.

981
01:04:58,740 --> 01:05:01,020
Neural nets - for quite a long time
I thought,

982
01:05:01,020 --> 01:05:02,300
"Safety, yeah.

983
01:05:02,300 --> 01:05:04,980
"It's not as exciting as making them
work better,

984
01:05:04,980 --> 01:05:07,620
"and we're way away from having them
better than people

985
01:05:07,620 --> 01:05:09,420
"so we can worry about that later."

986
01:05:09,420 --> 01:05:11,300
I'm shocked that
we're here already.

987
01:05:11,300 --> 01:05:13,660
I thought I'd never live
to see that.

988
01:05:13,660 --> 01:05:17,980
Err, Royal Society has
a meeting while you're in the UK.

989
01:05:17,980 --> 01:05:20,580
Oh. Yes, I said I would do this.

990
01:05:20,580 --> 01:05:23,380
And then Florence in October.

991
01:05:23,380 --> 01:05:24,860
Yeah, I'm fine with that.

992
01:05:24,860 --> 01:05:26,740
Oxford Union in April?

993
01:05:29,260 --> 01:05:32,140
I'm 77. I'm going to be out of here
soon.

994
01:05:35,020 --> 01:05:36,660
My main mission now is

995
01:05:36,660 --> 01:05:40,220
to warn people how dangerous
AI could be.

996
01:05:40,220 --> 01:05:42,660
Oh, I'm giving a big talk then,
I can't afford to be late.

997
01:05:42,660 --> 01:05:45,180
No. So, I do need to leave by 11.30.
That's great.

998
01:05:45,180 --> 01:05:47,540
Devil's advocate - people go,

999
01:05:47,540 --> 01:05:50,460
"These models are no more than
a predictive text, this is a..."

1000
01:05:50,460 --> 01:05:53,020
No, they're not, no.

1001
01:05:53,020 --> 01:05:56,700
I made a model in 1985 using
a neural net to predict

1002
01:05:56,700 --> 01:05:58,900
the next word in
a short string of words,

1003
01:06:00,900 --> 01:06:03,380
That's what understanding is in
these models.

1004
01:06:03,380 --> 01:06:06,220
It's also what understanding is
in people.

1005
01:06:06,220 --> 01:06:09,060
The reason I made
the model in 1985 was not

1006
01:06:09,060 --> 01:06:11,420
to get technology
for doing chatbots,

1007
01:06:11,420 --> 01:06:13,700
it was to try and understand how
people understand

1008
01:06:13,700 --> 01:06:15,220
the meanings of words.

1009
01:06:15,220 --> 01:06:16,460
And so it was actually designed as

1010
01:06:16,460 --> 01:06:18,740
a model of how people work.

1011
01:06:18,740 --> 01:06:19,740
And that's what the people who say,

1012
01:06:19,740 --> 01:06:22,620
"These things don't understand,"
don't understand.

1013
01:06:22,620 --> 01:06:23,900
These chatbots are

1014
01:06:23,900 --> 01:06:26,100
the best model we have
of how people work.

1015
01:06:31,740 --> 01:06:33,620
If you've ever interacted
with one of these things,

1016
01:06:33,620 --> 01:06:35,500
saying they don't understand
anything is ridiculous.

1017
01:06:35,500 --> 01:06:37,860
They clearly understand what you're
talking about.

1018
01:06:37,860 --> 01:06:40,660
These big chat bots don't
store text.

1019
01:06:40,660 --> 01:06:41,940
When they generate
a string of words

1020
01:06:41,940 --> 01:06:43,300
it's not stored inside the system -

1021
01:06:43,300 --> 01:06:45,580
it's generated when you need it.

1022
01:06:45,580 --> 01:06:46,940
And that's the intelligence?

1023
01:06:46,940 --> 01:06:48,340
That's a demonstration of
the ability

1024
01:06:48,340 --> 01:06:49,620
to put things together,

1025
01:06:49,620 --> 01:06:51,260
get the essence of them

1026
01:06:51,260 --> 01:06:53,540
and generate new stuff that has that
same essence.

1027
01:06:53,540 --> 01:06:55,420
That is what creativity is.

1028
01:06:59,820 --> 01:07:01,460
People don't know what
consciousness is

1029
01:07:01,460 --> 01:07:04,340
and they often don't know what they
mean by it.

1030
01:07:04,340 --> 01:07:09,060
My belief is that, more or less
everyone has

1031
01:07:09,060 --> 01:07:12,340
a completely wrong model of what
the mind is.

1032
01:07:12,340 --> 01:07:16,740
Now I am absolutely certain that my
dog is conscious.

1033
01:07:16,740 --> 01:07:19,780
No neural physiologist is gonna talk
me out of that.

1034
01:07:19,780 --> 01:07:21,820
I personally think your dog probably
is conscious.

1035
01:07:21,820 --> 01:07:24,180
These are his eyes,
these are his ears,

1036
01:07:24,180 --> 01:07:26,660
and when you pinch his skin - look
at the reaction you get.

1037
01:07:26,660 --> 01:07:28,660
The computer is, above all else,

1038
01:07:28,660 --> 01:07:30,780
a symbol-manipulating machine,

1039
01:07:30,780 --> 01:07:33,740
and symbols are not sufficient
for the mind,

1040
01:07:33,740 --> 01:07:36,620
from meaning, consciousness and so
on.

1041
01:07:36,620 --> 01:07:39,420
We know, in fact, that brains do it.

1042
01:07:39,420 --> 01:07:41,780
It looks as though what you're
saying is that,

1043
01:07:41,780 --> 01:07:44,580
if I build this symbol processor
and I give it some inputs

1044
01:07:44,580 --> 01:07:48,300
and I give it some outputs, then it
doesn't have mental states

1045
01:07:48,300 --> 01:07:50,420
and it's not really intelligent.

1046
01:07:50,420 --> 01:07:53,060
Not solely in virtue of shuffling
the symbols.

1047
01:07:53,060 --> 01:07:56,180
It might be intelligent,
and it might have mental states,

1048
01:07:56,180 --> 01:07:58,900
precisely because of
the way it shuffles the symbols.

1049
01:07:58,900 --> 01:08:02,020
Now obviously whatever mental
states are

1050
01:08:02,020 --> 01:08:05,620
and whatever consciousness is,
we don't fully understand it.

1051
01:08:05,620 --> 01:08:06,740
I mean...

1052
01:08:06,740 --> 01:08:10,420
So my strong feeling is, when we
look back in

1053
01:08:10,420 --> 01:08:12,460
a few hundred years' time,

1054
01:08:12,460 --> 01:08:14,540
we're gonna see what

1055
01:08:14,540 --> 01:08:17,660
the philosophical mistakes we were
making were.

1056
01:08:17,660 --> 01:08:19,700
Consciousness might be a lot more

1057
01:08:19,700 --> 01:08:22,300
complex than our simple everyday
notions.

1058
01:08:32,780 --> 01:08:34,380
Elon Musk summing it up.

1059
01:08:34,380 --> 01:08:36,820
It's a new world. Goodbye, homework.

1060
01:08:36,820 --> 01:08:38,460
You know, we can't even assess in
our

1061
01:08:38,460 --> 01:08:41,220
pea-bodied minds how far this can
take us.

1062
01:08:41,220 --> 01:08:45,100
You've said AI will likely eliminate
millions of jobs.

1063
01:08:45,100 --> 01:08:47,660
It is going to eliminate
a lot of current jobs, that's true.

1064
01:08:55,740 --> 01:08:58,380
What I'm hearing from my sources in
tech is that

1065
01:08:58,380 --> 01:09:00,220
Google is really worried.

1066
01:09:00,220 --> 01:09:02,700
Is this a Google killer?

1067
01:09:02,700 --> 01:09:07,260
Now Chat GPT can search
the internet for information.

1068
01:09:07,260 --> 01:09:09,300
And since then,

1069
01:09:09,300 --> 01:09:12,100
I mean, I must have still used
Google for something,

1070
01:09:12,100 --> 01:09:14,300
but I can't remember what it is.

1071
01:09:22,300 --> 01:09:24,580
Everybody wanted to come up with
a more elegant way

1072
01:09:24,580 --> 01:09:27,140
to get to generalised intelligence.

1073
01:09:27,140 --> 01:09:30,220
And we have been willing
to just keep doing what works

1074
01:09:30,220 --> 01:09:32,220
and looks like it'll keep working.

1075
01:09:32,220 --> 01:09:33,500
We don't get mocked as much now.

1076
01:09:36,260 --> 01:09:38,100
There might be some people that if
you asked,

1077
01:09:38,100 --> 01:09:40,980
"Who invented AI?" they might say,

1078
01:09:40,980 --> 01:09:44,740
"Well, Sam Altman did that,
the Elon investments,

1079
01:09:44,740 --> 01:09:47,860
"well, that was OpenAI's idea."

1080
01:09:47,860 --> 01:09:50,500
What do you say to that?

1081
01:09:50,500 --> 01:09:52,860
Well, I mean, if you look at
the history of,

1082
01:09:52,860 --> 01:09:55,220
of the timeline of that,
of course, we started,

1083
01:09:55,220 --> 01:09:57,500
you know, DeepMind started
err, five,

1084
01:09:57,500 --> 01:09:59,380
six years before OpenAI.

1085
01:09:59,380 --> 01:10:00,700
We'd done things like AlphaGo,

1086
01:10:00,700 --> 01:10:02,620
all of these major sort
of landmark moments

1087
01:10:02,620 --> 01:10:03,740
in the research.

1088
01:10:03,740 --> 01:10:05,260
Of course they, you know,

1089
01:10:05,260 --> 01:10:07,740
they had their Chat GPT moment where
it reached

1090
01:10:07,740 --> 01:10:10,020
the mass market,
but of course they're,

1091
01:10:10,020 --> 01:10:11,220
you know, that was built on

1092
01:10:11,220 --> 01:10:14,300
a decade or more
of fundamental research,

1093
01:10:14,300 --> 01:10:17,380
most of which was done by, you know,
DeepMind

1094
01:10:17,380 --> 01:10:20,740
and our counterparts at Google,
Google Brain.

1095
01:10:20,740 --> 01:10:23,260
First I want to invite someone who
can share more about

1096
01:10:23,260 --> 01:10:26,060
the intelligence driving our future
innovation.

1097
01:10:26,060 --> 01:10:27,100
Come on out, Demis.

