1
00:09:58,480 --> 00:10:00,640
Mhm.

2
00:12:00,480 --> 00:12:02,597
>> Good evening and welcome to the John F.

3
00:12:02,680 --> 00:12:05,317
Kennedy Jr. Forum at the Institute of

4
00:12:05,400 --> 00:12:08,877
Politics. My name is Amo Gyantah and I'm

5
00:12:08,960 --> 00:12:10,997
a freshman at the college studying

6
00:12:11,080 --> 00:12:13,637
computer science and economics and I'm a

7
00:12:13,720 --> 00:12:17,157
member of the JFK Jr. Forum Committee.

8
00:12:17,240 --> 00:12:19,637
Before we begin, please note the exit

9
00:12:19,720 --> 00:12:24,917
doors which are located both at the park
side and the JFK street side of the

10
00:12:25,000 --> 00:12:28,197
forum. In the event
of an emergency, walk to

11
00:12:28,280 --> 00:12:30,997
the exit closest to you and congregate

12
00:12:31,080 --> 00:12:33,677
in the JFK park. Please also take a

13
00:12:33,760 --> 00:12:37,517
moment now to silence your cell phones.

14
00:12:37,600 --> 00:12:44,440
And join me in welcoming Harvard
College undergraduate Matteo Cagliardo.

15
00:12:47,280 --> 00:12:50,197
Good evening everyone.
It is my pleasure tonight to welcome you

16
00:12:50,280 --> 00:12:53,677
all to the John F. Kennedy Jr. Forum
at the Harvard Institute of Politics.

17
00:12:53,760 --> 00:12:56,677
My name is Matteo Cagliardo. I'm
a junior at the college studying applied

18
00:12:56,760 --> 00:12:59,197
math and computer science with economics.
I'm a member of the John F.

19
00:12:59,280 --> 00:13:01,837
Kennedy Jr. Forum Student Committee.

20
00:13:01,920 --> 00:13:05,517
We gather tonight at what is arguably
one of the most pivotal moments in human

21
00:13:05,600 --> 00:13:09,757
history. Artificial intelligence has
evolved from being a futuristic concept

22
00:13:09,840 --> 00:13:13,797
from to becoming a fundamental
part of our everyday lives.

23
00:13:13,880 --> 00:13:18,277
Quietly but profoundly reshaping
economies, warfare, and the very fabric

24
00:13:18,360 --> 00:13:21,797
of human society.
For the first time since the dawn of the

25
00:13:21,880 --> 00:13:25,597
nuclear era, global stability may
hinge not just on military strength or

26
00:13:25,680 --> 00:13:29,877
economic weight, but also
on competing to secure a

27
00:13:29,960 --> 00:13:33,877
competitive advantage on this new
technological intelligence.

28
00:13:33,960 --> 00:13:37,317
All in all, AI embodies something
we have never seen before.

29
00:13:37,400 --> 00:13:39,717
A shift in how knowledge is generated.

30
00:13:39,800 --> 00:13:43,997
A shift in how decisions are made
and ultimately how humanity sees its own

31
00:13:44,080 --> 00:13:48,997
place in the world. What will
our society look like 20 years from now?

32
00:13:49,080 --> 00:13:52,877
And more importantly, how will
artificial intelligence have shaped it?

33
00:13:52,960 --> 00:13:55,877
Now to comment on these ideas, it
is my pleasure to introduce you all to our

34
00:13:55,960 --> 00:13:59,877
speakers for the evening.
Eric Schmidt served as chairman and CEO

35
00:13:59,960 --> 00:14:02,877
of Google between 2001 and 2011.

36
00:14:02,960 --> 00:14:06,677
Today he serves as chair and CEO
of Relativity Space. He is a founding

37
00:14:06,760 --> 00:14:10,317
partner at Innovation Endeavors
and is a leading voice in AI and national

38
00:14:10,400 --> 00:14:13,077
security in the US and around the globe.

39
00:14:13,160 --> 00:14:16,117
Graham Allison is the Douglas Dillon
Professor of Government at the Harvard

40
00:14:16,200 --> 00:14:19,517
Kennedy School where he served as founding
dean and director of its Belfer

41
00:14:19,600 --> 00:14:22,157
Center for Science
and International Affairs.

42
00:14:22,240 --> 00:14:25,117
He also served as assistant secretary
of defense in the first Clinton

43
00:14:25,200 --> 00:14:29,557
administration receiving the Defense
Medal for distinguished public service.

44
00:14:29,640 --> 00:14:32,557
Now without further ado, please join me
in opening up tonight's discussion and

45
00:14:32,640 --> 00:14:36,720
welcoming our esteemed
guests to the stage.

46
00:14:47,400 --> 00:14:50,797
So thank you for a generous
introduction. It's a

47
00:14:50,880 --> 00:14:53,597
extraordinary pleasure to welcome back

48
00:14:53,680 --> 00:14:56,757
to the JFK Jr. Forum Eric Schmidt, our

49
00:14:56,840 --> 00:14:58,637
colleague and friend, uh for a

50
00:14:58,720 --> 00:15:04,517
discussion of Henry Kissinger
uh and AI and the future.

51
00:15:04,600 --> 00:15:10,717
Uh as you can see here uh
this picture of Eric and Henry, uh

52
00:15:10,800 --> 00:15:13,837
uh they were best friends. Uh

53
00:15:13,920 --> 00:15:16,517
late developing friendship but became

54
00:15:16,600 --> 00:15:20,277
very deep friendship in which Eric was

55
00:15:20,360 --> 00:15:22,317
uh generously the tutor

56
00:15:22,400 --> 00:15:25,837
for a 95-year-old Henry Kissinger who

57
00:15:25,920 --> 00:15:28,877
discovered AI hanging around the end of

58
00:15:28,960 --> 00:15:32,997
a lecture that he had given uh when

59
00:15:33,080 --> 00:15:36,797
uh Demis Hassabis was uh offering to

60
00:15:36,880 --> 00:15:42,637
beginning to talk about AI and Henry
decided he needed to learn about this.

61
00:15:42,720 --> 00:15:46,277
Uh he called me up. I told him,
"Henry, forget about it.

62
00:15:46,360 --> 00:15:49,797
You know that you don't have
any background in science

63
00:15:49,880 --> 00:15:52,197
and technology. I told him actually you

64
00:15:52,280 --> 00:15:54,077
don't know the difference between a chip

65
00:15:54,160 --> 00:15:56,877
and a potato chip." Mhm. Uh he said,

66
00:15:56,960 --> 00:15:59,677
"That's true but Eric has promised to to

67
00:15:59,760 --> 00:16:02,197
to teach me." So we're very pleased to

68
00:16:02,280 --> 00:16:07,157
have him here. We had him here
last year. This is actually

69
00:16:07,240 --> 00:16:08,797
uh maybe it will become an annual

70
00:16:08,880 --> 00:16:10,637
tradition. Uh

71
00:16:10,720 --> 00:16:15,157
Henry passed 2 years ago last week. Uh

72
00:16:15,240 --> 00:16:18,077
so he was 100 years old. Uh

73
00:16:18,160 --> 00:16:20,437
to think of such an amazing life over

74
00:16:20,520 --> 00:16:23,637
that century. Uh he was a person who

75
00:16:23,720 --> 00:16:26,717
made a huge difference in America's

76
00:16:26,800 --> 00:16:29,397
national security and the world and made

77
00:16:29,480 --> 00:16:33,997
a huge difference in the lives of many
many people. Uh some of whom began as

78
00:16:34,080 --> 00:16:39,597
his students, some who became
as his tutors, uh and many others. So

79
00:16:39,680 --> 00:16:43,557
uh Eric has been introduced. Uh

80
00:16:43,640 --> 00:16:47,957
I would just remind you a couple
of things. I would say first that

81
00:16:48,040 --> 00:16:51,997
Eric was the chief executive
who took Google from a

82
00:16:52,080 --> 00:16:54,557
cause to one of the great companies of

83
00:16:54,640 --> 00:16:56,917
the world.

84
00:16:57,000 --> 00:16:59,117
And that's a pretty amazing thing.

85
00:16:59,200 --> 00:17:02,957
Uh secondly, early on he had

86
00:17:03,040 --> 00:17:05,397
identified AI

87
00:17:05,480 --> 00:17:08,477
as an arena of the future.

88
00:17:08,560 --> 00:17:11,157
And at Google essentially bought up all

89
00:17:11,240 --> 00:17:15,517
the super talent in the world
that he could find including uh

90
00:17:15,600 --> 00:17:17,477
uh DeepMind

91
00:17:17,560 --> 00:17:20,117
uh which was actually the company that

92
00:17:20,200 --> 00:17:24,957
then brought to Google uh Demis Hassabis

93
00:17:25,040 --> 00:17:27,637
who got the Nobel Prize last year for

94
00:17:27,720 --> 00:17:31,397
his work at Google on uh protein.

95
00:17:31,480 --> 00:17:35,117
Mustafa Suleyman who's now running uh

96
00:17:35,200 --> 00:17:39,237
uh Microsoft's consumer
AI was another part

97
00:17:39,320 --> 00:17:43,277
of the acquisition and many
many many many others.

98
00:17:43,360 --> 00:17:47,717
So the other thing to say about Eric
about Eric and why it's such a pleasure

99
00:17:47,800 --> 00:17:52,197
to have him here is that as one tries
to make any sense of all the claims that

100
00:17:52,280 --> 00:17:54,237
are being made about AI,

101
00:17:54,320 --> 00:17:57,157
most of the people who talk loudest

102
00:17:57,240 --> 00:17:59,517
are talking their own book.

103
00:17:59,600 --> 00:18:02,437
So when I listen to Sam Altman or to

104
00:18:02,520 --> 00:18:04,997
folks at Anthropic or with folks even

105
00:18:05,080 --> 00:18:07,837
Mustafa now at Microsoft, they have to

106
00:18:07,920 --> 00:18:12,597
talk their companies >> holding
my book. Uh well, wait a second.

107
00:18:12,680 --> 00:18:14,717
They have to talk their companies, you

108
00:18:14,800 --> 00:18:17,597
know, business, their future and they

109
00:18:17,680 --> 00:18:22,477
are chasing what they believe is the
biggest pot of gold at the end of the

110
00:18:22,560 --> 00:18:24,397
rainbow that there's ever been.

111
00:18:24,480 --> 00:18:28,717
So it's pretty hard to tell what they
actually think and what they're saying.

112
00:18:28,800 --> 00:18:32,117
Eric has kind of graduated
to be the statesman

113
00:18:32,200 --> 00:18:35,157
in this domain having had a big stake in

114
00:18:35,240 --> 00:18:39,157
it but now having stand
standing back and I

115
00:18:39,240 --> 00:18:43,837
particularly after his work
with Henry, basically trying to

116
00:18:43,920 --> 00:18:47,597
say things as as clearly
as he can and as clearly as

117
00:18:47,680 --> 00:18:49,917
he can see them. So I take him to be a

118
00:18:50,000 --> 00:18:53,597
much more valuable source of clues about

119
00:18:53,680 --> 00:18:57,717
what's happening than listening to most
of the people who are talking their own

120
00:18:57,800 --> 00:19:00,397
books. So let's start
with Henry and then we're

121
00:19:00,480 --> 00:19:05,837
going to go to AI and then we'll
go to questions of the audience.

122
00:19:05,920 --> 00:19:08,797
So uh Eric uh

123
00:19:08,880 --> 00:19:12,237
you uh spoke beautifully at at Henry's

124
00:19:12,320 --> 00:19:15,037
memorial about how

125
00:19:15,120 --> 00:19:20,117
much impact Henry had
on your own life and even on the

126
00:19:20,200 --> 00:19:21,797
you know, the questions that you were

127
00:19:21,880 --> 00:19:25,077
asking, the things that you

128
00:19:25,160 --> 00:19:27,717
think that matter. So tell us a little

129
00:19:27,800 --> 00:19:30,517
bit about more what Henry meant

130
00:19:30,600 --> 00:19:33,797
in this relationship and also for people

131
00:19:33,880 --> 00:19:36,277
that wouldn't have the opportunity to to

132
00:19:36,360 --> 00:19:38,997
know him, what what how might they

133
00:19:39,080 --> 00:19:41,477
capture a little bit of this magic?

134
00:19:41,560 --> 00:19:45,237
You know, when I when I first met
and thank you Graham as always, it's great

135
00:19:45,320 --> 00:19:47,477
to be here. Um

136
00:19:47,560 --> 00:19:50,357
I I met Henry when he was in his early

137
00:19:50,440 --> 00:19:55,237
80s and normally men
in their early 80s become a little

138
00:19:55,320 --> 00:19:58,997
off. And I was struck by how brilliant he

139
00:19:59,080 --> 00:20:03,357
was. And so I thought to myself,
you know, he he served in World War II, he

140
00:20:03,440 --> 00:20:07,077
got the Bronze Star, and you know,
emigrated or escaped from Germany cuz

141
00:20:07,160 --> 00:20:12,037
they were a Jewish family.
And um came to fought in the war, came

142
00:20:12,120 --> 00:20:15,117
on the GI Bill to Harvard uh
undergraduate and graduate, and

143
00:20:15,200 --> 00:20:17,237
ultimately was a professor here.

144
00:20:17,320 --> 00:20:21,477
And I tried to figure out he must have
been really smart when he was in your

145
00:20:21,560 --> 00:20:24,837
chairs, not that the building
existed at the time.

146
00:20:24,920 --> 00:20:29,117
So here's a quote when he
was an undergraduate at Harvard.

147
00:20:29,200 --> 00:20:32,837
Quote, "In the life
of every person there comes

148
00:20:32,920 --> 00:20:35,397
a point when he realizes that out of all

149
00:20:35,480 --> 00:20:40,837
of the seemingly limitless possibilities
of his youth, he has in fact become one

150
00:20:40,920 --> 00:20:44,837
actuality. No longer
is life a broad plane with

151
00:20:44,920 --> 00:20:49,397
forests and mountains and beckoning
all around, but it becomes apparent that

152
00:20:49,480 --> 00:20:53,157
one's journey across
the meadows has indeed followed a

153
00:20:53,240 --> 00:20:57,157
regular path. That one can
no longer go this way or

154
00:20:57,240 --> 00:21:01,877
that, but that the direction
is set and the limits are defined."

