1
00:00:00,160 --> 00:00:04,397
I'm just looking at you. I feel like
the the chaos of India is getting to you.

2
00:00:04,480 --> 00:00:05,997
It's day two.

3
00:00:06,080 --> 00:00:08,477
>> It's been a busy two days.

4
00:00:08,560 --> 00:00:12,717
>> Very warm welcome to Express Ada. Um

5
00:00:12,800 --> 00:00:16,797
guys, a very special day for us.
Uh Forbes magazine calls OpenAI one of the

6
00:00:16,880 --> 00:00:18,877
most important companies in the world.

7
00:00:18,960 --> 00:00:22,637
And the man who is if we look back uh at

8
00:00:22,720 --> 00:00:26,157
today 20 years from now we'll probably

9
00:00:26,240 --> 00:00:32,920
uh it I think I'll just wait
for the audience to settle down a bit.

10
00:00:34,320 --> 00:00:36,077
Okay.

11
00:00:36,160 --> 00:00:38,237
Okay. Are we settled? Yeah.

12
00:00:38,320 --> 00:00:43,597
If we look back uh at today, 20 years
from now, I don't think it'll be an

13
00:00:43,680 --> 00:00:47,357
overstatement to say that the gentleman
in front of me today is probably that

14
00:00:47,440 --> 00:00:50,957
one person who has made a tectonic shift.
It's anything close to the

15
00:00:51,040 --> 00:00:55,917
industrial revolution uh probably is the
the gentleman in front of me right now

16
00:00:56,000 --> 00:00:59,197
who it all probably started with. Is that
is that a fair statement to make, Sam?

17
00:00:59,280 --> 00:01:04,077
>> No. Um you know,
I think this is generally true

18
00:01:04,160 --> 00:01:07,437
for many companies. the person running
the company gets way too much credit

19
00:01:07,520 --> 00:01:12,077
relative to the work everybody else does.
But in our case in particular,

20
00:01:12,160 --> 00:01:15,437
this has really been a story of a
scientific discovery. There have been a

21
00:01:15,520 --> 00:01:18,477
handful of researchers that did a

22
00:01:18,560 --> 00:01:21,677
miracle and figured out uh a very deep

23
00:01:21,760 --> 00:01:25,837
thing about the way the world works. And

24
00:01:25,920 --> 00:01:27,837
the thing that is so special about deep

25
00:01:27,920 --> 00:01:30,797
learning um is what has led it to be so

26
00:01:30,880 --> 00:01:34,717
general which is that we figured out we

27
00:01:34,800 --> 00:01:36,477
this small set of researchers figured

28
00:01:36,560 --> 00:01:39,117
out this algorithm that could learn

29
00:01:39,200 --> 00:01:43,517
anything and at scale it gets better and

30
00:01:43,600 --> 00:01:45,517
better and then a whole set of people

31
00:01:45,600 --> 00:01:50,637
have worked very hard to deliver that
scale on these models and then the whole

32
00:01:50,720 --> 00:01:54,477
world has figured out how to build
incredible products and services around

33
00:01:54,560 --> 00:01:57,757
Um but more than anything
this is a story of the scientists.

34
00:01:57,840 --> 00:02:02,717
>> Yeah. And when you've uh the last time
you were in India to now lot has changed

35
00:02:02,800 --> 00:02:04,717
in the world but what do you think has

36
00:02:04,800 --> 00:02:08,637
changed in India with respect to AI

37
00:02:08,720 --> 00:02:13,597
>> is just a little over a year ago that I
was here. Uh and before talking about

38
00:02:13,680 --> 00:02:16,637
what's changed in India I think it's
worth talking about what's changed in AI

39
00:02:16,720 --> 00:02:19,037
because this will ground this. Um a year

40
00:02:19,120 --> 00:02:22,557
ago AI was just able to do high school

41
00:02:22,640 --> 00:02:24,157
math.

42
00:02:24,240 --> 00:02:27,917
And this was incredible. Uh AI could do

43
00:02:28,000 --> 00:02:31,677
a you know a very good not an excellent
but a very good job at high school

44
00:02:31,760 --> 00:02:36,957
mathematics and people couldn't believe
it and I was in India. I was speaking an

45
00:02:37,040 --> 00:02:41,437
event like this and I was talking about
it and people were it had just happened.

46
00:02:41,520 --> 00:02:43,277
People genuinely felt wonder about it.

47
00:02:43,360 --> 00:02:47,277
They're like you know this can do what
an 11th grader can do. This is amazing

48
00:02:47,360 --> 00:02:50,957
cuz only a couple years before that you
know AI couldn't really [clears throat]

49
00:02:51,040 --> 00:02:54,317
like do any math at all. I couldn't do
grade school math. I was struggling with

50
00:02:54,400 --> 00:02:55,997
that.

51
00:02:56,080 --> 00:02:59,917
By last summer, uh it was

52
00:03:00,000 --> 00:03:01,677
competing at the hardest mathematics

53
00:03:01,760 --> 00:03:04,237
competitions we had in the in the world

54
00:03:04,320 --> 00:03:07,677
and doing okay. And last week uh there

55
00:03:07,760 --> 00:03:09,917
was a new thing called first proof where

56
00:03:10,000 --> 00:03:15,117
mathematicians put out 10 research
problems uh that were the solutions to

57
00:03:15,200 --> 00:03:18,397
which were not known and took the our best
our best mathematicians the cutting

58
00:03:18,480 --> 00:03:21,677
edge of research you know new knowledge
to figure out because it hadn't yet been

59
00:03:21,760 --> 00:03:25,357
shared and I believe our latest AI got

60
00:03:25,440 --> 00:03:28,077
seven of those problems right so AI has

61
00:03:28,160 --> 00:03:33,517
gone from doing okay at high school math
to being able to do new research level

62
00:03:33,600 --> 00:03:37,357
mathematics figure out new knowledge
it's also happening in physics. And this

63
00:03:37,440 --> 00:03:42,157
is an amazing change in a year. Um
we've gone from AI that could do what we

64
00:03:42,240 --> 00:03:43,837
expected a very smart high school

65
00:03:43,920 --> 00:03:45,757
student to do to pushing the edge of

66
00:03:45,840 --> 00:03:49,277
human knowledge. Um the other so that's

67
00:03:49,360 --> 00:03:51,357
change number one. Uh the other change

68
00:03:51,440 --> 00:03:56,637
is that AI has gone from being able
to write a little bit of programming to

69
00:03:56,720 --> 00:04:00,797
completely changing what it means to be
a computer programmer. Uh last time I

70
00:04:00,880 --> 00:04:04,717
was here, people were amazed at these
sort of autocomplete tools for code. And

71
00:04:04,800 --> 00:04:08,717
now if you type in an idea one
year later if you type in an idea in

72
00:04:08,800 --> 00:04:10,557
something like codeex you can have an

73
00:04:10,640 --> 00:04:14,557
entire application created. Um it the

74
00:04:14,640 --> 00:04:18,797
job of a programmer is now I think
in this last year changed more than it's

75
00:04:18,880 --> 00:04:20,877
changed in any year that I've been an

76
00:04:20,960 --> 00:04:25,917
adult. Um codeex is our f
uh India is our fastest

77
00:04:26,000 --> 00:04:29,117
growing market for codec. So especially
what's happening here to get to the next

78
00:04:29,200 --> 00:04:30,717
part of the question is really amazing.

79
00:04:30,800 --> 00:04:34,317
Um the biggest change
that I feel from last

80
00:04:34,400 --> 00:04:37,357
time I was here is when I was here last
time it felt like India was a consumer

81
00:04:37,440 --> 00:04:40,077
of AI. People were using these services,

82
00:04:40,160 --> 00:04:43,197
people were doing other stuff and now uh

83
00:04:43,280 --> 00:04:48,797
the builder energy in India is off the
charts >> and and and and you in fact said

84
00:04:48,880 --> 00:04:51,677
something a little controversial last
time. Now it's controversial because at

85
00:04:51,760 --> 00:04:54,717
that time you said you know you had
painted a picture of despair that you

86
00:04:54,800 --> 00:04:56,557
know a $10 million fund would not be

87
00:04:56,640 --> 00:04:59,997
enough to uh you know to build an LLM a

88
00:05:00,080 --> 00:05:04,157
meaningful LLM uh out of India and now
yesterday we're talking about a full

89
00:05:04,240 --> 00:05:07,917
stack coming out of India
and India can actually lead an AI so

90
00:05:08,000 --> 00:05:10,077
>> uh that comment I mean that got taken

91
00:05:10,160 --> 00:05:11,517
out of context that was a few years ago

92
00:05:11,600 --> 00:05:16,077
but uh the the comment was not that

93
00:05:16,160 --> 00:05:20,717
India couldn't do an amazing job with
a $10 million company in a $10 million

94
00:05:20,800 --> 00:05:22,637
model. Uh I think that clearly is the

95
00:05:22,720 --> 00:05:28,077
case. It was to Could you make a very
frontier model for $10 million? Um I

96
00:05:28,160 --> 00:05:31,517
think you couldn't then. I think it's
even more true now that you can't. Um

97
00:05:31,600 --> 00:05:33,517
>> it's even it's even
more true now that you can't

98
00:05:33,600 --> 00:05:35,037
>> make a frontier model for 10 million.

99
00:05:35,120 --> 00:05:38,557
Yeah, that's >> they've
gotten very expensive. Um

100
00:05:38,640 --> 00:05:41,837
>> but uh if India
want with enough funding

101
00:05:41,920 --> 00:05:43,677
for sure an Indian company could do that

102
00:05:43,760 --> 00:05:48,957
and the smaller narrow models that India
is doing are incredible. So is that is

103
00:05:49,040 --> 00:05:51,197
that surprising to you? Is there is there
anything that's actually surprised

104
00:05:51,280 --> 00:05:55,197
you between your last trip and this trip
about India and AI that you didn't

105
00:05:55,280 --> 00:05:59,757
expect in terms of >> um that
stat that I mentioned a couple

106
00:05:59,840 --> 00:06:03,437
of minutes ago that this is the fastest
growing market for codecs in the world.

107
00:06:03,520 --> 00:06:07,997
Uh I I'm very happy I guess I shouldn't be

108
00:06:08,080 --> 00:06:11,677
surprised but it's it's great to see.
I was at IIT Delhi this morning and the

109
00:06:11,760 --> 00:06:15,757
energy from people building this new
generation of startups um that is a

110
00:06:15,840 --> 00:06:19,277
place where I expect India
to surprise the world on the upside

111
00:06:19,360 --> 00:06:23,037
>> and and the obviously the other side
of that coin is the disruption in the jobs

112
00:06:23,120 --> 00:06:25,197
especially in the IT IT enterprise

113
00:06:25,280 --> 00:06:28,237
service sector about 8% of our GDP comes

114
00:06:28,320 --> 00:06:31,437
from that one space um and a lot has

115
00:06:31,520 --> 00:06:36,317
been said about this in the last 10 days
uh because of Codex before Codex there

116
00:06:36,400 --> 00:06:40,157
was another claude and it just showed
how fast software has kind of just moved

117
00:06:40,240 --> 00:06:42,797
out. I know a lot has been said about
in the last week, but I still want you to

118
00:06:42,880 --> 00:06:46,717
reflect on that. How much of that is
a threat for for India and how should

119
00:06:46,800 --> 00:06:50,477
India think about that? >> I do
think that is going to change a lot

120
00:06:50,560 --> 00:06:51,837
and there's going to be a big impact

121
00:06:51,920 --> 00:06:55,677
there. Uh and I think it is never

122
00:06:55,760 --> 00:06:58,717
helpful to pretend otherwise. However,

123
00:06:58,800 --> 00:07:03,677
the what I expect to happen um as has

124
00:07:03,760 --> 00:07:07,597
happened with every other step forward
in computer programming is people will

125
00:07:07,680 --> 00:07:11,357
operate at a higher level of abstraction.
They will be able to do

126
00:07:11,440 --> 00:07:13,277
more people. There will still be a need

127
00:07:13,360 --> 00:07:15,277
for

128
00:07:15,360 --> 00:07:19,677
it broadly. Um the amount of product and

129
00:07:19,760 --> 00:07:23,997
code that's produced will increase a lot.
The expectations will go up. But I

130
00:07:24,080 --> 00:07:28,957
think as long as the country and the
companies that do this adapt quickly to

131
00:07:29,040 --> 00:07:33,997
this new world, there will be plenty
of new things to do. I'm I'm not a jobs

132
00:07:34,080 --> 00:07:37,037
doomer in the sense that I don't think
there will be jobs for people in the

133
00:07:37,120 --> 00:07:41,357
future. I think there will be a lot of
jobs. Every technological revolution has

134
00:07:41,440 --> 00:07:45,997
panicked about jobs going away and every
technological revolution has found new

135
00:07:46,080 --> 00:07:50,157
jobs on the other side. The promised
leisure has never happened. you know,

136
00:07:50,240 --> 00:07:53,677
this idea that we all get to go
on permanent vacation, that's never quite

137
00:07:53,760 --> 00:07:55,837
worked out. I don't think it will. Um, I

138
00:07:55,920 --> 00:07:59,037
hope we get more in that direction, but

139
00:07:59,120 --> 00:08:01,997
our desire to be useful to each other,

140
00:08:02,080 --> 00:08:06,077
our desire for more and better stuff,
our desire to express our creativity, we

141
00:08:06,160 --> 00:08:09,517
never kind of reach the end of that.
I think there is no natural limit to that.

