1
00:00:00,000 --> 00:00:01,917
OpenAI president and co-founder Greg

2
00:00:02,000 --> 00:00:04,077
Brockman joins us to discuss OpenAI's

3
00:00:04,160 --> 00:00:08,277
newest model, Spud, aka GPT 5.5, [music]

4
00:00:08,360 --> 00:00:11,757
and where it leaves OpenAI competitively.
That's coming up right

5
00:00:11,840 --> 00:00:15,517
after this. Open the Big Technology
podcast, today we have an emergency

6
00:00:15,600 --> 00:00:17,837
episode with OpenAI president and

7
00:00:17,920 --> 00:00:20,117
co-founder Greg Brockman all about GPT

8
00:00:20,200 --> 00:00:25,077
5.5, the famous Spud model. Looking
at what it does and what it means for

9
00:00:25,160 --> 00:00:27,917
OpenAI. Greg, great to see you.
Welcome back to the show.

10
00:00:28,000 --> 00:00:31,477
>> Thank you for having me.
Hope it's not too much of an emergency.

11
00:00:31,560 --> 00:00:36,757
>> Well, I am definitely recording
in in a Vegas hotel room, so

12
00:00:36,840 --> 00:00:39,957
more emergency than our last
conversation, but we had some time to

13
00:00:40,040 --> 00:00:45,037
prepare. So, it's great to be on with you.
So, let's let's just start with

14
00:00:45,120 --> 00:00:49,837
with this Can you confirm GPT 5.5 is Spud?

15
00:00:49,920 --> 00:00:52,397
>> Yes. >> Okay.

16
00:00:52,480 --> 00:00:56,037
What is GPT 5.5?

17
00:00:56,120 --> 00:00:57,997
>> Well, it's it's an amazing model. I

18
00:00:58,080 --> 00:01:01,477
think in many ways it is a step towards

19
00:01:01,560 --> 00:01:05,197
a new way of getting work done
with a computer. It's a new class of

20
00:01:05,280 --> 00:01:09,797
intelligence. It's extremely useful
at things like programming, right? And all

21
00:01:09,880 --> 00:01:14,277
the different aspects of debugging
and solving very hard and gnarly problems,

22
00:01:14,360 --> 00:01:18,077
just being very proactive, and really
being able to solve problems end to end

23
00:01:18,160 --> 00:01:22,797
with little instruction. But, the thing
that's to me most remarkable is not

24
00:01:22,880 --> 00:01:26,877
necessarily the fact that it got better
at coding. Like that, I think is is what

25
00:01:26,960 --> 00:01:29,637
everyone kind of expects. But,
the fact that it's now really crossed the

26
00:01:29,720 --> 00:01:33,877
threshold of usefulness for general
kinds of applications. And so, it's much

27
00:01:33,960 --> 00:01:38,317
better at creating slides, spreadsheets,
much better at computer use, using your

28
00:01:38,400 --> 00:01:39,997
browser, being able to kind of click

29
00:01:40,080 --> 00:01:45,517
through applications that are
otherwise hard to to have an AI operate.

30
00:01:45,600 --> 00:01:48,837
And so, I think that we're really seeing
the emergence of this new way of using a

31
00:01:48,920 --> 00:01:52,757
computer, and it starts with this
kind of intelligence at the core.

32
00:01:52,840 --> 00:01:54,357
>> When we spoke last, you mentioned that

33
00:01:54,440 --> 00:01:56,677
this was effectively the culmination of

34
00:01:56,760 --> 00:01:59,917
a two-year research process. So, was

35
00:02:00,000 --> 00:02:04,437
this planned two years ago?
Is that how far back OpenAI plans?

36
00:02:04,520 --> 00:02:06,477
>> I would say that yes, we do have very

37
00:02:06,560 --> 00:02:08,997
long horizons for how we plan. Now, one

38
00:02:09,080 --> 00:02:11,917
note is that we stack together many

39
00:02:12,000 --> 00:02:17,317
research ideas and bets on a variety
of timescales. And so, the way to think

40
00:02:17,400 --> 00:02:20,637
about it is that we are making constant
progress across every single part of the

41
00:02:20,720 --> 00:02:24,117
stack. And so, what GPT-5.5 represents

42
00:02:24,200 --> 00:02:27,877
is not an end point. In many ways,
it's a beginning point. It's really a step

43
00:02:27,960 --> 00:02:32,477
towards the kinds of models that we see
coming over even just upcoming months.

44
00:02:32,560 --> 00:02:36,597
And I think that you should expect
that we are going to have even larger

45
00:02:36,680 --> 00:02:38,757
improvements in the capability across a

46
00:02:38,840 --> 00:02:43,797
wide variety of of of these uh uh
of these uh aspects of what the model

47
00:02:43,880 --> 00:02:47,317
can do. And that's something that I
think will be very exciting, and we're

48
00:02:47,400 --> 00:02:50,797
just always thinking about how can we
make what we're producing more useful

49
00:02:50,880 --> 00:02:55,077
for real-world use, for real
users, and real applications.

50
00:02:55,160 --> 00:02:58,757
>> Can you share specifically what those
uh aspects are that we should be looking

51
00:02:58,840 --> 00:03:01,357
out for over the next few months?

52
00:03:01,440 --> 00:03:04,237
Um if this is the beginning,
what is it the beginning of?