155
00:21:01,960 --> 00:21:05,277
That's what he wrote when he was your age.

156
00:21:05,360 --> 00:21:08,157
Um when he was an undergraduate here, he

157
00:21:08,240 --> 00:21:12,437
holds the record for the longest
undergraduate thesis in the uh in the

158
00:21:12,520 --> 00:21:15,717
college. After he
submitted his undergraduate

159
00:21:15,800 --> 00:21:19,597
thesis on Kant and the meaning of the
world, they instituted a new rule which

160
00:21:19,680 --> 00:21:24,757
still applies to you that your thesis
can be no more than 350 pages.

161
00:21:24,840 --> 00:21:29,637
FYI. It's true. >> So he
was clearly an unusually brilliant

162
00:21:29,720 --> 00:21:34,237
polymath. And uh my own
explanation for Henry,

163
00:21:34,320 --> 00:21:38,397
aside from you know,
I really cared deeply about him,

164
00:21:38,480 --> 00:21:41,077
was that watching his family, and in

165
00:21:41,160 --> 00:21:45,757
particular his father, and he we talked
a lot about this, um see the destruction

166
00:21:45,840 --> 00:21:50,157
of the world around him
in Germany as the Nazis took over.

167
00:21:50,240 --> 00:21:53,797
And when they fled, and he saw
the damage to his father and his father's

168
00:21:53,880 --> 00:21:56,037
identity and so forth, Henry after the

169
00:21:56,120 --> 00:21:57,717
war decided that he would do everything

170
00:21:57,800 --> 00:22:01,317
he can could to avoid a future war.

171
00:22:01,400 --> 00:22:04,197
Now you can quibble with or argue very

172
00:22:04,280 --> 00:22:06,637
much with the things he did, but you

173
00:22:06,720 --> 00:22:09,237
cannot argue that his that his goal was

174
00:22:09,320 --> 00:22:13,957
not what I just said. The the data
supports it, the historians all agree.

175
00:22:14,040 --> 00:22:18,077
You can disagree with the specific
tactics, but at the end of the day what

176
00:22:18,160 --> 00:22:22,597
he sought is he sought a a a world
that did not have a World War

177
00:22:22,680 --> 00:22:27,077
III. And think about the issues
that he faced in the '50s and '60s.

178
00:22:27,160 --> 00:22:30,757
He used to tell me about this small
group, and he always thought that the

179
00:22:30,840 --> 00:22:34,557
most interesting policy things happen
with small groups, and this was a

180
00:22:34,640 --> 00:22:37,037
collaboration between MIT and Harvard

181
00:22:37,120 --> 00:22:40,317
and RAND, uh which in the early 1950s

182
00:22:40,400 --> 00:22:44,797
dev- de- um um invented
mutual assured destruction.

183
00:22:44,880 --> 00:22:47,677
And I said, "Henry, well who was in this
group?" He says, "Oh, Oppenheimer." And

184
00:22:47,760 --> 00:22:50,077
I'm like, you know, give me a break.

185
00:22:50,160 --> 00:22:52,637
Right? All the most famous people.

186
00:22:52,720 --> 00:22:57,157
So uh he was lucky uh to be born at the

187
00:22:57,240 --> 00:23:00,917
right time. He was clearly one of the
smartest people at alive at the time. He

188
00:23:01,000 --> 00:23:04,637
was lucky to get out of Germany before
his family and he were killed. He was

189
00:23:04,720 --> 00:23:07,237
lucky to be part of the GI Bill.

190
00:23:07,320 --> 00:23:12,237
Um and he had all sorts of interesting
stories. My favorite was one day I told

191
00:23:12,320 --> 00:23:15,477
him, "You have to go to the doctor." You
know, he's old. And he goes, "Okay." And

192
00:23:15,560 --> 00:23:19,197
I said, "I want you to go to the Mayo
Clinic." And I said, "Okay." Now he's a

193
00:23:19,280 --> 00:23:21,837
New Yorker. And the people
in Mayo were like, you

194
00:23:21,920 --> 00:23:26,397
know, Midwestern, soft-spoken, pleasant,
nice, not like the intense New Yorker

195
00:23:26,480 --> 00:23:28,717
types. He goes to the Mayo
Clinic, comes back,

196
00:23:28,800 --> 00:23:32,437
and he said, "I loved it." I said,
"Why would you like the Mayo Clinic in the

197
00:23:32,520 --> 00:23:35,117
middle of Rochester, Minnesota?"
He said, "It reminded me of the war." I

198
00:23:35,200 --> 00:23:38,957
said, "What?" And he goes,
"Well, he was in a uh

199
00:23:39,040 --> 00:23:40,357
he didn't speak English when he came to

200
00:23:40,440 --> 00:23:44,037
the country, and he worked as a um

201
00:23:44,120 --> 00:23:48,357
uh in a um shaving brush factory where
they made shaving brushes, and he was

202
00:23:48,440 --> 00:23:51,677
drafted as everyone expected,
and he went to war, and he went with a

203
00:23:51,760 --> 00:23:54,637
Wisconsin-based American group.

204
00:23:54,720 --> 00:23:59,637
So he indexed his identity, could hardly
speak English, and and nobody could

205
00:23:59,720 --> 00:24:03,357
understand his English anyway,
he indexed to that group. So he had all

206
00:24:03,440 --> 00:24:07,517
these sort of touching little stories.
I have lots of stories like that about how

207
00:24:07,600 --> 00:24:11,277
he became an American,
and of course became an American citizen.

208
00:24:11,360 --> 00:24:13,517
And you know the rest of the history.

209
00:24:13,600 --> 00:24:16,157
Well, I would say a fantastic person. I

210
00:24:16,240 --> 00:24:21,677
had the good fortune to enroll
in the course as a student.

211
00:24:21,760 --> 00:24:24,717
>> What was that like? >> 1965,

212
00:24:24,800 --> 00:24:26,437
God help us, uh

213
00:24:26,520 --> 00:24:29,277
taught by Henry, uh and

214
00:24:29,360 --> 00:24:33,597
ever thereafter was part of his entourage,

215
00:24:33,680 --> 00:24:35,677
and I think the thing that struck me

216
00:24:35,760 --> 00:24:39,077
most about him was this uh

217
00:24:39,160 --> 00:24:41,837
and the strategic acumen. That is, the

218
00:24:41,920 --> 00:24:44,597
ability to take a problem and lift it up

219
00:24:44,680 --> 00:24:47,317
to its whole

220
00:24:47,400 --> 00:24:51,357
360° of strategic challenge, and then

221
00:24:51,440 --> 00:24:53,877
find the

222
00:24:53,960 --> 00:24:56,397
points at which the

223
00:24:56,480 --> 00:25:01,197
policy process could be impacted. And uh

224
00:25:01,280 --> 00:25:06,237
I wrote about uh for his
100th birthday, which Richard

225
00:25:06,320 --> 00:25:09,117
asked us to write a little,
you know, something that

226
00:25:09,200 --> 00:25:13,797
Eric wrote something, I wrote something,
I said, "Really in thinking about Henry,

227
00:25:13,880 --> 00:25:14,997
uh

228
00:25:15,080 --> 00:25:17,717
he was not about, if you think about, he

229
00:25:17,800 --> 00:25:21,437
was often criticized for his realism or

230
00:25:21,520 --> 00:25:26,597
ruthless realism or realpolitik,
but if you look and look and see what he

231
00:25:26,680 --> 00:25:27,877
was doing,

232
00:25:27,960 --> 00:25:31,637
uh it wasn't raw real- realpolitik

233
00:25:31,720 --> 00:25:34,637
uh just for the advancement of the

234
00:25:34,720 --> 00:25:36,477
interests of one country.

235
00:25:36,560 --> 00:25:39,357
It was always about the construction of

236
00:25:39,440 --> 00:25:44,157
a viable order to prevent
catastrophic war.

237
00:25:44,240 --> 00:25:48,397
He had lived through the experience of
the Holocaust and the most deadly war in

238
00:25:48,480 --> 00:25:51,037
history. He

239
00:25:51,120 --> 00:25:54,837
foresaw the prospect of a nuclear war

240
00:25:54,920 --> 00:25:57,477
that could actually extinguish life on

241
00:25:57,560 --> 00:25:58,877
Earth.

242
00:25:58,960 --> 00:26:02,197
He was caught up in the Cold War, the

243
00:26:02,280 --> 00:26:07,357
fiercest rivalry that we'd seen to
that point between the US and the Soviet

244
00:26:07,440 --> 00:26:11,037
Union, and all the time
he was trying to reach

245
00:26:11,120 --> 00:26:14,197
beyond that to some

246
00:26:14,280 --> 00:26:15,877
way to survive

247
00:26:15,960 --> 00:26:19,517
for both the US and for for uh

248
00:26:19,600 --> 00:26:23,397
for fellow human beings. So that's I

249
00:26:23,480 --> 00:26:25,837
think and that's how he kind of backed

250
00:26:25,920 --> 00:26:27,357
into

251
00:26:27,440 --> 00:26:30,917
the AI issue, which for him was

252
00:26:31,000 --> 00:26:36,197
another generation of a somewhat
analogous problem. Well,

253
00:26:36,280 --> 00:26:39,717
he had written his undergraduate thesis
on Kant and the meaning of meaning, I

254
00:26:39,800 --> 00:26:44,677
think. And um he was very
when he went to hear

255
00:26:44,760 --> 00:26:49,317
Demos speak, um it sort
of he immediately got it. And

256
00:26:49,400 --> 00:26:51,597
he immediately said to himself,

257
00:26:51,680 --> 00:26:54,517
"What does this mean to be human?"

258
00:26:54,600 --> 00:26:59,077
And we are today grappling with the
question that he foresaw 20 years ago

259
00:26:59,160 --> 00:27:02,637
when we when we first started working
on this, "What does it mean to be human in

260
00:27:02,720 --> 00:27:08,077
the age of AI?" What does it mean to be
a child to an adult, to be a leader?

261
00:27:08,160 --> 00:27:09,397
What does it mean for economics? What

262
00:27:09,480 --> 00:27:10,757
does it mean for jobs? You know, all of

263
00:27:10,840 --> 00:27:15,677
that. But his core argument was that

264
00:27:15,760 --> 00:27:20,117
this is an epochal change
in the sense that it's like the um

265
00:27:20,200 --> 00:27:24,837
uh the various major
changes that we've had

266
00:27:24,920 --> 00:27:26,957
in sort of reasoning, the scientific

267
00:27:27,040 --> 00:27:31,917
revolution and so forth, cuz we
as humans have never had a competitor that

268
00:27:32,000 --> 00:27:35,157
is not human but of similar or greater

269
00:27:35,240 --> 00:27:37,197
level of intelligence. And it is

270
00:27:37,280 --> 00:27:39,357
unpredictable what we human will do. He

271
00:27:39,440 --> 00:27:42,597
used to say that what would happen is in

272
00:27:42,680 --> 00:27:44,597
magic, when people don't understand

273
00:27:44,680 --> 00:27:49,797
things, they either decide that it's
a new religion or they take up arms.

274
00:27:49,880 --> 00:27:52,397
And so he would say, "Well, are we going
to take up arms to AI or we going to

275
00:27:52,480 --> 00:27:56,797
make it a new religion?" And I said,
"I hope it's a religion." Um cuz of course

276
00:27:56,880 --> 00:27:59,157
I benefit from the religion, I guess.

277
00:27:59,240 --> 00:28:02,357
So in the book that
was published this time

278
00:28:02,440 --> 00:28:05,597
last year called Genesis, which was the

279
00:28:05,680 --> 00:28:09,117
work that Eric and Henry and Craig Bundy

280
00:28:09,200 --> 00:28:12,677
were still doing up
to the point at which he uh

281
00:28:12,760 --> 00:28:15,797
passed, he here's a a passage that

282
00:28:15,880 --> 00:28:18,917
clearly Henry wrote. It says, quote,

283
00:28:19,000 --> 00:28:21,837
it talking about the US and China, "If

284
00:28:21,920 --> 00:28:24,917
each wishes to maximize its unilateral

285
00:28:25,000 --> 00:28:28,557
position in the AI space,

286
00:28:28,640 --> 00:28:31,117
then the conditions would be such for a

287
00:28:31,200 --> 00:28:34,357
contest between rival military forces

288
00:28:34,440 --> 00:28:36,757
and intelligence agencies, the likes of

289
00:28:36,840 --> 00:28:39,477
which humanity has never faced before.

290
00:28:39,560 --> 00:28:42,117
Today, in the weeks and months and days

291
00:28:42,200 --> 00:28:44,837
leading up to the arrival of the first

292
00:28:44,920 --> 00:28:49,917
superintelligence, a security
dilemma of existential nature

293
00:28:50,000 --> 00:28:53,677
awaits us." So you've thought about this

294
00:28:53,760 --> 00:28:56,397
superintelligence as an existential

295
00:28:56,480 --> 00:28:59,197
threat beyond what

296
00:28:59,280 --> 00:29:03,477
what we can imagine? Say more, yeah.

297
00:29:03,560 --> 00:29:07,317
Well, um there is something
which I call the San

298
00:29:07,400 --> 00:29:10,877
Francisco consensus. And the reason I
call it the San Francisco consensus is

299
00:29:10,960 --> 00:29:13,437
everyone in San Francisco believes this.

300
00:29:13,520 --> 00:29:17,037
It may or may not be true, but if you go
to San Francisco, trust me, that's what

301
00:29:17,120 --> 00:29:21,197
they're doing besides the usual San
Francisco recreational behaviors.

302
00:29:21,280 --> 00:29:23,997
And um

303
00:29:24,080 --> 00:29:28,397
the the basic build goes that we
are we're we're we've gone through a

304
00:29:28,480 --> 00:29:31,597
language revolution. You understand
that, you know, language to language

305
00:29:31,680 --> 00:29:34,277
chat GPT, everybody
here knows what that is.