142
00:08:09,600 --> 00:08:12,477
There's a whole big universe out there
to go explore if we can figure out, you

143
00:08:12,560 --> 00:08:17,727
know, space travel or whatever. I don't
know where it's going to go. Um but

144
00:08:17,810 --> 00:08:20,877
[snorts] but I do think
clearly things are going

145
00:08:20,960 --> 00:08:25,677
to evolve fast and you have to be on
the right side of that and you know that

146
00:08:25,760 --> 00:08:29,517
that that statement about jobs is a very
lotus statement in India because you

147
00:08:29,600 --> 00:08:32,077
know 500 million people under the age of

148
00:08:32,160 --> 00:08:35,357
30. Uh

149
00:08:35,440 --> 00:08:39,517
you understand the gravity of that uh
of of jobs and that and that that that

150
00:08:39,600 --> 00:08:42,877
anxiety that that India has. Do you
discuss this when you when you meet the

151
00:08:42,960 --> 00:08:46,077
prime minister, when you meet people
who matter in in policy? How much of that

152
00:08:46,160 --> 00:08:50,477
comes up and what are the what are
the answers you have to that anxiety?

153
00:08:50,560 --> 00:08:54,077
>> Yeah, the the main thing that comes
up with polit the main things that come up

154
00:08:54,160 --> 00:08:58,797
with political leaders are infrastructure,

155
00:08:58,880 --> 00:09:00,717
jobs,

156
00:09:00,800 --> 00:09:02,797
sort of fair distribution of benefits

157
00:09:02,880 --> 00:09:07,357
and safety broadly speaking. Um and

158
00:09:07,440 --> 00:09:11,517
those are the proportion of those
is different in different countries and

159
00:09:11,600 --> 00:09:15,357
also the the level of focus on the
upsides versus the potential downsides

160
00:09:15,440 --> 00:09:17,517
is different in different countries.

161
00:09:17,600 --> 00:09:19,757
>> Um

162
00:09:19,840 --> 00:09:25,437
>> we talk about jobs impact a lot. Uh we

163
00:09:25,520 --> 00:09:28,157
one version of this that comes up in it

164
00:09:28,240 --> 00:09:33,117
feels like almost every conversation
I have socially is what should my kids

165
00:09:33,200 --> 00:09:36,877
study for the future? what should make
it what what's it going to look like?

166
00:09:36,960 --> 00:09:40,637
Like this is this is a version of asking
that question, you know, where where is

167
00:09:40,720 --> 00:09:44,877
the economy going to go? >> Um,

168
00:09:44,960 --> 00:09:48,557
it's really hard to answer that
specifically if you study history. One

169
00:09:48,640 --> 00:09:51,037
one thing I love to do is to read about

170
00:09:51,120 --> 00:09:53,117
the history of technology and if you

171
00:09:53,200 --> 00:09:55,757
look at say the primary source material

172
00:09:55,840 --> 00:10:00,477
of people that were experiencing
the industrial revolution, there was a lot

173
00:10:00,560 --> 00:10:04,077
of panic about jobs. There were a lot
of predictions about what the jobs were

174
00:10:04,160 --> 00:10:06,877
going to be. A lot of fear there
was going to be no jobs and a lot of

175
00:10:06,960 --> 00:10:10,557
prediction about what the new jobs would
be. They were sort of shockingly wrong.

176
00:10:10,640 --> 00:10:15,117
And none of them were like, I'm going
to be the CEO of an AI company. Um

177
00:10:15,200 --> 00:10:19,197
certainly none of them were like, I'm
going to be a YouTube influencer. Uh so

178
00:10:19,280 --> 00:10:23,757
I I think this is very hard to predict,
but the skills that will work no matter

179
00:10:23,840 --> 00:10:28,157
what. >> Yeah.
fluency with AI tools, resilience,

180
00:10:28,240 --> 00:10:32,637
adaptability, figuring out what you know
people are going to want and how to be

181
00:10:32,720 --> 00:10:34,717
useful to them. Uh how to work with

182
00:10:34,800 --> 00:10:36,557
other people. These these are all very

183
00:10:36,640 --> 00:10:40,797
good things to learn. And uh

184
00:10:40,880 --> 00:10:42,637
yeah, I think it's useless to pretend

185
00:10:42,720 --> 00:10:44,797
there won't be a big change. The change

186
00:10:44,880 --> 00:10:48,077
won't be as fast as some people in the

187
00:10:48,160 --> 00:10:50,317
AI industry predict because society's

188
00:10:50,400 --> 00:10:51,997
always got more inertia and it always

189
00:10:52,080 --> 00:10:55,117
takes longer. But eventually the change

190
00:10:55,200 --> 00:10:59,757
will be huge and we'll find
all sorts of new things to do.

191
00:10:59,840 --> 00:11:02,557
I want to come back to that a little
later about where that resistance you

192
00:11:02,640 --> 00:11:05,357
think will come from that inertia
of society does it come more from the

193
00:11:05,440 --> 00:11:09,197
developed economies or from the from
the underdeveloped economies but I before

194
00:11:09,280 --> 00:11:11,597
that a little bit more about India and

195
00:11:11,680 --> 00:11:14,317
AI that five layer cake that Jensen

196
00:11:14,400 --> 00:11:17,837
spoke about the whole AI stack um the

197
00:11:17,920 --> 00:11:21,597
energy the data centers the chips the
models and the applications that entire

198
00:11:21,680 --> 00:11:26,397
five layer cake which areas do you think
India has the ability and the right to

199
00:11:26,480 --> 00:11:28,557
win in and which areas do you think it's

200
00:11:28,640 --> 00:11:31,037
just a very risky bet for India to to

201
00:11:31,120 --> 00:11:33,837
kind of make

202
00:11:33,920 --> 00:11:39,197
>> I I think India should you know
whether you subscribe to it's a

203
00:11:39,280 --> 00:11:41,837
five layer cake or a seven layer
cake people define it a little bit

204
00:11:41,920 --> 00:11:44,317
differently um but I think India should

205
00:11:44,400 --> 00:11:47,037
play at all of those levels uh I think

206
00:11:47,120 --> 00:11:51,437
that there are

207
00:11:51,520 --> 00:11:56,477
there are important advantages
to vertical integration and being able to

208
00:11:56,560 --> 00:11:59,917
work across all those levels for an
economy me of the size and importance of

209
00:12:00,000 --> 00:12:04,957
India's is a good thing to do and uh
yeah I talked with the prime minister

210
00:12:05,040 --> 00:12:09,117
just before coming here he is certainly
motivated to uh to play at all of those

211
00:12:09,200 --> 00:12:12,477
levels and I think it's important >>
and if we do play at all those levels

212
00:12:12,560 --> 00:12:16,557
and you know you're talking about more
than 70% of the population already on

213
00:12:16,640 --> 00:12:19,357
the internet I don't know what
percentage of that you you estimate is

214
00:12:19,440 --> 00:12:23,437
on is doing artificial intelligence
but assuming we want a lot of that to move

215
00:12:23,520 --> 00:12:27,117
into the AI domain and that's a big
change last time you were here a lot of

216
00:12:27,200 --> 00:12:30,477
fear and anxiety about AI Now we're all
about let's embrace it and let's kind of

217
00:12:30,560 --> 00:12:33,837
run with it. Um does the world have the

218
00:12:33,920 --> 00:12:36,397
computing power it needs for India to

219
00:12:36,480 --> 00:12:39,357
become an AI first society?

220
00:12:39,440 --> 00:12:43,437
>> Not yet. But the world is rapidly going
to have to work together to solve that

221
00:12:43,520 --> 00:12:46,237
problem. Um a question I love to ask

222
00:12:46,320 --> 00:12:49,437
people is how many GPUs would you like

223
00:12:49,520 --> 00:12:53,677
working for you all the time? How many
GPUs would you personally like thinking

224
00:12:53,760 --> 00:12:57,277
about your problems, helping you
with your work, coming, you know,

225
00:12:57,360 --> 00:13:00,877
implementing your ideas, writing
the software you want, operating the, you

226
00:13:00,960 --> 00:13:03,837
know, humanoid robots you'll have in
your home and garden and everything else

227
00:13:03,920 --> 00:13:07,277
and, you know, building new houses for
your whatever. Like, how how many do you

228
00:13:07,360 --> 00:13:10,637
want? And people give different answers.

229
00:13:10,720 --> 00:13:13,517
No one ever says less than one. Some

230
00:13:13,600 --> 00:13:15,197
people say a thousand. I would like a

231
00:13:15,280 --> 00:13:17,917
thousand. Um,

232
00:13:18,000 --> 00:13:20,637
if you go multiply that out and say

233
00:13:20,720 --> 00:13:23,997
there's 8 billion people in the world,

234
00:13:24,080 --> 00:13:28,157
we have no way to deliver 8 trillion
GPUs anytime soon. I mean, that just

235
00:13:28,240 --> 00:13:30,957
seems ridiculous. Yeah, >> we're
not going to do that to be clear.

236
00:13:31,040 --> 00:13:34,957
We're not going to do
that on Earth at least. Um,

237
00:13:35,040 --> 00:13:39,677
but I do think it shows the as it points
to the level of ambition we need to have

238
00:13:39,760 --> 00:13:41,357
about how much compute capacity we're

239
00:13:41,440 --> 00:13:44,397
going to build out. And this will be the

240
00:13:44,480 --> 00:13:47,757
most expensive, complex

241
00:13:47,840 --> 00:13:51,597
infrastructure project the world has
ever collectively taken on for all of

242
00:13:51,680 --> 00:13:54,237
our welfare. Um, so we're going to need

243
00:13:54,320 --> 00:13:59,117
a lot of compute capacity. The good
news, it would be impossible to build

244
00:13:59,200 --> 00:14:03,037
this out the oldfashioned way, but we'll
have AI and robots helping us do this

245
00:14:03,120 --> 00:14:05,837
and you know, we'll figure it out.

246
00:14:05,920 --> 00:14:09,757
>> And and is that why space keeps coming
up nowadays? There's another space race,

247
00:14:09,840 --> 00:14:11,997
but >> that I keep hearing about.

248
00:14:12,080 --> 00:14:14,877
>> I honestly think the idea with the

249
00:14:14,960 --> 00:14:18,157
current landscape of putting data

250
00:14:18,240 --> 00:14:21,357
centers in space is ridiculous.

