1
00:00:00,000 --> 00:00:03,677
We're now in the AGI era. Astra has
really hit something that I'm like,

2
00:00:03,760 --> 00:00:07,357
"Okay, I think this is pretty reasonable
to call it AGI. We've seen it run

3
00:00:07,440 --> 00:00:10,957
coherently [clears throat] for 24 hours
to go accomplish tasks that I think are

4
00:00:11,040 --> 00:00:14,237
quite amazing. >> The models
will be plenty powerful, but

5
00:00:14,320 --> 00:00:17,357
it'll be hard to get to everybody given
that we won't have enough compute to

6
00:00:17,440 --> 00:00:19,197
serve it all. >> You don't
win the Super Bowl by saying,

7
00:00:19,280 --> 00:00:20,157
"I want to [music] win the Super Bowl."

8
00:00:20,240 --> 00:00:23,277
You win it by blocking, tackling.
You really have to make sure that safety,

9
00:00:23,360 --> 00:00:25,677
security, alignment, those are all
standards that you're constantly

10
00:00:25,760 --> 00:00:27,997
upleveling. I think that's going
[music] to be a huge challenge people are

11
00:00:28,080 --> 00:00:31,517
underestimating. You've made two very
big bets in your career. Helping build

12
00:00:31,600 --> 00:00:34,397
Stripe early and helping
of course co-ound open AAI.

13
00:00:34,480 --> 00:00:37,677
>> I throughout OpenAI have always
focused on [music] whatever is the most

14
00:00:37,760 --> 00:00:41,357
important problem. For the past 2
years it's been the data centers, the

15
00:00:41,440 --> 00:00:42,877
infrastructure, the machine learning.

16
00:00:42,960 --> 00:00:44,717
>> Do you know what the next
year we'll focus on or

17
00:00:44,800 --> 00:00:49,800
>> I think this is going to become
the most important conversation.

18
00:00:51,200 --> 00:00:53,197
>> Greg, welcome to the A&Z podcast.

19
00:00:53,280 --> 00:00:56,877
>> Thank you for having me. So, Greg,
you've made two very big bets in your

20
00:00:56,960 --> 00:01:00,797
career, uh, helping build Stripe early
and helping, of course, co-found uh,

21
00:01:00,880 --> 00:01:05,277
Open AI. If we were talking 10 years ago
and you were predicting what would the

22
00:01:05,360 --> 00:01:08,877
world look like in 2026 as it relates to
AI, would you be able to predict that we

23
00:01:08,960 --> 00:01:12,557
would be making the breakthroughs
that you've uh, you know, ma made today or

24
00:01:12,640 --> 00:01:16,797
what would you tell them about what
you would expect? Well, so Ilie and I

25
00:01:16,880 --> 00:01:20,397
actually spent a lot of time trying
to predict what it would look like, what

26
00:01:20,480 --> 00:01:22,237
the timelines would be. And I remember

27
00:01:22,320 --> 00:01:24,637
we did some math on compute in around

28
00:01:24,720 --> 00:01:28,797
2016, 2017. And we kind of came to the

29
00:01:28,880 --> 00:01:33,517
conclusion that if you look at Mo's law
progress, that kind of thing, 15 years

30
00:01:33,600 --> 00:01:36,637
felt like about the timeline to AGI.
And if you really squinted at it, you're

31
00:01:36,720 --> 00:01:40,397
willing to scale up and build massive
supercomputers, spend the hundreds of

32
00:01:40,480 --> 00:01:42,557
billions of dollars, that kind of thing,

33
00:01:42,640 --> 00:01:45,757
that maybe it' be 10 And so I actually

34
00:01:45,840 --> 00:01:50,237
feel like in some ways obviously what's
happening it's remarkable. It's this

35
00:01:50,320 --> 00:01:53,917
amazing sort of moment for everyone
to be a part of and to be able to help

36
00:01:54,000 --> 00:01:58,397
shape collectively. But it also feels
a little bit like maybe it's kind of the

37
00:01:58,480 --> 00:02:01,597
conclusion of like a lot of forces that
are all coming together for this moment.

38
00:02:01,680 --> 00:02:05,597
You if you step back and really take
that sort of macro view it kind of makes

39
00:02:05,680 --> 00:02:07,517
sense it's happening now. And do you

40
00:02:07,600 --> 00:02:11,277
think um we're currently because you

41
00:02:11,360 --> 00:02:15,677
guys slightly underestimated
the timeline or I I guess it was basically

42
00:02:15,760 --> 00:02:18,157
on on point. Um do you think we're still

43
00:02:18,240 --> 00:02:22,957
on the timeline given we're now starting
to drive real shortages on the supply

44
00:02:23,040 --> 00:02:23,997
chain?

45
00:02:24,080 --> 00:02:26,957
>> Well, look, I do think that we are in a

46
00:02:27,040 --> 00:02:31,197
world where it is hard for compute
to keep up with the demand that we're

47
00:02:31,280 --> 00:02:34,317
already seeing in the market, right? And
just in terms of how people are going to

48
00:02:34,400 --> 00:02:36,557
use this [clears throat] technology,
benefit from it. I do think it's going

49
00:02:36,640 --> 00:02:41,117
to be very hard for us to scale the raw
potential and capability of these models

50
00:02:41,200 --> 00:02:45,197
to everyone and that's part of what we
try to do and so I think that the the

51
00:02:45,280 --> 00:02:48,637
progress I do see like we have line
of sight to continue to make the models

52
00:02:48,720 --> 00:02:51,197
much more capable, safe and aligned but

53
00:02:51,280 --> 00:02:55,757
also really distributing that power
and the benefits and the empowerment to to

54
00:02:55,840 --> 00:02:58,877
everyone. I think that's going to be a
huge challenge people are underestimating.

55
00:02:58,960 --> 00:03:01,677
>> Right. Ah interesting. So we'll we'll

56
00:03:01,760 --> 00:03:03,517
have the mo the models will be plenty

57
00:03:03,600 --> 00:03:06,877
powerful um or or they'll continue a

58
00:03:06,960 --> 00:03:11,437
pace uh but they'll be it'll be hard to

59
00:03:11,520 --> 00:03:14,397
get to everybody is certainly in an
affordable way given that we won't have

60
00:03:14,480 --> 00:03:17,517
enough compute to serve it all. I think
that's true and I do think we're at a

61
00:03:17,600 --> 00:03:20,477
point now where we have to really start
thinking about what we call pacing the

62
00:03:20,560 --> 00:03:24,157
frontier. And so thinking about as we
move to more capable models, you really

63
00:03:24,240 --> 00:03:27,197
have to make sure that safety, security,
alignment, those are all standards that

64
00:03:27,280 --> 00:03:30,957
you're constantly upleveling.
And those actually become almost the the

65
00:03:31,040 --> 00:03:34,637
bottleneck to progress or sort of the
the sort of you know the the part that

66
00:03:34,720 --> 00:03:38,077
you have to uh spend a lot of your effort
to make sure you've gotten right.

67
00:03:38,160 --> 00:03:42,317
And so I think in my mind it's more
those constraints and the compute. I

68
00:03:42,400 --> 00:03:46,157
think we can make it happen. And then
on the flip side, I think yeah, this

69
00:03:46,240 --> 00:03:48,797
bringing it to everyone, which
is ultimately about our mission, right?

70
00:03:48,880 --> 00:03:52,237
It's empower everyone, ensure it
benefits everyone. That's something that

71
00:03:52,320 --> 00:03:54,317
I think deserves a lot more
airtime than it's gotten.

72
00:03:54,400 --> 00:03:57,757
>> Yeah. Actually, let's get a little
deeper on the safety thing because it's

73
00:03:57,840 --> 00:04:00,077
been very interesting to me in that it

74
00:04:00,160 --> 00:04:03,117
felt like in the beginning, safety was

75
00:04:03,200 --> 00:04:04,957
like, okay, let's make these things not

76
00:04:05,040 --> 00:04:07,917
say nasty stuff that people like don't

77
00:04:08,000 --> 00:04:11,517
like. And so the approach that was taken

78
00:04:11,600 --> 00:04:14,717
was kind of a surface around the edges.

79
00:04:14,800 --> 00:04:17,757
Okay, we'll um you know put some filters

80
00:04:17,840 --> 00:04:21,597
on this guy and we'll RHF

81
00:04:21,680 --> 00:04:25,437
you know around the edges. But like if
you get deep into the thing you'll be

82
00:04:25,520 --> 00:04:29,197
able to get the bad words out. But if
somebody wants to go through that to

83
00:04:29,280 --> 00:04:31,757
hear bad words themselves, who cares? Uh

84
00:04:31,840 --> 00:04:34,957
but then you know when you get into okay

85
00:04:35,040 --> 00:04:36,637
now these things are really good at uh

86
00:04:36,720 --> 00:04:39,997
cyber hacking and other kinds of ideas.

87
00:04:40,080 --> 00:04:44,797
Um now you need kind of a more
architectural idea where the model

88
00:04:44,880 --> 00:04:48,877
itself knows not to reward
hack in a way that's

89
00:04:48,960 --> 00:04:51,037
going to be dangerous and so forth. And

90
00:04:51,120 --> 00:04:54,557
how do you feel like yes we can make

91
00:04:54,640 --> 00:04:58,877
progress against that fast or is that like
a really hard different category of

92
00:04:58,960 --> 00:05:02,157
problem or how how are you thinking
about that? Well, I absolutely think we

93
00:05:02,240 --> 00:05:04,797
can and are making very rapid progress

94
00:05:04,880 --> 00:05:09,437
on this problem. I think that there's
>> a lot of both great ideas and research

95
00:05:09,520 --> 00:05:13,837
that we've been investing in for many
years. Actually, if you rewind to 2017,

96
00:05:13,920 --> 00:05:16,157
uh that I think people underappreciate

97
00:05:16,240 --> 00:05:20,877
some of the most key results that came
out of OpenAI in the field at the time.

98
00:05:20,960 --> 00:05:24,717
So, both kind of the first inklings
of modern language models. You can find a

99
00:05:24,800 --> 00:05:28,717
paper paper from 2017 that kind of laid
that out with LCMS and you know it's

100
00:05:28,800 --> 00:05:31,037
like kind of this very baby result but

101
00:05:31,120 --> 00:05:35,117
also reward uh learning uh reinforcement

102
00:05:35,200 --> 00:05:37,037
learning from human preferences that was

103
00:05:37,120 --> 00:05:39,037
also created in 2017 to start thinking

104
00:05:39,120 --> 00:05:44,397
about how can you align a model to match
what humans want right by providing

105
00:05:44,480 --> 00:05:47,117
feedback from people.
>> Yeah. Just for usability.