53
00:03:04,320 --> 00:03:08,517
>> Well, I think that the big vision
we have, and you can see it reflected in

54
00:03:08,600 --> 00:03:11,837
many things, not just the models,
but the the kind of you know, you think

55
00:03:11,920 --> 00:03:16,317
about the models as the brain,
you can think about the systems and the

56
00:03:16,400 --> 00:03:19,397
harnesses like Codex and the
applications like the super app as

57
00:03:19,480 --> 00:03:23,357
almost the body around it to make it
into a useful AI. And that's really

58
00:03:23,440 --> 00:03:26,957
what's happening is a shift from language
models being the thing that is

59
00:03:27,040 --> 00:03:32,077
produced by labs like ourselves to an AI
that's actually useful. It's actually an

60
00:03:32,160 --> 00:03:33,437
assistant that's out there trying to

61
00:03:33,520 --> 00:03:36,357
solve your goal, that's really operating

62
00:03:36,440 --> 00:03:40,117
according to your instruction. And you

63
00:03:40,200 --> 00:03:44,277
can see right now Codex is becoming

64
00:03:44,360 --> 00:03:45,997
this app that's not just for the coders.

65
00:03:46,080 --> 00:03:48,517
It's really for anyone using a computer.

66
00:03:48,600 --> 00:03:52,957
And that it's not perfect, right? That
there are still some tasks where that it

67
00:03:53,040 --> 00:03:56,037
should be able to do it and it doesn't
quite get it right. Sometimes the

68
00:03:56,120 --> 00:03:59,277
personality isn't quite what you wanted,
right? That it doesn't quite, you know,

69
00:03:59,360 --> 00:04:02,597
it's like extremely powerful and do
out there doing a lot of really amazing

70
00:04:02,680 --> 00:04:06,277
things, but the way it communicates back
to you that you have to still spend some

71
00:04:06,360 --> 00:04:08,757
time really trying to read through,
okay, exactly how did it solve this

72
00:04:08,840 --> 00:04:12,957
problem? And so, these aspects, we know
exactly how to make them much better.

73
00:04:13,040 --> 00:04:16,197
And we've already had a pretty
remarkable improvement from 5.4

74
00:04:16,280 --> 00:04:20,357
to 5.5. I think we're going to have
even more remarkable improvements across

75
00:04:20,440 --> 00:04:24,397
every single aspect of what makes these
models useful. And one thing to know

76
00:04:24,480 --> 00:04:28,917
internally is that we think a lot about
the end application. Like that is one

77
00:04:29,000 --> 00:04:31,117
thing that changed for us over the past

78
00:04:31,200 --> 00:04:34,157
12, you know, 18 months, something like

79
00:04:34,240 --> 00:04:38,277
that, is that we used to really just be
focused on let's be let's improve on the

80
00:04:38,360 --> 00:04:43,037
benchmarks, let's make these models
more cerebrally capable, but we now are

81
00:04:43,120 --> 00:04:45,757
really focused on let's bring them
to real-world application. Let's think

82
00:04:45,840 --> 00:04:50,357
about finance, sales, marketing,
every single function that someone uses a

83
00:04:50,440 --> 00:04:53,597
computer, how can we help with their
computer work? How can we actually make

84
00:04:53,680 --> 00:04:56,157
the model have not just the theoretical

85
00:04:56,240 --> 00:05:00,637
capability to help, but is actually
experienced those kinds of tasks, it's

86
00:05:00,720 --> 00:05:02,197
actually been able to see what good

87
00:05:02,280 --> 00:05:04,837
looks like. And I think that the place

88
00:05:04,920 --> 00:05:08,117
we're going is one where you as a person

89
00:05:08,200 --> 00:05:12,597
doing work that you are the overseer,
you are the the CEO of almost this

90
00:05:12,680 --> 00:05:16,237
autonomous corporation or you know,
of this this fleet of agents perhaps is

91
00:05:16,320 --> 00:05:20,237
more is is the way to say it.
And that they are operating according to

92
00:05:20,320 --> 00:05:22,837
your goals. Now, you are still
accountable, right? You're still in the

93
00:05:22,920 --> 00:05:26,677
driver's seat. You're still the person
who thinks about, well, is this what I

94
00:05:26,760 --> 00:05:30,877
actually wanted? Was this work up to up
to standard? But that the details of

95
00:05:30,960 --> 00:05:34,557
exactly what buttons were clicked
and exactly the kind of code that was

96
00:05:34,640 --> 00:05:37,837
written or exactly how the formula
in the spreadsheet works, that you can

97
00:05:37,920 --> 00:05:42,357
abstract yourself from those if they're
not important to the evaluation of

98
00:05:42,440 --> 00:05:45,397
whether or not something was what
you wanted. And so, I think it's like

99
00:05:45,480 --> 00:05:48,477
increasing leverage for every worker.

100
00:05:48,560 --> 00:05:51,717
>> Okay, uh let me take my best guess
as to what's happening and you tell me how

101
00:05:51,800 --> 00:05:55,237
close I am. I mean, I'm thinking about
this. This is like a like you mentioned

102
00:05:55,320 --> 00:05:57,877
a combination of two years of work.

103
00:05:57,960 --> 00:06:00,797
There's two different types of I mean
not to tell you you know this but for

104
00:06:00,880 --> 00:06:04,677
our audience two different types of AI
training. There's the pre-training or at

105
00:06:04,760 --> 00:06:07,917
least the ones that have been pertinent
for these models. The pre-training where

106
00:06:08,000 --> 00:06:11,117
you just make the model generally smart
by having it predict the next word and

107
00:06:11,200 --> 00:06:15,477
then reinforcement training where you
have it like go out and actually take

108
00:06:15,560 --> 00:06:18,477
you know try to accomplish different
tasks and you reward it when it does a

109
00:06:18,560 --> 00:06:21,517
good job with those tasks
and effectively it sort of

110
00:06:21,600 --> 00:06:25,077
teaches it or it learns how
to how to do those tasks. Um

111
00:06:25,160 --> 00:06:28,997
is is what you're saying basically that
like this is the first result that we're