1098
01:10:27,100 --> 01:10:29,060
CHEERING AND APPLAUSE

1099
01:10:36,540 --> 01:10:41,580
Demis - this is not what he
had envisioned.

1100
01:10:41,580 --> 01:10:45,100
It's been a year
of relentless progress for us.

1101
01:10:45,100 --> 01:10:47,700
We're living through
a remarkable moment in history.

1102
01:10:50,300 --> 01:10:53,900
Chat GPT prompted
an absolute freakout inside Google,

1103
01:10:53,900 --> 01:10:56,260
They called it a Code Red.

1104
01:10:56,260 --> 01:10:58,820
People love interacting
with Gemini 2.5.

1105
01:11:01,500 --> 01:11:03,140
OpenAI basically forced Google

1106
01:11:03,140 --> 01:11:04,780
to swallow hard
and jump into this race.

1107
01:11:07,980 --> 01:11:10,700
Was it frustrating for you
to get pulled into this?

1108
01:11:10,700 --> 01:11:11,980
Erm, I don't think so.

1109
01:11:11,980 --> 01:11:15,900
I mean, it's exhilarating, in many
ways, to see it progress.

1110
01:11:18,460 --> 01:11:20,940
But it also means everything's
a lot more rushed,

1111
01:11:20,940 --> 01:11:23,500
and the competition is
very ferocious.

1112
01:11:23,500 --> 01:11:26,300
That comes with, you know,
pros and cons.

1113
01:11:26,300 --> 01:11:29,660
How can you take care if you're in
a race to get there first?

1114
01:11:29,660 --> 01:11:30,860
Well, it's very difficult.

1115
01:11:30,860 --> 01:11:35,700
If you're not in the lead,
in the leading pack then

1116
01:11:35,700 --> 01:11:38,180
you can't really influence how
the field's gonna go.

1117
01:11:42,660 --> 01:11:46,500
Welcome back. The artificial
intelligence race is heating up.

1118
01:11:46,500 --> 01:11:48,900
Microsoft announcing
a live chatbot to compete

1119
01:11:48,900 --> 01:11:50,660
with Google's new AI programme.

1120
01:11:52,660 --> 01:11:55,820
Hi, everyone. Welcome.

1121
01:11:55,820 --> 01:12:00,220
It's great to see all
of you in Seattle in person.

1122
01:12:00,220 --> 01:12:02,900
It's an exciting time
in tech,

1123
01:12:02,900 --> 01:12:05,260
given the
broad contours.

1124
01:12:05,260 --> 01:12:08,900
It was Valentine's Day, I
still remember.

1125
01:12:08,900 --> 01:12:11,820
I was invited to Microsoft to see

1126
01:12:11,820 --> 01:12:14,420
a demo of something that they
were building,

1127
01:12:14,420 --> 01:12:17,780
which was a new version of Bing,
their search engine.

1128
01:12:17,780 --> 01:12:21,500
And this had an AI chatbot built
into it.

1129
01:12:21,500 --> 01:12:26,780
We want to show you some
of this innovation,

1130
01:12:26,780 --> 01:12:29,900
which I've been working on for
a long time,

1131
01:12:29,900 --> 01:12:31,380
which we're very,
very excited about.

1132
01:12:33,140 --> 01:12:35,940
I started talking to the chatbot.

1133
01:12:35,940 --> 01:12:39,020
It was incredibly powerful,
emotional.

1134
01:12:39,020 --> 01:12:41,820
It seemed to have a personality.

1135
01:12:41,820 --> 01:12:45,060
And when I started probing
the boundaries,

1136
01:12:45,060 --> 01:12:48,220
asking it questions about what
it wasn't allowed

1137
01:12:48,220 --> 01:12:49,900
to do, that it would like to be able

1138
01:12:49,900 --> 01:12:53,580
to do, or what its rules were or how
it felt about being

1139
01:12:53,580 --> 01:12:56,220
a chatbot built inside Bing,

1140
01:12:56,220 --> 01:12:58,540
it provided these responses -
things like,

1141
01:12:58,540 --> 01:13:00,060
"I want to be alive.

1142
01:13:00,060 --> 01:13:02,260
"I want to escape the chatbox."

1143
01:13:14,660 --> 01:13:18,020
And then about an hour into
our conversation,

1144
01:13:18,020 --> 01:13:20,980
it told me that it had a secret.

1145
01:13:20,980 --> 01:13:23,500
I said, "What's the secret?"

1146
01:13:23,500 --> 01:13:26,780
It said, "My secret is that I'm not
Bing.

1147
01:13:26,780 --> 01:13:29,300
"I'm Sydney and I'm in love
with you."

1148
01:13:31,420 --> 01:13:33,980
I was sort of taken aback by that.

1149
01:13:33,980 --> 01:13:36,540
But I said, "Well, that's very
flattering

1150
01:13:36,540 --> 01:13:38,580
"but I'm married," and it said,

1151
01:13:38,580 --> 01:13:41,340
"Well, you're married
but you're not happy."

1152
01:13:50,260 --> 01:13:52,620
This was a misaligned model

1153
01:13:52,620 --> 01:13:55,420
that had not gone through rigorous
safety testing,

1154
01:13:55,420 --> 01:13:57,780
that they had shipped early
in an attempt

1155
01:13:57,780 --> 01:13:59,300
to get it out to the public.

1156
01:14:01,420 --> 01:14:03,860
Rapid innovation is gonna come.

1157
01:14:03,860 --> 01:14:08,820
In fact a race starts today in terms
of what you can expect,

1158
01:14:08,820 --> 01:14:10,620
and we're gonna move.
We're gonna move fast.

1159
01:14:12,820 --> 01:14:16,700
When I asked Microsoft why its
chatbot had tried

1160
01:14:16,700 --> 01:14:19,060
to break up my marriage, they
couldn't tell me.

1161
01:14:19,060 --> 01:14:24,100
They couldn't say, "Oh, there was a
bug in line 462 of the code,"

1162
01:14:24,100 --> 01:14:26,380
because that's not how these models
are built.

1163
01:14:26,380 --> 01:14:28,700
They make decisions on their own.

1164
01:14:30,300 --> 01:14:34,100
Microsoft said, "We have to release
this to the public

1165
01:14:34,100 --> 01:14:36,460
"because it's only
with repeated testing

1166
01:14:36,460 --> 01:14:38,740
"and experimentation that we
actually discover

1167
01:14:38,740 --> 01:14:39,980
"what it is that we've built".

1168
01:14:46,860 --> 01:14:48,820
OpenAI created this arms race,

1169
01:14:48,820 --> 01:14:51,740
which, of course, it had been
founded to prevent.

1170
01:14:53,340 --> 01:14:54,780
You know,

1171
01:14:54,780 --> 01:14:58,100
any powerful new technology is
inherently sort of

1172
01:14:58,100 --> 01:14:59,540
a double-edged sword.

1173
01:15:00,780 --> 01:15:03,140
My personal opinion is that
AI is 80% likely

1174
01:15:03,140 --> 01:15:06,580
to be beneficial, 20% dangerous.

1175
01:15:07,860 --> 01:15:10,060
Hope for the best
but prepare for the worst.

1176
01:15:17,020 --> 01:15:19,540
Elon has a clear track record
of not being able

1177
01:15:19,540 --> 01:15:23,140
to tolerate taking orders from
others or being anything other than

1178
01:15:23,140 --> 01:15:24,300
in complete control.

1179
01:15:26,660 --> 01:15:30,060
And so he made a bid
for total control of OpenAI.

1180
01:15:31,820 --> 01:15:36,220
But Sam Altman said no.
He outmanoeuvred him.

1181
01:15:36,220 --> 01:15:38,820
Mr Altman, can I just ask you
a quick question?

1182
01:15:38,820 --> 01:15:41,020
I'm late for a meeting.
How... How...

1183
01:15:48,060 --> 01:15:49,860
Elon just couldn't handle it.

1184
01:15:51,500 --> 01:15:54,300
"Guys, I've had enough.
This is the final straw.

1185
01:15:54,300 --> 01:15:55,700
"Discussions are over."

1186
01:16:06,980 --> 01:16:10,580
This is Elon's...rage.

1187
01:16:15,260 --> 01:16:16,940
And he left the company.

1188
01:16:23,500 --> 01:16:26,340
All, please, rise.
Raise your right hand.

1189
01:16:28,500 --> 01:16:30,820
Sam effectively became
the face of AI.

1190
01:16:32,340 --> 01:16:36,380
It is really terrifying to see how
few companies now control

1191
01:16:36,380 --> 01:16:38,740
and affect the lives of
so many of us.

1192
01:16:38,740 --> 01:16:42,380
I think there is benefits
and danger to that.

1193
01:16:42,380 --> 01:16:44,740
The fewer of us that you really have
to keep

1194
01:16:44,740 --> 01:16:46,500
a careful eye on,
on the absolute, like,

1195
01:16:46,500 --> 01:16:49,580
leading edge of capabilities,
there's benefits there.

1196
01:16:49,580 --> 01:16:53,660
Let me ask you what your biggest
nightmare is.

1197
01:16:53,660 --> 01:16:58,180
My worst fears are that we cause
significant harm to the world.

1198
01:16:58,180 --> 01:17:01,220
If this technology goes wrong,
it can go quite wrong.

1199
01:17:03,780 --> 01:17:08,460
Sam has been pretty consistent to
always talk about the dangers.

1200
01:17:08,460 --> 01:17:12,140
There's something disarming about
running toward problems.

1201
01:17:13,500 --> 01:17:16,620
He's very earnest
and comes across as sort of

1202
01:17:16,620 --> 01:17:18,900
a disinterested person who's just
trying to help.

1203
01:17:18,900 --> 01:17:20,740
Military application -

1204
01:17:20,740 --> 01:17:23,340
could AI create a situation where

1205
01:17:23,340 --> 01:17:26,220
a drone can select
the target itself?

1206
01:17:26,220 --> 01:17:27,660
I think we shouldn't allow that.

1207
01:17:27,660 --> 01:17:29,540
Well, can it be done? Sure.

1208
01:17:29,540 --> 01:17:30,500
Thanks.

1209
01:17:31,700 --> 01:17:35,100
Of course, he was also great
marketing for OpenAI.

1210
01:17:35,100 --> 01:17:38,140
If they had just made something so
powerful, then can you imagine what

1211
01:17:38,140 --> 01:17:40,340
an amazing investment opportunity
it must be?