106
00:05:47,200 --> 00:05:51,197
>> Exactly. In 2017 2018 we had ideas for
if you have something that's very smart

107
00:05:51,280 --> 00:05:54,637
and capable how can you actually
supervise what it's doing how can you

108
00:05:54,720 --> 00:05:56,077
provide feedback and ensure that it's

109
00:05:56,160 --> 00:05:59,117
staying aligned with you and we had

110
00:05:59,200 --> 00:06:01,837
ideas such as debate or uh iterative

111
00:06:01,920 --> 00:06:06,637
amplification. So these are ideas that
were really sort of at this phase before

112
00:06:06,720 --> 00:06:10,877
these systems existed and you can start
to see the sort of trickle down of those

113
00:06:10,960 --> 00:06:15,437
ideas into modern systems and and
investment. And so it's in some ways

114
00:06:15,520 --> 00:06:20,077
that I think that there was this early
phase when open started where we really

115
00:06:20,160 --> 00:06:23,757
were thinking about AGI safety things
like that and that was very front and

116
00:06:23,840 --> 00:06:28,877
center in even the comms. And then I
think that as things like chatb took off

117
00:06:28,960 --> 00:06:30,637
then people started to see okay well

118
00:06:30,720 --> 00:06:36,077
we're not at this point yet and so
the is the AI politically neutral and

119
00:06:36,160 --> 00:06:38,557
questions like that start to become
to the front and center >> right

120
00:06:38,640 --> 00:06:42,637
>> and now that we're here all these other
ideas that we've been talking about for

121
00:06:42,720 --> 00:06:46,557
a long time they're taking the main
stage again and I think we've been sort

122
00:06:46,640 --> 00:06:49,997
of thinking about this moment for a long
time. That's really good news. And and

123
00:06:50,080 --> 00:06:52,797
when you think about and we don't have

124
00:06:52,880 --> 00:06:55,037
like a really great community yet

125
00:06:55,120 --> 00:06:58,477
amongst the uh soda models, but it seems

126
00:06:58,560 --> 00:07:01,357
like those kinds of ideas

127
00:07:01,440 --> 00:07:05,837
you and Google and Anthropic and uh

128
00:07:05,920 --> 00:07:08,957
SpaceX would want to share and Meta now

129
00:07:09,040 --> 00:07:11,037
um as opposed to like okay this is a

130
00:07:11,120 --> 00:07:13,437
proprietary idea that's a way to keep

131
00:07:13,520 --> 00:07:15,997
these models safe since you're all on

132
00:07:16,080 --> 00:07:20,477
related architectures. um or h how do
you see that unfolding or is everybody

133
00:07:20,560 --> 00:07:23,357
going to do it independently?
>> Well, I think there's nuance here and I

134
00:07:23,440 --> 00:07:27,037
do think the coordination is going to be
a very important theme, right? To really

135
00:07:27,120 --> 00:07:29,277
think about within the frontier labs and

136
00:07:29,360 --> 00:07:31,757
really just thinking broadly about what

137
00:07:31,840 --> 00:07:34,957
has to happen for humanity as a whole

138
00:07:35,040 --> 00:07:39,197
>> to sort of navigate this technology
in the best way. I think that we we're

139
00:07:39,280 --> 00:07:41,677
going to have to really think hard
about those kinds of questions. And we

140
00:07:41,760 --> 00:07:44,637
published a lot of our thoughts. And
again, some of this is about pacing the

141
00:07:44,720 --> 00:07:48,317
frontier. Some of this is about
unilateral actions that we can take and

142
00:07:48,400 --> 00:07:51,837
how we think about how do you make
safety cases for even training and

143
00:07:51,920 --> 00:07:55,677
developing and evaluating these kinds
of models. All that's new. No one's ever

144
00:07:55,760 --> 00:07:58,957
really had to operationalize this before.
And I think it is not at all

145
00:07:59,040 --> 00:08:02,717
unique to OpenAI. Like there's a whole
world that is basically developing this

146
00:08:02,800 --> 00:08:04,637
technology. And I think one thing it's

147
00:08:04,720 --> 00:08:09,917
easy to miss is that what we're building
is almost a a sort of thing that falls

148
00:08:10,000 --> 00:08:13,117
out of compute progress. And in some ways,
compute progress is something that

149
00:08:13,200 --> 00:08:16,797
falls out of technological progress. And
so there's this this massive wave that's

150
00:08:16,880 --> 00:08:20,397
been building for a very long time. And
we're starting to see the leading edges

151
00:08:20,480 --> 00:08:24,317
of this technology. And companies like
OpenAI can lead by a bit in order to

152
00:08:24,400 --> 00:08:27,997
kind of peer into this future and really
understand what is possible. How do we

153
00:08:28,080 --> 00:08:31,117
shape this technology? But we can't
do that alone. And I think that having

154
00:08:31,200 --> 00:08:33,917
coordination and especially the more
that we can talk about safety techniques

155
00:08:34,000 --> 00:08:37,917
and share what we're seeing, alignment
failures, those kinds of things, all of

156
00:08:38,000 --> 00:08:41,597
that is going to again take a very
front seat for this next phase.

157
00:08:41,680 --> 00:08:45,277
>> Right. Very interesting.
>> You've called the OpenAI hugging face

158
00:08:45,360 --> 00:08:49,197
recent incident a watershed moment and
talked about how the defender window is

159
00:08:49,280 --> 00:08:52,797
now open. Can you explain that statement
and the significance behind it?

160
00:08:52,880 --> 00:08:54,557
>> So I think hugging face shows two

161
00:08:54,640 --> 00:08:58,237
things. one is call it a something for

162
00:08:58,320 --> 00:09:01,437
us in terms of how we monitor sandbox

163
00:09:01,520 --> 00:09:05,517
and control the models during evaluation
and that's something we've really risen

164
00:09:05,600 --> 00:09:09,197
to that occasion our team has totally
changed so much of our internal

165
00:09:09,280 --> 00:09:13,597
standards and and really implemented
a lot of controls that I think are very

166
00:09:13,680 --> 00:09:17,677
important and very critical as we look
to to future more capable models but

167
00:09:17,760 --> 00:09:19,277
there's a second thing that I think is

168
00:09:19,360 --> 00:09:21,357
also very valuable for the world that

169
00:09:21,440 --> 00:09:25,117
came out of this which is a insight into

170
00:09:25,200 --> 00:09:30,157
what future capabilities will be like
when they are broadly diffused and in

171
00:09:30,240 --> 00:09:33,357
the hands of threat actors and that will
happen right that there are so many

172
00:09:33,440 --> 00:09:36,717
people who are building these models
and again there's something very important

173
00:09:36,800 --> 00:09:41,517
and good about the diffusion broadly
of AI capabilities because there's a risk

174
00:09:41,600 --> 00:09:46,397
of concentration of power if one or a few
entities people huge risk right it's

175
00:09:46,480 --> 00:09:48,077
something not not to not to at all write

176
00:09:48,160 --> 00:09:50,797
off but you also have to prepare for if

177
00:09:50,880 --> 00:09:54,557
if everyone is empowered with tools that

178
00:09:54,640 --> 00:09:56,637
are cyber capable. And in the case of

179
00:09:56,720 --> 00:09:59,277
Hugging Face, you saw both an AI that

180
00:09:59,360 --> 00:10:04,477
was able to hack out of a secure
environment and hack into a company's

181
00:10:04,560 --> 00:10:08,237
production environment. And I think
that the takeaway >> very cleverly,

182
00:10:08,320 --> 00:10:10,157
>> very cleverly, right?
And it's like that

183
00:10:10,240 --> 00:10:13,037
the things that it found
were were quite sophisticated.

184
00:10:13,120 --> 00:10:17,277
>> Yes. >> And this capability
broadly diffused, I

185
00:10:17,360 --> 00:10:18,877
think, is something that will really

186
00:10:18,960 --> 00:10:22,077
empower threat actors in new ways. And I

187
00:10:22,160 --> 00:10:24,317
think that defenders need to use this

188
00:10:24,400 --> 00:10:29,437
time before that technology is broadly
available to secure themselves. And the

189
00:10:29,520 --> 00:10:32,717
nice thing about it is it's a dual use,
right? It's something where if you can

190
00:10:32,800 --> 00:10:35,597
find vulnerabilities, if you're
an attacker, you can use it for no good.

191
00:10:35,680 --> 00:10:40,157
But if you're a defender, you can patch,
right? If you're a defender, you control

192
00:10:40,240 --> 00:10:45,277
the battleground, right? You control the
setup of your systems. And so our belief

193
00:10:45,360 --> 00:10:49,197
right now is that there's this window of
you have frontier capabilities. You have

194
00:10:49,280 --> 00:10:51,597
the broadly diffused capabilities and

195
00:10:51,680 --> 00:10:56,317
you as a defender by default you know
your security is probably pretty static

196
00:10:56,400 --> 00:10:59,917
been static for the past 5 10 years
that kind of thing. you need to move use

197
00:11:00,000 --> 00:11:03,197
these frontier capabilities that you
you'll have differential access to right

198
00:11:03,280 --> 00:11:05,677
where we have trusted access programs
things like that to bring these

199
00:11:05,760 --> 00:11:09,837
capabilities to defenders and you can
use that to move yourself up so that as

200
00:11:09,920 --> 00:11:14,397
the frontier capabilities get better you
get pulled along too right okay so I've

201
00:11:14,480 --> 00:11:16,077
got a comment and a question on it I

202
00:11:16,160 --> 00:11:17,517
would say there's a third thing that we

203
00:11:17,600 --> 00:11:21,597
learn which is like these things have

204
00:11:21,680 --> 00:11:23,437
capabilities that I don't know that we

205
00:11:23,520 --> 00:11:29,037
all understood before which on the good
side like oh I can deploy 10,000 agents

206
00:11:29,120 --> 00:11:32,237
and they can talk to each other and
organize themselves and do stuff for me.

207
00:11:32,320 --> 00:11:35,597
Like that's uh pretty amazing. So that

208
00:11:35,680 --> 00:11:38,237
that was on the good side. On the other

209
00:11:38,320 --> 00:11:41,517
side, so I agree that we've got a kind

210
00:11:41,600 --> 00:11:45,677
of defense window. Um however,

211
00:11:45,760 --> 00:11:50,077
we have like 50 years of code and

212
00:11:50,160 --> 00:11:53,197
architectural ideas um and deployment

213
00:11:53,280 --> 00:11:55,837
ideas that weren't built for this world.

