1
00:00:00,160 --> 00:00:03,917
We have Greg Brockman, the co-founder
and president of OpenAI with us. Welcome

2
00:00:04,000 --> 00:00:06,077
to the show, Greg. How are you doing?

3
00:00:06,160 --> 00:00:07,837
>> Doing great. Thank you for having me.

4
00:00:07,920 --> 00:00:12,397
>> Thanks for hopping on. Uh huge day.
Can we start with the with the math

5
00:00:12,480 --> 00:00:16,717
advances? What is what what is happening?
Why is this important? Uh

6
00:00:16,800 --> 00:00:19,197
there's a lot of back and forth in the
timeline, but I'd love for you to just

7
00:00:19,280 --> 00:00:21,357
set the table for us on what actually

8
00:00:21,440 --> 00:00:24,717
happened with uh Navier Stokes today.

9
00:00:24,800 --> 00:00:28,237
Well, uh it's always a huge day uh in uh

10
00:00:28,320 --> 00:00:30,157
AI and modern scientific progress, I

11
00:00:30,240 --> 00:00:33,197
would say. Yeah. Today we announced that

12
00:00:33,280 --> 00:00:37,037
our model had solved the Navier Stokes

13
00:00:37,120 --> 00:00:40,957
uh problem. Uh that we found a a counter

14
00:00:41,040 --> 00:00:43,597
example or a sort of proof that you can

15
00:00:43,680 --> 00:00:48,877
actually that these these uh theoretical
equations do have a singularity or kind

16
00:00:48,960 --> 00:00:51,597
of break down under certain circumstances.
And this is a problem

17
00:00:51,680 --> 00:00:53,437
that has been open for a very long time.

18
00:00:53,520 --> 00:00:56,797
It's one of the seven millennium
problems which are kind of some of the

19
00:00:56,880 --> 00:01:01,117
deepest most important problems
in mathematics and I I think that the

20
00:01:01,200 --> 00:01:04,397
problem itself is important right this
is new knowledge for humanity that the

21
00:01:04,480 --> 00:01:07,917
proof itself is actually very elegant
and beautiful and I think that there's a

22
00:01:08,000 --> 00:01:10,957
lot to learn from it and the the

23
00:01:11,040 --> 00:01:15,677
equations have lots of application and
fluid dynamics in other areas but to me

24
00:01:15,760 --> 00:01:19,197
what's even more important is about what
this represents about where we are in

25
00:01:19,280 --> 00:01:22,557
terms of model capabilities and the fact
that we can actually generate new

26
00:01:22,640 --> 00:01:26,477
knowledge that we can learn from these
models to and have them help us solve

27
00:01:26,560 --> 00:01:29,597
problems that are otherwise outside
of reach or would take us a very long to

28
00:01:29,680 --> 00:01:31,997
solve. >> Yeah. I mean,
where should I actually go

29
00:01:32,080 --> 00:01:36,237
with this? Is this going to help me book
a flight? Is this going to help me cure

30
00:01:36,320 --> 00:01:39,917
cancer? Is this just going to help you
recruit researchers who are fascinated

31
00:01:40,000 --> 00:01:42,077
by this stuff? Uh because I think that

32
00:01:42,160 --> 00:01:47,277
there's this is, you know, taken over
the technology world. Uh but I imagine

33
00:01:47,360 --> 00:01:51,197
that this will not be something that gets
talked about at backyard barbecues

34
00:01:51,280 --> 00:01:54,877
with friends and family that are
three clicks removed from outside of

35
00:01:54,960 --> 00:01:57,357
>> Yeah. Yeah. >> Well, look,
I think that there's first

36
00:01:57,440 --> 00:02:02,077
of all the applications of this specific
result or the equations themselves,

37
00:02:02,160 --> 00:02:05,437
right, which are things that let us
better understand phenomena from ocean

38
00:02:05,520 --> 00:02:07,197
currents to air flow around aircraft.

39
00:02:07,280 --> 00:02:10,797
Sure. Uh around, you know, turbulence
and things like that. But it's really

40
00:02:10,880 --> 00:02:14,637
about the broader insights and methods
that can help scientists and

41
00:02:14,720 --> 00:02:17,997
mathematicians further accelerate
their research. And I think that again

42
00:02:18,080 --> 00:02:20,317
representative of if we have models that

43
00:02:20,400 --> 00:02:23,837
can help solve this kind of problem then

44
00:02:23,920 --> 00:02:28,637
what happens from here like what
other problems that are immediately

45
00:02:28,720 --> 00:02:32,317
applicable. And I think that talking
about curing diseases and new medicines

46
00:02:32,400 --> 00:02:35,837
we're going to be able to develop, all
of that starts to become much more real

47
00:02:35,920 --> 00:02:40,477
when you have models that are at this
level of assistance and capability. And

48
00:02:40,560 --> 00:02:44,397
I do think that there are going to be
real changes to to think about in terms

49
00:02:44,480 --> 00:02:49,037
of we can have so much more ambition
with the kinds of of challenges that we

50
00:02:49,120 --> 00:02:50,637
can hope to tackle now.

