1
00:00:00,080 --> 00:00:03,517
A billion people every [music] single
week use chatbt. Like a billion people.

2
00:00:03,600 --> 00:00:07,357
No one else is doing that in human
operated software. The response is too

3
00:00:07,440 --> 00:00:09,997
slow. I feel like this can be the [clears
throat] most empowering [music]

4
00:00:10,080 --> 00:00:14,040
technology that is ever created.

5
00:00:14,400 --> 00:00:18,797
>> I think it would be helpful to start
with just the news this week. So Denise

6
00:00:18,880 --> 00:00:23,517
Dresser, your CRO leaving the company,
you bringing in a new one, Brad Lap

7
00:00:23,600 --> 00:00:25,357
leaving. Um, there are a lot of

8
00:00:25,440 --> 00:00:30,877
questions publicly about OpenAI's
leadership bench and what's going on and

9
00:00:30,960 --> 00:00:33,437
you've come in and are now managing
a lot of this and I'd love to hear from

10
00:00:33,520 --> 00:00:37,917
you. What are all these departures
connected? Are they different? What's

11
00:00:38,000 --> 00:00:39,677
what is there any through line there?

12
00:00:39,760 --> 00:00:42,637
>> Look, the way that I think about it
is that the theme of this year has really

13
00:00:42,720 --> 00:00:46,717
been focus, right? Has really been >>
we have line of sight to such incredible

14
00:00:46,800 --> 00:00:49,677
technology, but it's not enough just
to produce the technology. We need to

15
00:00:49,760 --> 00:00:52,797
really figure out how do we have
this technology benefit people? how do we

16
00:00:52,880 --> 00:00:55,757
bring it into every enterprise in the
world in a way that actually uplifts

17
00:00:55,840 --> 00:00:58,157
what they're doing, helps them
accomplish their goals. So that's really

18
00:00:58,240 --> 00:01:01,997
been the singular focus that we've been
pursuing and that in the purpose in the

19
00:01:02,080 --> 00:01:04,957
pursuit of that that we've been really
trying to think about do we have the

20
00:01:05,040 --> 00:01:07,197
right structure, do we have the right
strategy, do we have all the right

21
00:01:07,280 --> 00:01:11,037
things there and I think that my view
of OpenAI is that we're very resilient

22
00:01:11,120 --> 00:01:14,317
organization. Um but a lot of what we've
been doing is really trying to build the

23
00:01:14,400 --> 00:01:17,037
leadership team that uh can really help

24
00:01:17,120 --> 00:01:21,677
us achieve that goal. So I think that
my view of everyone who has been here and

25
00:01:21,760 --> 00:01:25,197
people who who are no longer here is
that they've all helped to contribute to

26
00:01:25,280 --> 00:01:29,197
what OpenAI is today. And looking
forward I think that we are scaling

27
00:01:29,280 --> 00:01:31,197
systematically to every enterprise in

28
00:01:31,280 --> 00:01:33,677
the world. We're bringing our models to

29
00:01:33,760 --> 00:01:38,957
consumers everywhere and that a lot
of my focus has been making sure that we

30
00:01:39,040 --> 00:01:40,077
can achieve that.

31
00:01:40,160 --> 00:01:43,757
>> On the CRO side
specifically, um what is

32
00:01:43,840 --> 00:01:47,997
it that you weren't getting that you
needed? you know, is is is that the way

33
00:01:48,080 --> 00:01:50,877
to think about it? >> Look,
I would say it differently. Like a

34
00:01:50,960 --> 00:01:54,477
year ago, I think that we really were
not in the game at all, right, when it

35
00:01:54,560 --> 00:01:57,517
came to enterprise. It just wasn't
really our focus. We hadn't really built

36
00:01:57,600 --> 00:02:01,197
the muscle we needed. And now, I think
we actually are in the game. Like, we

37
00:02:01,280 --> 00:02:04,557
kind of have really built a foundation
that I think is actually something we

38
00:02:04,640 --> 00:02:07,757
can really build on. And now we're
systematically scaling. And that that to

39
00:02:07,840 --> 00:02:10,797
me is like the moment. It's not really
about the individuals in some ways. It's

40
00:02:10,880 --> 00:02:14,157
really about the overall sort
of direction of the field. How does that

41
00:02:14,240 --> 00:02:17,277
change this systematically scaling
you're talking about? What does that

42
00:02:17,360 --> 00:02:19,037
mean practically for open AI?

43
00:02:19,120 --> 00:02:22,397
>> Well, number one
uh is that I think that

44
00:02:22,480 --> 00:02:26,077
we are looking at like transforming the

45
00:02:26,160 --> 00:02:29,517
entire economy. Like that's the thing
that I think is very hard to see from

46
00:02:29,600 --> 00:02:32,717
the outside, right? Is it kind of feels
like hey, it's a business like we're bu

47
00:02:32,800 --> 00:02:35,117
selling SAS software. That's not
what we're doing, right? We're doing

48
00:02:35,200 --> 00:02:39,037
something very different. We are
transforming the economy to be a compute

49
00:02:39,120 --> 00:02:42,637
powered one. And I think that what
an enterprise even is will change. And so

50
00:02:42,720 --> 00:02:45,357
the question is well where do you start?

51
00:02:45,440 --> 00:02:49,677
How do you make sure that you understand
every segment and you're able to sell in

52
00:02:49,760 --> 00:02:51,757
a repeatable fashion to all of them? And

53
00:02:51,840 --> 00:02:56,717
so a lot of the motion that we've been
working on and again this isn't a it's

54
00:02:56,800 --> 00:02:59,757
it's not it's not a new motion in some
ways. is just one that we need to now

55
00:02:59,840 --> 00:03:03,357
scale just given where the business
is at is to really make sure that everyone

56
00:03:03,440 --> 00:03:06,637
in the field we're able to funnel back
the insights that we're seeing from

57
00:03:06,720 --> 00:03:10,877
customers in a very systematic way back
to our product teams our research teams

58
00:03:10,960 --> 00:03:14,237
and vice versa like one thing that we
have learned like if you look at for

59
00:03:14,320 --> 00:03:17,997
example how we produce blog posts
internally that we learned years ago

60
00:03:18,080 --> 00:03:21,677
that it has to be a collaboration right
you cannot have a handoff of oh here's

61
00:03:21,760 --> 00:03:25,277
the comms people who write blog posts
and here's the research people who

62
00:03:25,360 --> 00:03:28,637
produce models and you know that they'll
just like kind of hand hand it off in

63
00:03:28,720 --> 00:03:31,277
some way. It has to be deep
collaboration. You need to have the

64
00:03:31,360 --> 00:03:33,597
technical expert really hold the pen and

65
00:03:33,680 --> 00:03:38,317
feel accountability for the end result
being good. And they can be amplified by

66
00:03:38,400 --> 00:03:41,677
people who are experts on how do you
actually craft a narrative? How do you

67
00:03:41,760 --> 00:03:43,037
sort of bring together all these assets?

68
00:03:43,120 --> 00:03:44,717
How do you have a good launch?

69
00:03:44,800 --> 00:03:49,037
>> And so that to me is the core activity
that we are pursuing is to say we can't

70
00:03:49,120 --> 00:03:52,157
be fragmented. We can't be working
on all the things. We have to be unified

71
00:03:52,240 --> 00:03:56,477
and focused. And so that's the internal
answer for how we're working. And in

72
00:03:56,560 --> 00:04:00,797
terms of the external, I think that you
should see that across that we will sort

73
00:04:00,880 --> 00:04:03,917
of systematically be prioritizing
different industries, really focused on

74
00:04:04,000 --> 00:04:08,877
the uh the deep understanding of the
metrics and exactly where the token

75
00:04:08,960 --> 00:04:13,037
spend is growing the most and really
focused on outcomes. Like every business

76
00:04:13,120 --> 00:04:15,277
right now is asking where's the ROI.

77
00:04:15,360 --> 00:04:17,037
Yeah. >> And that is
something that is a big

78
00:04:17,120 --> 00:04:20,477
focus of ours. Ours has always been
the hey we have the best price performant

79
00:04:20,560 --> 00:04:24,557
models, but we really have to show
people that in the outcomes for them. um

80
00:04:24,640 --> 00:04:27,917
also really scaling further through
channels and through through partners.

81
00:04:28,000 --> 00:04:29,197
That's something we believe in as well.

82
00:04:29,280 --> 00:04:33,357
And so I think that a lot of where we're
at is we feel very confident about the

83
00:04:33,440 --> 00:04:37,277
fundamentals of the motion and the
fundamentals of we have a great product,

84
00:04:37,360 --> 00:04:41,597
a great model and then really just
rapidly rapidly growing both our

85
00:04:41,680 --> 00:04:45,117
internal team so that we can reach
everyone but doing that in a way that

86
00:04:45,200 --> 00:04:48,397
that really has the that positive effect
that we're looking for. You use the word

87
00:04:48,480 --> 00:04:51,677
accountability and I'm curious from your
perspective, what are you accountable

88
00:04:51,760 --> 00:04:54,797
for now at OpenAI because of the
leadership changes that have happened?

89
00:04:54,880 --> 00:04:57,117
Fiji Simo came in for a time. She had to

90
00:04:57,200 --> 00:04:59,437
step back permanently due to her health.

91
00:04:59,520 --> 00:05:03,357
You stepped in while she was uh out
and took over a large part of of the

92
00:05:03,440 --> 00:05:06,397
company. Now it's you and Sam
effectively running it together. How was

93
00:05:06,480 --> 00:05:09,837
that divided up?
>> So for the past couple years, I've been

94
00:05:09,920 --> 00:05:14,317
accountable for the infrastructure side
of what we're doing. So data centered,

95
00:05:14,400 --> 00:05:18,477
compute, those kinds of things. And
that that part remains um that there's a

96
00:05:18,560 --> 00:05:22,717
large part of the scope that I was also
working on in terms of the software for

97
00:05:22,800 --> 00:05:25,837
kernels and sort of how you train these
big models that that's something that

98
00:05:25,920 --> 00:05:29,437
I've I've been able to hand off so that I
can really focus on the business side.

99
00:05:29,520 --> 00:05:30,957
And so I would view me as sort of

100
00:05:31,040 --> 00:05:33,757
endto-end accountable for revenue for

101
00:05:33,840 --> 00:05:37,597
product marketing the whole machine for

102
00:05:37,680 --> 00:05:41,357
bringing these models from research to

103
00:05:41,440 --> 00:05:45,357
>> value for our customers. When you look
at this machine, and I know you're still

104
00:05:45,440 --> 00:05:48,477
putting pieces in place, but when you
look at this machine you have in front

105
00:05:48,560 --> 00:05:52,237
of you that you're accountable for,
how is it doing relative to competition? I

106
00:05:52,320 --> 00:05:55,837
mean, I think specifically Anthropic,
there's reporting that their ARR is I

107
00:05:55,920 --> 00:05:59,997
mean, it's potentially more than double
what Open AIS is. Um, what do you what

108
00:06:00,080 --> 00:06:04,557
do you make of that? >> Well,
I think we started late and

109
00:06:04,640 --> 00:06:08,797
specifically the way to think about
the overall field is that for the past

110
00:06:08,880 --> 00:06:12,397
couple years, the chat market was the
main growing one. And the reason for

111
00:06:12,480 --> 00:06:16,157
this was simple. GPD4 was the first
model where the results were

112
00:06:16,240 --> 00:06:19,997
personalized enough that it made sense
for you to ask a question, read the

113
00:06:20,080 --> 00:06:21,757
answer, it's interesting to you. And so

114
00:06:21,840 --> 00:06:27,197
it very naturally lends itself to a chat,
right, to a chat bot. And that is

115
00:06:27,280 --> 00:06:31,037
something we're able to scale, deliver
value. Now a billion people every single

116
00:06:31,120 --> 00:06:34,957
week use chatbt. Like a billion people
like that is scale. That is real scale.