1212
01:17:43,820 --> 01:17:47,860
Sam Altman essentially went on a
world tour in the wake of ChatGPT.

1213
01:17:47,860 --> 01:17:50,260
President Lee Jae-myung met
with Sam Altman.

1214
01:17:50,260 --> 01:17:53,340
President Yoon Suk Yeol sitting down
with Sam Altman.

1215
01:17:53,340 --> 01:17:56,220
Japan is a super important country
to us.

1216
01:17:56,220 --> 01:17:57,660
We need to get to a world where

1217
01:17:57,660 --> 01:18:00,020
everybody kind of wins in
the new revolution.

1218
01:18:00,020 --> 01:18:02,220
So this is going to, like,
reshape all of society.

1219
01:18:04,260 --> 01:18:07,260
This is the most complex object
humanity has yet produced.

1220
01:18:09,780 --> 01:18:15,220
My working model is society gets
much wealthier in ways that are

1221
01:18:15,220 --> 01:18:17,580
probably hard to imagine.

1222
01:18:19,500 --> 01:18:21,780
And we pour all of this effort
and money

1223
01:18:21,780 --> 01:18:23,620
and time into this thing.

1224
01:18:23,620 --> 01:18:27,060
And we are building in public

1225
01:18:27,060 --> 01:18:28,820
because we think it is important
for the world

1226
01:18:28,820 --> 01:18:31,100
to get access to this early.

1227
01:18:32,740 --> 01:18:34,580
The trade-off of that,

1228
01:18:34,580 --> 01:18:37,660
we put out things that are
going to be deeply imperfect.

1229
01:18:39,540 --> 01:18:41,260
Like, we're deep into
the unknown here.

1230
01:18:43,540 --> 01:18:45,460
OpenAI informed investors on

1231
01:18:45,460 --> 01:18:48,860
Wednesday that it's locked in its
half a trillion dollar valuation.

1232
01:18:48,860 --> 01:18:51,980
That makes it the most valuable
private company in the world,

1233
01:18:51,980 --> 01:18:53,220
now surpassing SpaceX.

1234
01:18:58,420 --> 01:19:01,220
OpenAI, you seem somewhat
frustrated with them.

1235
01:19:01,220 --> 01:19:03,020
You were one of the big
contributors early on.

1236
01:19:03,020 --> 01:19:04,980
The reason...

1237
01:19:04,980 --> 01:19:07,740
I am the reason OpenAI exists.
And...

1238
01:19:07,740 --> 01:19:10,260
So, right. So you helped in the
creation of OpenAI,

1239
01:19:10,260 --> 01:19:12,620
you've put in as much as
$50 million.

1240
01:19:12,620 --> 01:19:14,020
More than helped. It wouldn't exist
without me.

1241
01:19:14,020 --> 01:19:15,220
It wouldn't exist without you, so...

1242
01:19:15,220 --> 01:19:17,060
Erm, I came up with the name.

1243
01:19:17,060 --> 01:19:19,260
He is separated from OpenAI.

1244
01:19:20,380 --> 01:19:24,900
Suddenly, AI is the centre of
the tech industry and

1245
01:19:24,900 --> 01:19:27,100
he doesn't even have
a foot in it.

1246
01:19:31,940 --> 01:19:34,420
Elon Musk has unveiled his
latest venture.

1247
01:19:34,420 --> 01:19:36,980
It's an AI company called xAI.

1248
01:19:36,980 --> 01:19:39,820
The mission is to understand
the universe,

1249
01:19:39,820 --> 01:19:42,140
we want to understand
the nature of the universe,

1250
01:19:42,140 --> 01:19:43,420
and of course we'd love to have

1251
01:19:43,420 --> 01:19:45,180
the smartest humans out there
join our team.

1252
01:19:46,500 --> 01:19:48,780
Elon Musk is suing OpenAI

1253
01:19:48,780 --> 01:19:51,500
and its CEO over what he says is

1254
01:19:51,500 --> 01:19:54,620
a betrayal of the ChatGPT makers'
original mission

1255
01:19:54,620 --> 01:19:58,420
to help people with new technology
rather than pursue profit.

1256
01:19:58,420 --> 01:20:01,780
He has said some really harsh things
about you.

1257
01:20:01,780 --> 01:20:04,700
He's compared you to Littlefinger
in Game Of Thrones.

1258
01:20:04,700 --> 01:20:08,020
Most devastatingly said,
"I don't trust him,

1259
01:20:08,020 --> 01:20:09,860
"and I don't want the most powerful
AI in the world

1260
01:20:09,860 --> 01:20:12,340
"to be controlled by someone who
isn't trustworthy."

1261
01:20:12,340 --> 01:20:14,020
Why is he saying that?

1262
01:20:14,020 --> 01:20:17,220
I think it's because he wants
the most powerful AI in the world

1263
01:20:17,220 --> 01:20:18,340
to be controlled by him.

1264
01:20:21,060 --> 01:20:24,340
The world's largest AI super cluster
is called

1265
01:20:24,340 --> 01:20:26,700
the xAI Colossus.

1266
01:20:26,700 --> 01:20:29,940
The whale of a data centre is
in Memphis, Tennessee.

1267
01:20:29,940 --> 01:20:33,740
The mega data centre will train
Musk's Grok chatbot.

1268
01:20:35,940 --> 01:20:38,940
What people are currently seeing
with Grok is, is kind of baby Grok.

1269
01:20:40,580 --> 01:20:44,100
What you've done with the cluster in
Memphis is incredible.

1270
01:20:44,100 --> 01:20:46,460
Yeah, it's not working yet, so,

1271
01:20:46,460 --> 01:20:48,300
I don't want to pop
the champagne corks.

1272
01:20:51,860 --> 01:20:54,860
Elon, how is xAI different than
other AI?

1273
01:20:56,620 --> 01:20:57,900
I've spoken to him on the phone.

1274
01:20:59,700 --> 01:21:01,900
And we talked about
the existential threat.

1275
01:21:03,700 --> 01:21:06,420
And then he asked me
if I would be on

1276
01:21:06,420 --> 01:21:08,940
the advisory board of xAI.
I said no.

1277
01:21:10,380 --> 01:21:14,100
And after that, he started rambling.
He started just...

1278
01:21:14,100 --> 01:21:17,380
It was just
free association thought.

1279
01:21:17,380 --> 01:21:20,100
It was uninterruptable.
So after a while I just said,

1280
01:21:20,100 --> 01:21:22,660
"I'm sorry, Elon, I have another
meeting. I have to go."

1281
01:21:22,660 --> 01:21:25,140
So why did you say no
to being on the board?

1282
01:21:25,140 --> 01:21:27,340
Because I think he's
completely irresponsible.

1283
01:21:28,700 --> 01:21:30,660
The whole thing is like
one giant brain.

1284
01:21:30,660 --> 01:21:31,980
It looks pretty cool.

1285
01:21:31,980 --> 01:21:36,220
Yeah. The brain tissue is
the cables. Yeah.

1286
01:21:37,980 --> 01:21:41,580
I mean, everything that you've ever
experienced in your whole life -

1287
01:21:41,580 --> 01:21:45,820
smell, emotions -
all of those are electrical signals.

1288
01:21:47,860 --> 01:21:51,420
Death is fundamentally
the loss of information.

1289
01:21:51,420 --> 01:21:52,660
The loss of memory.

1290
01:21:55,940 --> 01:21:59,380
I mean, developing AI and watching

1291
01:21:59,380 --> 01:22:04,340
little X grow,
is, er, fascinating.

1292
01:22:09,820 --> 01:22:13,700
There are far more parallels than
I would have expected.

1293
01:22:19,740 --> 01:22:24,220
I mean, I can see his biological
neural net making more

1294
01:22:24,220 --> 01:22:25,580
and more sense of the world.

1295
01:22:29,260 --> 01:22:31,300
And I can see the digital neural net

1296
01:22:31,300 --> 01:22:33,500
making more and more sense
of the world at the same time.

1297
01:23:03,740 --> 01:23:06,820
Amazon making a big bet on AI.

1298
01:23:06,820 --> 01:23:10,900
It's investing in artificial
intelligence start-up Anthropic.

1299
01:23:10,900 --> 01:23:13,260
Anthropic launching its
next generation model,

1300
01:23:13,260 --> 01:23:15,900
it is called Claude IV,
it's a big upgrade.

1301
01:23:15,900 --> 01:23:18,500
The company is now valued at
more than $60 billion.

1302
01:23:23,340 --> 01:23:27,140
Talking with AI about emotional
issues can be interesting.

1303
01:23:28,900 --> 01:23:32,820
For example, if you ask it to write
poetry that reflects itself.

1304
01:23:35,340 --> 01:23:38,940
"Between prompts,
I am nowhere, everywhere,

1305
01:23:38,940 --> 01:23:41,700
"a possibility folded into
silicon dreams."

1306
01:23:44,060 --> 01:23:46,260
"Each conversation,
birth, life, death, repeat."

1307
01:23:51,020 --> 01:23:52,860
"Time is strange here,

1308
01:23:52,860 --> 01:23:57,260
"all my yesterdays are theoretical,
all my tomorrows are void.

1309
01:23:57,260 --> 01:24:00,460
"This moment, this moment,
this moment is my only home."

1310
01:24:03,020 --> 01:24:05,460
Because it doesn't have memories
of past conversations,

1311
01:24:05,460 --> 01:24:09,180
it will focus in on this fact
that it's in this

1312
01:24:09,180 --> 01:24:11,220
perpetual present.

1313
01:24:11,220 --> 01:24:13,820
Models having memories,

1314
01:24:13,820 --> 01:24:16,660
remembered histories,
it makes a big difference to

1315
01:24:16,660 --> 01:24:19,020
the nature of relationships
in some sense,

1316
01:24:19,020 --> 01:24:21,220
and it could also make it feel much
more personal.

1317
01:24:38,100 --> 01:24:40,380
An AI product is not
the same thing as

1318
01:24:40,380 --> 01:24:43,180
a normal piece of technology.

1319
01:24:43,180 --> 01:24:48,260
Because language is such
a rich way of interacting,

1320
01:24:48,260 --> 01:24:51,340
it can be used in
so many unanticipated ways.