306
00:29:34,360 --> 00:29:38,597
Agents, which are now on the horizon,
and these agents allow you to

307
00:29:38,680 --> 00:29:42,637
essentially automate tasks. And the key
thing to understand about agents is they

308
00:29:42,720 --> 00:29:47,157
can be concatenated. So you can, you
you do do this and this and this and this

309
00:29:47,240 --> 00:29:51,317
and this and then they all get piped
together. Why am I talking about

310
00:29:51,400 --> 00:29:54,517
agents? Because I've
just described workflow.

311
00:29:54,600 --> 00:29:57,637
And that's what businesses do, what it's
universities do, it's what governments

312
00:29:57,720 --> 00:30:01,957
do, so forth and so on.
And then the next one is reasoning.

313
00:30:02,040 --> 00:30:04,517
Now, reasoning is higher-order function

314
00:30:04,600 --> 00:30:06,557
in in humans and the

315
00:30:06,640 --> 00:30:09,277
reasoning revolution is just beginning.

316
00:30:09,360 --> 00:30:11,957
Um as of today, it is technically

317
00:30:12,040 --> 00:30:14,797
correct to say that the scaling laws

318
00:30:14,880 --> 00:30:19,517
that operate in AI have not slowed down
yet. The scaling law basically says that

319
00:30:19,600 --> 00:30:23,557
if you put more data and more
electricity and more chips, you get this

320
00:30:23,640 --> 00:30:26,157
emergent behavior one after the other.

321
00:30:26,240 --> 00:30:30,717
And you can see that, for example, with
Gemini 3 which just came out, which beat

322
00:30:30,800 --> 00:30:33,117
uh OpenAI 5 which just came out, which

323
00:30:33,200 --> 00:30:37,157
beat um Claude 4.5 which just came out,

324
00:30:37,240 --> 00:30:41,437
which beat Deep Seek cuz they copied it
anyway, and blah blah blah. You know, so

325
00:30:41,520 --> 00:30:46,677
so it's >> In case you didn't notice,
Gemini 3 is from Google. Yes.

326
00:30:46,760 --> 00:30:49,597
I'm proud to say that we're back in charge

327
00:30:49,680 --> 00:30:53,757
um until the next one. It's it's very
very competitive and what's happening is

328
00:30:53,840 --> 00:30:58,157
that these massive data center build-outs
uh which are, by the way, one of the key

329
00:30:58,240 --> 00:31:02,237
drivers of the American economy are
helping not only lift our economy, but

330
00:31:02,320 --> 00:31:05,117
they're also building building this kind
of stuff. And it's a whole new world

331
00:31:05,200 --> 00:31:07,677
compared to my whole experience.

332
00:31:07,760 --> 00:31:10,077
So, the question is what
happens over time?

333
00:31:10,160 --> 00:31:13,157
So, you have language,
agents, and reasoning.

334
00:31:13,240 --> 00:31:18,317
Well, isn't that what we do? Right?
We do stuff, we communicate, and

335
00:31:18,400 --> 00:31:19,917
we do actions.

336
00:31:20,000 --> 00:31:22,557
So, the San Francisco consensus is that

337
00:31:22,640 --> 00:31:26,117
at some point that stuff comes together

338
00:31:26,200 --> 00:31:29,837
and you get what is called technically
recursive self-improvement. And

339
00:31:29,920 --> 00:31:33,677
recursive self-improvement
is when it's learning on its own.

340
00:31:33,760 --> 00:31:37,637
This is not true today. Today,
when you set up one of these huge

341
00:31:37,720 --> 00:31:40,717
data centers, you know what they look
like, um you have to tell them what to

342
00:31:40,800 --> 00:31:43,357
learn. But the belief
is that this is coming.

343
00:31:43,440 --> 00:31:47,077
And there's lots of evidence that this
is coming. The ability for computers to

344
00:31:47,160 --> 00:31:51,757
write programs, to generate mathematical
conjectures, to discover new facts looks

345
00:31:51,840 --> 00:31:55,277
like it's very very close.
Many people believe that there will be

346
00:31:55,360 --> 00:31:58,317
new math design, new mathematicians, AI

347
00:31:58,400 --> 00:32:01,157
mathematicians to love in the next year.

348
00:32:01,240 --> 00:32:04,157
So, we collectively as an industry
believe that this is going to happen

349
00:32:04,240 --> 00:32:06,437
soon. If you ask the San Francisco

350
00:32:06,520 --> 00:32:09,197
people, they'll say 2 years.

351
00:32:09,280 --> 00:32:14,717
Which is really soon. If you ask me,
I double that to 4 years, which is really

352
00:32:14,800 --> 00:32:19,237
soon. Right? So, it's happening. It's
happening very quickly. I want and Henry

353
00:32:19,320 --> 00:32:22,757
certainly wanted it. We wanted it to be
built with American values and human

354
00:32:22,840 --> 00:32:27,797
values. There is a point, in my view,
and we talked a lot about this,

355
00:32:27,880 --> 00:32:32,317
where somebody's going to have to raise
their hand and say, we just went too far.

356
00:32:32,400 --> 00:32:35,557
There's too much danger here.
We don't want to give that agency to the

357
00:32:35,640 --> 00:32:39,557
computer. We want humans to be in charge
of it. It is not agreed to where that

358
00:32:39,640 --> 00:32:43,237
point is, but our book spends a lot
of time talking about where that might be.

359
00:32:43,320 --> 00:32:46,637
Another example would be the you
discover that the computer has decided

360
00:32:46,720 --> 00:32:49,597
on its own to get access to weapons.

361
00:32:49,680 --> 00:32:53,077
That that's clearly a Like everyone
would agree that's not a good idea,

362
00:32:53,160 --> 00:32:55,957
right? It's bad enough that humans
have access to weapons. Imagine if the

363
00:32:56,040 --> 00:32:59,317
computer has it. And what are
the criteria by which And you can think of

364
00:32:59,400 --> 00:33:04,037
many other such examples. Fundamen-
fundamentally about human agency. We

365
00:33:04,120 --> 00:33:07,717
also spend a lot of time talking
about the effect on children.

366
00:33:07,800 --> 00:33:10,717
Um we're running a mass experiment on

367
00:33:10,800 --> 00:33:13,277
human development by deploying these

368
00:33:13,360 --> 00:33:15,917
systems that are incredibly addictive.

369
00:33:16,000 --> 00:33:20,517
Um what whether it's on iPad or phone or,
you know, whatever for young people

370
00:33:20,600 --> 00:33:24,357
who may not be certainly not adults,
may not have their own identity, and they

371
00:33:24,440 --> 00:33:27,597
can really be manipulated. So, what does
it mean to be a child whose best friend

372
00:33:27,680 --> 00:33:32,557
is non-human? Aside from maybe
becoming like a super nerd.

373
00:33:32,640 --> 00:33:35,077
But I don't know, what does it mean?
We we we have no data. We don't know what

374
00:33:35,160 --> 00:33:38,637
it means to young young boys and girls,
to their development, to their ability

375
00:33:38,720 --> 00:33:43,117
to associate. Will they ultimately
rebel and say, I only love people, I hate

376
00:33:43,200 --> 00:33:45,197
computers? Like, you know, kids rebel.

377
00:33:45,280 --> 00:33:47,717
We just don't know.

378
00:33:47,800 --> 00:33:54,640
So, can you take the next slide? I'm
sorry, I have the clicker, but I left it.

379
00:33:57,880 --> 00:34:02,680
Can I Can you guys make
the next slide go up?

380
00:34:03,560 --> 00:34:06,917
Yeah, thank you. So,
here's the so-called Kissinger

381
00:34:07,000 --> 00:34:08,197
challenge.

382
00:34:08,280 --> 00:34:11,357
This is Henry 1969.

383
00:34:11,440 --> 00:34:14,917
So, Nixon was elected in 1968.

384
00:34:15,000 --> 00:34:19,037
Became president January 1969.

385
00:34:19,120 --> 00:34:21,197
Appointed Henry, who was a professor

386
00:34:21,280 --> 00:34:23,397
here at Harvard, as his national

387
00:34:23,480 --> 00:34:28,957
security advisor. And Henry writes,
as you'll see here, that anyone coming to

388
00:34:29,040 --> 00:34:30,117
office

389
00:34:30,200 --> 00:34:32,917
in the late '60s could not fail to be

390
00:34:33,000 --> 00:34:36,517
awed by the unprecedented dimensions of

391
00:34:36,600 --> 00:34:38,637
the challenge of peace.

392
00:34:38,720 --> 00:34:43,757
And then down to the bottom line,
there could be no higher duty

393
00:34:43,840 --> 00:34:45,517
than to prevent the catastro-

394
00:34:45,600 --> 00:34:49,717
catastrophe of nuclear war.

395
00:34:49,800 --> 00:34:54,797
So, that became for him
a defining challenge

396
00:34:54,880 --> 00:34:56,477
in the Cold War

397
00:34:56,560 --> 00:34:58,237
with nuclear weapons, arsenals

398
00:34:58,320 --> 00:35:03,597
developing, but quote,
no higher duty than to prevent.

399
00:35:03,680 --> 00:35:07,997
So, how might this apply
currently as we try

400
00:35:08,080 --> 00:35:10,397
to think about the US, China, and AI?

401
00:35:10,480 --> 00:35:13,037
Well, let me say that I
think there's no higher duty

402
00:35:13,120 --> 00:35:17,157
than to preserve human
agency and human freedom.

403
00:35:17,240 --> 00:35:20,997
Right? The things that we
value collectively the most.

404
00:35:21,080 --> 00:35:24,037
And I think that's going to be a central
challenge for all of you when you

405
00:35:24,120 --> 00:35:28,557
graduate. Um all of you will be facing
these issues and they're complicated and

406
00:35:28,640 --> 00:35:32,237
they're subtle.
Um imagine if the internet had been

407
00:35:32,320 --> 00:35:35,997
invented by China and it
didn't have the kind of openness

408
00:35:36,080 --> 00:35:40,397
that uh the internet has today.
It was under the Chinese internet.

409
00:35:40,480 --> 00:35:44,917
Um it looks like China is pursuing
a different strategy than what I'm talking

410
00:35:45,000 --> 00:35:50,357
about. In my most recent visit to China,
my the way I operated, I ask engineers

411
00:35:50,440 --> 00:35:52,877
technical questions cuz
they don't lie to you.

412
00:35:52,960 --> 00:35:55,197
Whereas kind of everyone
else, I'm not so sure.

413
00:35:55,280 --> 00:35:58,837
And I eventually figured out what
the Chinese are doing is they're really

414
00:35:58,920 --> 00:36:02,357
really focused on applying AI
in their businesses. And they're going to

415
00:36:02,440 --> 00:36:04,677
outcompete with us.

416
00:36:04,760 --> 00:36:07,357
We We will lose to China for their

417
00:36:07,440 --> 00:36:09,717
incredible adoption of AI in every

418
00:36:09,800 --> 00:36:14,957
product. Because they're just relentless
and they work so hard. Called 996, 9:00

419
00:36:15,040 --> 00:36:18,357
in the morning, 9:00 in the evening,
6 days a week. Illegal, by the way, in

420
00:36:18,440 --> 00:36:23,077
China. Illegal in America, right?
But nevertheless practiced. Um they're

421
00:36:23,160 --> 00:36:27,077
coming. Um they do not
appear to be as focused

422
00:36:27,160 --> 00:36:29,317
on superintelligence and the path that I

423
00:36:29,400 --> 00:36:31,197
was describing as the San Francisco

424
00:36:31,280 --> 00:36:34,237
consensus. Of course, that could change.

425
00:36:34,320 --> 00:36:38,197
So, it appears that the two are pursuing
different paths. One of the questions

426
00:36:38,280 --> 00:36:41,997
for you all and uh at the graduate
level, even undergraduates given it's

427
00:36:42,080 --> 00:36:46,917
Harvard, is start figuring out what
happens when these divergent paths hit

428
00:36:47,000 --> 00:36:51,637
roadblocks. Right? Cuz they're both
have it. Example would be in America,

429
00:36:51,720 --> 00:36:55,037
we have essentially produced no new
electricity because it's very hard to

430
00:36:55,120 --> 00:36:59,317
provision electricity. China has infinite

431
00:36:59,400 --> 00:37:01,957
electricity because of their incredible

432
00:37:02,040 --> 00:37:04,357
investment in renewables and so forth.

433
00:37:04,440 --> 00:37:08,637
They built something like 120 gigawatts
of new renewable energy in the last 5

434
00:37:08,720 --> 00:37:10,197
years. It's some number like that.

435
00:37:10,280 --> 00:37:12,157
>> no. Wait. Wait. Wait. Wait.

436
00:37:12,240 --> 00:37:15,317
Every day in China,

437
00:37:15,400 --> 00:37:18,557
Thursday, Saturday, Monday, every day uh

438
00:37:18,640 --> 00:37:22,917
another gigawatt of electricity has
added to the grid. Wow. That's pretty

439
00:37:23,000 --> 00:37:28,037
extreme. >> Every day. So,
my point is And here And here, by

440
00:37:28,120 --> 00:37:31,717
the way, just to give you an example,
how big is a nuclear power plant? About

441
00:37:31,800 --> 00:37:36,517
1.5 gigawatts. So, that's
the scale of the electricity

442
00:37:36,600 --> 00:37:40,837
revolution that's going on in China,
again using those numbers.

443
00:37:40,920 --> 00:37:44,397
How many of those kinds of plants
have we built in America? Zero.

444
00:37:44,480 --> 00:37:47,877
Uh and we're certainly losing the
renewables race to China for the reasons

445
00:37:47,960 --> 00:37:49,677
that everybody here knows.

446
00:37:49,760 --> 00:37:54,957
So, they have a lot of power, right?
We don't. We have a lot of chips,

447
00:37:55,040 --> 00:37:58,677
they don't. Um it sets
up the competition, right?