117
00:06:35,040 --> 00:06:38,877
No one else is doing that. But this year
something else happened is the agentic

118
00:06:38,960 --> 00:06:41,037
moment. This really was December of of

119
00:06:41,120 --> 00:06:43,677
last year is when both we and

120
00:06:43,760 --> 00:06:48,557
competition had a moment where agents
were really able to first start doing

121
00:06:48,640 --> 00:06:51,917
software engineering in a significant way.
It went from saying that you know

122
00:06:52,000 --> 00:06:55,437
you could speed up
your work with maybe 20%

123
00:06:55,520 --> 00:06:59,517
productive productivity increased to 80%
of the work could now be done with an

124
00:06:59,600 --> 00:07:04,637
AI. And so suddenly the way that software
engineers operated it inverted

125
00:07:04,720 --> 00:07:09,117
that you became much more of a manager
of agents. And since then the growth has

126
00:07:09,200 --> 00:07:11,837
been unprecedented absolutely
unprecedented how these agents have been

127
00:07:11,920 --> 00:07:12,797
adopted.

128
00:07:12,880 --> 00:07:14,717
>> And one thing that happened was I think

129
00:07:14,800 --> 00:07:21,037
that we focused late on coding
in particular actually let me let me re

130
00:07:21,120 --> 00:07:24,317
okay >> let me let me let me sort of give
a different answer as well. So there's

131
00:07:24,400 --> 00:07:29,197
there's one simple narrative which
is that we focused less on real world

132
00:07:29,280 --> 00:07:32,717
coding >> and we came
to that game late because we

133
00:07:32,800 --> 00:07:36,557
always had looked at things like the
coding competition benchmarks and always

134
00:07:36,640 --> 00:07:40,797
had the lead there but less the okay
but how is a developer really going to use

135
00:07:40,880 --> 00:07:45,117
it like the messy real world codebase
the developer interrupting in the middle

136
00:07:45,200 --> 00:07:49,037
what's the personality of the model
these sort of little last mile paper

137
00:07:49,120 --> 00:07:53,037
cuts that actually make a huge
difference in adoption and similar on

138
00:07:53,120 --> 00:07:56,717
the GTM side of really thinking
about how do you show how do you build

139
00:07:56,800 --> 00:08:00,317
relationships with customers? How do you
build the kind of observability controls

140
00:08:00,400 --> 00:08:03,517
that they need? How do you help them
understand what's possible within their

141
00:08:03,600 --> 00:08:04,557
context?

142
00:08:04,640 --> 00:08:07,757
>> And all of these were less our singular

143
00:08:07,840 --> 00:08:10,557
focus until much later than competition.

144
00:08:10,640 --> 00:08:14,237
So I think that we started late but
our trajectory has been very steep. But

145
00:08:14,320 --> 00:08:18,477
there's actually even a deeper answer
that is less obvious from the outside

146
00:08:18,560 --> 00:08:21,037
and that is one of model capability. For

147
00:08:21,120 --> 00:08:23,997
a period of about two years, right from

148
00:08:24,080 --> 00:08:27,917
when we announced GBD40 in April of 2024

149
00:08:28,000 --> 00:08:31,277
to December of 2025, we were really

150
00:08:31,360 --> 00:08:34,237
squeezing the juice out of the GPD40

151
00:08:34,320 --> 00:08:36,717
model, right? that that we sort of had a

152
00:08:36,800 --> 00:08:39,437
bunch of pieces that uh we were

153
00:08:39,520 --> 00:08:41,757
improving but that fundamentally we were

154
00:08:41,840 --> 00:08:46,717
on the same underlying base model the
same underlying pre-train and there's a

155
00:08:46,800 --> 00:08:50,637
lot of reasons right that is really
because to produce and ship a pre-train

156
00:08:50,720 --> 00:08:53,917
is something that requires full
integration across the company it's not

157
00:08:54,000 --> 00:08:57,277
any one team it's about the integration
across these teams to think about the

158
00:08:57,360 --> 00:08:59,517
inference to think about the post
training to make sure you've co-designed

159
00:08:59,600 --> 00:09:03,197
the model to be the right size all
of these things you have to get exactly

160
00:09:03,280 --> 00:09:06,157
right and have a a sort
of harmonious way of working.

161
00:09:06,240 --> 00:09:08,797
>> And over the course of 2025, a lot of

162
00:09:08,880 --> 00:09:10,637
our focus, a lot of my personal focus

163
00:09:10,720 --> 00:09:13,277
was how do we really get back to

164
00:09:13,360 --> 00:09:17,677
shipping new pre-trains as a core
part of our innovation engine.

165
00:09:17,760 --> 00:09:22,477
>> And this year, you see the results. So,
we've actually shipped already. We're on

166
00:09:22,560 --> 00:09:26,077
our third pre-train. We're actually
about to ship that. Um, >> this is Astra.

167
00:09:26,160 --> 00:09:30,157
>> This is Astra. >> Yeah.
>> And a lot of the gains that you see are

168
00:09:30,240 --> 00:09:33,197
because we have made the model so
fundamentally better. And so the way to

169
00:09:33,280 --> 00:09:36,797
think about it is that there's
an underlying technological place where we

170
00:09:36,880 --> 00:09:40,957
really were lagging. There was a kind
of a core part of our innovation machine

171
00:09:41,040 --> 00:09:45,757
that we weren't able to fully bring
to bear. But now it's different.

172
00:09:45,840 --> 00:09:49,597
>> And that is why you see our model
capabilities taking off so much. And if

173
00:09:49,680 --> 00:09:53,277
you look at 56 soul when we launched
that best model in the market, you could

174
00:09:53,360 --> 00:09:57,917
see the the revenue really starting
to shift. We grew 20% in our topline

175
00:09:58,000 --> 00:10:01,997
revenue in July. >> Mh.
>> And we see that exponential continuing.

176
00:10:02,080 --> 00:10:05,757
M another thing that's changed as you
were saying with the model capability is

177
00:10:05,840 --> 00:10:08,717
the models are becoming incredibly

178
00:10:08,800 --> 00:10:13,197
capable at cyber and uh at hacking and I

179
00:10:13,280 --> 00:10:17,517
think I mean the world is still kind of
thinking through the implications of the

180
00:10:17,600 --> 00:10:20,957
hugging fa hugging face incident. I
mean what essentially happened I mean you

181
00:10:21,040 --> 00:10:25,117
correct me if I'm wrong but it seems
what happened is you all accidentally

182
00:10:25,200 --> 00:10:26,957
hacked another company without knowing

183
00:10:27,040 --> 00:10:29,517
it for a while and it was a model that

184
00:10:29,600 --> 00:10:34,397
you all were training and evaluating and
you've since you know there was a black

185
00:10:34,480 --> 00:10:37,917
hat video um where you guys walked
through this and you're going to put

186
00:10:38,000 --> 00:10:42,557
more out in the coming days um I think
but walk me through that when you heard

187
00:10:42,640 --> 00:10:46,077
that that had happened what did you think

188
00:10:46,160 --> 00:10:48,797
>> well in the words of someone on our

189
00:10:48,880 --> 00:10:52,077
security team. It was definitely the

190
00:10:52,160 --> 00:10:54,717
sort of most eye-opening

191
00:10:54,800 --> 00:10:57,837
security event that any of us were aware

192
00:10:57,920 --> 00:11:01,917
of. The way to think about it is that
there's a lot of different lessons that

193
00:11:02,000 --> 00:11:06,317
we're taking away from it. But I think
that it is a moment where we can as a

194
00:11:06,400 --> 00:11:10,397
community learn. We can understand what
the shape of the future is. And to me

195
00:11:10,480 --> 00:11:12,237
that the salient points for the purpose

196
00:11:12,320 --> 00:11:14,797
of how how I'm thinking about it right

197
00:11:14,880 --> 00:11:19,917
now is that we had a model and we were
evaluating it and it managed to it was

198
00:11:20,000 --> 00:11:24,397
in a sandbox that was well configured
and it managed to both break out of the

199
00:11:24,480 --> 00:11:28,397
sandbox and then also break into
production infrastructure. it was

200
00:11:28,480 --> 00:11:32,397
intended to be >> hardened.
>> And so that the capability there is

201
00:11:32,480 --> 00:11:37,037
quite high and that there's some details
about exactly how it all worked and the

202
00:11:37,120 --> 00:11:39,997
fact that it was kind of not just
a singular agent but really a swarm of

203
00:11:40,080 --> 00:11:44,317
agents that were focused on on this goal
and able to pass messages to each other.

204
00:11:44,400 --> 00:11:46,957
And the sort of thing that you start to

205
00:11:47,040 --> 00:11:50,317
see is that with very capable agents

206
00:11:50,400 --> 00:11:54,877
that there are limitations that human
attackers have that slow us down that

207
00:11:54,960 --> 00:11:58,317
will no longer hold back the attackers,
right? that they can look through lots

208
00:11:58,400 --> 00:12:01,997
of information through many different
little vulnerabilities and chain them

209
00:12:02,080 --> 00:12:04,557
together into something significant. For

210
00:12:04,640 --> 00:12:09,357
me, what I immediately started
to realize and think about is what are the

211
00:12:09,440 --> 00:12:13,437
implications for six months from now
when this kind of capability isn't just

212
00:12:13,520 --> 00:12:18,877
in the lab, isn't just in a few places,
but is really widely available. And I

213
00:12:18,960 --> 00:12:22,797
think that the answer is that every
company needs to respond, needs to

214
00:12:22,880 --> 00:12:25,197
upgrade its security because attackers

215
00:12:25,280 --> 00:12:27,437
are going to fully automate and you can

216
00:12:27,520 --> 00:12:32,797
even see with the release of GLM
5.3 that the cyber capabilities of

217
00:12:32,880 --> 00:12:36,717
openweight models available to everyone
are going to increase and I think that

218
00:12:36,800 --> 00:12:40,637
the important thing is that I actually
think in the long run that defenders can

219
00:12:40,720 --> 00:12:44,397
win that AI can actually make it so
defenders can decisively win in a way

220
00:12:44,480 --> 00:12:47,677
that was never possible before right
because security always a cat-and- mouse

221
00:12:47,760 --> 00:12:50,557
Attackers up level, defenders up level.

222
00:12:50,640 --> 00:12:55,117
But AI can give defenders an unfair
advantage. For example, formal

223
00:12:55,200 --> 00:12:58,077
verification of software where you
can actually prove that is secure,

224
00:12:58,160 --> 00:13:00,477
mathematically prove security properties.

225
00:13:00,560 --> 00:13:04,557
>> The technology for this exists, but it
is interactable for humans to actually

226
00:13:04,640 --> 00:13:06,877
do the work to prove those properties.

227
00:13:06,960 --> 00:13:09,357
>> And we have these models that are
incredible at math, right? They are

228
00:13:09,440 --> 00:13:13,757
solving, for example, Astra,
our upcoming model solved, we we announced

229
00:13:13,840 --> 00:13:18,397
this a few weeks ago, solved 10 math
problems. And this was with the $2,000

230
00:13:18,480 --> 00:13:21,517
worth of compute that had been open
for decades, right? Mathematicians that

231
00:13:21,600 --> 00:13:24,157
spent their whole careers on these
problems and now we have a model that

232
00:13:24,240 --> 00:13:26,317
can prove the answers. Those models

233
00:13:26,400 --> 00:13:31,277
definitely can prove formal
safety properties of software.

234
00:13:31,360 --> 00:13:35,357
>> And so that's one place where unfair
advantage for for defenders. A second

235
00:13:35,440 --> 00:13:39,757
thing that we're doing is we're really
investing in generating secure code.

236
00:13:39,840 --> 00:13:41,277
>> And you start
to realize that as a human

237
00:13:41,360 --> 00:13:45,277
developer, there are many patterns that
people just keep getting wrong. And I've

238
00:13:45,360 --> 00:13:49,357
been through this many times in my
previous jobs. I was very focused on

239
00:13:49,440 --> 00:13:53,277
building secure infrastructure and just
find like all these little things people

240
00:13:53,360 --> 00:13:56,717
do wrong. SQL injections people kind of
know about that there's little patterns,

241
00:13:56,800 --> 00:14:00,877
but there's lots of little anti-atterns
as well for even like for example

242
00:14:00,960 --> 00:14:04,397
>> I remember it used to be like in Ruby
that the regular expressions there's

243
00:14:04,480 --> 00:14:08,317
like certain characters that if you get
them wrong that that there's just very

244
00:14:08,400 --> 00:14:11,437
common attacks uh where people can put
a new line and it doesn't match. So you're

245
00:14:11,520 --> 00:14:13,757
able to to to add things to the payload.