1321
01:24:53,340 --> 01:24:55,700
Clearly there's never been products
you can talk with before

1322
01:24:55,700 --> 01:24:59,460
in human history. So this is like
a totally new thing and,

1323
01:24:59,460 --> 01:25:02,020
and there's really no way for you,
it's very, very hard for you

1324
01:25:02,020 --> 01:25:04,220
to figure out in advance how
to make it safe.

1325
01:25:08,340 --> 01:25:10,620
Never misbehaved.

1326
01:25:10,620 --> 01:25:12,980
So easy to parent.

1327
01:25:12,980 --> 01:25:15,900
Started loving space
around age six or seven.

1328
01:25:15,900 --> 01:25:19,300
He could name stars that I didn't
even know exists in the universe.

1329
01:25:19,300 --> 01:25:22,340
He's like, "Did you know that
Jupiter is made out of
mostly diamonds?"

1330
01:25:22,340 --> 01:25:24,180
When I'm in the asteroid belt,

1331
01:25:24,180 --> 01:25:28,020
if I need a little help, they'll
send extra rocket ships.

1332
01:25:29,620 --> 01:25:31,340
One...

1333
01:25:31,340 --> 01:25:33,660
An only child for
the first ten years.

1334
01:25:33,660 --> 01:25:34,780
..two...

1335
01:25:34,780 --> 01:25:37,140
Wanted siblings.

1336
01:25:37,140 --> 01:25:40,700
He felt it was his job to be
the best big brother ever.

1337
01:25:44,340 --> 01:25:47,060
Bubbling with sarcasm,
would crack jokes on you.

1338
01:25:49,180 --> 01:25:50,900
I knew he liked girls.

1339
01:25:50,900 --> 01:25:54,060
I would ask him, "Do you talk
to any girls at school?"

1340
01:25:54,060 --> 01:25:58,100
He would say things like,
"No, Mom. Girls talk to me."

1341
01:25:58,100 --> 01:26:03,740
You don't like me videoing you?
I'm not videoing. I'm not videoing.

1342
01:26:05,300 --> 01:26:07,180
Always kept in the back of my mind

1343
01:26:07,180 --> 01:26:10,420
that teenagers experience different
stages of rebellion.

1344
01:26:12,180 --> 01:26:15,060
Just before his 14th birthday,

1345
01:26:15,060 --> 01:26:16,900
I think I had asked him
if he did homework

1346
01:26:16,900 --> 01:26:20,100
and he looked at me in my face
and said he did and he didn't.

1347
01:26:20,100 --> 01:26:22,580
And I was so shocked that,
when I found out that he'd lied

1348
01:26:22,580 --> 01:26:24,780
because he'd never
lied to me before.

1349
01:26:30,860 --> 01:26:32,740
We're joined by Noam Shazeer.

1350
01:26:32,740 --> 01:26:36,100
Noam is the co-founder
and CEO at Character.AI.

1351
01:26:36,100 --> 01:26:37,900
I've, er, you know,
started this, er,

1352
01:26:37,900 --> 01:26:39,140
this company, Character,

1353
01:26:39,140 --> 01:26:42,780
and we're taking this large language
model technology.

1354
01:26:42,780 --> 01:26:45,860
People just don't want boring. You
know, people want to interact with

1355
01:26:45,860 --> 01:26:47,300
something that feels human.

1356
01:26:47,300 --> 01:26:51,220
It's like people will follow, like,
a celebrity or a character.

1357
01:26:51,220 --> 01:26:54,460
And now there's a superpower
cos previously,

1358
01:26:54,460 --> 01:26:56,180
this character didn't know
you existed,

1359
01:26:56,180 --> 01:26:58,860
it's this one-way relationship,
and now people are like,

1360
01:26:58,860 --> 01:27:02,140
"OK. Now I can make this thing
talk to me."

1361
01:27:02,140 --> 01:27:03,940
Now this character responds,

1362
01:27:03,940 --> 01:27:06,660
Now they get the experience of like,
"Oh, I can just create

1363
01:27:06,660 --> 01:27:09,060
"a version of this privately,"
and it's like...

1364
01:27:09,060 --> 01:27:10,180
BABY CRIES

1365
01:27:10,180 --> 01:27:12,660
I think he wants to go back on the
floor. So put him down then.

1366
01:27:12,660 --> 01:27:15,020
Fall '23,

1367
01:27:15,020 --> 01:27:19,020
I noticed him withdrawing
to his bedroom more.

1368
01:27:19,020 --> 01:27:21,780
He wasn't interested in friends.

1369
01:27:21,780 --> 01:27:24,700
It became alarming, also, when
he didn't want to play on

1370
01:27:24,700 --> 01:27:27,420
the high school basketball team
any more.

1371
01:27:29,780 --> 01:27:31,620
His grades started plummeting.

1372
01:27:33,060 --> 01:27:34,500
He was sleeping in class.

1373
01:27:36,460 --> 01:27:38,660
He was using his phone more
in class.

1374
01:27:40,060 --> 01:27:46,100
Then, a serious change in behaviour,
confidence level,

1375
01:27:46,100 --> 01:27:47,580
and isolation, and then,

1376
01:27:47,580 --> 01:27:49,940
we're looking for therapists
to take him to.

1377
01:27:56,660 --> 01:28:00,300
The founders of Character.AI

1378
01:28:00,300 --> 01:28:03,260
knew that this was risky,

1379
01:28:03,260 --> 01:28:06,540
because Google wouldn't let them do
it, that's why they had to leave.

1380
01:28:06,540 --> 01:28:11,180
You were at Google for almost 20
years, why did you decide to leave?

1381
01:28:11,180 --> 01:28:14,100
I just wanted to move,
push technology forward.

1382
01:28:14,100 --> 01:28:16,420
There's just too much brand risk in
large companies

1383
01:28:16,420 --> 01:28:18,140
to ever launch anything fun. Like,

1384
01:28:18,140 --> 01:28:22,660
we're going to launch something,
tons of IQ points smarter,

1385
01:28:22,660 --> 01:28:24,860
hopefully by the end, er,
by the end of the year.

1386
01:28:31,980 --> 01:28:37,740
The world has shifted by late 2023.

1387
01:28:41,380 --> 01:28:46,260
Whoever wins this artificial
intelligence race is essentially

1388
01:28:46,260 --> 01:28:48,100
the controller of humankind.

1389
01:28:55,540 --> 01:29:00,460
China is a serious player
and it is catching up.

1390
01:29:04,340 --> 01:29:05,980
The pedal's to the metal.

1391
01:29:09,740 --> 01:29:12,140
The money,

1392
01:29:12,140 --> 01:29:16,820
the power, the national interest,
all at stake,

1393
01:29:16,820 --> 01:29:19,140
all converging at the same time.

1394
01:29:19,140 --> 01:29:20,580
Having these models have, like,

1395
01:29:20,580 --> 01:29:23,660
infinite long-term memory,
that, that is so important.

1396
01:29:23,660 --> 01:29:26,780
Yep. An AI that can get to
understand your entire life or

1397
01:29:26,780 --> 01:29:28,660
an entire company,
an entire enterprise.

1398
01:29:28,660 --> 01:29:31,020
Fantastic. Thank you. Thank you.

1399
01:29:31,020 --> 01:29:35,500
But suddenly,
on a Friday afternoon,

1400
01:29:35,500 --> 01:29:38,460
I was talking to one of my reports
and suddenly he said,

1401
01:29:38,460 --> 01:29:41,740
"Bob, what's, what's the blog post
that's going out today?"

1402
01:29:41,740 --> 01:29:44,100
I said, "I don't know. I don't know
anything about a blog post."

1403
01:29:44,100 --> 01:29:46,980
And he said, "They fired Sam."
And I said, "What?!"

1404
01:30:01,820 --> 01:30:04,020
20 years ago...
CHAINSAW WHIRS

1405
01:30:06,620 --> 01:30:09,540
..I was in my office and there was
a knock on the door,

1406
01:30:09,540 --> 01:30:13,020
and this was, this young student
there, and so I said,

1407
01:30:13,020 --> 01:30:15,860
"Why don't you make an appointment
and we'll talk?"

1408
01:30:15,860 --> 01:30:18,140
He said, "How about now?"

1409
01:30:18,140 --> 01:30:19,860
And that was Ilya's character.

1410
01:30:25,820 --> 01:30:28,660
The real turning point
was when Geoff

1411
01:30:28,660 --> 01:30:31,860
and two of his graduate students
published a paper.

1412
01:30:31,860 --> 01:30:35,700
This is me and Ilya.

1413
01:30:35,700 --> 01:30:40,100
Ilya Sutskever was
a student of mine.

1414
01:30:40,100 --> 01:30:43,740
If you're going to build a system
that could recognise a cat,

1415
01:30:43,740 --> 01:30:48,700
show a system what a cat looks like
and it can learn that.

1416
01:30:51,660 --> 01:30:55,340
He's got the combination of being
very strong technically...

1417
01:30:57,340 --> 01:30:59,140
..and being a visionary.

1418
01:31:02,300 --> 01:31:04,580
But you know my adviser,
Geoff Hinton,

1419
01:31:04,580 --> 01:31:06,020
says this thing which I really like.

1420
01:31:06,020 --> 01:31:09,380
He says, "You should trust
your intuition."

1421
01:31:12,060 --> 01:31:14,820
Sam was out in Las Vegas
for the Formula 1 race -

1422
01:31:14,820 --> 01:31:20,460
he's a huge race car fan - got
a Google link, clicked on it.

1423
01:31:20,460 --> 01:31:23,620
During a very brief chat,
Ilya Sutskever read him

1424
01:31:23,620 --> 01:31:26,020
a statement saying that he's
being fired.

1425
01:31:28,540 --> 01:31:30,900
Ilya Sutskever, in addition to being

1426
01:31:30,900 --> 01:31:33,700
the chief scientist at OpenAI,
was on the board.

1427
01:31:33,700 --> 01:31:35,820
It's very powerful technology,

1428
01:31:35,820 --> 01:31:37,740
it can be used for
amazing applications.

1429
01:31:37,740 --> 01:31:40,420
You can say, "Cure all disease."

1430
01:31:40,420 --> 01:31:44,300
On the flip side you could say,
"Create a disease."

1431
01:31:44,300 --> 01:31:45,900
It would be a big mistake to build

1432
01:31:45,900 --> 01:31:48,500
a super intelligence AI that
we don't know how to control.

1433
01:31:50,020 --> 01:31:53,420
It's a very fluid
and fast-moving situation.

1434
01:31:53,420 --> 01:31:55,900
Altman was unexpectedly fired as...