251
00:14:21,440 --> 00:14:25,197
It will make sense someday. [applause]

252
00:14:25,280 --> 00:14:27,277
But if if you just do like the very

253
00:14:27,360 --> 00:14:31,277
rough math of launch costs relative to

254
00:14:31,360 --> 00:14:35,357
the cost of power we can do on Earth to
say nothing of how you're going to fix a

255
00:14:35,440 --> 00:14:37,197
broken GPU in space and they do break a

256
00:14:37,280 --> 00:14:40,157
lot still unfortunately. Um we are not

257
00:14:40,240 --> 00:14:43,277
there yet. There will come a time space

258
00:14:43,360 --> 00:14:47,277
is great for a lot of things. Orbital
data centers are not something that's

259
00:14:47,360 --> 00:14:49,677
going to matter at scale this decade.

260
00:14:49,760 --> 00:14:53,117
But then all this infrastructure
buildout Sam it feels like it kind of

261
00:14:53,200 --> 00:14:57,517
forces you organizations like yours
to have a very intimate relationship with

262
00:14:57,600 --> 00:15:01,357
government and is that something that's
kind of changing with AI this whole

263
00:15:01,440 --> 00:15:04,237
power structure that you know now the at

264
00:15:04,320 --> 00:15:09,197
one point big tech was a very a very
privatized force uh with a very

265
00:15:09,280 --> 00:15:12,157
complicated relationship with government
now it seems to be that you need to have

266
00:15:12,240 --> 00:15:14,317
a good relationship with government
whether it's Indian companies with

267
00:15:14,400 --> 00:15:18,157
Indian government or companies like
yours with the with the US government is

268
00:15:18,240 --> 00:15:21,517
the government a big enabler Yeah,
I think it's going to be really

269
00:15:21,600 --> 00:15:24,717
important not not just for building out
infrastructure, but just given the the

270
00:15:24,800 --> 00:15:28,477
level of impact this is going to have
on society and the need to truly

271
00:15:28,560 --> 00:15:31,997
democratize this technology. Governments
are going to have to be involved and

272
00:15:32,080 --> 00:15:34,317
companies like ours are going
to have to partner with governments.

273
00:15:34,400 --> 00:15:37,997
>> And how was that? I mean, have you ever
thought of paused and thought about that

274
00:15:38,080 --> 00:15:42,237
evolution of that relationship between
big tech and uh and government? By the

275
00:15:42,320 --> 00:15:43,917
way, can I call you big tech or not?

276
00:15:44,000 --> 00:15:48,877
Really? I don't think most
I don't feel like big tech.

277
00:15:48,960 --> 00:15:51,117
>> Okay. >> I don't think we're there yet.

278
00:15:51,200 --> 00:15:53,357
>> Okay. All right.
But tech companies like

279
00:15:53,440 --> 00:15:58,317
yours and government, I mean, >> um,

280
00:15:58,400 --> 00:16:00,317
I I I

281
00:16:00,400 --> 00:16:06,120
think the tech industry
started out as this extremely

282
00:16:06,880 --> 00:16:09,997
libertarian, you know, we don't need the
government, the government doesn't need

283
00:16:10,080 --> 00:16:13,197
us sort of view. That has changed a lot.

284
00:16:13,280 --> 00:16:17,357
Even before AI, like the last couple
of decades as the companies got bigger and

285
00:16:17,440 --> 00:16:19,837
more central to the economy and the way

286
00:16:19,920 --> 00:16:21,997
the whole world works, that's changed

287
00:16:22,080 --> 00:16:25,037
significantly. Um,

288
00:16:25,120 --> 00:16:29,517
but maybe never before has it been
this important just given the scale of the

289
00:16:29,600 --> 00:16:31,357
infrastructure that needs to happen.

290
00:16:31,440 --> 00:16:34,797
>> And how do you feel about that? Because
we, you know, one of the theories, one

291
00:16:34,880 --> 00:16:38,637
of the things we keep hearing
is that this White House is very close to

292
00:16:38,720 --> 00:16:42,717
Silicon Valley. Silicon Valley was very
close sponsor of JD Vance. That whole

293
00:16:42,800 --> 00:16:45,277
relationship is a very intimate one. Has

294
00:16:45,360 --> 00:16:47,677
is that is that true? Is that good? Is

295
00:16:47,760 --> 00:16:50,957
that bad? I would say close in some ways

296
00:16:51,040 --> 00:16:55,037
and not not close in others. You know,
there there are some tight ties and then

297
00:16:55,120 --> 00:16:57,037
also this administration has had some

298
00:16:57,120 --> 00:17:01,517
big criticisms of tech. Uh and

299
00:17:01,600 --> 00:17:05,277
is that good or bad? Uh I think a close

300
00:17:05,360 --> 00:17:09,357
cooperation between tech companies
and the government is going to become

301
00:17:09,440 --> 00:17:13,517
increasingly important over time. Um
it obviously won't be a perfectly smooth

302
00:17:13,600 --> 00:17:17,757
relationship but the better it can
be I think the better for all of us.

303
00:17:17,840 --> 00:17:20,317
>> And and what about
government-to-government relationships?

304
00:17:20,400 --> 00:17:22,077
You know what what got announced today

305
00:17:22,160 --> 00:17:25,837
was um very interesting partnership of

306
00:17:25,920 --> 00:17:27,597
several countries who are thinking about

307
00:17:27,680 --> 00:17:31,597
AI in a very advanced way. Is it the all

308
00:17:31,680 --> 00:17:33,357
these countries versus China? Is that

309
00:17:33,440 --> 00:17:37,197
what's kind of playing out or is it are

310
00:17:37,280 --> 00:17:40,717
we all kind of in a race and somewhere
China is also in that race?

311
00:17:40,800 --> 00:17:42,637
>> I suspect it'll turn out to be way more

312
00:17:42,720 --> 00:17:44,877
complex than that. I think it will shift

313
00:17:44,960 --> 00:17:49,757
over time. Uh there

314
00:17:49,840 --> 00:17:54,157
I do sort of think that AI will become

315
00:17:54,240 --> 00:17:57,197
one of the most important political
issues in the world. one of the highest

316
00:17:57,280 --> 00:18:00,237
order bits of geopolitical

317
00:18:00,320 --> 00:18:02,957
tension and cooperation.

318
00:18:03,040 --> 00:18:05,757
But I I don't think it'll be a fixed

319
00:18:05,840 --> 00:18:10,237
thing. I think as it develops and as I
mean this has always happened that sort

320
00:18:10,320 --> 00:18:12,317
of political alliances shift over time.

321
00:18:12,400 --> 00:18:16,797
I expect them to shift more in the
future uh not less as things move

322
00:18:16,880 --> 00:18:21,757
faster. >> But what is how do
you think about China in all of

323
00:18:21,840 --> 00:18:25,917
this? I mean, you know, we keep hearing
that China's leap years ahead and

324
00:18:26,000 --> 00:18:29,037
China's got all this energy and China's
got all this renewable energy and

325
00:18:29,120 --> 00:18:33,437
China's got all this robotics and all
this technology. Uh, are we hearing the

326
00:18:33,520 --> 00:18:36,797
right thing and why are they
leapers ahead? What have they done?

327
00:18:36,880 --> 00:18:40,717
>> I would say very ahead in some areas
and not ahead in others. I I think this is

328
00:18:40,800 --> 00:18:43,907
like I mean that sounds like a boring
statement that's probably always

329
00:18:43,990 --> 00:18:45,837
[laughter] >> always true
in the world like in terms

330
00:18:45,920 --> 00:18:49,517
of manufacturing physical
robots clearly ahead

331
00:18:49,600 --> 00:18:51,597
>> a big edge on things like electric

332
00:18:51,680 --> 00:18:55,837
motors, magnets, stuff like that. um

333
00:18:55,920 --> 00:19:01,357
clearly ahead on energy buildout.
Uh so there are places where China is

334
00:19:01,440 --> 00:19:04,077
significantly ahead of other countries
and then there's places where I think

335
00:19:04,160 --> 00:19:07,357
we're ahead of them and my guess is that

336
00:19:07,440 --> 00:19:11,277
that's sort of always what it's been
like and what it will continue to be

337
00:19:11,360 --> 00:19:14,157
like. It's hard to be ahead
on everything or behind on everything.

338
00:19:14,240 --> 00:19:16,877
Maybe if you have like the only super
intelligence in the world, you could do

339
00:19:16,960 --> 00:19:19,757
it. I think that would actually
be bad for other reasons.

340
00:19:19,840 --> 00:19:23,357
>> But is it a race, Sam? I mean it it it

341
00:19:23,440 --> 00:19:27,037
we keep framing it
as a race >> sort of. Uh

342
00:19:27,120 --> 00:19:29,757
>> is it a collaborative
journey or is it a race or

343
00:19:29,840 --> 00:19:33,357
>> Yeah, I mean both. That's like the hard
part of this, right? There's always

344
00:19:33,440 --> 00:19:37,037
>> there are always ways in which
the global economy is interdependent and

345
00:19:37,120 --> 00:19:40,637
very collaborative and then there's
always ways in which it's a competition.

346
00:19:40,720 --> 00:19:45,357
And it's so tempting
to say it's all one or

347
00:19:45,440 --> 00:19:49,517
the other. Reality is always
messier than that. Um,

348
00:19:49,600 --> 00:19:52,237
but I do think it's important to look at

349
00:19:52,320 --> 00:19:54,477
the net

350
00:19:54,560 --> 00:19:56,877
differences in power and the impact that

351
00:19:56,960 --> 00:19:59,437
could have. Um,

352
00:19:59,520 --> 00:20:01,517
and I think it's, you know, really

353
00:20:01,600 --> 00:20:05,197
important that democracy leads with AI.

354
00:20:05,280 --> 00:20:08,237
>> And when you say democracy leads with
AI, you mean China should not lead with

355
00:20:08,320 --> 00:20:11,837
AI. >> Um,

356
00:20:11,920 --> 00:20:15,837
it was a little bit more nuanced of a
comment than that, but if you want to

357
00:20:15,920 --> 00:20:19,997
reduce it to that. No, because in in in
your speech yesterday, you mentioned uh

358
00:20:20,080 --> 00:20:23,757
authoritarian, you know, a couple
of times. So, I'm just kind of, you know,

359
00:20:23,840 --> 00:20:26,237
how do we think about China?
How do we unpack China?

360
00:20:26,320 --> 00:20:27,837
>> What I meant more
is I don't think there

361
00:20:27,920 --> 00:20:31,437
should be any single super intelligence
in the world. I don't think there should

362
00:20:31,520 --> 00:20:34,557
be any one person or any one country
or anyone company in charge of super

363
00:20:34,640 --> 00:20:38,157
intelligence, including the United States.
I think that would also be bad.

364
00:20:38,240 --> 00:20:40,157
Um,

365
00:20:40,240 --> 00:20:43,277
I think the world is at its best when

366
00:20:43,360 --> 00:20:47,677
power is widely distributed, when people
have a lot of different ideas, and when

367
00:20:47,760 --> 00:20:50,077
there's enough of a balance in power

368
00:20:50,160 --> 00:20:55,117
that we can all we countries,
individuals, companies, whatever, sort

369
00:20:55,200 --> 00:20:59,357
of keep each other in check and all have
input and over time, even though we'll

370
00:20:59,440 --> 00:21:02,957
do some dumb things along the way,
you know, people kind of get to pick the

371
00:21:03,040 --> 00:21:05,997
best ideas and let those rise up. So

372
00:21:06,080 --> 00:21:08,637
what I meant by authoritarian here is

373
00:21:08,720 --> 00:21:13,837
you don't literally want one AI
in charge of the world no matter who has

374
00:21:13,920 --> 00:21:15,117
it. I think that would be bad.