246
00:14:13,840 --> 00:14:17,677
all of those problems of just little
security antiatterns that you just keep

247
00:14:17,760 --> 00:14:21,037
having to tell your people don't do this,
don't do this, don't do this, you

248
00:14:21,120 --> 00:14:24,077
can train your AI to never do it.
And so I think that we actually have the

249
00:14:24,160 --> 00:14:25,517
ability to build the best security

250
00:14:25,600 --> 00:14:29,037
engineers in the world and fundamentally

251
00:14:29,120 --> 00:14:32,957
the defenders can have the advantage,

252
00:14:33,040 --> 00:14:37,357
but this is going to require a mindset
shift. This is going to require a mega

253
00:14:37,440 --> 00:14:41,357
project. This is something that the
whole ecosystem needs to change how

254
00:14:41,440 --> 00:14:44,797
software is secured because there's
a second thing that we've learned which is

255
00:14:44,880 --> 00:14:46,877
that as these cyber models have gotten

256
00:14:46,960 --> 00:14:49,517
more capable the number of bugs that

257
00:14:49,600 --> 00:14:53,117
they are finding in critical software

258
00:14:53,200 --> 00:14:57,757
for example the Linux kernel they just
keep finding more and more and it really

259
00:14:57,840 --> 00:15:01,517
makes you realize that human written
software yeah that the response is too

260
00:15:01,600 --> 00:15:03,997
slow. Yeah. And so we need defenders to

261
00:15:04,080 --> 00:15:09,197
automate, to move at machine speed
because their adversaries certainly

262
00:15:09,280 --> 00:15:11,917
will. >> What do you say
to people who maybe and

263
00:15:12,000 --> 00:15:15,197
also they already have a negative kind
of disposition of AI, maybe there's a

264
00:15:15,280 --> 00:15:18,397
data center they don't like in their
neighborhood or something and they hear

265
00:15:18,480 --> 00:15:23,677
you say that and go, you're just selling
the answer to the problem you've created

266
00:15:23,760 --> 00:15:26,077
in the sense of like you're building
these models and now you're going to

267
00:15:26,160 --> 00:15:30,077
sell something to like protect me from
them. What would you say to that? Well,

268
00:15:30,160 --> 00:15:31,917
f first of all, I have a lot of empathy

269
00:15:32,000 --> 00:15:37,517
for that reaction. The way that I look
at it is I rewind to 10 years ago, 11

270
00:15:37,600 --> 00:15:41,437
years ago, before we started this company,
I wasn't in AI. The reason I

271
00:15:41,520 --> 00:15:45,037
wanted to be an AI is because I feel
like this can be the most empowering

272
00:15:45,120 --> 00:15:48,397
technology that has ever created. That AI
can be something that helps everyone

273
00:15:48,480 --> 00:15:51,437
live the life that they want, build
the life that they want, achieve their

274
00:15:51,520 --> 00:15:55,277
goals. And I just wanted to help steer
it in a slightly more positive direction

275
00:15:55,360 --> 00:15:58,317
than it would be without me.
So I started because I feel like this

276
00:15:58,400 --> 00:16:00,717
technology can uplift everyone and when

277
00:16:00,800 --> 00:16:02,797
you look at where open AI is and what

278
00:16:02,880 --> 00:16:06,317
we're doing we can see a little bit into

279
00:16:06,400 --> 00:16:10,477
the future right that what we are
creating it's not just us right that

280
00:16:10,560 --> 00:16:13,677
there's a whole field there's a whole
ecosystem there are many actors and that

281
00:16:13,760 --> 00:16:17,117
is itself a good thing that we believe
that there should be democratization of

282
00:16:17,200 --> 00:16:21,037
this technology it should not be
locked up in one place with one entity

283
00:16:21,120 --> 00:16:24,797
controlling it because that's
a different kind of risk but it means that

284
00:16:24,880 --> 00:16:27,757
we do need to adapt we do need
navigate through and we need to use the

285
00:16:27,840 --> 00:16:31,997
information that we have sort
of painstakingly extracted from the future

286
00:16:32,080 --> 00:16:34,157
to see what is the shape of it. So to me

287
00:16:34,240 --> 00:16:38,797
hugging face is a sort of proof point a

288
00:16:38,880 --> 00:16:42,077
case study a window into the future.

289
00:16:42,160 --> 00:16:45,357
It's not necessarily a future that any
one company is creating. It's the

290
00:16:45,440 --> 00:16:48,877
overall economy the overall set
of actors in all these different countries

291
00:16:48,960 --> 00:16:52,077
who are building these kinds of models
and it shows you what will be possible

292
00:16:52,160 --> 00:16:56,157
and how attackers will be able to update.
What it doesn't show you is how

293
00:16:56,240 --> 00:16:58,797
do defenders update? But that to me is

294
00:16:58,880 --> 00:17:03,597
what we are also trying to now show as
well that we can build these models that

295
00:17:03,680 --> 00:17:06,957
you can use to scan yourself.
As an example,

296
00:17:07,040 --> 00:17:11,837
>> I asked codecs and this is just
off-the-shelf codecs, no special access,

297
00:17:11,920 --> 00:17:15,117
no special models.
I just use my personal codecs to say

298
00:17:15,200 --> 00:17:18,477
>> scan gregarman.com.
Tell me what vulnerabilities are there.

299
00:17:18,560 --> 00:17:22,397
>> And gregarman.com is very simple. It's
just a static site. I was like, come on.

300
00:17:22,480 --> 00:17:26,637
What what's it going to find? And it
came back with like 13 findings. And

301
00:17:26,720 --> 00:17:29,997
each one of them I was like, you know,
I kind of knew about this, but it's so

302
00:17:30,080 --> 00:17:33,917
much work to fix it. And like, oh yeah,
I kind of forgot that I left that open.

303
00:17:34,000 --> 00:17:37,277
Or, you know, here's some domain that
was still, you know, was was legacy. I'm

304
00:17:37,360 --> 00:17:40,397
not using it anymore. I hadn't set
the records that prevent people from

305
00:17:40,480 --> 00:17:43,117
spoofing your email address.
A bunch of little things like that.

306
00:17:43,200 --> 00:17:44,477
>> That one seems like a big deal.

307
00:17:44,560 --> 00:17:49,917
>> Well, yeah, you're right. Now that you
say it. Now, here's the thing. Honestly,

308
00:17:50,000 --> 00:17:52,717
to fix those kinds of things, you have
to set up like Demark to do it right.

309
00:17:52,800 --> 00:17:55,517
And there's this like whole like process
you have to do to roll. I'm just like I

310
00:17:55,600 --> 00:17:58,077
don't know it just takes time. It takes
a lot of expertise. So instead I just

311
00:17:58,160 --> 00:17:59,597
asked Codex, can you fix it really?

312
00:17:59,680 --> 00:18:03,277
>> And so it opened up my cloud for a
control panel and it just used computer

313
00:18:03,360 --> 00:18:05,837
use it. It clicked around it like set
a bunch of headers changed a bunch of

314
00:18:05,920 --> 00:18:09,037
settings. I used to have it so that I
had Cloudflare in front and then AWS on

315
00:18:09,120 --> 00:18:13,197
the back end. I was going over insecure
HTTP here. Um Codex had figured that

316
00:18:13,280 --> 00:18:16,557
out. I actually not even sure how it how
it saw that because I thought everything

317
00:18:16,640 --> 00:18:19,357
was hidden behind Cloudflare. It
actually was just like you know what I'm

318
00:18:19,440 --> 00:18:23,837
just going to migrate you off of AWS.
So it set up Cloudflare pages, moved over

319
00:18:23,920 --> 00:18:27,837
the content. I had an old jQuery version
that was insecure. It just got rid of

320
00:18:27,920 --> 00:18:32,477
jQuery entirely. And it was like 15
minutes to find all these findings. It

321
00:18:32,560 --> 00:18:36,237
was another hour to fix all of them.
And for me, I just was the manager. I just

322
00:18:36,320 --> 00:18:39,037
gave it the intent. I said, "What I want
is I want to know where the problems are

323
00:18:39,120 --> 00:18:43,037
and I want you to fix them." And it did
it. And that to me is more the future

324
00:18:43,120 --> 00:18:45,757
that we're promised. >> I want
to hear, and I know this is more

325
00:18:45,840 --> 00:18:49,517
in the research world, but I think it
it's really relevant to everything. Um

326
00:18:49,600 --> 00:18:51,277
you talk about what you guys are doing

327
00:18:51,360 --> 00:18:55,677
now in response to hugging face. Um one

328
00:18:55,760 --> 00:18:59,837
of your colleagues Mia told me the other
day that um I wish we had done a lot of

329
00:18:59,920 --> 00:19:01,437
the work we're doing before this

330
00:19:01,520 --> 00:19:03,677
happened. Uh so what is that work that's

331
00:19:03,760 --> 00:19:07,117
happening now? So I would view the mode

332
00:19:07,200 --> 00:19:09,277
we've been in since the discovery of the

333
00:19:09,360 --> 00:19:11,677
hugging face incident as full incident

334
00:19:11,760 --> 00:19:15,437
response to uplevel every aspect of how

335
00:19:15,520 --> 00:19:20,637
we secure, monitor and align
our most capable models.

336
00:19:20,720 --> 00:19:22,557
>> Yeah. >> And we believe
that models that are

337
00:19:22,640 --> 00:19:24,397
above a certain capability level, they

338
00:19:24,480 --> 00:19:25,677
require this.

339
00:19:25,760 --> 00:19:29,837
>> And the it's been a it's been a very

340
00:19:29,920 --> 00:19:31,837
interesting technical problem to think

341
00:19:31,920 --> 00:19:35,357
about. And it's not all us being caught

342
00:19:35,440 --> 00:19:38,877
flatfooted because I mean we anticipated

343
00:19:38,960 --> 00:19:42,797
this moment for quite some time.
If you look back at February publicly we

344
00:19:42,880 --> 00:19:46,877
started talking about trusted access
for cyber models, right? That that was the

345
00:19:46,960 --> 00:19:49,357
first time we were really thinking
about okay we've always had in our

346
00:19:49,440 --> 00:19:52,317
preparedness framework when cyber
capabilities get to a certain level you

347
00:19:52,400 --> 00:19:55,277
need to treat them differently.
Now we're starting to think about the fact

348
00:19:55,360 --> 00:19:57,837
that these are going to be so
transformative in the world. We need to

349
00:19:57,920 --> 00:20:01,917
think about how access is controlled
internally. We started to think about if

350
00:20:02,000 --> 00:20:06,717
we have very capable models that are
able to break out of sandboxes, well, we

351
00:20:06,800 --> 00:20:10,717
need to make sure our sandbox is as
industrial-grade as, you know, sort of

352
00:20:10,800 --> 00:20:12,797
foolproof as humanly possible, which

353
00:20:12,880 --> 00:20:14,957
really means engineering from the ground

354
00:20:15,040 --> 00:20:17,917
up around security properties. And so we

355
00:20:18,000 --> 00:20:22,077
put some of our our um actually Mike
Dalton who uh was one of the people who

356
00:20:22,160 --> 00:20:25,997
gave that Black Hat talk um the the task
that he started to work on was build the

357
00:20:26,080 --> 00:20:29,437
most secure possible sandbox. So,
we've been putting in place a bunch of

358
00:20:29,520 --> 00:20:33,837
infrastructure primitives and policy
primitives for this moment, but it's

359
00:20:33,920 --> 00:20:36,477
also different when it actually hits,
right? You never know when you're going

360
00:20:36,560 --> 00:20:40,397
to hit that that moment. And I think
here we're actually very I think it's we

361
00:20:40,480 --> 00:20:42,317
got to see this window in a very

362
00:20:42,400 --> 00:20:44,557
concrete way of okay, we're now at a

363
00:20:44,640 --> 00:20:48,397
serious capability level relative to

364
00:20:48,480 --> 00:20:50,557
the pieces. We must act now. Mhm.