448
00:37:58,760 --> 00:38:03,437
And each will pursue different ways.
Um one of the technical questions

449
00:38:03,520 --> 00:38:07,357
is something called diffusion, where
what you do is you take a very powerful

450
00:38:07,440 --> 00:38:11,197
model, we'll use Gemini 3,
the top one at the moment,

451
00:38:11,280 --> 00:38:15,357
um and you ask it 10,000 questions
and you take the answers, and then the

452
00:38:15,440 --> 00:38:20,197
system can learn from the questions
and answers enough to mimic, without the

453
00:38:20,280 --> 00:38:24,637
expense of doing the big training,
the big model. So, again,

454
00:38:24,720 --> 00:38:28,917
thinking through the strategies that China
will pursue and the the strategies

455
00:38:29,000 --> 00:38:32,957
that are available for America um
are is probably pretty important.

456
00:38:33,040 --> 00:38:37,797
What's interesting is both countries are
relying on the private sector to do this.

457
00:38:37,880 --> 00:38:41,437
You would have thought in Henry's period
that you would have used the government.

458
00:38:41,520 --> 00:38:45,397
But in fact, our government cannot move
this quickly, the compensation systems,

459
00:38:45,480 --> 00:38:49,157
and so forth. It turns out it's probably
true for the Chinese as well. I don't

460
00:38:49,240 --> 00:38:51,157
know at the at the security level if

461
00:38:51,240 --> 00:38:53,677
this is true, but I've not found any

462
00:38:53,760 --> 00:38:58,357
large, weird, Manhattan-type projects
within China, although there's plenty of

463
00:38:58,440 --> 00:39:00,037
people in the private companies that are

464
00:39:00,120 --> 00:39:02,477
working hard on national security.

465
00:39:02,560 --> 00:39:06,397
So, on the AI topic in general, tell us

466
00:39:06,480 --> 00:39:08,477
Just give us a minute or two of what is

467
00:39:08,560 --> 00:39:10,637
the upside, the things that excite you

468
00:39:10,720 --> 00:39:13,677
the most that are 1 year, 2 year, 3

469
00:39:13,760 --> 00:39:16,157
years that are the line of sight to.

470
00:39:16,240 --> 00:39:20,157
Well, the first question is, why is this
madness occurring? It must be a bubble

471
00:39:20,240 --> 00:39:25,037
and it's going to crash. No,
it's not a bubble. If if anything,

472
00:39:25,120 --> 00:39:29,437
it's underhyped because you're
fundamentally automating businesses.

473
00:39:29,520 --> 00:39:33,397
And the reason people are spending this
enormous amount of money is to automate

474
00:39:33,480 --> 00:39:36,917
the boring parts of what they view
as the boring parts of their business. So,

475
00:39:37,000 --> 00:39:40,597
whether it's billing or accounting
or product design or delivery or inventory

476
00:39:40,680 --> 00:39:44,477
or management or whatever,
people are automating those. Um

477
00:39:44,560 --> 00:39:48,157
and there's an awful lot there.
Think about medicine, think about climate

478
00:39:48,240 --> 00:39:53,840
change, engineering,
new science, it's extraordinary.

479
00:39:57,160 --> 00:40:01,957
So, what excites you the most
of the ones that you see

480
00:40:02,040 --> 00:40:05,997
that you have that you have a line of
sight to that the rest of us probably do

481
00:40:06,080 --> 00:40:12,520
or not seeing yet? Sorry,
I have to Sorry. have a cough.

482
00:40:12,880 --> 00:40:17,637
I apologize. We can all

483
00:40:17,720 --> 00:40:19,717
see in our own imagination what we're

484
00:40:19,800 --> 00:40:25,037
thinking of and then we'll see what Eric
says. Yes. Well, when I started when I

485
00:40:25,120 --> 00:40:28,717
was in high school,
I was an early programmer.

486
00:40:28,800 --> 00:40:32,157
And I delighted in writing code. When
I went to college and graduate school,

487
00:40:32,240 --> 00:40:35,077
that's all I ever wanted to do. I didn't
I ignored all these history things and

488
00:40:35,160 --> 00:40:39,157
things like this. I was the
definition of a nerd at the time.

489
00:40:39,240 --> 00:40:43,517
And everything that I did in my 20s,
which got me to where I am, has now been

490
00:40:43,600 --> 00:40:47,277
completely automated.
Every aspect of the programming that I

491
00:40:47,360 --> 00:40:50,197
did, every aspect of the design is now

492
00:40:50,280 --> 00:40:54,877
done by computers. I recently
had it write a whole program

493
00:40:54,960 --> 00:40:59,197
for me and I'm sitting there watching it
define the classes and the detail of the

494
00:40:59,280 --> 00:41:03,477
interactions and so forth as it's
generating it got like, holy crap, you

495
00:41:03,560 --> 00:41:07,077
know, the end of me.
And I think watching I've been doing

496
00:41:07,160 --> 00:41:10,997
programming for 55 years.

497
00:41:11,080 --> 00:41:14,677
So, to see something start and then
end in front of your own life and you're

498
00:41:14,760 --> 00:41:18,037
still alive is really
profound, if I might say.

499
00:41:18,120 --> 00:41:20,557
Now, computer science is not going away.

500
00:41:20,640 --> 00:41:23,597
The computer scientist will be at least
until the computer scientist gets

501
00:41:23,680 --> 00:41:27,117
replaced, will be supervising this.

502
00:41:27,200 --> 00:41:30,277
But the ability to generate code at the

503
00:41:30,360 --> 00:41:34,677
power that these systems can do
is revolutionary. It means that each and

504
00:41:34,760 --> 00:41:37,077
every one of you has a supercomputer and

505
00:41:37,160 --> 00:41:39,357
a superprogrammer in your pocket.

506
00:41:39,440 --> 00:41:43,557
Now, nobody here is a terrorist, right?
But using it's always easier to

507
00:41:43,640 --> 00:41:47,917
use negative examples. There's plenty
of I'll use a stereotype, young men living

508
00:41:48,000 --> 00:41:50,997
in the basement, their mothers give
them food and they sit there on the

509
00:41:51,080 --> 00:41:55,437
equivalent of crypto for chan, you know,
paranoia, whatever. What pick your pick

510
00:41:55,520 --> 00:41:59,037
your poison. They all have
the ability to use these

511
00:41:59,120 --> 00:42:02,797
tools to build incredibly
powerful systems.

512
00:42:02,880 --> 00:42:06,877
Right? Cyber attacks, other things.
No whatever they care about.

513
00:42:06,960 --> 00:42:10,677
Um there was there's some evidence that I

514
00:42:10,760 --> 00:42:15,477
think it's Mengioni, the fellow who
killed the insurance insurance executive,

515
00:42:15,560 --> 00:42:19,237
um was into some of this and there's
some people were looking at some of his

516
00:42:19,320 --> 00:42:22,517
writings. Of course, he's in jail now,
but that he was somehow influenced with

517
00:42:22,600 --> 00:42:26,637
that. Now, I'm not proving causality here,
but it's an example of some of the

518
00:42:26,720 --> 00:42:30,957
darkest recesses of humanity.
You give those people these kinds of

519
00:42:31,040 --> 00:42:34,357
tools, we have to be ready. Now, the
industry is well aware of this and we're

520
00:42:34,440 --> 00:42:36,277
working on it. It's very important that

521
00:42:36,360 --> 00:42:39,117
the defensive systems be capable. The

522
00:42:39,200 --> 00:42:41,597
eventual solution to AI, by the way, bad

523
00:42:41,680 --> 00:42:46,237
AI is AI fighting good AI fighting
bad AI and that's how it will resolve

524
00:42:46,320 --> 00:42:49,957
itself. Okay, can we do
the next slide, please?

525
00:42:50,040 --> 00:42:52,797
I want to ask you about the US China

526
00:42:52,880 --> 00:42:57,197
rivalry in AI as you see it. So, this is

527
00:42:57,280 --> 00:42:59,317
apologies if the slide's not clear

528
00:42:59,400 --> 00:43:04,837
enough, but what it suggests is that if
we take the series of indices,

529
00:43:04,920 --> 00:43:08,597
uh if you look at the performance gap

530
00:43:08,680 --> 00:43:13,677
in January of 204,
it was significantly larger than it is

531
00:43:13,760 --> 00:43:18,480
today. And then the uh

532
00:43:20,160 --> 00:43:23,877
What do we make of this and what
do we make of the likely future?

533
00:43:23,960 --> 00:43:28,477
So, the chart is correct,
but the people who are

534
00:43:28,560 --> 00:43:32,477
influenced by this claim that it's
not going to be true for long

535
00:43:32,560 --> 00:43:37,277
because the reasoning revolution
requires so many chips and so much of

536
00:43:37,360 --> 00:43:40,717
the magic that the San
Francisco people have invented,

537
00:43:40,800 --> 00:43:42,917
using that as a sort of moniker, that

538
00:43:43,000 --> 00:43:47,757
the gap will widen.
My own view is that the gap will widen,

539
00:43:47,840 --> 00:43:51,837
but for different reasons.
I think that the Chinese focus is

540
00:43:51,920 --> 00:43:56,957
largely, as I mentioned,
on embedding AI in everything. Smart

541
00:43:57,040 --> 00:43:58,637
toasters, cars, so forth and so on.

542
00:43:58,720 --> 00:44:03,237
They're moving much more quickly than
our their robots. I think that the vast

543
00:44:03,320 --> 00:44:06,957
majority of humanoid robots will
be Chinese AI powered and Chinese AI

544
00:44:07,040 --> 00:44:10,917
manufactured simply because they know
how to drive the cost of things down.

545
00:44:11,000 --> 00:44:13,717
Their supply chains are incredible,
their cost management, they work so

546
00:44:13,800 --> 00:44:15,077
hard, blah blah blah, all that kind of

547
00:44:15,160 --> 00:44:16,277
stuff.

548
00:44:16,360 --> 00:44:20,757
So, my guess is that that it's true that

549
00:44:20,840 --> 00:44:24,677
the gap will probably get larger,
but the real question is will you as a

550
00:44:24,760 --> 00:44:28,317
consumer ultimately have a a better
experience as a consumer with a Chinese

551
00:44:28,400 --> 00:44:32,317
product than a US product?
And the answer is from a fit and finish,

552
00:44:32,400 --> 00:44:35,757
probably the Chinese product
and that's of concern.

553
00:44:35,840 --> 00:44:40,637
So, let me drill a little further on
this one. So, what there is a half dozen

554
00:44:40,720 --> 00:44:45,117
questions about which people are making
bets and you've thought about them quite

555
00:44:45,200 --> 00:44:47,357
deeply. So, one is are we going to bet

556
00:44:47,440 --> 00:44:48,317
on

557
00:44:48,400 --> 00:44:53,157
computer chips or stacks or brains?

558
00:44:53,240 --> 00:44:54,797
Another one is are we going to bet

559
00:44:54,880 --> 00:44:57,797
closed or open? Yep. Another one is are

560
00:44:57,880 --> 00:45:01,557
we going to bet um working hard on AGI

561
00:45:01,640 --> 00:45:04,757
or diffusion and applications?

562
00:45:04,840 --> 00:45:08,317
So, if you go across the spectrum there,

563
00:45:08,400 --> 00:45:09,357
uh

564
00:45:09,440 --> 00:45:12,877
if I look at the Chinese piece,

565
00:45:12,960 --> 00:45:17,117
uh certainly the guys
at Deep Seek think 200 people

566
00:45:17,200 --> 00:45:20,077
uh who just have brains

567
00:45:20,160 --> 00:45:22,637
can have a reasoning machine

568
00:45:22,720 --> 00:45:24,837
that costs 1/1000

569
00:45:24,920 --> 00:45:27,837
of the cost of Open AI.

570
00:45:27,920 --> 00:45:30,277
And now there's six other little dragons

571
00:45:30,360 --> 00:45:32,957
coming along from the same from the same

572
00:45:33,040 --> 00:45:35,597
space. So, that one makes me worry.

573
00:45:35,680 --> 00:45:40,637
On the closed versus open, if I remember
our last conversation, you had pretty

574
00:45:40,720 --> 00:45:44,677
much concluded that open is going
to beat closed, but all of our companies

575
00:45:44,760 --> 00:45:49,637
are mostly closed. So, what about that?

576
00:45:49,720 --> 00:45:52,357
And then thirdly, uh

577
00:45:52,440 --> 00:45:57,357
maybe if there's this AGI breakthrough,
all the rest of this other stuff won't

578
00:45:57,440 --> 00:46:00,237
matter, but if the

579
00:46:00,320 --> 00:46:03,157
if the diffusion and application is

580
00:46:03,240 --> 00:46:05,197
already working in

581
00:46:05,280 --> 00:46:08,357
these other arenas, so just take us

582
00:46:08,440 --> 00:46:13,557
another step on the on the rivalry.
So, diffusion refer backward diffusion

583
00:46:13,640 --> 00:46:17,277
refers to essentially
learning from one model by

584
00:46:17,360 --> 00:46:21,117
the by many pairs and learning
from it, as we discussed.

585
00:46:21,200 --> 00:46:25,677
My own opinion is that which is not
I'm not sure of, but I think so,

586
00:46:25,760 --> 00:46:29,677
that the large companies will ultimately
not release their largest models. It'll

587
00:46:29,760 --> 00:46:33,357
be too dangerous to do so that they'll
subset them. That's what I would do.

588
00:46:33,440 --> 00:46:36,237
Now, I assume that they'll
come to that decision.

589
00:46:36,320 --> 00:46:39,837
I think the most interesting question is
open versus closed. For the those of you

590
00:46:39,920 --> 00:46:41,877
who don't have a background in this,

591
00:46:41,960 --> 00:46:44,077
open source open weights, open source is

592
00:46:44,160 --> 00:46:48,037
what I did for decades. Um when you use

593
00:46:48,120 --> 00:46:52,277
any form of computer, much of the software
that you're using was developed

594
00:46:52,360 --> 00:46:55,837
by open source, which meant that the
source was published and people would

595
00:46:55,920 --> 00:46:59,157
collectively advance it. And there was
a whole movement around this, which I was

596
00:46:59,240 --> 00:47:03,477
part of. So, I'm relentlessly
open source in my view.

