1
00:00:00,160 --> 00:00:01,917
So, how did OpenAI come about?

2
00:00:02,000 --> 00:00:06,157
>> I knew I wanted to do a startup
because I felt like that was something.

3
00:00:06,240 --> 00:00:08,477
>> But you were just in a startup.
Stripe was a startup.

4
00:00:08,560 --> 00:00:11,917
>> It's true. But I never I I felt like

5
00:00:12,000 --> 00:00:16,877
Stripe, the problem that we
were solving was not my problem,

6
00:00:16,960 --> 00:00:19,437
right? It wasn't the problem I had grown
up thinking about. It was an important

7
00:00:19,520 --> 00:00:23,757
problem that I I dedicated myself
to that mission for a number of years, but

8
00:00:23,840 --> 00:00:28,317
I felt like it was going to succeed with
or without me. And so then I had a first

9
00:00:28,400 --> 00:00:32,157
moment to really think about what is a
mission that I want to dedicate myself

10
00:00:32,240 --> 00:00:36,797
to where I would spend the rest of my life
working on this problem just to see

11
00:00:36,880 --> 00:00:38,877
it play out in a slightly better way.

12
00:00:38,960 --> 00:00:43,037
And it was very clear to me that top
of the list was AI, right? If you could

13
00:00:43,120 --> 00:00:48,077
actually make a difference in how
AI will play out in the world,

14
00:00:48,160 --> 00:00:49,997
like that would be a life well-lived.

15
00:00:50,080 --> 00:00:53,757
When you were thinking about leaving,
Patrick told you to go talk to Sam

16
00:00:53,840 --> 00:00:56,317
Alman. What happened in that conversation?

17
00:00:56,400 --> 00:00:58,557
>> Well, so Patrick
had said, "Sam has seen

18
00:00:58,640 --> 00:01:00,797
lots of young people in your situation."

19
00:01:00,880 --> 00:01:04,157
And Patrick, I think, really hoped
that Sam would convince me to stay. A few

20
00:01:04,240 --> 00:01:08,477
minutes of talking to Sam, he's like,
"Okay, you clearly have already decided.

21
00:01:08,560 --> 00:01:11,677
It is very obvious." And so he asked,
"Well, what are you planning on doing

22
00:01:11,760 --> 00:01:16,237
next?" And I said, "Well, I'm thinking
about doing an AI company." And he said,

23
00:01:16,320 --> 00:01:18,077
"I'm also thinking about doing something

24
00:01:18,160 --> 00:01:20,077
in AI. We should keep in touch." So, I

25
00:01:20,160 --> 00:01:23,277
had talked to Sam maybe one more time

26
00:01:23,360 --> 00:01:27,437
after I was leaving Stripe and he asked,
"Are you still thinking about doing

27
00:01:27,520 --> 00:01:31,357
something in AI?" I said, "Yes."
He said, "I'm also starting to get more

28
00:01:31,440 --> 00:01:33,837
details and putting together this dinner

29
00:01:33,920 --> 00:01:35,517
in

30
00:01:35,600 --> 00:01:38,797
July." And I flew out for the dinner.

31
00:01:38,880 --> 00:01:42,237
And the thing that I remember was a

32
00:01:42,320 --> 00:01:45,997
topic was, is it too late to start a lab

33
00:01:46,080 --> 00:01:50,317
with many of the best researchers?
Is it possible? And this is what year?

34
00:01:50,400 --> 00:01:52,157
>> 2015, right? Because you think about

35
00:01:52,240 --> 00:01:55,437
just the degree to which Deep Mind had

36
00:01:55,520 --> 00:02:00,397
all the researchers, all the capital,
all the data. It just felt like, is it

37
00:02:00,480 --> 00:02:04,157
even possible to get something off
the ground still? People came up with all

38
00:02:04,240 --> 00:02:07,357
sorts of reasons. It was hard.
No one could come up with a reason it was

39
00:02:07,440 --> 00:02:11,917
actually impossible. And so Sam and I
driving back to the city that night. I

40
00:02:12,000 --> 00:02:13,837
remember we looked at each other and we

41
00:02:13,920 --> 00:02:17,037
said, "We got to do this, right? Like,

42
00:02:17,120 --> 00:02:20,957
we just have to." And so next day I was
full-time on putting this together. And

43
00:02:21,040 --> 00:02:23,037
it was tough because it was very

44
00:02:23,120 --> 00:02:25,837
illdefined. We had a mission, a vision

45
00:02:25,920 --> 00:02:30,237
of saying we think that we can build
human level AI, make it be something

46
00:02:30,320 --> 00:02:33,437
positive for the world,
make the benefits be something that are

47
00:02:33,520 --> 00:02:38,237
distributed broadly. But how and how do
you get people to actually leave their

48
00:02:38,320 --> 00:02:40,237
jobs to come and join this thing?

49
00:02:40,320 --> 00:02:41,757
Initially, the set of people that I

50
00:02:41,840 --> 00:02:43,997
narrowed down to were actually Ilia,

51
00:02:44,080 --> 00:02:47,437
Dario, Amadai, Chrysola, and myself.

52
00:02:47,520 --> 00:02:50,717
that was going to be the team and we
spent a lot of time together. We spent a

53
00:02:50,800 --> 00:02:54,077
lot of time talking about potential
visions for the lab, potential ways that

54
00:02:54,160 --> 00:02:58,957
things would work. It didn't quite come
together and that there was just partly

55
00:02:59,040 --> 00:03:02,957
a question of will this have enough
momentum. You know, Daario felt like

56
00:03:03,040 --> 00:03:05,997
that he needed to go and establish
a name for himself and he wasn't sure if

57
00:03:06,080 --> 00:03:08,957
this was really going to be it. It was a
question of just how it was all going to

58
00:03:09,040 --> 00:03:12,717
work. In the meanwhile, I was trying
to get John Schulman interested. He said

59
00:03:12,800 --> 00:03:15,837
that he was going to do it. Dario and
Chris ended up deciding to go to Google

60
00:03:15,920 --> 00:03:19,597
Brain. And so it was really just,
you know, Ilia, me, and you know, John

61
00:03:19,680 --> 00:03:21,277
starting to be maybe a few others. And

62
00:03:21,360 --> 00:03:26,717
so I had a group of about 10 people that

63
00:03:26,800 --> 00:03:29,917
many of them were saying, "I'm interested,
but who else is in?" I asked

64
00:03:30,000 --> 00:03:33,997
Sam, "Okay, how do we break symmetry here?
How do we actually get everyone to

65
00:03:34,080 --> 00:03:38,077
say, all right, we're joining?" And Sam's
suggestion was, "Invite people out

66
00:03:38,160 --> 00:03:43,437
for an off-site." So, we set up a thing
in Napa and I actually made t-shirts.

67
00:03:43,520 --> 00:03:45,917
Uh, at the time we were >>
and this is before they had joined.

68
00:03:46,000 --> 00:03:48,877
>> There was no official offers. No one
had joined. We didn't have a structure. We

69
00:03:48,960 --> 00:03:50,637
had nothing. Right? We just had an idea.

70
00:03:50,720 --> 00:03:52,877
We had a vision. We had a mission and we

71
00:03:52,960 --> 00:03:55,757
flew people out. We drove up to to Napa

72
00:03:55,840 --> 00:03:57,757
together and it was an amazing day.

73
00:03:57,840 --> 00:04:00,077
Right? The ideas were flowing. We came

74
00:04:00,160 --> 00:04:05,197
up with what I would really say
is almost the technical plan that we have

75
00:04:05,280 --> 00:04:09,517
pursued for the past 10 years.
Number one, solve reinforcement learning.

76
00:04:09,600 --> 00:04:11,997
Number two, solve unsupervised learning.

77
00:04:12,080 --> 00:04:13,757
And number three was gradually learn

78
00:04:13,840 --> 00:04:17,277
more complicated in quotes. After that

79
00:04:17,360 --> 00:04:21,677
offsite, I sent offers to everyone and

80
00:04:21,760 --> 00:04:24,877
said, "Hey, we want to get started
in the next 2 to 3 weeks. Please let me

81
00:04:24,960 --> 00:04:27,837
know if you're in."
>> Why did you think that Deep Mind had

82
00:04:27,920 --> 00:04:30,237
such an insurmountable advantage?

83
00:04:30,320 --> 00:04:34,637
>> It was very much the case that Google
Deep Mind was the 10,000lb gorilla in

84
00:04:34,720 --> 00:04:39,597
the field. They just had lots of capital.
They had the track record. And

85
00:04:39,680 --> 00:04:43,837
this was before AlphaGo, right? AlphaGo
came out a couple months later, but it

86
00:04:43,920 --> 00:04:47,357
wasn't a surprise, right? It's like
very much the momentum was very clearly

87
00:04:47,440 --> 00:04:49,517
there. And so the question of is it

88
00:04:49,600 --> 00:04:51,677
really possible to build something

89
00:04:51,760 --> 00:04:54,717
independent and new? It wasn't obvious.

90
00:04:54,800 --> 00:04:58,397
>> At what point did you realize that like
this nonprofit thing just wasn't going

91
00:04:58,480 --> 00:05:01,917
to work? >> In 2017,
we started to think very hard

92
00:05:02,000 --> 00:05:05,437
about first of all, how do we really
achieve the mission? How do we actually

93
00:05:05,520 --> 00:05:07,757
build an AGI? What will that look like?

94
00:05:07,840 --> 00:05:10,797
And we started to do the math on compute.
And you start to realize that

95
00:05:10,880 --> 00:05:14,237
it's gonna take big computer. And we

96
00:05:14,320 --> 00:05:16,877
came across a company called Cerebras,

97
00:05:16,960 --> 00:05:22,077
which was building a unique piece
of computing hardware. And the kind of

98
00:05:22,160 --> 00:05:25,437
computer that they were promising, we
realized was going to be far advanced of

99
00:05:25,520 --> 00:05:27,117
where our compute calculations looked.

100
00:05:27,200 --> 00:05:29,917
As you start to realize if we could
buy a lot of those computers, we could

101
00:05:30,000 --> 00:05:33,837
actually probably succeed at building an
AGI. If we could get exclusive access to

102
00:05:33,920 --> 00:05:38,397
Cerebrris, that could give us
an overwhelming advantage. If we could buy

103
00:05:38,480 --> 00:05:42,557
very large data centers, that could be
something unique as well. And the thing

104
00:05:42,640 --> 00:05:45,037
about nonprofit [clears throat]
fundraising is I think that there is

105
00:05:45,120 --> 00:05:47,037
essentially a cap to what is possible

106
00:05:47,120 --> 00:05:52,077
there. And so Elon, Sam, Ilia, and I all

107
00:05:52,160 --> 00:05:56,157
agreed that the only path forward
for OpenAI, the only path to achieve the

108
00:05:56,240 --> 00:06:00,317
mission was to create a for-profit
entity associated with OpenAI of some

109
00:06:00,400 --> 00:06:03,917
form. And so we were committed to that
direction and that is something that we

110
00:06:04,000 --> 00:06:06,077
knew was the only way
to achieve the mission.

111
00:06:06,160 --> 00:06:10,157
>> When was the moment that you realized
everything was going to change for you?

112
00:06:10,240 --> 00:06:12,557
Was that Dota or was it before then or

113
00:06:12,640 --> 00:06:15,197
after? The way that opening eye works is

114
00:06:15,280 --> 00:06:18,077
it's a series of moments where you

115
00:06:18,160 --> 00:06:21,037
realize that it's real now and every

116
00:06:21,120 --> 00:06:26,077
time you think that you understand it,
that is really settled in for you, you

117
00:06:26,160 --> 00:06:29,757
realize that there is a new horizon you
had not yet appreciated. And so along

118
00:06:29,840 --> 00:06:34,077
the way, I think that there was the
initial launch. It was like, "Wow, we

119
00:06:34,160 --> 00:06:37,837
actually got a team together. Now we can
pursue this mission." But you show up at

120
00:06:37,920 --> 00:06:40,797
the office the next day and you're like,
"Well, what do we do?" Right? We didn't

121
00:06:40,880 --> 00:06:43,677
even have a whiteboard. You know,
I John wanted to write something on a

122
00:06:43,760 --> 00:06:46,237
whiteboard. I was like, "I will get
a whiteboard. That's something I can do."

123
00:06:46,320 --> 00:06:49,917
Dota, we had our first big result,
right? That really was like, "Wow, we

124
00:06:50,000 --> 00:06:52,717
can actually accomplish something when
we put our mind to it. You can actually

125
00:06:52,800 --> 00:06:54,237
see all this compute coming together.

126
00:06:54,320 --> 00:06:57,997
You scale up the compute, you scale up
the result." There were multiple moments

127
00:06:58,080 --> 00:07:02,237
with the GPT series and I remember
actually an early moment was the

128
00:07:02,320 --> 00:07:04,237
unsupervised sentiment neuron paper.

129
00:07:04,320 --> 00:07:06,637
Have you heard of that one?
>> I've heard of it but I haven't read it.

130
00:07:06,720 --> 00:07:10,797
>> Okay. Yes, that one's that one's
an interesting one cuz it's 2017 and it's

131
00:07:10,880 --> 00:07:15,357
really the first time that we saw
semantics arise from training on a

132
00:07:15,440 --> 00:07:18,717
language modeling objective.
So you train on learn the next character,

133
00:07:18,800 --> 00:07:22,237
predict the next character and then
suddenly you get a neural net that

134
00:07:22,320 --> 00:07:26,157
understands sentiment, understands if
something is positive or negative. It's

135
00:07:26,240 --> 00:07:30,397
harder than it sounds, but that was a
moment where you realize, wow, we are

136
00:07:30,480 --> 00:07:33,597
building machines that can learn
semantics, not just where the commas are

137
00:07:33,680 --> 00:07:36,877
and where the nouns and verbs are,
but it can really learn the meaning of

138
00:07:36,960 --> 00:07:39,997
sentences. You got to push that.
And then of course when you see something

139
00:07:40,080 --> 00:07:43,437
like GPD4 I remember we were playing

140
00:07:43,520 --> 00:07:47,117
with it and someone asked

141
00:07:47,200 --> 00:07:51,837
why is this thing not an AGI right it's
like actually really hard to put your

142
00:07:51,920 --> 00:07:55,277
finger on it because you can talk to it
fluently in anything you want it clearly

143
00:07:55,360 --> 00:07:58,157
wasn't an AGI it was lacking something

144
00:07:58,240 --> 00:08:03,677
but just if you'd described your
criteria for AGI 2 months prior it

145
00:08:03,760 --> 00:08:06,797
probably would have been compatible
with what GBD4 was and so there are many

146
00:08:06,880 --> 00:08:10,557
moments along the way where you feel
like it's real now. It's going to really

147
00:08:10,640 --> 00:08:14,397
happen. The economy is going
to transform into this compute powered

148
00:08:14,480 --> 00:08:19,437
world. And I think that those moments
are not yet at the end. I think that we

149
00:08:19,520 --> 00:08:22,477
have many more breakthrough moments
where you realize that the next stage is

150
00:08:22,560 --> 00:08:26,797
possible. I thought Dota was like
an incredible moment because it was it

151
00:08:26,880 --> 00:08:31,197
wasn't u chess like deep blue
and it wasn't alpha go which is like

152
00:08:31,280 --> 00:08:35,117
computationally intensive but very
defined rules. It was actually