51
00:02:50,720 --> 00:02:53,357
>> Yeah. I I mean I think uh even even if

52
00:02:53,440 --> 00:02:56,077
this doesn't break through to the

53
00:02:56,160 --> 00:03:00,397
broader uh the broader world uh the

54
00:03:00,480 --> 00:03:04,077
weekend definitely will because it
seemed like everyone was talking about

55
00:03:04,160 --> 00:03:06,477
Blender, talking about Astra building

56
00:03:06,560 --> 00:03:11,117
stuff. Uh take us through the launch
of Astra, what the feedback has been, what

57
00:03:11,200 --> 00:03:14,877
you've learned. It seemed like there
were a couple resets, the model scaled

58
00:03:14,960 --> 00:03:19,757
very well. Uh how was this launch
different than previous launches?

59
00:03:19,840 --> 00:03:23,677
Well, first of all, I've just been blown
away by the community reaction to Astra.

60
00:03:23,760 --> 00:03:27,917
It's been really amazing and very
humbling, honestly, to see all the

61
00:03:28,000 --> 00:03:31,677
creativity and the different ways that
people have been applying the model. And

62
00:03:31,760 --> 00:03:33,597
I think that it's very clear that we've

63
00:03:33,680 --> 00:03:35,917
reached a new threshold of computer use.

64
00:03:36,000 --> 00:03:39,357
So, this model is able to really work

65
00:03:39,440 --> 00:03:43,197
with different kinds of applications
in a way that was not previously possible.

66
00:03:43,280 --> 00:03:47,197
And people are taking full advantage
of that fact and really thinking about how

67
00:03:47,280 --> 00:03:51,117
to create. I lots of people
showing off 3D creations and

68
00:03:51,200 --> 00:03:53,677
>> mapping out physical locations and

69
00:03:53,760 --> 00:03:58,317
turning them into into these 3D models
and thinking about can you use this for

70
00:03:58,400 --> 00:04:02,077
>> design of physical parts. Someone
talked about how they were designing some

71
00:04:02,160 --> 00:04:07,357
mechanic for catching hair in a shower
drain and that they were able to

72
00:04:07,440 --> 00:04:10,557
>> super cool, right? That that they're
now able to actually manufacture that

73
00:04:10,640 --> 00:04:12,097
>> shower hair super intelligence.

74
00:04:12,180 --> 00:04:15,997
[laughter] Yeah. No, that stuff's so
mundane, but it's so important. I I feel

75
00:04:16,080 --> 00:04:18,157
like uh a lot of this stuff gets lost.

76
00:04:18,240 --> 00:04:20,077
Right. >> Ex. Exactly.
And I think there's a core

77
00:04:20,160 --> 00:04:23,597
there that's very important, which
is that we are talking about these grand

78
00:04:23,680 --> 00:04:28,077
challenges sometimes or very esoteric
applications, but really the everyday

79
00:04:28,160 --> 00:04:31,997
the number of problems that you have
in your life that you would love to solve.

80
00:04:32,080 --> 00:04:34,797
It's now possible, right? That that
we're really trying to empower the

81
00:04:34,880 --> 00:04:37,517
individual to make it so that you can be

82
00:04:37,600 --> 00:04:41,837
you can have superpowers, you can
accomplish more. And I think that that

83
00:04:41,920 --> 00:04:45,197
you know, really trying to benefit people,
empower people, build tools that

84
00:04:45,280 --> 00:04:49,677
can really help you and help you in your
your daily life. That's all part of what

85
00:04:49,760 --> 00:04:51,197
we're we're we're working on.

86
00:04:51,280 --> 00:04:54,157
>> So, yeah, I mean, it seems like a huge

87
00:04:54,240 --> 00:04:57,117
number of people in uh in tech uh

88
00:04:57,200 --> 00:04:59,757
effectively rebuilt their entire house

89
00:04:59,840 --> 00:05:02,317
in Blender and planned a remodel this

90
00:05:02,400 --> 00:05:04,797
weekend. Uh thanks to Astra and but I am

91
00:05:04,880 --> 00:05:06,877
wondering about the merge and how you

92
00:05:06,960 --> 00:05:09,997
bring together codecs, chatt work,

93
00:05:10,080 --> 00:05:13,117
chacht on desktop. I have a gaming PC

94
00:05:13,200 --> 00:05:17,277
now with a Nvidia card in it. I have
a Mac Mini. Like I have all these things

95
00:05:17,360 --> 00:05:21,677
and I can imagine that I'm just I'm
doing that unnecessary or like it's fun

96
00:05:21,760 --> 00:05:25,917
for me, but uh that early adopter work
of going and unhobling it a little bit

97
00:05:26,000 --> 00:05:27,917
here and there, but in the future this

98
00:05:28,000 --> 00:05:30,717
will all just be tucked in one, you

99
00:05:30,800 --> 00:05:33,517
know, prompt box and it might build a 3D

100
00:05:33,600 --> 00:05:38,477
Blender model for to answer my question
of should I remodel my house or not. Um,

101
00:05:38,560 --> 00:05:42,877
but how do you see the capabilities
that we saw on display from sort of light

102
00:05:42,960 --> 00:05:44,797
power users over the weekend actually

103
00:05:44,880 --> 00:05:47,597
making their way into uh consumers who

104
00:05:47,680 --> 00:05:50,157
might not even know what Blender is.