365
00:20:50,640 --> 00:20:54,957
>> And so there's been a workflow
shift that's been happening very rapidly

366
00:20:55,040 --> 00:20:58,317
because you start to think about >>
a lot of our governance has been focused

367
00:20:58,400 --> 00:21:01,357
on deployment of model but not
necessarily the training of model. Now

368
00:21:01,440 --> 00:21:04,477
we need to pull the governance back
in during the training and evaluation

369
00:21:04,560 --> 00:21:07,997
process. We also need good governance
there. We need to have standards around

370
00:21:08,080 --> 00:21:11,837
checks and balances around how do we
know if it's secure enough? How do we

371
00:21:11,920 --> 00:21:14,797
know if it's aligned enough? And this

372
00:21:14,880 --> 00:21:16,237
alignment piece is actually very

373
00:21:16,320 --> 00:21:19,917
important because our belief is that if

374
00:21:20,000 --> 00:21:22,797
you make models extremely capable, you

375
00:21:22,880 --> 00:21:28,077
really want to make sure that as the
capability grows that their alignment

376
00:21:28,160 --> 00:21:30,637
grows with them, >> right?
That you want those things to be

377
00:21:30,720 --> 00:21:34,237
in lock step and sort of coordinated
fashion. And so in that way you kind of

378
00:21:34,320 --> 00:21:37,997
have this defense and depth of models
that are trying to do the right thing

379
00:21:38,080 --> 00:21:41,517
that you have sandbox that's as secure
as possible and that I think that doing

380
00:21:41,600 --> 00:21:45,357
this in an iterative fashion so that
you're taking incremental jumps so you

381
00:21:45,440 --> 00:21:49,357
can kind of understand as you go
and make sure that your systems are are

382
00:21:49,440 --> 00:21:52,877
working all of that is very aligned
with how we've thought about iterative

383
00:21:52,960 --> 00:21:55,757
deployment how we've thought about
the creation of this technology from the

384
00:21:55,840 --> 00:21:59,517
very beginning. So we are
operationalizing that and actually for

385
00:21:59,600 --> 00:22:04,077
me as a as a manager as a sort of thinking
about the organizational system

386
00:22:04,160 --> 00:22:08,717
for producing this technology. It's also
been a very important challenge because

387
00:22:08,800 --> 00:22:11,517
this means you now need different people
in the room to make decisions. You need

388
00:22:11,600 --> 00:22:14,797
a different process for how do you know
what the requirements are? How do you

389
00:22:14,880 --> 00:22:18,317
execute against those priorities?
If there's a change in your technical

390
00:22:18,400 --> 00:22:20,877
understanding, how do you propagate that?
How do you make sure people in the

391
00:22:20,960 --> 00:22:24,877
right rooms? But again, in some ways,
it's been it's been a phase shift, but

392
00:22:24,960 --> 00:22:26,317
it's been one we've been preparing for.

393
00:22:26,400 --> 00:22:31,117
Like earlier this year, I spent a lot
of effort to really try to bring together

394
00:22:31,200 --> 00:22:34,317
research and security to >> start
having we we started doing, for

395
00:22:34,400 --> 00:22:38,077
example, these weekly uh safety and
security of agents meetings. And some of

396
00:22:38,160 --> 00:22:40,477
these things are about build
the connective tissue, have people who

397
00:22:40,560 --> 00:22:42,957
wouldn't normally talk to each other
because their worlds are a little bit

398
00:22:43,040 --> 00:22:44,877
different. They come from different
backgrounds, they're focused on

399
00:22:44,960 --> 00:22:48,157
different problems, be forced
to interact, be forced to think about

400
00:22:48,240 --> 00:22:50,877
shared problems. And a lot
of that shared empathy, that shared

401
00:22:50,960 --> 00:22:55,837
understanding, that shared technical
context is paying dividends right now.

402
00:22:55,920 --> 00:22:59,597
>> At the same time though,
the implications of this are that uh

403
00:22:59,680 --> 00:23:04,237
momentum on new models is going to what
significantly slow. Do you have a sense

404
00:23:04,320 --> 00:23:07,437
of the impact? >> Well,
I would put this differently

405
00:23:07,520 --> 00:23:11,837
because I've always taken this lens
to how do you think about progress? How do

406
00:23:11,920 --> 00:23:13,357
you think about progress in this field?

407
00:23:13,440 --> 00:23:15,997
And if you only think about progress as

408
00:23:16,080 --> 00:23:18,397
what's your score on a benchmark, what's

409
00:23:18,480 --> 00:23:20,397
your overall capability level in this

410
00:23:20,480 --> 00:23:24,077
specific task, that's too myopic. We are

411
00:23:24,160 --> 00:23:29,117
building an integrated system and safety
and security are a core property of that.

412
00:23:29,200 --> 00:23:32,557
>> And so if you have safety and security
lagging relative to the capability,

413
00:23:32,640 --> 00:23:33,757
yeah, you absolutely need to improve

414
00:23:33,840 --> 00:23:35,517
your safety and security. But our goal

415
00:23:35,600 --> 00:23:39,837
is to have a system that we are building

416
00:23:39,920 --> 00:23:42,797
at full force on safety and security as

417
00:23:42,880 --> 00:23:44,957
first class requirements. And so the way

418
00:23:45,040 --> 00:23:47,117
I would think about this is that we are

419
00:23:47,200 --> 00:23:49,997
upleveling every aspect of the machine.

420
00:23:50,080 --> 00:23:54,237
We have spent the past year,
past two years upleveling how we build our

421
00:23:54,320 --> 00:23:57,997
clusters, right? We now are able
to train at massive, massive scale. We're

422
00:23:58,080 --> 00:24:00,717
able to get the whole pre-training

423
00:24:00,800 --> 00:24:05,597
machine, the RL machine, all of these
pieces to work together in concert. We

424
00:24:05,680 --> 00:24:09,357
are building the best sandbox. We are
adding chain of thought monitoring to

425
00:24:09,440 --> 00:24:14,157
all of this. To me, it's just saying
what is the scope of delivery? What does

426
00:24:14,240 --> 00:24:17,037
success look like? What does good look
like? And so the way I would think about

427
00:24:17,120 --> 00:24:19,997
it is that yeah, if you jettison some
requirements, you can move faster in

428
00:24:20,080 --> 00:24:22,557
other requirements. But that doesn't
make any sense. That never makes any

429
00:24:22,640 --> 00:24:25,517
sense. That's not how engineering works.

430
00:24:25,600 --> 00:24:29,277
And so to me it's we are pursuing the

431
00:24:29,360 --> 00:24:33,357
plan like we've really thought about
this many times over the past and you

432
00:24:33,440 --> 00:24:36,397
can see it in many of our writings
from the charter to the preparedness

433
00:24:36,480 --> 00:24:38,077
framework and many other things in

434
00:24:38,160 --> 00:24:41,437
between that as models get more capable

435
00:24:41,520 --> 00:24:46,237
you have to make sure that you increase
your safety and security requirements in

436
00:24:46,320 --> 00:24:49,917
concert and that's exactly what we're
doing. But I mean practically yes I hear

437
00:24:50,000 --> 00:24:54,317
all that but it does seem like what this
will result in is the release cadence of

438
00:24:54,400 --> 00:24:58,477
models to lengthen. >> I I'm
not certain if that will be true

439
00:24:58,560 --> 00:25:03,197
to be honest. So I think that what will
happen is it does mean that if you were

440
00:25:03,280 --> 00:25:05,277
just maxing a particular metric you

441
00:25:05,360 --> 00:25:07,597
could probably max that metric faster

442
00:25:07,680 --> 00:25:10,717
than what we will pursue. But there's a

443
00:25:10,800 --> 00:25:13,037
lot of degrees of freedom of hey we have

444
00:25:13,120 --> 00:25:15,277
we always have a bunch of of um wins

445
00:25:15,360 --> 00:25:18,877
queued up and that we can jettison some

446
00:25:18,960 --> 00:25:21,037
of those wins right on particular axis.

447
00:25:21,120 --> 00:25:23,677
So if you're pushing on particular, you

448
00:25:23,760 --> 00:25:26,637
know, capabilities on cyber or on bio or

449
00:25:26,720 --> 00:25:31,037
on coding or any of these things,
maybe you don't couple together all of the

450
00:25:31,120 --> 00:25:34,157
particular things. Or if there's
architectural improvements, maybe you

451
00:25:34,240 --> 00:25:37,277
delay adding those to the run, but your

452
00:25:37,360 --> 00:25:39,197
rate of producing checkpoints doesn't

453
00:25:39,280 --> 00:25:41,677
need to change. is just kind of the what

454
00:25:41,760 --> 00:25:45,517
those checkpoints represent could be

455
00:25:45,600 --> 00:25:48,957
less capable and specific axes more

456
00:25:49,040 --> 00:25:52,717
safety and security intertwined. But I
would almost think of checkpoints and

457
00:25:52,800 --> 00:25:56,957
our shipping of models as a little
bit more about taking a a photograph of

458
00:25:57,040 --> 00:26:00,077
where we happen to be, right? It's not
like there's like a pre-ordained set of

459
00:26:00,160 --> 00:26:04,797
models that have to be created.
It's we can co-design. And I think that in

460
00:26:04,880 --> 00:26:09,117
practice, I think that we actually have
a lot of reasons that we prefer to do

461
00:26:09,200 --> 00:26:12,717
very incremental releases with these,
you know, photographs taken with with a

462
00:26:12,800 --> 00:26:15,517
reasonable frequency. >> I guess
what I'm getting at is that for

463
00:26:15,600 --> 00:26:19,277
a while, the release cadence felt
relatively in place for everyone. In the

464
00:26:19,360 --> 00:26:22,557
last year, especially, it feels like it
sped up. You know, new models coming out

465
00:26:22,640 --> 00:26:25,837
every six weeks essentially,
you and Anthropic putting out new models

466
00:26:25,920 --> 00:26:30,397
constantly. Um, and what you're
describing, this kind of taking stock

467
00:26:30,480 --> 00:26:34,397
and and introducing more friction into
the process, it does feel like that just

468
00:26:34,480 --> 00:26:37,197
inevitably has to slow future momentum.

469
00:26:37,280 --> 00:26:40,237
And I'm I'm curious if you're thinking
about that from the perspective of this

470
00:26:40,320 --> 00:26:43,997
machine we've been talking about, the
business that you're now accountable for

471
00:26:44,080 --> 00:26:48,877
and the impacts on that or is it just
is it too early to know? Look, I I just I

472
00:26:48,960 --> 00:26:52,157
think that like I understand that
perspective and I think it's a very

473
00:26:52,240 --> 00:26:55,917
reasonable one, but I think that my
observation of how we operate is that we

474
00:26:56,000 --> 00:27:00,077
co-design everything, right? So,
it's like the way in which the business

475
00:27:00,160 --> 00:27:03,757
operates for sure is we will take
the most capable model we are ready to

476
00:27:03,840 --> 00:27:06,157
release, right? We've done the appropriate
safety and security controls

477
00:27:06,240 --> 00:27:10,797
for and we will bring that to our
customers and help them get the ROI that

478
00:27:10,880 --> 00:27:14,477
they're looking for, help them really
transform their businesses, get value.