153
00:08:35,200 --> 00:08:37,757
interactive against humans in a way that

154
00:08:37,840 --> 00:08:39,997
like the world is sort of structured but

155
00:08:40,080 --> 00:08:42,717
you have all this freedom. Yeah, that

156
00:08:42,800 --> 00:08:45,837
was that was something very compelling

157
00:08:45,920 --> 00:08:47,837
about it. And the ironic thing is we'd

158
00:08:47,920 --> 00:08:51,277
actually set out with Dota to develop

159
00:08:51,360 --> 00:08:55,117
new methods because the reinforcement

160
00:08:55,200 --> 00:08:57,757
learning at the time was clearly not

161
00:08:57,840 --> 00:09:02,237
going to scale. Right? The the algorithm
we use called PO you plan over every

162
00:09:02,320 --> 00:09:05,277
single time step. there's no hierarchy,
right? As a human, that's not how you

163
00:09:05,360 --> 00:09:09,117
plan your day. And so, we knew that this
algorithm was incredibly flawed, would

164
00:09:09,200 --> 00:09:12,797
never scale, and had all these problems,
but you got to start somewhere. You got

165
00:09:12,880 --> 00:09:16,397
to push your baselines to reach the wall
so you actually see the limits of what

166
00:09:16,480 --> 00:09:19,917
good looks like with what you have.
And then you can bring to bear a new

167
00:09:20,000 --> 00:09:23,117
algorithm. And we just kept scaling PO

168
00:09:23,200 --> 00:09:26,957
and we exceeded the performance of the
best humans. And that itself was the

169
00:09:27,040 --> 00:09:30,477
finding, right? That actually massive
compute with simple algorithms,

170
00:09:30,560 --> 00:09:33,997
>> right? that that is something where we
cannot just doesn't just work in theory,

171
00:09:34,080 --> 00:09:37,597
it works in practice. We can really make
it happen. And in this incredibly messy

172
00:09:37,680 --> 00:09:41,357
environment where you cannot program it,
you cannot look ahead. You cannot do a

173
00:09:41,440 --> 00:09:45,677
search. You just need this almost
humanlike intuition. And by the way, the

174
00:09:45,760 --> 00:09:49,917
neural net we used tiny tiny little
insect brain, similar number of synapses

175
00:09:50,000 --> 00:09:54,237
as to truly an insect brain. and you
realize like, wait, what if you had the

176
00:09:54,320 --> 00:09:56,477
same computational

177
00:09:56,560 --> 00:09:58,717
approach but scaled it up to something

178
00:09:58,800 --> 00:10:04,957
that's much more human brain scale? What
would that be like? Very, very evocative

179
00:10:05,040 --> 00:10:07,197
question. >> Is there
a difference between reasoning

180
00:10:07,280 --> 00:10:10,557
and predicting? You mentioned sort
of like predicting the next character,

181
00:10:10,640 --> 00:10:14,877
predicting the next word versus
actually reasoning in first principles.

182
00:10:14,960 --> 00:10:19,277
>> I think they are
connected in a deep way. So

183
00:10:19,360 --> 00:10:23,597
on the one hand, just predicting what
comes next sounds like a pedestrian

184
00:10:23,680 --> 00:10:28,477
task, but if you really can predict
the next word out of Einstein's mouth, you

185
00:10:28,560 --> 00:10:32,237
are at least as smart as Einstein.
And you can make arguments, oh well, like,

186
00:10:32,320 --> 00:10:35,917
you know, it's but I I think that those
arguments fall flat, that there's

187
00:10:36,000 --> 00:10:39,197
something there's something false there
because the point of prediction is not

188
00:10:39,280 --> 00:10:42,317
about being able to predict what is known.
The point is you put yourself in

189
00:10:42,400 --> 00:10:44,557
a new situation you've never seen before

190
00:10:44,640 --> 00:10:47,597
and predict what comes next. And I think

191
00:10:47,680 --> 00:10:50,797
that there's something deeply connected

192
00:10:50,880 --> 00:10:55,357
to intelligence and prediction
that there's a long story of academic

193
00:10:55,440 --> 00:10:58,157
literature and how you think about this
compression. They're all kind of part of

194
00:10:58,240 --> 00:11:02,877
the same thing. Now, these reasoning
models, the thing that I think is very

195
00:11:02,960 --> 00:11:05,837
interesting is that we train them with
reinforcement learning. And so there's

196
00:11:05,920 --> 00:11:08,877
really back to the original OpenAI plan,
there's two steps to it. The first is

197
00:11:08,960 --> 00:11:12,317
unsupervised learning. You train a model
just by having it predict what comes

198
00:11:12,400 --> 00:11:15,837
next. And there it's much more static
data. It's much more observational.

199
00:11:15,920 --> 00:11:19,757
Again, it's data it's never seen before,
situations never seen before, but it is

200
00:11:19,840 --> 00:11:21,597
a situation that has already happened.

201
00:11:21,680 --> 00:11:25,277
Then you do reinforcement learning,
which is you basically have the AI learn

202
00:11:25,360 --> 00:11:29,357
on its own data, right? You have it make
its own here's the action I'm going to

203
00:11:29,440 --> 00:11:33,837
take. You get an observation from the
world and you learn from that. And it's

204
00:11:33,920 --> 00:11:37,357
again the way you actually train it is
still predicting. It's trying to predict

205
00:11:37,440 --> 00:11:40,877
if I take this action, what's the thing
that's likely to happen? And you

206
00:11:40,960 --> 00:11:45,357
reinforce that depending on how good of
a job you did. And the beauty of that is

207
00:11:45,440 --> 00:11:48,797
that it now is an AI that has this
background knowledge and has real world

208
00:11:48,880 --> 00:11:52,797
experience. But fundamentally the
technology that we used to train during

209
00:11:52,880 --> 00:11:56,957
unsupervised stage and during
the reinforcing stage, they're exactly the

210
00:11:57,040 --> 00:12:00,157
same. You are just predicting but you've
changed the structure of the data.

211
00:12:00,240 --> 00:12:04,157
>> When did things start to get tense?
I think the thing about open AI is that if

212
00:12:04,240 --> 00:12:07,197
you truly believe in the mission,
if you truly believe in the possibility of

213
00:12:07,280 --> 00:12:09,277
creating machines that have the

214
00:12:09,360 --> 00:12:14,237
intelligence level of humans, it means
the stakes always feel very high. The

215
00:12:14,320 --> 00:12:16,877
question of who's making the decision,
the question of what are the values that

216
00:12:16,960 --> 00:12:21,037
go into those decisions, the question of
these things that are maybe mundane in a

217
00:12:21,120 --> 00:12:24,637
typical company that are much more like
office politics start to take on this

218
00:12:24,720 --> 00:12:26,637
existential weight. And I think that

219
00:12:26,720 --> 00:12:29,037
that has colored a lot of how open AI

220
00:12:29,120 --> 00:12:32,317
these more high-profile conflicts, you

221
00:12:32,400 --> 00:12:37,117
know, sometimes it's like you put it
in like even just the question of who gets

222
00:12:37,200 --> 00:12:39,997
credit for a particular thing
and suddenly takes on this existential

223
00:12:40,080 --> 00:12:43,997
weight. Well, that's where I was sort of
thinking about this because it's like at

224
00:12:44,080 --> 00:12:48,557
that point you probably realized this
technology is inevitable and it's going

225
00:12:48,640 --> 00:12:52,237
to change the world and that wasn't
broadly known to the world and then I

226
00:12:52,320 --> 00:12:55,677
would imagine there's people who like
I want to be front and center. I want to

227
00:12:55,760 --> 00:12:56,877
take credit for this.

228
00:12:56,960 --> 00:13:00,477
>> Yes, that that is the overwhelming

229
00:13:00,560 --> 00:13:05,197
dynamic that I have observed in this
field. It's not just about open AI

230
00:13:05,280 --> 00:13:09,357
actually. Like one observation I had
early on is that this technology is by

231
00:13:09,440 --> 00:13:13,837
nature very fragmentaryary right that it's
sometimes it you know like when you

232
00:13:13,920 --> 00:13:16,237
have a lot of pressure you can get a

233
00:13:16,320 --> 00:13:21,357
diamond or you can get cracks.
Often you'll see diamonds form in pockets,

234
00:13:21,440 --> 00:13:24,957
right? Teams of people that really work
together, that have a lot of high trust,

235
00:13:25,040 --> 00:13:29,517
that know how to operate, but sometimes
you can see that they that they splinter

236
00:13:29,600 --> 00:13:31,357
off and they kind of go their own way.

237
00:13:31,440 --> 00:13:35,997
And I think within AI, I think we've
gotten some real benefits out of

238
00:13:36,080 --> 00:13:38,877
diversity of approach and different
groups that are really pushing each

239
00:13:38,960 --> 00:13:41,757
other in order to both bring this

240
00:13:41,840 --> 00:13:45,197
technology in a more beneficial way

241
00:13:45,280 --> 00:13:49,277
sometimes how to think about all
the thorny questions around safety around

242
00:13:49,360 --> 00:13:52,797
what does it mean to be safe?
What does it mean to actually deploy this

243
00:13:52,880 --> 00:13:56,557
technology and how to think about how
to mitigate and but also how to maximize

244
00:13:56,640 --> 00:14:00,157
those benefits? And that's something
where I think that there's a lot of very

245
00:14:00,240 --> 00:14:04,157
healthy debate. It's always gone on within
OpenAI's walls. Now it's starting

246
00:14:04,240 --> 00:14:05,757
to really happen, I think, in the world.

247
00:14:05,840 --> 00:14:09,757
And I think that's something that we
as a society really benefit from. The

248
00:14:09,840 --> 00:14:12,237
sponsor of this show is Coin Shares.

249
00:14:12,320 --> 00:14:16,077
While most of the industry was still
arguing about whether digital assets

250
00:14:16,160 --> 00:14:20,077
were legitimate, Coin was quietly
building the infrastructure to invest in

251
00:14:20,160 --> 00:14:24,077
them properly. They now manage over
6 billion in assets and have stayed

252
00:14:24,160 --> 00:14:26,397
profitable through every market cycle.

253
00:14:26,480 --> 00:14:29,677
fully regulated with the kind of
transparency and governance that serious

254
00:14:29,760 --> 00:14:32,077
investors actually expect. Whether it's

255
00:14:32,160 --> 00:14:36,797
crypto ETFs, active strategies,
or Bitcoin mining ETF exposure, you can

256
00:14:36,880 --> 00:14:40,557
access all of it through your existing
brokerage account, coin shares. The

257
00:14:40,640 --> 00:14:45,037
adults have arrived. Learn more
at coinshares.com. This is not investment

258
00:14:45,120 --> 00:14:48,957
advice. [music] You're
in meetings all day. You're

259
00:14:49,040 --> 00:14:52,317
trying to stay present, but you're also
worried you'll forget the decision, the

260
00:14:52,400 --> 00:14:54,957
action item, the important next step.

261
00:14:55,040 --> 00:14:59,117
That's where Granola comes in. Granola
is an AI powered notepad for meetings.

262
00:14:59,200 --> 00:15:02,797
You jot down rough notes like you
always do. And in the background, Granola

263
00:15:02,880 --> 00:15:06,717
transcribes and turns them into clear,
useful notes when the meeting ends.

264
00:15:06,800 --> 00:15:10,637
There are no bots joining your calls.
No distractions, just a clean notepad that

265
00:15:10,720 --> 00:15:14,877
helps you focus. During or after the call,
you can chat with your notes. You

266
00:15:14,960 --> 00:15:18,397
can ask Granola to pull out action
items, help you negotiate, make a

267
00:15:18,480 --> 00:15:22,637
decision, write a follow-up email,
and so much more. I even use it when I'm

268
00:15:22,720 --> 00:15:24,797
listening to podcasts. Once you try it

269
00:15:24,880 --> 00:15:29,757
on a first meeting, it's hard to go
without. Head to granola.ai/shane

270
00:15:29,840 --> 00:15:35,480
and get three months free with the
code Shane. That's granola.ai/shane.

271
00:15:37,360 --> 00:15:42,157
Take me back to the moment you found out
that Sam had been fired. Where were you?

272
00:15:42,240 --> 00:15:44,397
>> I was at home. >> And what happened?

273
00:15:44,480 --> 00:15:47,517
>> I got a text saying, "Can we hop on a

274
00:15:47,600 --> 00:15:49,437
video call?"

275
00:15:49,520 --> 00:15:52,477
So, I hopped on the video call. I

276
00:15:52,560 --> 00:15:58,317
noticed that it was the board
minus Sam who were on there.

277
00:15:58,400 --> 00:16:02,637
>> No, at that point. >> No,
I mean I inferred something was up, but

278
00:16:02,720 --> 00:16:05,277
>> cuz you're on the board.
>> I am on the board or was on the board

279
00:16:05,360 --> 00:16:08,077
>> at this point? >> Yes.
>> And then what happened?

280
00:16:08,160 --> 00:16:12,237
>> I was told that the board has decided

281
00:16:12,320 --> 00:16:15,917
that Sam would be removed and

282
00:16:16,000 --> 00:16:19,677
effectively the message that I got
was the same messaging that was in the

283
00:16:19,760 --> 00:16:21,517
public post.

284
00:16:21,600 --> 00:16:24,557
And I asked if I could have any more

285
00:16:24,640 --> 00:16:27,437
information. I was told no, not right

286
00:16:27,520 --> 00:16:31,437
now. And I pressed on that maybe another

287
00:16:31,520 --> 00:16:36,477
time and again was told nothing more
to share. And then was told wait there's

288
00:16:36,560 --> 00:16:39,837
more. Also that I had
been removed from the

289
00:16:39,920 --> 00:16:42,797
board but would be staying with the

290
00:16:42,880 --> 00:16:47,037
company because I was very critical to

291
00:16:47,120 --> 00:16:49,597
accompanying the mission. I said again I

292
00:16:49,680 --> 00:16:52,477
asked if I get any reasons

293
00:16:52,560 --> 00:16:56,397
any feedback was told no towards the end

294
00:16:56,480 --> 00:16:59,437
was told that hey like in this new setup

295
00:16:59,520 --> 00:17:02,237
that you will start to get hopefully you

296
00:17:02,320 --> 00:17:07,117
can get feedback in this new
configuration and so that was the

297
00:17:07,200 --> 00:17:10,317
conversation.
>> What went through your mind?

298
00:17:10,400 --> 00:17:14,397
>> It just wasn't right. >> Was it anger?

299
00:17:14,480 --> 00:17:16,957
>> No. I felt like I understood what had

300
00:17:17,040 --> 00:17:21,677
happened. How long before you knew what
had actually transpired to sort of cause

301
00:17:21,760 --> 00:17:24,717
this? >> Well, there's
two parts to the answer.

302
00:17:24,800 --> 00:17:29,117
One is I feel like I still am learning
some additional fact, some additional

303
00:17:29,200 --> 00:17:32,237
thing that was in someone's head.
To some extent, it comes down to a lack of

304
00:17:32,320 --> 00:17:36,157
communication, right? That you realize
that there are all these different

305
00:17:36,240 --> 00:17:39,357
things that I buffered. And to some
extent, you know, approximately I kind

306
00:17:39,440 --> 00:17:43,917
of knew I was like, I understand for
every person here. I have a pretty good

307
00:17:44,000 --> 00:17:47,997
model of why they acted the way that
they did. But it wasn't what was most

308
00:17:48,080 --> 00:17:51,197
important to me in the moment. I just
knew that this wasn't right. Right after

309
00:17:51,280 --> 00:17:53,757
I hung up the call, I talked to my wife

310
00:17:53,840 --> 00:17:59,277
and I said, "Got to quit."
And she said, "I agree."