112
00:06:29,080 --> 00:06:33,397
seeing where open AI has just loaded
a ton of reinforcement learning on task

113
00:06:33,480 --> 00:06:37,357
specific stuff uh into this model
and that's what's producing the results

114
00:06:37,440 --> 00:06:40,757
you're talking about. >> Well,
I would actually say it a little

115
00:06:40,840 --> 00:06:44,597
differently. I would say that there's
many steps in the pipeline, right? That

116
00:06:44,680 --> 00:06:47,477
there's pre-training, mid-training,
reinforcement learning. There's you

117
00:06:47,560 --> 00:06:50,637
know, the data collection. There's like
a lot of these different things that all

118
00:06:50,720 --> 00:06:54,157
come together to produce the end result
and the way in which it's connected to

119
00:06:54,240 --> 00:06:58,277
the world, that's also very key to making
it useful. And the thing that I'm

120
00:06:58,360 --> 00:07:00,197
really saying is we have been investing

121
00:07:00,280 --> 00:07:03,877
on every single one of these and have a

122
00:07:03,960 --> 00:07:06,037
repeatable we have like a team, right?

123
00:07:06,120 --> 00:07:09,277
That it's not just about individuals
working on these pieces but a team that

124
00:07:09,360 --> 00:07:12,357
really comes together and looks across
the whole stack to say, "How do we make

125
00:07:12,440 --> 00:07:15,757
this more useful for real world
applications?" And so it's not really

126
00:07:15,840 --> 00:07:20,917
any one thing that we do. It's really
about the the overall effort of trying

127
00:07:21,000 --> 00:07:23,917
to like if you think about if you're
building a car, right? That there's it's

128
00:07:24,000 --> 00:07:26,197
not just about do you have like a better
engine, right? You could build a great

129
00:07:26,280 --> 00:07:30,197
engine but if the rest of the car is not
up to the quality level of the engine,

130
00:07:30,280 --> 00:07:33,517
it's not going to matter. And so I think
that that is the real innovation. It's

131
00:07:33,600 --> 00:07:37,637
really the end-to-end co-design and all
coming together in a repeatable fashion

132
00:07:37,720 --> 00:07:41,397
to make these models better
and better for our users.

133
00:07:41,480 --> 00:07:45,157
>> Um you were on a media call earlier
today with myself and a number of

134
00:07:45,240 --> 00:07:49,117
members of the press and uh one
of the interesting things that you said or

135
00:07:49,200 --> 00:07:51,797
basically I think you said this right
off the bat is that the model more

136
00:07:51,880 --> 00:07:54,957
intuitively knows uh what
you want and you don't have to

137
00:07:55,040 --> 00:07:59,357
spell it out exactly um as you as you
would in the past. Uh here's a tweet

138
00:07:59,440 --> 00:08:01,557
from from Rune, "There are early signs

139
00:08:01,640 --> 00:08:06,397
of 5.5 being a competent AI research
partner. Several researchers let 5.5 run

140
00:08:06,480 --> 00:08:10,677
variations of experiments overnight
given only a high-level uh algorithmic

141
00:08:10,760 --> 00:08:14,797
idea waking up to find a completed
sweep dashboards and samples never having

142
00:08:14,880 --> 00:08:17,677
touched the code or terminal at all."

143
00:08:17,760 --> 00:08:22,117
Um just uh if you can answer briefly
on this, a two-parter. How do you do that?

144
00:08:22,200 --> 00:08:26,317
Uh and um does that mean
prompt engineering is dead?

145
00:08:26,400 --> 00:08:29,877
>> Um number one, I think it
really comes down to when we

146
00:08:29,960 --> 00:08:33,797
say there's a new class of capability,
new class of intelligence, that's really

147
00:08:33,880 --> 00:08:35,077
what we mean, right? The models are

148
00:08:35,160 --> 00:08:39,237
becoming much more intuitive to use

149
00:08:39,320 --> 00:08:43,357
because they have deeper understanding
of what it is you're asking of them,

150
00:08:43,440 --> 00:08:46,517
right? That they really look at the
context, try to understand and puzzle

151
00:08:46,600 --> 00:08:50,477
out, "What am I being asked to do?"
And it really makes you realize, you know,

152
00:08:50,560 --> 00:08:52,877
to the second part, is prompt
engineering dead, which I actually think

153
00:08:52,960 --> 00:08:57,997
that prompt engineering in some ways
may be even more vibrant than before. Um

154
00:08:58,080 --> 00:09:01,877
but you spend so much time right now
trying to explain to your computer what

155
00:09:01,960 --> 00:09:05,277
you even want. You try to like pack
in this context and be like, "Well, here's

156
00:09:05,360 --> 00:09:08,157
what's going on, here's the situation,
here's the thing I want from you." And

157
00:09:08,240 --> 00:09:10,917
you're like, "Why do I have to explain
this to my computer? Right? Like the

158
00:09:11,000 --> 00:09:13,997
whole thing is the computer should be
doing the work to help me. Like I don't

159
00:09:14,080 --> 00:09:16,957
want to have to be sort of, you know,
breaking down the task, trying to

160
00:09:17,040 --> 00:09:19,957
explain to it step by step how
to do things. I want to point it in a

161
00:09:20,040 --> 00:09:23,917
direction and I want it to be able
to take care of the details and to get me

162
00:09:24,000 --> 00:09:27,357
the result. Again, in a way that I can
observe and kind of provide feedback

163
00:09:27,440 --> 00:09:31,517
along the way, but I want it to be
the driver of the of the of those like

164
00:09:31,600 --> 00:09:35,997
low-level execution. And so I think that
in some ways where prompt engineering is

165
00:09:36,080 --> 00:09:39,837
going to go is it's going to be about
you can get so much more out of these

166
00:09:39,920 --> 00:09:44,637
models with so much less effort,
but with the same amount of effort

167
00:09:44,720 --> 00:09:47,797
you still have a multiplier. Think about
how much more you could even get. And I

168
00:09:47,880 --> 00:09:50,877
think that we're just at the leading
edge right now of seeing the ceiling of

169
00:09:50,960 --> 00:09:54,317
what is capable what even
today's models are capable of.