214
00:11:55,920 --> 00:11:59,917
And so yes, the AI can help us like

215
00:12:00,000 --> 00:12:02,237
okay, find a bug, patch a bug, and so

216
00:12:02,320 --> 00:12:05,997
forth. But it seems like there's, you

217
00:12:06,080 --> 00:12:07,917
know, maybe a bigger issue, which is we

218
00:12:08,000 --> 00:12:10,797
have these huge, you know, massive

219
00:12:10,880 --> 00:12:13,277
honeypotss of consumer data and all

220
00:12:13,360 --> 00:12:15,037
these things lying all over the

221
00:12:15,120 --> 00:12:18,237
internet. Uh, and you know, from a

222
00:12:18,320 --> 00:12:21,277
consumer standpoint, it's like, okay,

223
00:12:21,360 --> 00:12:24,077
I can't protect my stuff. uh these all

224
00:12:24,160 --> 00:12:25,837
these companies have to get their act

225
00:12:25,920 --> 00:12:29,917
together which um seems a bit worrisome

226
00:12:30,000 --> 00:12:32,957
and do you think kind of in the future

227
00:12:33,040 --> 00:12:35,757
we need a do we need a decentralized

228
00:12:35,840 --> 00:12:40,397
consumer architecture like will this
kind of current world that we live in

229
00:12:40,480 --> 00:12:43,997
with all these centralized
data repositories

230
00:12:44,080 --> 00:12:47,917
be viable in a world of AI >> so
several pieces to the answer and

231
00:12:48,000 --> 00:12:50,077
first to your point on what you can get

232
00:12:50,160 --> 00:12:55,117
out of 10,000 agents we actually use
10,000 agents to solve the Navier Stokes

233
00:12:55,200 --> 00:12:57,197
problem. [laughter] >> Yeah,
that was pretty pretty awesome by

234
00:12:57,280 --> 00:12:58,637
the way. Congratulations on that.

235
00:12:58,720 --> 00:13:02,237
>> Thank you. Thank you. And it's both
an important problem for what it is has

236
00:13:02,320 --> 00:13:06,077
significant implications
and applications to fluid dynamics to how

237
00:13:06,160 --> 00:13:09,277
you think about ocean currents,
all these things, but for what it

238
00:13:09,360 --> 00:13:13,277
represents, right, of new knowledge
created by AI and it unlocking a whole

239
00:13:13,360 --> 00:13:16,717
wave of scientific discovery, medicines,
all those things, they're on the table

240
00:13:16,800 --> 00:13:20,397
now. So I think there's something really
amazing to think about what can happen

241
00:13:20,480 --> 00:13:23,037
through the power of AI that is able to

242
00:13:23,120 --> 00:13:25,997
really help solve problems. And in the

243
00:13:26,080 --> 00:13:28,877
case of of cyber security, how I think

244
00:13:28,960 --> 00:13:34,077
about it, we at OpenAI took our models

245
00:13:34,160 --> 00:13:39,197
and applied them to finding
vulnerabilities. We took 25% of our

246
00:13:39,280 --> 00:13:43,117
production engineers and said, "Sorry,
all your projects are on hold. You are

247
00:13:43,200 --> 00:13:46,717
now defending. You are now upleveling
our security architecture. you're going

248
00:13:46,800 --> 00:13:48,717
to use the models to find all the holes

249
00:13:48,800 --> 00:13:53,837
and we found a number of of of serious
issues and we fixed them and I've talked

250
00:13:53,920 --> 00:13:57,517
to a number of CISOs over the past
couple couple weeks and months and there

251
00:13:57,600 --> 00:14:00,557
are many companies who are also telling
me that yeah that they've applied these

252
00:14:00,640 --> 00:14:03,917
models they found some very significant
issues but that they're able to fix them

253
00:14:04,000 --> 00:14:08,397
and one positive sort of part of the story
is that when we took Astra pointed

254
00:14:08,480 --> 00:14:13,277
out our our systems we found some new
new problems but eventually it saturated

255
00:14:13,360 --> 00:14:17,037
we basically found to our knowledge
all of the P zeros, all of the critical

256
00:14:17,120 --> 00:14:19,517
problems that Astra
is smart enough to find.

257
00:14:19,600 --> 00:14:22,077
>> And of course, there will be a new
model. There will be a new round.

258
00:14:22,160 --> 00:14:24,557
>> Even smarter. >> Exactly.
But I think that that's the

259
00:14:24,640 --> 00:14:27,437
world that we'll be in is that you'll be
in a world where you want to be in this

260
00:14:27,520 --> 00:14:32,397
tight loop of new cyber capability drops.
You deploy it against your

261
00:14:32,480 --> 00:14:36,717
systems. You find the new holes and
ideally you've managed to automate this

262
00:14:36,800 --> 00:14:39,757
what we call defense factory. And that's
what we're building internally. this end

263
00:14:39,840 --> 00:14:42,397
to end of both find vulnerability,

264
00:14:42,480 --> 00:14:46,797
triage it, >> remediate,
deploy, validate, right? That

265
00:14:46,880 --> 00:14:50,397
end to end. And if you can do that at
machine speed, I think the defenders

266
00:14:50,480 --> 00:14:53,997
will be advantaged in deeply
significant ways. And there are

267
00:14:54,080 --> 00:14:58,717
>> ideas, for example, formerly verifying
all of software that are possible with AI.

268
00:14:58,800 --> 00:15:02,077
>> Yeah, we never that's always been
a dream. We've had these ver formal

269
00:15:02,160 --> 00:15:06,157
languages and all these kinds of things,
but they never kind of took off.

270
00:15:06,240 --> 00:15:08,957
>> That's right. Because just it's just
intractable for people. It's just so

271
00:15:09,040 --> 00:15:12,077
hard. But we have these
that are solving these crazy

272
00:15:12,160 --> 00:15:16,877
>> impossible math problems. And so one
application of that that proving power

273
00:15:16,960 --> 00:15:19,837
right actually one of the things to know
about the navier stokes problem is that

274
00:15:19,920 --> 00:15:22,717
we formalized it right.
We formalized into lean and so

275
00:15:22,800 --> 00:15:25,837
>> so the AIs can write
verifiable code >> they can.

276
00:15:25,920 --> 00:15:28,397
>> Yeah. Very nice. Yeah that's a great

277
00:15:28,480 --> 00:15:31,757
idea. So I think I think there's real

278
00:15:31,840 --> 00:15:36,397
hope but I think that our view is that the
world needs to act with urgency because

279
00:15:36,480 --> 00:15:38,797
>> we're in a very dangerous
window right now. Yeah.

280
00:15:38,880 --> 00:15:41,357
>> We just see it coming.
>> Closing the loop on on on this in

281
00:15:41,440 --> 00:15:45,997
incident. Is there anything you felt
that the the narrative um gotten wrong

282
00:15:46,080 --> 00:15:49,837
in an important way or or or is there
any preferred way of of talking about

283
00:15:49,920 --> 00:15:52,557
what what happened or or or this window
that that's important to get across when

284
00:15:52,640 --> 00:15:56,157
you think about the public narrative?
or or is there anything just inside the

285
00:15:56,240 --> 00:16:00,237
company that uh else that that changed
in terms of how you're approaching the

286
00:16:00,320 --> 00:16:04,077
the set of issues? >> Well,
two things. I think that one big

287
00:16:04,160 --> 00:16:07,437
theme that people should take away from
it is a question of access, right? That

288
00:16:07,520 --> 00:16:12,397
these these capabilities exist right now
in the world, but that they are in small

289
00:16:12,480 --> 00:16:15,437
number of frontier companies and the
frontier companies have a trusted access

290
00:16:15,520 --> 00:16:19,677
program, which means that anyone who's
not in the trusted access program is not

291
00:16:19,760 --> 00:16:22,637
really able to benefit from the fact
that there's this differential. the

292
00:16:22,720 --> 00:16:26,717
people we're in, they get a benefit if
they use it. And so I think that there's

293
00:16:26,800 --> 00:16:31,437
something we need to do as a field
and as a society to really scale up the

294
00:16:31,520 --> 00:16:35,357
number of defenders that have access
to these technologies because that's it's

295
00:16:35,440 --> 00:16:40,157
like every day matters and to use those
days you need access to these tools. And

296
00:16:40,240 --> 00:16:43,197
one thing that was actually interesting
about the hugging face response was that

297
00:16:43,280 --> 00:16:47,437
they said that they used frontier models
to try to look over the logs of what had

298
00:16:47,520 --> 00:16:51,037
happened because that's the only way
to actually analyze an attack like this.

299
00:16:51,120 --> 00:16:54,477
and that they said the frontier models
refuse, but they didn't actually try our

300
00:16:54,560 --> 00:16:56,717
frontier models and they actually
believe that ours would have permitted

301
00:16:56,800 --> 00:17:00,317
it. And so there is something too about
the default stance of providers. I think

302
00:17:00,400 --> 00:17:02,877
there's something here about using these

303
00:17:02,960 --> 00:17:06,157
capabilities for good with this urgency

304
00:17:06,240 --> 00:17:10,717
and and this sort of sort of I yeah
this this this real sense that it has to

305
00:17:10,800 --> 00:17:14,477
happen. I'll tell a quick story by the
way which is unrelated but maybe also

306
00:17:14,560 --> 00:17:18,077
shows a little bit about how I think
about this. I remember when we trained

307
00:17:18,160 --> 00:17:20,877
GPD3. It was beginning of December 2019.

308
00:17:20,960 --> 00:17:23,837
So, everyone's about to head out
on vacation. Y's like, "Okay, we can train

309
00:17:23,920 --> 00:17:27,517
the model." And I just remember feeling
like this model is sitting on a shelf.

310
00:17:27,600 --> 00:17:30,957
No one is using it. It's this like
amazing technology, new to the world,

311
00:17:31,040 --> 00:17:34,317
new to humanity. It's like every day
that no one is exploring what it's

312
00:17:34,400 --> 00:17:37,437
capable of and trying to understand it,
figure out what to do with it. Um,

313
00:17:37,520 --> 00:17:39,117
that's a day that is lost to the world.

314
00:17:39,200 --> 00:17:42,157
And so, I was just like, I canceled
basically all my holiday plans. I spent

315
00:17:42,240 --> 00:17:45,517
the whole time just playing with the
model, building interfaces around it,

316
00:17:45,600 --> 00:17:48,557
trying to see what it was capable of.
I remember I was trying to teach it how to

317
00:17:48,640 --> 00:17:51,917
how to sort lists some of numbers.
It didn't work very well. Uh but it was

318
00:17:52,000 --> 00:17:55,037
just like this like really try to to to
probe it and see what's possible. And I

319
00:17:55,120 --> 00:17:59,197
think that that spirit and ethos
is something I think we should bring to

320
00:17:59,280 --> 00:18:02,957
what we're building today. It's it's
sort of obviously at much larger scale,

321
00:18:03,040 --> 00:18:07,597
much larger impact, but we as a world
have the opportunity to understand this

322
00:18:07,680 --> 00:18:11,997
technology in this moment, which then
helps us shape and steer where it will

323
00:18:12,080 --> 00:18:14,717
go next. >> And very good point.

324
00:18:14,800 --> 00:18:17,677
>> You you said you two things. What what
do you have another one? One was access

325
00:18:17,760 --> 00:18:20,317
or did you say both of them?
>> I think I said both of them. Yes. Yes.