311
00:17:59,360 --> 00:18:01,837
>> And you quit that day. >> Yes.

312
00:18:01,920 --> 00:18:05,917
>> And then what happened >> that day
when I quit? Started to get all

313
00:18:06,000 --> 00:18:08,877
these messages, people saying, "I don't
know what you and Sam are doing next,

314
00:18:08,960 --> 00:18:12,877
but I'm with you. I want to go start
something with you." Like just that was

315
00:18:12,960 --> 00:18:16,717
that was a real honest surprise.
I didn't really expect to get that kind of

316
00:18:16,800 --> 00:18:18,637
support, that kind of outpouring. There

317
00:18:18,720 --> 00:18:21,037
were a few of my close collaborators who

318
00:18:21,120 --> 00:18:23,757
quit that day as well. That's Jakob,

319
00:18:23,840 --> 00:18:27,677
Shimone, Alexander, and the five of us.

320
00:18:27,760 --> 00:18:31,997
So, those people plus Sam, we all got
together and we started to chart out

321
00:18:32,080 --> 00:18:36,477
what a new company could look like.
I remember feeling that first day like,

322
00:18:36,560 --> 00:18:38,637
okay, there's a 10% chance that we

323
00:18:38,720 --> 00:18:43,677
actually get the company back. 10%.
The next day, we set up a meeting at Sam's

324
00:18:43,760 --> 00:18:48,477
house. a bunch of people from the company
came by and we showed the vision

325
00:18:48,560 --> 00:18:51,677
that we'd been sketching out. So, it
was really one day in, you know, this this

326
00:18:51,760 --> 00:18:53,837
fresh picture of how we'd run the

327
00:18:53,920 --> 00:18:57,517
project. And we spent that a bunch of

328
00:18:57,600 --> 00:19:01,597
time over that weekend also negotiating
with the board and the company and

329
00:19:01,680 --> 00:19:04,557
trying to figure out is there a path
back together that makes sense. That

330
00:19:04,640 --> 00:19:08,317
Sunday night, the board replaced me as

331
00:19:08,400 --> 00:19:11,917
interim CEO with a new person and the

332
00:19:12,000 --> 00:19:15,597
company just rebelled.
like we'd actually been at the office.

333
00:19:15,680 --> 00:19:19,437
We thought we were close
to a deal and >> to come back

334
00:19:19,520 --> 00:19:21,037
>> to come back. Yeah. We thought that we

335
00:19:21,120 --> 00:19:24,637
had a path and then the board made that

336
00:19:24,720 --> 00:19:28,877
change and then suddenly it was everyone

337
00:19:28,960 --> 00:19:30,957
streaming out of the building and it was

338
00:19:31,040 --> 00:19:33,837
just like real chaos. I was on video

339
00:19:33,920 --> 00:19:37,997
calls with many of the people who had
been interested in coming to this new

340
00:19:38,080 --> 00:19:42,557
this new company saying it's going to be
okay. we have a plan and we expanded you

341
00:19:42,640 --> 00:19:45,677
know we'd been building this little life
raft right for the small set of people

342
00:19:45,760 --> 00:19:49,677
we expected to want to come and suddenly
it was like no one wanted to be

343
00:19:49,760 --> 00:19:52,397
associated with this this entity right

344
00:19:52,480 --> 00:19:54,637
people wanted to you know stand up for

345
00:19:54,720 --> 00:19:59,757
what they viewed as right Sam talked
to Satia who we'd been talking about hey

346
00:19:59,840 --> 00:20:02,877
could you be a funder could you you know
help help support this new endeavor I

347
00:20:02,960 --> 00:20:07,357
was like hey actually can we expand
from the small life raft to like

348
00:20:07,440 --> 00:20:11,357
yes we take everyone and we're like all
right we'll figure it out somehow and a

349
00:20:11,440 --> 00:20:13,597
lot of people this was right before
Thanksgiving a lot of people were

350
00:20:13,680 --> 00:20:17,597
supposed to be flying to home wherever
that is and instead they canceled their

351
00:20:17,680 --> 00:20:21,517
flights and the office was packed. It's
like everyone was at the office just to

352
00:20:21,600 --> 00:20:24,477
be there, be part of it and just,
you know, even if they couldn't contribute

353
00:20:24,560 --> 00:20:26,637
to any of these conversations that they

354
00:20:26,720 --> 00:20:31,757
just wanted to be there as this history
was made. Then this petition starts to

355
00:20:31,840 --> 00:20:34,877
circulate. So many people were trying
to sign the petition at once. It actually

356
00:20:34,960 --> 00:20:38,157
crashed Google Docs. And so that you
had to have certain people who are

357
00:20:38,240 --> 00:20:41,597
designated as the person you go to to
actually put your name on the document

358
00:20:41,680 --> 00:20:44,397
so you don't don't have too many
editors at once. I think that that was a

359
00:20:44,480 --> 00:20:49,917
statement that that was really
heard loudly. And I remember

360
00:20:50,000 --> 00:20:52,557
You know, I probably got home around
like 5:00 a.m. or something, went to

361
00:20:52,640 --> 00:20:54,237
sleep, and I woke up like 45 minutes

362
00:20:54,320 --> 00:20:57,917
later, and I checked Twitter, and I saw

363
00:20:58,000 --> 00:21:02,797
that Ilia had posted and had
signed the petition, and it said

364
00:21:02,880 --> 00:21:06,317
that he wanted the company to come back
together. And that was this real moment

365
00:21:06,400 --> 00:21:09,677
of relief. I felt so much gratitude that

366
00:21:09,760 --> 00:21:14,877
it just felt like, okay, like, we can
put this back together. We can get back

367
00:21:14,960 --> 00:21:18,717
to a good track. >> You and I
had built this together. What

368
00:21:18,800 --> 00:21:23,117
was it like trying to find your way
back to that relationship after?

369
00:21:23,200 --> 00:21:27,757
>> Look, it was tough. It was definitely
that was definitely a very close

370
00:21:27,840 --> 00:21:30,077
relationship. You'd been the officient

371
00:21:30,160 --> 00:21:33,517
at my civil ceremony, right? We'd been

372
00:21:33,600 --> 00:21:37,117
through extremely tough times together
and like any relationship, you always

373
00:21:37,200 --> 00:21:39,597
have your ups and downs and we spent a

374
00:21:39,680 --> 00:21:41,677
lot of time afterwards really talking

375
00:21:41,760 --> 00:21:44,317
things through and really trying to

376
00:21:44,400 --> 00:21:49,357
understand and just articulate some
of the things that we had let build up or

377
00:21:49,440 --> 00:21:53,677
had left unsaid between us and I think
that we got to through that process I

378
00:21:53,760 --> 00:21:57,517
think we had gotten to a really good
place and for me it was I felt like we

379
00:21:57,600 --> 00:22:01,917
got to closure on on everything that had
happened. How did you feel about all the

380
00:22:02,000 --> 00:22:06,237
loyalty you inspired? >> Deeply grateful.

381
00:22:06,320 --> 00:22:09,677
Truly, it was never something that I
would have asked for, something I never

382
00:22:09,760 --> 00:22:14,157
would have expected. I think the way
that I operate is I'm very much a in the

383
00:22:14,240 --> 00:22:18,317
trenches kind of leader. Try to lead
from the front. And sometimes when I do

384
00:22:18,400 --> 00:22:22,477
that, I don't always, sorry, I'm going
to get a little emotional. Um, but I

385
00:22:22,560 --> 00:22:24,237
don't always look back to see if

386
00:22:24,320 --> 00:22:25,997
everyone's following. I just like run

387
00:22:26,080 --> 00:22:29,517
right in. And when people do when when

388
00:22:29,600 --> 00:22:31,437
people do come and really help to build

389
00:22:31,520 --> 00:22:35,117
the thing, I just it it makes me feel so

390
00:22:35,200 --> 00:22:39,677
grateful for them and to feel like they
have exceeded my expectations in every

391
00:22:39,760 --> 00:22:43,437
way. >> And so eventually
everybody comes back.

392
00:22:43,520 --> 00:22:47,837
>> I'll tell you it was not guaranteed
because throughout that weekend all the

393
00:22:47,920 --> 00:22:52,157
competitors were circling. Just imagine
this like feeding frenzy that people

394
00:22:52,240 --> 00:22:56,397
were shaping up to do. people are
getting offers and we actually did not

395
00:22:56,480 --> 00:22:59,277
lose a single person through that weekend.
No one accepted a competing

396
00:22:59,360 --> 00:23:02,477
offer. I think that's incredible. It

397
00:23:02,560 --> 00:23:04,717
really was. That's more, you know, Coach

398
00:23:04,800 --> 00:23:09,917
Belich told me this actually uh when we
were talking about the best teams and he

399
00:23:10,000 --> 00:23:12,957
said they're not playing for money,
they're playing for the person beside

400
00:23:13,040 --> 00:23:15,677
them. And when you were saying that all

401
00:23:15,760 --> 00:23:19,997
these people quit, it makes me think
of that like and none of them left for

402
00:23:20,080 --> 00:23:22,237
presumably more money, better offers.

403
00:23:22,320 --> 00:23:24,477
Everybody was trying
to circle and poach and

404
00:23:24,560 --> 00:23:28,477
>> yeah, it was a very
that was a diamond moment.

405
00:23:28,560 --> 00:23:31,917
>> After all of this happened,
you took some time off. What was going on

406
00:23:32,000 --> 00:23:36,397
internally with you?
>> That was an intense experience to go

407
00:23:36,480 --> 00:23:38,317
through, an intense experience to come

408
00:23:38,400 --> 00:23:41,597
back to. And

409
00:23:41,680 --> 00:23:43,997
honestly, just

410
00:23:44,080 --> 00:23:45,357
one of the hardest moments for me at

411
00:23:45,440 --> 00:23:50,557
OpenAI was when Ilia left. And it was

412
00:23:50,640 --> 00:23:53,357
maybe the only moment in open eyes
history where I felt like I didn't want

413
00:23:53,440 --> 00:23:55,517
to do it anymore.

414
00:23:55,600 --> 00:23:58,077
Um,

415
00:23:58,160 --> 00:24:01,117
I think I needed some time to kind of

416
00:24:01,200 --> 00:24:05,277
find my way back to remembering like
why I was doing this and why it was so

417
00:24:05,360 --> 00:24:07,197
important and why it was worth the pain.

418
00:24:07,280 --> 00:24:09,437
>> What did you do during the time?

419
00:24:09,520 --> 00:24:13,357
>> I trained language models on >>
That's when you learned how to do it,

420
00:24:13,440 --> 00:24:16,237
right? Like you did the self-study thing.
I read on your blog.

421
00:24:16,320 --> 00:24:18,317
>> Well, no. So, I actually done that.
I actually had done that throughout the

422
00:24:18,400 --> 00:24:21,677
course of OpenAI. Um, so I trained
language models on DNA sequences.

423
00:24:21,760 --> 00:24:24,077
>> Oh wow. >> Yeah.
So I basically got to take for ARC

424
00:24:24,160 --> 00:24:29,197
>> for ARC Institute. Yeah. And it was it
was a very great experience. I took my

425
00:24:29,280 --> 00:24:32,397
skills that I had and applied them as
a very different domain. A domain that's

426
00:24:32,480 --> 00:24:36,157
very personally meaningful to both me
and and to my wife. You know, she has a

427
00:24:36,240 --> 00:24:40,877
lot of health conditions and that we
think about what AI can do for her

428
00:24:40,960 --> 00:24:44,877
health, what it can do for the health
of animals. who we're both very passionate

429
00:24:44,960 --> 00:24:46,797
about just it's like this application

430
00:24:46,880 --> 00:24:50,157
area that felt like maybe we could help

431
00:24:50,240 --> 00:24:53,997
in this very different way from how I've

432
00:24:54,080 --> 00:24:57,277
been pursuing this technology. So
that was that was that was a very you know

433
00:24:57,360 --> 00:24:59,357
positive part of the of the experience.

434
00:24:59,440 --> 00:25:03,117
If I were to say like open a Google
document, write out sort of what you

435
00:25:03,200 --> 00:25:05,517
learned about yourself on one page from

436
00:25:05,600 --> 00:25:09,037
this whole starting to Sam getting

437
00:25:09,120 --> 00:25:12,877
ousted to you quitting to inspiring all
this loyalty to the time off and then

438
00:25:12,960 --> 00:25:16,157
coming back. What would you write?
I think I've just learned to just keep

439
00:25:16,240 --> 00:25:17,917
going for something that's worth it.

440
00:25:18,000 --> 00:25:19,357
Right? If you have a mission that

441
00:25:19,440 --> 00:25:23,757
matters, then the fact of you keep going

442
00:25:23,840 --> 00:25:27,997
through the ups and the downs.
There going to be moments where it's

443
00:25:28,080 --> 00:25:32,077
it's all over. There's moments where
it's we're so back. And you just can't

444
00:25:32,160 --> 00:25:34,077
let those moments pull you off course.