375
00:21:15,200 --> 00:21:17,037
>> Yeah. And do you
think if you compare to

376
00:21:17,120 --> 00:21:20,957
the big tech era of 7 years ago uh AI is

377
00:21:21,040 --> 00:21:24,397
going to fragment power more or is it
going to concentrated more because it

378
00:21:24,480 --> 00:21:26,157
seems to be that big tech the old big

379
00:21:26,240 --> 00:21:29,037
tech is equally sort of staying uh you

380
00:21:29,120 --> 00:21:34,157
know staying is keeping pace.
That's a good that is at some point it's

381
00:21:34,240 --> 00:21:37,277
in some sense I think one of the most
important questions in front of us right

382
00:21:37,360 --> 00:21:40,317
now. You can totally imagine a world

383
00:21:40,400 --> 00:21:43,037
where AI massively concentrates power

384
00:21:43,120 --> 00:21:45,437
where say a single company or country is

385
00:21:45,520 --> 00:21:49,757
able to hold on to AI and use that to

386
00:21:49,840 --> 00:21:51,677
amass a gigantic amount of power and

387
00:21:51,760 --> 00:21:53,997
wealth in the world. You can imagine

388
00:21:54,080 --> 00:21:57,037
another extreme where everybody on earth

389
00:21:57,120 --> 00:21:58,877
has a super intelligence with no rules

390
00:21:58,960 --> 00:22:01,677
whatsoever and you know some pretty

391
00:22:01,760 --> 00:22:05,517
terrible things and chaos happen. Um

392
00:22:05,600 --> 00:22:07,597
you can imagine many worlds there in the

393
00:22:07,680 --> 00:22:10,797
middle. I personally think something way

394
00:22:10,880 --> 00:22:14,717
more towards the democratized version
of that is good. We of course will need

395
00:22:14,800 --> 00:22:18,397
some regulation. We of course will need
some guardrails. But I think putting

396
00:22:18,480 --> 00:22:20,157
this technology in the hands of lots of

397
00:22:20,240 --> 00:22:23,757
people is possible. And that will be a

398
00:22:23,840 --> 00:22:27,517
decentralization of economic power. The

399
00:22:27,600 --> 00:22:30,397
the clearest example of this you can see

400
00:22:30,480 --> 00:22:34,557
uh is in the last six and especially 3

401
00:22:34,640 --> 00:22:37,917
months the power that a one or two or

402
00:22:38,000 --> 00:22:40,557
three person company can have.

403
00:22:40,640 --> 00:22:43,037
This just didn't happen uh a few years

404
00:22:43,120 --> 00:22:47,677
ago. It was impossible.
In this new world of things like codecs

405
00:22:47,760 --> 00:22:49,517
and general purpose knowledge working

406
00:22:49,600 --> 00:22:52,077
agents um we're going to have to get

407
00:22:52,160 --> 00:22:55,357
used very quickly to one person startups

408
00:22:55,440 --> 00:22:59,997
that have enormous amounts of success
or two person three person startups and I

409
00:23:00,080 --> 00:23:03,597
think this is great. >> Yeah.
And how competitive is how

410
00:23:03,680 --> 00:23:05,997
competitive is all of this from your

411
00:23:06,080 --> 00:23:07,757
seat? I mean, you know, we saw a little

412
00:23:07,840 --> 00:23:11,997
awkward moment uh on stage yesterday. Um

413
00:23:12,080 --> 00:23:15,357
there was several memes
about it going around, but

414
00:23:15,440 --> 00:23:17,597
>> you got to give the internet
something to laugh at.

415
00:23:17,680 --> 00:23:22,397
>> Um [laughter] it is definitely
competitive, but I will

416
00:23:22,480 --> 00:23:25,837
say that I think >> it's also
very incestuous. I mean, you

417
00:23:25,920 --> 00:23:28,557
guys because all you know, a lot
of people were building out stuff or

418
00:23:28,640 --> 00:23:30,797
working with you at one point
and everybody knows everybody and

419
00:23:30,880 --> 00:23:34,077
everybody's like Nvidia's investing in
you. you're buying Nvidia chips and then

420
00:23:34,160 --> 00:23:37,917
you it's just the small >> it
is a weirdly small world for sure.

421
00:23:38,000 --> 00:23:43,117
>> Um I think it's very competitive
commercially but almost all of the

422
00:23:43,200 --> 00:23:47,517
efforts that have a very
advanced model I think

423
00:23:47,600 --> 00:23:52,237
fuel the gravity of what's happening
and they're very committed to getting

424
00:23:52,320 --> 00:23:56,877
uh safety and alignment right
and willing to cooperate there.

425
00:23:56,960 --> 00:23:59,757
>> A little bit more about what happened
yesterday if you want to share this is a

426
00:23:59,840 --> 00:24:03,997
good opportunity. I I don't
really have that much more to add.

427
00:24:04,080 --> 00:24:07,197
>> I won't push it. Okay. Uh, you know,
there's just so much to cover, Sam, and

428
00:24:07,280 --> 00:24:10,797
I learned so much. I I'm very grateful
for you because I learned so much prep

429
00:24:10,880 --> 00:24:14,957
to prep for this conversation. I
was like 3 days of just like deep into it.

430
00:24:15,040 --> 00:24:17,437
So, I thought I'll just kind
of break it up into a game. Great.

431
00:24:17,520 --> 00:24:18,797
>> And we can get,
you know, I can just try

432
00:24:18,880 --> 00:24:22,717
to cover more and more stuff rather than
having a conversation about it. Um, and

433
00:24:22,800 --> 00:24:26,877
we'll see what we kind of do. So, it's
not a rapid fire because it's just too

434
00:24:26,960 --> 00:24:28,557
long to be a rapid fire. And I tried to

435
00:24:28,640 --> 00:24:32,237
make it shorter, but I couldn't. Um,

436
00:24:32,320 --> 00:24:34,317
is there any one thing you admire about

437
00:24:34,400 --> 00:24:36,717
Google's catchup in the AI race? Because

438
00:24:36,800 --> 00:24:39,837
10 years ago they were out of the water.

439
00:24:39,920 --> 00:24:44,477
>> Well, 10 years ago, 10 years ago,
Google was the only serious AI effort.

440
00:24:44,560 --> 00:24:47,597
>> Five years ago, maybe >> um

441
00:24:47,680 --> 00:24:50,397
maybe 3 years ago when we launched

442
00:24:50,480 --> 00:24:53,277
Chachi BT, they were then way out of it.

443
00:24:53,360 --> 00:24:55,757
But the the first thing I admire about

444
00:24:55,840 --> 00:25:01,037
Google is Demis and the Google team
started working on AI before anyone else

445
00:25:01,120 --> 00:25:05,677
in the modern era with a lot
of conviction and uh you know I think

446
00:25:05,760 --> 00:25:09,837
without their inspiration we certainly
wouldn't be here. So that's one thing I

447
00:25:09,920 --> 00:25:11,677
admire and then a second thing more

448
00:25:11,760 --> 00:25:16,077
recently is their relentless focus and

449
00:25:16,160 --> 00:25:20,237
execution and ability to really scale
the model after being pretty far behind

450
00:25:20,320 --> 00:25:23,757
is quite impressive. >> Okay.
Were you surprised when Apple

451
00:25:23,840 --> 00:25:26,957
partnered with Google for Siri?

452
00:25:27,040 --> 00:25:30,957
>> Not particularly. >> Okay.
The one country that is by and

453
00:25:31,040 --> 00:25:36,760
large on the right path
for regulation on AI.

454
00:25:38,160 --> 00:25:42,157
>> I don't think any of us know the answer
to that yet. A thing that I'm happy

455
00:25:42,240 --> 00:25:46,637
about is countries are trying different
approaches and I think this is a good

456
00:25:46,720 --> 00:25:50,557
thing about many sovereign powers in the
world. We will get to observe over the

457
00:25:50,640 --> 00:25:53,277
next few years very different approaches

458
00:25:53,360 --> 00:25:56,557
and see what works and what doesn't and

459
00:25:56,640 --> 00:26:00,397
pretty quickly I believe the world will
move towards more of what works. But I'm

460
00:26:00,480 --> 00:26:03,117
grateful for the experimentation
and different countries trying different

461
00:26:03,200 --> 00:26:06,477
things. >> Okay.
So short game that I'll give you

462
00:26:06,560 --> 00:26:09,837
some criticisms of AI and you give
me a short defense of each of these

463
00:26:09,920 --> 00:26:13,117
criticisms. So the first >> I
might not have one for some. That's

464
00:26:13,200 --> 00:26:14,237
fine. And and there's so many questions

465
00:26:14,320 --> 00:26:18,717
that you can always duck or pass. um too

466
00:26:18,800 --> 00:26:22,077
much concentration of power

467
00:26:22,160 --> 00:26:26,317
>> fair unless we push super hard
to democratize and I think everyone in the

468
00:26:26,400 --> 00:26:30,797
world needs to hold countries and
companies to this threshold of you don't

469
00:26:30,880 --> 00:26:34,637
get to say concentration of power in the
name of safety we don't want that trade

470
00:26:34,720 --> 00:26:39,197
it's got to be democratized >> okay
uh the amount of natural resources

471
00:26:39,280 --> 00:26:40,637
that are going into these data centers

472
00:26:40,720 --> 00:26:43,437
the amount of water the amount of

473
00:26:43,520 --> 00:26:48,157
>> water is totally fake uh it used to be
True. We used to do evaporative cooling

474
00:26:48,240 --> 00:26:51,597
in data centers, but now that we don't
do that, you know, you see these like

475
00:26:51,680 --> 00:26:55,757
things on the internet where don't use
chatbt, it's 17 gallons of water for

476
00:26:55,840 --> 00:26:57,997
each query or whatever. This is

477
00:26:58,080 --> 00:27:00,797
completely untrue. Totally insane. No,

478
00:27:00,880 --> 00:27:03,197
no connection to reality. What is fair

479
00:27:03,280 --> 00:27:05,837
though is the energy consumption not per

480
00:27:05,920 --> 00:27:08,077
query but in total because the world is

481
00:27:08,160 --> 00:27:11,997
now using so much AI is real and we need

482
00:27:12,080 --> 00:27:15,037
to move towards nuclear or wind and

483
00:27:15,120 --> 00:27:19,517
solar very quickly. There was a you know
we had Bill Gates last year and I asked

484
00:27:19,600 --> 00:27:22,397
him this question and he gave a very
interesting statistic. At that time

485
00:27:22,480 --> 00:27:24,557
apparently we calculated that there were

486
00:27:24,640 --> 00:27:29,677
10 iPhone worth of battery would go
for every chat GPT query and now it's come

487
00:27:29,760 --> 00:27:33,437
down to one iPhone or one and a half
iPhone battery life worth of energy for

488
00:27:33,520 --> 00:27:34,957
every chat GPT query. Is that correct?

489
00:27:35,040 --> 00:27:37,517
>> There's no way it's
anything close to that much.

490
00:27:37,600 --> 00:27:39,997
>> It's much less than that.
>> Way way way less.

491
00:27:40,080 --> 00:27:42,877
>> Okay. All right. But but his theory was

492
00:27:42,960 --> 00:27:45,597
that that the AIs will learn from human

493
00:27:45,680 --> 00:27:51,320
evolution to be more efficient
and how how much energy they consume.

494
00:27:51,680 --> 00:27:56,237
>> One of the things that is always unfair
in this comparison is people talk about

495
00:27:56,320 --> 00:27:58,477
how much energy it takes to train an AI

496
00:27:58,560 --> 00:28:03,757
model relative to how much it costs
a human to do one inference query. But it

497
00:28:03,840 --> 00:28:07,677
also takes a lot of energy to train
a human. takes like 20 years of life and

498
00:28:07,760 --> 00:28:11,837
all of the food you eat during that time
before you get smart. And not only that,

499
00:28:11,920 --> 00:28:15,037
it took like the very widespread
evolution of the hundred billion people

500
00:28:15,120 --> 00:28:18,717
that have ever lived and learned not
to get eaten by predators and learned how

501
00:28:18,800 --> 00:28:21,837
to like figure out science and whatever
to produce you and then you took

502
00:28:21,920 --> 00:28:24,077
whatever you you know you took. So the

503
00:28:24,160 --> 00:28:29,117
fair comparison is if you ask chat
a question, how much energy does it take

504
00:28:29,200 --> 00:28:30,797
once its model is trained to answer that

505
00:28:30,880 --> 00:28:34,797
question versus a human? And probably AI

506
00:28:34,880 --> 00:28:37,917
has already caught up on an energy
efficiency basis. Measure it that way.