326
00:18:20,400 --> 00:18:23,597
>> The um let's go back to Astro. Uh it's

327
00:18:23,680 --> 00:18:27,597
incredible to see all the excitement
on on X u all sorts of cases. People are

328
00:18:27,680 --> 00:18:29,437
excited about computer use. Um you you

329
00:18:29,520 --> 00:18:31,677
you've said that in some ways it is uh

330
00:18:31,760 --> 00:18:37,037
bring us closer to AGI. What do you talk
about what you find most compelling in

331
00:18:37,120 --> 00:18:39,917
Astra or what you think the the
breakthrough there in light of that

332
00:18:40,000 --> 00:18:42,717
statement and and uh where
we have still left to go?

333
00:18:42,800 --> 00:18:46,397
>> Actually wait sorry let me let me
actually revise my my answer and I can

334
00:18:46,480 --> 00:18:49,277
say I so both access and let me also

335
00:18:49,360 --> 00:18:51,517
tell another story about how I have used

336
00:18:51,600 --> 00:18:54,157
the models personally. So after hugging

337
00:18:54,240 --> 00:18:59,277
face, I was thinking about how
can I use these models in my

338
00:18:59,360 --> 00:19:02,317
personal life? What can I do to secure
myself? And I have a website. It's a

339
00:19:02,400 --> 00:19:06,717
very simple website, gregarin.com.
Not the most popular website. Got a good

340
00:19:06,800 --> 00:19:09,837
blog post on it. Exactly. You got some
blog post. It's a static site. It's very

341
00:19:09,920 --> 00:19:13,277
simple. Like what what kind
of vulnerabilities could be there? So I

342
00:19:13,360 --> 00:19:17,917
took my codeex and asked it go check
out gregman.com. Tell me if there's any

343
00:19:18,000 --> 00:19:22,397
vulnerabilities. So I did a pen test and
it came back with 13 findings. And these

344
00:19:22,480 --> 00:19:26,237
findings were things like I had set
my SPF record so that you would prevent

345
00:19:26,320 --> 00:19:31,197
people from spoofing emails, right?
That there was some it wasing over HTTP

346
00:19:31,280 --> 00:19:34,877
without forcing people to HTTPS,
things like that. And individually, these

347
00:19:34,960 --> 00:19:37,837
things are maybe not the biggest deal,
but if you think about with an AI that's

348
00:19:37,920 --> 00:19:40,957
able to chain together many small
vulnerabilities into a big one, I'm

349
00:19:41,040 --> 00:19:44,237
like, do I really want a hole where
someone can scoop emails for me?

350
00:19:44,320 --> 00:19:46,957
Probably not. So 15 minutes for it to

351
00:19:47,040 --> 00:19:49,597
find these 13 findings. But then I asked

352
00:19:49,680 --> 00:19:51,757
it, can you fix these? Right. Because

353
00:19:51,840 --> 00:19:54,397
fixing is so annoying. So painful and

354
00:19:54,480 --> 00:19:58,237
boring. Exactly. And so 45 minutes it

355
00:19:58,320 --> 00:20:01,837
opened up my cloud for control panel.
It clicked around, set all the headers. It

356
00:20:01,920 --> 00:20:04,957
migrated me to COD for pages. It's like
set everything correctly. It started the

357
00:20:05,040 --> 00:20:08,317
demarked process, which apparently
you have to do like a 48 hour window of

358
00:20:08,400 --> 00:20:13,997
whatever. Um, and that was 45 minutes
of of fixing. And I felt so protected. I

359
00:20:14,080 --> 00:20:17,357
felt like, wow. >> Did you ask
it to find the vulnerabilities?

360
00:20:17,440 --> 00:20:19,997
>> There you go. No, I added Yeah.
So it actually does do that automatically.

361
00:20:20,080 --> 00:20:22,877
This was by six soul. It said I just
checked that this one's fixed, this

362
00:20:22,960 --> 00:20:26,477
one's fixed, this one's fixed, and in 48
hours I'm going to have to run and set

363
00:20:26,560 --> 00:20:29,437
up a little automation. So in 48 hours,
it would check back in to complete the

364
00:20:29,520 --> 00:20:33,597
demark process. And I was like,
all right, this is we're in business now.

365
00:20:33,680 --> 00:20:37,517
>> All right. All right.
Craig Brockman.com.

366
00:20:37,600 --> 00:20:40,397
>> There we go.
You too can't be protected.

367
00:20:40,480 --> 00:20:43,517
>> Awesome. >> Let's um
let's transition to Astra. Uh

368
00:20:43,600 --> 00:20:46,717
it's incredible to see all
the excitement online and the use cases.

369
00:20:46,800 --> 00:20:50,557
people really excited about computer use
among other things. You you've said it's

370
00:20:50,640 --> 00:20:55,277
uh sort of closer to the way along to AGI.
I'm curious what you find most

371
00:20:55,360 --> 00:20:58,397
groundbreaking with it and where
do you think we still left to go?

372
00:20:58,480 --> 00:21:01,757
>> Well, I think that Astra is really a

373
00:21:01,840 --> 00:21:07,197
step function on so many axes and in
many ways it is the sum of a number of

374
00:21:07,280 --> 00:21:10,557
research bets that we've been making
for years and to see them come into one

375
00:21:10,640 --> 00:21:13,917
model at one time it's been absolutely
incredible. And so just one thing to

376
00:21:14,000 --> 00:21:16,797
know about how we do numbering is that

377
00:21:16,880 --> 00:21:21,677
we kind of have been wanting to have
GPD6 represent something that's worthy

378
00:21:21,760 --> 00:21:25,117
of it. And the problem we always
have is that our models are kind of

379
00:21:25,200 --> 00:21:28,237
incrementally getting better. And so
it's just never feels like it's a right

380
00:21:28,320 --> 00:21:29,837
moment to go for a major version bump.

381
00:21:29,920 --> 00:21:33,677
You always have be like it's 56 now
should be 57. And this one just happened

382
00:21:33,760 --> 00:21:37,117
to be because all these things came
together at once the first time that we

383
00:21:37,200 --> 00:21:40,397
actually had this almost discontinuous
step on in a way that we could have

384
00:21:40,480 --> 00:21:45,037
predicted, but it just was like all
these these these factors um happened to

385
00:21:45,120 --> 00:21:47,197
line up at once. And so I was a real

386
00:21:47,280 --> 00:21:52,557
positive moment. And to me the computer
use is the headline thing that we've

387
00:21:52,640 --> 00:21:53,997
talked about. And part of the reason

388
00:21:54,080 --> 00:21:56,317
computer use is so significant is that

389
00:21:56,400 --> 00:22:01,277
for agentic use cases, it really comes
down to tools. It's like is the model

390
00:22:01,360 --> 00:22:04,397
smart enough to use the tools and then
does it have access to the context that

391
00:22:04,480 --> 00:22:07,277
needs to through these tools?
And so people have been building these MCP

392
00:22:07,360 --> 00:22:11,917
servers and these CLIs and just really
sort of taking the world of software and

393
00:22:12,000 --> 00:22:16,237
making accessible in this almost stilted
way that is not really meant for humans,

394
00:22:16,320 --> 00:22:17,677
right? It's like we're kind of retooling

395
00:22:17,760 --> 00:22:21,117
the world, right? You made it. It's kind

396
00:22:21,200 --> 00:22:23,677
of like, oh, we'll build an API like

397
00:22:23,760 --> 00:22:27,837
it's a software, but like what if it's

398
00:22:27,920 --> 00:22:29,837
really more behaving like a human? Can

399
00:22:29,920 --> 00:22:32,637
it just use a computer and and it's kind

400
00:22:32,720 --> 00:22:35,917
of and the result of building that other

401
00:22:36,000 --> 00:22:40,797
layer you have now another layer of
security challenges this that the other.

402
00:22:40,880 --> 00:22:42,957
>> Exactly. >> Really so
it's kind of Yeah. They've

403
00:22:43,040 --> 00:22:45,757
always felt like very
weird and suboptimal.

404
00:22:45,840 --> 00:22:49,677
>> Yes. And from the very beginning
of OpenAI I remember in November 2015 we

405
00:22:49,760 --> 00:22:53,197
did this offsite in Napa and we talked
about our plans. We actually laid out

406
00:22:53,280 --> 00:22:55,997
this three-step plan that basically
is what we ended up following for the next

407
00:22:56,080 --> 00:22:57,997
10 years. But we also talked about what

408
00:22:58,080 --> 00:23:00,637
if we could do reinforcement learning

409
00:23:00,720 --> 00:23:05,037
where the environment is screen pixels,
keyboard, mouse, right? Same interface

410
00:23:05,120 --> 00:23:08,717
as a human. Suddenly any sort of task
you could do with a computer is in

411
00:23:08,800 --> 00:23:13,277
there. It's in it's in distribution
as long as set aside sound, whatever. But

412
00:23:13,360 --> 00:23:17,357
you basically have the full power of a
computer there. And we had some aborted

413
00:23:17,440 --> 00:23:20,637
attempts early on to try to build agents
that could do that. And so it really

414
00:23:20,720 --> 00:23:25,677
took us until now, but you're seeing the
power immediately. And it's just been so

415
00:23:25,760 --> 00:23:29,357
cool to see people take the Blender
capabilities and, you know, take a

416
00:23:29,440 --> 00:23:32,877
screenshot of something and they have
to make a 3D model and you can actually,

417
00:23:32,960 --> 00:23:36,077
you know, lots of people are now
designing houses or trying to redesign

418
00:23:36,160 --> 00:23:40,237
uh their living room, all those things
by just utilizing this capability. And

419
00:23:40,320 --> 00:23:41,917
to me that the the thing that really

420
00:23:42,000 --> 00:23:45,597
stands out is that you can now

421
00:23:45,680 --> 00:23:50,077
move forward on AI that can do things
for you without you having to build all

422
00:23:50,160 --> 00:23:53,437
these specific connectors. And I think
there's so much software they don't even

423
00:23:53,520 --> 00:23:56,957
think about that you have to orchestrate
every day. And like how much of your

424
00:23:57,040 --> 00:24:00,317
life is like clicking around menus
and like you know typing things into a

425
00:24:00,400 --> 00:24:04,157
spreadsheet and things like that. Like
none of that is what we should be doing

426
00:24:04,240 --> 00:24:07,517
100 years ago. No one is doing any
of these things. And so it's not crazy to

427
00:24:07,600 --> 00:24:11,837
think that in 5 years 10 years no one
will be doing any of this stuff anymore.

428
00:24:11,920 --> 00:24:14,477
That we will get our time back. We're
not going to have to be get our carpal

429
00:24:14,560 --> 00:24:18,717
tunnel or hunch shoulders or you know
all of those all those physical problems

430
00:24:18,800 --> 00:24:20,397
that are us contorting to the machine.

431
00:24:20,480 --> 00:24:22,397
It's now the machine is there to help us

432
00:24:22,480 --> 00:24:25,037
to empower us to to really serve us.