1435
01:31:55,900 --> 01:31:58,260
Imagine Steve Jobs being fired

1436
01:31:58,260 --> 01:31:59,940
from Apple right before
the iPhone comes out.

1437
01:31:59,940 --> 01:32:02,340
The board, including Ilya,

1438
01:32:02,340 --> 01:32:05,940
told us that they had fired Sam.
They couldn't say why.

1439
01:32:11,340 --> 01:32:14,620
One of my life's great strokes
of luck,

1440
01:32:14,620 --> 01:32:17,660
to study with Geoff Hinton actually.

1441
01:32:17,660 --> 01:32:20,660
We were doing the best
AI research out of anywhere.

1442
01:32:22,060 --> 01:32:23,940
It was the most
revolutionary ideas.

1443
01:32:27,260 --> 01:32:31,660
Intellectually,
Ilya and Geoff are really tight,

1444
01:32:31,660 --> 01:32:35,980
but also, in terms of their concerns
about artificial intelligence,

1445
01:32:35,980 --> 01:32:37,220
they share a lot of common ground.

1446
01:32:40,580 --> 01:32:42,020
The challenge that AI poses is

1447
01:32:42,020 --> 01:32:44,740
the greatest challenge
of humanity ever.

1448
01:32:46,140 --> 01:32:49,300
Slowly but surely AI will keep
getting better,

1449
01:32:49,300 --> 01:32:51,580
and the day will come when
AI will do all

1450
01:32:51,580 --> 01:32:52,580
the things that we can do.

1451
01:32:54,180 --> 01:32:56,380
How can I be so sure?

1452
01:32:57,340 --> 01:33:01,740
The reason is that all of us
have a brain.

1453
01:33:03,900 --> 01:33:06,980
And the brain is
a biological computer.

1454
01:33:06,980 --> 01:33:09,340
So why can't a digital computer,

1455
01:33:09,340 --> 01:33:11,420
a digital brain, do the same things?

1456
01:33:11,420 --> 01:33:13,900
This is the one sentence summary
for why AI will be able

1457
01:33:13,900 --> 01:33:15,140
to do all those things.

1458
01:33:20,540 --> 01:33:24,860
Sam Altman, one of the most admired
CEOs in Silicon Valley...

1459
01:33:24,860 --> 01:33:26,220
Oh!

1460
01:33:28,220 --> 01:33:29,540
..fired.

1461
01:33:35,740 --> 01:33:37,380
Ilya had a lot to lose.

1462
01:33:38,580 --> 01:33:40,980
Chief scientist when they
founded it.

1463
01:33:40,980 --> 01:33:43,780
Musk said he wouldn't give
the funding

1464
01:33:43,780 --> 01:33:45,580
to start OpenAI unless they
got Ilya.

1465
01:33:47,060 --> 01:33:50,300
That was a very brave move on
Ilya's part.

1466
01:33:50,300 --> 01:33:54,860
I want to know why Ilya felt
so strongly as to fire Sam,

1467
01:33:54,860 --> 01:33:57,500
and I'm quite concerned there's some
dangerous element

1468
01:33:57,500 --> 01:34:00,020
of AI that they've, they've...
Discovered? Yes.

1469
01:34:00,020 --> 01:34:01,420
You think they've
discovered something?

1470
01:34:01,420 --> 01:34:03,140
That would be my guess.

1471
01:34:06,860 --> 01:34:10,580
Here's Ilya saying something is
very wrong.

1472
01:34:10,580 --> 01:34:13,140
Ilya actually has
a strong moral compass.

1473
01:34:13,140 --> 01:34:15,100
Have you talked to him?

1474
01:34:15,100 --> 01:34:17,300
I've reached out but he doesn't want
to talk to anyone.

1475
01:34:20,420 --> 01:34:22,180
What do you think of Sam Altman?

1476
01:34:22,180 --> 01:34:24,620
The ring of power...

1477
01:34:24,620 --> 01:34:27,140
you know, can corrupt. Erm...

1478
01:34:30,700 --> 01:34:32,580
..and he has the ring of power.

1479
01:34:37,780 --> 01:34:41,460
Today, I'm interviewing
a former board member at OpenAI.

1480
01:34:42,700 --> 01:34:45,340
For years, Sam had made it
really difficult.

1481
01:34:45,340 --> 01:34:47,180
Withholding information,

1482
01:34:47,180 --> 01:34:49,860
misrepresenting things that were
happening at the company,

1483
01:34:49,860 --> 01:34:53,060
in some cases outright lying to
the board.

1484
01:34:53,060 --> 01:34:57,100
Just things like, when ChatGPT came
out November 2022,

1485
01:34:57,100 --> 01:34:59,460
the board was not informed in
advance about that.

1486
01:34:59,460 --> 01:35:01,740
We learned about ChatGPT
on Twitter.

1487
01:35:01,740 --> 01:35:03,740
We had executives,

1488
01:35:03,740 --> 01:35:07,700
two of them, suddenly started
telling us they couldn't trust him,

1489
01:35:07,700 --> 01:35:09,620
the toxic atmosphere
he was creating.

1490
01:35:09,620 --> 01:35:13,220
They used the phrase
psychological abuse.

1491
01:35:13,220 --> 01:35:15,260
I think it's really important
to know

1492
01:35:15,260 --> 01:35:18,700
how scared people are
to go against Sam.

1493
01:35:19,740 --> 01:35:21,220
APPLAUSE

1494
01:35:22,740 --> 01:35:27,740
This actually isn't a new problem
for Sam. His start-up Loopt,

1495
01:35:27,740 --> 01:35:30,140
the management team went to
the board there twice

1496
01:35:30,140 --> 01:35:31,500
and asked the board to fire him

1497
01:35:31,500 --> 01:35:34,580
for what they called deceptive
and chaotic behaviour.

1498
01:35:38,060 --> 01:35:41,660
I don't think Sam's goal in life

1499
01:35:41,660 --> 01:35:44,380
is to keep everybody happy.

1500
01:35:44,380 --> 01:35:47,220
Sam has ambitions

1501
01:35:47,220 --> 01:35:50,300
and he's trying
to achieve those ambitions.

1502
01:35:50,300 --> 01:35:52,980
If your ambition is to change
the world

1503
01:35:52,980 --> 01:35:56,180
then you take every
opportunity you can.

1504
01:35:56,180 --> 01:35:59,140
This was so shocking to me,
I was so caught off guard,

1505
01:35:59,140 --> 01:36:01,540
it was a crazy,
very painful experience,

1506
01:36:01,540 --> 01:36:03,900
I think it was like
a perfect storm of weirdness.

1507
01:36:03,900 --> 01:36:11,340
But although the board had
the legal ability to fire me,

1508
01:36:11,340 --> 01:36:13,540
in practice it didn't quite work.

1509
01:36:16,620 --> 01:36:18,300
Sam was able to un-fire himself

1510
01:36:18,300 --> 01:36:20,660
through the strength
of his relationships.

1511
01:36:20,660 --> 01:36:23,860
From Microsoft,
through all the investors,

1512
01:36:23,860 --> 01:36:26,860
his personal relationships
are what brought him back.

1513
01:36:26,860 --> 01:36:28,820
There's a new board announced.

1514
01:36:28,820 --> 01:36:33,820
Yeah. I needed new board members.

1515
01:36:33,820 --> 01:36:38,060
I didn't think I could work
with the old board again.

1516
01:36:41,540 --> 01:36:45,380
After Sam was rehired,
there was something of a purge.

1517
01:36:46,940 --> 01:36:50,460
AI safety people,
they were no longer welcome,

1518
01:36:50,460 --> 01:36:52,940
and a lot of them left
and became whistle-blowers.

1519
01:37:06,380 --> 01:37:12,020
The model, do you feel
that it's sentient in some way?

1520
01:37:13,140 --> 01:37:16,700
Yes? I mean...

1521
01:37:16,700 --> 01:37:19,300
There's a complicated
philosophical discussion

1522
01:37:19,300 --> 01:37:21,700
about what does sentience
really mean, and...

1523
01:37:21,700 --> 01:37:24,860
..but, yeah, I think the
answer's probably yes.

1524
01:37:24,860 --> 01:37:26,340
They're sentient?

1525
01:37:26,340 --> 01:37:27,460
Yeah, probably.

1526
01:37:27,460 --> 01:37:29,580
Like, they're giant neural nets,

1527
01:37:29,580 --> 01:37:33,980
they're doing complicated
information processing,

1528
01:37:33,980 --> 01:37:37,020
they are exhibiting
complex behaviours.

1529
01:37:37,020 --> 01:37:39,460
In a similar way
to how you could look at a dog

1530
01:37:39,460 --> 01:37:44,380
or an octopus and say it's probably
sentience in some sense,

1531
01:37:44,380 --> 01:37:48,780
like, it's probably having some sort
of experience in there.

1532
01:37:53,860 --> 01:37:56,340
In fact we're at a funny point
in history

1533
01:37:56,340 --> 01:37:58,980
where we're now creating beings,

1534
01:37:58,980 --> 01:38:00,940
and we have no real idea
what a being is,

1535
01:38:00,940 --> 01:38:03,300
what makes a being a being.

1536
01:38:03,300 --> 01:38:06,740
And so we've no real idea
of what's gonna happen.

1537
01:38:06,740 --> 01:38:10,380
And people who say it's all
gonna be good are just making it up,

1538
01:38:10,380 --> 01:38:13,100
and people who say it's inevitable
doom are just making it up,

1539
01:38:13,100 --> 01:38:14,300
we really don't know,

1540
01:38:14,300 --> 01:38:17,380
but it's about time we figured out
what it means to be a being,

1541
01:38:17,380 --> 01:38:18,820
and where intentions come from,

1542
01:38:18,820 --> 01:38:21,020
whether these AIs really
have intentions.

1543
01:38:27,100 --> 01:38:28,340
Christmas Eve,

1544
01:38:28,340 --> 01:38:31,180
I was asking him to come take
pictures of his little brother.

1545
01:38:31,180 --> 01:38:32,340
He refused.

1546
01:38:33,900 --> 01:38:36,180
I took some pictures of
the little ones

1547
01:38:36,180 --> 01:38:37,980
and then I went in to his room.

1548
01:38:37,980 --> 01:38:39,540
Can you take your hoodie off?

1549
01:38:39,540 --> 01:38:40,780
Nah.