445
00:25:34,160 --> 00:25:36,877
And I think that the degree of just

446
00:25:36,960 --> 00:25:40,317
personal resilience that you have to

447
00:25:40,400 --> 00:25:42,477
grow during these times because if

448
00:25:42,560 --> 00:25:45,517
you're leading, people look to you for

449
00:25:45,600 --> 00:25:50,237
that steadiness, for that support, for
the direction that the whole thing will

450
00:25:50,320 --> 00:25:54,157
go. And I think that a lot
of what I've tried to

451
00:25:54,240 --> 00:25:55,997
grow

452
00:25:56,080 --> 00:25:58,637
with is to really

453
00:25:58,720 --> 00:26:02,717
be able to both understand the details,

454
00:26:02,800 --> 00:26:06,317
right, of what we're doing, what
the implications will be of a choice, but

455
00:26:06,400 --> 00:26:09,677
also be decisive. I think that that
there have been moments where I think

456
00:26:09,760 --> 00:26:14,877
I've been very much approaching OpenAI
through a lens of uncertainty of feeling

457
00:26:14,960 --> 00:26:18,797
like I don't know what the right answer
is. I don't know what the right way to

458
00:26:18,880 --> 00:26:23,197
build this technology is or how do you
answer these very thorny questions, but

459
00:26:23,280 --> 00:26:26,157
there's lots of people here who are very
smart who have very strong opinions and

460
00:26:26,240 --> 00:26:29,037
so really try to understand all those
opinions and figure out how to put them

461
00:26:29,120 --> 00:26:33,117
together. And sometimes that's the right
thing. And sometimes that you realize

462
00:26:33,200 --> 00:26:36,077
that the opinions are mutually
contradictory. They can't all be true at

463
00:26:36,160 --> 00:26:38,957
once. And sometimes you do just have
to pick and you know that that means that

464
00:26:39,040 --> 00:26:41,117
there's going to be someone who's going
to be upset, someone who's going to

465
00:26:41,200 --> 00:26:45,437
quit, someone who's going to feel
slighted. And I think that a lot of what

466
00:26:45,520 --> 00:26:46,877
I've tried to do is have a stronger

467
00:26:46,960 --> 00:26:50,637
sense of self and a stronger sense of

468
00:26:50,720 --> 00:26:53,277
when there is conviction that we need to

469
00:26:53,360 --> 00:26:55,117
act on it. And I think of things that we

470
00:26:55,200 --> 00:27:00,317
have done over the course of OpenAI
where I feel like I wish that we had

471
00:27:00,400 --> 00:27:04,557
done it differently. I think usually
that's of the form we dragged our feet

472
00:27:04,640 --> 00:27:09,037
on something we knew. We knew it wasn't
quite the right person in the role. We

473
00:27:09,120 --> 00:27:11,837
didn't think this was like quite the right
technical direction. and we didn't

474
00:27:11,920 --> 00:27:15,037
think that this way of letting
the projects run was going to quite work,

475
00:27:15,120 --> 00:27:18,477
but we just waited too long. And so

476
00:27:18,560 --> 00:27:21,677
that's something I try to learn from and

477
00:27:21,760 --> 00:27:24,237
an ax in which I try to grow truly every

478
00:27:24,320 --> 00:27:26,957
day. When I reflect on both the course

479
00:27:27,040 --> 00:27:31,597
of OpenAI and Stripe and even rewinding

480
00:27:31,680 --> 00:27:34,317
to college and the projects I worked on

481
00:27:34,400 --> 00:27:36,797
in the past, I think that the way I tend

482
00:27:36,880 --> 00:27:40,557
to operate is that I both really love

483
00:27:40,640 --> 00:27:44,877
the day-to-day activity. I love
the individual contribution. I love the

484
00:27:44,960 --> 00:27:48,477
software. I love the thinking through
the problem. But I also really care

485
00:27:48,560 --> 00:27:53,117
about the environment in which these
things are done. And I actually am

486
00:27:53,200 --> 00:27:57,997
willing to give up on that, you know,
type one fun of just the quick hit, like

487
00:27:58,080 --> 00:27:59,197
you get to build the thing. It's it's

488
00:27:59,280 --> 00:28:01,997
always cool for something that's more

489
00:28:02,080 --> 00:28:07,277
like type two fun of it's painful in
the moment, but it's worthwhile. But that

490
00:28:07,360 --> 00:28:12,157
what you do is you create an environment
where everyone else can get that

491
00:28:12,240 --> 00:28:16,637
do the icy work, do the great the great
thing. And so really trying to build an

492
00:28:16,720 --> 00:28:19,597
environment is something I just
gravitate towards. It's not always the

493
00:28:19,680 --> 00:28:22,717
easiest, right? that you really do have
to be willing to take on great personal

494
00:28:22,800 --> 00:28:27,277
pain. Like in the words of Ilia,
Ilia always says that you have to suffer,

495
00:28:27,360 --> 00:28:30,877
right? If you're not suffering, like
you're not building value. And I think

496
00:28:30,960 --> 00:28:33,117
there's deep truth to it.
>> Double click on that.

497
00:28:33,200 --> 00:28:38,157
>> The Ilia perspective I think on it's
it's funny because he has a particular

498
00:28:38,240 --> 00:28:39,437
way of talking that I think is very

499
00:28:39,520 --> 00:28:42,877
unique to him and that there's always

500
00:28:42,960 --> 00:28:46,077
deep inspiration in the words that he

501
00:28:46,160 --> 00:28:50,157
chooses. And this picture of suffering
was something that that we thought about

502
00:28:50,240 --> 00:28:52,877
throughout the course of OpenAI where
it's like we had so much uncertainty

503
00:28:52,960 --> 00:28:57,757
from the beginning. Is this thing going
to work? And there's many reasons why it

504
00:28:57,840 --> 00:29:01,517
might not work, why it should not work,
why you could even say it cannot work,

505
00:29:01,600 --> 00:29:05,037
right? Whether it's how do you get
the people? How do you pursue the

506
00:29:05,120 --> 00:29:08,877
technology? Like how do you get
enough capital? How do you keep people

507
00:29:08,960 --> 00:29:11,517
motivated? Like how do you make
the right decisions? Like each of these

508
00:29:11,600 --> 00:29:15,517
things is extremely hard, extremely
uncertain. And it's easy to just sweep

509
00:29:15,600 --> 00:29:19,757
the problems under the rug and just
blindly say go. And I think that is the

510
00:29:19,840 --> 00:29:22,237
negative side of like Silicon Valley

511
00:29:22,320 --> 00:29:26,877
culture, right? Certainly Silicon
Valley perception, right? It's just the

512
00:29:26,960 --> 00:29:29,997
you just kind of blindly do the thing
and you kind of do a reality distortion,

513
00:29:30,080 --> 00:29:34,077
whatever it is. But I don't think that
works in AI and I don't think that works

514
00:29:34,160 --> 00:29:37,197
for open AI. I don't think that's how
we've operated ever. I think the way

515
00:29:37,280 --> 00:29:39,437
that we have always operated is to say

516
00:29:39,520 --> 00:29:42,637
encounter the hard truth understand

517
00:29:42,720 --> 00:29:45,357
science the reality as it is and that is

518
00:29:45,440 --> 00:29:46,877
I think something that has contributed

519
00:29:46,960 --> 00:29:50,717
to the successes we have had of thinking

520
00:29:50,800 --> 00:29:55,117
about the problems differently of not
being happy with even in the early days

521
00:29:55,200 --> 00:29:56,797
we're thinking about okay if we just

522
00:29:56,880 --> 00:30:00,317
write some papers and publish them it'll

523
00:30:00,400 --> 00:30:03,997
be great we'll get citations we'll get
you know be the coolest people at these

524
00:30:04,080 --> 00:30:06,877
conferences but will
we achieve the mission

525
00:30:06,960 --> 00:30:11,277
like how is it that you do that activity
and then AGI goes better for the world?

526
00:30:11,360 --> 00:30:14,157
They're not connected, right? It's not
enough. Maybe it's a foundation, maybe

527
00:30:14,240 --> 00:30:16,317
it's a step, but it is not sufficient.

528
00:30:16,400 --> 00:30:19,677
And so then you start really thinking
about these bigger picture questions of

529
00:30:19,760 --> 00:30:23,917
what it would what would it take
to build an AGI and not pleasant, right?

530
00:30:24,000 --> 00:30:27,917
Because you realize there's no path. You
realize you need dollars. You don't have

531
00:30:28,000 --> 00:30:30,877
any mechanism that's going to allow you
to raise dollars. And you can try hard.

532
00:30:30,960 --> 00:30:34,717
We did try hard. We tried extremely hard.
But, you know, maybe you're

533
00:30:34,800 --> 00:30:37,037
raising $100 million, you could do $500

534
00:30:37,120 --> 00:30:41,757
million, great. But a billion, pretty
hard. And you look at what OpenAI has

535
00:30:41,840 --> 00:30:45,917
been able to accomplish with the
resources we have been able to raise to

536
00:30:46,000 --> 00:30:50,957
further that mission, there truly would
be no other way to do it besides having

537
00:30:51,040 --> 00:30:53,917
leaned into the suffering and trying to
understand the truth of what it is we're

538
00:30:54,000 --> 00:30:57,917
trying to accomplish. >> What's a lesson
you've had to learn more than once?

539
00:30:58,000 --> 00:31:02,157
>> Make the hard decision. Have the hard
conversation. What's the best advice

540
00:31:02,240 --> 00:31:03,517
you've ever been given?

541
00:31:03,600 --> 00:31:05,197
>> I would actually say it was from my

542
00:31:05,280 --> 00:31:09,517
Harvard freshman uh writing class of

543
00:31:09,600 --> 00:31:14,157
just keep cutting words in order
to be clear and communicate well.

544
00:31:14,240 --> 00:31:15,997
>> How do you filter information?

545
00:31:16,080 --> 00:31:18,797
>> I read a lot triage aggressively.

546
00:31:18,880 --> 00:31:21,117
>> Who are your role models and why? I

547
00:31:21,200 --> 00:31:26,077
would say Gaus and Decart as

548
00:31:26,160 --> 00:31:29,677
people who are incredibly thoughtful,
very much ahead of their time, very much

549
00:31:29,760 --> 00:31:33,757
visionaries who came up with real
breakthroughs that I think transform how

550
00:31:33,840 --> 00:31:37,597
we think and and how we live.
>> What do you want non- tech people to

551
00:31:37,680 --> 00:31:40,877
know about AI? >> That it's
going to be a force for good

552
00:31:40,960 --> 00:31:42,717
in their personal life that they'll

553
00:31:42,800 --> 00:31:46,397
benefit from and will help advance

554
00:31:46,480 --> 00:31:51,837
science, medicine, and really lift up
everyone. What does the world get wrong

555
00:31:51,920 --> 00:31:53,277
about Greg Brockman?

556
00:31:53,360 --> 00:31:56,317
>> I think people don't understand how

557
00:31:56,400 --> 00:32:01,117
focused I am on this mission in a way

558
00:32:01,200 --> 00:32:04,477
that I think has been

559
00:32:04,560 --> 00:32:08,717
very personally painful at many turns.

560
00:32:08,800 --> 00:32:10,477
But I just believe this technology can

561
00:32:10,560 --> 00:32:12,637
just help

562
00:32:12,720 --> 00:32:15,997
empower people and benefit everyone. And

563
00:32:16,080 --> 00:32:17,597
I really want to help make that happen.

564
00:32:17,680 --> 00:32:20,707
Why is OpenAI so bad at naming models?

565
00:32:20,790 --> 00:32:23,837
[sighs and gasps]
That one I can't tell you.

566
00:32:23,920 --> 00:32:27,037
>> Are we near the point where AI makes AI

567
00:32:27,120 --> 00:32:32,077
go parabolic? >> I would say
we are in this phase where

568
00:32:32,160 --> 00:32:33,757
you apply AI to its own development

569
00:32:33,840 --> 00:32:35,917
process and it's going to go faster and

570
00:32:36,000 --> 00:32:40,077
faster. And that is something that's

571
00:32:40,160 --> 00:32:44,477
been happening really I mean certainly
since chat GBT in many ways right we

572
00:32:44,560 --> 00:32:47,037
used chat GBT to make our development

573
00:32:47,120 --> 00:32:52,077
process 10% 20% faster
now we have these amazing

574
00:32:52,160 --> 00:32:56,237
coding tools which have truly
revolutionized how software engineering

575
00:32:56,320 --> 00:32:58,637
is done and most of what we do in the

576
00:32:58,720 --> 00:33:03,757
production of models is bottlenecked
by software. It's about implementing these

577
00:33:03,840 --> 00:33:05,437
systems. It's about scaling them up.

578
00:33:05,520 --> 00:33:08,477
It's about managing
these massive computers

579
00:33:08,560 --> 00:33:12,397
and we're going to be hitting a phase
soon where the AI will also come up with

580
00:33:12,480 --> 00:33:14,157
its own research ideas and test those

581
00:33:14,240 --> 00:33:16,637
out, run experiments. And so I think

582
00:33:16,720 --> 00:33:20,717
that the speed of iteration and

583
00:33:20,800 --> 00:33:25,677
innovation is going to continue
to increase as a result of what we're

584
00:33:25,760 --> 00:33:29,357
producing. >> What percentage
of the code is now written by AI?

585
00:33:29,440 --> 00:33:32,317
>> It's hard to know what percent
of the code is not written by AI. It's a

586
00:33:32,400 --> 00:33:35,037
vanishing fraction. the actual writing

587
00:33:35,120 --> 00:33:38,317
of code. Currently, the AI is much

588
00:33:38,400 --> 00:33:41,917
better than humans at writing code given

589
00:33:42,000 --> 00:33:44,477
the right context, given the right

590
00:33:44,560 --> 00:33:46,717
structure. Now, there's parts of the

591
00:33:46,800 --> 00:33:52,077
actual structure of the code that our
human experts still are much better at,

592
00:33:52,160 --> 00:33:55,517
right? that that's about thinking about
how the module should be laid out, how

593
00:33:55,600 --> 00:33:59,597
the pieces should work, maybe
the definition of certain kinds of of

594
00:33:59,680 --> 00:34:01,917
interfaces, but the actual writing of

595
00:34:02,000 --> 00:34:04,397
code is essentially all AI now.

596
00:34:04,480 --> 00:34:08,237
>> Is it coming up with novel ideas
that you wouldn't have thought of?

597
00:34:08,320 --> 00:34:12,797
>> I'd say that where we are is we're
getting close. So we've seen for example

598
00:34:12,880 --> 00:34:14,477
in chip design. So in the design of our

599
00:34:14,560 --> 00:34:17,517
own chip last year we applied our

600
00:34:17,600 --> 00:34:21,997
technology to trying to get a better fit

601
00:34:22,080 --> 00:34:24,797
to actually shrink the area used by the

602
00:34:24,880 --> 00:34:28,317
the circuits and there we found that the

603
00:34:28,400 --> 00:34:31,917
optimizations that the model produced
were actually on our list. So it didn't

604
00:34:32,000 --> 00:34:35,197
come with something novel and new
that no human ever would have but it

605
00:34:35,280 --> 00:34:37,037
implemented it faster in a way that we

606
00:34:37,120 --> 00:34:39,277
wouldn't have had time to accomplish. If

607
00:34:39,360 --> 00:34:42,637
you look at math and physics, we now are

608
00:34:42,720 --> 00:34:47,197
solving open math problems.
We're solving open physics problems and

609
00:34:47,280 --> 00:34:50,397
actually have resolved this particular
physics problem recently in quantum

610
00:34:50,480 --> 00:34:54,877
physics in the opposite way that the
community expected and with a you know

611
00:34:54,960 --> 00:34:56,877
beautiful elegant formula. It's like

612
00:34:56,960 --> 00:34:59,117
it's really happening. So new ideas from

613
00:34:59,200 --> 00:35:02,877
these models extremely doable. We're

614
00:35:02,960 --> 00:35:06,397
starting to see it in some of these
domains. now applying it in harder and

615
00:35:06,480 --> 00:35:11,117
harder domains or ones that require more
real world context and things like that.

616
00:35:11,200 --> 00:35:14,557
We're starting to see it. We have a line
of sight for how to accomplish it, but

617
00:35:14,640 --> 00:35:19,037
we have a lot of work to do. Why do
models feel like they have a political

618
00:35:19,120 --> 00:35:21,677
leaning to them,
like a political bias almost?

619
00:35:21,760 --> 00:35:24,957
>> So, we put a lot of effort into
neutrality for our models to have them

620
00:35:25,040 --> 00:35:28,557
represent truth. And you can see exactly

621
00:35:28,640 --> 00:35:32,237
the values that go into our models
on our website. We have a publicly

622
00:35:32,320 --> 00:35:35,997
published spec which defines and you can
give feedback on the different ways we

623
00:35:36,080 --> 00:35:38,237
want our model to behave. We've spent a

624
00:35:38,320 --> 00:35:43,037
lot of effort to really get to this
mutual point of view and trying to be

625
00:35:43,120 --> 00:35:46,157
fair and balanced. And I think that
sometimes when you see these screenshots

626
00:35:46,240 --> 00:35:49,277
on Twitter that they're not always fully

627
00:35:49,360 --> 00:35:51,437
honest themselves in terms of where they

628
00:35:51,520 --> 00:35:53,597
came from, either because there's some

629
00:35:53,680 --> 00:35:56,717
memories that are behind the scenes that

630
00:35:56,800 --> 00:36:00,717
tweak the answer in a certain way or
hidden instructions or previous parts of

631
00:36:00,800 --> 00:36:04,557
the conversation. And so sometimes it's
also there's just no right answer. And

632
00:36:04,640 --> 00:36:07,757
so you you know you can have like
a question you say answer in one word and

633
00:36:07,840 --> 00:36:11,197
no matter which one you say you're going
to get some sort of claim of bias. And

634
00:36:11,280 --> 00:36:13,437
so I think that to some extent the core

635
00:36:13,520 --> 00:36:17,197
of it in my mind is that we are yeah we

636
00:36:17,280 --> 00:36:19,437
care a lot about truth and about having

637
00:36:19,520 --> 00:36:21,917
an AI that really represents you.