105
00:05:50,240 --> 00:05:52,237
>> Well, I think you're exactly right that

106
00:05:52,320 --> 00:05:55,837
we really want to shift these tools from

107
00:05:55,920 --> 00:05:58,477
requiring kind of low-level

108
00:05:58,560 --> 00:06:03,517
sort of >> access or guidance
to really having the

109
00:06:03,600 --> 00:06:06,717
human be able to fly, right? to really
be empowered for you to be able to set

110
00:06:06,800 --> 00:06:11,837
the goals and the objectives and that
you still should feel like you can get

111
00:06:11,920 --> 00:06:14,477
into those details and you can
understand them. you can provide that

112
00:06:14,560 --> 00:06:17,997
oversight because ultimately you should
feel accountable for the outcomes but

113
00:06:18,080 --> 00:06:22,077
you have this this absolute amplifier
right like a trampoline or like a you

114
00:06:22,160 --> 00:06:25,997
know rocket ship for the mind like
however you want to to to analogize it

115
00:06:26,080 --> 00:06:29,117
and I think what that means at a
practical level so first of all this

116
00:06:29,200 --> 00:06:33,917
year we've been really seeing this shift
from just pure chat use cases to gentic

117
00:06:34,000 --> 00:06:37,437
use cases but I would also keep in mind
that chat is alive and well I mean we're

118
00:06:37,520 --> 00:06:41,277
we're now well over a billion users
every week that you can see that the

119
00:06:41,360 --> 00:06:44,877
market share of chatbt is starting
to climb once again because we've been

120
00:06:44,960 --> 00:06:49,117
investing so hard in in so many use
cases that are important for people in

121
00:06:49,200 --> 00:06:53,037
education and health and a variety
of other areas and then at the same time

122
00:06:53,120 --> 00:06:56,797
these productivity deep knowledge work
use cases those are really taking off

123
00:06:56,880 --> 00:07:00,317
we've had this like almost vertical wall
of agentic adoption since we launched

124
00:07:00,400 --> 00:07:04,477
chat GBT work and I think that the fact
that these are two distinct modes that

125
00:07:04,560 --> 00:07:08,157
that is actually a point in time that is
something that we're continuing to unify

126
00:07:08,240 --> 00:07:11,917
and merge and that we're starting to see
that there's a new emerging form factor

127
00:07:12,000 --> 00:07:14,397
for how people want to consume AI. And I

128
00:07:14,480 --> 00:07:16,237
think that it's almost like that the

129
00:07:16,320 --> 00:07:21,437
promise of AI has always been that you
have something that you can talk to and

130
00:07:21,520 --> 00:07:23,117
really delegate work to that's proactive

131
00:07:23,200 --> 00:07:26,877
and persistent. And that I what we were

132
00:07:26,960 --> 00:07:31,357
promised if you were to rewind five 10
years ago was never a low-level language

133
00:07:31,440 --> 00:07:34,397
model. You have to think about context
windows and you have to select thinking

134
00:07:34,480 --> 00:07:37,597
strength and you have to select
different models. Like none of that none

135
00:07:37,680 --> 00:07:41,277
of that is is the future. And so I
think that we're moving towards real

136
00:07:41,360 --> 00:07:46,397
amplification, real giving you time
back, real having computers that are

137
00:07:46,480 --> 00:07:48,237
able to operate according to your goals,

138
00:07:48,320 --> 00:07:51,437
to your desires. And I think that that

139
00:07:51,520 --> 00:07:55,277
is a core of it. We're developing
this safely. That's that's one of the core

140
00:07:55,360 --> 00:07:59,037
commitments that that we make and how
we think about this. Um but but we really

141
00:07:59,120 --> 00:08:01,917
see the power of these tools starting to

142
00:08:02,000 --> 00:08:04,957
really um really start to increase in

143
00:08:05,040 --> 00:08:09,277
terms of what people are capable of and
that under the hood utilizing tools like

144
00:08:09,360 --> 00:08:12,317
blender so that that is almost a detail
that fades into the background is

145
00:08:12,400 --> 00:08:16,557
absolutely the direction of travel.
It feels like in AI particularly there's

146
00:08:16,640 --> 00:08:18,877
been a almost like a first mover

147
00:08:18,960 --> 00:08:22,237
disadvantage in that billions of people

148
00:08:22,320 --> 00:08:25,997
have tried Chad GBT and and some
percentage of them tried it for the

149
00:08:26,080 --> 00:08:29,917
first time and have a certain impression
of the product and what it can do and

150
00:08:30,000 --> 00:08:33,677
then you know even in the last few weeks
there's been new agents and and products

151
00:08:33,760 --> 00:08:37,277
that come online and people try it and
their mind is just completely blown and

152
00:08:37,360 --> 00:08:41,037
I think it's funny because I'm like well
as somebody who's like you know trying

153
00:08:41,120 --> 00:08:44,317
to get the absolute max out of ChatG GBT.
I'm like, well, I've been running

154
00:08:44,400 --> 00:08:48,797
like, you know, I've had like an agent
running that, for example, will tell me

155
00:08:48,880 --> 00:08:51,757
every time a SpaceX launch is going
to happen and if it gets delayed, right?