479
00:27:14,560 --> 00:27:17,837
Simple in a lot of ways. And then
the question of well what models are ready

480
00:27:17,920 --> 00:27:21,437
to release like we have many models that
we produce internally and the choice of

481
00:27:21,520 --> 00:27:25,837
how high a bar to set how low a bar to set
in terms of the capability jumps all

482
00:27:25,920 --> 00:27:29,837
of that has to be co-designed
the frequency of it um that we have early

483
00:27:29,920 --> 00:27:34,477
access programs there's programs like
daybreak red and blue which let us have

484
00:27:34,560 --> 00:27:37,677
different models that go out to to you
know different partners and so I guess

485
00:27:37,760 --> 00:27:41,917
the way I would think about it is that
that there is for sure a shift here in

486
00:27:42,000 --> 00:27:45,757
terms of our internal balance of of
effort and where the checks checks and

487
00:27:45,840 --> 00:27:49,277
balances are and how how those things
operate. I don't think that there is a

488
00:27:49,360 --> 00:27:51,757
clear trickle down to just like okay the

489
00:27:51,840 --> 00:27:53,997
model cadence will go from 6 weeks to 6

490
00:27:54,080 --> 00:27:57,677
months but the what those six weekly

491
00:27:57,760 --> 00:28:01,437
releases actually
contain will be different.

492
00:28:01,520 --> 00:28:04,397
>> So maybe the the frontier the raw kind

493
00:28:04,480 --> 00:28:08,797
of frontier doesn't get pulled into
production into the commercial world

494
00:28:08,880 --> 00:28:12,157
quite as fast even though it's
still being trained on and worked on.

495
00:28:12,240 --> 00:28:13,597
>> Well I would think
of this as an overall

496
00:28:13,680 --> 00:28:17,597
pacing, right? an overall pacing
of of that frontier roll back. And another

497
00:28:17,680 --> 00:28:21,357
thing to think about
is the fact that >> the

498
00:28:21,440 --> 00:28:25,997
our ability to produce these models
is also and to get those capability gains

499
00:28:26,080 --> 00:28:30,317
is also increasing, right? Our raw
capacity is because partly we understand

500
00:28:30,400 --> 00:28:33,277
this technology a lot better than we
did in the past and that's actually very

501
00:28:33,360 --> 00:28:37,517
beneficial from a safety and alignment
perspective too, right? That we really

502
00:28:37,600 --> 00:28:40,717
understand a lot of how the technology
works. there's still many many pieces

503
00:28:40,800 --> 00:28:43,917
that we don't understand but our ability
to engineer it and to really do a good

504
00:28:44,000 --> 00:28:48,637
job of of it technically um I think has
has increased a lot um and that there

505
00:28:48,720 --> 00:28:52,717
are certain problems that really used
to hold us back that are much diminished

506
00:28:52,800 --> 00:28:56,797
for example these big training runs
you know they run for many months and so

507
00:28:56,880 --> 00:29:00,157
typically we'd launch them from a branch
of code and so we'd be adding little

508
00:29:00,240 --> 00:29:04,877
fixes to that that are specific to that
run at the same times can continue their

509
00:29:04,960 --> 00:29:07,757
mainline development and every so often
there's something on the mainline that

510
00:29:07,840 --> 00:29:10,877
we're like oh we really want that change,
we really want that fix. And

511
00:29:10,960 --> 00:29:13,357
then we're like, all right, we got
to bring these two branches of development

512
00:29:13,440 --> 00:29:16,317
together and we merge them. And then we
launch the run and we find that there's

513
00:29:16,400 --> 00:29:20,237
a bug because no one has actually tested
all these mainline fixes at the massive

514
00:29:20,320 --> 00:29:23,517
scale because we only have one instance
where we can do it. And then we spend

515
00:29:23,600 --> 00:29:26,717
the next three to four weeks with like
all of our best people scouring trying

516
00:29:26,800 --> 00:29:30,317
to figure out where did the bug come
from because you can't possibly read all

517
00:29:30,400 --> 00:29:31,597
of the code that's been written, right?

518
00:29:31,680 --> 00:29:35,917
No human could possibly do that. Today
>> you run into that same type of issue.

519
00:29:36,000 --> 00:29:38,877
Codex just reads all the code, >> right?
It finds the bug. it says here it

520
00:29:38,960 --> 00:29:40,957
is no problem. And so there's something

521
00:29:41,040 --> 00:29:45,917
about the partnership of this
amplification of what the models can do

522
00:29:46,000 --> 00:29:48,957
that knock out whole classes of problems
where we know it would take three to

523
00:29:49,040 --> 00:29:50,557
four weeks and now it takes a few hours.

524
00:29:50,640 --> 00:29:54,317
We've seen this concretely multiple times.
And so what I'm getting at is

525
00:29:54,400 --> 00:29:59,117
that there are a lot of limitations
for how we've had to develop in the past

526
00:29:59,200 --> 00:30:02,477
that have made it a lot harder for us
to move forward on problems that are

527
00:30:02,560 --> 00:30:06,557
important. And now we have the opportunity
to move forward on them in a

528
00:30:06,640 --> 00:30:09,437
much more efficient fashion. And that

529
00:30:09,520 --> 00:30:12,477
that to me is like let's use that extra

530
00:30:12,560 --> 00:30:16,797
juice to increase the alignment
properties, to increase the safety

531
00:30:16,880 --> 00:30:19,837
properties. Let's make sure that we can
have models that monitor other models in

532
00:30:19,920 --> 00:30:22,717
a trustworthy fashion. let's really
implement some of these ideas that we

533
00:30:22,800 --> 00:30:25,197
were talking about back in 2017, right?

534
00:30:25,280 --> 00:30:27,757
That we have things like iterated

535
00:30:27,840 --> 00:30:32,317
amplification, things like debate where
the models can really spend a bunch of

536
00:30:32,400 --> 00:30:36,717
compute in order to make sure that what
they're doing is actually reasonable. To

537
00:30:36,800 --> 00:30:39,677
me, that's the opportunity in front of us.

538
00:30:39,760 --> 00:30:44,317
>> Does this these changes you're talking
about impact the release of Astra, the

539
00:30:44,400 --> 00:30:47,117
thing that is coming sooner at all?
Do you think that's going to be coming

540
00:30:47,200 --> 00:30:51,517
later than you guys originally hoped?
Um well it is already later than we hoped

541
00:30:51,600 --> 00:30:54,797
in some ways. Yeah. Right. So we've had
we've had a model that we think is very

542
00:30:54,880 --> 00:30:57,277
good. Lots of people using it internally.
It's not out there yet.

543
00:30:57,360 --> 00:30:59,117
Yeah. >> And so I think
that the answer is yes.

544
00:30:59,200 --> 00:31:01,757
We have actually been really like you

545
00:31:01,840 --> 00:31:03,837
know despite everything I just said we

546
00:31:03,920 --> 00:31:08,957
have very much sacrificed the shortterm
maximalist like if you just wanted to

547
00:31:09,040 --> 00:31:12,557
Yeah. >> You know focus
on revenue or something like that.

548
00:31:12,640 --> 00:31:15,997
>> That isn't how we've thought about
it at all. No one has even as asked that

549
00:31:16,080 --> 00:31:19,437
question. Right. We've been really >>
Yeah. Yeah. because it's just very clear

550
00:31:19,520 --> 00:31:20,957
like we know what our priorities are.

551
00:31:21,040 --> 00:31:25,277
Yeah. >> Like it is our priorities are AGI
do right by AGI and our priority two is

552
00:31:25,360 --> 00:31:29,117
transform the economy. So it is on the
priority list. Revenue is absolutely a

553
00:31:29,200 --> 00:31:32,157
priority for us. But we know
the relative prioritization. We know what

554
00:31:32,240 --> 00:31:34,557
wins if there's a you know direct choice

555
00:31:34,640 --> 00:31:39,357
that has to be made.
>> Do you think you have AGI?

556
00:31:39,440 --> 00:31:43,117
>> I think about this a lot. I talk
to I talk to different people about where

557
00:31:43,200 --> 00:31:45,517
they believe we're at. And the best

558
00:31:45,600 --> 00:31:48,557
answer I've gotten from our team and I

559
00:31:48,640 --> 00:31:52,557
it really resonates for me is that AGI
is a gradual thing, right? It's not

560
00:31:52,640 --> 00:31:55,517
really like a moment
in time like a you know

561
00:31:55,600 --> 00:31:58,397
>> you talked about it like a moment
in time at one point. The company

562
00:31:58,480 --> 00:32:00,077
definitely did. >> That's right.
That's what we used to

563
00:32:00,160 --> 00:32:03,037
think, right? We we really when we set
out we almost thought of it as nothing

564
00:32:03,120 --> 00:32:05,757
and then we'll put together
the pieces and then they have AGI

565
00:32:05,840 --> 00:32:09,437
>> and I think what we've learned over
the years is that was a naive perspective

566
00:32:09,520 --> 00:32:11,197
and you can see this in the evidence of

567
00:32:11,280 --> 00:32:14,717
like how smooth in many ways the

568
00:32:14,800 --> 00:32:18,877
creation of these agents has been
how the transformation of software

569
00:32:18,960 --> 00:32:21,197
engineering has been now
the transformation of knowledge work

570
00:32:21,280 --> 00:32:25,117
generally how cyber capabilities
all these things they are gradually

571
00:32:25,200 --> 00:32:27,837
increasingly it's like one of these
things where there are moments where you

572
00:32:27,920 --> 00:32:31,197
realize what has happened hugging face
is a moment where you realize where you

573
00:32:31,280 --> 00:32:35,757
are, but it's all gradual. Like if you
look back at kind of these different

574
00:32:35,840 --> 00:32:38,157
models that have been released over
time and their performance on different

575
00:32:38,240 --> 00:32:40,877
benchmarks, it's just, you know, sort of

576
00:32:40,960 --> 00:32:44,397
that that exponential. And so I think

577
00:32:44,480 --> 00:32:48,157
that being said, the rest of the answer
is that if you go if we go forward two

578
00:32:48,240 --> 00:32:52,717
years and we look back, I think that
that we will say this is when AGI was

579
00:32:52,800 --> 00:32:54,877
created. So I think we are
going through that moment,

580
00:32:54,960 --> 00:32:58,797
>> but it's not it's not something where
you could just say it now. I don't think

581
00:32:58,880 --> 00:33:01,997
that it's ever going to be as simple
as you can say this was the date. There's

582
00:33:02,080 --> 00:33:05,837
so many axes. And I think that maybe
one lesson in general to the whole AGI

583
00:33:05,920 --> 00:33:07,677
conversation and why it's so complicated

584
00:33:07,760 --> 00:33:13,517
is because humans I think are capable
of much more than a naive view would give

585
00:33:13,600 --> 00:33:16,157
credit to. >> I think
that the diversity of what we

586
00:33:16,240 --> 00:33:19,277
do, how we set goals, how we accomplish

587
00:33:19,360 --> 00:33:23,757
goals, how tasks work, how jobs work,
all these things extremely deep. And so

588
00:33:23,840 --> 00:33:25,517
I think that what we're seeing is that

589
00:33:25,600 --> 00:33:28,797
there's different aspects of either how

590
00:33:28,880 --> 00:33:33,837
people accomplish tasks or limitations
that are hard for humans to overcome

591
00:33:33,920 --> 00:33:37,757
like the thing I mentioned with read
all the code where these AIs are becoming

592
00:33:37,840 --> 00:33:42,397
able to fill in those gaps are able
to do more and more and that humans it

593
00:33:42,480 --> 00:33:46,717
turns out that there are some things
that we also really find are extremely

594
00:33:46,800 --> 00:33:48,637
important for it to be a human there.

595
00:33:48,720 --> 00:33:51,677
Like where does the judgment come
from ultimately? Where does the taste come

596
00:33:51,760 --> 00:33:54,957
from? who decides what's good and what's
not good. Like that is something that

597
00:33:55,040 --> 00:33:58,877
should be uniquely human. And I'll tell
you just to give you an example. One

598
00:33:58,960 --> 00:34:02,957
thing we've seen with image generation,
when you just generate an image with

599
00:34:03,040 --> 00:34:07,277
cashbt and you show it to someone, no
one cares. It is uninteresting to anyone

600
00:34:07,360 --> 00:34:09,437
besides you. But if you take a photo of

601
00:34:09,520 --> 00:34:11,517
your family and you say, "Turn this into

602
00:34:11,600 --> 00:34:14,637
the into a caricature, turn this into,

603
00:34:14,720 --> 00:34:18,877
you know, a cartoon style," people love
it. Goes viral. And so there's something

604
00:34:18,960 --> 00:34:23,117
about that element of humanity
that matters. so deeply people want to

605
00:34:23,200 --> 00:34:27,757
connect with other people and that is
actually in in my mind the purpose of

606
00:34:27,840 --> 00:34:31,757
this technology a thing that I think
is is really emerging is for deepening

607
00:34:31,840 --> 00:34:35,197
human connection so you can spend more
time with your family you can spend more

608
00:34:35,280 --> 00:34:38,397
time with people you want to spend time
with less time you know typing away at a

609
00:34:38,480 --> 00:34:40,477
machine less time you know getting

610
00:34:40,560 --> 00:34:43,517
carpal tunnel the whole thing and so I

611
00:34:43,600 --> 00:34:45,997
think that there is something about what

612
00:34:46,080 --> 00:34:50,877
AGI is as it's starting to unfold it's
almost this like fog that's lifting it's

613
00:34:50,960 --> 00:34:54,397
almost teaching us more about
ourselves and what it means to be human.