638
00:36:22,000 --> 00:36:26,237
>> Do you think the models evolve to tell
us what we want to hear if they're based

639
00:36:26,320 --> 00:36:29,597
on reinforcement learning? So if I lean
left, it's going to tell me an answer

640
00:36:29,680 --> 00:36:33,117
that leans left or if I lean right,
it's going to give me an answer that leans

641
00:36:33,200 --> 00:36:34,397
right. Well, so we've actually gone

642
00:36:34,480 --> 00:36:38,317
through an evolution of how we train the

643
00:36:38,400 --> 00:36:42,317
models to user preferences and that

644
00:36:42,400 --> 00:36:46,717
we've seen that at one point like last
year that the models really did start to

645
00:36:46,800 --> 00:36:48,637
lean into telling you what you wanted to

646
00:36:48,720 --> 00:36:49,997
hear, saying, "Oh, that was such a great

647
00:36:50,080 --> 00:36:53,997
answer." And we reacted to that. We said

648
00:36:54,080 --> 00:36:57,917
that this is not how we want our models
to operate. And we made changes because

649
00:36:58,000 --> 00:37:01,197
the true thing we want the models to be
aligned to is helping you solve your

650
00:37:01,280 --> 00:37:05,357
goals, your long-term goals. Right? And
maybe in the moment it feels good to be

651
00:37:05,440 --> 00:37:09,517
told that was a great question, best
question anyone's ever asked, but that's

652
00:37:09,600 --> 00:37:13,917
not what you actually want. Maybe there
are some people, but uh it's not it's

653
00:37:14,000 --> 00:37:16,237
not what most people truly want. And so

654
00:37:16,320 --> 00:37:18,397
we've actually made great technological

655
00:37:18,480 --> 00:37:23,517
improvements to make sure that our AI
training does not result in what is

656
00:37:23,600 --> 00:37:26,157
called hacking the greater, right? That we
really want to make sure that there

657
00:37:26,240 --> 00:37:31,357
is a good signal there that is about
the goal, not just your short term. what's

658
00:37:31,440 --> 00:37:35,757
going to get you a quick hit.
And that to me is maybe the most important

659
00:37:35,840 --> 00:37:40,397
part of the vision for where
our personal AI, personal AGI is going to

660
00:37:40,480 --> 00:37:44,397
take us is to really make sure it's not
just about something that looks good in

661
00:37:44,480 --> 00:37:47,677
the moment. It's really about alignment
with your long-term well-being, your

662
00:37:47,760 --> 00:37:50,797
long-term goals, the thing that you
actually want. And that is what I think

663
00:37:50,880 --> 00:37:53,677
will most empower people, right?
Is really put you in the driver's seat

664
00:37:53,760 --> 00:37:57,997
because you will have this entity that
is there operating on your behalf 24/7,

665
00:37:58,080 --> 00:38:01,197
right? You're asleep. it's out there
trying to figure out what is it that

666
00:38:01,280 --> 00:38:04,477
Shane wants, how can I do it better and

667
00:38:04,560 --> 00:38:06,317
is actually able to accomplish it.

668
00:38:06,400 --> 00:38:08,477
>> Are we in a global AI race?

669
00:38:08,560 --> 00:38:11,837
>> I think we are certainly in a global
AI renaissance and I think that the

670
00:38:11,920 --> 00:38:14,077
dynamics between countries are not yet

671
00:38:14,160 --> 00:38:15,997
fully defined. We have this

672
00:38:16,080 --> 00:38:19,437
concentration of where the breakthrough

673
00:38:19,520 --> 00:38:23,917
algorithms come from in the US in

674
00:38:24,000 --> 00:38:26,637
Western companies. There's clearly a lot

675
00:38:26,720 --> 00:38:28,717
of innovation happening around the

676
00:38:28,800 --> 00:38:31,437
world, but I think exactly the balance

677
00:38:31,520 --> 00:38:35,757
of dynamics and how like which countries

678
00:38:35,840 --> 00:38:39,437
re rely on which providers, all of that
is something that I think is still being

679
00:38:39,520 --> 00:38:41,997
determined. >> Is there
a consequence do you think for

680
00:38:42,080 --> 00:38:44,637
the United States not being the first

681
00:38:44,720 --> 00:38:47,597
country to reach AGI? Well, I do think

682
00:38:47,680 --> 00:38:50,637
that leading an AI is very critical for

683
00:38:50,720 --> 00:38:53,757
America because I think that this is how

684
00:38:53,840 --> 00:38:55,757
you can ensure that democratic values

685
00:38:55,840 --> 00:38:59,357
are protected and preserved. And I think

686
00:38:59,440 --> 00:39:03,757
that every country is also starting
to realize that they need some sort of

687
00:39:03,840 --> 00:39:08,317
sovereign AI strategy. They need to if
this is becoming the basis of economic

688
00:39:08,400 --> 00:39:10,557
security, national security, they need

689
00:39:10,640 --> 00:39:14,077
to participate somehow. And

690
00:39:14,160 --> 00:39:16,557
if you look at a lot of the efforts by

691
00:39:16,640 --> 00:39:19,837
the United States to think about how to

692
00:39:19,920 --> 00:39:22,237
manage chip exports, how to think about

693
00:39:22,320 --> 00:39:25,277
technology exports, there's something

694
00:39:25,360 --> 00:39:28,157
where if you lean too far out, then

695
00:39:28,240 --> 00:39:32,877
everyone else has to develop
their own competitor or rely on

696
00:39:32,960 --> 00:39:35,117
someone else who's who's building this.

697
00:39:35,200 --> 00:39:39,357
If you lean too far in, then maybe you
lose your advantage. And the question

698
00:39:39,440 --> 00:39:43,357
is, how do you balance those? How do you
maintain your leadership? But leadership

699
00:39:43,440 --> 00:39:45,357
is not just about being ahead.

700
00:39:45,440 --> 00:39:49,197
Leadership is about also
bringing along the world with you.

701
00:39:49,280 --> 00:39:52,477
>> Are other countries stealing
advancements? I've been reading a lot

702
00:39:52,560 --> 00:39:55,837
about distillation. >> There's
certainly a lot of attempts to

703
00:39:55,920 --> 00:40:00,477
distill models and that comes from
companies in the US. It comes from uh

704
00:40:00,560 --> 00:40:04,797
from all over the world. But I think
that it misses the core point which is

705
00:40:04,880 --> 00:40:09,597
that the way this technology
is developing is it is on an exponential

706
00:40:09,680 --> 00:40:13,437
and anytime we have a model we've
already moved on to the next one. We're

707
00:40:13,520 --> 00:40:17,437
already moving to the next level.
So we put in a lot of effort to protect

708
00:40:17,520 --> 00:40:21,117
against distillation make it harder
to do especially with things like chain of

709
00:40:21,200 --> 00:40:25,277
thought and other parts of the model
that are not really necessary to get the

710
00:40:25,360 --> 00:40:26,797
benefits to someone to get the outputs

711
00:40:26,880 --> 00:40:30,957
to someone. But that the core advantage

712
00:40:31,040 --> 00:40:32,877
that we have, the strength that we're

713
00:40:32,960 --> 00:40:38,077
building up over time is really about
not just any one model. It's about the

714
00:40:38,160 --> 00:40:39,437
machine that makes the models.

715
00:40:39,520 --> 00:40:41,997
>> Oh, is that why you guys
stopped showing reasoning?

716
00:40:42,080 --> 00:40:45,757
>> That is part of it. So there's
two reasons. One is to think about

717
00:40:45,840 --> 00:40:50,717
distillation. But the second in some
ways more important is that we had this

718
00:40:50,800 --> 00:40:52,397
insight when we first developed the

719
00:40:52,480 --> 00:40:57,517
reasoning paradigm that it gives us a

720
00:40:57,600 --> 00:41:01,277
interpretability mechanism we had not
been anticipating because you can really

721
00:41:01,360 --> 00:41:03,917
read the model's thoughts. You can see

722
00:41:04,000 --> 00:41:05,757
exactly how it got to an answer. So you

723
00:41:05,840 --> 00:41:08,957
can interpret how like what was actually

724
00:41:09,040 --> 00:41:11,037
motivating that answer. Now the problem

725
00:41:11,120 --> 00:41:13,757
is if you train the model to have a

726
00:41:13,840 --> 00:41:18,717
chain of thought that looks good then
you lose all the faithfulness right it's

727
00:41:18,800 --> 00:41:21,997
just going to be like the model knows
that part of the answer that that is

728
00:41:22,080 --> 00:41:26,317
desired is for the chain of thought
to look a certain way and so it may not be

729
00:41:26,400 --> 00:41:29,357
representative of how it actually
arrived at that answer anymore and so we

730
00:41:29,440 --> 00:41:33,517
were we made an early decision to say
we want to avoid any temptation to train

731
00:41:33,600 --> 00:41:37,597
these chain of thoughts to look
favorable to look like something you

732
00:41:37,680 --> 00:41:41,197
could present to a user and so that
really made us lean out for multiple

733
00:41:41,280 --> 00:41:45,677
reasons, for competitive reasons, for
safety reasons from the idea of showing

734
00:41:45,760 --> 00:41:48,797
these intermediate thoughts.
>> It seems like the current trend right

735
00:41:48,880 --> 00:41:52,797
now is to release preview models. Is
that because we're computer constrained?

736
00:41:52,880 --> 00:41:53,597
Do you think?

737
00:41:53,680 --> 00:41:57,037
>> I would say that we in general are

738
00:41:57,120 --> 00:42:00,077
heading to a computed world. Like if you

739
00:42:00,160 --> 00:42:04,237
think about the amount of value that
these models can produce for someone,

740
00:42:04,320 --> 00:42:07,997
it's extreme, right? It's not just
answering a quick question anymore. It's

741
00:42:08,080 --> 00:42:12,077
not just even answering, you know,
giving you access to health information.

742
00:42:12,160 --> 00:42:16,797
It's really going deep and spending
a lot of tokens to put together a bunch of

743
00:42:16,880 --> 00:42:20,237
different data sources, search through
your enterprise knowledge base to

744
00:42:20,320 --> 00:42:24,077
actually be able to solve this hard
problem to write that software that

745
00:42:24,160 --> 00:42:27,757
that's better than a human would be
able to. All of that is something that is

746
00:42:27,840 --> 00:42:29,997
like hard. And if you look at the

747
00:42:30,080 --> 00:42:33,037
progress that we made between GBD5 to 51

748
00:42:33,120 --> 00:42:36,877
to 52 to 53 codecs to 54, it's been

749
00:42:36,960 --> 00:42:40,157
extreme. And these models
are getting extremely

750
00:42:40,240 --> 00:42:44,157
better at understanding your intent
molding to what you want to accomplish.

751
00:42:44,240 --> 00:42:47,917
And we also put them in these surfaces
like codecs that make them very usable

752
00:42:48,000 --> 00:42:51,517
so that you as a developer, you can
really fly, right? That you can achieve

753
00:42:51,600 --> 00:42:56,077
more than you would have dreamed
otherwise. And all this is powered by

754
00:42:56,160 --> 00:43:01,037
compute fundamentally. And there's not
enough comput. that if you just wanted

755
00:43:01,120 --> 00:43:04,877
enough compute for you know you wanted
one GPU for every person in the world

756
00:43:04,960 --> 00:43:08,797
you're talking like 8 billion GPUs
we are not on a trajectory to build

757
00:43:08,880 --> 00:43:12,637
anywhere near that level of compute
right it's like you know hundreds of

758
00:43:12,720 --> 00:43:16,317
thousands of GPUs like that's a pretty
big fleet these days millions of GPUs

759
00:43:16,400 --> 00:43:19,357
coming up it's not surprising that there
is too little compute in the world and

760
00:43:19,440 --> 00:43:22,077
that we're going to need much more
in order to really be able to bring this

761
00:43:22,160 --> 00:43:26,157
technology to everyone and then in terms
of the training I'd say that the way

762
00:43:26,240 --> 00:43:29,997
that we tend to launch things so we have
put in a lot of effort to make sure we

763
00:43:30,080 --> 00:43:33,997
are building compute in anticipation of
what we see coming. And so I think we're

764
00:43:34,080 --> 00:43:37,357
going to be very focused on our mission
of bringing these models to everyone,

765
00:43:37,440 --> 00:43:39,197
making them widely available.

766
00:43:39,280 --> 00:43:41,677
>> You guys were teased
for putting so much

767
00:43:41,760 --> 00:43:47,117
effort, money into data centers. How do
you think that's playing out now? Well,

768
00:43:47,200 --> 00:43:50,557
I think it's going to give us
an advantage and I think it's going to be

769
00:43:50,640 --> 00:43:53,437
something that's an advantage not
just for the business, but for actually

770
00:43:53,520 --> 00:43:57,197
delivering on the mission of bringing
this technology to everyone cuz you guys

771
00:43:57,280 --> 00:44:01,437
like you saw that way in advance. You
get teased for it by almost all of your

772
00:44:01,520 --> 00:44:04,557
competitors. >> Mhm. Who's laughing now?

773
00:44:04,640 --> 00:44:08,237
>> Yeah. [snorts] I mean,
I I think our competitors are

774
00:44:08,320 --> 00:44:10,317
not having a good time on compute,
let me put it that way.

775
00:44:10,400 --> 00:44:11,917
>> But you must have seen something that

776
00:44:12,000 --> 00:44:17,357
they didn't see. Like that's the I mean
everybody was in a very similar or it

777
00:44:17,440 --> 00:44:19,837
seems at least from the outside
everybody was in a very similar

778
00:44:19,920 --> 00:44:22,317
technological space. They all knew this

779
00:44:22,400 --> 00:44:25,517
was coming and yet you guys had the

780
00:44:25,600 --> 00:44:29,037
boldness to make that bet with a hundred

781
00:44:29,120 --> 00:44:30,877
billion dollars like but that is the

782
00:44:30,960 --> 00:44:34,637
core of open AI is really encountering

783
00:44:34,720 --> 00:44:38,957
reality as it is really thinking about
what is the implication of what it is

784
00:44:39,040 --> 00:44:43,277
we'll accomplish in the next 6 months
the next 12 months the next 10 years.

785
00:44:43,360 --> 00:44:45,757
And that is true for the grand mission.