170
00:09:54,400 --> 00:09:58,237
>> Okay. Let me briefly speak with you
about the economics of building a model

171
00:09:58,320 --> 00:10:02,797
like this. There's been this pattern
where these big massive models now

172
00:10:02,880 --> 00:10:06,517
you're not saying how much money your
computer used used you've used to train

173
00:10:06,600 --> 00:10:09,397
this but I think we
can be safe in assuming

174
00:10:09,480 --> 00:10:12,477
it was a lot and there's been this
pattern where these massive models come

175
00:10:12,560 --> 00:10:16,237
out they get distilled by open source

176
00:10:16,320 --> 00:10:19,997
model makers and then open source is

177
00:10:20,080 --> 00:10:23,837
just a couple months behind
the leading foundational models

178
00:10:23,920 --> 00:10:27,717
um and you know I guess like when
the investment was smaller being a couple

179
00:10:27,800 --> 00:10:31,237
months ahead you know
mattered a lot but I'm curious

180
00:10:31,320 --> 00:10:35,677
now that the investment is so big
and the models are the capabilities are

181
00:10:35,760 --> 00:10:39,357
increasing you know fairly
dramatically you know as

182
00:10:39,440 --> 00:10:43,397
you go how is this defensible
in the long term

183
00:10:43,480 --> 00:10:46,597
if you're just going to have
that pattern repeat over and over?

184
00:10:46,680 --> 00:10:48,077
>> Well I look at it a little differently.

185
00:10:48,160 --> 00:10:52,557
Like I think that the real
investment that we are making is into

186
00:10:52,640 --> 00:10:57,157
that end-to-end code design right of
having a system a system of people right

187
00:10:57,240 --> 00:11:01,557
who are producing this technology right
a way of working together and some of

188
00:11:01,640 --> 00:11:03,157
this is about how you leverage these

189
00:11:03,240 --> 00:11:05,317
massive supercomputers produce these

190
00:11:05,400 --> 00:11:10,637
models. Now it is also the case that
it's not as simple as you can take the

191
00:11:10,720 --> 00:11:13,637
output of these models and distill them
and you have exactly the model the same

192
00:11:13,720 --> 00:11:16,917
capability. It's just
smaller and could run fast. If

193
00:11:17,000 --> 00:11:19,797
that were the case we would just do that
and then we would also have a model that

194
00:11:19,880 --> 00:11:23,557
would be you know much more easy
to serve it in many ways and of

195
00:11:23,640 --> 00:11:26,717
course there's a lot of art behind
distillation there's a lot of great

196
00:11:26,800 --> 00:11:30,477
things there but the point that I'm
getting at is that the real thing that

197
00:11:30,560 --> 00:11:34,637
we are investing in is the machine
that makes the machine. Now, the at the

198
00:11:34,720 --> 00:11:37,757
deployment side, we think a lot
about safeguards. We think a lot about

199
00:11:37,840 --> 00:11:39,677
mitigations. And we do that for many,

200
00:11:39,760 --> 00:11:42,157
many different aspects of how these

201
00:11:42,240 --> 00:11:47,557
models could be misused um in real
situations. And that's something

202
00:11:47,640 --> 00:11:50,917
that we have been investing in for many
years. And we think about that across

203
00:11:51,000 --> 00:11:52,677
areas like cyber or thinking about that

204
00:11:52,760 --> 00:11:56,157
in areas like bio that we have a I a

205
00:11:56,240 --> 00:11:59,757
long-standing effort that you can see
in our preparedness framework, which is

206
00:11:59,840 --> 00:12:03,837
public, about how we approach these
kinds of uses of the model and how we

207
00:12:03,920 --> 00:12:07,597
try to make maximize the benefits,
mitigate the risks. And so, I think it's

208
00:12:07,680 --> 00:12:10,077
a real motion that every piece of what

209
00:12:10,160 --> 00:12:14,637
we do needs to connect to the question
of how do we continue to make progress,

210
00:12:14,720 --> 00:12:17,557
but also how do we make these models
broadly available? Because that's

211
00:12:17,640 --> 00:12:20,637
something that we really believe in,
that we believe this technology empowers

212
00:12:20,720 --> 00:12:24,837
people, and that we want it
to benefit people and lift everyone up.

213
00:12:24,920 --> 00:12:27,397
>> Yeah. But, just to go back on that, um

214
00:12:27,480 --> 00:12:29,237
the pricing on this model is, I think,

215
00:12:29,320 --> 00:12:33,717
double uh last model, GPT-5.4.

216
00:12:33,800 --> 00:12:37,477
Um and so, from an economics or a
business standpoint, the question would

217
00:12:37,560 --> 00:12:41,197
be, you know, let's say you keep
on progressing, but because there's been

218
00:12:41,280 --> 00:12:43,997
all this infrastructure that
infrastructure that's been put towards

219
00:12:44,080 --> 00:12:46,837
training the models, if open source uh

220
00:12:46,920 --> 00:12:48,837
can deliver not as good performance, but

221
00:12:48,920 --> 00:12:52,557
almost as good, uh and do it cheaper, um

222
00:12:52,640 --> 00:12:54,437
how do you handle that threat?