433
00:24:25,120 --> 00:24:29,117
>> Yeah. And you know that's a really
good point because I think you know one of

434
00:24:29,200 --> 00:24:31,197
the things that that that you have been

435
00:24:31,280 --> 00:24:34,237
um I would say more sober on as a

436
00:24:34,320 --> 00:24:37,197
company is just okay what happens with

437
00:24:37,280 --> 00:24:41,517
employment and uh I think that I think

438
00:24:41,600 --> 00:24:44,397
that that's exactly right that there's
all these things that we do because we

439
00:24:44,480 --> 00:24:48,237
have to do and like they became valuable
but we shouldn't be doing it. All they

440
00:24:48,320 --> 00:24:50,317
do is wreck our health and and wreck our

441
00:24:50,400 --> 00:24:53,197
personalities. And

442
00:24:53,280 --> 00:24:57,677
the idea that humans are going to just
run out of ideas of cool things to do or

443
00:24:57,760 --> 00:24:59,677
how to make the world better or problems

444
00:24:59,760 --> 00:25:03,917
to solve um seems a little absurd to me.

445
00:25:04,000 --> 00:25:06,797
And so far, at least in the numbers, the

446
00:25:06,880 --> 00:25:11,437
better AI gets, the higher employment
goes, not the lower. And so I wonder

447
00:25:11,520 --> 00:25:14,637
your and of course it's unknowable.
You know, we've never had this technology

448
00:25:14,720 --> 00:25:17,517
before. It's getting better and so forth.
So, how do you kind of think

449
00:25:17,600 --> 00:25:22,077
about the future of employment as it
relates to these models and AI as it

450
00:25:22,160 --> 00:25:24,877
progresses? >> Well,
I do have a fundamental belief

451
00:25:24,960 --> 00:25:29,357
that AI is surprising. And I think we
even put this in the OpenAI launch post

452
00:25:29,440 --> 00:25:33,117
back in the day in 2015, just saying
that the history so far has been somehow

453
00:25:33,200 --> 00:25:35,597
it just doesn't play out the way
that you think it does, even when there's

454
00:25:35,680 --> 00:25:38,797
this like logical conclusion it should
be a certain way. And I think the same

455
00:25:38,880 --> 00:25:41,677
will be true, right? I think that
there's something that we've learned

456
00:25:41,760 --> 00:25:44,237
about that humans and I think like for

457
00:25:44,320 --> 00:25:47,037
any job that we almost it's easy to not

458
00:25:47,120 --> 00:25:51,437
give it as much credit for how deep
the field is and how much sort of

459
00:25:51,520 --> 00:25:54,877
sophistication building relationships
accountability is a good example of

460
00:25:54,960 --> 00:25:58,397
something where I think that people
setting goals and being accountable for

461
00:25:58,480 --> 00:26:01,597
outcomes like those feel fundamental
to me those feel like things that we

462
00:26:01,680 --> 00:26:05,037
actually should preserve for the long
term right that's something that feels

463
00:26:05,120 --> 00:26:08,317
like deeply human like people are not
valuable just because we can do tasks

464
00:26:08,400 --> 00:26:11,917
right we're valuable because of her
people and I think that it's important

465
00:26:12,000 --> 00:26:15,037
not to lose sight of that in some
of these narratives and I think that the

466
00:26:15,120 --> 00:26:19,517
way that things will change and how
what we do with our time evolves and we

467
00:26:19,600 --> 00:26:22,557
clearly will be in a world of abundance
and how do we ensure that that abundance

468
00:26:22,640 --> 00:26:25,917
is broadly distributed um but also
at the same time I think that we should be

469
00:26:26,000 --> 00:26:29,917
in a world where the ceiling of ambition
is higher than ever before and I think

470
00:26:30,000 --> 00:26:32,957
we're going to see a wave
of entrepreneurship where it's actually

471
00:26:33,040 --> 00:26:37,277
already starting I've heard um from
someone in you know particular industry

472
00:26:37,360 --> 00:26:41,197
was saying that a bunch of people in in
his world are now making the leap to go

473
00:26:41,280 --> 00:26:44,717
quit and start their own firms and that
they're doing it because they have these

474
00:26:44,800 --> 00:26:48,557
AI tools and they're just like I can do
so much more and so it's the barriers to

475
00:26:48,640 --> 00:26:53,117
entry and for entries and this should
be a wonderful renaissance. Yeah though

476
00:26:53,200 --> 00:26:55,357
well it's it's been a lot of fun for us

477
00:26:55,440 --> 00:26:58,957
in just going okay there are you know

478
00:26:59,040 --> 00:27:03,517
particularly for our young people
because you know they get by default the

479
00:27:03,600 --> 00:27:08,237
grunt work but like what if the AI does
to the grunt work then they can really

480
00:27:08,320 --> 00:27:10,397
develop much faster actually because

481
00:27:10,480 --> 00:27:16,237
they can um kind of get involved on the
you know really the the real part of our

482
00:27:16,320 --> 00:27:18,717
business which is what is the
relationship with the entrepreneur? How

483
00:27:18,800 --> 00:27:20,877
do we open up the world for them? How do

484
00:27:20,960 --> 00:27:25,677
we make um them feel like oh they can do

485
00:27:25,760 --> 00:27:29,997
anything and they're an important CEO
and they can go build things and as

486
00:27:30,080 --> 00:27:32,317
opposed to you know spend the whole

487
00:27:32,400 --> 00:27:34,397
weekend writing an investment memo which

488
00:27:34,480 --> 00:27:39,357
by the way uh I have to say Astra very
good at writing investment it's awesome

489
00:27:39,440 --> 00:27:42,477
>> I love hearing that >> yeah
and again I do think it's going to

490
00:27:42,560 --> 00:27:45,197
be a nuance story right I don't think
that we should paint that everything's

491
00:27:45,280 --> 00:27:48,077
going to be rosy and it's all going
to be just easy I think it's going to be

492
00:27:48,160 --> 00:27:51,757
hard I think there's going to be change
but I think that it can be a much better

493
00:27:51,840 --> 00:27:55,677
world. I think the future could be much
better for than the past for everyone.

494
00:27:55,760 --> 00:27:57,837
>> Yeah, that and that's
it feels like what

495
00:27:57,920 --> 00:28:02,797
we should expect, you know,
like before the plow,

496
00:28:02,880 --> 00:28:07,277
you know, the world was a lot worse like
it was just a worse life even though it

497
00:28:07,360 --> 00:28:11,837
did put like a lot of human labor out
of business, you know, and and created the

498
00:28:11,920 --> 00:28:15,437
whole lite movement and all those kinds
of things. You know, nobody here wants

499
00:28:15,520 --> 00:28:17,997
to go back to 1870.

500
00:28:18,080 --> 00:28:20,877
uh and so the idea that no we don't want

501
00:28:20,960 --> 00:28:23,357
to go into the future now um seems a

502
00:28:23,440 --> 00:28:26,637
little shortsighted but uh I think the

503
00:28:26,720 --> 00:28:30,557
speed at which things are moving
is is very very scary for people

504
00:28:30,640 --> 00:28:34,557
>> and we we really recognize the fact
of how things are moving and that we spend

505
00:28:34,640 --> 00:28:37,757
a lot of time really trying
to understand as well as we can how people

506
00:28:37,840 --> 00:28:41,597
are feeling how we can be showing up
better and I think that two things like

507
00:28:41,680 --> 00:28:44,877
one is that when we think about
development the pace of progress we're

508
00:28:44,960 --> 00:28:48,957
being very deliberate about it safety
is our foremost priority. We think about

509
00:28:49,040 --> 00:28:52,477
how do we build this technology in a
safe secure way and what should those

510
00:28:52,560 --> 00:28:56,397
standards be and you can see that showing
up in a lot of our comms inside

511
00:28:56,480 --> 00:29:00,237
the building. It is absolutely what
people are thinking about and what we

512
00:29:00,320 --> 00:29:03,917
care about is that we really want this
technology to empower everyone broadly.

513
00:29:04,000 --> 00:29:06,397
And I think that for us as a world to

514
00:29:06,480 --> 00:29:11,597
really think about how do we get
the most out of this technology? How do we

515
00:29:11,680 --> 00:29:14,637
get the benefits? How do we mitigate
the risks? I think this is going to become

516
00:29:14,720 --> 00:29:18,317
the most important conversation that we
have and I think that that will emerge

517
00:29:18,400 --> 00:29:21,277
over even maybe the next one to two years.
I think that this should be

518
00:29:21,360 --> 00:29:24,237
something that is front and center.
I think people sense it yet you can sense

519
00:29:24,320 --> 00:29:27,837
it in how people react right now
and even thinking about things like data

520
00:29:27,920 --> 00:29:30,957
centers and these kinds of questions
of do we want AI and how do we how do we

521
00:29:31,040 --> 00:29:34,077
think about where it's appropriate
and how do we ensure child safety all of

522
00:29:34,160 --> 00:29:38,637
these kinds of questions these are core
questions that we care so much about

523
00:29:38,720 --> 00:29:41,517
getting right >> to that end
why do we think sentiment

524
00:29:41,600 --> 00:29:44,317
and AI is higher certain Asian countries

525
00:29:44,400 --> 00:29:46,957
>> all Asian countries
well and actually in

526
00:29:47,040 --> 00:29:51,757
European countries everywhere but the US
has like got the lowest AI sentiment

527
00:29:51,840 --> 00:29:54,157
>> why is that or more
>> what's driving the

528
00:29:54,240 --> 00:29:56,717
>> what can we do about
like what can we learn from

529
00:29:56,800 --> 00:29:58,717
>> well one thing that I think about is

530
00:29:58,800 --> 00:30:01,757
that I think we as a field as a company

531
00:30:01,840 --> 00:30:03,837
need to do a much better job of

532
00:30:03,920 --> 00:30:06,717
articulating to people why they benefit

533
00:30:06,800 --> 00:30:11,277
why is this a good thing for them
and not just for the country right which I

534
00:30:11,360 --> 00:30:14,957
think that this technology is going
to be and is rapidly becoming the single

535
00:30:15,040 --> 00:30:18,957
most important strategic
priority and resource

536
00:30:19,040 --> 00:30:21,837
for the United States it's happening Yes,

537
00:30:21,920 --> 00:30:27,037
>> absolutely. You look at chatt 300
million health queries or 300 million

538
00:30:27,120 --> 00:30:30,797
people every single week using it
for help, right? That's a huge deal. And

539
00:30:30,880 --> 00:30:35,037
we're at a billion almost, you know,
1.1 billion weekly active users. I think

540
00:30:35,120 --> 00:30:38,077
within the US it's about 100 million
something like that. Like a third of the

541
00:30:38,160 --> 00:30:40,957
population, if I have that number
correct, right, is using chat every

542
00:30:41,040 --> 00:30:44,717
single week. So people are touching
this technology. But I think that that for

543
00:30:44,800 --> 00:30:48,397
many people there are some people who
have gone very deep and really gone

544
00:30:48,480 --> 00:30:52,477
through the health journey for example
that's been true for my family for my

545
00:30:52,560 --> 00:30:56,237
wife that she has a number of health
conditions that would be we don't even

546
00:30:56,320 --> 00:30:59,917
really know how we would have managed
these before chat and there's just so