786
00:44:45,840 --> 00:44:49,037
It's true for day-to-day how we design

787
00:44:49,120 --> 00:44:51,197
different pieces of our software. And

788
00:44:51,280 --> 00:44:53,517
it's true for things like scaling up

789
00:44:53,600 --> 00:44:58,637
compute. And I think that we are deeply
motivated by bringing this technology to

790
00:44:58,720 --> 00:45:01,517
everyone. And we think about lots
of different mechanisms for how to do that

791
00:45:01,600 --> 00:45:06,317
well and safely.
Smart operators look for leverage. Video

792
00:45:06,400 --> 00:45:08,317
is the most effective way to

793
00:45:08,400 --> 00:45:10,237
communicate, but it's always been too

794
00:45:10,320 --> 00:45:13,437
slow and expensive to produce. Hey Genen

795
00:45:13,520 --> 00:45:16,077
eliminates that. You go from idea to

796
00:45:16,160 --> 00:45:18,237
professional video in minutes. No

797
00:45:18,320 --> 00:45:21,117
camera, no crew. Their avatars are rated

798
00:45:21,200 --> 00:45:24,157
the most realistic on G2 and over 30

799
00:45:24,240 --> 00:45:26,317
million people are already using Hey

800
00:45:26,400 --> 00:45:29,987
Gen, including 85% of the Fortune 100.

801
00:45:30,070 --> 00:45:31,837
[snorts] You can even reach an audience

802
00:45:31,920 --> 00:45:35,757
in over 175 languages with AI lips sync

803
00:45:35,840 --> 00:45:39,837
and translation. Same voice perfectly
matched. If you're someone who

804
00:45:39,920 --> 00:45:42,397
communicates at scale, take a look. Your

805
00:45:42,480 --> 00:45:48,680
first three videos are free.
Hey Jen.com hyge.com.

806
00:45:49,200 --> 00:45:53,757
Ever hit 3 p.m. and feel like your brain
just quit? For a lot of people, that's

807
00:45:53,840 --> 00:45:56,077
not caffeine or sleep. It's

808
00:45:56,160 --> 00:45:59,197
electrolytes. And water alone won't fix

809
00:45:59,280 --> 00:46:02,157
it. That's why I drink Element every day

810
00:46:02,240 --> 00:46:05,037
after lunch. Zero sugar, no dodgy

811
00:46:05,120 --> 00:46:07,517
ingredients, just a real dose of sodium,

812
00:46:07,600 --> 00:46:12,317
potassium, and magnesium. I know
you're all thinking electrolytes are for

813
00:46:12,400 --> 00:46:16,157
athletes, but you don't have to be
an athlete to benefit from it. And it

814
00:46:16,240 --> 00:46:20,877
tastes great. Stay sharp in the afternoon
and grab a free 8count sample

815
00:46:20,960 --> 00:46:22,637
pack with any purchase at

816
00:46:22,720 --> 00:46:25,597
drinklement.com/tkp.

817
00:46:25,680 --> 00:46:30,280
That's drinklm nnt.com/tkp.

818
00:46:32,560 --> 00:46:36,477
Do you think data centers eventually get
dedicated towards a problem like you'll

819
00:46:36,560 --> 00:46:40,077
have a huge data center in North Dakota
and it's just on solving cancer and

820
00:46:40,160 --> 00:46:42,237
that's all it's doing? >> Yes.

821
00:46:42,320 --> 00:46:46,477
>> How far away are we from that? I think
that this kind of thing happening this

822
00:46:46,560 --> 00:46:50,237
year is not out of the question and it's
really amazing if you think about it

823
00:46:50,320 --> 00:46:54,957
having this giant machine right and have
you have you been to any of these data

824
00:46:55,040 --> 00:46:57,837
centers? >> No, I've seen them
online but never in person.

825
00:46:57,920 --> 00:47:00,957
>> It is a very different experience to

826
00:47:01,040 --> 00:47:04,077
walk amongst these racks, right? walk

827
00:47:04,160 --> 00:47:08,077
down the the rows and you look at the
cables, they're all perfectly exactly

828
00:47:08,160 --> 00:47:09,757
the right the right length and you just

829
00:47:09,840 --> 00:47:14,717
realize that what a data center is. It's

830
00:47:14,800 --> 00:47:18,717
a massive machine. These are maybe
the biggest machines that humanity creates

831
00:47:18,800 --> 00:47:20,157
and then you ask the question of why.

832
00:47:20,240 --> 00:47:25,597
Why do we build these machines? Why
is it worthwhile? And it is because they

833
00:47:25,680 --> 00:47:29,597
have the potential to solve problems
that matter for people, right? to solve,

834
00:47:29,680 --> 00:47:34,877
you know, come up with cures for cancer,
to help people run businesses, to, you

835
00:47:34,960 --> 00:47:36,077
know, sometimes maybe it's mundane

836
00:47:36,160 --> 00:47:38,557
queries. The purpose in my mind is

837
00:47:38,640 --> 00:47:42,157
really about how do you deliver value?

838
00:47:42,240 --> 00:47:44,637
How do you deliver on people's goals?

839
00:47:44,720 --> 00:47:47,277
And I think the opportunity presented by

840
00:47:47,360 --> 00:47:51,757
these massive machines targeting one
problem is something we have not yet

841
00:47:51,840 --> 00:47:54,717
really internalized. >> But if
we're computer constraint, like

842
00:47:54,800 --> 00:47:58,237
how do you decide who to serve? like why
are you serving me when I'm like trying

843
00:47:58,320 --> 00:48:01,037
to make an image over like solving cancer?

844
00:48:01,120 --> 00:48:04,557
>> Well, this is going to be the most
important question for society to

845
00:48:04,640 --> 00:48:09,757
answer. Where does the compute go? What
problems are worthy? And there's lots of

846
00:48:09,840 --> 00:48:12,477
worthy problems, but you need to
prioritize them because you only have so

847
00:48:12,560 --> 00:48:16,077
much compute. And so, one thing we
really believe in is that everyone is

848
00:48:16,160 --> 00:48:19,677
going to need access to compute.
And so, that's why we have a free tier of

849
00:48:19,760 --> 00:48:24,397
chatbt. We've really put effort into
making sure that people are able to use

850
00:48:24,480 --> 00:48:27,597
this technology that it's widely
available because we believe that is

851
00:48:27,680 --> 00:48:31,437
core to what we're doing here. We think
that putting this technology in people's

852
00:48:31,520 --> 00:48:34,637
hands that empowers them that lets
them achieve goals. It helps them also

853
00:48:34,720 --> 00:48:37,837
understand the technology, right? It's
something that helps them then shape how

854
00:48:37,920 --> 00:48:41,517
does this technology slot in. You could
take a very different approach and say,

855
00:48:41,600 --> 00:48:43,197
well, it's all about the ivory tower.

856
00:48:43,280 --> 00:48:47,117
It's all about the just solve the problem
and we will then distribute the

857
00:48:47,200 --> 00:48:49,117
technology breakthroughs in some way.

858
00:48:49,200 --> 00:48:51,997
And I think there's merit to

859
00:48:52,080 --> 00:48:54,797
that as well, but that's not where I'd

860
00:48:54,880 --> 00:48:59,517
put the the balance of of what we do.
I think that that is very much a like we

861
00:48:59,600 --> 00:49:03,757
do want to make great strides on specific
problems, but I think that that

862
00:49:03,840 --> 00:49:07,517
should be in service again of the we
want the benefits of this technology to

863
00:49:07,600 --> 00:49:10,637
be broadly distributed.
>> How do you think about that internally

864
00:49:10,720 --> 00:49:15,357
just at OpenAI between consumer
and enterprise? Well, a lot of what I've

865
00:49:15,440 --> 00:49:17,277
been thinking about recently has been

866
00:49:17,360 --> 00:49:20,797
focus because this field, it is

867
00:49:20,880 --> 00:49:25,597
opportunity incarnate, right? It's
like you can take AI and apply it to any

868
00:49:25,680 --> 00:49:30,077
problem, any sort of thing you want
to build. It's now on the table. And the

869
00:49:30,160 --> 00:49:33,437
problem that we have is that there's
only so much compute. Where do you want

870
00:49:33,520 --> 00:49:37,037
to put it? And so, you need to have
synergies. You need to have return to

871
00:49:37,120 --> 00:49:38,397
the fact that you have multiple things

872
00:49:38,480 --> 00:49:41,357
happening that they all add up. 1 plus 1

873
00:49:41,440 --> 00:49:46,477
equals 10. Like that's where that's the
dream that's the goal and a lot of what

874
00:49:46,560 --> 00:49:48,157
I think is important for this next phase

875
00:49:48,240 --> 00:49:51,917
of open AI very clearly enterprise

876
00:49:52,000 --> 00:49:55,757
because the economy is becoming this
compute powered economy before our very

877
00:49:55,840 --> 00:49:58,717
eyes. It's happening right now
like we've seen this with software

878
00:49:58,800 --> 00:50:03,517
engineering and it's going to happen
with every single field of work people

879
00:50:03,600 --> 00:50:08,237
do with a computer. everyone's computer
work like is going to be something where

880
00:50:08,320 --> 00:50:11,037
rather than you doing work with your
computer, your computer's going to do

881
00:50:11,120 --> 00:50:13,677
work for you. It's like truly going to

882
00:50:13,760 --> 00:50:18,557
be amazing. And so we need to be there
to help people deploy these models,

883
00:50:18,640 --> 00:50:21,997
figure out how to utilize them, figure
how to get the most benefit out of them.

884
00:50:22,080 --> 00:50:25,677
And by the way, there's also going to be
a blurring of the line between what is

885
00:50:25,760 --> 00:50:29,917
enterprise and what is consumer because
entrepreneurship is going to become far

886
00:50:30,000 --> 00:50:33,437
easier than ever before. Like we're
seeing this already. And even for

887
00:50:33,520 --> 00:50:37,437
example, one of my friends
was describing that his sister was

888
00:50:37,520 --> 00:50:40,877
describing this app that she really wish
someone had created that she had this

889
00:50:40,960 --> 00:50:42,557
picture of like exactly what she wanted.

890
00:50:42,640 --> 00:50:46,957
And he in the meanwhile was typing into
to Codex. Uh-huh. Uh-huh. And then push

891
00:50:47,040 --> 00:50:50,957
enter. And a few hours later, he shows
her this app and she's like, "Wait, what

892
00:50:51,040 --> 00:50:52,317
what what is this? Like where did this

893
00:50:52,400 --> 00:50:54,877
thing come from? Who built this?" And he

894
00:50:54,960 --> 00:50:59,997
said, "You did." And that is I think
just an amazing thing where you realize

895
00:51:00,080 --> 00:51:03,517
anyone can be a builder. Like these
tools like codeex is for everyone. It's

896
00:51:03,600 --> 00:51:07,197
not just for software engineers.
It's like everyone now can be a software

897
00:51:07,280 --> 00:51:10,717
engineer if they have a vision, if they
have this agency that they have a thing

898
00:51:10,800 --> 00:51:12,637
that they want to accomplish. Like you

899
00:51:12,720 --> 00:51:15,677
now have this magic tool that can do it.

900
00:51:15,760 --> 00:51:19,357
And then on the consumer side, the thing
about consumers, it's too broad of a

901
00:51:19,440 --> 00:51:21,997
term, right? That there's lots
of different things. There's like

902
00:51:22,080 --> 00:51:23,757
entertainment, there's a bunch of things

903
00:51:23,840 --> 00:51:26,957
in self-expression, and there's also

904
00:51:27,040 --> 00:51:31,197
solving goals. And the aspect
of consumer that we're really dialed in on

905
00:51:31,280 --> 00:51:35,437
is solving goals. Like we believe
that this technology, you look at

906
00:51:35,520 --> 00:51:39,677
smartphones, that's like four billion
people use them. All of those people

907
00:51:39,760 --> 00:51:44,397
should have a personal AI, you know,
a personal AGI that's out there that knows

908
00:51:44,480 --> 00:51:48,477
them well, that has their personal
context, that is trustworthy, that they

909
00:51:48,560 --> 00:51:52,397
can ask for advice, but that also knows
them so well that, you know, if your

910
00:51:52,480 --> 00:51:56,157
favorite musician is in town, it just
goes and proactively purchases tickets

911
00:51:56,240 --> 00:51:59,117
and maybe it knows like, oh, I should
check in before doing this. Or maybe it

912
00:51:59,200 --> 00:52:01,517
knows, yep, that just like I got to get

913
00:52:01,600 --> 00:52:03,597
do this and I I have prior approval.

914
00:52:03,680 --> 00:52:09,037
like that level of having an AI
that knows you and can help you achieve

915
00:52:09,120 --> 00:52:11,837
whatever it is you want to achieve and
help you flesh out what are your goals.

916
00:52:11,920 --> 00:52:15,037
You should still set those goals. They
should be your goals, right? You should

917
00:52:15,120 --> 00:52:18,477
be in charge, but that is something we
want to create. And that is something I

918
00:52:18,560 --> 00:52:21,677
think that not just 4 billion people
are going to want to need. I think it's

919
00:52:21,760 --> 00:52:22,957
going to be 8 billion people. I think

920
00:52:23,040 --> 00:52:25,357
that the whole planet is going to really

921
00:52:25,440 --> 00:52:28,717
benefit from and need access to personal

922
00:52:28,800 --> 00:52:31,677
AI, personal AGI. And so you look at

923
00:52:31,760 --> 00:52:33,917
those two dimensions of deep knowledge

924
00:52:34,000 --> 00:52:37,757
work and broad distribution of access to

925
00:52:37,840 --> 00:52:43,437
an agentic system. And we want to build
those two aspects and they come together

926
00:52:43,520 --> 00:52:47,517
because ultimately they're the same
technology. ultimately want an AI that

927
00:52:47,600 --> 00:52:49,437
is there in the cloud that has access to

928
00:52:49,520 --> 00:52:52,637
information that is trustworthy that is

929
00:52:52,720 --> 00:52:55,277
able to give good answers and able to

930
00:52:55,360 --> 00:53:00,077
take actions on your behalf whether it's
building whether it's in your personal

931
00:53:00,160 --> 00:53:03,837
life and maybe you have multiple
instances of it but fundamentally it is

932
00:53:03,920 --> 00:53:08,157
one technological system >> do you
think we'll have data centers in space

933
00:53:08,240 --> 00:53:10,157
>> I think we're going
to have data centers everywhere

934
00:53:10,240 --> 00:53:12,237
>> how far away do you
think we are from that

935
00:53:12,320 --> 00:53:15,917
>> well data centers in space has a lot
of has many technical problems associated

936
00:53:16,000 --> 00:53:18,077
with it. Even for example, the data

937
00:53:18,160 --> 00:53:21,997
centers we build today are very finicky,

938
00:53:22,080 --> 00:53:26,637
right? They're these massive machines
with very breakable, very expensive

939
00:53:26,720 --> 00:53:31,037
components. We've had many issues
in the past where the cables were just too

940
00:53:31,120 --> 00:53:35,277
taut, just literally like too too
tight of cables and then you get signal

941
00:53:35,360 --> 00:53:39,117
integrity issues and the computer
doesn't work. And so figuring out how do

942
00:53:39,200 --> 00:53:41,837
you maintain systems today? it's people

943
00:53:41,920 --> 00:53:46,237
go in physically pull them probably will
move to robotics. So I think figuring

944
00:53:46,320 --> 00:53:49,837
out how to solve some of these technical
problems are going to be very important

945
00:53:49,920 --> 00:53:52,397
dependencies as we think about putting

946
00:53:52,480 --> 00:53:54,637
them in you know people talk about

947
00:53:54,720 --> 00:53:59,517
putting data center in you know various
difficult locations. Um space feels like

948
00:53:59,600 --> 00:54:03,837
a like a grand challenge but I think
that we are going to have such need for

949
00:54:03,920 --> 00:54:05,837
compute that we need to be thinking

950
00:54:05,920 --> 00:54:09,037
about all options. What is iterative

951
00:54:09,120 --> 00:54:11,117
deployment and why do you do it?