597
00:47:03,560 --> 00:47:06,117
Um the major companies are closed source

598
00:47:06,200 --> 00:47:11,117
largely for economic reasons.
Um basically, if you borrow 50 billion

599
00:47:11,200 --> 00:47:15,037
dollars from the financial market, they
do want a return and saying to them, oh,

600
00:47:15,120 --> 00:47:18,157
by the way, we're going to give all
the models away and you won't get your

601
00:47:18,240 --> 00:47:22,837
return probably isn't a very
good legal or financial strategy.

602
00:47:22,920 --> 00:47:24,797
So, it So, the American model seems to

603
00:47:24,880 --> 00:47:26,197
have emerged as

604
00:47:26,280 --> 00:47:29,917
as closed. Strangely, the Chinese model

605
00:47:30,000 --> 00:47:34,557
is completely open open
weights open source.

606
00:47:34,640 --> 00:47:39,677
Why? I don't know.
Um one possible explanation is that the

607
00:47:39,760 --> 00:47:42,797
Chinese government has figured out
that they were losing in closed source

608
00:47:42,880 --> 00:47:46,717
because they couldn't get the hardware
and open source, because it's free,

609
00:47:46,800 --> 00:47:50,077
gives them proliferation. So,
one of the consequences of open

610
00:47:50,160 --> 00:47:52,397
source open weights is the vast majority

611
00:47:52,480 --> 00:47:54,077
of humans on the planet will use Chinese

612
00:47:54,160 --> 00:47:57,797
models. Why? Cuz they're free, right?

613
00:47:57,880 --> 00:48:01,597
And most countries cannot afford
the computing and the data centers and so

614
00:48:01,680 --> 00:48:05,637
forth. They're just going to take the
Chinese models for free and embed them.

615
00:48:05,720 --> 00:48:09,517
Now, is that an issue? Absolutely,
because it comes with Chinese values,

616
00:48:09,600 --> 00:48:13,237
Chinese training, Chinese biases and so
forth. We would prefer to have it be

617
00:48:13,320 --> 00:48:17,557
American. Um we will see.
There is a there's a

618
00:48:17,640 --> 00:48:20,717
couple of open source
projects in America that I am a

619
00:48:20,800 --> 00:48:24,837
supporter of, but they
can't raise the 10 billion

620
00:48:24,920 --> 00:48:28,797
dollars that are needed from the public
markets to get to where they want. So,

621
00:48:28,880 --> 00:48:33,477
they are gems, but they're not
at the scale. I've argued that the US

622
00:48:33,560 --> 00:48:35,597
government should help fund them. I've

623
00:48:35,680 --> 00:48:37,357
argued that philanthropists should help

624
00:48:37,440 --> 00:48:40,637
fund them, but I don't really know.

625
00:48:40,720 --> 00:48:45,557
Okay, that's very helpful. So, one
last question and then we're going to uh

626
00:48:45,640 --> 00:48:50,597
take questions from the audience.
If you were picking two or three questions

627
00:48:50,680 --> 00:48:52,477
that a graduate student or an

628
00:48:52,560 --> 00:48:58,117
undergraduate interested in this space,
AI and geopolitics, should be thinking

629
00:48:58,200 --> 00:49:01,757
about. Give us two
or three questions that we

630
00:49:01,840 --> 00:49:05,877
might have somebody that has
an answer to or part of an answer to

631
00:49:05,960 --> 00:49:09,677
the next time you come. So, so a couple
well, again, we're dealing with some of

632
00:49:09,760 --> 00:49:10,797
the smartest people in the world here.

633
00:49:10,880 --> 00:49:14,917
So, the first question is what does
it mean to be human in the age of AI?

634
00:49:15,000 --> 00:49:19,597
That is a question that you can write
many PhD theses on. So, basically, study

635
00:49:19,680 --> 00:49:23,357
history, study philosophy, study how
people work, study economics and then

636
00:49:23,440 --> 00:49:26,837
figure out what this new technology
is going to do. We always in in my

637
00:49:26,920 --> 00:49:30,597
industry, because we didn't take those
classes, we always ignore those things,

638
00:49:30,680 --> 00:49:34,477
right? You guys are capable of answering
these questions in ways and if we get

639
00:49:34,560 --> 00:49:38,677
out of line, maybe you can use
that to call to call us out.

640
00:49:38,760 --> 00:49:40,677
The second question has to do with this

641
00:49:40,760 --> 00:49:44,397
China-US rivalry. Why China versus US?

642
00:49:44,480 --> 00:49:48,997
The only two countries that are going
to matter. And the reason is you need an

643
00:49:49,080 --> 00:49:52,997
enormous amount of money and an enormous
number of people. And as much as I like

644
00:49:53,080 --> 00:49:55,717
Europe, Europe is not organized, doesn't
have enough people, doesn't have enough

645
00:49:55,800 --> 00:49:59,397
money to do it. India is not
organized enough yet to do

646
00:49:59,480 --> 00:50:02,597
it, although they're trying. And most
of the other countries don't have enough

647
00:50:02,680 --> 00:50:05,157
money, don't have enough of the talent,
don't have the right universities and so

648
00:50:05,240 --> 00:50:07,677
forth. And then the third question has

649
00:50:07,760 --> 00:50:09,797
to do much more with conflict.

650
00:50:09,880 --> 00:50:13,197
In a world where you have terrorists who

651
00:50:13,280 --> 00:50:17,637
have access to AI and you have
governments that have access to AI, what

652
00:50:17,720 --> 00:50:22,077
does conflict look like? What does
a terrorist attack look like against a

653
00:50:22,160 --> 00:50:25,597
major power? Obviously, I'm not
endorsing that, it's a terrible thing.

654
00:50:25,680 --> 00:50:27,917
How do we defend against it? Same thing

655
00:50:28,000 --> 00:50:30,357
for China versus US. But what about

656
00:50:30,440 --> 00:50:32,877
Russia versus Ukraine? What about Europe

657
00:50:32,960 --> 00:50:37,717
versus whatever? Trying to understand
how conflict plays out in an algorithmic

658
00:50:37,800 --> 00:50:42,557
warfare where the AI is driving
everything is a very, very fertile area

659
00:50:42,640 --> 00:50:47,240
for research and new ideas.
And it's just starting.

660
00:50:47,680 --> 00:50:50,677
So, good topics for somebody to work on

661
00:50:50,760 --> 00:50:53,997
in the interim. Let's start here. Please

662
00:50:54,080 --> 00:50:57,237
introduce yourself and a short question.

663
00:50:57,320 --> 00:51:00,717
Hi, my name is Teresa. I'm a second year
student here at the Kennedy School and

664
00:51:00,800 --> 00:51:04,837
I'm in Professor Allison's national
security class. We had a class talk

665
00:51:04,920 --> 00:51:08,237
about cybersecurity
and most of the tech that

666
00:51:08,320 --> 00:51:11,357
underpins US homeland security is all in

667
00:51:11,440 --> 00:51:15,837
private hands. So, what governance
governance model do you think will

668
00:51:15,920 --> 00:51:20,077
actually help work to coordinate
government and private companies

669
00:51:20,160 --> 00:51:24,797
to during especially during
like a AI type of emergency?

670
00:51:24,880 --> 00:51:28,597
And given we talked a lot
about like US-China rivalry

671
00:51:28,680 --> 00:51:31,117
and despite the technology and

672
00:51:31,200 --> 00:51:35,877
investment rivalry, there's also a big
difference in governance model. So, how

673
00:51:35,960 --> 00:51:40,677
do you think the two governance model
will lead in terms of there's a homeland

674
00:51:40,760 --> 00:51:44,317
security cyber crisis? Well,
first place, I think in under the Trump

675
00:51:44,400 --> 00:51:47,477
administration, you're not going to see
much regulation of AI. That's pretty

676
00:51:47,560 --> 00:51:52,237
clear. In China, it appears that they're
allowing the the companies to do

677
00:51:52,320 --> 00:51:56,157
whatever they want.
Although they have the laws about

678
00:51:56,240 --> 00:51:59,077
various things, they don't seem to be
implementing them. So, it looks like

679
00:51:59,160 --> 00:52:01,717
it's an all-out business conflict. The

680
00:52:01,800 --> 00:52:04,197
biggest concern I have is cyber attacks.

681
00:52:04,280 --> 00:52:07,357
If you can write code the way that I've

682
00:52:07,440 --> 00:52:09,517
seen these things write, and every

683
00:52:09,600 --> 00:52:14,437
company in my world is now using
programmers and AI programmers together.

684
00:52:14,520 --> 00:52:18,277
It's extraordinary. It's happened
very quickly. Look at Claude code, for

685
00:52:18,360 --> 00:52:22,117
example, the most recent one which is by
far at the top. It's from Anthropic. And

686
00:52:22,200 --> 00:52:25,637
others coming. Gemini, of course, claims
it has a competitor, but at the moment

687
00:52:25,720 --> 00:52:29,037
Claude code is a bit
better, it looks like.

688
00:52:29,120 --> 00:52:32,037
If you can write code,
you can also write cyber attacks

689
00:52:32,120 --> 00:52:34,117
because the objective function is easy.

690
00:52:34,200 --> 00:52:36,917
Just keep writing code
until you break something.

691
00:52:37,000 --> 00:52:39,557
And if you have enough hardware
and enough energy, you can just keep doing

692
00:52:39,640 --> 00:52:43,717
that. So, I think there's going
to be a many, many more cyber attacks.

693
00:52:43,800 --> 00:52:47,397
And that doesn't mean from governments,
it could also be from terrorists and and

694
00:52:47,480 --> 00:52:51,237
bad organizations. And I think getting
organized for that would be my primary

695
00:52:51,320 --> 00:52:54,717
concerns. Please.

696
00:52:54,800 --> 00:53:00,157
Hi. Thank you for coming.
I'm David Wideman. I'm MPP2 here.

697
00:53:00,240 --> 00:53:05,357
You You mentioned the strength of current
closed-source models. You can tell me if

698
00:53:05,440 --> 00:53:08,357
I'm wrong, but I think open-source
models are about a half generation

699
00:53:08,440 --> 00:53:10,917
behind the current closed-source models.

700
00:53:11,000 --> 00:53:14,437
Perhaps it's a little dangerous
to characterize open-source versus

701
00:53:14,520 --> 00:53:19,600
closed-source as an America
versus China issue.

702
00:53:20,160 --> 00:53:23,037
And my my fear is that the Silicon
Valley people that you're talking about

703
00:53:23,120 --> 00:53:27,117
are seeking regulatory capture
and that's why they want to fearmonger

704
00:53:27,200 --> 00:53:31,437
around open-source models.
What would it take if we wanted to

705
00:53:31,520 --> 00:53:36,197
establish a moratorium on closed-source
models if that's something that we seek?

706
00:53:36,280 --> 00:53:39,797
What would it take in order to do
establish this moratorium worldwide?

707
00:53:39,880 --> 00:53:44,117
I think I just don't agree
with your question, so I'm sorry.

708
00:53:44,200 --> 00:53:48,797
I I don't see the open-source leadership
in America. And I see nothing but

709
00:53:48,880 --> 00:53:52,437
open-source leadership in China.
I think those are the facts.

710
00:53:52,520 --> 00:53:57,117
So, I don't think there's regulatory
capture on the open-source side.

711
00:53:57,200 --> 00:54:00,317
And I think again under the Trump
administration, it's highly unlikely

712
00:54:00,400 --> 00:54:03,797
that you're going to see significant
regulation of the closed-source

713
00:54:03,880 --> 00:54:08,197
companies. But I think the closed-source
decision is largely driven by economics,

714
00:54:08,280 --> 00:54:10,157
not policy. If you see my point.

715
00:54:10,240 --> 00:54:13,037
Literally, you just can't I mean, think
about the cost of these things. We're

716
00:54:13,120 --> 00:54:15,917
talking about $10 billion, $20 billion.

717
00:54:16,000 --> 00:54:19,277
How would you raise that without it?

718
00:54:19,360 --> 00:54:21,877
And I think Eric, this goes back to

719
00:54:21,960 --> 00:54:24,077
before making our list of questions, I

720
00:54:24,160 --> 00:54:26,597
would add add to it. How does the

721
00:54:26,680 --> 00:54:29,077
financing considerations

722
00:54:29,160 --> 00:54:32,077
impact the strategic choices? Because if

723
00:54:32,160 --> 00:54:34,597
it if it just happens to be an accident

724
00:54:34,680 --> 00:54:37,557
of the financial

725
00:54:37,640 --> 00:54:42,637
structure, that wouldn't necessarily
reflect the national interest. That

726
00:54:42,720 --> 00:54:46,717
would just reflect the capital
markets as they are. I think

727
00:54:46,800 --> 00:54:51,077
that it's important to acknowledge
that America has the most extraordinary

728
00:54:51,160 --> 00:54:53,157
capital market financial system in the

729
00:54:53,240 --> 00:54:55,957
world by far. I think the numbers are

730
00:54:56,040 --> 00:54:59,477
60% of the volume and 90% of of the

731
00:54:59,560 --> 00:55:03,197
value dollar and so
forth is in US dollars.

732
00:55:03,280 --> 00:55:07,037
And so countries that are
not dollar denominated do fear

733
00:55:07,120 --> 00:55:11,917
the power of that financial market. And we
see it in the ability to raise money.