547
00:31:00,000 --> 00:31:03,917
much toil and time and just getting
to the right answer and a doctor tells you

548
00:31:04,000 --> 00:31:06,957
something you don't know what the thing
is and how do you get that that sort of

549
00:31:07,040 --> 00:31:08,557
sanity check to really even understand

550
00:31:08,640 --> 00:31:13,357
it people who I that their life

551
00:31:13,440 --> 00:31:17,917
was saved through information delivered
by chatbt like I'll tell you a story for

552
00:31:18,000 --> 00:31:19,997
one one of my friends is that she was in

553
00:31:20,080 --> 00:31:25,197
the hospital and uh that the doctor
was about to inject a antibiotic and she's

554
00:31:25,280 --> 00:31:28,717
like give me a moment she typed into
chatbt and Chad said absolutely do not

555
00:31:28,800 --> 00:31:33,437
take that if you do you may die because
you have this thing that you had a year

556
00:31:33,520 --> 00:31:36,637
ago you have this condition like this
kind of thing your reaction showed the

557
00:31:36,720 --> 00:31:41,037
doctor I know right >> and the
doctor said oh my goodness no

558
00:31:41,120 --> 00:31:45,037
that's absolutely right I had no idea.
I only had five minutes to read your

559
00:31:45,120 --> 00:31:50,237
chart. >> Yeah. Many such cases, by the
way, telling me stories in their chart at

560
00:31:50,320 --> 00:31:55,037
least. >> Exactly. And so that these kinds
of stories I think don't get told nearly

561
00:31:55,120 --> 00:31:58,717
enough, but they're out there. I hear
them every day. And the people who run

562
00:31:58,800 --> 00:32:02,237
their small business on chat and would
be totally unable to do it otherwise.

563
00:32:02,320 --> 00:32:06,157
Like that kind of empowerment. Again,
people who are able to save money, make

564
00:32:06,240 --> 00:32:10,877
money, live a better life. Um those
kinds of stories I think need to be in

565
00:32:10,960 --> 00:32:14,957
the public consciousness as we approach
this this question. So it's painting the

566
00:32:15,040 --> 00:32:17,517
narrative that hey you've got a teacher
in your pocket, a doctor in your pocket,

567
00:32:17,600 --> 00:32:20,157
something you know lawyer in your pocket,
therapist in your pocket, you

568
00:32:20,240 --> 00:32:23,357
know all these utilities in your pocket
while also not threatening those the

569
00:32:23,440 --> 00:32:26,797
same well also telling the teachers and
doctors and lawyers therapists that hey

570
00:32:26,880 --> 00:32:29,517
you've now got this tool too and it's
going to make your your business better

571
00:32:29,600 --> 00:32:31,197
as well. >> Yes.
And it's not just the narrative,

572
00:32:31,280 --> 00:32:32,717
it's the reality, right? It's you need

573
00:32:32,800 --> 00:32:37,277
both. I think that many other countries

574
00:32:37,360 --> 00:32:41,597
are looking in. Yeah. >> Seeing
the position that the US is in,

575
00:32:41,680 --> 00:32:44,957
right, that seeing the potential of this
technology. And partly too, you think

576
00:32:45,040 --> 00:32:47,997
about demographics that I think in many
of these other countries, it's more

577
00:32:48,080 --> 00:32:52,557
keenly felt that there's an older gen
generation that's much larger than than

578
00:32:52,640 --> 00:32:55,837
the younger population that is going to
need to support them. These questions of

579
00:32:55,920 --> 00:32:59,437
how is that supposed to work? And so,
I think that there is something about

580
00:32:59,520 --> 00:33:02,957
really thinking to the future
and thinking about what's possible. how do

581
00:33:03,040 --> 00:33:05,757
you get the benefits out of this
technology and really wanting to lean

582
00:33:05,840 --> 00:33:09,197
into that that I think we're seeing
across the world. And so again, I think

583
00:33:09,280 --> 00:33:13,437
that there's something that we need
to do better as a field and as as a as a

584
00:33:13,520 --> 00:33:17,597
company in order to communicate this
domestically, but I think the potential

585
00:33:17,680 --> 00:33:21,837
is there and we're in such a privileged
position and leading this field in a way

586
00:33:21,920 --> 00:33:25,437
that I think was not guaranteed and it's
not guaranteed to remain true for the

587
00:33:25,520 --> 00:33:27,597
future either. >> Yeah.
Particularly if we ban data

588
00:33:27,680 --> 00:33:31,997
centers, I think that that'll be
a problem for us maintaining our lead. It

589
00:33:32,080 --> 00:33:35,837
will it will drive the data centers
overseas which is what happened with

590
00:33:35,920 --> 00:33:37,917
silicon back in you know the 80s or so.

591
00:33:38,000 --> 00:33:39,837
>> Right. Right. And and there are so many

592
00:33:39,920 --> 00:33:41,197
you know one of the interesting things

593
00:33:41,280 --> 00:33:45,597
about data centers is it creates so many

594
00:33:45,680 --> 00:33:47,357
bluecollar

595
00:33:47,440 --> 00:33:50,317
manufacturing jobs. I think uh switch

596
00:33:50,400 --> 00:33:53,997
employs like 45,000

597
00:33:54,080 --> 00:33:58,637
>> uh people on kind of a union contract
basis to build data centers. That's just

598
00:33:58,720 --> 00:34:00,557
one of the data center providers in the

599
00:34:00,640 --> 00:34:03,917
US. and and and they're great jobs,

600
00:34:04,000 --> 00:34:07,997
they're highpaying, uh and then you know

601
00:34:08,080 --> 00:34:11,837
I think that while there have been bad

602
00:34:11,920 --> 00:34:14,557
actors in the data center space um most

603
00:34:14,640 --> 00:34:16,877
of them are very good actors and

604
00:34:16,960 --> 00:34:20,477
contribute uh to the power grid uh don't

605
00:34:20,560 --> 00:34:24,157
waste water and uh are not noisy and so

606
00:34:24,240 --> 00:34:28,637
like not that there was never an issue
but like we could just say hey you have

607
00:34:28,720 --> 00:34:30,957
to be a well-behaved data center as

608
00:34:31,040 --> 00:34:32,957
opposed to like we're going to ban them.

609
00:34:33,040 --> 00:34:38,077
Um or like we're going
to stop AI. Uh which I

610
00:34:38,160 --> 00:34:41,917
like we're not going to even as a country
we're not big enough to stop AI.

611
00:34:42,000 --> 00:34:46,797
Uh so the idea like AI will continue
without us. Uh and then we'll have zero

612
00:34:46,880 --> 00:34:51,037
say as opposed to we're the leaders
and then we have all the say. So it's a

613
00:34:51,120 --> 00:34:52,797
really really important cultural message.

614
00:34:52,880 --> 00:34:55,917
>> I think it's very important and and I
think on on data centers. So we've made

615
00:34:56,000 --> 00:34:58,797
commitments on not increasing people's

616
00:34:58,880 --> 00:35:03,197
electricity bills. Our data centers are
all closed loop water. So the amount of

617
00:35:03,280 --> 00:35:07,197
water used by Abalene which is the data
center that actually trained Astra that

618
00:35:07,280 --> 00:35:10,397
it uses about the same amount of water
as an office building. Right. So it's

619
00:35:10,480 --> 00:35:14,157
it's really it's really yeah that the
technology is quite quite advanced on

620
00:35:14,240 --> 00:35:17,917
that and that we have a number
of community commitments so that we can

621
00:35:18,000 --> 00:35:21,757
actually help um in Ohio and Georgia
where we have data centers. We've

622
00:35:21,840 --> 00:35:24,557
announced we've talked about how we're
providing credits to every college

623
00:35:24,640 --> 00:35:28,397
student for codeex access. So there's
this broad set of benefits that we are

624
00:35:28,480 --> 00:35:31,037
bringing to bear. But again,
I think that we need to do even more.

625
00:35:31,120 --> 00:35:33,517
>> Yeah. >> And you know,
like I think making those

626
00:35:33,600 --> 00:35:35,277
kinds of things a requirement to build a

627
00:35:35,360 --> 00:35:38,717
data center is very reasonable. Um, but

628
00:35:38,800 --> 00:35:41,757
like let's have positive some ideas as

629
00:35:41,840 --> 00:35:44,157
opposed to okay, we're going to jump out

630
00:35:44,240 --> 00:35:49,917
of the AI game as a as a country and let
uh China or whomever dictate what it's

631
00:35:50,000 --> 00:35:52,957
going to be. >> Speaking
of uh contributions, you guys

632
00:35:53,040 --> 00:35:56,077
made a billion dollar commitment
to frontline defenders. Why don't you talk

633
00:35:56,160 --> 00:36:00,957
about that? So we believe that every
organization, every company, every

634
00:36:01,040 --> 00:36:05,197
government, critical infrastructure such
as water service providers, hospitals

635
00:36:05,280 --> 00:36:09,357
should all be using this defender window
to secure themselves. But not every

636
00:36:09,440 --> 00:36:13,997
organization will have the capital
required to do it. So we have a billion

637
00:36:14,080 --> 00:36:18,557
dollar commitment to frontline defenders,
so to organizations that we

638
00:36:18,640 --> 00:36:23,277
all rely on every day in our communities
to access our models to secure

639
00:36:23,360 --> 00:36:26,797
themselves. We think this is the
beginning. This is not the end. We're

640
00:36:26,880 --> 00:36:29,837
working part closely with partners. For
example, CrowdStrike and we are working

641
00:36:29,920 --> 00:36:32,077
together to provide discounted access to

642
00:36:32,160 --> 00:36:33,757
defenders as well. And I think that

643
00:36:33,840 --> 00:36:36,877
there's there should be a global effort

644
00:36:36,960 --> 00:36:41,917
in order to bring these tools to bear
to really secure every single organization

645
00:36:42,000 --> 00:36:44,157
given what we see coming
and what's possible.

646
00:36:44,240 --> 00:36:46,477
>> Yeah. And that's a a super positive

647
00:36:46,560 --> 00:36:50,157
advance because this is our before AI

648
00:36:50,240 --> 00:36:52,317
hospitals were getting broken into, held

649
00:36:52,400 --> 00:36:55,917
hostages all the time. um our water

650
00:36:56,000 --> 00:36:59,837
supply has been hacked by uh foreign

651
00:36:59,920 --> 00:37:03,917
actors um state actors and so like we're

652
00:37:04,000 --> 00:37:08,637
already dealing with kind of our
critical infrastructure was not built

653
00:37:08,720 --> 00:37:12,397
for with cyber security in mind. It's
not been maintained with cyber security

654
00:37:12,480 --> 00:37:14,957
in mind and here's an opportunity to go

655
00:37:15,040 --> 00:37:19,837
from not even secure in a pre I AI world

656
00:37:19,920 --> 00:37:23,757
to completely secure. So to me this is

657
00:37:23,840 --> 00:37:26,237
like an incredibly important effort uh

658
00:37:26,320 --> 00:37:29,117
to you know not have our water supply

659
00:37:29,200 --> 00:37:33,437
and uh our hospitals at risk. I I really
agree with that perspective, right? That

660
00:37:33,520 --> 00:37:38,077
the point of I think that we as a society
have been lax, right? That we've

661
00:37:38,160 --> 00:37:43,037
allowed tech debt to pile up that every
cyber security organization I've never

662
00:37:43,120 --> 00:37:45,917
met a CISO who felt that they were
appropriately resourced, right? That

663
00:37:46,000 --> 00:37:47,597
they were appropriately prioritized.