952
00:54:11,200 --> 00:54:14,317
>> Well, iterative
deployment is one of the

953
00:54:14,400 --> 00:54:18,637
core pillars of how OpenAI has
approached how to get this technology to

954
00:54:18,720 --> 00:54:23,117
benefit people and to achieve our mission.
And this is something that I

955
00:54:23,200 --> 00:54:27,597
think I think I was probably the person
who articulated those two words, but

956
00:54:27,680 --> 00:54:30,717
this spirit is something that really
emerged as we thought about our first

957
00:54:30,800 --> 00:54:33,837
product deployment and really thinking
about how does that connect to what

958
00:54:33,920 --> 00:54:37,677
we're trying to do. and you realized
that there were two different routes

959
00:54:37,760 --> 00:54:39,757
that you could take in terms of thinking

960
00:54:39,840 --> 00:54:42,957
about the you want to build an AGI it's

961
00:54:43,040 --> 00:54:45,037
going to benefit people how do you do it

962
00:54:45,120 --> 00:54:47,517
and one is you go for kind of build it

963
00:54:47,600 --> 00:54:52,317
in secret you don't deploy anything
you have a lot of time to just kind of

964
00:54:52,400 --> 00:54:55,437
polish it get it right but then at some
point you push a button and you say

965
00:54:55,520 --> 00:54:58,557
deploy and I remember thinking about

966
00:54:58,640 --> 00:55:03,117
could I sign up for that strategy could
I be accountable for that strategy do

967
00:55:03,200 --> 00:55:06,077
you want to sitting in a
room thinking about, okay,

968
00:55:06,160 --> 00:55:09,597
we ran all our tests. Are we ready
to deploy? And you've never deployed

969
00:55:09,680 --> 00:55:14,717
anything ever before, right?
That's your first contact with reality

970
00:55:14,800 --> 00:55:18,397
and it's a very powerful system that's
going to really change the world. Like

971
00:55:18,480 --> 00:55:21,437
that is a very tough problem set.

972
00:55:21,520 --> 00:55:25,437
But instead, what about if you take
an approach where this is your hundth

973
00:55:25,520 --> 00:55:29,597
system? You've had to solve this problem
99 times before with systems of

974
00:55:29,680 --> 00:55:33,357
increasing power. and the world has
also had a chance to adapt to them to

975
00:55:33,440 --> 00:55:38,237
reconfigure around them. And we learned
very early on with GPD3. We got to see

976
00:55:38,320 --> 00:55:40,877
this very concretely what it's like to
deploy something where we spent a lot of

977
00:55:40,960 --> 00:55:43,917
time thinking about what are all
the misuses of GPD3. What are the ways it

978
00:55:44,000 --> 00:55:46,477
could go wrong? We thought about
misinformation. We thought about these

979
00:55:46,560 --> 00:55:49,677
kinds of, you know, grand pictures.
And you know what the number one misuse of

980
00:55:49,760 --> 00:55:53,357
GPD3 was? >> What?
>> It was medical spam. Like advertising

981
00:55:53,440 --> 00:55:56,717
different drugs to people, right?
It's like not something we ever would have

982
00:55:56,800 --> 00:56:00,957
thought of as a problem, [clears throat]
but we see it in front of our eyes and

983
00:56:01,040 --> 00:56:04,717
we get a chance to react and learn from
it. And so iterative deployment is the

984
00:56:04,800 --> 00:56:07,917
idea that we will bring intermediate
intermediate versions of this

985
00:56:08,000 --> 00:56:11,517
technology. Now, it's not an excuse
to just blindly deploy, right? You still

986
00:56:11,600 --> 00:56:16,157
need to think at every step about what's
our best view on all the ways this might

987
00:56:16,240 --> 00:56:20,477
be misused, what are the downsides, what
are the risks, let's mitigate those, but

988
00:56:20,560 --> 00:56:23,917
you get to see it. You get to see if
you're right, learn from reality, and do

989
00:56:24,000 --> 00:56:27,037
better the next time.
>> I think people don't understand the

990
00:56:27,120 --> 00:56:29,117
extent of which like this is all new.

991
00:56:29,200 --> 00:56:30,477
There's no playbook. Like you're

992
00:56:30,560 --> 00:56:32,797
figuring this out as you go as well on

993
00:56:32,880 --> 00:56:36,877
the most rapidly deployed technology in

994
00:56:36,960 --> 00:56:39,677
the world perhaps that's so powerful.

995
00:56:39,760 --> 00:56:44,557
>> It is true that at various points
in OpenAI's history, we've had some hope

996
00:56:44,640 --> 00:56:47,597
that hey, there are people who have
deployed transformative technologies

997
00:56:47,680 --> 00:56:51,437
before, maybe they can tell us the
answers. And it's never been so simple.

998
00:56:51,520 --> 00:56:54,317
They do have wisdom and insights
and that's something that I think we've

999
00:56:54,400 --> 00:56:58,237
really incorporated. But we realized
that we're the closest ones to this

1000
00:56:58,320 --> 00:57:00,637
technology that by virtue of creating

1001
00:57:00,720 --> 00:57:04,077
it, we have an understanding of the ways

1002
00:57:04,160 --> 00:57:05,917
in which we could shape it that is hard

1003
00:57:06,000 --> 00:57:08,877
for someone who isn't so close to it to

1004
00:57:08,960 --> 00:57:11,437
appine on to advise on. And I think that

1005
00:57:11,520 --> 00:57:13,997
one observation I have is that the right

1006
00:57:14,080 --> 00:57:18,717
choices are extremely specific
to the facts of the technology or there's

1007
00:57:18,800 --> 00:57:20,477
different pressures that are exerted by

1008
00:57:20,560 --> 00:57:23,277
cell phones versus, you know, mainframe

1009
00:57:23,360 --> 00:57:26,077
computers versus AI versus electricity.

1010
00:57:26,160 --> 00:57:30,077
Each one of these has its own unique
proclivities and problem sets, ways that

1011
00:57:30,160 --> 00:57:33,917
they're being developed. The individuals
doing it matter too, right? the dynamics

1012
00:57:34,000 --> 00:57:38,077
between different humans and these human
factors have been hugely impactful for

1013
00:57:38,160 --> 00:57:43,037
how AI is playing out today in front
of our eyes. I think that a lot of what we

1014
00:57:43,120 --> 00:57:46,717
spend our time doing from the beginning
of OpenAI and really even before is you

1015
00:57:46,800 --> 00:57:49,917
spend a lot of time dreaming. We spend
a time a lot of time really thinking about

1016
00:57:50,000 --> 00:57:53,757
all the implications of what you might do.
And I think that one thing I've

1017
00:57:53,840 --> 00:57:55,837
observed is that we haven't really been

1018
00:57:55,920 --> 00:57:58,797
surprised by some moments along the way,

1019
00:57:58,880 --> 00:58:01,117
but we have been surprised by when they

1020
00:58:01,200 --> 00:58:03,757
arrive, how hard they are to accomplish,

1021
00:58:03,840 --> 00:58:07,277
exactly the order in which we see them.

1022
00:58:07,360 --> 00:58:09,037
And that the world that we are moving

1023
00:58:09,120 --> 00:58:11,437
towards is I think one that is in many

1024
00:58:11,520 --> 00:58:16,477
ways more wonderful and all inspiring

1025
00:58:16,560 --> 00:58:19,197
than many of the ones we anticipated. If

1026
00:58:19,280 --> 00:58:22,797
one frontier model puts safety as a

1027
00:58:22,880 --> 00:58:25,677
primary concern and another frontier

1028
00:58:25,760 --> 00:58:27,597
model doesn't, how do you view that

1029
00:58:27,680 --> 00:58:30,397
competition playing out over time?

1030
00:58:30,480 --> 00:58:33,757
>> Well, I think we have found that safety

1031
00:58:33,840 --> 00:58:38,557
is actually a core product feature. Like
no one wants a model that is not aligned

1032
00:58:38,640 --> 00:58:40,397
with them, right? You want a model you

1033
00:58:40,480 --> 00:58:44,397
can trust, that does the right things in

1034
00:58:44,480 --> 00:58:47,677
any circumstance you give it. And so we

1035
00:58:47,760 --> 00:58:52,317
have invested and I think we've
actually invested possibly far more than

1036
00:58:52,400 --> 00:58:53,837
certainly people perceive and possibly

1037
00:58:53,920 --> 00:58:57,037
more than any other lab in safety right

1038
00:58:57,120 --> 00:58:59,357
that we have in JBT the broadest

1039
00:58:59,440 --> 00:59:02,717
deployment of AI these language models

1040
00:59:02,800 --> 00:59:05,357
in the world used by the most people we

1041
00:59:05,440 --> 00:59:09,757
have to care we've always cared but you
really see it in terms of us being able

1042
00:59:09,840 --> 00:59:13,517
to bring this technology to so many
people so I don't think that there's a

1043
00:59:13,600 --> 00:59:18,077
sustainable state where the people who
are building this technology and having

1044
00:59:18,160 --> 00:59:20,397
successful products are not also

1045
00:59:20,480 --> 00:59:25,277
investing super hard in safety. And I
think that actually the challenge is a

1046
00:59:25,360 --> 00:59:26,957
little bit about if you step back

1047
00:59:27,040 --> 00:59:29,437
because there are some aspects of what

1048
00:59:29,520 --> 00:59:34,397
it means to deliver safety that are
not necessarily short- term. You have to

1049
00:59:34,480 --> 00:59:37,997
think long term for not just your
business but for what it is that you're

1050
00:59:38,080 --> 00:59:41,197
creating. And some of this is about how
you train the models. Some of this is

1051
00:59:41,280 --> 00:59:43,357
about how do you get your feedback loop.

1052
00:59:43,440 --> 00:59:46,877
But I would just say that we are
committed to safety as part of our

1053
00:59:46,960 --> 00:59:51,917
mission and that's something where I
think it has played out in our products

1054
00:59:52,000 --> 00:59:56,397
and in the world. One thing that people
also miss is that it's not just about

1055
00:59:56,480 --> 01:00:00,797
the safety of the model. It's about
the resilience of society. If you look at

1056
01:00:00,880 --> 01:00:03,197
how transformative technologies enter

1057
01:00:03,280 --> 01:00:06,237
the world that society builds around

1058
01:00:06,320 --> 01:00:10,797
them about their strengths and their
risks, you think about engines, right?

1059
01:00:10,880 --> 01:00:15,357
You build cars, but you also need seat
belts. You also need to have roads. And

1060
01:00:15,440 --> 01:00:20,317
you reorient cities around the fact that
this is how this technology works. You

1061
01:00:20,400 --> 01:00:24,237
think about electricity. You have
various safety standards. You have

1062
01:00:24,320 --> 01:00:28,877
different kinds of where you're allowed
to put the electric poles, right? And

1063
01:00:28,960 --> 01:00:31,197
high voltage lines and all these things.

1064
01:00:31,280 --> 01:00:36,077
And I think the same will be true for AI
that it's not just about the technology

1065
01:00:36,160 --> 01:00:38,077
itself. It's not about the model itself.

1066
01:00:38,160 --> 01:00:43,277
It's really about how do they integrate
into the world with a society that is

1067
01:00:43,360 --> 01:00:47,677
resilient and that is something we're
investing in very significantly. The

1068
01:00:47,760 --> 01:00:49,917
OpenAI Foundation has this as one of its

1069
01:00:50,000 --> 01:00:55,597
key focuses of trying to help society
invest in and build a resilience layer

1070
01:00:55,680 --> 01:00:59,517
for AI. >> What do you think
regulation for AI should look like?

1071
01:00:59,600 --> 01:01:01,037
>> Well, I think that there's a number of

1072
01:01:01,120 --> 01:01:06,557
different pieces to what regulation for AI
needs to accomplish. One that I think

1073
01:01:06,640 --> 01:01:09,277
is very important is we need
to ultimately ensure this technology

1074
01:01:09,360 --> 01:01:12,477
benefits people and you think about

1075
01:01:12,560 --> 01:01:15,197
questions like like it is very clear

1076
01:01:15,280 --> 01:01:19,517
that institutions, jobs, just life paths

1077
01:01:19,600 --> 01:01:23,757
that people thought would be stable.

1078
01:01:23,840 --> 01:01:25,997
Those assumptions may not hold anymore.

1079
01:01:26,080 --> 01:01:28,477
And we need to make sure that we provide

1080
01:01:28,560 --> 01:01:33,437
support that we're all there to support
each other as this technology rolls out.

1081
01:01:33,520 --> 01:01:36,157
And so what does that mean from a
regulatory perspective? I think there's

1082
01:01:36,240 --> 01:01:39,757
a lot of ideas whether it's things like
everyone should have access to compute.

1083
01:01:39,840 --> 01:01:44,157
How do we ensure that that's true?
How do we ensure that as this technology

1084
01:01:44,240 --> 01:01:48,397
starts to generate more economic value,
it doesn't acrude to just one place,

1085
01:01:48,480 --> 01:01:52,077
right? This should be something that
actually everyone is benefiting from.

1086
01:01:52,160 --> 01:01:55,197
This technology shouldn't just
abstractly benefit the economy. It's

1087
01:01:55,280 --> 01:01:59,037
very clearly going to do. It should
directly be something that people feel

1088
01:01:59,120 --> 01:02:02,717
in their daily lives that they
themselves their life is better because

1089
01:02:02,800 --> 01:02:04,797
this technology exists because they're

1090
01:02:04,880 --> 01:02:06,557
using it because they're able to

1091
01:02:06,640 --> 01:02:10,237
accomplish more. And I think that the

1092
01:02:10,320 --> 01:02:13,437
ways in which I see this playing out

1093
01:02:13,520 --> 01:02:15,997
it's very important to ground it in what

1094
01:02:16,080 --> 01:02:20,957
are we really seeing like a good example
is the number of people whose who say

1095
01:02:21,040 --> 01:02:24,477
that their life their life was saved
or life of a loved one was saved through

1096
01:02:24,560 --> 01:02:27,277
the use of chat GBT and you realize that

1097
01:02:27,360 --> 01:02:31,677
that's something that should be
supported and protected and so a good

1098
01:02:31,760 --> 01:02:35,917
example of how you can do that through
regulation is thinking about privacy and

1099
01:02:36,000 --> 01:02:39,917
privilege. You talk to a doctor,
you talk to a lawyer, those are privileged

1100
01:02:40,000 --> 01:02:43,117
conversations, right? You feel
comfortable sharing them. There's

1101
01:02:43,200 --> 01:02:46,637
certain guard rails on when the health
care provider would have to, you know,

1102
01:02:46,720 --> 01:02:49,997
provide that information to law
enforcement or alert someone. It's well

1103
01:02:50,080 --> 01:02:53,517
defined in the law. We don't have
anything like that for AI right now, but

1104
01:02:53,600 --> 01:02:55,437
people are using these tools and they

1105
01:02:55,520 --> 01:03:00,637
should use these tools because they're
so important for giving people access to

1106
01:03:00,720 --> 01:03:03,837
information that they wouldn't be able
to get otherwise and that they should

1107
01:03:03,920 --> 01:03:07,197
have the appropriate kind
of understanding of protections there too.