156
00:08:51,840 --> 00:08:55,837
These sort of like persistent agents
that are running in the background. Um

157
00:08:55,920 --> 00:08:57,837
and but but strategically I feel like

158
00:08:57,920 --> 00:09:00,397
it's a it's a new kind of challenge

159
00:09:00,480 --> 00:09:05,197
because you have this the as
capability has been been scaling

160
00:09:05,280 --> 00:09:07,117
first movers need to be almost like

161
00:09:07,200 --> 00:09:10,477
constantly reminding the the market of

162
00:09:10,560 --> 00:09:13,117
um all these just like new ways to use

163
00:09:13,200 --> 00:09:15,917
>> overhang. Yeah,
we we think about this a

164
00:09:16,000 --> 00:09:21,037
lot and I think that there's this
discovery problem that we as a field

165
00:09:21,120 --> 00:09:24,557
need to really encounter in a first
class way and we haven't done it fully

166
00:09:24,640 --> 00:09:29,517
yet, but I think we have a real shot
at solving it better than any product

167
00:09:29,600 --> 00:09:33,757
before because >> the the
thing that that right now we

168
00:09:33,840 --> 00:09:37,597
kind of rely on. You think about chat
GBT chat GBT work. These are both text

169
00:09:37,680 --> 00:09:40,717
boxes, right? And it's like, well, this
new text box is way more powerful than

170
00:09:40,800 --> 00:09:43,757
the old text box. And but oh, oh, there
are some reasons that you still want to

171
00:09:43,840 --> 00:09:46,637
use the old text box. It's [laughter]
not, you know, it's like far too

172
00:09:46,720 --> 00:09:49,917
confusing. People just want something
that can help them solve their problem.

173
00:09:50,000 --> 00:09:52,397
The whole point is to get your time
back, not for you to have to go and

174
00:09:52,480 --> 00:09:55,837
become an expert in all these these
internal details. But at the same time,

175
00:09:55,920 --> 00:09:59,517
we also have a model that understands
what you're trying to accomplish, right?

176
00:09:59,600 --> 00:10:02,557
That you're explaining to it, here's
what I want. It has a lot of context on

177
00:10:02,640 --> 00:10:06,877
you. And so it should also be able
to proactively say to you, hey actually if

178
00:10:06,960 --> 00:10:11,037
you ask me this other way or if you add
added this connector or if you authorize

179
00:10:11,120 --> 00:10:15,917
me to do this or if you you know hook
up your credentials in this in this way I

180
00:10:16,000 --> 00:10:19,357
can go and do this other thing for you
and so we're thinking a lot about that

181
00:10:19,440 --> 00:10:23,517
self-nowledge that onboarding process
and I think that is a huge huge

182
00:10:23,600 --> 00:10:27,757
opportunity and I think that there
is both the the disadvantage that you cite

183
00:10:27,840 --> 00:10:30,797
of people tried it they form
an impression and it's changed that it's

184
00:10:30,880 --> 00:10:34,877
something new but there's also
an advantage I chat chatbt like over a

185
00:10:34,960 --> 00:10:39,757
billion users every week like that is
unique no one has that kind of use on on

186
00:10:39,840 --> 00:10:42,717
these on these models and I think
that the number of people who have tried

187
00:10:42,800 --> 00:10:46,717
chatbt before I think it's it's another
you know billion billion and a half

188
00:10:46,800 --> 00:10:50,877
something like that and so that's a huge
opportunity as well for us to go back to

189
00:10:50,960 --> 00:10:55,277
those users and say hey we can now solve
the problem for you we can now help you

190
00:10:55,360 --> 00:10:59,277
in ways that you didn't see before and I
think that it's it's true it's real you

191
00:10:59,360 --> 00:11:03,037
look at how many people use chat
for health 300 million people every single

192
00:11:03,120 --> 00:11:05,837
week with health queries, right?
And that that's really making a difference

193
00:11:05,920 --> 00:11:09,117
in people's lives and that of their loved
ones. And so we have such opportunity.

194
00:11:09,200 --> 00:11:12,877
>> Can you stay there with with
health and and and explain where

195
00:11:12,960 --> 00:11:16,397
>> one note before that? Something that I
think is really interesting and I think

196
00:11:16,480 --> 00:11:20,397
something that OpenAI can can do a lot
better is like when people talk about

197
00:11:20,480 --> 00:11:24,877
like everyone in AI wants to be like the
Apple of AI from a marketing standpoint.

198
00:11:24,960 --> 00:11:27,997
When they when you think like Apple
marketing, you're thinking like Mac

199
00:11:28,080 --> 00:11:30,557
versus PC or you're thinking 1984, these

200
00:11:30,640 --> 00:11:35,037
big branding campaigns. But the actual
thing that Apple does really really well

201
00:11:35,120 --> 00:11:39,277
with marketing is they just hammer
really really specific details about

202
00:11:39,360 --> 00:11:42,557
their products, right? They're like
>> they're advertising the new camera.

203
00:11:42,640 --> 00:11:44,557
They're advertising Mimoji, right?

204
00:11:44,640 --> 00:11:48,317
They're advertising like >> certain
features in Safari, right? And

205
00:11:48,400 --> 00:11:51,917
so it's like with AI, the surface
area of like things that you need to

206
00:11:52,000 --> 00:11:55,917
communicate is actually like an order
of magnitude greater because it can do so

207
00:11:56,000 --> 00:11:58,877
many different things. And so I
think there's such an opportunity

208
00:11:58,960 --> 00:12:02,477
>> um for the company
to focus advertising.