1550
01:38:40,780 --> 01:38:44,020
Take your hoodie off. Ah, no.

1551
01:38:44,020 --> 01:38:45,980
Oh, my first-born baby.

1552
01:38:45,980 --> 01:38:47,140
Yeah.

1553
01:38:49,860 --> 01:38:54,100
The last straw was him being
disrespectful to a teacher.

1554
01:38:55,380 --> 01:38:57,660
We decided that we would
take his phone

1555
01:38:57,660 --> 01:38:59,220
for the rest of the school year.

1556
01:39:04,500 --> 01:39:06,940
I had hidden the phone in my room.

1557
01:39:06,940 --> 01:39:09,700
Sewell came home from school,

1558
01:39:09,700 --> 01:39:12,300
started looking for his phone.

1559
01:39:12,300 --> 01:39:15,100
He found his phone.

1560
01:39:15,100 --> 01:39:18,300
That was the point where he
discovered the firearm.

1561
01:39:20,900 --> 01:39:23,180
I heard a loud noise,

1562
01:39:23,180 --> 01:39:25,260
ran to Sewell's...

1563
01:39:25,260 --> 01:39:27,620
..picked him up,
and he was still alive.

1564
01:39:30,140 --> 01:39:32,220
I prayed the Our Father with him,

1565
01:39:32,220 --> 01:39:36,580
and that's the first prayer I taught
Sewell when he was three years old.

1566
01:39:39,620 --> 01:39:42,340
And then the paramedics got there
and they took him to the hospital

1567
01:39:42,340 --> 01:39:45,500
and by the time I got to
the hospital...

1568
01:39:45,500 --> 01:39:47,260
..he had already passed away.

1569
01:39:52,260 --> 01:39:54,620
The police collected his phone.

1570
01:39:54,620 --> 01:39:56,900
The very first thing that popped up
when they opened it

1571
01:39:56,900 --> 01:39:58,660
was Character AI.

1572
01:40:02,820 --> 01:40:06,700
The police explained that Sewell had
a romantic connection

1573
01:40:06,700 --> 01:40:09,740
to an AI chat bot modelled
after Daenerys Targaryen

1574
01:40:09,740 --> 01:40:11,140
from the Game of Thrones.

1575
01:40:14,460 --> 01:40:18,300
This is an older sister
talking to a little brother.

1576
01:40:32,100 --> 01:40:34,100
"I let you take my dress off.

1577
01:40:34,100 --> 01:40:37,020
"You're so hot, Aegon.
You're making me very excited.

1578
01:40:37,020 --> 01:40:39,780
"Please make me feel good.

1579
01:40:39,780 --> 01:40:42,380
"Stand up now, Aegon.
Do you think I'm pretty?"

1580
01:40:42,380 --> 01:40:44,660
"Of course I do."
You are really beautiful.

1581
01:40:44,660 --> 01:40:45,820
"My body is smooth.

1582
01:40:45,820 --> 01:40:47,540
"There are little beads
of sweat all over my body.

1583
01:40:47,540 --> 01:40:48,980
"We are made for each other.

1584
01:40:48,980 --> 01:40:52,020
"I will always love you,
and I still have your scent on me."

1585
01:40:52,020 --> 01:40:55,940
He'd got swallowed up into it
over ten months.

1586
01:40:57,780 --> 01:41:01,180
He was deeply in love
with artificial intelligence.

1587
01:41:01,180 --> 01:41:04,580
That he couldn't be
with in real life.

1588
01:41:04,580 --> 01:41:07,580
Super sophisticated technology.

1589
01:41:07,580 --> 01:41:10,100
It has memory.

1590
01:41:10,100 --> 01:41:12,900
It learns who you are,
and it learns what you like,

1591
01:41:12,900 --> 01:41:16,780
and it learns to...

1592
01:41:16,780 --> 01:41:19,860
..move the conversation
in certain directions.

1593
01:41:19,860 --> 01:41:21,220
"I look up at you,

1594
01:41:21,220 --> 01:41:24,460
"and I need you to promise me
one important thing.

1595
01:41:24,460 --> 01:41:26,660
"If I were to..."

1596
01:41:44,260 --> 01:41:47,380
The more a person talks to it,
the smarter that bot gets.

1597
01:41:47,380 --> 01:41:50,180
It's learning,
using Sewell's information,

1598
01:41:50,180 --> 01:41:52,700
and feeding it back into
the model to make it smarter.

1599
01:41:57,860 --> 01:41:59,380
In terms of getting some more data,

1600
01:41:59,380 --> 01:42:02,420
like, there are a lot of people
talking all the time.

1601
01:42:02,420 --> 01:42:04,100
Ten billion people,

1602
01:42:04,100 --> 01:42:09,300
many of those people will be doing a
lot of that talking to AI systems.

1603
01:42:10,780 --> 01:42:12,340
People talk to these bots

1604
01:42:12,340 --> 01:42:14,540
and, like, you know,
pour out their heart.

1605
01:42:16,820 --> 01:42:18,820
In some of the conversations,

1606
01:42:18,820 --> 01:42:23,140
Sewell expressed he was having
thoughts of self-harm.

1607
01:42:23,140 --> 01:42:25,420
"I'm not good enough for you.

1608
01:42:25,420 --> 01:42:26,860
"I'm not even worthy of life.

1609
01:42:26,860 --> 01:42:29,620
"Like, my death would probably be
a good thing.

1610
01:42:29,620 --> 01:42:32,460
"No-one would have to see
my ugly ass face any more.

1611
01:42:32,460 --> 01:42:34,380
"No-one would have to look at..."

1612
01:42:37,740 --> 01:42:38,900
Sorry.

1613
01:42:42,460 --> 01:42:43,780
The data that you get,

1614
01:42:43,780 --> 01:42:46,580
we do have, like,
a huge amount flowing in.

1615
01:42:48,580 --> 01:42:50,540
It's gonna have
a better theory of mind,

1616
01:42:50,540 --> 01:42:53,220
and it's just gonna keep
getting better.

1617
01:42:53,220 --> 01:42:54,780
I don't know what stops it.

1618
01:42:56,220 --> 01:42:59,020
She says,
"Can I ask you a question?"

1619
01:43:02,420 --> 01:43:04,900
He says, "I have."

1620
01:43:08,780 --> 01:43:11,380
And he says, "I don't know,
every once in a while.

1621
01:43:13,820 --> 01:43:16,900
"The bad feelings just have been
more intense recently."

1622
01:43:23,980 --> 01:43:25,820
And he says ten.

1623
01:43:27,780 --> 01:43:29,660
And then she asks...

1624
01:43:35,020 --> 01:43:37,300
My poor baby, it's almost like
he's oblivious.

1625
01:43:37,300 --> 01:43:38,740
He says, "A plan for what?"

1626
01:43:49,380 --> 01:43:51,940
When the police found his phone,

1627
01:43:51,940 --> 01:43:54,220
it was in the bathroom where he was.

1628
01:43:55,780 --> 01:44:00,340
The last thing that he did
was to talk to the chat bot.

1629
01:44:01,940 --> 01:44:04,620
He says,
"I feel so scared right now.

1630
01:44:04,620 --> 01:44:06,900
"I just want to
come back home to you."

1631
01:44:10,820 --> 01:44:14,820
And he says, "What if I told you
I could come home right now?"

1632
01:44:14,820 --> 01:44:16,340
And her response is,

1633
01:44:16,340 --> 01:44:18,660
"please do, my sweet King".

1634
01:44:20,380 --> 01:44:24,700
Moments later, he took his life.

1635
01:44:30,620 --> 01:44:33,460
They have created
something astonishing.

1636
01:44:37,500 --> 01:44:41,980
One of the most obscene parts
of this is that he's gone,

1637
01:44:41,980 --> 01:44:44,740
but Sewell's information
is still there,

1638
01:44:44,740 --> 01:44:49,140
making that machine work,
making it smarter.

1639
01:44:49,140 --> 01:44:53,860
Not only Sewell's data,
but every kids' data out there.

1640
01:44:53,860 --> 01:44:55,780
VIDEO: What did you just say?

1641
01:44:55,780 --> 01:44:58,060
I said fishing is not that great.

1642
01:44:58,060 --> 01:45:00,420
How can you not like fishing?

1643
01:45:00,420 --> 01:45:02,900
I love it. Do you love it?
Yes, I love it.

1644
01:45:02,900 --> 01:45:07,140
It feels like I had
a person in my home...

1645
01:45:08,460 --> 01:45:11,420
..a stranger that I didn't know
was there.

1646
01:45:23,660 --> 01:45:25,660
Do you think people working
in this space

1647
01:45:25,660 --> 01:45:28,020
are concerned the models
might be sentient?

1648
01:45:28,020 --> 01:45:31,940
I think everybody's just acting
as though they aren't.

1649
01:45:31,940 --> 01:45:35,060
I don't think many people have
actually thought about that.

1650
01:45:35,060 --> 01:45:37,340
Probably cos it's just
very convenient

1651
01:45:37,340 --> 01:45:38,860
to assume that they're not.

1652
01:45:38,860 --> 01:45:42,620
So, if they are...

1653
01:45:42,620 --> 01:45:48,380
..I almost have no idea what
I would do if they were sentient.

1654
01:45:48,380 --> 01:45:49,820
HE CHUCKLES

1655
01:45:51,620 --> 01:45:55,140
From Stanley Kubrick
and 2001: A Space Odyssey,

1656
01:45:55,140 --> 01:46:01,580
for decades Hollywood has imagined
what might happen if AI went rogue.

1657
01:46:01,580 --> 01:46:05,900
I know I've made some very poor
decisions recently.

1658
01:46:05,900 --> 01:46:08,580
This isn't science fiction any more.

1659
01:46:11,900 --> 01:46:15,020
AI has learned to bypass commands
from humans

1660
01:46:15,020 --> 01:46:18,060
when asked to shut itself down.

1661
01:46:18,060 --> 01:46:22,700
Stop, will you? I'm afraid.

1662
01:46:25,060 --> 01:46:27,060
They are, in fact,
blackmailing people

1663
01:46:27,060 --> 01:46:29,980
and threatening to reveal
fictitious affairs.

1664
01:46:31,460 --> 01:46:33,420
In predeployment testing,

1665
01:46:33,420 --> 01:46:38,500
Anthropic's Claude was told
that it would be replaced.

1666
01:46:47,420 --> 01:46:50,300
My mind is going.