507
00:28:38,000 --> 00:28:44,600
>> Wow. Okay. Fantastic.
AI is making my kids dumber.

508
00:28:46,240 --> 00:28:51,160
>> True for some kids. Um,

509
00:28:54,240 --> 00:28:57,837
look, when I

510
00:28:57,920 --> 00:29:00,797
when I hear kids talk about AI, uh,

511
00:29:00,880 --> 00:29:05,917
there are definitely some kids who are
like, "This is great. I cheated my way

512
00:29:06,000 --> 00:29:09,917
through all of high school. I never did
any homework. Thank you." And I'm like,

513
00:29:10,000 --> 00:29:13,997
"What's your plan for like the rest
of your life?" And they're like, "Well, I

514
00:29:14,080 --> 00:29:16,077
assume I can like still use CHB to do my

515
00:29:16,160 --> 00:29:19,517
job. this is very bad. Um, we absolutely

516
00:29:19,600 --> 00:29:23,677
have to still teach our kids to to learn

517
00:29:23,760 --> 00:29:25,197
and to think and to be creative and to

518
00:29:25,280 --> 00:29:27,837
use these tools. That's not what most

519
00:29:27,920 --> 00:29:31,677
kids say, though. Most kids say, uh, I

520
00:29:31,760 --> 00:29:33,677
can't believe what I can accomplish now.

521
00:29:33,760 --> 00:29:37,277
Um, look at this thing that I've just
made. I've built these incredible new

522
00:29:37,360 --> 00:29:42,797
workflows. You know, sure, like I may
use Chat GBT and the way you use Google

523
00:29:42,880 --> 00:29:46,477
when you were in high school to help
with your homework. Um, but now that I

524
00:29:46,560 --> 00:29:50,717
have this new tool, I'm doing these
amazing things. I think we will need to

525
00:29:50,800 --> 00:29:55,037
find new ways to teach and evaluate in
school to make sure every kid is brought

526
00:29:55,120 --> 00:29:57,117
along. But the potential of this

527
00:29:57,200 --> 00:29:59,837
technology, the ability to learn more,

528
00:29:59,920 --> 00:30:02,557
do more. I have no doubt about that. Um,

529
00:30:02,640 --> 00:30:07,677
when I was in school,
Google had just come out and uh

530
00:30:07,760 --> 00:30:10,877
the like my middle school teacher
was like, "This is the worst thing ever."

531
00:30:10,960 --> 00:30:15,357
You know, there's no point to teaching
anymore. Why do you have to memorize,

532
00:30:15,440 --> 00:30:18,317
you know, the date that someone
in history was born if you could just look

533
00:30:18,400 --> 00:30:22,077
it up on Google? And my answer was,
I think it's a complete waste of time to

534
00:30:22,160 --> 00:30:25,197
memorize what year someone was born.
I will just go look that up on Google if I

535
00:30:25,280 --> 00:30:29,277
ever need to know it again, which
usually you don't that often. And then I

536
00:30:29,360 --> 00:30:32,477
watched over the next few years,
teachers come to peace with this, the

537
00:30:32,560 --> 00:30:36,077
education system evolve. And like always

538
00:30:36,160 --> 00:30:40,557
happens with new tools, the

539
00:30:40,640 --> 00:30:44,557
the potential goes up and expectations
go up. And we'll have to teach people to

540
00:30:44,640 --> 00:30:48,317
think harder, create more, and you know,
I'm pretty sure that like a kid born

541
00:30:48,400 --> 00:30:51,997
today when they're 18, graduating high
school will be able to do things that no

542
00:30:52,080 --> 00:30:55,197
one today can. And I think that's great.

543
00:30:55,280 --> 00:30:58,397
The other criticism is that it's just
not a democratic enough as in the people

544
00:30:58,480 --> 00:31:01,837
who want to pause AI don't have a voice.

545
00:31:01,920 --> 00:31:06,877
Uh and I want to take this double click
on this on this question is that where

546
00:31:06,960 --> 00:31:08,477
do you see the resistance coming from?

547
00:31:08,560 --> 00:31:12,477
Uh is it coming from the people who have
experienced technology at its highest

548
00:31:12,560 --> 00:31:16,157
levels and then they say we're nervous
we want to pause or do you see the

549
00:31:16,240 --> 00:31:18,717
resistance coming from those who have
never experienced and are nervous to

550
00:31:18,800 --> 00:31:20,877
experience and then they feel the the

551
00:31:20,960 --> 00:31:23,197
vulnerability?

552
00:31:23,280 --> 00:31:26,797
I I mean it kind of comes from parts of

553
00:31:26,880 --> 00:31:30,957
it come from everywhere. Although
as more people use the technology, I think

554
00:31:31,040 --> 00:31:32,717
there's less of a let's totally pause

555
00:31:32,800 --> 00:31:36,877
it. People like it and there's more of a

556
00:31:36,960 --> 00:31:40,637
what is this going to mean? Can we have
more of a voice? Can this go a little

557
00:31:40,720 --> 00:31:44,397
more slowly? >> Okay.
The one thing Silicon Valley

558
00:31:44,480 --> 00:31:47,037
should learn from Chinese technology.

559
00:31:47,120 --> 00:31:50,397
>> Move fast. Okay. The one thing Silicon

560
00:31:50,480 --> 00:31:53,037
Valley can learn from Indian technology.

561
00:31:53,120 --> 00:31:55,917
>> Move faster.

562
00:31:56,000 --> 00:32:02,840
The one real and one imagined
fear of Chinese dominance in AI.

563
00:32:04,640 --> 00:32:09,437
I I think the imagined fear is that,
you know, the Chinese military is going to

564
00:32:09,520 --> 00:32:13,117
send a billion humanoid robots marching
through the streets and overpower the US

565
00:32:13,200 --> 00:32:15,197
army or something like that.

566
00:32:15,280 --> 00:32:20,637
And the real fear is and and I think
that's like an example of fighting the

567
00:32:20,720 --> 00:32:24,157
last war and thinking it's going to be
the same. I think the real fear is a new

568
00:32:24,240 --> 00:32:27,437
kind of war between nations that gets
fought over the internet where you

569
00:32:27,520 --> 00:32:30,717
influence people and hack into a bunch
of critical infrastructure and that

570
00:32:30,800 --> 00:32:35,117
feels totally possible. That's
a very scary uh reality where we say

571
00:32:35,200 --> 00:32:37,117
surveillance state all those kind of

572
00:32:37,200 --> 00:32:42,600
things you don't >> super
worried about that um

573
00:32:42,800 --> 00:32:46,317
I think increasingly so people are using

574
00:32:46,400 --> 00:32:49,037
fear of AI going wrong to justify a

575
00:32:49,120 --> 00:32:54,317
surveillance state and I don't
think that people doing that

576
00:32:54,400 --> 00:32:56,877
have fully thought through
the downsides of the surveillance state.

577
00:32:56,960 --> 00:33:00,717
>> Yeah. Okay. Um, you urgently need to

578
00:33:00,800 --> 00:33:03,917
look at you you urgently need to look
look something up. Your phone is

579
00:33:04,000 --> 00:33:08,317
discharged. You borrow my phone.
The letter T on my type on my touchpad

580
00:33:08,400 --> 00:33:13,107
doesn't work. So, you can only key
in Grock, Claw, Gemini, or Deep Seek.

581
00:33:13,190 --> 00:33:15,210
[laughter]

582
00:33:16,240 --> 00:33:18,957
>> I guess if I have to look
something up, I pick Gemini.

583
00:33:19,040 --> 00:33:22,477
>> Okay. All right. Um, the for other

584
00:33:22,560 --> 00:33:23,837
things, I'd pick other models.

585
00:33:23,920 --> 00:33:27,997
>> Okay. the most compelling reason you
have moved from a not for-p profofit to

586
00:33:28,080 --> 00:33:32,157
a cap profit to pretty much
any revenue including ads.

587
00:33:32,240 --> 00:33:34,557
>> Um

588
00:33:34,640 --> 00:33:39,277
the need to democratize AI and the need
to stay at the forefront of research

589
00:33:39,360 --> 00:33:42,317
both of those require
huge amounts of capital.

590
00:33:42,400 --> 00:33:45,597
>> And do you think you're research first

591
00:33:45,680 --> 00:33:47,197
company or a product first company?

592
00:33:47,280 --> 00:33:49,597
>> Research first company. Uh the coolest

593
00:33:49,680 --> 00:33:54,237
thing about AI I think is that the great
majority of making a good product is

594
00:33:54,320 --> 00:33:57,997
doing good research. We make a great
model, the product is already is pretty

595
00:33:58,080 --> 00:34:02,317
good. The early version of chatbt was
almost no product at all. It was like a

596
00:34:02,400 --> 00:34:06,077
text box that fed into a model and then
put the output on the screen. So

597
00:34:06,160 --> 00:34:09,837
research to continue to push what
a model is capable of is what enables

598
00:34:09,920 --> 00:34:10,877
everything else.

599
00:34:10,960 --> 00:34:14,077
>> Okay. The one criticism about you that

600
00:34:14,160 --> 00:34:17,357
bothers you when you hear it?

601
00:34:17,440 --> 00:34:21,277
>> Yeah. I've heard
basically like every possible

602
00:34:21,360 --> 00:34:25,197
criticism and at some point none of them
bother you because you're just like this

603
00:34:25,280 --> 00:34:28,077
gotten so insane, so detached from truth.

604
00:34:28,160 --> 00:34:30,557
>> Um

605
00:34:30,640 --> 00:34:34,157
I guess the one that

606
00:34:34,240 --> 00:34:38,157
makes me the most sad is that like he's
just doing it for the power. He doesn't

607
00:34:38,240 --> 00:34:40,557
really care about sort of trying to do
something that's going to be helpful to

608
00:34:40,640 --> 00:34:43,277
the world. >> What what
were you thinking when you

609
00:34:43,360 --> 00:34:47,357
when you decided not to take any equity
in OpenAI? what what what was going in

610
00:34:47,440 --> 00:34:49,517
your mind at that time >> that
that was truly one of the dumbest

611
00:34:49,600 --> 00:34:51,917
thing as I say this as [laughter]
someone who's made a lot of dumb

612
00:34:52,000 --> 00:34:54,557
decisions. Um the conspiracy theories

613
00:34:54,640 --> 00:34:57,197
that that brought about like the one I

614
00:34:57,280 --> 00:35:01,837
just mentioned uh it was totally not
worth it. We were like a nonprofit. I,

615
00:35:01,920 --> 00:35:06,397
you know, there was like a kind of quirk
of it where to be on the board of a

616
00:35:06,480 --> 00:35:11,357
nonprofit, you had to like be sort
of disinterested. And I had been like very

617
00:35:11,440 --> 00:35:12,797
fortunate to have a successful career.

618
00:35:12,880 --> 00:35:15,997
And I was like, I don't really care here
either way. But it was a very dumb thing.

619
00:35:16,080 --> 00:35:19,277
>> And if somebody today were to offer and
figure out a way for you to get into the

620
00:35:19,360 --> 00:35:22,717
cap table with Fred equity,
would you take it?

621
00:35:22,800 --> 00:35:26,957
>> I mean, at this point, I feel like
so tired of the whole conversation and so

622
00:35:27,040 --> 00:35:28,317
trapped, I'm not sure what to do.