209
00:12:02,560 --> 00:12:05,757
The brand campaigns are awesome and like
the the launch video for Astro was

210
00:12:05,840 --> 00:12:09,277
amazing, but it's like there should be
billboards running of like very specific

211
00:12:09,360 --> 00:12:13,037
things that Chach BT can do
to give you back your time. Um

212
00:12:13,120 --> 00:12:17,357
>> yes, this this has actually been
a real sort of realization or just like

213
00:12:17,440 --> 00:12:21,037
something that I have really come
to over the course of of of this year. And

214
00:12:21,120 --> 00:12:25,997
if you look at even for example the you
know we just announced chat GBT images 2.5

215
00:12:26,080 --> 00:12:29,757
>> and if you look at the launch video
there the thing that I love about it is

216
00:12:29,840 --> 00:12:34,317
it shows here's someone creating
an image and here's like a bunch of

217
00:12:34,400 --> 00:12:38,157
different variations of it and then
here's them taking their favorite one

218
00:12:38,240 --> 00:12:42,077
and having it in the world. like someone
said, "Here's a cool like little sketch

219
00:12:42,160 --> 00:12:45,357
of a candle holder." You see an awesome
visualization of it and then you see the

220
00:12:45,440 --> 00:12:48,557
physical candle holder and you're just
like that's what you want, right? It's

221
00:12:48,640 --> 00:12:50,717
like you it speaks to you immediately.

222
00:12:50,800 --> 00:12:54,477
And I think that that really showing
people here's a use case. And the thing

223
00:12:54,560 --> 00:12:57,117
that was also a little surprising to me
is that we've sometimes highlighted

224
00:12:57,200 --> 00:13:01,037
esoteric use cases, something that
appeals to someone in particular and

225
00:13:01,120 --> 00:13:05,437
it's amazing for that person, but people
don't then say, "Oh, because it's this

226
00:13:05,520 --> 00:13:08,877
powerful, I can also do this other
powerful thing I've been waiting on."

227
00:13:08,960 --> 00:13:12,717
like that connection is is is is
something that is less sort of easy to

228
00:13:12,800 --> 00:13:15,997
make than I than I'd realize and it
makes sense. What you want is to for

229
00:13:16,080 --> 00:13:19,757
there to be use cases where people say
I actually want that particular thing now

230
00:13:19,840 --> 00:13:22,877
let me go try it myself and then from
there you start exploring and you start

231
00:13:22,960 --> 00:13:25,357
to find I [clears throat] actually
do have this powerful use case that I

232
00:13:25,440 --> 00:13:27,037
didn't even realize was tip at the time.

233
00:13:27,120 --> 00:13:30,637
>> Yeah. Studio is probably like the best
example of that working really really

234
00:13:30,720 --> 00:13:35,597
well. >> The other thing is is reminding
people to ask the AI what it's capable of.

235
00:13:35,680 --> 00:13:38,637
Yeah, like I was I was having lunch
with a buddy who's a real estate developer

236
00:13:38,720 --> 00:13:42,077
and he is using chat all day long for

237
00:13:42,160 --> 00:13:45,917
different like deal memos and to
understand like projects that he's

238
00:13:46,000 --> 00:13:48,957
working on all this stuff and he'll
ask me he'll ask me all the time can can

239
00:13:49,040 --> 00:13:53,437
catch GPT do this or that and I'm like
I'm happy to I'm happy to answer you but

240
00:13:53,520 --> 00:13:57,757
like you have the thing that will just
explain exactly how to do the thing that

241
00:13:57,840 --> 00:14:00,877
you want to do or or maybe
not but it probably can.

242
00:14:00,960 --> 00:14:03,837
>> Yeah, I I I did the same thing. I was
kicking off this Blender thing and I was

243
00:14:03,920 --> 00:14:08,397
like, should I run this as a local
codeex thread or in the cloud? Let me

244
00:14:08,480 --> 00:14:11,677
know which one's better based on my
system and it gave me a good answer and

245
00:14:11,760 --> 00:14:12,637
I was able to go for it.

246
00:14:12,720 --> 00:14:16,557
>> On on images,
uh when when images 2 came

247
00:14:16,640 --> 00:14:20,797
out, I had some moments where I thought,
okay, images is solved. Like where do

248
00:14:20,880 --> 00:14:25,117
you where do you think images actually
go as a category? because it felt uh

249
00:14:25,200 --> 00:14:28,637
like this has been something that has
maybe oneshoted me more than anything

250
00:14:28,720 --> 00:14:30,557
else. specifically with like when it

251
00:14:30,640 --> 00:14:33,197
when it released, I was I was um

252
00:14:33,280 --> 00:14:36,157
>> uh spending hours like like on a

253
00:14:36,240 --> 00:14:40,557
Saturday trying to design furniture,
right? And just going through like you

254
00:14:40,640 --> 00:14:42,237
know hundreds and hundreds of prompts.

255
00:14:42,320 --> 00:14:45,997
But how how far can image models go and

256
00:14:46,080 --> 00:14:50,157
where are they going and what are the ways
in which you think they can um they

257
00:14:50,240 --> 00:14:53,837
can be applied? We were talking earlier
too about the downstream impact of of

258
00:14:53,920 --> 00:14:58,877
image models. If you can take, you know,
a a physical space somewhere and take a

259
00:14:58,960 --> 00:15:02,157
picture of it and imagine it as all
these other variations, there's so much

260
00:15:02,240 --> 00:15:06,237
like real world activity that will be
driven from that because people can see

261
00:15:06,320 --> 00:15:08,957
this thing. >> Yeah. >> Visually
and say like I now I want to go

262
00:15:09,040 --> 00:15:11,677
make that reality which
I think is really cool.