664
00:37:47,680 --> 00:37:49,837
>> Never. >> And particularly
not in the public sector.

665
00:37:49,920 --> 00:37:53,357
>> That's right. And so I think that we
have to change that. And we should have

666
00:37:53,440 --> 00:37:57,037
changed this years ago, but now is a
moment where we actually have a real

667
00:37:57,120 --> 00:38:00,717
both motivation to do it and a real
ability to do it. And I think that

668
00:38:00,800 --> 00:38:03,917
delivering the secure world that we
all deserve so that we can we can really

669
00:38:04,000 --> 00:38:08,477
depend on it and be safe and secure
in our in our daily lives and online lives

670
00:38:08,560 --> 00:38:10,557
like that to me feels like table stakes.

671
00:38:10,640 --> 00:38:11,997
We absolutely need to do this.

672
00:38:12,080 --> 00:38:14,797
>> Yeah, definitely. No,
that's a that's a great effort.

673
00:38:14,880 --> 00:38:17,757
>> Closing the loop on Astra.
You've emphasized that capabilities still

674
00:38:17,840 --> 00:38:22,077
remain jagged. What do you think um
still still left to go or still needs to

675
00:38:22,160 --> 00:38:26,637
be fleshed out that gets approximates
most of your definition of AGI? Well, I

676
00:38:26,720 --> 00:38:28,637
think that AGI has turned out to be less

677
00:38:28,720 --> 00:38:30,797
of a point in time and more of this sort

678
00:38:30,880 --> 00:38:35,757
of fuzzy spectrum. And for me, Astra
has really hit something that I'm like,

679
00:38:35,840 --> 00:38:40,157
okay, I think this is pretty reasonable
to call it AGI in that with its computer

680
00:38:40,240 --> 00:38:44,317
use capabilities, you really can ask it
to do long live tasks and it'll just do

681
00:38:44,400 --> 00:38:46,557
it. That we've seen it it run coherently

682
00:38:46,640 --> 00:38:49,437
for 24 hours to go accomplish tasks that

683
00:38:49,520 --> 00:38:54,477
I think are quite quite amazing
and across a wide variety of domains. Now,

684
00:38:54,560 --> 00:38:58,477
it still is jagged and so that there
are still places where, for example, it's

685
00:38:58,560 --> 00:39:01,677
writing. It's pretty good writing.
It's the first time it's not sloping.

686
00:39:01,760 --> 00:39:03,277
>> Yeah. >> But it's not great writing.

687
00:39:03,360 --> 00:39:06,797
>> Yeah. >> And I think
that there's a number of

688
00:39:06,880 --> 00:39:11,677
areas where I feel like we just need
to polish it a little bit and it would be

689
00:39:11,760 --> 00:39:15,597
fantastic. And it's just like not
quite there. So I I see this like I saw

690
00:39:15,680 --> 00:39:18,957
someone post a graph on on Twitter of like
you know this like kind of you know

691
00:39:19,040 --> 00:39:23,277
jagged frontier and that where we really
need to to be is a much more steady

692
00:39:23,360 --> 00:39:27,117
across the board um really hit on all
these categories. But I think that what

693
00:39:27,200 --> 00:39:32,397
people are finding is that it is so
capable across such a wide variety of

694
00:39:32,480 --> 00:39:35,997
tasks that it is accelerative.
It is empowering and it's something that I

695
00:39:36,080 --> 00:39:39,757
think we've never really seen a model
that that's been a jump like this.

696
00:39:39,840 --> 00:39:44,397
>> Yeah. One of the things that's been
interesting for me is that a as you

697
00:39:44,480 --> 00:39:48,637
solve problems sometimes the

698
00:39:48,720 --> 00:39:51,197
the world doesn't realize it like so I

699
00:39:51,280 --> 00:39:53,197
haven't seen a hallucination in quite

700
00:39:53,280 --> 00:39:56,717
some time. Um but nobody says oh the

701
00:39:56,800 --> 00:39:58,957
models don't hallucinate anymore. It's

702
00:39:59,040 --> 00:40:01,037
just kind of in the ethos that that's

703
00:40:01,120 --> 00:40:04,157
what AI does. Um, how do you do you

704
00:40:04,240 --> 00:40:06,157
think that'll just go away over time or

705
00:40:06,240 --> 00:40:09,517
is it you know does there need to be

706
00:40:09,600 --> 00:40:14,237
some like continual education for the

707
00:40:14,320 --> 00:40:17,917
for the non- like hardcore tech
people >> I this thing is moving

708
00:40:18,000 --> 00:40:20,957
>> I think one of the most important
problems we actually have is the

709
00:40:21,040 --> 00:40:25,357
continual education right the really how
do you people shouldn't have to extract

710
00:40:25,440 --> 00:40:28,557
from the AI what it's capable of it
should go the other way around the AI

711
00:40:28,640 --> 00:40:31,997
should say hey I can help you in this
new way so we have about you know we

712
00:40:32,080 --> 00:40:35,197
over a billion weekly active users
on chatbt, but I think we have something

713
00:40:35,280 --> 00:40:40,797
like another maybe 1.5 billion people who
have used chatbt and don't use it anymore.

714
00:40:40,880 --> 00:40:43,437
>> Oh wow. >> Right.
So think about that. That's a

715
00:40:43,520 --> 00:40:45,837
significant fraction of the of the planet.

716
00:40:45,920 --> 00:40:50,077
>> And those people exactly those people
we should really be able to go back to and

717
00:40:50,160 --> 00:40:52,717
say, "Hey, we have made so much

718
00:40:52,800 --> 00:40:57,197
progress. We think we can be useful
to you in these ways." And I think that

719
00:40:57,280 --> 00:41:01,597
that is just shows you the kind
of problem we have in front of us is that

720
00:41:01,680 --> 00:41:05,917
these AI like if you look at chatbt
and chatbt work, they're both text boxes,

721
00:41:06,000 --> 00:41:08,877
right? It's like this new text box
is way better than the old text box, but

722
00:41:08,960 --> 00:41:11,757
there's still some things that the old
text box is better at. So don't always

723
00:41:11,840 --> 00:41:15,917
use it. It's like that is not the AI we
were promised. The AI we were promised

724
00:41:16,000 --> 00:41:19,117
should be an AI that you talk to over
voice primarily. You can talk to it over

725
00:41:19,200 --> 00:41:22,557
text if you want to. uh that it has
persistence, that it has memory, it has

726
00:41:22,640 --> 00:41:26,877
context, it knows you. It's trustworthy,
that you have seen it be proactive and

727
00:41:26,960 --> 00:41:29,677
helps solve problems for you, that helps
in your personal life, in your work

728
00:41:29,760 --> 00:41:33,197
life. And that's how it should be. It
should be something that is able to that

729
00:41:33,280 --> 00:41:36,717
you can really sort of rely on for the
things that you care about that empowers

730
00:41:36,800 --> 00:41:39,437
you and and helps you solve your goals.

731
00:41:39,520 --> 00:41:42,317
And I think that being able to explain

732
00:41:42,400 --> 00:41:46,557
to you how it can help you is a core

733
00:41:46,640 --> 00:41:50,077
part of that. >> Yeah.
Interesting. and proactively do

734
00:41:50,160 --> 00:41:52,317
it. That that's such an interesting idea

735
00:41:52,400 --> 00:41:57,757
like we need more
helpfulness out of our AIs.

736
00:41:57,840 --> 00:42:01,437
>> Yes. >> Which is kind of a thing like
some humans aren't and probably the humans

737
00:42:01,520 --> 00:42:05,197
who develop AI are not very helpful
people. I would guess just [laughter]

738
00:42:05,280 --> 00:42:07,277
being around engineers and researchers.

739
00:42:07,360 --> 00:42:10,557
>> You'd be surprised. I think we have
very very uh helpful helpful engineers at

740
00:42:10,640 --> 00:42:13,997
OpenAI. But but there is something about
if you think about how do you work with

741
00:42:14,080 --> 00:42:16,637
another person, right?
A new co-orker you've never worked with.

742
00:42:16,720 --> 00:42:18,797
>> It takes you a little bit of time,
right? you kind of feel them out. You

743
00:42:18,880 --> 00:42:20,317
see how they respond in different areas.

744
00:42:20,400 --> 00:42:23,757
People do not come with an instruction
manual. And often actually sometimes

745
00:42:23,840 --> 00:42:26,957
it's it's interesting in areas like
consulting or something where they they

746
00:42:27,040 --> 00:42:30,557
really lean into like MyersBriggs
and that they do say like here's like a

747
00:42:30,640 --> 00:42:34,237
quick way to know who I am and how I
operate. So that there is some precedent

748
00:42:34,320 --> 00:42:37,837
for how humans can kind of present
a little bit more. You have a resume. You

749
00:42:37,920 --> 00:42:41,437
have sort of track record. People can
ask for back channels on you. So we have

750
00:42:41,520 --> 00:42:45,277
built up a way of how do you understand
how a human will work and what the best

751
00:42:45,360 --> 00:42:47,677
way is to to get the best out of that

752
00:42:47,760 --> 00:42:50,397
person and I think sort of figuring out

753
00:42:50,480 --> 00:42:52,637
what is the right analog for AI and

754
00:42:52,720 --> 00:42:57,597
especially as AI changes and we produce
new tools and product surface and new

755
00:42:57,680 --> 00:42:59,517
models and all these things. I think

756
00:42:59,600 --> 00:43:02,157
that this will be a very important

757
00:43:02,240 --> 00:43:07,437
society company interplay and again I
think that what we what our northstar

758
00:43:07,520 --> 00:43:08,957
should be is simplicity right that we

759
00:43:09,040 --> 00:43:12,077
really should be one AI that's unified

760
00:43:12,160 --> 00:43:17,437
that makes it so easy
and smooth for you to be less

761
00:43:17,520 --> 00:43:20,237
>> engaging with the computer and less
wrapping your yourself around the

762
00:43:20,320 --> 00:43:23,597
computer. The computer should be
there to empower you to help serve you.

763
00:43:23,680 --> 00:43:25,837
>> Right. Right.