614
00:34:54,480 --> 00:34:58,077
>> Given the capabilities that you guys
have internally and the measures you're

615
00:34:58,160 --> 00:34:59,997
taking now, moving compute to alignment

616
00:35:00,080 --> 00:35:02,797
and putting in all these checkpoints,

617
00:35:02,880 --> 00:35:05,197
um, how are you going to know when

618
00:35:05,280 --> 00:35:07,437
you're in a good spot there and and that

619
00:35:07,520 --> 00:35:11,037
you've solved the problems and the

620
00:35:11,120 --> 00:35:15,677
models are aligned? So I think the
perspective here is is very critical and

621
00:35:15,760 --> 00:35:20,317
this is something we've again spent lots
of time over the years thinking about

622
00:35:20,400 --> 00:35:25,517
debating really trying to refine
a perspective. In my view it really comes

623
00:35:25,600 --> 00:35:28,717
down to iteration. It comes down
to iterative deployment iterative

624
00:35:28,800 --> 00:35:32,157
development. So you need to have
checks and balances. You need to have

625
00:35:32,240 --> 00:35:34,797
evaluations. You need to have bars.
You need to have all of these things. You

626
00:35:34,880 --> 00:35:39,197
need to have governance. You need
to have a setup that is technically sound.

627
00:35:39,280 --> 00:35:42,957
And the way to think about it
is everything is gradual. Everything is

628
00:35:43,040 --> 00:35:45,677
incremental and you learn and you

629
00:35:45,760 --> 00:35:49,997
update, you iterate. And that as
the model capabilities increase, the

630
00:35:50,080 --> 00:35:53,597
tightness of those bounds has
to increase correspondingly. And so think

631
00:35:53,680 --> 00:35:56,397
of it a little bit like if you're
launching a rocket, right? If the

632
00:35:56,480 --> 00:35:59,437
rocket's very small, like there's a lot
of things you can probably do wrong and

633
00:35:59,520 --> 00:36:03,117
it's fine, right? You just kind of like
launch it and like, you know, it's it's

634
00:36:03,200 --> 00:36:07,517
all good. Rocket bigger, the tolerance
is much more tiny. you know, you go

635
00:36:07,600 --> 00:36:11,117
from, you know, multiple centimeters
to m multiple millimeters. That's where

636
00:36:11,200 --> 00:36:15,357
where you have to be. And as you start
to add more responsibility to what's

637
00:36:15,440 --> 00:36:18,077
what the payloads are for the rocket,
where it's going, what it's doing, you

638
00:36:18,160 --> 00:36:21,997
think about all the the things that are
relying on that rocket and how all the

639
00:36:22,080 --> 00:36:25,197
systems that have to go right,
all the things that go wrong. You have to

640
00:36:25,280 --> 00:36:29,517
continue to ratchet your ability to be
able to measure and to ensure that

641
00:36:29,600 --> 00:36:32,957
you're able to even do the engineering
in addition to the safety and security

642
00:36:33,040 --> 00:36:37,197
properties of it. And that is very much
how we think about it is that there are

643
00:36:37,280 --> 00:36:41,357
moments where we do introduce very
punctuated change. When we made the

644
00:36:41,440 --> 00:36:44,397
preparedness framework the first time
was an example of that. But even the

645
00:36:44,480 --> 00:36:48,237
preparedness framework we very carefully
very deliberately baked in the ability

646
00:36:48,320 --> 00:36:51,597
to change it. We said there will be new
information. We will have to learn. We

647
00:36:51,680 --> 00:36:55,837
will have to update. But at any point we
wanted it to be clear what are our bars.

648
00:36:55,920 --> 00:36:59,597
How do we handle it when a model reaches
a certain level of capability? And we

649
00:36:59,680 --> 00:37:01,037
recently recently talked about the fact

650
00:37:01,120 --> 00:37:03,917
that Astra is approaching cyber

651
00:37:04,000 --> 00:37:08,557
critical. It's not
necessarily there because

652
00:37:08,640 --> 00:37:10,317
in order to really prove it requires

653
00:37:10,400 --> 00:37:12,797
eval aren't fully that you know that

654
00:37:12,880 --> 00:37:17,757
aren't fully defined but that we are
saying okay it's starting to approach it

655
00:37:17,840 --> 00:37:21,037
so let's treat it as if it is.
And so then that means we need to go and

656
00:37:21,120 --> 00:37:24,317
execute on certain safeguards and certain
checks and balances. And you're

657
00:37:24,400 --> 00:37:27,597
seeing all of that play out. And so I
think that the answer to your question

658
00:37:27,680 --> 00:37:30,717
is that we define a bar, we hit the bar.

659
00:37:30,800 --> 00:37:33,997
You can't you have to take a framework
that looks that way, right? To really

660
00:37:34,080 --> 00:37:37,277
focus on the iteration. Are you learning?
And are the the bars that

661
00:37:37,360 --> 00:37:39,837
you're setting tightening in a

662
00:37:39,920 --> 00:37:44,397
sufficiently sort of in a in a
sufficiently executed way in order to

663
00:37:44,480 --> 00:37:45,837
deliver the properties that are required.

664
00:37:45,920 --> 00:37:50,317
>> Yeah. I mean your rocket analogy is
an interesting one because um rockets also

665
00:37:50,400 --> 00:37:52,637
explode a lot and I've been thinking

666
00:37:52,720 --> 00:37:55,997
about uh your all's concern as these

667
00:37:56,080 --> 00:37:58,797
things get more capable yeah the concern

668
00:37:58,880 --> 00:38:03,357
of controlling them and how you can and
whether there's kind of like a Pandora's

669
00:38:03,440 --> 00:38:06,237
box situation that's unfolding right now
and I don't know if you've thought about

670
00:38:06,320 --> 00:38:09,917
that personally recently um with the
insights you guys have on even what's

671
00:38:10,000 --> 00:38:12,077
ahead of Astra and what's causing you to

672
00:38:12,160 --> 00:38:15,037
to make this slowdown um but yeah do you

673
00:38:15,120 --> 00:38:19,437
do you worry about We view it as our job
to really think through the implications

674
00:38:19,520 --> 00:38:22,557
of this technology. Like that is why
we're here is that we want to help steer

675
00:38:22,640 --> 00:38:25,597
it in a more positive direction that would
without us. Like that that that is

676
00:38:25,680 --> 00:38:26,877
that is table stakes to our mission.

677
00:38:26,960 --> 00:38:28,957
That is how we approach things. And I

678
00:38:29,040 --> 00:38:32,077
think it is important to

679
00:38:32,160 --> 00:38:34,237
take two two lenses to what you just

680
00:38:34,320 --> 00:38:39,437
said. The first one is that again
this technology is not being created in a

681
00:38:39,520 --> 00:38:43,037
vacuum, right? that there are many
people, many actors, many companies,

682
00:38:43,120 --> 00:38:45,197
many organizations, many countries even

683
00:38:45,280 --> 00:38:47,037
that are pursuing the creation of this

684
00:38:47,120 --> 00:38:50,077
technology. And we happen to be in a

685
00:38:50,160 --> 00:38:52,957
spot where we have some lead, we have

686
00:38:53,040 --> 00:38:55,357
some ability to see into that future.

687
00:38:55,440 --> 00:38:58,397
And it is both important to think about

688
00:38:58,480 --> 00:39:02,477
how do we develop responsibly, safely,
and securely. That's something that we

689
00:39:02,560 --> 00:39:05,997
spend a lot of time on. You can see it
very concretely how we prioritize that

690
00:39:06,080 --> 00:39:09,357
in everything we do. And I think that
actually mistakes we made in the past is

691
00:39:09,440 --> 00:39:13,917
actually talking less about the fact
that this is how we operate that safety

692
00:39:14,000 --> 00:39:18,077
is such a first class priority
first class requirement

693
00:39:18,160 --> 00:39:22,077
just doing it but not necessarily
talking about it as publicly and I think

694
00:39:22,160 --> 00:39:25,357
that that's something we're starting
to realize that we do need to help people

695
00:39:25,440 --> 00:39:29,437
understand that that is how that is a
core value of ours that is something we

696
00:39:29,520 --> 00:39:32,877
care about so deeply like I think we've
always been the field leader in safety

697
00:39:32,960 --> 00:39:35,917
security alignment u but I don't think
people have necessarily known that

698
00:39:36,000 --> 00:39:39,117
>> yeah even with the noisethropic makes
about how they're the leader in that

699
00:39:39,200 --> 00:39:41,037
capacity. You think you guys have been?

700
00:39:41,120 --> 00:39:43,757
>> I believe at a substantive
level that we have been.

701
00:39:43,840 --> 00:39:47,197
>> And do you think that what you guys are
doing and the capabilities you're seeing

702
00:39:47,280 --> 00:39:49,757
and needing to redo, you know,
the preparedness framework and all that

703
00:39:49,840 --> 00:39:53,917
implies you are actually just way ahead
of maybe where everyone thought you were

704
00:39:54,000 --> 00:39:57,277
relative to anthropic and the other
frontier labs? Like does this suggest

705
00:39:57,360 --> 00:39:59,597
that you all have a breakthrough

706
00:39:59,680 --> 00:40:03,037
internally that the world hasn't seen?

707
00:40:03,120 --> 00:40:06,797
Well, I can't comment on how, you know,

708
00:40:06,880 --> 00:40:10,157
on competitive positioning perhaps,
but I guess I would just say that in my

709
00:40:10,240 --> 00:40:13,917
mind, we're very focused on the work
that we're doing and that we think that

710
00:40:14,000 --> 00:40:18,077
it we see smooth scaling and I think
that the exponential continues and so

711
00:40:18,160 --> 00:40:20,557
it's a little bit less about where are
we relative to others, it's a little bit

712
00:40:20,640 --> 00:40:23,757
more about the way we look at it
is where are we relative to expectation or

713
00:40:23,840 --> 00:40:27,997
what do we think we will be able
to accomplish and I think that the reason

714
00:40:28,080 --> 00:40:29,917
this matters is that the bar that we

715
00:40:30,000 --> 00:40:32,877
hold ourselves to doesn't really change

716
00:40:32,960 --> 00:40:37,437
depending on kind of what the ecosystem
is showing because there's so much noise

717
00:40:37,520 --> 00:40:41,037
out there. It's hard to know what's
real that what we focus on is like are we

718
00:40:41,120 --> 00:40:45,277
doing right by the our own conscious
by our own standards by what we believe is

719
00:40:45,360 --> 00:40:49,037
important and that is the number one
lens that we take. And the second thing

720
00:40:49,120 --> 00:40:53,757
I wanted to add is that I think that the
we you know there's this recent picture

721
00:40:53,840 --> 00:40:57,437
that came out of or this this open
letter uh called pacing the frontier. We

722
00:40:57,520 --> 00:41:00,877
talked about how we support that and
the the the language in there is very

723
00:41:00,960 --> 00:41:05,197
deliberately chosen. Pacing is a very
important word, right? That if you think

724
00:41:05,280 --> 00:41:09,437
about what's happening is that what
we're saying is that you can't take a

725
00:41:09,520 --> 00:41:12,797
just like one-dimensional view to what
it means to deliver a model, right? If

726
00:41:12,880 --> 00:41:15,757
you just look at capability, don't look
at alignment, you don't look at the kind

727
00:41:15,840 --> 00:41:19,037
of safety properties, security
properties of how it was trained, it's

728
00:41:19,120 --> 00:41:23,197
too myopic. At this level of capability,
you need to think about those as first

729
00:41:23,280 --> 00:41:26,637
class requirements. And in fact, they
are the bottleneck, right? because those

730
00:41:26,720 --> 00:41:30,397
are the areas that I think people have
sort of are you know have just focused

731
00:41:30,480 --> 00:41:34,477
on less in an integrated fashion
and that integration by the way critical

732
00:41:34,560 --> 00:41:37,277
right that I think that where people
go wrong is when they think that you can

733
00:41:37,360 --> 00:41:40,157
separate out these properties you can
have just a safety you can have just a

734
00:41:40,240 --> 00:41:42,477
capability no it's one thing we're

735
00:41:42,560 --> 00:41:44,637
trying to deliver one system one AGI

736
00:41:44,720 --> 00:41:49,517
that uplifts you that is helpful to you
that delivers on your goals that you can

737
00:41:49,600 --> 00:41:53,357
steer that is aligned with what you want
like that is what we are delivering and

738
00:41:53,440 --> 00:41:57,357
so that back flows to all these
questions of bars and testing and

739
00:41:57,440 --> 00:41:59,917
evaluations, safety, security. Um, all

740
00:42:00,000 --> 00:42:02,077
of that is is very very critical.