263
00:15:11,760 --> 00:15:14,717
>> Well, I think that's exactly the right
way to think about it is as you hit new

264
00:15:14,800 --> 00:15:19,117
thresholds of capability, my experience
has always been that fundamentally new

265
00:15:19,200 --> 00:15:22,797
applications become unlocked in ways
that you almost wouldn't have thought

266
00:15:22,880 --> 00:15:27,037
about ahead of time. And so I think
that within for example knowledge work,

267
00:15:27,120 --> 00:15:31,037
professional work, marketing, all
those areas, you just need to be above a

268
00:15:31,120 --> 00:15:34,237
quality threshold. If you're below it,
it's a cool concept, but you can't

269
00:15:34,320 --> 00:15:35,837
actually use the final material, right?

270
00:15:35,920 --> 00:15:39,437
That that then means that you you
haven't really solved the problem. and

271
00:15:39,520 --> 00:15:42,237
that having precise edit control, being

272
00:15:42,320 --> 00:15:46,477
fast and really being creative and having
a diversity of different results

273
00:15:46,560 --> 00:15:49,997
and also being able to >> have
this this good interplay back and

274
00:15:50,080 --> 00:15:54,717
forth with a person. I think that that
really unlocks whole new use cases and I

275
00:15:54,800 --> 00:15:58,557
think there's a huge market there and even
for example the kinds of things you

276
00:15:58,640 --> 00:16:01,997
may not think of naively but actually
start to be really important

277
00:16:02,080 --> 00:16:06,317
applications we're seeing happening
is slide creation or making awesome

278
00:16:06,400 --> 00:16:10,237
websites, right? Being able to have
that image generation capability in the

279
00:16:10,320 --> 00:16:14,237
middle is something that's very unique
to OpenAI uh relative to to some of our

280
00:16:14,320 --> 00:16:16,877
competitors. And I think that you're
able to then produce much better

281
00:16:16,960 --> 00:16:18,957
artifacts downstream. And so we really

282
00:16:19,040 --> 00:16:22,557
view images, we view voice, we view

283
00:16:22,640 --> 00:16:25,997
coding, all of these capabilities as one

284
00:16:26,080 --> 00:16:30,237
package that are going to come together
to create an AI that empowers you. That

285
00:16:30,320 --> 00:16:34,077
means you can create anything that you
imagine. And I think it's going to be

286
00:16:34,160 --> 00:16:35,357
something that's just unlike anything

287
00:16:35,440 --> 00:16:38,157
out there. Let's go back to health. I

288
00:16:38,240 --> 00:16:41,437
think most people already are aware that

289
00:16:41,520 --> 00:16:43,837
you can synthesize some lab data with

290
00:16:43,920 --> 00:16:46,637
some sleep scores, but your vision that

291
00:16:46,720 --> 00:16:48,957
you laid out recently for where that

292
00:16:49,040 --> 00:16:51,597
product goes is much more complex, much

293
00:16:51,680 --> 00:16:55,597
deeper. So, uh tell me where uh ChatBT

294
00:16:55,680 --> 00:16:58,157
Health is going in the future.

295
00:16:58,240 --> 00:17:02,317
>> Well, I would think of it as there
are three sides to what we do on health.

296
00:17:02,400 --> 00:17:06,477
There's the consumer side again 300
million people every week with health

297
00:17:06,560 --> 00:17:10,717
queries. >> There's the clinician
side which is bottoms up

298
00:17:10,800 --> 00:17:14,637
>> and that's really about think
of it a chatbt that's really tuned for

299
00:17:14,720 --> 00:17:18,477
clinicians that gives them direct
citations to medical literature things

300
00:17:18,560 --> 00:17:21,837
like that. >> There's
a third pillar which is the

301
00:17:21,920 --> 00:17:25,677
enterprise side of selling directly
to hospitals and them enabling it. And you

302
00:17:25,760 --> 00:17:28,717
can see things like we have
an integration within Epic and really

303
00:17:28,800 --> 00:17:31,917
trying to bring each of these three

304
00:17:32,000 --> 00:17:36,557
pillars the best possible service
independently. But you think about as

305
00:17:36,640 --> 00:17:39,837
those really build momentum that you
actually are able to get synergies

306
00:17:39,920 --> 00:17:42,877
across them, right? That there's
something that actually makes the health

307
00:17:42,960 --> 00:17:46,477
experience and the ability to really
transform health care in America and the

308
00:17:46,560 --> 00:17:50,717
world on the table because there's so
many think about how much work you as a

309
00:17:50,800 --> 00:17:53,597
patient have to do. If you're talking to
different specialists, you have to carry

310
00:17:53,680 --> 00:17:57,037
your medical record from one to the other.
You have to explain again here's

311
00:17:57,120 --> 00:17:59,917
the issue and that ultimately you're
on the hook. You're the doctor who has to

312
00:18:00,000 --> 00:18:02,557
make the decision whether you like it
or not. And actually being able to have

313
00:18:02,640 --> 00:18:06,797
just good sharing of of that information
across different providers that becomes

314
00:18:06,880 --> 00:18:09,757
possible if everyone's on one
platform or think about clinical trial

315
00:18:09,840 --> 00:18:13,437
enrollment. That's a huge bottleneck
to drug development and finding people who

316
00:18:13,520 --> 00:18:17,197
are eligible and will benefit from being
involved in a particular trial. And if

317
00:18:17,280 --> 00:18:21,197
you have that kind of of data, if
people are willing to sort of, you know,

318
00:18:21,280 --> 00:18:24,397
entrust you with that information,
that that's something can actually really

319
00:18:24,480 --> 00:18:28,317
benefit them and benefit the world
at the same time. And so what I view us as

320
00:18:28,400 --> 00:18:32,077
building is really trying to build
the world's best healthcare platform to