223
00:12:54,520 --> 00:12:57,877
>> Well, again, I I look at it a little
differently. So, first of all, if you

224
00:12:57,960 --> 00:13:01,637
look at our history, which
really is not driven by anything

225
00:13:01,720 --> 00:13:03,277
in competition. It's just like our our

226
00:13:03,360 --> 00:13:07,077
own sort of progress and and desire, we

227
00:13:07,160 --> 00:13:11,677
have dropped prices on the same level
of intelligence year over year, sometimes

228
00:13:11,760 --> 00:13:13,997
by literally a factor of 100. Right?

229
00:13:14,080 --> 00:13:17,237
It's like at least an order of magnitude
year over year, sometimes literally 100.

230
00:13:17,320 --> 00:13:19,757
But, the thing that keeps happening,
it's real Jevons paradox, where it's

231
00:13:19,840 --> 00:13:22,357
like you lower the cost of something,

232
00:13:22,440 --> 00:13:25,077
way more activity happens, right? And I

233
00:13:25,160 --> 00:13:29,797
think that what we keep seeing
is that there are returns to intelligence,

234
00:13:29,880 --> 00:13:33,797
right? That for the kinds of tasks that
these models are now capable of doing,

235
00:13:33,880 --> 00:13:37,277
that a little bit more intelligence
goes a long way. And I think that is the

236
00:13:37,360 --> 00:13:41,917
story of 5-5, that in some ways you can
almost look at it as like, oh, there's

237
00:13:42,000 --> 00:13:44,597
just an incremental improvement
in intelligence, but I think there's going

238
00:13:44,680 --> 00:13:48,397
to be a massive improvement in terms
of what people use it for. And by the way,

239
00:13:48,480 --> 00:13:51,997
I actually think that incremental is
actually very much an understatement for

240
00:13:52,080 --> 00:13:56,877
this model relative to 5-5. You know,
it's a 0.1 improvement in some ways, um

241
00:13:56,960 --> 00:14:01,197
but I think that that that actually
really undersells the magic that we see

242
00:14:01,280 --> 00:14:04,477
within this model and that the that our
early testers have have really seen in

243
00:14:04,560 --> 00:14:08,877
their practical work. >> So,
if people see these numbers and they say,

244
00:14:08,960 --> 00:14:09,597
um

245
00:14:09,680 --> 00:14:13,917
uh there's IPO pressure on uh on OpenAI,

246
00:14:14,000 --> 00:14:16,277
and therefore the you know, we've been

247
00:14:16,360 --> 00:14:21,077
getting a great deal on intelligence
and the free ride is over.

248
00:14:21,160 --> 00:14:24,437
You would argue against that.
>> I I Yeah, look, the way I think about

249
00:14:24,520 --> 00:14:25,957
this is that we have a very simple

250
00:14:26,040 --> 00:14:28,637
business in some ways, right? We rent,

251
00:14:28,720 --> 00:14:31,717
build, buy compute, and we resell it

252
00:14:31,800 --> 00:14:35,877
with some positive margin.
And as long as it's

253
00:14:35,960 --> 00:14:39,517
you know, positive operating margin,
and as long as there's scalable demand

254
00:14:39,600 --> 00:14:42,757
for intelligence, which I think is true
as long as there's problems to solve,

255
00:14:42,840 --> 00:14:45,197
like no one's going to run out
of problems to solve, and we've seen this

256
00:14:45,280 --> 00:14:48,597
at every step that the demand outstrips
our supply, then we can scale that

257
00:14:48,680 --> 00:14:52,837
compute all day. And I think that that
in my mind, that's the main directive

258
00:14:52,920 --> 00:14:56,837
that I that I ask of the team. It's just
like, just think about we need to add

259
00:14:56,920 --> 00:14:58,557
value on top of the raw compute and make

260
00:14:58,640 --> 00:15:00,237
sure that we are at positive operating

261
00:15:00,320 --> 00:15:03,957
margin on it, and that that is something

262
00:15:04,040 --> 00:15:08,517
where it's actually not even about like
like the different competition in the

263
00:15:08,600 --> 00:15:11,797
marketplace.
It's just a question of can you

264
00:15:11,880 --> 00:15:16,917
like have compute that gets turned into
intelligence, and that's just how, you

265
00:15:17,000 --> 00:15:20,157
know, that it does that at at a you
know, slightly improved uh you know,

266
00:15:20,240 --> 00:15:23,877
value coming out relative to the cost
going in. And I think that that is

267
00:15:23,960 --> 00:15:26,597
something where again we're
always trying to make more

268
00:15:26,680 --> 00:15:30,277
efficient models, but then we just want
more of them, and then we want the more

269
00:15:30,360 --> 00:15:33,237
intelligent models, and regardless
of where they're coming from, it's kind of

270
00:15:33,320 --> 00:15:35,717
all the same compute that's going in.

271
00:15:35,800 --> 00:15:39,037
And so, I think that it's actually
a great like competition this marketplace

272
00:15:39,120 --> 00:15:41,597
has been great for innovation, um but I

273
00:15:41,680 --> 00:15:46,277
think that it's actually something where
it's driving more usage and more overall

274
00:15:46,360 --> 00:15:51,197
spend in the ecosystem, and you can see
that in the revenue numbers of us and uh

275
00:15:51,280 --> 00:15:53,037
you know, others in this industry.

276
00:15:53,120 --> 00:15:55,837
>> Okay, I want to take a quick break
and come back and talk to talk with you

277
00:15:55,920 --> 00:16:00,797
about cybersecurity, trust, and whatever
else we can get to in our time in this

278
00:16:00,880 --> 00:16:04,637
emergency show. We'll be back right
after this. And we're back here on Big

279
00:16:04,720 --> 00:16:07,157
Technology Podcast with OpenAI president

280
00:16:07,240 --> 00:16:09,277
and co-founder Greg Brockman.