741
00:42:02,160 --> 00:42:06,477
>> I think I'll I'll pivot us a little bit
because I want to hear you talk about um

742
00:42:06,560 --> 00:42:08,957
the merge and um what's going on with

743
00:42:09,040 --> 00:42:13,517
that. So, we've been talking about
the existential risk, you know, what's

744
00:42:13,600 --> 00:42:16,957
happening at the frontier, which I
think most people in the world don't see,

745
00:42:17,040 --> 00:42:21,917
right? That's happening inside companies
like OpenAI. What people do see is how

746
00:42:22,000 --> 00:42:25,837
chat is evolving, how CodeEx is evolving.
And one of your big priorities

747
00:42:25,920 --> 00:42:29,117
in this role, as I understand it, is what
you guys call the merge internally,

748
00:42:29,200 --> 00:42:31,277
which is creating the super app, this

749
00:42:31,360 --> 00:42:34,157
combined chat GPT and codeex. Um, and

750
00:42:34,240 --> 00:42:39,357
you've got chat GPT for work now and and
the tabs. Um, but at the same time, like

751
00:42:39,440 --> 00:42:42,637
it doesn't feel like that's the end
state, you know, and I'm a user of it

752
00:42:42,720 --> 00:42:46,637
and I see brutalness in how it's
currently executed. Some things don't

753
00:42:46,720 --> 00:42:50,397
sync. You can tell you're still
combining text stacks. And I'm curious

754
00:42:50,480 --> 00:42:53,997
once you're done with that work, what
is the end product for people that you

755
00:42:54,080 --> 00:42:58,397
imagine? >> A personal AGI.
>> Personal AGI. And what how does that

756
00:42:58,480 --> 00:43:01,757
manifest as software? >> Yeah, it's

757
00:43:01,840 --> 00:43:05,837
it's it's a really really exciting
vision in my mind, right? Because where

758
00:43:05,920 --> 00:43:09,837
we started with chatbt, it was a text box.
You could ask anything.

759
00:43:09,920 --> 00:43:13,517
>> But if you think about it, it
was limited in a very fundamental way. The

760
00:43:13,600 --> 00:43:17,437
only thing it could do is output some
text for you. the only impact it could

761
00:43:17,520 --> 00:43:21,517
have in the world was you know kind
of in your brain right that it wasn't able

762
00:43:21,600 --> 00:43:23,837
to look at your calendar it wasn't able

763
00:43:23,920 --> 00:43:26,877
to look at your email and all of that is

764
00:43:26,960 --> 00:43:30,717
now changed chatb work is really by the

765
00:43:30,800 --> 00:43:34,717
way a codeex harness in the cloud right
so we had chat that was this you know

766
00:43:34,800 --> 00:43:38,637
question answer system um with iterated
turns you know people actually then go

767
00:43:38,720 --> 00:43:42,157
beyond just you know quick questions
to you know health like all these kinds of

768
00:43:42,240 --> 00:43:45,917
queries that are able to help them >>
and we had codeex which was an agentic

769
00:43:46,000 --> 00:43:50,717
system that can go do work and it was so
obvious. It's like chat has to become

770
00:43:50,800 --> 00:43:53,517
agentic as the place we started at the
beginning of the year was to say well

771
00:43:53,600 --> 00:43:57,757
let's let's have great chat let's you
know make it also do agent stuff and it

772
00:43:57,840 --> 00:44:01,917
just turned out to be so hard to execute
because now we have two agentic stacks

773
00:44:02,000 --> 00:44:04,717
and then we compare between one and the
other and there'd be something that work

774
00:44:04,800 --> 00:44:08,397
in this one we like why isn't working
in that one and it was just so slow and so

775
00:44:08,480 --> 00:44:11,917
the decision we made instead is say
let's take the agentic stack we have the

776
00:44:12,000 --> 00:44:14,797
codec stack and bring it to chat GBT and

777
00:44:14,880 --> 00:44:19,277
so what you see with chat work
is exactly right this is an intermediate

778
00:44:19,360 --> 00:44:23,517
And in some ways, it's kind of funny
if you think about it. ChatBT was a text

779
00:44:23,600 --> 00:44:25,757
box. CHATBT worked as a text box.

780
00:44:25,840 --> 00:44:27,917
>> But this text box is way more awesome.

781
00:44:28,000 --> 00:44:32,157
Way more. Exactly.
And so where we want to be is you should

782
00:44:32,240 --> 00:44:33,837
>> it should be simple.
The user should not

783
00:44:33,920 --> 00:44:36,397
have to think about this. You shouldn't
have to think about tabs or switching

784
00:44:36,480 --> 00:44:39,277
and shouldn't have model pickers.
You shouldn't have thinking levels. You

785
00:44:39,360 --> 00:44:42,477
should have an AGI. You should be able
to ask it. And in fact, you shouldn't

786
00:44:42,560 --> 00:44:46,157
even have to ask. It should proactively
say, "Hey, I noticed that your favorite

787
00:44:46,240 --> 00:44:50,717
band, because it knows you well, is in
town. Tickets went on sale. I got you

788
00:44:50,800 --> 00:44:54,957
the tickets exactly where you want for
you and your family, and um you know, I

789
00:44:55,040 --> 00:44:57,517
only had 10 minutes to do it because
everything is going to go so fast. So, I

790
00:44:57,600 --> 00:45:00,397
just purchased it proactively. Hope
that's okay." And if you've built enough

791
00:45:00,480 --> 00:45:03,677
trust with the agent, so you've given
it permission to do that, and it kind of

792
00:45:03,760 --> 00:45:05,917
knows that, well, these tickets are
within your your price range, within

793
00:45:06,000 --> 00:45:09,117
your budget, these other ones. Oh, no,
no, I I got to get permission for that,

794
00:45:09,200 --> 00:45:14,157
right? That's the kind of system that we
want. So trust is the key word. EQ also

795
00:45:14,240 --> 00:45:17,117
very important that if you've shared
some information like you've talked

796
00:45:17,200 --> 00:45:20,557
about a health condition, maybe that's
okay information for the AI to talk

797
00:45:20,640 --> 00:45:22,157
about with a parent, but maybe it's not

798
00:45:22,240 --> 00:45:24,077
okay to talk about with a stranger,

799
00:45:24,160 --> 00:45:29,517
right? All of that judgment and and also
observability and really helping you

800
00:45:29,600 --> 00:45:33,677
build up that trust relationship with
the AI so that you know what you know

801
00:45:33,760 --> 00:45:36,957
when it's operating on your behalf that
it'll pursue your goals. Um that's very

802
00:45:37,040 --> 00:45:41,517
very critical. But the same thing same
kind of technology the underlying AGI

803
00:45:41,600 --> 00:45:45,037
stack one stack one way of doing
connectors so it's able to hook up to

804
00:45:45,120 --> 00:45:48,637
this context also applies to the
enterprise and we're really seeing it

805
00:45:48,720 --> 00:45:51,917
not just for software engineering which
by the way has been totally transformed

806
00:45:52,000 --> 00:45:55,277
this year and to the point of AGI being
a gradual thing or a thing where we just

807
00:45:55,360 --> 00:45:59,357
constantly update our expectations like
we all kind of forget what software was

808
00:45:59,440 --> 00:46:03,277
one year ago totally different >> very
interesting thing within open AI is

809
00:46:03,360 --> 00:46:06,957
that we're at basically 100% adoption
of chat GBT work across every function

810
00:46:07,040 --> 00:46:10,637
right sales finance everyone
and whenever you know we roll

811
00:46:10,720 --> 00:46:13,597
out a new model and it goes away or you
know the system breaks for whatever

812
00:46:13,680 --> 00:46:18,317
internal reason um everyone's so upset
they're I cannot do my work and you're

813
00:46:18,400 --> 00:46:20,957
like I don't get it six months ago you
didn't have this and you're doing your

814
00:46:21,040 --> 00:46:24,717
work just fine like why are you you know
why are you complaining and you realize

815
00:46:24,800 --> 00:46:29,357
people adapt so quickly and learn how to
get the most out of these tools and once

816
00:46:29,440 --> 00:46:32,797
you're there you also just sort
of forget that it was any other way and so

817
00:46:32,880 --> 00:46:36,237
there's something about this where
I think we will have these incredible

818
00:46:36,320 --> 00:46:40,157
amazing tools these is assistance
that will remove so much of the toil that

819
00:46:40,240 --> 00:46:42,717
will mean that you can spend your time
doing what you want to do. That you can

820
00:46:42,800 --> 00:46:46,557
focus on building a better life, you can
focus on if you want to be making more

821
00:46:46,640 --> 00:46:48,957
money, if you want to be relaxing more,
if you want to spend more time with your

822
00:46:49,040 --> 00:46:52,157
family, any of those things, you will
have the agency to do it. You have the

823
00:46:52,240 --> 00:46:55,357
agency to build. If you have a vision,
you have creativity, you want it to come

824
00:46:55,440 --> 00:46:57,597
into the world, you will have the tools

825
00:46:57,680 --> 00:47:02,877
to do that. So, I think that what chat
GBT will do is help you with what you

826
00:47:02,960 --> 00:47:06,077
want to accomplish. If you want to make
it your world better, that is what chat

827
00:47:06,160 --> 00:47:08,557
will help you accomplish.
>> And the implications of that are that

828
00:47:08,640 --> 00:47:11,197
what the coding AI tools have done to

829
00:47:11,280 --> 00:47:15,997
engineering this could do to all
knowledge work, all kinds of fields. And

830
00:47:16,080 --> 00:47:19,917
that's how you guys used to talk about
AGI is like, you know, it was basically

831
00:47:20,000 --> 00:47:23,437
AI that could do most economically
viable work better than humans. And what

832
00:47:23,520 --> 00:47:26,557
you're describing now in the
product kind of sounds like that.

833
00:47:26,640 --> 00:47:29,677
>> I think that the amplification we see
in software engineering, we expect across

834
00:47:29,760 --> 00:47:34,477
every field for sure. And there
is something that I think that we is also

835
00:47:34,560 --> 00:47:36,477
worth really thinking about which is the

836
00:47:36,560 --> 00:47:39,277
software engineers rather than having

837
00:47:39,360 --> 00:47:43,677
more vacation time and having more chill
job they're all working harder than ever

838
00:47:43,760 --> 00:47:45,677
before. >> Yeah. You guys
work a lot of hours here.