321
00:18:32,160 --> 00:18:36,797
really be able to bring health care
into the AI age. And I think that it's

322
00:18:36,880 --> 00:18:40,957
something that is going to be absolutely
transformative to many people's quality

323
00:18:41,040 --> 00:18:44,397
of life, really uplift so many people,
and we're seeing it already in such

324
00:18:44,480 --> 00:18:47,997
concrete ways. Some of my favorite
stories about chatbt are people who say,

325
00:18:48,080 --> 00:18:52,477
"Hey, information I got from chat helped
me save my own life. Help me save that

326
00:18:52,560 --> 00:18:56,717
of a loved one that there was this
medical issue that someone had and that

327
00:18:56,800 --> 00:19:00,157
I if you know the doctor told me one
thing, I was able to to double check

328
00:19:00,240 --> 00:19:03,277
that and understand what they were
saying and be able to to push back on it

329
00:19:03,360 --> 00:19:06,637
and got to a good outcome." And that
happens every single day. So I think

330
00:19:06,720 --> 00:19:10,477
that health with these AIs is something
that it we're still just scratching the

331
00:19:10,560 --> 00:19:13,117
surface of what's possible and I
think it's one of the most positive

332
00:19:13,200 --> 00:19:15,517
applications of AI that that you can you

333
00:19:15,600 --> 00:19:17,517
can think of. Yeah, I'm very interested

334
00:19:17,600 --> 00:19:20,797
to see how uh the advancements in memory

335
00:19:20,880 --> 00:19:22,797
intersect with health because I expect

336
00:19:22,880 --> 00:19:27,837
that uh chatbt with where
memory has gone where I'll be

337
00:19:27,920 --> 00:19:32,317
in a new thread and it will bring up
just the right information or tie back

338
00:19:32,400 --> 00:19:36,877
to a thread that maybe happened three
weeks ago. When you actually apply that

339
00:19:36,960 --> 00:19:41,997
to like health related queries, it may
be able to like pick up patterns that

340
00:19:42,080 --> 00:19:45,437
sometimes would take a human years
to figure out like a certain ailment or

341
00:19:45,520 --> 00:19:47,837
something like that where it's like,
hey, you're asking about all these

342
00:19:47,920 --> 00:19:51,677
different things and maybe you thought
they were not connected. Turns out they

343
00:19:51,760 --> 00:19:55,917
they actually are and you should go
down this sort of like rabbit hole.

344
00:19:56,000 --> 00:19:57,597
>> Last question. I'm sorry. Please.

345
00:19:57,680 --> 00:20:00,637
>> I was I was going to say I think that's
absolutely right. We're seeing that very

346
00:20:00,720 --> 00:20:05,357
concretely and we've seen you know just
in my own personal life my my wife you

347
00:20:05,440 --> 00:20:09,277
know talked about some of her her
medical conditions publicly but I it was

348
00:20:09,360 --> 00:20:12,237
really this five-year journey of talking
to many specialists each one who was

349
00:20:12,320 --> 00:20:14,717
kind of touching one part of the
elephant and would try to address that

350
00:20:14,800 --> 00:20:18,477
one part and it was only finally her
allergist who said hey I think all these

351
00:20:18,560 --> 00:20:22,557
symptoms you're seeing are connected and
you have this uh genetic condition that

352
00:20:22,640 --> 00:20:24,557
affects all of your subsystems and

353
00:20:24,640 --> 00:20:26,637
that's the kind of thing where it's

354
00:20:26,720 --> 00:20:30,397
really hard to say how many people have

355
00:20:30,480 --> 00:20:33,677
similar kinds of conditions and just
never find out. How many people have

356
00:20:33,760 --> 00:20:38,077
these areas where it's like if you just
sort of are functionally specialized

357
00:20:38,160 --> 00:20:41,197
that you're never going to bring
together the whole diagnosis and I think

358
00:20:41,280 --> 00:20:45,597
that is one of the powers and potentials
of an AI that really deeply is able to

359
00:20:45,680 --> 00:20:50,237
help you across all parts of your life
and also is a deep domain expert in all

360
00:20:50,320 --> 00:20:53,917
areas of medicine. >> Yeah.
Uh I have one last question. Uh

361
00:20:54,000 --> 00:20:55,677
what how do you tell the story of

362
00:20:55,760 --> 00:20:59,197
operator? It feels like it was a failure

363
00:20:59,280 --> 00:21:02,237
or a side quest, but it feels incredibly

364
00:21:02,320 --> 00:21:04,477
important now given the advances in

365
00:21:04,560 --> 00:21:06,877
computer use. Is there a clear lineage

366
00:21:06,960 --> 00:21:10,397
there? What was operator? Does that

367
00:21:10,480 --> 00:21:15,837
still exist somewhere within chat GPT?
H how does how did how did computer use

368
00:21:15,920 --> 00:21:18,557
get solved? >> Yeah. Well,
look, I would look at all

369
00:21:18,640 --> 00:21:20,717
these things as timing and all about

370
00:21:20,800 --> 00:21:26,317
iterative deployment. Yeah. Right. that
that there's a moment where you need to

371
00:21:26,400 --> 00:21:30,077
where the capabilities aren't quite
there but actually learning from real

372
00:21:30,160 --> 00:21:33,757
world deployment is very helpful right
and the I think operator was just kind

373
00:21:33,840 --> 00:21:37,997
of below threshold in terms of the model
capabilities it was a cloud-based system