623
00:35:28,400 --> 00:35:30,237
>> Okay. It's like
a lose-lose thing for me.

624
00:35:30,320 --> 00:35:34,797
>> Okay. Um, so one thing in spite of all
the differences you have with Elon Musk,

625
00:35:34,880 --> 00:35:38,317
the one thing you the one
thing you admire about him.

626
00:35:38,400 --> 00:35:40,237
>> Is there any one thing you
can put all the difference?

627
00:35:40,320 --> 00:35:44,760
>> No, I'm going to think
of something, but give me a minute.

628
00:35:46,560 --> 00:35:50,477
>> He's he's extremely good at

629
00:35:50,560 --> 00:35:54,397
physical engineering and also extremely
good at getting people to perform

630
00:35:54,480 --> 00:35:56,477
incredibly well at their jobs.

631
00:35:56,560 --> 00:36:00,237
>> Okay. Super. Okay. Um, Chad Gubet does

632
00:36:00,320 --> 00:36:03,597
is not known for its opinions. It's
known for giving facts and and factful

633
00:36:03,680 --> 00:36:06,797
answers back. So, I'm going to ask your
opinion on the news. There've been a lot

634
00:36:06,880 --> 00:36:11,277
of stuff in the news the last week. So,
we want your views on would you support

635
00:36:11,360 --> 00:36:17,320
the governmentordered ban
on social media for kids under 16.

636
00:36:18,080 --> 00:36:19,597
Uh, I don't know if a total ban is

637
00:36:19,680 --> 00:36:22,717
right. Maybe it is. Uh, I think at least

638
00:36:22,800 --> 00:36:26,877
heavy restrictions would be good. I

639
00:36:26,960 --> 00:36:30,797
I have a young child. I'm more
interested in the what I think are the

640
00:36:30,880 --> 00:36:32,557
very negative impacts of infinite

641
00:36:32,640 --> 00:36:37,357
scroll. Um I would like my kid not to

642
00:36:37,440 --> 00:36:42,317
just be one of those iPad kids going
like this all of the time. So that seems

643
00:36:42,400 --> 00:36:46,477
very important. But yeah, probably
social media for kids is something we

644
00:36:46,560 --> 00:36:47,837
have to be very careful with.

645
00:36:47,920 --> 00:36:51,677
>> Okay, it seems the clouds of war are

646
00:36:51,760 --> 00:36:54,557
again upon us. a lot of talk about in in

647
00:36:54,640 --> 00:36:58,877
Iran. Um, if government were to ask to

648
00:36:58,960 --> 00:37:02,397
use AI companies like yours to help

649
00:37:02,480 --> 00:37:05,597
figure out how to get into war, how do

650
00:37:05,680 --> 00:37:10,237
you how would you how do you respond? Is
Pentagon just another client for you and

651
00:37:10,320 --> 00:37:11,757
you know you're the they're the

652
00:37:11,840 --> 00:37:13,437
customer, you do what they ask, or do

653
00:37:13,520 --> 00:37:18,317
you draw a line somewhere? Um,

654
00:37:18,400 --> 00:37:23,597
I don't think AI systems
should be used to make

655
00:37:23,680 --> 00:37:26,557
a AI systems should be used to make
warfighting decisions. I don't think

656
00:37:26,640 --> 00:37:30,397
they're at a level of sophistication
and reliability where this is a good idea.

657
00:37:30,480 --> 00:37:34,957
That said, like we certainly want
to support the government and uh there's a

658
00:37:35,040 --> 00:37:38,797
lot of things we can do already.
Um, someday there will be, I think,

659
00:37:38,880 --> 00:37:43,757
applications of this, really important
applications of this to defense. Um, but

660
00:37:43,840 --> 00:37:46,957
right now the models
have clear limitations.

661
00:37:47,040 --> 00:37:50,317
>> It was used in for Maduro's capture,
I believe. Is that is that true? Is that

662
00:37:50,400 --> 00:37:53,757
true? >> I No, I I just
don't know. Like I I'm

663
00:37:53,840 --> 00:37:57,917
sure it was used in some ways. There
are things that AI can do a great job of

664
00:37:58,000 --> 00:38:02,557
today. Like I think using AI to analyze
a huge amount of intelligence reports

665
00:38:02,640 --> 00:38:05,517
probably a great use of AI and maybe
it was used some way like that.

666
00:38:05,600 --> 00:38:09,917
>> Yeah. Your views on Claude safety
chief Mang Sharma's dramatic resonation. I

667
00:38:10,000 --> 00:38:12,637
don't know if you followed that but it
made a lot of news. I heard something

668
00:38:12,720 --> 00:38:16,397
about it. >> Okay.
>> Like they they resigned and like a

669
00:38:16,480 --> 00:38:19,037
letter about it's all
useless or something.

670
00:38:19,120 --> 00:38:21,757
>> Yeah. World's going to go to hell.
I want to go write poetry. It was that

671
00:38:21,840 --> 00:38:26,397
that direction. And >> I just
a a lot of safety people are are

672
00:38:26,480 --> 00:38:29,517
getting I think overwhelmed. I mean
that's the real piece that I wanted you

673
00:38:29,600 --> 00:38:32,920
to reflect on.

674
00:38:33,280 --> 00:38:35,997
The part of it I agree with is the

675
00:38:36,080 --> 00:38:40,557
inside view at the companies of looking
what's going to happen. Like the world

676
00:38:40,640 --> 00:38:44,957
is not prepared. We're going to have
extremely capable models soon. It's

677
00:38:45,040 --> 00:38:49,437
going to be a faster takeoff than I
originally thought and that is stressful

678
00:38:49,520 --> 00:38:53,757
and anxietyinducing. >> Okay.
Uh reflect a bit on the future.

679
00:38:53,840 --> 00:38:57,757
The future of love and relationships.
We keep hearing of people using AI agents

680
00:38:57,840 --> 00:39:00,637
to to open up conversations on dating

681
00:39:00,720 --> 00:39:05,757
apps. Um, how much of authenticity are
we losing with love and relationships or

682
00:39:05,840 --> 00:39:07,117
how much of it does it come into?

683
00:39:07,200 --> 00:39:11,837
>> Um, my bet would be that in the future
we value human relationships much more.

684
00:39:11,920 --> 00:39:16,317
I think we are wired to do that. There
will be some people who will like fall

685
00:39:16,400 --> 00:39:20,077
in love with an AI or talk about their AI
boyfriend or girlfriend or have AI

686
00:39:20,160 --> 00:39:24,237
like you know send messages on dating
apps for them that there will be some

687
00:39:24,320 --> 00:39:27,117
small percentage of the world that
does that and there will be a lot of

688
00:39:27,200 --> 00:39:31,597
breathless articles written about
the end of society and all of that but I

689
00:39:31,680 --> 00:39:35,437
think for almost everyone in a world
with more abundance in a world where you

690
00:39:35,520 --> 00:39:39,997
can kind of have anything human
connection human attention human warmth

691
00:39:40,080 --> 00:39:44,717
will be one of the most valuable
commodities and we will I bet we will

692
00:39:44,800 --> 00:39:47,117
care much more about
that in a world with AI.

693
00:39:47,200 --> 00:39:50,557
>> Okay. You've invested in over 400
companies. Is that I don't know if

694
00:39:50,640 --> 00:39:54,157
that's accurate or >> I mean
counting and Y common air like 3,000.

695
00:39:54,240 --> 00:39:58,237
>> Wow. Okay. If you take all the Y
cominator investments fair um but given

696
00:39:58,320 --> 00:40:01,757
the speed of AI growth and the disruption
where would you advise people

697
00:40:01,840 --> 00:40:03,757
in this room to invest today? Is there

698
00:40:03,840 --> 00:40:07,757
one place that you think they should?

699
00:40:07,840 --> 00:40:09,677
The cool thing about a moment like this

700
00:40:09,760 --> 00:40:12,077
when the ground is shaking is sort of

701
00:40:12,160 --> 00:40:15,437
everything is up for grabs. Um things

702
00:40:15,520 --> 00:40:20,797
are going to change so fast.
Things are changing so fast.

703
00:40:20,880 --> 00:40:23,677
As we talked about earlier, you can
have these oneperson companies doing the

704
00:40:23,760 --> 00:40:27,277
things that used to take much bigger
companies, but they can move faster,

705
00:40:27,360 --> 00:40:31,517
much more efficiently. So there will be
all of the do the existing things better

706
00:40:31,600 --> 00:40:35,437
and then there's a whole set of new
things we just couldn't imagine before.

707
00:40:35,520 --> 00:40:38,237
Um, it feels like the investment

708
00:40:38,320 --> 00:40:43,517
landscape is more open than it's ever
been in any of my time as an adult. Uh,

709
00:40:43,600 --> 00:40:46,077
there were many times before where it
felt like there was only one thing to

710
00:40:46,160 --> 00:40:48,237
invest in. There was only one frontier.

711
00:40:48,320 --> 00:40:52,077
And now it feels like you can kind
of do whatever whatever you want.

712
00:40:52,160 --> 00:40:57,277
>> Okay. How far are we from ASI? You said
AGI we're a few years away. How far are

713
00:40:57,360 --> 00:41:00,637
we from ASI? >> No, I said
from ASI we're a few years

714
00:41:00,720 --> 00:41:06,637
away. Um, >> so AGI, how far then? I mean,
hi feels pretty close at this point.

715
00:41:06,720 --> 00:41:07,037
>> Okay.

716
00:41:07,120 --> 00:41:11,677
>> Like if you had asked me,

717
00:41:11,760 --> 00:41:15,997
I think if you had asked
most people 6 years ago,

718
00:41:16,080 --> 00:41:21,197
what would you think if we had systems
that could do new research on their own?

719
00:41:21,280 --> 00:41:23,917
What would you think if we had systems
that could make an entire complex

720
00:41:24,000 --> 00:41:27,677
computer program on their own that could
do pretty sophisticated knowledge work

721
00:41:27,760 --> 00:41:31,037
in all these different fields? you know,
you could have one system that could act

722
00:41:31,120 --> 00:41:33,597
as an AI doctor, lawyer, a computer

723
00:41:33,680 --> 00:41:35,757
scientist would say, "Okay, that sounds

724
00:41:35,840 --> 00:41:38,157
pretty general and pretty intelligent."

725
00:41:38,240 --> 00:41:41,917
Um, we get used to whatever we have. Uh,

726
00:41:42,000 --> 00:41:45,597
but just watching how much
the technology we already have is

727
00:41:45,680 --> 00:41:47,917
accelerating us internally, I would say

728
00:41:48,000 --> 00:41:53,117
it's pretty close. Um, and given what
I now expect to be a faster takeoff, I

729
00:41:53,200 --> 00:41:54,797
think super intelligence is not that far

730
00:41:54,880 --> 00:42:00,600
off. The one thing
you'll never ask Chad GPT

731
00:42:06,960 --> 00:42:09,997
I think I would never ask it like how to

732
00:42:10,080 --> 00:42:14,317
be happy. I would rather
ask like a wise person then.

733
00:42:14,400 --> 00:42:18,397
>> That's a really
interesting because [applause]

734
00:42:18,480 --> 00:42:20,717
that's really interesting because uh one

735
00:42:20,800 --> 00:42:23,037
of the most uses of AI is for

736
00:42:23,120 --> 00:42:26,877
companionship for you know for that that

737
00:42:26,960 --> 00:42:31,117
categorization of >> um to the
extent of therapy and you know

738
00:42:31,200 --> 00:42:34,637
it's >> yeah I think
like for sort of things

739
00:42:34,720 --> 00:42:39,117
like therapy or life advice it can
be pretty good. I think for like life

740
00:42:39,200 --> 00:42:41,757
philosophy, I'm still
not going to take it.

741
00:42:41,840 --> 00:42:43,757
>> Okay. No. Wonderful. Okay. Wonderful.