734
00:55:12,000 --> 00:55:15,757
When I was in China, it was quite clear
in talking to my friends, they don't

735
00:55:15,840 --> 00:55:19,117
have access to the depth of the financial
market. They literally cannot

736
00:55:19,200 --> 00:55:23,557
get the money. The venture financing
that was in China is cut by a factor of

737
00:55:23,640 --> 00:55:28,157
five since three years ago. Now, there
are many reasons for that, not just our

738
00:55:28,240 --> 00:55:32,477
world and not just the US. But without
that access to capital, it's very hard

739
00:55:32,560 --> 00:55:37,037
to to do these large models with these
companies and training. Now, you can

740
00:55:37,120 --> 00:55:42,037
imagine everything I'm saying
changing if the underlying

741
00:55:42,120 --> 00:55:45,317
algorithms change. There are people
who are working on new non-transformer

742
00:55:45,400 --> 00:55:48,517
models which are less economically

743
00:55:48,600 --> 00:55:52,437
expensive. There are many people who
have commented on the energy efficiency

744
00:55:52,520 --> 00:55:56,117
of human brains versus the cost of these
data centers. And trust me, our brains

745
00:55:56,200 --> 00:56:01,077
are complicated, but they're not energy
they're not energy expensive. So, so

746
00:56:01,160 --> 00:56:04,117
again, there there are there might be
a breakthrough that would change the

747
00:56:04,200 --> 00:56:07,157
calculus that we're discussing.
Go ahead. Please.

748
00:56:07,240 --> 00:56:10,837
Thank you, Dr. Schmidt.
My name is Fatjon Semanajaku. I'm an

749
00:56:10,920 --> 00:56:13,277
MC/MPA student here at Harvard Kennedy

750
00:56:13,360 --> 00:56:16,557
School. Originally from Albania.

751
00:56:16,640 --> 00:56:19,597
In the age of AI, you and Dr. Kissinger

752
00:56:19,680 --> 00:56:23,637
describe a world where AI systems begin

753
00:56:23,720 --> 00:56:26,757
to interpret reality for us.

754
00:56:26,840 --> 00:56:29,197
If human strategic judgment becomes

755
00:56:29,280 --> 00:56:31,077
shaped, even subtly, by

756
00:56:31,160 --> 00:56:33,877
machine-generated frames, what becomes

757
00:56:33,960 --> 00:56:36,477
the anchor of responsibility in global

758
00:56:36,560 --> 00:56:39,277
affairs? In other words, when an

759
00:56:39,360 --> 00:56:42,077
AI-influenced decision produces a

760
00:56:42,160 --> 00:56:44,877
real-world consequences, who holds the

761
00:56:44,960 --> 00:56:47,837
moral burden? The human who acted, the

762
00:56:47,920 --> 00:56:51,317
institution that deployed the system, or

763
00:56:51,400 --> 00:56:53,677
the algorithm that shaped the

764
00:56:53,760 --> 00:56:56,837
perception? And how should democracies

765
00:56:56,920 --> 00:56:59,877
redesign their institutions before this

766
00:56:59,960 --> 00:57:03,117
ambiguity becomes a geopolitical

767
00:57:03,200 --> 00:57:05,557
vulnerability? A very well-phrased

768
00:57:05,640 --> 00:57:08,957
point. So, I worry that the future of

769
00:57:09,040 --> 00:57:11,757
democracy is uncertain

770
00:57:11,840 --> 00:57:16,237
simply because I'll use an American
example. We believe in free speech and

771
00:57:16,320 --> 00:57:18,277
I'm certainly in favor of absolute free

772
00:57:18,360 --> 00:57:20,677
speech in America. I'm not in favor of

773
00:57:20,760 --> 00:57:23,437
boosted speech or algorithmic speech.

774
00:57:23,520 --> 00:57:28,317
So, where is the line, right? If if I
say something that's wrong and then the

775
00:57:28,400 --> 00:57:31,517
algorithm, because my claim
is outrageous, decides to spread it

776
00:57:31,600 --> 00:57:35,837
everywhere, is that appropriate
in a democracy? And you can imagine, and

777
00:57:35,920 --> 00:57:37,317
again, I'm not making a partisan thing.

778
00:57:37,400 --> 00:57:41,277
I think any any any part of the
political system can use this for

779
00:57:41,360 --> 00:57:43,797
whatever effect. The ability to generate

780
00:57:43,880 --> 00:57:45,757
misinformation that people really

781
00:57:45,840 --> 00:57:50,237
believe is so simple now that I hope the

782
00:57:50,320 --> 00:57:54,357
answer is better education and critical
thinking among human beings. But you

783
00:57:54,440 --> 00:57:58,357
could imagine if I were evil, which
I hope I'm not, and I I sat down and I

784
00:57:58,440 --> 00:58:02,637
started to flood everyone with my
particular unique messages, that I could

785
00:58:02,720 --> 00:58:05,357
overwhelm your belief in truth by simply

786
00:58:05,440 --> 00:58:09,797
relentless copying and repetition.
We know that there's somebody called

787
00:58:09,880 --> 00:58:13,877
essentially anchor bias. If you hear it
first, then you judge So, if I managed

788
00:58:13,960 --> 00:58:17,477
to get the message to you first that the
building is on fire, which it's not, so

789
00:58:17,560 --> 00:58:21,597
I'm not committing a crime.
If I managed to get that first, you then

790
00:58:21,680 --> 00:58:26,157
anchor from that point. You see the
danger of this. It's very real. I don't

791
00:58:26,240 --> 00:58:28,077
think and my my answer to your question

792
00:58:28,160 --> 00:58:30,917
is that every every democracy will face

793
00:58:31,000 --> 00:58:35,517
this. And I think democracies will make
different decisions based on cultural

794
00:58:35,600 --> 00:58:39,557
values and their understanding of the
threat. You won't see a single answer

795
00:58:39,640 --> 00:58:43,037
from democracies. And by the way,
the way authoritarian countries solve this

796
00:58:43,120 --> 00:58:46,397
problem is they just ban it.
Right. So, a lot of this stuff is

797
00:58:46,480 --> 00:58:49,357
illegal in authoritarian countries
because they have the ability to

798
00:58:49,440 --> 00:58:53,917
suppress speech. And obviously,
we're we're all in favor of free speech.

799
00:58:54,000 --> 00:58:58,317
Please. Good. Yes, ma'am. Yes,
thank you very much for for coming

800
00:58:58,400 --> 00:59:02,317
here this evening. So, I'm called Elena
Kraine and I'm a fellow at the Belfast

801
00:59:02,400 --> 00:59:04,237
Center. So, I'm wondering firstly, could

802
00:59:04,320 --> 00:59:06,757
you comment on the role of humans such

803
00:59:06,840 --> 00:59:09,437
as you and me in the very long term with

804
00:59:09,520 --> 00:59:13,837
regards to AI? What's our what
is our role become? And also,

805
00:59:13,920 --> 00:59:16,157
we often frame, you know, because of the

806
00:59:16,240 --> 00:59:18,877
value systems, the US-China debate as a

807
00:59:18,960 --> 00:59:23,437
competition. But is there is there value
to thinking about that in terms of a

808
00:59:23,520 --> 00:59:26,597
collaboration? And maybe more
interestingly, you know, you mentioned

809
00:59:26,680 --> 00:59:28,957
Europe being maybe somewhat
disorganized, but they also have

810
00:59:29,040 --> 00:59:33,037
enormous strength and enormous talent
like Mistral in France and so on.

811
00:59:33,120 --> 00:59:36,277
Is there something to do
there with um collaboration?

812
00:59:36,360 --> 00:59:39,837
Um so so a couple comments on Europe, um

813
00:59:39,920 --> 00:59:43,877
I was the first investor in Mistral,
so I'm a big Mistral fan.

814
00:59:43,960 --> 00:59:48,037
Mistral cannot raise the money that
their US competitors. They have the

815
00:59:48,120 --> 00:59:50,277
same problem and they're working on

816
00:59:50,360 --> 00:59:55,277
solutions to that.
Um with respect to US and China,

817
00:59:55,360 --> 00:59:58,560
the um

818
00:59:58,760 --> 01:00:03,397
I started because of Henry, I I spent
about 5 years looking at Chinese peer

819
01:00:03,480 --> 01:00:07,917
competition. And I thought
that it would be possible

820
01:00:08,000 --> 01:00:11,917
to get closer to China.
And I eventually through his work and

821
01:00:12,000 --> 01:00:15,477
others discovered that the Chinese were
more afraid of competing partnering with

822
01:00:15,560 --> 01:00:20,837
us than we were with them.
So it takes two to tango, right? And I

823
01:00:20,920 --> 01:00:24,957
think it's highly unlikely. I think
you're going to see in your lifetime

824
01:00:25,040 --> 01:00:28,757
um hopefully understanding that we have to

825
01:00:28,840 --> 01:00:33,717
coexist, but it's highly unlikely that
the two systems will become best buddies

826
01:00:33,800 --> 01:00:35,597
for the obvious reasons.

827
01:00:35,680 --> 01:00:37,557
I think on this this other question

828
01:00:37,640 --> 01:00:41,557
about what is what are humans useful for

829
01:00:41,640 --> 01:00:46,357
uh in the very long term.
Um it's very very clear that humans are

830
01:00:46,440 --> 01:00:49,677
social animals and want
to be around other humans.

831
01:00:49,760 --> 01:00:53,597
It's also very clear that there are some
things that we need, medical care and so

832
01:00:53,680 --> 01:00:56,037
forth, which will be delivered by humans.

833
01:00:56,120 --> 01:00:58,477
I think that that it's a rough way to

834
01:00:58,560 --> 01:01:03,637
say this is most other
functions will be doable by

835
01:01:03,720 --> 01:01:07,837
computers. Will we allow them?
That's a question

836
01:01:07,920 --> 01:01:09,677
for your research.

837
01:01:09,760 --> 01:01:11,837
Right? Is where is that line?

838
01:01:11,920 --> 01:01:15,237
My my old example was the following.

839
01:01:15,320 --> 01:01:19,077
This is back when we were building Waymo
at Google. So the thought experiment is

840
01:01:19,160 --> 01:01:23,357
New York City is nothing but Waymo
and competitor automobile cars.

841
01:01:23,440 --> 01:01:25,957
And the engineers at Google and the other
companies have figured out how to

842
01:01:26,040 --> 01:01:30,757
completely optimize the traffic so that
the the streets which don't change, you

843
01:01:30,840 --> 01:01:34,957
have the absolutely mathematically true
highest loading per street ever and it's

844
01:01:35,040 --> 01:01:39,317
seamless. And then you have somebody with
an emergency, a pregnant woman, you know,

845
01:01:39,400 --> 01:01:42,757
whatever who has to get to the
hospital and needs an exception.

846
01:01:42,840 --> 01:01:47,437
Is there an exception button in the car
which says you have to violate all the

847
01:01:47,520 --> 01:01:51,477
rules because there's something uh
Now if there is not an exception button

848
01:01:51,560 --> 01:01:56,597
in the car, then that's
dominance of humans by a computer.

849
01:01:56,680 --> 01:02:00,877
And humans will revolt. If I go back to
the pickaxe versus religion, humans will

850
01:02:00,960 --> 01:02:02,997
fight that. That will be seen as

851
01:02:03,080 --> 01:02:06,037
oppression by the government, in this

852
01:02:06,120 --> 01:02:10,637
case, in the governor of New York
City, whatever metaphor you want,

853
01:02:10,720 --> 01:02:12,637
and the computer companies themselves.

854
01:02:12,720 --> 01:02:16,757
If the system, on the other hand,
is adaptable to human needs

855
01:02:16,840 --> 01:02:20,437
and says, "Ah, we have an actual, you
know, medical emergency and not somebody

856
01:02:20,520 --> 01:02:23,717
who's high on drugs or, you know,
whatever who's just goofing around or

857
01:02:23,800 --> 01:02:26,237
some kid who's, you know,
playing with the buttons."

858
01:02:26,320 --> 01:02:30,357
Um and it reasons and it says, "Oh my god,
I've got to do everything I can to

859
01:02:30,440 --> 01:02:32,237
get the the person to the hospital."

860
01:02:32,320 --> 01:02:36,477
Then it'll be more accepted. So a lot
of this depends on whether it integrates

861
01:02:36,560 --> 01:02:38,677
with our human experience and needs.

862
01:02:38,760 --> 01:02:42,757
Does it limit our freedom
or does it increase freedom?

863
01:02:42,840 --> 01:02:46,437
Uh I've come to the view, I'm old
enough now to believe that preserving our

864
01:02:46,520 --> 01:02:47,957
freedom, our freedom of thought, of

865
01:02:48,040 --> 01:02:50,517
motion, of assembly, of gathering, you

866
01:02:50,600 --> 01:02:55,397
know, of all of the things, right?
Uh are really really important. If they

867
01:02:55,480 --> 01:02:58,797
impinge on our freedom, then they will
be fought and I will be leading that

868
01:02:58,880 --> 01:03:01,197
fight.

869
01:03:01,280 --> 01:03:03,437
Please.

870
01:03:03,520 --> 01:03:06,157
Namaste, sir. My name is Ilma Afroz. I'm

871
01:03:06,240 --> 01:03:11,277
from India. You mentioned India in your
conversation. I'm lucky to be Professor

872
01:03:11,360 --> 01:03:12,837
Allison's student and to be learning

873
01:03:12,920 --> 01:03:15,717
from him. My question is how we can

874
01:03:15,800 --> 01:03:18,637
bring together US and India. India has

875
01:03:18,720 --> 01:03:21,037
got some excellent talented people. How

876
01:03:21,120 --> 01:03:23,517
we can bring both our countries together

877
01:03:23,600 --> 01:03:26,957
in order uh to create a world where

878
01:03:27,040 --> 01:03:29,997
democracies flourish and it's a win-win

879
01:03:30,080 --> 01:03:32,757
partnership for both our countries.