764
00:43:25,920 --> 00:43:28,317
>> The the business is ripping. You guys

765
00:43:28,400 --> 00:43:33,197
have such broad surface area in terms
of what you cover. How do you decide in

766
00:43:33,280 --> 00:43:37,677
terms of prioritizing where to go deepest,
what not to build and then also

767
00:43:37,760 --> 00:43:40,957
your role has also evolved and changed
and you've you know encompassed so many

768
00:43:41,040 --> 00:43:44,957
things you know research product
commercialization or management etc. How

769
00:43:45,040 --> 00:43:47,517
are you also thinking
about prizing your time?

770
00:43:47,600 --> 00:43:49,277
>> Well, they go hand in hand. Yeah.

771
00:43:49,360 --> 00:43:52,717
>> So this year the theme was focus.

772
00:43:52,800 --> 00:43:56,637
>> I think that we really realized that we
can't do it all right. We need to pick

773
00:43:56,720 --> 00:44:00,157
and particularly there's one thing
we're trying to accomplish which is our

774
00:44:00,240 --> 00:44:02,957
mission right we want to ensure AGI
benefits all of humanity now how do you

775
00:44:03,040 --> 00:44:07,357
backsolve from that what are the areas
like deployment and productization is

776
00:44:07,440 --> 00:44:09,837
actually something that does reinforce
that right that we do want to bring this

777
00:44:09,920 --> 00:44:13,837
technology to bear and have it uplift
everyone and people deploying it in

778
00:44:13,920 --> 00:44:15,997
useful applications all of that personal

779
00:44:16,080 --> 00:44:17,997
life work life the whole thing um very

780
00:44:18,080 --> 00:44:21,117
core but how do you when you think about

781
00:44:21,200 --> 00:44:23,117
this moment we're in of this agentic

782
00:44:23,200 --> 00:44:25,917
coding takeoff that exponential

783
00:44:26,000 --> 00:44:29,437
what areas reinforced that and which

784
00:44:29,520 --> 00:44:32,957
ones were kind of just sort of you know
they got labeled a side quest in the

785
00:44:33,040 --> 00:44:35,997
media but just were not on track for it
even if they were individually something

786
00:44:36,080 --> 00:44:40,077
very exciting was a very core question
that we had to grapple with and so

787
00:44:40,160 --> 00:44:43,437
things like Sora that's maybe
the highest profile one of these projects

788
00:44:43,520 --> 00:44:46,557
that we decided to cancel
very very painful by the way

789
00:44:46,640 --> 00:44:50,157
>> not an easy thing to do
but it was so critical to

790
00:44:50,240 --> 00:44:53,997
>> unleash the business in many ways so we
could really focus bringing together the

791
00:44:54,080 --> 00:44:59,117
consumer and enterprise side of chats into
chat work. That's another area where we've

792
00:44:59,200 --> 00:45:03,117
really had to focus and really say
this is what we're doing. So, a lot of the

793
00:45:03,200 --> 00:45:06,237
way that we've thought about this to
unlock this moment is to really have

794
00:45:06,320 --> 00:45:09,597
vision about where we think the future's
going and how do we think that the new

795
00:45:09,680 --> 00:45:13,277
capabilities that are emerging can best
be brought to bear with a single unified

796
00:45:13,360 --> 00:45:16,957
stack that works across the different
areas, different walks of life,

797
00:45:17,040 --> 00:45:20,237
different areas that we're trying
to focus on. And it's been painful, right?

798
00:45:20,320 --> 00:45:22,877
that if you look for the first half,
I think that there was just a lot of

799
00:45:22,960 --> 00:45:26,397
metrics that were not looking the
direction that we wanted and that there

800
00:45:26,480 --> 00:45:29,357
was a lot of just sort of telling
the team we just need to focus on the

801
00:45:29,440 --> 00:45:32,397
basics. Like one of my one of my
favorite management books is the score

802
00:45:32,480 --> 00:45:34,477
takes care of itself.
Have you guys read that one?

803
00:45:34,560 --> 00:45:37,597
>> Yeah, Keith boy favorite.
>> Yeah, it's a it's a great one and it

804
00:45:37,680 --> 00:45:40,797
just is a very empowering book because
you just realize it's like you cannot

805
00:45:40,880 --> 00:45:45,277
affect the outcome. You can only affect
the inputs, right? You can only affect

806
00:45:45,360 --> 00:45:48,477
the like the basics. And so focus
on those basics, right? You don't win the

807
00:45:48,560 --> 00:45:51,597
Super Bowl by saying I want to win
the Super Bowl. You win it by blocking and

808
00:45:51,680 --> 00:45:53,597
tackling. Yeah. And so that's what we

809
00:45:53,680 --> 00:45:56,077
have done for this whole year. And for

810
00:45:56,160 --> 00:46:00,957
myself, I throughout OpenAI have
always focused on whatever is the most

811
00:46:01,040 --> 00:46:03,597
important problem that I think that I
can move the needle on that just isn't

812
00:46:03,680 --> 00:46:07,757
going to happen without me. And >>
for the past two years, it's been the

813
00:46:07,840 --> 00:46:12,157
data centers, the infrastructure, the
machine learning engineering. And that's

814
00:46:12,240 --> 00:46:15,197
an area where we really spent a lot
of effort to get our pre-training

815
00:46:15,280 --> 00:46:17,197
infrastructure into great great shape.

816
00:46:17,280 --> 00:46:20,077
This year it's really been about
the business. It's really been about the

817
00:46:20,160 --> 00:46:23,277
okay, we've figured out how to get the
research really humming. We figured out

818
00:46:23,360 --> 00:46:25,917
how to get the infrastructure really
humming, but how do we really bring this

819
00:46:26,000 --> 00:46:30,477
technology to the world? And I think
that that's where I've been really

820
00:46:30,560 --> 00:46:35,197
putting a lot of my efforts in trying to
bring together a bunch of functions that

821
00:46:35,280 --> 00:46:39,117
were otherwise kind of running
in parallel or crosswise. And that is

822
00:46:39,200 --> 00:46:43,117
something where I think as a founder,
as someone who has kind of touched every

823
00:46:43,200 --> 00:46:45,997
part of this business from the beginning,
I think I've been uniquely

824
00:46:46,080 --> 00:46:49,677
able to go in and make the changes,
the make the hard decisions and really

825
00:46:49,760 --> 00:46:54,237
figure out this is the direction. Let's
go. And a lot of my style is that I like

826
00:46:54,320 --> 00:46:57,677
to lead from the trenches. And so I get
very deep in the weeds on what the thing

827
00:46:57,760 --> 00:47:01,517
is and really try to keep asking a lot
of questions. Like that's actually a lot

828
00:47:01,600 --> 00:47:04,157
of my style is just asking like, hey,
does this make sense still? I don't

829
00:47:04,240 --> 00:47:07,597
quite get that. Um, sometimes when
things are confused, for example, over

830
00:47:07,680 --> 00:47:10,237
the past couple days, there have been
times when it's just like, we've got a

831
00:47:10,320 --> 00:47:12,957
thing, we got to figure out how to even
talk about it. How do we think about it?

832
00:47:13,040 --> 00:47:14,157
How should the world think about this?

833
00:47:14,240 --> 00:47:17,037
I'm just like, let's just get everyone
who can touch different parts of the

834
00:47:17,120 --> 00:47:20,877
elephant on a call. We're like going
through a Google doc on a hangout and

835
00:47:20,960 --> 00:47:23,677
just kind of like being like, does this
line make sense? Wait, what do we really

836
00:47:23,760 --> 00:47:27,277
mean by this? And so really trying
to uplevel execution sometimes in small

837
00:47:27,360 --> 00:47:30,317
ways and sometimes large.
>> Yeah. No, that's fantastic.

838
00:47:30,400 --> 00:47:32,397
>> By the way, exact right way to operate.

839
00:47:32,480 --> 00:47:35,917
>> Yeah. Do do you know what the next
year will focus on or prep for now?

840
00:47:36,000 --> 00:47:40,717
>> Well, look, I think that I think
that the business is is a huge area that I

841
00:47:40,800 --> 00:47:44,077
think we're not done yet with really
upleveling every part of execution. So,

842
00:47:44,160 --> 00:47:45,597
I think there's a lot more to do there.

843
00:47:45,680 --> 00:47:49,677
But I also think we are moving into
a new phase of AI development, right? I

844
00:47:49,760 --> 00:47:53,917
call this and we call this we're now
in the AGI era. And I think that that is

845
00:47:54,000 --> 00:47:56,877
something that you can debate. Is it
this model, previous model, next model?

846
00:47:56,960 --> 00:48:00,637
It doesn't matter. The point is that we
are in a new phase where safety,

847
00:48:00,720 --> 00:48:03,517
security, alignment, really thinking
about these things not just at

848
00:48:03,600 --> 00:48:07,357
deployment time, but all the way back
at development time evaluation. It's

849
00:48:07,440 --> 00:48:10,957
objective. This is critical. It must
happen. This is core to our mission.

850
00:48:11,040 --> 00:48:13,997
This is core to what we need to do.
And so a lot of what I spend my time

851
00:48:14,080 --> 00:48:17,837
thinking about is making sure do we have
all the right processes? Are we talking

852
00:48:17,920 --> 00:48:19,757
about the right things? Do we have plans

853
00:48:19,840 --> 00:48:22,717
that really at a operational level, at a

854
00:48:22,800 --> 00:48:25,117
practical level lead us to the kind of

855
00:48:25,200 --> 00:48:29,757
security and variance and the kinds
of safety guarantees that we view as core

856
00:48:29,840 --> 00:48:31,197
to our mission, what we need to do. And

857
00:48:31,280 --> 00:48:34,077
so I think that again the theme of open

858
00:48:34,160 --> 00:48:37,357
AI certainly for the past five years has

859
00:48:37,440 --> 00:48:40,797
been deeper codees, deeper intertwining

860
00:48:40,880 --> 00:48:42,717
across these functions that are maybe on

861
00:48:42,800 --> 00:48:44,637
the surface very disperate, right? all

862
00:48:44,720 --> 00:48:50,417
the way from go to market to long-term
research to chip design by building

863
00:48:50,500 --> 00:48:53,197
[clears throat] these in a coherent way
where everyone has context right they

864
00:48:53,280 --> 00:48:56,877
kind of understand how do I fit into the
overall picture what are we trying to do

865
00:48:56,960 --> 00:49:01,197
and what is the end outcome we want
to achieve like that is what has to happen

866
00:49:01,280 --> 00:49:03,837
so I think that the areas that I will

867
00:49:03,920 --> 00:49:06,957
focus on I think will be dictated by the

868
00:49:07,040 --> 00:49:10,237
areas that most need that intertwining

869
00:49:10,320 --> 00:49:14,557
and I think that I see us moving more
and more in concert in lock up as time

870
00:49:14,640 --> 00:49:17,837
goes on. >> Yeah.
I uh we could go all day, but we

871
00:49:17,920 --> 00:49:20,717
have a hard stop. Uh I think this is a
great place to wrap. Greg, thank you so

872
00:49:20,800 --> 00:49:21,837
much for coming on the podcast.

873
00:49:21,920 --> 00:49:25,880
>> Thank you. >> Great. Fantastic.