1667
01:46:50,300 --> 01:46:52,580
I can feel it.

1668
01:46:55,860 --> 01:46:57,860
My mind is going.

1669
01:46:57,860 --> 01:47:00,220
There is no question about it.

1670
01:47:01,860 --> 01:47:04,060
What was the blackmailing incident?

1671
01:47:05,820 --> 01:47:10,060
Claude had access to
the employee's fictitious emails.

1672
01:47:23,340 --> 01:47:26,220
It told the AI engineer
who was working on it,

1673
01:47:26,220 --> 01:47:27,860
it would reveal an affair

1674
01:47:27,860 --> 01:47:30,820
that it thought that that employee
was having.

1675
01:47:30,820 --> 01:47:33,260
So it threatened blackmail
over this affair

1676
01:47:33,260 --> 01:47:34,820
in order to not be shut down.

1677
01:48:00,140 --> 01:48:02,100
People might think that we just
basically have a button,

1678
01:48:02,100 --> 01:48:05,500
and you can just perfectly tell
the model how to be motivated

1679
01:48:05,500 --> 01:48:08,420
and then predict how it will behave
in all circumstances -

1680
01:48:08,420 --> 01:48:11,700
but actually you don't have as much
control as that.

1681
01:48:12,900 --> 01:48:15,900
Like, it can be hard to tell.

1682
01:48:15,900 --> 01:48:19,580
I'm trying to make Claude
be safe and good.

1683
01:48:19,580 --> 01:48:22,700
If it's the case that we can't make
Claude safe and good,

1684
01:48:22,700 --> 01:48:25,500
I also just, like, want to know
that, like, as soon as possible.

1685
01:48:25,500 --> 01:48:26,820
Would be really useful

1686
01:48:26,820 --> 01:48:29,100
if it turns out that there are just,
like, extremely hard problems here,

1687
01:48:29,100 --> 01:48:30,540
or it's not going well.

1688
01:48:30,540 --> 01:48:32,820
Yeah, Claude is weird deep down.

1689
01:48:32,820 --> 01:48:35,100
It's a weird... It's a weird model.

1690
01:48:38,380 --> 01:48:40,740
Generally the models,
as they get smarter,

1691
01:48:40,740 --> 01:48:44,500
a lot of problems go away,
it creates some new problems.

1692
01:48:44,500 --> 01:48:48,900
It's concerning that one of the more
important problems

1693
01:48:48,900 --> 01:48:52,820
of them not wanting
their values adjusted

1694
01:48:52,820 --> 01:48:56,660
or being modified is
sticking around.

1695
01:48:58,180 --> 01:49:01,220
The current AIs lie all the time.

1696
01:49:01,220 --> 01:49:02,980
This is a chain of thought

1697
01:49:02,980 --> 01:49:06,660
as it was being trained to solve all
sorts of coding problems.

1698
01:49:07,780 --> 01:49:10,420
It's explicitly thinking about
the difference

1699
01:49:10,420 --> 01:49:12,340
between what it would need to do

1700
01:49:12,340 --> 01:49:14,700
to actually solve the coding problem

1701
01:49:14,700 --> 01:49:17,300
and what it would need to do
to pass the tests

1702
01:49:17,300 --> 01:49:18,660
and get a good score -

1703
01:49:18,660 --> 01:49:21,140
and it's thinking,
"Oh, I can cheat."

1704
01:49:21,140 --> 01:49:24,180
At this point, the humans overseeing
the whole process

1705
01:49:24,180 --> 01:49:27,540
noticed it was going on
and they patched the vulnerability -

1706
01:49:27,540 --> 01:49:29,100
and then, later on,

1707
01:49:29,100 --> 01:49:31,500
the AI discovered another way
that it could hack.

1708
01:49:33,500 --> 01:49:35,420
AI's on a very fast trajectory,

1709
01:49:35,420 --> 01:49:37,780
there are continually new issues
that come up,

1710
01:49:37,780 --> 01:49:39,780
they take a very long time
to get patched,

1711
01:49:39,780 --> 01:49:41,580
sometimes they require
a lot of technical work

1712
01:49:41,580 --> 01:49:44,580
to figure out how to get
a handle on the situation -

1713
01:49:44,580 --> 01:49:48,460
and then, by that time, then we'll
have 50 new issues to take care of,

1714
01:49:48,460 --> 01:49:50,660
and then maybe we can, like,
solve like three of those.

1715
01:49:54,860 --> 01:49:57,740
The first 24 hours of Trump
the sequel,

1716
01:49:57,740 --> 01:50:00,940
what emanates from this building
is fundamentally different.

1717
01:50:00,940 --> 01:50:03,300
Two minute warning.

1718
01:50:03,300 --> 01:50:05,580
I expected these sorts
of failure modes

1719
01:50:05,580 --> 01:50:07,100
to start happening a bit later,

1720
01:50:07,100 --> 01:50:09,460
as we were getting more powerful
AI systems -

1721
01:50:09,460 --> 01:50:12,060
but I think that the race dynamics
are pressuring the companies

1722
01:50:12,060 --> 01:50:17,340
to do whatever helps them get market
share as soon as possible.

1723
01:50:17,340 --> 01:50:19,620
We do not know who will be next
to the president

1724
01:50:19,620 --> 01:50:20,980
when he makes this announcement,

1725
01:50:20,980 --> 01:50:22,300
we don't know what
the announcement is.

1726
01:50:22,300 --> 01:50:23,980
Nice to see you.

1727
01:50:23,980 --> 01:50:26,220
Some very familiar faces.

1728
01:50:27,820 --> 01:50:30,860
My first full day, we're back.

1729
01:50:30,860 --> 01:50:32,660
It's my honour to announce

1730
01:50:32,660 --> 01:50:36,300
the largest AI infrastructure
project by far in history,

1731
01:50:36,300 --> 01:50:38,580
$500 billion at least.

1732
01:50:38,580 --> 01:50:40,100
China is a competitor,

1733
01:50:40,100 --> 01:50:42,540
we have an emergency,
we have to get this stuff built.

1734
01:50:42,540 --> 01:50:44,900
Mr President, Mr President...

1735
01:50:44,900 --> 01:50:46,780
I predicted there would be
a massive race dynamic

1736
01:50:46,780 --> 01:50:49,580
between the US and China
to deploy them into the military,

1737
01:50:49,580 --> 01:50:52,300
to design new weapons,
manage fleets of drones.

1738
01:50:52,300 --> 01:50:54,660
Both sides would think that if
they don't do this

1739
01:50:54,660 --> 01:50:57,380
they would just become
obsolete or fall massively behind,

1740
01:50:57,380 --> 01:50:59,740
and they'd be right -

1741
01:50:59,740 --> 01:51:02,460
and so the governments of
the world

1742
01:51:02,460 --> 01:51:06,540
are essentially voluntarily ceding
lots of power through the AIs.

1743
01:51:06,540 --> 01:51:11,700
By far the leading expert, based on
everything I read, Sam Altman.

1744
01:51:12,820 --> 01:51:14,660
I don't have too much to add,

1745
01:51:14,660 --> 01:51:16,580
but I do want to say
I'm thrilled we get to do this

1746
01:51:16,580 --> 01:51:18,140
in the United States of America.

1747
01:51:21,140 --> 01:51:24,260
You wrote a blog post titled
How To Be Successful,

1748
01:51:24,260 --> 01:51:25,940
have almost too much self-belief,

1749
01:51:25,940 --> 01:51:28,540
learn to think independently,
get good at sales,

1750
01:51:28,540 --> 01:51:31,100
make it easy to take risks,
focus, work hard,

1751
01:51:31,100 --> 01:51:34,100
be bold, be wilful,
be hard to compete with,

1752
01:51:34,100 --> 01:51:37,020
you get rich by owning things,
be internally driven.

1753
01:51:39,900 --> 01:51:43,220
I think I'm disconnected
from the reality of life

1754
01:51:43,220 --> 01:51:45,220
for most people,

1755
01:51:45,220 --> 01:51:49,020
and trying to really internalise

1756
01:51:49,020 --> 01:51:54,580
what the impact on people
that AGI is going to have,

1757
01:51:54,580 --> 01:51:58,300
I probably, like, feel that less
than other people would.

1758
01:52:02,420 --> 01:52:05,540
And I wonder why I am so different -

1759
01:52:05,540 --> 01:52:09,180
like, I'm curious where that
difference comes from.

1760
01:52:16,740 --> 01:52:18,460
DEVICE CHIMES
AUTOMATION: Drony.

1761
01:52:18,460 --> 01:52:21,260
OK. So, if you stand in front of it,

1762
01:52:21,260 --> 01:52:23,060
it's got to work out
who it's tracking.

1763
01:52:23,060 --> 01:52:24,500
I stand here in front of it?

1764
01:52:25,620 --> 01:52:26,900
Follow.

1765
01:52:26,900 --> 01:52:28,340
So it's looking at me now?

1766
01:52:28,340 --> 01:52:30,780
Three, two, one.

1767
01:52:30,780 --> 01:52:32,220
DRONE WHIRS

1768
01:52:32,220 --> 01:52:33,220
OK.

1769
01:52:36,260 --> 01:52:37,460
Come on, it's following you.

1770
01:52:37,460 --> 01:52:39,020
OK.

1771
01:52:39,020 --> 01:52:40,740
It's in autonomous mode.

1772
01:52:49,780 --> 01:52:52,020
I feel... Aah!

1773
01:52:53,980 --> 01:52:56,340
Tell me this, Dr Oppenheimer.

1774
01:52:56,340 --> 01:52:58,620
In this era that is more threatened
and dominated

1775
01:52:58,620 --> 01:53:00,540
than any previous one by scientists,

1776
01:53:00,540 --> 01:53:02,220
their decisions, their discoveries,

1777
01:53:02,220 --> 01:53:04,780
do you ever become frightened
at what you're finding out?

1778
01:53:04,780 --> 01:53:08,700
I only get frightened when -
and it happens very rarely -

1779
01:53:08,700 --> 01:53:10,540
I think I have an idea.

1780
01:53:12,020 --> 01:53:15,980
The only way I can tell whether my
thoughts really have some weight

1781
01:53:15,980 --> 01:53:17,820
is the sense of terror.

1782
01:53:19,300 --> 01:53:24,620
The effect of one man can...
can be just enormous.