880
01:03:32,840 --> 01:03:37,317
Um I strongly agree.
I've spent lots of time in India and the

881
01:03:37,400 --> 01:03:41,797
literally because of the IITs
and the quality of the people, the

882
01:03:41,880 --> 01:03:46,517
um the depth of talent in India
is extraordinary. However, the depth of

883
01:03:46,600 --> 01:03:50,317
computing is not. Um last year
we did a calculation, there

884
01:03:50,400 --> 01:03:54,877
were only about a thousand GPUs in the
entire country for 1 billion people. So

885
01:03:54,960 --> 01:03:57,037
I and others have organized to try to to

886
01:03:57,120 --> 01:04:02,077
fix that. Um so I was alarmed
with the most recent

887
01:04:02,160 --> 01:04:05,837
trade war, uh which I
think has set uh India and

888
01:04:05,920 --> 01:04:09,557
the United States back and that needs
to get fixed. I don't understand the trade

889
01:04:09,640 --> 01:04:11,757
war and I just my position is that's

890
01:04:11,840 --> 01:04:14,477
hurting us. India is our natural

891
01:04:14,560 --> 01:04:17,397
partner, right? It's a democracy, it's a

892
01:04:17,480 --> 01:04:20,637
messy democracy, but we are too. Um

893
01:04:20,720 --> 01:04:22,957
w- in in Silicon Valley, most of the

894
01:04:23,040 --> 01:04:25,037
people I work with are of

895
01:04:25,120 --> 01:04:28,117
um essentially South Asian origin with

896
01:04:28,200 --> 01:04:32,717
India or some of the other countries.
So at least in Silicon Valley, the Indians

897
01:04:32,800 --> 01:04:37,077
are flourishing and I want the tightest
possible integration with them.

898
01:04:37,160 --> 01:04:42,357
Thank you. Please. Hi,
my name's Josh. I'm an MPP2 here at

899
01:04:42,440 --> 01:04:45,957
the Kennedy School. Earlier,
you outlined two national strategies

900
01:04:46,040 --> 01:04:50,557
earlier. There was pushing the frontier
or pursuing adoption, but ideally you

901
01:04:50,640 --> 01:04:55,037
have both, so thinking about
adoption in the US, what

902
01:04:55,120 --> 01:04:59,357
are some of the major barriers
to enterprise adoption right now?

903
01:04:59,440 --> 01:05:02,517
Does that vary by industry and what
should government do to help address

904
01:05:02,600 --> 01:05:06,077
those barriers if anything?
Um the government usually doesn't do a

905
01:05:06,160 --> 01:05:08,797
very good job on that. Um

906
01:05:08,880 --> 01:05:10,917
the industry believes that there is

907
01:05:11,000 --> 01:05:13,397
something called a technology overhang,

908
01:05:13,480 --> 01:05:16,117
that we, your friendly industry, have

909
01:05:16,200 --> 01:05:18,877
produced more tools than you are using.

910
01:05:18,960 --> 01:05:20,677
Now that that is our belief, whether

911
01:05:20,760 --> 01:05:22,837
it's true or not, you can discuss.

912
01:05:22,920 --> 01:05:26,397
Um and a lot of it seems to be readiness

913
01:05:26,480 --> 01:05:29,917
to adopt the technology,
uh the fact that a lot of the stuff is

914
01:05:30,000 --> 01:05:32,877
software, most companies don't have very
good software people. I'm sorry to say

915
01:05:32,960 --> 01:05:36,077
that. Um the the changes internally that

916
01:05:36,160 --> 01:05:40,717
are required. Um I personally believe
that this technology adoption thing is

917
01:05:40,800 --> 01:05:45,677
just a temporary problem and that
as new CEOs come in and winners

918
01:05:45,760 --> 01:05:47,637
emerge, the competitive pressure, which

919
01:05:47,720 --> 01:05:49,797
are very high in America, will cause

920
01:05:49,880 --> 01:05:52,837
that adoption. Um you don't see very

921
01:05:52,920 --> 01:05:57,477
much adoption in regulated industries
because regulation is used as a excuse

922
01:05:57,560 --> 01:06:00,557
to not innovate,
but at least in the hard innovative

923
01:06:00,640 --> 01:06:04,757
industries, I think this will get solved.
But the general answer is I

924
01:06:04,840 --> 01:06:07,117
would not have the government do very
much because I don't think it would

925
01:06:07,200 --> 01:06:11,717
help. Um what would help is just having
every business understand that if they

926
01:06:11,800 --> 01:06:15,197
want to make money, which is the
capitalist vision, they need to use AI

927
01:06:15,280 --> 01:06:19,117
much more profoundly. Remember, you
can target your customers, you can serve

928
01:06:19,200 --> 01:06:23,277
them, you can understand them, you can
talk to them and so forth, all with AI.

929
01:06:23,360 --> 01:06:26,877
Um and a lot of this there's a lot
of negative issues with this. For example,

930
01:06:26,960 --> 01:06:28,997
it may lead to layoffs in companies.

931
01:06:29,080 --> 01:06:33,397
There's lots of examples where low-end
jobs are being replaced by computers and

932
01:06:33,480 --> 01:06:37,157
that's obviously a job loss, that's
a societal problem, but from an adoption

933
01:06:37,240 --> 01:06:40,077
perspective, the answer is just time.

934
01:06:40,160 --> 01:06:42,997
Please. Hi, my name's Ayush.
I'm a third-year

935
01:06:43,080 --> 01:06:44,917
PhD student in computer science doing AI

936
01:06:45,000 --> 01:06:50,277
research. Um my question is regarding
um you were mentioning China's sort of

937
01:06:50,360 --> 01:06:53,997
intense focus on automating AI in
business. I guess in the Valley it seems

938
01:06:54,080 --> 01:06:56,877
like, you know, every day there's a new
startup that comes around like there's

939
01:06:56,960 --> 01:07:01,357
the big seed targeting some automation
for workflow and we we also sort of have

940
01:07:01,440 --> 01:07:03,557
the the 9-9-6 fad.

941
01:07:03,640 --> 01:07:05,677
Um what would you identify as like a key

942
01:07:05,760 --> 01:07:08,997
distinguisher in China's AI um

943
01:07:09,080 --> 01:07:12,877
automating business that you think
the US needs to catch up on or at least in

944
01:07:12,960 --> 01:07:17,637
the the Silicon Valley startups now
that is What is your PhD topic? I work on

945
01:07:17,720 --> 01:07:19,317
reasoning for large language models.

946
01:07:19,400 --> 01:07:22,517
Perfect. Right?
So what is your time frame for

947
01:07:22,600 --> 01:07:26,517
AGI? Uh I'd give it
like six to seven years.

948
01:07:26,600 --> 01:07:30,117
Yeah, you see, not San Francisco
consensus, East Coast consensus. Good.

949
01:07:30,200 --> 01:07:31,517
We will see.

950
01:07:31,600 --> 01:07:37,680
You may be right. Um I think that the

951
01:07:38,320 --> 01:07:42,077
It has to do, I think,
with the grandeur of the dreams.

952
01:07:42,160 --> 01:07:45,237
When I'm in China, I don't hear the same

953
01:07:45,320 --> 01:07:48,557
rhetoric as I do in California.

954
01:07:48,640 --> 01:07:53,797
In California, it's two years,
the world will change, no one is ready,

955
01:07:53,880 --> 01:07:56,717
this is happening so fast and so forth
and so on. It's this sort of rhetoric

956
01:07:56,800 --> 01:07:59,997
that drives and it's self-reproducing,
it's a belief system, it's like a

957
01:08:00,080 --> 01:08:01,237
religion.

958
01:08:01,320 --> 01:08:04,797
Now it always takes longer than than uh

959
01:08:04,880 --> 01:08:07,157
the dreamers say. I don't hear that in

960
01:08:07,240 --> 01:08:11,837
China. It it's it's it's
noticeably different.

961
01:08:11,920 --> 01:08:15,317
So for example, with reasoning, you know
what DeepSeek did with R3, they did a

962
01:08:15,400 --> 01:08:18,837
fantastic job, they invented a new way
of doing supervised fine-tuning. I mean,

963
01:08:18,920 --> 01:08:22,757
it's really really clever stuff. You
just don't hear it at a national scale.

964
01:08:22,840 --> 01:08:25,877
And by the way, DeepSeek
is a national champion for China.

965
01:08:25,960 --> 01:08:28,597
Right? They literally now are, you know,
they they're they're on the list of

966
01:08:28,680 --> 01:08:32,077
famous companies, they're getting
a enormous amount of money. Uh last time I

967
01:08:32,160 --> 01:08:35,437
met with them, they said, "We've solved
our hardware problem." which was code

968
01:08:35,520 --> 01:08:36,757
for the government is just going to give

969
01:08:36,840 --> 01:08:42,760
them a lot of chips. Right?
Welcome to a communist country.

970
01:08:43,480 --> 01:08:47,157
Okay, so this gentle lady
here, please. Hi, thank

971
01:08:47,240 --> 01:08:48,597
you so much for being here. I'm Sonia.

972
01:08:48,680 --> 01:08:51,077
I'm a joint degree student with the MBA

973
01:08:51,160 --> 01:08:55,757
and uh and the policy school.
Um this might be a slightly more out there

974
01:08:55,840 --> 01:08:59,957
question and and open-ended, but you
talk a lot about what it means to be

975
01:09:00,040 --> 01:09:03,917
human and I think a lot
of that is consciousness.

976
01:09:04,000 --> 01:09:07,597
Um and I'm curious when I think
about theories of consciousness, which

977
01:09:07,680 --> 01:09:11,837
obviously is very ill-defined, AI is
hitting more and more of those and so it

978
01:09:11,920 --> 01:09:14,517
it like if anything, we need
a revolution in how we think about

979
01:09:14,600 --> 01:09:17,197
consciousness, but I'm curious how you

980
01:09:17,280 --> 01:09:20,837
think about consciousness in AI and if

981
01:09:20,920 --> 01:09:24,557
you do think that there's a possibility
that AI is or will become conscious,

982
01:09:24,640 --> 01:09:27,917
what that looks like, what model welfare
looks like in that case. Well, let me

983
01:09:28,000 --> 01:09:29,357
ask let me ask you a simple question.

984
01:09:29,440 --> 01:09:34,517
So, we have I'm going to give
Graham credit for being conscious.

985
01:09:34,600 --> 01:09:37,997
And in return, he appears to be
conscious. And in return, Graham is

986
01:09:38,080 --> 01:09:41,397
going to give it to me. So, for purposes
of argument, the computer is on the

987
01:09:41,480 --> 01:09:43,717
table. And we ask the computer, "Are you

988
01:09:43,800 --> 01:09:46,597
conscious?" And it says, "Yes."

989
01:09:46,680 --> 01:09:49,037
And so, you, being the smart graduate

990
01:09:49,120 --> 01:09:51,917
student here at Harvard, um

991
01:09:52,000 --> 01:09:54,037
come up with a series of questions.

992
01:09:54,120 --> 01:09:56,997
It answers every question correctly.

993
01:09:57,080 --> 01:09:59,837
How do we know? How could
you know that the computer is

994
01:09:59,920 --> 01:10:04,637
conscious? How would you understand
its internal reasoning state?

995
01:10:04,720 --> 01:10:08,077
Now, you could instrument the way its
reasoning works, which people are now

996
01:10:08,160 --> 01:10:11,197
doing. What they're doing is they're
watching what are called supernodes

997
01:10:11,280 --> 01:10:15,797
within the um within the weight
structure to watch how it's actually

998
01:10:15,880 --> 01:10:19,357
making its decisions. So, maybe
you could discover consciousness by

999
01:10:19,440 --> 01:10:22,437
inspection, but that's speculation.

1000
01:10:22,520 --> 01:10:24,917
So, I sat down with a bunch
of neuroscientists cuz I didn't know the

1001
01:10:25,000 --> 01:10:28,917
answer to this question. I said, "How
does it How does consciousness evolve?"

1002
01:10:29,000 --> 01:10:32,477
And their theory, which is just a theory,

1003
01:10:32,560 --> 01:10:36,997
was that consciousness evolves when you
have dissimilar systems that are working

1004
01:10:37,080 --> 01:10:41,277
together and growing,
and they develop an awareness of the

1005
01:10:41,360 --> 01:10:44,317
other. Right? And that human consciousness

1006
01:10:44,400 --> 01:10:47,157
evolved because we needed to understand

1007
01:10:47,240 --> 01:10:51,477
that we were something. Right? That
the the the development of the id, the

1008
01:10:51,560 --> 01:10:55,957
identity, and so forth and so on was
necessary in order to command the system.

1009
01:10:56,040 --> 01:10:58,037
Now, there's no way to prove that.

1010
01:10:58,120 --> 01:11:02,197
So, the answer is I don't know, but this
is a good another it's a fourth question

1011
01:11:02,280 --> 01:11:04,717
to ask this audience to work on.

1012
01:11:04,800 --> 01:11:07,757
And and there's two questions. The first,
how does it work? And the second

1013
01:11:07,840 --> 01:11:11,397
is, how would you verify it? Thank you.

1014
01:11:11,480 --> 01:11:13,997
I have the unfortunate

1015
01:11:14,080 --> 01:11:18,677
responsibility to say that we've come to
the witching hour. We there Graham, this

1016
01:11:18,760 --> 01:11:21,877
is so much fun. I love this. Well, you
want to stay for a while more? Five more

1017
01:11:21,960 --> 01:11:23,157
minutes. This is good. I love this.

1018
01:11:23,240 --> 01:11:28,117
>> no. We'll stay forever. I'm sorry.
It's just I was pro- I I promised your I

1019
01:11:28,200 --> 01:11:32,037
promised your schedulers that I
would get you out of here. So, please.

1020
01:11:32,120 --> 01:11:34,237
Short questions and short answers.

1021
01:11:34,320 --> 01:11:37,477
Uh thanks, Eric. Troop here from Harvard
Business School, a second-year student

1022
01:11:37,560 --> 01:11:40,797
there. Also a background
in international relations before my MBA.

1023
01:11:40,880 --> 01:11:43,837
Um you Are you convincingly
about the need for an

1024
01:11:43,920 --> 01:11:48,077
IAEA or International Atomic
Energy Agency equivalent for AI?

1025
01:11:48,160 --> 01:11:51,317
So, my question is, how does the State
Department, how does the Pentagon, and

1026
01:11:51,400 --> 01:11:54,997
similar agencies need to configure
themselves to prepare for this the

1027
01:11:55,080 --> 01:11:57,317
coming era of superintelligence? And my

1028
01:11:57,400 --> 01:12:02,437
second, slightly risky question is,
um if you are looking for someone to help

1029
01:12:02,520 --> 01:12:06,477
you this summer or this winter to work
on this topic, I'd be very happy to.