281
00:16:09,360 --> 00:16:12,557
Um Greg, let me ask you about uh the

282
00:16:12,640 --> 00:16:15,517
cybersecurity implications here. Um two

283
00:16:15,600 --> 00:16:18,197
very different approaches between OpenAI

284
00:16:18,280 --> 00:16:20,597
and Anthropic. Anthropic's

285
00:16:20,680 --> 00:16:23,397
uh latest massive model, Mythos, uh is

286
00:16:23,480 --> 00:16:25,757
not released to the public. Um

287
00:16:25,840 --> 00:16:30,437
this one, you know, Spud or 5.5,
is released to the public.

288
00:16:30,520 --> 00:16:33,637
I mean, let me just ask you straight up,
is there a chance that uh releasing this

289
00:16:33,720 --> 00:16:37,957
powerful model into you know, into the
public without this like step-by-step uh

290
00:16:38,040 --> 00:16:40,997
practice could lead
to some major cyber attacks?

291
00:16:41,080 --> 00:16:42,917
>> Well, I I actually
have a different view

292
00:16:43,000 --> 00:16:46,677
on the premise of the question. So, the
thing to understand is that we have been

293
00:16:46,760 --> 00:16:48,477
investing in cyber safeguards and

294
00:16:48,560 --> 00:16:51,677
cybersecurity as a part of our

295
00:16:51,760 --> 00:16:53,877
preparedness framework for years, right?

296
00:16:53,960 --> 00:16:57,597
The This is something we have invested
in far ahead of having the kinds of

297
00:16:57,680 --> 00:17:01,517
capabilities we see coming. And so,
we have been taking a very deliberate

298
00:17:01,600 --> 00:17:05,677
step-by-step approach. You can see even
just over the past couple weeks where

299
00:17:05,760 --> 00:17:09,637
we've expanded our trusted access
for cyber program, and in general, we

300
00:17:09,720 --> 00:17:11,517
believe in ecosystem resilience, right?

301
00:17:11,600 --> 00:17:14,197
That we think that you do want to go
step-by-step, that these models are

302
00:17:14,280 --> 00:17:18,037
going to continuously better. We have
line of sight to even more capable ones,

303
00:17:18,120 --> 00:17:22,157
and that you want to be able
to I put these models in the

304
00:17:22,240 --> 00:17:24,957
hands of defenders to make sure
that you're able to protect critical

305
00:17:25,040 --> 00:17:29,237
infrastructure. And we believe
in in that resilience of as you can bring

306
00:17:29,320 --> 00:17:32,637
these models into people's hands
that that then they're able to explore in

307
00:17:32,720 --> 00:17:36,957
ways that you would not be able to without
that kind of access. And so, you

308
00:17:37,040 --> 00:17:40,677
kind of want this graduated approach
and to make sure that you are moving down

309
00:17:40,760 --> 00:17:44,637
that pipeline as you can bring
in additional safeguards uh in order to

310
00:17:44,720 --> 00:17:47,997
make sure that you can maximize the
benefits and mitigate the risks. And so,

311
00:17:48,080 --> 00:17:50,637
we've really taken a deliberate approach.
I think our team has been

312
00:17:50,720 --> 00:17:54,917
working incredibly hard to think through
the cyber implications of this model.

313
00:17:55,000 --> 00:17:58,797
Um we also believe in iterative
deployment. Uh that's part of this

314
00:17:58,880 --> 00:18:01,757
really bringing the models as they
continuously get better. And we believe

315
00:18:01,840 --> 00:18:04,997
in democratic access.
We believe that ultimately the goal of

316
00:18:05,080 --> 00:18:08,597
creating this technology is to empower
people to ensure that it does benefit

317
00:18:08,680 --> 00:18:12,437
all of humanity. And so, we are
constantly trying to solve for how do we

318
00:18:12,520 --> 00:18:16,677
safely and responsibly bring this
technology to bear in the world in a

319
00:18:16,760 --> 00:18:21,157
broad way. >> Right.
And um I think suffice it to say

320
00:18:21,240 --> 00:18:24,757
that um your team hasn't been fans
of the way that Anthropic's deployed

321
00:18:24,840 --> 00:18:28,237
mythos. This is a quote from Sam. Uh
it's clearly incredible marketing to say

322
00:18:28,320 --> 00:18:31,357
we have built a bomb, we're about to drop
it on your head, we will sell you a

323
00:18:31,440 --> 00:18:34,677
bomb shelter for a hundred million
to run all your to run across all your

324
00:18:34,760 --> 00:18:38,517
stuff, but only if we
pick you as a customer. Um

325
00:18:38,600 --> 00:18:42,397
let me talk through the other
case and then get your response.

326
00:18:42,480 --> 00:18:46,677
Um the other case would be there are you
can't account for everything, and there

327
00:18:46,760 --> 00:18:50,197
are clearly going to be some
vulnerabilities that can't that will

328
00:18:50,280 --> 00:18:53,997
only be found by people
or entities deploying this and

329
00:18:54,080 --> 00:18:57,797
looking for them. So, maybe it makes
sense to start with a trusted group of

330
00:18:57,880 --> 00:19:00,557
testers before you deploy it before you

331
00:19:00,640 --> 00:19:03,037
deploy it broadly. What do you think?

332
00:19:03,120 --> 00:19:07,517
>> Well, I believe the correct answer here
is subtle, and I think it is rooted in

333
00:19:07,600 --> 00:19:11,877
the technical specifics of what you have
in front of you and many, many factors,

334
00:19:11,960 --> 00:19:15,437
right? You need to think about how are
the models progressing, right? Not just

335
00:19:15,520 --> 00:19:19,397
your own capabilities, but others
in the ecosystem. You need to think about

336
00:19:19,480 --> 00:19:22,877
what kind of benefit do you get from
having a small group that has access and

337
00:19:22,960 --> 00:19:26,917
are able to have, you know, are they
able to have high leverage by by being

338
00:19:27,000 --> 00:19:30,237
able to find and produce patches, but
then how do you actually coordinate the

339
00:19:30,320 --> 00:19:32,597
disclosure of those across an industry?