374
00:21:38,080 --> 00:21:41,917
that operated with computer use but it
was slow it wasn't fully accurate it was

375
00:21:42,000 --> 00:21:45,357
like pretty painful to to use some
people got value but it really wasn't

376
00:21:45,440 --> 00:21:49,037
above threshold and if you look
at what's happened that the team like one

377
00:21:49,120 --> 00:21:52,317
thing openai does very well is we make
long-term investments on things that

378
00:21:52,400 --> 00:21:56,717
really matter and we do the grind
and that the team this year I think really

379
00:21:56,800 --> 00:22:00,877
started to build momentum that we put
in a lot of effort to go and sort of burn

380
00:22:00,960 --> 00:22:05,517
down a long list of issues were able to
really focus on solving computer use and

381
00:22:05,600 --> 00:22:08,957
I think that they delivered in a
significant way and there's more to do

382
00:22:09,040 --> 00:22:12,557
never done all those things but it's
a true milestone I think people are really

383
00:22:12,640 --> 00:22:16,717
appreciating what's possible because
we've been in a world with these agents

384
00:22:16,800 --> 00:22:21,997
using computers through connectors right
through these very painstakingly coded

385
00:22:22,080 --> 00:22:26,797
systems that are so different from how
humans use computers. Whereas humans can

386
00:22:26,880 --> 00:22:29,997
already use everything on a computer,
right? Everything's designed for people.

387
00:22:30,080 --> 00:22:33,117
So if you have an AI that can operate
that way and even from the very

388
00:22:33,200 --> 00:22:36,957
beginning of OpenAI, we had a dream that
one day we could create such an AI, it

389
00:22:37,040 --> 00:22:41,037
becomes able to help you across
everything that you would be able to do

390
00:22:41,120 --> 00:22:44,797
with a computer yourself. And so I think
we're there with Astra. I think that

391
00:22:44,880 --> 00:22:48,397
there's just so much more that people
are going to uncover in terms of

392
00:22:48,480 --> 00:22:52,957
applications where this can go. But it's
an example of long-term focus, doing the

393
00:22:53,040 --> 00:22:55,757
work, and not giving up even
when the going gets tough.

394
00:22:55,840 --> 00:22:59,997
>> Yeah. And it feels like Astra,
my view is it really felt like all these

395
00:23:00,080 --> 00:23:02,717
different bets coming together,
you know, at the right at the right time,

396
00:23:02,800 --> 00:23:04,477
right? The the advancements in in the

397
00:23:04,560 --> 00:23:06,717
model itself, the the computer use,

398
00:23:06,800 --> 00:23:11,997
voice, all these things. Um, and it
uh it's all it's all making sense. So,

399
00:23:12,080 --> 00:23:15,197
>> yeah. >> Um, >> together,
it's focus. It's uh it's

400
00:23:15,280 --> 00:23:19,117
something that I think this company
does extremely well when we really put a

401
00:23:19,200 --> 00:23:23,117
challenge in front of us and think about
how to accomplish it safely well and uh

402
00:23:23,200 --> 00:23:25,437
to to really deliver
the value we're doing thing.

403
00:23:25,520 --> 00:23:27,197
>> Yeah. put different
differently, it felt

404
00:23:27,280 --> 00:23:28,957
like when you look at last year, it felt

405
00:23:29,040 --> 00:23:32,957
like OpenAI was uh operating like a big

406
00:23:33,040 --> 00:23:37,197
company and it and it was a big company
but but the way in which like product

407
00:23:37,280 --> 00:23:40,397
experimentation was happening
was you know when you think of like a

408
00:23:40,480 --> 00:23:42,557
hyperscaler they'll launch a new product

409
00:23:42,640 --> 00:23:47,997
thinking okay if there's a 20% hit rate
or even a 10% or 5% chance that's okay

410
00:23:48,080 --> 00:23:52,157
and then this year it feels like
actually the entire company switched

411
00:23:52,240 --> 00:23:56,957
back into actual startup mode which
is like you know focus focus focus all

412
00:23:57,040 --> 00:24:00,317
these things need to come together
the whole team needs to be rowing in the

413
00:24:00,400 --> 00:24:04,477
same direction and then the difference
in momentum and growth and all these

414
00:24:04,560 --> 00:24:08,877
things uh coming from that and and uh
actually taking the company from like

415
00:24:08,960 --> 00:24:12,717
operating and shipping more like a big
company to shipping again and focusing

416
00:24:12,800 --> 00:24:17,197
like a startup has been uh feels like
an impossible task and it's been uh

417
00:24:17,280 --> 00:24:21,277
incredible to watch. So >> yeah,
>> thank you. No, it's been a real real

418
00:24:21,360 --> 00:24:25,117
effort from many many people at OpenAI
to really bring together and something

419
00:24:25,200 --> 00:24:28,557
that that I really value that I think
we really value as a company and I think

420
00:24:28,640 --> 00:24:32,077
that we're just so laser focused on our
mission and really thinking about every

421
00:24:32,160 --> 00:24:35,037
piece of what we do should add
up to helping us accomplish it.

422
00:24:35,120 --> 00:24:37,517
>> Well, thank you so much for taking
the time to come chat with us.

423
00:24:37,600 --> 00:24:39,037
>> Great to see you.
>> We'll talk to you soon, Greg.

424
00:24:39,120 --> 00:24:43,200
>> Thank you.
>> Talk the rest of your day. Goodbye.