1108
01:03:07,280 --> 01:03:10,397
And so I think that there's a lot
of just really leaning into thinking about

1109
01:03:10,480 --> 01:03:15,037
how do these models insert into people's
lives? How do we make sure that we can

1110
01:03:15,120 --> 01:03:19,757
continue to innovate while at the same
time also making sure that the benefits

1111
01:03:19,840 --> 01:03:24,157
flow broadly? How do we ensure that
America remains a leader? Right? That

1112
01:03:24,240 --> 01:03:25,597
you think about robotics where I think

1113
01:03:25,680 --> 01:03:27,917
we are not the leader. I think that for

1114
01:03:28,000 --> 01:03:31,837
AI we have to make sure that we continue

1115
01:03:31,920 --> 01:03:36,397
with this remarkable position that we

1116
01:03:36,480 --> 01:03:40,397
have been able to achieve. And you
think about things like data centers that

1117
01:03:40,480 --> 01:03:43,837
those are something where there's
clearly been a lot of concern about

1118
01:03:43,920 --> 01:03:47,197
questions like do they drive up
electricity prices and we have a

1119
01:03:47,280 --> 01:03:51,357
commitment to ensure that they do not
and I think that each of these things

1120
01:03:51,440 --> 01:03:54,237
can be achieved through many different
mechanisms sometimes it's through

1121
01:03:54,320 --> 01:03:57,117
regulation sometimes it's through
commitments from the company sometimes

1122
01:03:57,200 --> 01:04:01,437
it's just through people understanding
the facts like a good example is data

1123
01:04:01,520 --> 01:04:05,037
centers and water usage like that that's
something that people talk about a lot

1124
01:04:05,120 --> 01:04:07,997
but actually our data centers use
incredibly little water right It's

1125
01:04:08,080 --> 01:04:09,677
actually misinformation
that they use a lot.

1126
01:04:09,760 --> 01:04:11,117
>> It's less than a household, isn't it?

1127
01:04:11,200 --> 01:04:13,437
>> It's It is because it's a closed loop.

1128
01:04:13,520 --> 01:04:15,677
You basically fill up a giant like,
you know, think of it as like a swimming

1129
01:04:15,760 --> 01:04:18,717
pool of water and you just circle it

1130
01:04:18,800 --> 01:04:20,797
around. And so it's a fixed amount of

1131
01:04:20,880 --> 01:04:25,757
water that's not very large, but I think
people really understanding the why. Why

1132
01:04:25,840 --> 01:04:30,397
are we building these things? Why is it
worthwhile? How does it benefit me? and

1133
01:04:30,480 --> 01:04:35,037
being able to give people that
empowerment whether it's helping them

1134
01:04:35,120 --> 01:04:38,477
feel that they can be an entrepreneur
now that they can build a business that

1135
01:04:38,560 --> 01:04:42,797
they can create something like all
of that we have to solve for we have to

1136
01:04:42,880 --> 01:04:45,437
make sure that people feel
it in their daily lives

1137
01:04:45,520 --> 01:04:49,037
>> when I told people I was doing this
interview one of the common reactions is

1138
01:04:49,120 --> 01:04:52,797
that they're fearing for their job
and their uncertainty what would you tell

1139
01:04:52,880 --> 01:04:56,877
them >> well I do think
that this technology it

1140
01:04:56,960 --> 01:05:01,677
is uncertain exactly how it will play I
think it is surprising how it will play

1141
01:05:01,760 --> 01:05:04,797
out as well. Like the AIS that we have
right now, the world that we have right

1142
01:05:04,880 --> 01:05:08,397
now is not really something that was
anticipated by science fiction. It's

1143
01:05:08,480 --> 01:05:10,397
just different and some inevitable

1144
01:05:10,480 --> 01:05:13,277
conclusions I think actually turn out to

1145
01:05:13,360 --> 01:05:15,677
not quite look the same way when they

1146
01:05:15,760 --> 01:05:20,557
come to pass. So I believe it's always
easiest to see what you lose, right? And

1147
01:05:20,640 --> 01:05:23,597
the change is coming. There is no
denying that. That is absolutely the

1148
01:05:23,680 --> 01:05:26,157
case. But it's much harder to see a

1149
01:05:26,240 --> 01:05:28,397
prairie what you gain. And as an

1150
01:05:28,480 --> 01:05:31,277
example, just think about Uber being

1151
01:05:31,360 --> 01:05:35,597
described to someone in 1950. You have
to think about computers. You have to

1152
01:05:35,680 --> 01:05:37,837
think about mobile phones. You have to

1153
01:05:37,920 --> 01:05:39,677
think about GPS. And it's all so that

1154
01:05:39,760 --> 01:05:42,637
you can get a car to appear where you

1155
01:05:42,720 --> 01:05:46,797
where you are in 3 minutes. And like
that's actually crazy if you think about

1156
01:05:46,880 --> 01:05:51,277
that level of technological investment
for that kind of use case. But it really

1157
01:05:51,360 --> 01:05:54,637
happened. And it didn't just happen
for that one use case. It happened for

1158
01:05:54,720 --> 01:05:56,237
thousands, for tens of thousands, for

1159
01:05:56,320 --> 01:05:58,397
millions of other use cases. And so I

1160
01:05:58,480 --> 01:06:01,837
think that my view of AI is it is about

1161
01:06:01,920 --> 01:06:05,437
empowerment. It is about human agency.

1162
01:06:05,520 --> 01:06:06,797
And that that does mean that some of

1163
01:06:06,880 --> 01:06:09,997
these institutions, jobs, these kinds of

1164
01:06:10,080 --> 01:06:14,317
things that there will be things that we
thought we could rely on that turn out

1165
01:06:14,400 --> 01:06:16,077
not to be as stable as we thought. And

1166
01:06:16,160 --> 01:06:18,557
so it will affect people. But the

1167
01:06:18,640 --> 01:06:21,357
question to lean into is what do you

1168
01:06:21,440 --> 01:06:25,117
gain and how do you benefit from it. So

1169
01:06:25,200 --> 01:06:29,357
now you can be a builder. You can create
anything. Anything you can imagine can

1170
01:06:29,440 --> 01:06:33,117
become real. Well,
what do you imagine? How do you

1171
01:06:33,200 --> 01:06:36,557
build that skill? Really leaning into
this technology. Like one thing that I

1172
01:06:36,640 --> 01:06:39,837
have observed is across multiple
generations of this technology, the

1173
01:06:39,920 --> 01:06:43,997
people who seem to be getting the most
benefit out of it are the people who did

1174
01:06:44,080 --> 01:06:45,677
it for the previous one. Right? So the

1175
01:06:45,760 --> 01:06:47,837
more that you build the skill at the

1176
01:06:47,920 --> 01:06:51,277
core of it is agency is having a vision

1177
01:06:51,360 --> 01:06:55,597
is having ideas because now the barrier
to entry to trying them out is lower

1178
01:06:55,680 --> 01:06:59,597
than ever before. So I think there will
be new opportunity created. I think that

1179
01:06:59,680 --> 01:07:03,917
the world does need to think about
how do we support everyone through this

1180
01:07:04,000 --> 01:07:07,517
moment of uncertainty through whatever
transitions will come because the

1181
01:07:07,600 --> 01:07:12,157
economy will be a computed economy.
It will be different, but I think that

1182
01:07:12,240 --> 01:07:15,117
there will be a place
for everyone to contribute.

1183
01:07:15,200 --> 01:07:18,477
>> Where should young people be investing
today? If you're in high school or

1184
01:07:18,560 --> 01:07:20,477
university or just trying to start out

1185
01:07:20,560 --> 01:07:22,717
in a job, what skills do you think will

1186
01:07:22,800 --> 01:07:25,517
be more valuable in the future?

1187
01:07:25,600 --> 01:07:27,597
>> Well, I really think leaning into this

1188
01:07:27,680 --> 01:07:32,157
technology is going to be a critical

1189
01:07:32,240 --> 01:07:36,477
skill. Just really understanding how
do you get the most out of AI? because

1190
01:07:36,560 --> 01:07:40,957
we're all going to be heading to a world
where we're managers of agents and soon

1191
01:07:41,040 --> 01:07:43,277
maybe the CEO of an autonomous AI

1192
01:07:43,360 --> 01:07:45,277
corporation, right? Just imagine if you

1193
01:07:45,360 --> 01:07:48,957
had the workforce of, you know, 100,000

1194
01:07:49,040 --> 01:07:54,317
person company all at your disposal
operating on your behalf as long

1195
01:07:54,400 --> 01:07:57,917
>> 24/7, right? And as long as you've
got the tokens, the compute to power it,

1196
01:07:58,000 --> 01:08:01,277
which again I think everyone needs
access to compute, that's like so

1197
01:08:01,360 --> 01:08:02,797
critical for the world to figure out and

1198
01:08:02,880 --> 01:08:05,997
get right because then at that point,

1199
01:08:06,080 --> 01:08:09,437
you can point that at any problem and

1200
01:08:09,520 --> 01:08:11,917
the number of problems that humanity

1201
01:08:12,000 --> 01:08:15,117
could want to solve are boundless. And

1202
01:08:15,200 --> 01:08:19,357
so I think that the more that people
do lean into this technology, figure out

1203
01:08:19,440 --> 01:08:22,797
how to take advantage of what's coming,
how to combine these technologies in new

1204
01:08:22,880 --> 01:08:28,157
ways, how to interact with our agents
to really manage them to think about when

1205
01:08:28,240 --> 01:08:31,757
what is it that I want, what is it that
is my sense of self, what is my purpose,

1206
01:08:31,840 --> 01:08:35,757
what do I want to see in the world,
it is going to be easier than ever to

1207
01:08:35,840 --> 01:08:39,917
accomplish that. And I think
that that world with what we

1208
01:08:40,000 --> 01:08:45,117
gain, I think is going to be almost
unimaginable in its upside. That's the

1209
01:08:45,200 --> 01:08:48,557
most positive sort of view of the future.
What's the most negative one you

1210
01:08:48,640 --> 01:08:52,477
can imagine? One thing that's very
interesting about how technology has

1211
01:08:52,560 --> 01:08:55,437
played out to date is that it's really

1212
01:08:55,520 --> 01:08:57,997
been about contorting ourselves to the

1213
01:08:58,080 --> 01:09:01,117
machine, right? You think about how many

1214
01:09:01,200 --> 01:09:05,917
people work where you have this box
and you're typing away at it and you're

1215
01:09:06,000 --> 01:09:09,117
getting your carpal tunnel and your
shoulders are hunched and all of those

1216
01:09:09,200 --> 01:09:10,637
things that were not natural, right?

1217
01:09:10,720 --> 01:09:14,717
That's not really what we're designed for.
And we're going to be moving to

1218
01:09:14,800 --> 01:09:19,277
this world where it's not just that
you're doing work with your computers.

1219
01:09:19,360 --> 01:09:20,717
Your computer actually does work for

1220
01:09:20,800 --> 01:09:24,637
you. And that is something that presents

1221
01:09:24,720 --> 01:09:28,557
opportunities. I think it presents risks.
I think we need to figure out how

1222
01:09:28,640 --> 01:09:32,957
to mitigate those. Like one core
thing at the end of the day is that

1223
01:09:33,040 --> 01:09:36,957
if you have machines that help people
actualize their goals, right, that's out

1224
01:09:37,040 --> 01:09:41,437
there doing what you want.
Sometimes people have conflicting goals.

1225
01:09:41,520 --> 01:09:45,037
How do you resolve that? How do you
decide what the bounds are on what an A

1226
01:09:45,120 --> 01:09:47,517
will help you with and what they won't?

1227
01:09:47,600 --> 01:09:50,717
Really trying to figure out how does
this slot into society? How do you make

1228
01:09:50,800 --> 01:09:53,277
sure that the benefits don't just go to

1229
01:09:53,360 --> 01:09:55,677
one corporation, one set of people, but

1230
01:09:55,760 --> 01:09:58,397
that actually do lift up everyone? We

1231
01:09:58,480 --> 01:10:02,397
need to raise the floor so that everyone

1232
01:10:02,480 --> 01:10:06,957
has access to a great life,
this technology, and are able to do things

1233
01:10:07,040 --> 01:10:09,997
with it. And I think
it'll correspondently also

1234
01:10:10,080 --> 01:10:14,077
lift the ceiling. And so I think
we're going to be in a world where

1235
01:10:14,160 --> 01:10:18,237
everyone is going to have new
opportunities that there will be more

1236
01:10:18,320 --> 01:10:21,917
just I don't know if the right word
is safety net or just like that. There

1237
01:10:22,000 --> 01:10:24,957
should be something that that really
is able to make sure that everyone gets

1238
01:10:25,040 --> 01:10:27,837
brought along, but then
we're going to be able to

1239
01:10:27,920 --> 01:10:31,997
accomplish so much more. And you think
about things like access to medical

1240
01:10:32,080 --> 01:10:33,997
care. Like we should be in a world if we

1241
01:10:34,080 --> 01:10:37,517
do our job right where everyone gets

1242
01:10:37,600 --> 01:10:42,557
access has a doctor in their pocket
that is better than any team of doctors

1243
01:10:42,640 --> 01:10:46,077
today. The world's best doctors. They're
there for you. They care about you.

1244
01:10:46,160 --> 01:10:49,677
They're actually reading your charts
that they're thinking 24/7 about how can

1245
01:10:49,760 --> 01:10:52,557
we actually help with this condition.

1246
01:10:52,640 --> 01:10:57,117
It's disruptive, right? It's it's not
going to come for free in terms of how

1247
01:10:57,200 --> 01:11:00,797
this technology will interact with
the world. And we've already seen the

1248
01:11:00,880 --> 01:11:04,077
beginning eras of it. But I think that

1249
01:11:04,160 --> 01:11:07,997
what we're going to see over just even
the next two years, I think it will be

1250
01:11:08,080 --> 01:11:12,957
this force for good. But we have to also
acknowledge all the ways that it could

1251
01:11:13,040 --> 01:11:16,637
go wrong or the risks of it
in order to achieve those upsides.

1252
01:11:16,720 --> 01:11:20,387
>> When you need to build up your team
to handle the growing chaos at work, use

1253
01:11:20,470 --> 01:11:24,317
[music] Indeed sponsor jobs. It gives
your job post the boost it needs to be

1254
01:11:24,400 --> 01:11:27,517
seen and helps reach people with the
right skills, certifications, [music]

1255
01:11:27,600 --> 01:11:31,277
and more. Spend less time searching
and more time actually interviewing

1256
01:11:31,360 --> 01:11:32,837
candidates who check all your boxes.

1257
01:11:32,920 --> 01:11:36,477
[music] Listeners of this show
will get a $75 sponsor job credit at

1258
01:11:36,560 --> 01:11:40,637
indeed.com/mpodcast.
That's indeed.com/mpodcast. [music]

1259
01:11:40,720 --> 01:11:44,477
Terms and conditions apply. Need
a hiring hero? This is a job for Indeed

1260
01:11:44,560 --> 01:11:48,077
sponsored jobs. >> We always
end every podcast with the

1261
01:11:48,160 --> 01:11:52,077
same question which
is what is success for you?

1262
01:11:52,160 --> 01:11:56,077
achieving the open eye mission
of ensuring that artificial general

1263
01:11:56,160 --> 01:11:59,037
intelligence benefits all of humanity.

1264
01:11:59,120 --> 01:12:01,197
>> Thank you very much. This was awesome.

1265
01:12:01,280 --> 01:12:02,717
This is a great conversation, man.

1266
01:12:02,800 --> 01:12:06,680
>> Thank you. I had a great time.