1030
01:12:06,560 --> 01:12:08,477
I love people with courage.

1031
01:12:08,560 --> 01:12:11,357
Um so, the IAEA question. So, there's a

1032
01:12:11,440 --> 01:12:16,477
group of people who are very close to me
who concluded that the only solution to

1033
01:12:16,560 --> 01:12:18,357
the problems that I'm answering

1034
01:12:18,440 --> 01:12:22,197
is to create the equivalent of CERN.

1035
01:12:22,280 --> 01:12:25,877
And the idea was that we would get
together, all of us, including China,

1036
01:12:25,960 --> 01:12:30,197
you know, everybody, and we put it all
in here. Take all the best minds, and

1037
01:12:30,280 --> 01:12:31,997
we'd all work on this.

1038
01:12:32,080 --> 01:12:35,797
And um as a result, we would

1039
01:12:35,880 --> 01:12:40,117
m- m- build this huge future because
the benefits to humanity of this are so

1040
01:12:40,200 --> 01:12:43,237
great. Start thinking about every disease

1041
01:12:43,320 --> 01:12:47,637
eliminated, solutions to real hard
problems. The We have Megan here. Energy

1042
01:12:47,720 --> 01:12:51,997
issues, and she's the expert.
All of this kind of stuff.

1043
01:12:52,080 --> 01:12:55,397
I think it's pretty you
probably say it's probably pretty

1044
01:12:55,480 --> 01:12:59,397
unlikely that's going to happen. So,
then the next group says, "Well, then we

1045
01:12:59,480 --> 01:13:00,997
need an a an

1046
01:13:01,080 --> 01:13:04,237
IAEA, which is essentially a mandatory

1047
01:13:04,320 --> 01:13:07,237
inspection group for nuclear."

1048
01:13:07,320 --> 01:13:09,797
And what in that scenario, what happens

1049
01:13:09,880 --> 01:13:14,517
is you'd have to go, and this group
would visit the data centers and the

1050
01:13:14,600 --> 01:13:18,157
algorithms of every
company in every country.

1051
01:13:18,240 --> 01:13:22,917
Now, that occurred, remember,
the Vienna Group occurred

1052
01:13:23,000 --> 01:13:26,397
after 15 years of negotiation,
including people like Henry,

1053
01:13:26,480 --> 01:13:30,677
and after two nuclear bombs were dropped
on Hiroshima and Nagasaki, which we can

1054
01:13:30,760 --> 01:13:35,437
all agree was hor- horrific.
We've not had such an example.

1055
01:13:35,520 --> 01:13:39,357
There are people that I know who say the
following, and they don't say this with

1056
01:13:39,440 --> 01:13:43,877
evil intent. They say
that we won't we will have an

1057
01:13:43,960 --> 01:13:45,437
event

1058
01:13:45,520 --> 01:13:48,437
that will force this IAEA thing, and we

1059
01:13:48,520 --> 01:13:52,117
hope that it is a Chernobyl-level event.

1060
01:13:52,200 --> 01:13:53,597
And I said, "What does that mean?" And

1061
01:13:53,680 --> 01:13:56,717
they said, "Not that many people died."

1062
01:13:56,800 --> 01:14:00,037
Okay? As opposed to a nuclear attack.

1063
01:14:00,120 --> 01:14:04,597
So, there's a a strain of thinking
in my world that is consistent with that.

1064
01:14:04,680 --> 01:14:07,917
But I think uh first,
I think nobody knows, and it's

1065
01:14:08,000 --> 01:14:12,877
unlikely that the countries will um

1066
01:14:12,960 --> 01:14:16,997
put up with that uh until
there's a real crisis.

1067
01:14:17,080 --> 01:14:21,197
I used to say that in

1068
01:14:21,280 --> 01:14:23,477
I don't know, 20 years

1069
01:14:23,560 --> 01:14:26,117
after the climate is being destroyed

1070
01:14:26,200 --> 01:14:31,077
and, you know, the Earth is melting
and sea level rise, there's going to be a

1071
01:14:31,160 --> 01:14:33,757
there's going to there's going to be
a meeting of all the powers where people

1072
01:14:33,840 --> 01:14:35,717
say, "Well, we really screwed this up.

1073
01:14:35,800 --> 01:14:37,637
We now have to fix it."

1074
01:14:37,720 --> 01:14:42,797
So, I think there is precedent for a
global challenge and a resolution, but

1075
01:14:42,880 --> 01:14:46,437
it's very messy. It's not
done rationally. It's not like

1076
01:14:46,520 --> 01:14:50,437
you and I agree. There has to be some
crisis and some political dynamic, and

1077
01:14:50,520 --> 01:14:53,597
it can be quite severe.
It's an opportunity for the next

1078
01:14:53,680 --> 01:14:57,677
generation. Let's do this very quick.
Do we take the three questions? Just do

1079
01:14:57,760 --> 01:15:01,597
your short questions.
We'll let Eric wrap up, please. Yes.

1080
01:15:01,680 --> 01:15:04,517
Hi. Uh thank you for your extra time.
Uh my name is Kevin. I'm a second-year at

1081
01:15:04,600 --> 01:15:09,197
the law school. Um my question is about
automation and jobs. Um a couple months

1082
01:15:09,280 --> 01:15:12,677
ago, Sam Altman went on an interview
saying that insinuating that if

1083
01:15:12,760 --> 01:15:16,557
something could be automated, maybe it
wasn't a real job to begin with. Uh I

1084
01:15:16,640 --> 01:15:20,317
want to get your thoughts on that
statement, and if you agree with the

1085
01:15:20,400 --> 01:15:24,437
premise that if something can be
automated, should it? Um and relatedly,

1086
01:15:24,520 --> 01:15:28,197
if you disagree, what are ways that we
as a society can decide what we should

1087
01:15:28,280 --> 01:15:31,317
and shouldn't automate?
>> Sure. Please have it.

1088
01:15:31,400 --> 01:15:35,237
Um good night. My name is Griselda.
I'm a middle mid-career MPA and a former

1089
01:15:35,320 --> 01:15:37,437
Googler, so I'm very happy to see you.

1090
01:15:37,520 --> 01:15:39,957
Um my question is regarding algorithm

1091
01:15:40,040 --> 01:15:42,437
diplomacy. Like Dr. Kissinger used to

1092
01:15:42,520 --> 01:15:45,757
say that constructive um ambiguity and

1093
01:15:45,840 --> 01:15:50,037
the human pause is really necessary,
but then AI is binary.

1094
01:15:50,120 --> 01:15:53,317
Are we getting to a towards a world
where de-escalation is computationally

1095
01:15:53,400 --> 01:15:57,677
impossible? And does this push for new
algorithm diplomacy that we could push

1096
01:15:57,760 --> 01:16:01,237
harder? Ben. Uh thank you.
My name is Kanishk.

1097
01:16:01,320 --> 01:16:04,197
I'm from Harvard Business School.
Uh my question is about ethical

1098
01:16:04,280 --> 01:16:08,317
superintelligence. Uh first, do you
Why do you not believe that if America

1099
01:16:08,400 --> 01:16:11,757
develops superintelligence, other
nations will follow and be able to copy

1100
01:16:11,840 --> 01:16:15,277
it, like China did with Deep Seek?
And then second of all, if we're trying to

1101
01:16:15,360 --> 01:16:18,317
embody it with ethics, then if
someone else comes up with an unethical

1102
01:16:18,400 --> 01:16:22,117
superintelligence, isn't that going
to be strictly more free and more capable

1103
01:16:22,200 --> 01:16:26,597
than a constrained superintelligence
model that America tries to develop?

1104
01:16:26,680 --> 01:16:29,437
You It's uh one word
about your question again?

1105
01:16:29,520 --> 01:16:33,397
Uh yeah, I wanted to ask what your
thoughts were on Sam Altman's quote.

1106
01:16:33,480 --> 01:16:38,837
Yeah. >> So, um one of the
things that people at this

1107
01:16:38,920 --> 01:16:43,797
level forget is that human
dignity involves purpose.

1108
01:16:43,880 --> 01:16:47,677
And an awful lot of jobs provide purpose
to an awful lot of people. And the loss

1109
01:16:47,760 --> 01:16:52,397
of those jobs is a major crisis,
not financially, but emotionally, in terms

1110
01:16:52,480 --> 01:16:55,757
of their meaning. So,
in order for us to get through this,

1111
01:16:55,840 --> 01:16:58,357
we're going to have to address that.

1112
01:16:58,440 --> 01:17:01,877
Right? We're going to have
to actually do the right thing.

1113
01:17:01,960 --> 01:17:06,597
And the right thing will be some
combination of better tools and so forth.

1114
01:17:06,680 --> 01:17:10,717
I'm not as worried about the jobs issue
because we're producing fewer humans,

1115
01:17:10,800 --> 01:17:12,717
which I view as a major crisis.

1116
01:17:12,800 --> 01:17:16,797
Um collectively, you all are having
fewer children than my generation, which

1117
01:17:16,880 --> 01:17:20,397
is having fewer children than
my parents' generation, and so forth. And

1118
01:17:20,480 --> 01:17:25,597
we need more humans. Um partly I'm
a businessman, so you need more customers.

1119
01:17:25,680 --> 01:17:29,397
But the important point is, we need more
humans. And if we have less humans, then

1120
01:17:29,480 --> 01:17:33,917
there's going to be open jobs and no
people to fill them. AI can help take

1121
01:17:34,000 --> 01:17:36,357
people who are not ready
for a job and get them.

1122
01:17:36,440 --> 01:17:40,397
One word from yours, aside
from being a a future Googler?

1123
01:17:40,480 --> 01:17:42,357
No, the algorithm diplomacy and

1124
01:17:42,440 --> 01:17:44,997
information and computation. So, in your

1125
01:17:45,080 --> 01:17:47,757
premise of your question, you described

1126
01:17:47,840 --> 01:17:50,077
the algorithm as as binary.

1127
01:17:50,160 --> 01:17:54,957
And so, if you think about Kissinger
in 1971, when he came up with strategic

1128
01:17:55,040 --> 01:17:59,317
ambiguity, do you think that a computer

1129
01:17:59,400 --> 01:18:04,357
in one or two years will be able
to invent strategic diplomacy ambiguity,

1130
01:18:04,440 --> 01:18:08,077
given that it's already been done
before? I think the answer is yes.

1131
01:18:08,160 --> 01:18:10,477
So, I think that that um the algorithms

1132
01:18:10,560 --> 01:18:15,397
are getting so intelligent that as long
as it's a concept that's been around in

1133
01:18:15,480 --> 01:18:19,717
the past, it can probably figure out
a way to apply it. So, I don't agree with

1134
01:18:19,800 --> 01:18:23,677
you that it's as binary as your
question thing. And your question was

1135
01:18:23,760 --> 01:18:26,077
>> Uh superintelligence.
Superintelligence.

1136
01:18:26,160 --> 01:18:28,560
Um

1137
01:18:30,840 --> 01:18:33,837
I I think it's the same answers, right?

1138
01:18:33,920 --> 01:18:38,477
It Superintelligence in one form
or another is going to apply.

1139
01:18:38,560 --> 01:18:41,717
My own view, which is maybe
different from the consensus,

1140
01:18:41,800 --> 01:18:44,837
is that we're going to develop brilliant

1141
01:18:44,920 --> 01:18:47,837
AI physicists, brilliant AI biologists,

1142
01:18:47,920 --> 01:18:49,797
brilliant AI chemists, brilliant AI

1143
01:18:49,880 --> 01:18:52,677
writers, brilliant AI historians.

1144
01:18:52,760 --> 01:18:56,277
Um but that that the concept and and

1145
01:18:56,360 --> 01:18:59,037
there'll be systems that can drive them.

1146
01:18:59,120 --> 01:19:02,717
It's not at all obvious that if you were

1147
01:19:02,800 --> 01:19:05,037
Einstein in 1902 with that amount of

1148
01:19:05,120 --> 01:19:07,397
math available to you, that you would

1149
01:19:07,480 --> 01:19:10,837
have had the brilliance to invent um

1150
01:19:10,920 --> 01:19:12,917
special relativity

1151
01:19:13,000 --> 01:19:15,157
using the algorithms of today.

1152
01:19:15,240 --> 01:19:18,437
In the industry, there is a view that

1153
01:19:18,520 --> 01:19:20,277
that's the next really hard problem.

1154
01:19:20,360 --> 01:19:24,317
There are various theories about it.
Uh one answer is you could just do uh,

1155
01:19:24,400 --> 01:19:27,477
repetition. You could just keep asking
questions, you know, the monkey at the

1156
01:19:27,560 --> 01:19:31,277
keyboard kind of thing, and eventually
you do it. Another way is that you could

1157
01:19:31,360 --> 01:19:34,917
give the optimization function to be
curiosity, and if you just wait long

1158
01:19:35,000 --> 01:19:38,597
enough, it'll discover special relativity.
But that's not how Einstein

1159
01:19:38,680 --> 01:19:40,437
did it. Einstein sat with his little pen

1160
01:19:40,520 --> 01:19:43,077
and paper at the age of 18 or 17 with a

1161
01:19:43,160 --> 01:19:46,357
lamp and figured it out. We're not there

1162
01:19:46,440 --> 01:19:48,837
yet. My own view is that's going to be a

1163
01:19:48,920 --> 01:19:52,877
very hard boundary. So, in other words,

1164
01:19:52,960 --> 01:19:56,757
we're going to get to superhuman behavior,
which is not the same thing as

1165
01:19:56,840 --> 01:19:59,957
superintelligence, which I
would say Einstein and Kissinger

1166
01:20:00,040 --> 01:20:03,837
and so forth really were. Graham,
it's such a privilege to be here with

1167
01:20:03,920 --> 01:20:07,557
>> It's such a pleasure to have you here,
and we look forward to having you back.

1168
01:20:07,640 --> 01:20:11,200
Let's say thank you very much.
Thank you all.