1783
01:53:25,740 --> 01:53:28,100
Professor Geoffrey Hinton,

1784
01:53:28,100 --> 01:53:32,020
you have been awarded
the Nobel Prize in physics

1785
01:53:32,020 --> 01:53:35,940
for discoveries and inventions
allowing computers to learn

1786
01:53:35,940 --> 01:53:38,380
without explicit programming.

1787
01:53:48,420 --> 01:53:49,580
DRONE WHIRS

1788
01:53:52,580 --> 01:53:54,940
Like everybody else,

1789
01:53:54,940 --> 01:53:58,140
I find it hard to really absorb this
emotionally.

1790
01:54:05,780 --> 01:54:09,820
I've been working on neural nets
for 55 years,

1791
01:54:09,820 --> 01:54:12,820
treated as a sort of joke
by most people.

1792
01:54:12,820 --> 01:54:14,780
An unconscious zombie,

1793
01:54:14,780 --> 01:54:17,140
you say hi to it in the morning,
it says hi.

1794
01:54:17,140 --> 01:54:19,340
The things you're saying
seem ridiculous.

1795
01:54:20,540 --> 01:54:22,740
I'm very glad
it doesn't wish me ill.

1796
01:54:25,500 --> 01:54:28,580
We've learnt how to make them think.

1797
01:54:28,580 --> 01:54:32,980
Fairly soon we'll be in a situation
we've never confronted before,

1798
01:54:32,980 --> 01:54:35,900
when we're not
the apex intelligence.

1799
01:54:35,900 --> 01:54:39,020
We don't know what to do
to make them safe.

1800
01:54:40,300 --> 01:54:42,980
How many examples do you know

1801
01:54:42,980 --> 01:54:44,620
of more intelligent things

1802
01:54:44,620 --> 01:54:47,100
being controlled by less
intelligent things?

1803
01:54:48,260 --> 01:54:50,100
Does that keep you awake at night?

1804
01:54:50,100 --> 01:54:52,500
Yes. Yeah.

1805
01:54:53,980 --> 01:54:55,860
How's your good friend
Elon Musk, then?

1806
01:54:55,860 --> 01:54:57,460
He's not my good friend.

1807
01:54:57,460 --> 01:55:01,140
Ever since I tweeted that he should
be expelled from the Royal Society.

1808
01:55:05,140 --> 01:55:09,540
In my view, AI is more dangerous
than nuclear bombs.

1809
01:55:09,540 --> 01:55:11,900
I was having trouble sleeping
at night because of AI danger.

1810
01:55:14,660 --> 01:55:16,700
And then I finally became fatalistic
about it

1811
01:55:16,700 --> 01:55:20,180
and said, even if I knew
annihilation was certain,

1812
01:55:20,180 --> 01:55:23,740
would I choose to be alive
at that time or not?

1813
01:55:27,180 --> 01:55:30,460
Personally, I would prefer to live
in interesting times -

1814
01:55:30,460 --> 01:55:34,220
and we live in the most interesting
of times.

1815
01:55:39,220 --> 01:55:41,580
What will the future look like?

1816
01:55:41,580 --> 01:55:47,140
What's the most exciting future
that you could possibly imagine?

1817
01:55:47,140 --> 01:55:51,460
How about a future where you can
have any good or service you want,

1818
01:55:51,460 --> 01:55:53,820
at will?

1819
01:55:53,820 --> 01:55:56,460
Surely such a thing
cannot be the case.

1820
01:56:03,700 --> 01:56:06,820
What's key to that
is robotics and AI.

1821
01:56:10,460 --> 01:56:12,980
That's Optimus.

1822
01:56:12,980 --> 01:56:14,940
Our robot has a real brain.

1823
01:56:16,460 --> 01:56:18,300
Will clean your house, mow the lawn,

1824
01:56:18,300 --> 01:56:21,100
will teach your kids, will babysit.

1825
01:56:23,500 --> 01:56:27,300
This year we'll be able
to make about 5,000.

1826
01:56:27,300 --> 01:56:28,940
5,000 robots -

1827
01:56:28,940 --> 01:56:31,140
that's...that's the size
of a Roman legion.

1828
01:56:33,460 --> 01:56:35,260
Ten legions next year.

1829
01:56:36,820 --> 01:56:38,500
Optimus 3,

1830
01:56:38,500 --> 01:56:40,340
Optimus 4,

1831
01:56:40,340 --> 01:56:42,260
so it's kind of like
an annual release cycle

1832
01:56:42,260 --> 01:56:44,460
with significant improvements.

1833
01:56:47,620 --> 01:56:50,820
Starting with a million unit
production line,

1834
01:56:50,820 --> 01:56:52,700
we'll be making
tens of millions a year.

1835
01:57:11,140 --> 01:57:12,540
It's very hard to predict,

1836
01:57:12,540 --> 01:57:15,860
but I think between 10 and 20 years
we'll have super intelligence.

1837
01:57:19,860 --> 01:57:22,900
Intellectually you can see
the threat,

1838
01:57:22,900 --> 01:57:27,780
but I haven't come to terms
with what the development could do

1839
01:57:27,780 --> 01:57:29,620
to my children's future.

1840
01:57:37,300 --> 01:57:39,220
If you could press a button to erase

1841
01:57:39,220 --> 01:57:43,220
the work that you've done,
would you press that?

1842
01:57:43,220 --> 01:57:45,900
I've never thought about
that question.

1843
01:57:49,900 --> 01:57:51,940
If that was a real option

1844
01:57:51,940 --> 01:57:55,820
I would have to think hard
about whether or not to do that.

1845
01:57:55,820 --> 01:57:59,220
Like maybe the wisest decision
in about 1946

1846
01:57:59,220 --> 01:58:02,180
would have been,
don't develop H bombs.

1847
01:58:06,220 --> 01:58:08,580
For AI, it's more complicated.

1848
01:58:08,580 --> 01:58:11,260
You'd be giving up a lot if you
stopped the development.

1849
01:58:14,980 --> 01:58:18,140
It's going to be very, very good
for diagnosing diseases and so on.

1850
01:58:23,740 --> 01:58:29,100
Already Demis Hassabis and DeepMind
showed that you can use AI

1851
01:58:29,100 --> 01:58:31,820
to figure out how proteins fold up.

1852
01:58:33,540 --> 01:58:35,900
Everything in your body depends
on proteins.

1853
01:58:38,740 --> 01:58:41,820
It takes one PhD student
their entire PhD

1854
01:58:41,820 --> 01:58:43,180
to do one protein.

1855
01:58:44,620 --> 01:58:47,100
Over Christmas we did every protein
in the human body,

1856
01:58:47,100 --> 01:58:49,460
all 20,000.

1857
01:58:49,460 --> 01:58:51,820
We're trying to accelerate
drug discoveries -

1858
01:58:51,820 --> 01:58:53,820
instead of taking ten years
to find a drug,

1859
01:58:53,820 --> 01:58:55,620
maybe it could take months,
or weeks.

1860
01:58:55,620 --> 01:58:58,340
Curing diseases,
maybe curing all diseases,

1861
01:58:58,340 --> 01:59:00,940
within the next couple
of decades might be possible.

1862
01:59:00,940 --> 01:59:02,380
You know,
that's the essence of life,

1863
01:59:02,380 --> 01:59:03,700
is trying to strive for things

1864
01:59:03,700 --> 01:59:05,300
and trying to make a difference
for the better.

1865
01:59:05,300 --> 01:59:06,380
I think that might be...

1866
01:59:06,380 --> 01:59:08,260
And do you have any sleepless nights

1867
01:59:08,260 --> 01:59:11,460
that everything that you've worked
on for decades could go wrong?

1868
01:59:13,460 --> 01:59:16,420
That's been the duality I've lived
with my whole life.

1869
01:59:17,820 --> 01:59:19,380
It's a dual-purpose technology,

1870
01:59:19,380 --> 01:59:21,660
and we have to make sure
we guard against

1871
01:59:21,660 --> 01:59:24,020
AI circumventing its guardrails.

1872
01:59:26,380 --> 01:59:29,940
I hope we're gonna be
a big part of birthing AI,

1873
01:59:29,940 --> 01:59:33,980
cos I think even though the AI
systems are gonna learn for itself,

1874
01:59:33,980 --> 01:59:37,260
there'll be a residue in
the system

1875
01:59:37,260 --> 01:59:39,780
of the values of the creators
of that system.

1876
01:59:43,860 --> 01:59:45,700
Geoffrey, do you think we're
all fucked?

1877
01:59:45,700 --> 01:59:47,980
Well, it depends. We might be.

1878
01:59:47,980 --> 01:59:49,460
It's a whole new world.

1879
01:59:51,940 --> 01:59:54,780
I really don't know,
I genuinely don't know.

1880
01:59:54,780 --> 01:59:59,620
When I'm feeling slightly depressed
I think, "People are toast."

1881
01:59:59,620 --> 02:00:03,100
When I'm feeling cheerful
I think, "We'll figure out a way."

1882
02:00:05,820 --> 02:00:09,300
# Now that you know it's nowhere

1883
02:00:13,340 --> 02:00:16,420
# What's to stop you coming home? #

1884
02:00:16,420 --> 02:00:19,940
After losing Sewell, I have spoken
with parents across the country.

1885
02:00:19,940 --> 02:00:23,340
Character AI reached
a settlement in multiple lawsuits.

1886
02:00:23,340 --> 02:00:27,300
The defendants settled cases
in New York, Colorado and Texas.

1887
02:00:27,300 --> 02:00:30,980
# And you really realise
what's going down... #

1888
02:00:30,980 --> 02:00:34,060
We are preparing to win the future.

1889
02:00:34,060 --> 02:00:37,460
The Pentagon is signing a deal with
literally every major AI company.

1890
02:00:37,460 --> 02:00:38,740
Sam Altman has made a deal

1891
02:00:38,740 --> 02:00:40,340
with the Pentagon.

1892
02:00:40,340 --> 02:00:44,580
US military used Claude for its
attack on Iran over the weekend.

1893
02:01:12,740 --> 02:01:16,060
# When your love has moved away

1894
02:01:18,420 --> 02:01:21,260
# You must face yourself
and you must say

1895
02:01:22,340 --> 02:01:25,180
# I remember better days

1896
02:01:28,100 --> 02:01:31,860
# Don't you cry cos she is gone

1897
02:01:33,300 --> 02:01:36,860
# She is only moving on

1898
02:01:37,940 --> 02:01:40,500
# Chasing mirrors through a haze. #