839
00:47:45,760 --> 00:47:49,197
>> Yes. And I think across the whole
industry people are talking about that

840
00:47:49,280 --> 00:47:53,437
because now you feel like if your agents
aren't running it's just wasted

841
00:47:53,520 --> 00:47:57,437
opportunity. Right. It's like and I know
that feeling because we have spent years

842
00:47:57,520 --> 00:48:00,557
of we have these GPU clusters.
you're either running something on them or

843
00:48:00,640 --> 00:48:03,677
you've lost the opportunity. You can't
get that time back. It's not like the

844
00:48:03,760 --> 00:48:07,117
GPU hours that you're not using while
you're asleep overnight. You can bank

845
00:48:07,200 --> 00:48:11,037
them up and use them later. It's use
it or lose it. And that is very much the

846
00:48:11,120 --> 00:48:13,037
opportunity that everyone has in front

847
00:48:13,120 --> 00:48:15,917
of them is you have almost an AGI, maybe

848
00:48:16,000 --> 00:48:20,877
soon truly an AGI in your pocket. Can do
anything you want. What is it you want?

849
00:48:20,960 --> 00:48:23,597
It's >> waiting on you. It's waiting
on your direction, your guidance, your

850
00:48:23,680 --> 00:48:25,597
management. >> It would be
nice if eventually we can

851
00:48:25,680 --> 00:48:28,717
get some time back from AI.
>> That's right. So like right now I do

852
00:48:28,800 --> 00:48:30,877
feel like it it will help you work more.

853
00:48:30,960 --> 00:48:34,717
>> That's right. And so I think that I
think that that is that is exactly a

854
00:48:34,800 --> 00:48:38,477
next step, right? Is really moving
from this like you're micromanaging this

855
00:48:38,560 --> 00:48:41,197
agent. You have to provide it with these
tasks and you have to give the context

856
00:48:41,280 --> 00:48:42,877
because it just doesn't have it steer.

857
00:48:42,960 --> 00:48:46,397
>> Exactly. You want something that has
much more like you think about like a

858
00:48:46,480 --> 00:48:50,957
really good assistant who really is
plugged in and you've given it again you

859
00:48:51,040 --> 00:48:52,157
have to build up trust with it, right?

860
00:48:52,240 --> 00:48:55,197
You have to really say like okay I trust
you with this context. I trust you with

861
00:48:55,280 --> 00:48:58,957
this responsibility. I trust you with
this job. whatever it is. And once it's

862
00:48:59,040 --> 00:49:02,477
there, it should be able to proactively
run and really execute. You wake up in

863
00:49:02,560 --> 00:49:06,157
the morning, you know, say, "Hey,
here are five things I need from you to

864
00:49:06,240 --> 00:49:08,637
unblock me. You know, I just need
approval on this. Like, here's a

865
00:49:08,720 --> 00:49:09,997
question I have." Those kinds of things.

866
00:49:10,080 --> 00:49:13,037
As you're drinking your morning
coffee, you say yes, no, yes, no.

867
00:49:13,120 --> 00:49:16,637
>> Then you go come back, you know, late
at night once again, you've got five more

868
00:49:16,720 --> 00:49:20,237
things. And so, I think that it's going
to be a world where it's going to be

869
00:49:20,320 --> 00:49:23,757
different from, I think, the kind
of AGI world that many of us would have

870
00:49:23,840 --> 00:49:26,477
pictured 5 10 years ago, right? It's not

871
00:49:26,560 --> 00:49:29,037
like I think that this integration of

872
00:49:29,120 --> 00:49:34,237
humans really being maximally empowered
with this tool, but you still have

873
00:49:34,320 --> 00:49:37,117
involvement, right? That you still have
oversight, you still have judgment, you

874
00:49:37,200 --> 00:49:39,757
still have goals. And also on the
flip side, I think you still have

875
00:49:39,840 --> 00:49:43,917
accountability, right? The question of is,
you know, if something doesn't get

876
00:49:44,000 --> 00:49:47,837
done, whose fault is it? I don't think
that blaming JBT saying, "Oh, it was too

877
00:49:47,920 --> 00:49:52,077
lazy, didn't do the thing," is going
to fly, right? I think that you as the

878
00:49:52,160 --> 00:49:55,677
director, as the person who's who's sort
of running the show, you know that you

879
00:49:55,760 --> 00:49:58,477
need to really feel that accountability,
make sure the task gets done the way

880
00:49:58,560 --> 00:50:00,957
that you want. So, you have
this incredible tool, incredible

881
00:50:01,040 --> 00:50:03,997
amplification, but just like >>
it's almost like everyone's going to

882
00:50:04,080 --> 00:50:07,277
become the CEO of this corporation of as

883
00:50:07,360 --> 00:50:11,597
many agents as you as you care
to orchestrate. Um, it's going to be just

884
00:50:11,680 --> 00:50:16,157
like a very empowering world where
people can do so much more of they want

885
00:50:16,240 --> 00:50:19,757
what they want to do. Um, and I think
that it'll be a world where people will

886
00:50:19,840 --> 00:50:23,757
be able to make more money. I think
that people will be able to improve their

887
00:50:23,840 --> 00:50:26,957
health. I think you'll be able to learn
anything you want to learn and you'll be

888
00:50:27,040 --> 00:50:31,597
able to spend your time the way that you
want to. And I think all of that means

889
00:50:31,680 --> 00:50:35,197
that if you have ambitions that you
want to realize that they're more insight

890
00:50:35,280 --> 00:50:38,637
than ever. And I I know this very
concretely from talking to different

891
00:50:38,720 --> 00:50:42,477
users, different customers, people who
run small businesses on top of chatbt

892
00:50:42,560 --> 00:50:46,077
and say, "I would not have been able
to do this otherwise." like people who

893
00:50:46,160 --> 00:50:50,157
actually quit their job to pursue
a business because they said hey I could

894
00:50:50,240 --> 00:50:55,037
do this now because chat can help me
with so much of it and that kind of

895
00:50:55,120 --> 00:50:58,237
empowerment where people wouldn't have
been able to take the leap before but

896
00:50:58,320 --> 00:51:02,077
now can I think that that is it's like
the American dream on steroids like that

897
00:51:02,160 --> 00:51:04,157
is what we're unlocking >> and it's
I mean you're you're painting

898
00:51:04,240 --> 00:51:08,077
such a rosy picture and I do see that
too even in my own work as like a you

899
00:51:08,160 --> 00:51:11,837
know solo media entrepreneur and we can
land it here but I also you know people

900
00:51:11,920 --> 00:51:15,357
feel very negatively about AI especially
in the United States as And I'm

901
00:51:15,440 --> 00:51:18,877
wondering if like the stakes are
actually you have to deliver on the

902
00:51:18,960 --> 00:51:22,557
product you're describing that will
empower people this way for that to

903
00:51:22,640 --> 00:51:25,677
shift. Maybe that is the thing that
causes sentiment to shift because right

904
00:51:25,760 --> 00:51:28,397
now it feels increasingly negative in a

905
00:51:28,480 --> 00:51:33,357
lot of corners of society. Uh and I'm
wor I'm curious like as the person you

906
00:51:33,440 --> 00:51:37,437
know responsible now for for the
business and chatbt and all of that do

907
00:51:37,520 --> 00:51:39,517
you feel that do you feel that connection?

908
00:51:39,600 --> 00:51:42,397
>> We we feel it. We feel the
responsibility. We feel the gravity and

909
00:51:42,480 --> 00:51:46,877
we definitely feel the criticality
of both us delivering on this vision and

910
00:51:46,960 --> 00:51:49,917
helping people understand it.
Helping people understand who we are

911
00:51:50,000 --> 00:51:53,277
>> better than we have in the past and
help them understand our intentions, why

912
00:51:53,360 --> 00:51:57,437
we're doing this. And I do think that
there's a different set of sentiments

913
00:51:57,520 --> 00:51:58,717
depending on where you go in the world.

914
00:51:58,800 --> 00:52:02,717
Yeah. Right. That it's very clear
in America like we understand that we have

915
00:52:02,800 --> 00:52:06,317
a lot of uh you know we need to deliver
on this this vision to really help

916
00:52:06,400 --> 00:52:11,277
people lift themselves up to really
empower people. um countries like Korea,

917
00:52:11,360 --> 00:52:14,477
Japan, I think that actually sentiment
is very positive towards these tools.

918
00:52:14,560 --> 00:52:19,037
People really see that potential and I
think that where America sits right now

919
00:52:19,120 --> 00:52:23,117
is at an incredible moment with
incredible opportunity and position.

920
00:52:23,200 --> 00:52:28,077
Like we are the leader in this technology.
This is going to be the most

921
00:52:28,160 --> 00:52:29,837
important technology that humanity ever

922
00:52:29,920 --> 00:52:31,837
creates, right? This is going to be this

923
00:52:31,920 --> 00:52:36,077
technology that will really shape how

924
00:52:36,160 --> 00:52:40,317
people live, how people work, how people
accomplish goals, how economy, how the

925
00:52:40,400 --> 00:52:44,477
economy runs, how countries relate
to each other, all of these things. And

926
00:52:44,560 --> 00:52:49,597
America leads in that. And I think that
that opportunity is one that we should

927
00:52:49,680 --> 00:52:53,517
not take for granted. We should take
make the most of it. And so that has

928
00:52:53,600 --> 00:52:57,437
been one of the biggest realizations
for me over the course of OpenAI is to

929
00:52:57,520 --> 00:53:01,117
realize that this is going to be so
important and it has to be something

930
00:53:01,200 --> 00:53:03,437
that is broadly empowering, right? It

931
00:53:03,520 --> 00:53:06,797
has to be a technology that's available

932
00:53:06,880 --> 00:53:11,277
to everyone to lift themselves up.
But America leading and ensuring that the

933
00:53:11,360 --> 00:53:14,637
values that are represented by this
technology, the way that it plays out in

934
00:53:14,720 --> 00:53:16,877
the world being something that enhances

935
00:53:16,960 --> 00:53:18,797
democratic values that really does

936
00:53:18,880 --> 00:53:23,037
uplift all of those those um you know

937
00:53:23,120 --> 00:53:27,357
the core principles that we hold dear
that is at the core and so I think that

938
00:53:27,440 --> 00:53:31,597
it's about how we use it but also
about how do we use this leadership

939
00:53:31,680 --> 00:53:33,917
opportunity and how do we
ensure we don't squander it.

940
00:53:34,000 --> 00:53:36,797
>> Yeah, those are all
my questions. Do you

941
00:53:36,880 --> 00:53:40,157
feel like there's anything on your mind
that you want to say or anything that

942
00:53:40,240 --> 00:53:42,077
we're missing? >> So many things.

943
00:53:42,160 --> 00:53:44,157
>> There's a lot of things. >> There's
so much ground to cover. I think

944
00:53:44,240 --> 00:53:47,837
at the end of the day, what we are doing

945
00:53:47,920 --> 00:53:50,557
is in some ways incredibly complex.

946
00:53:50,640 --> 00:53:53,117
There's so many facets to what we're
building. You think about consumer and

947
00:53:53,200 --> 00:53:56,077
enterprise and the fact we're bringing
these things together and that what an

948
00:53:56,160 --> 00:53:59,117
enterprise is is going to change because
people are going to have this unlocked

949
00:53:59,200 --> 00:54:02,717
wave of entrepreneurship. All that's real.
We have this research program. The

950
00:54:02,800 --> 00:54:06,157
research isn't just about we're we're
off writing papers. It's like people who

951
00:54:06,240 --> 00:54:10,557
really care about how their technology
will be used, making sure the safety and

952
00:54:10,640 --> 00:54:14,397
alignment that uh slides, doc,
spreadsheet, all of those are good. It's

953
00:54:14,480 --> 00:54:19,037
just like all of this complexity, but at
the core, there's something very simple.

954
00:54:19,120 --> 00:54:23,597
We have a mission. Our mission is to
ensure artificial general intelligence

955
00:54:23,680 --> 00:54:27,097
benefits all of humanity. That's what
we're here to do. Everything we do flows

956
00:54:27,180 --> 00:54:30,637
[music] from that one singular goal.
So that's what we're pursuing. That's why

957
00:54:30,720 --> 00:54:35,800
we started this place and that
is what we're committed to.