340
00:19:32,680 --> 00:19:35,797
And so there's a lot of factors that go
into it. I think that the true answer is

341
00:19:35,880 --> 00:19:39,997
like if either extreme is not quite right.
There are tools that can be

342
00:19:40,080 --> 00:19:43,197
applied to a specific situation. And I
think that this is not the first time

343
00:19:43,280 --> 00:19:44,437
we've had to think about this problem.

344
00:19:44,520 --> 00:19:47,597
It's not the last time we will have to
think about it. But one thing to note is

345
00:19:47,680 --> 00:19:50,237
that we have had our model in the hands

346
00:19:50,320 --> 00:19:55,317
of defenders um for some time that we've
been building up our trusted access

347
00:19:55,400 --> 00:19:58,317
program. That the model that
we're releasing is actually not

348
00:19:58,400 --> 00:20:01,437
cyber-permissive, right? That it
actually has a number of safeguards

349
00:20:01,520 --> 00:20:03,717
built into it. And that you can then

350
00:20:03,800 --> 00:20:08,477
have a gap between what you're privately
sharing, testing, those kinds of things.

351
00:20:08,560 --> 00:20:11,517
And so I think my my short answer
is like it's there's definitely these

352
00:20:11,600 --> 00:20:16,277
different schools of thought in terms
of values of is the value that you want to

353
00:20:16,360 --> 00:20:20,237
get these models into people's hands
and empower them or is the value that you

354
00:20:20,320 --> 00:20:23,397
kind of want them to be centralized and
controlled and that you don't want them

355
00:20:23,480 --> 00:20:27,437
in people's hands? That is something
that is a maybe underlying tension in

356
00:20:27,520 --> 00:20:28,997
some of these debates. But I think that

357
00:20:29,080 --> 00:20:34,477
the tactics, right, the you know, that
those almost flow from the details and

358
00:20:34,560 --> 00:20:38,517
that they can be informed by these values.
But either extreme, reflexively,

359
00:20:38,600 --> 00:20:42,237
I don't think will yield
the best outcome for the world.

360
00:20:42,320 --> 00:20:46,197
>> Okay, I want to ask you about agents.
Uh back to agents if we could. Um these

361
00:20:46,280 --> 00:20:50,317
agents work work the best
if you sort of let them uh

362
00:20:50,400 --> 00:20:54,917
have a high degree of autonomy.
I mean, sort of makes sense. Um

363
00:20:55,000 --> 00:20:58,477
so I'm just kind of curious to hear your
perspective. As we get more agents that

364
00:20:58,560 --> 00:21:01,997
can do more things and access more files
and work across programs, what is the

365
00:21:02,080 --> 00:21:05,437
proper amount of trust
to put into agents right now?

366
00:21:05,520 --> 00:21:08,117
>> So, I think that right now, actually,

367
00:21:08,200 --> 00:21:11,957
agents tend to be quite reliable. Um and

368
00:21:12,040 --> 00:21:15,597
even things like prompt injections,
I think that there's still holes there,

369
00:21:15,680 --> 00:21:18,037
but that we're patching them
and that the models are becoming much more

370
00:21:18,120 --> 00:21:22,837
resilient. But, I also think that the flip
side is that as these models be are

371
00:21:22,920 --> 00:21:27,077
given increasing responsibility
and access to more important context, that

372
00:21:27,160 --> 00:21:31,197
you need to have some answer for just
like if you have employees, you know, if

373
00:21:31,280 --> 00:21:34,717
you have a team of five employees,
they're all kind of trustworthy,

374
00:21:34,800 --> 00:21:39,317
fine. But, if you have 500,000 of the
same employees, that some somehow those

375
00:21:39,400 --> 00:21:42,997
numbers, right, just like that there's
the law of large numbers, that you start

376
00:21:43,080 --> 00:21:46,077
to worry about, "Okay, how do I have
good governance and oversight?" Right.

377
00:21:46,160 --> 00:21:49,237
And so, this is something where as we're
investing in these capabilities and

378
00:21:49,320 --> 00:21:52,837
making the super app more accessible
not just to coders, but to to any person

379
00:21:52,920 --> 00:21:56,037
doing work with a computer, also
investing in governance and oversight.

380
00:21:56,120 --> 00:21:59,797
And you can see this very concretely
in Workspace agents, which we released

381
00:21:59,880 --> 00:22:04,077
recently. So, that's within your
enterprise, you can now define agents.

382
00:22:04,160 --> 00:22:07,357
So, you get a hosted Codex harness
in the cloud. You can hook up tools, you

383
00:22:07,440 --> 00:22:10,917
can hook up to your Slack, and it's
doing work. It's like awesome. A lot of

384
00:22:11,000 --> 00:22:14,717
people use it. It's been very cool
to see how sort of viral it goes within an

385
00:22:14,800 --> 00:22:17,077
organization when you see you use
someone else's agent, you're like,

386
00:22:17,160 --> 00:22:18,917
"Wait, I can build one of these, too."

387
00:22:19,000 --> 00:22:21,797
And you can just fork it
and and do your own thing.