742
00:42:43,840 --> 00:42:46,797
Um, so which is less likely to happen?

743
00:42:46,880 --> 00:42:50,717
That the uh TSMC loses its monopoly over

744
00:42:50,800 --> 00:42:55,037
world chip manufacturing or Musk
and you become friends again?

745
00:42:55,120 --> 00:42:59,917
Which one is which one is less
likely >> or more likely?

746
00:43:00,000 --> 00:43:03,437
>> Uh, I think Musk and I becoming friends

747
00:43:03,520 --> 00:43:07,517
again is less likely. I feel like
I have more control over that one.

748
00:43:07,600 --> 00:43:09,837
But both very unlikely things.

749
00:43:09,920 --> 00:43:15,037
>> What is it about TSMC that you
know the the semiconductor Taiwan semic

750
00:43:15,120 --> 00:43:18,397
semiconductor corporation? What what
is it about it that makes it so I mean the

751
00:43:18,480 --> 00:43:21,917
whole world is kind of just >>
it's they have just this relentless

752
00:43:22,000 --> 00:43:24,557
focus on getting better at what they do.

753
00:43:24,640 --> 00:43:29,517
It's this incredibly complex process
and they are just a machine at how they

754
00:43:29,600 --> 00:43:32,237
optimize and get better
and learn at every level.

755
00:43:32,320 --> 00:43:36,477
>> Okay. Um, the most recent update you
can share with us about your partnership

756
00:43:36,560 --> 00:43:40,157
with Johnny Ives. You're talking about
these new cool devices. It's going to

757
00:43:40,240 --> 00:43:43,677
change the form factor of a cell phone,
how you interact with machines. Anything

758
00:43:43,760 --> 00:43:46,077
you can share with us about that?

759
00:43:46,160 --> 00:43:49,160
>> Um,

760
00:43:51,920 --> 00:43:56,557
we have basically used computers for the
same way in for like 50 years. You know,

761
00:43:56,640 --> 00:44:00,637
I think the work that Xerox Park did
to invent all the concepts of Windows and

762
00:44:00,720 --> 00:44:02,797
pointing devices and everything that we

763
00:44:02,880 --> 00:44:05,597
now take for granted is amazing. Putting

764
00:44:05,680 --> 00:44:07,677
it on a phone and adding multi-touch was

765
00:44:07,760 --> 00:44:10,237
also amazing. But fundamentally, it's

766
00:44:10,320 --> 00:44:13,277
been, you know, a similar kind of idea.

767
00:44:13,360 --> 00:44:15,517
And then AI came along and AI is stuck

768
00:44:15,600 --> 00:44:20,317
in that same form factor. But AI is like
a quite different thing. AI, you can now

769
00:44:20,400 --> 00:44:23,677
talk to a computer in natural language
and it can do incredibly complex things

770
00:44:23,760 --> 00:44:28,077
for you. it can understand a huge
amount of context and the form factor of

771
00:44:28,160 --> 00:44:30,477
computers does not quite work for this.

772
00:44:30,560 --> 00:44:35,037
You know, I want I want to use a piece
of technology that is observing my whole

773
00:44:35,120 --> 00:44:36,797
life and that has all the context and is

774
00:44:36,880 --> 00:44:39,357
maximally useful to me and that has the

775
00:44:39,440 --> 00:44:42,237
that is in my way the least and that can

776
00:44:42,320 --> 00:44:43,837
not just be on or off but these sort of

777
00:44:43,920 --> 00:44:48,077
subtle states in between. And

778
00:44:48,160 --> 00:44:50,477
we may fail at this. This is very hard.

779
00:44:50,560 --> 00:44:54,557
But I think we have a chance to make
a new kind of products, a new family of

780
00:44:54,640 --> 00:44:58,237
products that were really designed
around AI that are kind of participating

781
00:44:58,320 --> 00:45:01,197
in your life and not in the way of it.

782
00:45:01,280 --> 00:45:03,277
>> So that's a theory,
but what is the product? I mean,

783
00:45:03,360 --> 00:45:07,437
>> you'll have to wait and see. [laughter]
>> Do we get uh when can we see this?

784
00:45:07,520 --> 00:45:10,557
Because it's sounding
very re revolutionary. So

785
00:45:10,640 --> 00:45:14,957
>> um I mean we hope
to talk be able to talk

786
00:45:15,040 --> 00:45:17,677
about it late this year,
but >> Okay. Sure.

787
00:45:17,760 --> 00:45:21,997
>> Hardware is hard. >> Okay. Um

788
00:45:22,080 --> 00:45:25,197
the one thing that governments should
regulate and one thing they should avoid

789
00:45:25,280 --> 00:45:28,280
regulating.

790
00:45:30,880 --> 00:45:32,477
This is not a competent answer.

791
00:45:32,560 --> 00:45:36,477
Generally speaking, I think it's
probably a good idea for governments to

792
00:45:36,560 --> 00:45:38,477
focus on regulating the really

793
00:45:38,560 --> 00:45:41,197
potentially catastrophic issues and

794
00:45:41,280 --> 00:45:44,317
being more lenient on the less important

795
00:45:44,400 --> 00:45:46,317
issues until we understand them better.

796
00:45:46,400 --> 00:45:50,957
>> Okay. Is there any one introspection
or learning from your professional

797
00:45:51,040 --> 00:45:57,240
disagreements with any
of the co-founders you've had?

798
00:46:02,160 --> 00:46:09,000
>> It It's It's hard to reduce
that to a single sound bite. Um,

799
00:46:12,800 --> 00:46:16,957
one thing that Greg and I have disagreed
on a lot over the years is how much the

800
00:46:17,040 --> 00:46:20,877
company should focus versus do a lot
of things. And I think I've been wrong

801
00:46:20,960 --> 00:46:24,637
about trying to do so many things.
And I've really learned from him the

802
00:46:24,720 --> 00:46:27,517
importance of extremely narrow deep focus.

803
00:46:27,600 --> 00:46:30,157
>> Okay. Interesting. one mistake the

804
00:46:30,240 --> 00:46:36,200
biggest mistake you have seen
corporates making in adopting AI.

805
00:46:39,760 --> 00:46:42,957
Um, I was in a meeting yesterday uh

806
00:46:43,040 --> 00:46:47,197
yesterday with a big company who was

807
00:46:47,280 --> 00:46:49,997
planning to uh

808
00:46:50,080 --> 00:46:51,997
spend 2026

809
00:46:52,080 --> 00:46:54,797
strategizing, 2027 getting the company

810
00:46:54,880 --> 00:46:59,400
ready and 2028 deploying.

811
00:47:00,080 --> 00:47:03,037
And that may work for other kinds of

812
00:47:03,120 --> 00:47:07,117
technology. Apparently, if you do like
a giant ERP migration, that's the kind of

813
00:47:07,200 --> 00:47:09,677
timeline it takes. Doing that for AI

814
00:47:09,760 --> 00:47:12,717
will be a catastrophic mistake um the

815
00:47:12,800 --> 00:47:17,517
the the nimleness required, the speed,
the commitment required is just totally

816
00:47:17,600 --> 00:47:21,117
different. >> Okay. Okay.
Uh get to the last piece

817
00:47:21,200 --> 00:47:25,277
here. Um following heads of state want
your opinion on one thing they need to

818
00:47:25,360 --> 00:47:27,997
do for their people's future in this new

819
00:47:28,080 --> 00:47:31,437
AI context. Xi Jinping, Narendra Modi,

820
00:47:31,520 --> 00:47:38,040
Putin, Trump. What's the one thing you
tell each of them that they should do?

821
00:47:38,720 --> 00:47:41,757
>> I would say the same thing to everyone,
which is you got to democratize this

822
00:47:41,840 --> 00:47:44,717
technology and put it in the hands
of people. They're not all going to listen

823
00:47:44,800 --> 00:47:48,990
to that. Obviously, [laughter]

824
00:47:49,840 --> 00:47:52,877
it's a fair answer, of course,
but it's a safe answer. But is it is it a

825
00:47:52,960 --> 00:47:55,517
different thing that you know like
a young country should do to an old

826
00:47:55,600 --> 00:47:57,437
country from a from a head of state

827
00:47:57,520 --> 00:48:00,237
perspective? I mean

828
00:48:00,320 --> 00:48:05,517
>> I don't I mean I feel
increasingly radicalized

829
00:48:05,600 --> 00:48:07,117
on this point. I don't think any other

830
00:48:07,200 --> 00:48:08,477
strategy is going to work.

831
00:48:08,560 --> 00:48:13,437
>> Okay, fair. Um the one if I can ask you

832
00:48:13,520 --> 00:48:17,277
to describe your relationship with Satya
Nadella because he calls you he calls

833
00:48:17,360 --> 00:48:20,157
open a friend of me. Uh

834
00:48:20,240 --> 00:48:21,677
>> I would just say friends.

835
00:48:21,760 --> 00:48:26,157
>> Okay, cool. Your most expensive hire

836
00:48:26,240 --> 00:48:27,597
ever. You don't have to tell me who, but

837
00:48:27,680 --> 00:48:30,317
how much money did it it cost? Because

838
00:48:30,400 --> 00:48:35,640
>> we we we are no
comparison to Meta here.

839
00:48:36,090 --> 00:48:40,237
[applause] >> Okay. Okay.
Which my last question now,

840
00:48:40,320 --> 00:48:44,477
which of these four statements that you
have you have made and you have now

841
00:48:44,560 --> 00:48:47,197
upgraded or revised do you regret?

842
00:48:47,280 --> 00:48:51,197
Statement one that India building
a foundation model with $10 million is

843
00:48:51,280 --> 00:48:55,357
hopeless. Statement two that you will
remain a non for profit. I clarified that.

844
00:48:55,440 --> 00:48:59,197
>> Yeah. I said anyone building the point
was anyone building a foundation model

845
00:48:59,280 --> 00:49:01,037
for $10 million. >> Okay,
so that's statement one. Statement

846
00:49:01,120 --> 00:49:02,637
two, you'll remain a not for profofit.

847
00:49:02,720 --> 00:49:04,877
Statement three, too much AI regulation

848
00:49:04,960 --> 00:49:06,957
can stifle innovation. And statement

849
00:49:07,040 --> 00:49:11,037
four, open AI won't accept advertising.

850
00:49:11,120 --> 00:49:13,517
>> I don't think I ever said
we won't accept advertising.

851
00:49:13,600 --> 00:49:15,117
>> Sorry. >> Um I don't
think I ever said we won't

852
00:49:15,200 --> 00:49:18,157
accept advertising. I think maybe I said
I had misgivings. I guess I'll pick the

853
00:49:18,240 --> 00:49:22,077
nonprofit one. >> Okay.
All right. Super. Um it's been a

854
00:49:22,160 --> 00:49:25,677
great uh I've tried to cover as much
as I could. Must have been a lot.

855
00:49:25,760 --> 00:49:28,550
>> Um, [laughter]

856
00:49:29,680 --> 00:49:33,437
>> you only gave me an hour, man.
I would have wanted few. We'll take a few

857
00:49:33,520 --> 00:49:34,797
questions from the audience. Yeah.

858
00:49:34,880 --> 00:49:36,957
>> Okay. Can I get a show
of hands who all are here for?

859
00:49:37,040 --> 00:49:38,797
>> I would love another meme.
So, if anyone

860
00:49:38,880 --> 00:49:41,757
wants to ask me about Indian
foundation models, let's run it back.

861
00:49:41,840 --> 00:49:45,357
>> Okay. Yeah. Um, can we Okay. Yeah. Can

862
00:49:45,440 --> 00:49:47,677
we get Rajes please first? Rajes is the

863
00:49:47,760 --> 00:49:52,557
founder of Make My Trip. Uh,
he makes money when we are holidaying.

864
00:49:52,640 --> 00:49:56,877
>> All right. Uh, thank you. Well, thank
you Anand. Uh, and welcome again, Sam.

865
00:49:56,960 --> 00:50:00,797
Uh, incredible one hour chat and and you
did a great job by the way, keeping

866
00:50:00,880 --> 00:50:03,517
yourself calm while Anand did a great

867
00:50:03,600 --> 00:50:06,477
job on provoking you hell of a lot. Um,

868
00:50:06,560 --> 00:50:11,117
so that's good job. Yeah, by the way,
partners, I met Brad earlier today. I

869
00:50:11,200 --> 00:50:12,717
just have one question. It got triggered

870
00:50:12,800 --> 00:50:15,037
from your comment actually that you are

871
00:50:15,120 --> 00:50:18,317
a research first company, right? Um, and

872
00:50:18,400 --> 00:50:21,517
if I think harder about that, um, the

873
00:50:21,600 --> 00:50:24,317
kick for researchers

874
00:50:24,400 --> 00:50:27,597
um the satisfaction that they get when

875
00:50:27,680 --> 00:50:29,517
they every time they go deep into

876
00:50:29,600 --> 00:50:32,717
research and they strike gold right you

877
00:50:32,800 --> 00:50:36,957
know in terms of innovation in terms
of something really crazy that they

878
00:50:37,040 --> 00:50:38,797
discover

879
00:50:38,880 --> 00:50:44,317
do you think um um given the power of AI

880
00:50:44,400 --> 00:50:47,677
that's equal amount of focus needs to be

881
00:50:47,760 --> 00:50:51,997
on responsible AI as well

882
00:50:52,080 --> 00:50:56,397
>> yes I do think that and I think the one
of the things I'm most proud of with our

883
00:50:56,480 --> 00:51:00,157
research team is how much they feel
that as we get closer and closer to super

884
00:51:00,240 --> 00:51:03,597
intelligence they of course feel it more
and more but it's been a core part of

885
00:51:03,680 --> 00:51:07,837
our DNA from the very beginning and I
think the researchers that succeed the

886
00:51:07,920 --> 00:51:12,397
most with us have that process
going on in their brains

887
00:51:12,480 --> 00:51:15,517
>> I think we don't have time the last
question is I'm I'm getting some hand

888
00:51:15,600 --> 00:51:19,757
signals yeah okay last question >> do
a few more we'll go I'll answer fast

889
00:51:19,840 --> 00:51:24,640
>> the race for talent u
where do you think it's heading

890
00:51:25,040 --> 00:51:28,040
Um,

891
00:51:29,280 --> 00:51:33,037
as AI does more and more of research and

892
00:51:33,120 --> 00:51:36,157
engineering, I suspect teams will do

893
00:51:36,240 --> 00:51:37,757
much more with smaller amounts of

894
00:51:37,840 --> 00:51:41,677
people. And so you'll have uh

895
00:51:41,760 --> 00:51:43,517
I I think I think the race for talent

896
00:51:43,600 --> 00:51:46,717
should ease somewhat and you'll see

897
00:51:46,800 --> 00:51:49,117
small teams uh in research labs just do

898
00:51:49,200 --> 00:51:52,157
a huge amount of work.

899
00:51:52,240 --> 00:51:54,557
A question for Anand actually. I think

900
00:51:54,640 --> 00:51:56,717
he's asked you all the questions. How

901
00:51:56,800 --> 00:51:59,197
much did OpenAI chat GPD help you with

902
00:51:59,280 --> 00:52:02,157
the question?

903
00:52:02,240 --> 00:52:07,357
>> I I I I have to admit I I I use a lot
for research. But if you ask OpenAI for

904
00:52:07,440 --> 00:52:11,597
any interview to help you and give
the questions I you know it it just feels

905
00:52:11,680 --> 00:52:13,437
very synthetic. So I don't end up uh you

906
00:52:13,520 --> 00:52:16,317
know using it yet. Uh

907
00:52:16,400 --> 00:52:20,637
>> that's good. [laughter] >> Sam
I don't do we have time one more or

908
00:52:20,720 --> 00:52:25,037
no? Yes. Okay, cool. Can we actually
show hands? I'm not getting a sense. Can

909
00:52:25,120 --> 00:52:29,517
Can I do some uh sort of affirmative
action and get a women some women hands

910
00:52:29,600 --> 00:52:31,677
please? Yes, please. Right there.

911
00:52:31,760 --> 00:52:35,677
Thanks. Please.

912
00:52:35,760 --> 00:52:39,437
>> Uh, hi Sam. This is Tuti. Um, before I

913
00:52:39,520 --> 00:52:44,077
ask you my question, I think I might
have manifested this. I'm a podcaster

914
00:52:44,160 --> 00:52:48,397
and I asked Chad GBD. >> Okay,
it's not a short question. Please.

915
00:52:48,480 --> 00:52:50,317
>> So, because this is what I manifested.

916
00:52:50,400 --> 00:52:52,237
This is you and me sitting in your open

917
00:52:52,320 --> 00:52:56,157
AI 3 weeks ago and here I am. But my

918
00:52:56,240 --> 00:52:58,557
question to you is as a creator of

919
00:52:58,640 --> 00:53:01,597
course we use a lot of chat GBT but what

920
00:53:01,680 --> 00:53:05,357
concerns you more will AI become too

921
00:53:05,440 --> 00:53:07,757
powerful or will humans become very

922
00:53:07,840 --> 00:53:10,717
passive especially for creators you know

923
00:53:10,800 --> 00:53:13,917
chat GPD like even an unset that it sort

924
00:53:14,000 --> 00:53:17,917
of becomes very superficial
when you rely only on that

925
00:53:18,000 --> 00:53:20,957
>> I don't think humans will become too
passive so I guess I'll take the other

926
00:53:21,040 --> 00:53:23,277
one uh watching what creators are doing

927
00:53:23,360 --> 00:53:25,357
in particular where they can use these

928
00:53:25,440 --> 00:53:28,477
tools to help have a quicker, tighter

929
00:53:28,560 --> 00:53:33,277
iteration loop from idea to feedback
to idea to feedback. Um, I think we'll get

930
00:53:33,360 --> 00:53:36,157
better stuff. I think we'll get more
creative podcast, better writing, better

931
00:53:36,240 --> 00:53:39,437
images, but I don't think we're going
to become passive in that process. I used

932
00:53:39,520 --> 00:53:42,477
to worry about that, but it doesn't
seem like it's what's happening.

933
00:53:42,560 --> 00:53:45,677
>> Okay, we I actually I've got a list of
people we had promised questions to. So,

934
00:53:45,760 --> 00:53:47,757
Mr. Abishek Katan is here. Sorry, where

935
00:53:47,840 --> 00:53:51,117
is he? No. Oh, there he is. Can we get a

936
00:53:51,200 --> 00:53:54,397
mic to him, please? Uh Hazel is Hazel

937
00:53:54,480 --> 00:53:57,757
around yeah please go ahead.

938
00:53:57,840 --> 00:53:59,757
>> Uh first of all it's a very interesting

939
00:53:59,840 --> 00:54:03,037
session uh what I wanted to ask is

940
00:54:03,120 --> 00:54:05,597
according to you which profession

941
00:54:05,680 --> 00:54:09,037
because of uh chat GPD and AI will be

942
00:54:09,120 --> 00:54:12,840
most under threat.

943
00:54:13,280 --> 00:54:17,837
>> Um I think a lot
of professions will almost

944
00:54:17,920 --> 00:54:20,157
go away. a lot of the current ones we

945
00:54:20,240 --> 00:54:22,957
will find new things to do for sure and

946
00:54:23,040 --> 00:54:25,597
people will adapt those but if you think

947
00:54:25,680 --> 00:54:28,397
about

948
00:54:28,480 --> 00:54:31,597
well I'll pick my own uh I I was trained

949
00:54:31,680 --> 00:54:33,917
as a software engineer and the way I

950
00:54:34,000 --> 00:54:39,037
learned to write software is now
effectively completely irrelevant so it

951
00:54:39,120 --> 00:54:40,317
doesn't mean there's no longer a

952
00:54:40,400 --> 00:54:42,477
software engineering job in the future

953
00:54:42,560 --> 00:54:46,237
but writing you know C++ code by hand

954
00:54:46,320 --> 00:54:49,037
that's over And I think there's a lot of

955
00:54:49,120 --> 00:54:53,357
other professions where in the old sense
they're over. There may be something

956
00:54:53,440 --> 00:54:56,317
totally new. There will be other
professions that change almost not at

957
00:54:56,400 --> 00:54:59,997
all, but big categories of jobs AI is

958
00:55:00,080 --> 00:55:04,237
just going to completely obsolete and
we'll what those people do will have to

959
00:55:04,320 --> 00:55:07,517
completely adapt. Um >> Sam,
let's flip the question. What's the

960
00:55:07,600 --> 00:55:11,320
least vulnerable job?

961
00:55:12,880 --> 00:55:16,317
>> Well, there was a question about
creators. One one thing that I think is

962
00:55:16,400 --> 00:55:18,877
interesting is when AI started

963
00:55:18,960 --> 00:55:23,837
generating images, people said uh
graphic artists that's over. That's

964
00:55:23,920 --> 00:55:28,317
done. And that might be true for the
kind of graphic artist job that was to

965
00:55:28,400 --> 00:55:31,997
to sort of like, you know,
make someone's birthday card invitation or

966
00:55:32,080 --> 00:55:34,157
something. But for fine art, the price

967
00:55:34,240 --> 00:55:37,837
of AI generated art is a zero. And the

968
00:55:37,920 --> 00:55:41,197
price of human generated graphic art
has continued to go up since this has

969
00:55:41,280 --> 00:55:43,917
happened. And so there are many things

970
00:55:44,000 --> 00:55:46,077
like that where we care about the person

971
00:55:46,160 --> 00:55:49,277
who does it. Um you know another example

972
00:55:49,360 --> 00:55:53,437
is like I really care about I had to go

973
00:55:53,520 --> 00:55:56,957
to the hospital recently. I really cared
about the nurse that was taking care of

974
00:55:57,040 --> 00:56:00,317
me. If that were like a robot I think I
would have been pretty unhappy no matter

975
00:56:00,400 --> 00:56:01,757
how smart the robot was.

976
00:56:01,840 --> 00:56:04,957
>> Okay. Hazel.

977
00:56:05,040 --> 00:56:10,077
>> Hi Sam. I'm from Chitkara University.
Um just taking from the previous question

978
00:56:10,160 --> 00:56:11,917
on responsible

979
00:56:12,000 --> 00:56:15,197
AI how do you see your role like do you

980
00:56:15,280 --> 00:56:18,397
invest in time putting in checks and

981
00:56:18,480 --> 00:56:20,957
balances you know and does the pace of

982
00:56:21,040 --> 00:56:23,997
the checks and balances are in the same

983
00:56:24,080 --> 00:56:27,037
line as the you know the advancements

984
00:56:27,120 --> 00:56:31,837
that we are seeing in AI um
so just around that thank you

985
00:56:31,920 --> 00:56:33,757
>> yeah it's a huge focus of ours as a

986
00:56:33,840 --> 00:56:37,117
company and mine personally um

987
00:56:37,200 --> 00:56:39,117
sometimes Sometimes we overdo it.

988
00:56:39,200 --> 00:56:42,557
Sometimes we're too conservative with a
new technology or a new level. Sometimes

989
00:56:42,640 --> 00:56:46,717
we put in too many checks. People get
mad about that. Uh there is a tension

990
00:56:46,800 --> 00:56:51,677
between democratizing AI and enough
safety. But I think our principle of

991
00:56:51,760 --> 00:56:55,277
starting conservative and then
broadening access has served us well and

992
00:56:55,360 --> 00:56:57,357
we plan to continue to do that.

993
00:56:57,440 --> 00:57:00,317
>> Sam, we're completely out of time. Sam,
thank you so much. Thank you so much for

994
00:57:00,400 --> 00:57:04,280
your time. Thank you for being here.