388
00:22:21,880 --> 00:22:25,677
And then that's an opportunity to have
great governance in that you can see

389
00:22:25,760 --> 00:22:29,477
that's baked into the product where your
IT organization can see all the agents

390
00:22:29,560 --> 00:22:31,957
that have been created that for an agent,
you can see the conversations

391
00:22:32,040 --> 00:22:35,477
it's had, and that you can think about
exactly what the guardrails are around

392
00:22:35,560 --> 00:22:39,637
it. So, I think that the short
answer is like you want to ramp the

393
00:22:39,720 --> 00:22:42,877
responsibility entrusted with the agent
and the diversity of things that agents

394
00:22:42,960 --> 00:22:45,917
are doing together with security,

395
00:22:46,000 --> 00:22:50,677
safety, observability, oversight. And if
you're not doing those in hand-in-hand,

396
00:22:50,760 --> 00:22:53,717
then I think that that that that's
a little bit out of balance. And I think

397
00:22:53,800 --> 00:22:56,317
it's important to to
think about both sides.

398
00:22:56,400 --> 00:22:59,797
>> Yeah, basically,
go ahead, but be careful.

399
00:22:59,880 --> 00:23:03,357
>> But, you and really lean in, right?
I think it's like as you scale, like you

400
00:23:03,440 --> 00:23:07,237
can prototype and that that it's just
the nature of scale that starts to bring

401
00:23:07,320 --> 00:23:11,317
in the do you still have the ability to
to oversee what's going on? So, you need

402
00:23:11,400 --> 00:23:15,037
to to kind of make sure at each step
do you feel like you're calibrated, you

403
00:23:15,120 --> 00:23:17,317
understand what your teams are up to.

404
00:23:17,400 --> 00:23:19,197
>> Greg, let's end with this.
Um you called

405
00:23:19,280 --> 00:23:22,597
this a compute-powered economy.
What does that mean?

406
00:23:22,680 --> 00:23:24,917
>> Well, I think we are heading to a world

407
00:23:25,000 --> 00:23:27,677
where the more compute is poured into a

408
00:23:27,760 --> 00:23:29,757
problem, the faster that problem will be

409
00:23:29,840 --> 00:23:34,757
solved. And that the ceiling of problem
that can be solved depends on how much

410
00:23:34,840 --> 00:23:37,877
compute is available. And you think
about things like drug discovery, right?

411
00:23:37,960 --> 00:23:39,837
Being able to solve complex diseases.

412
00:23:39,920 --> 00:23:43,717
Like, those are Solving complex diseases
like Alzheimer's is kind of outside of

413
00:23:43,800 --> 00:23:47,757
humanity's reach right now.
We've never really done it. But,

414
00:23:47,840 --> 00:23:49,037
imagine a world where you can take a

415
00:23:49,120 --> 00:23:51,557
gigawatt data center and have it just

416
00:23:51,640 --> 00:23:55,197
think about how to solve Alzheimer's.

417
00:23:55,280 --> 00:23:59,037
For a month, for a year, however long it
takes. And it may not be literally just

418
00:23:59,120 --> 00:24:03,037
cerebrally solving this problem, but it
may have to consult with world experts,

419
00:24:03,120 --> 00:24:06,277
maybe it has to suggest
experiments to get run in a wet

420
00:24:06,360 --> 00:24:11,397
lab. But, if you can actually solve
such a problem, that would be such a

421
00:24:11,480 --> 00:24:14,517
transformatively positive thing for
humanity. And I think we're heading to a

422
00:24:14,600 --> 00:24:18,717
world where that is how important
problems get solved. And that is how

423
00:24:18,800 --> 00:24:22,797
tasks in your daily life can also be
solved. Whether it's having an agent

424
00:24:22,880 --> 00:24:26,877
that knows you, that has your personal
context, that is trustworthy, that you

425
00:24:26,960 --> 00:24:30,037
can ask for advice
on health and you get back

426
00:24:30,120 --> 00:24:33,917
trustworthy information.
Um And that's just a thing that's a

427
00:24:34,000 --> 00:24:37,237
smartphone that's that's in your pocket,
right? You can just talk to and it'll be

428
00:24:37,320 --> 00:24:40,317
out there doing things. It proactively
knows what are your goals, what are your

429
00:24:40,400 --> 00:24:42,957
interest, and how it can help you. And I

430
00:24:43,040 --> 00:24:45,077
think that big and small compute is

431
00:24:45,160 --> 00:24:48,157
going to be the resource that shows how

432
00:24:48,240 --> 00:24:52,717
much computers can be used to help people,
to do work on behalf of people.

433
00:24:52,800 --> 00:24:55,677
And I think we're heading to that world
and it's one that we're all building

434
00:24:55,760 --> 00:24:58,797
collectively. >> Yeah,
and that I think would explain the

435
00:24:58,880 --> 00:25:03,277
massive investments that you've led
making these big infrastructure bets.

436
00:25:03,360 --> 00:25:06,517
>> Still not enough. We're going to feel
the scarcity. We're going to feel it.

437
00:25:06,600 --> 00:25:09,197
We're feeling it already. You can sense
it right now [music] on people who are

438
00:25:09,280 --> 00:25:12,157
trying to use these agents and just
simply cannot, you know, hitting the

439
00:25:12,240 --> 00:25:15,317
rate limits. Um so we're working
on behalf of our customers [music] on

440
00:25:15,400 --> 00:25:18,237
behalf of of everyone who wants to use
these agents to ensure that there is

441
00:25:18,320 --> 00:25:21,637
enough and I don't think we're going
to get there. We're going to do our best,

442
00:25:21,720 --> 00:25:24,997
but I think that we are headed to a world
of compute scarcity and uh again,

443
00:25:25,080 --> 00:25:28,117
I think this is something where >> [music]
>> we can all contribute to trying to help

444
00:25:28,200 --> 00:25:30,917
there just be more availability
of this in the world.

445
00:25:31,000 --> 00:25:33,557
>> Greg, busy day. Always appreciate
[music] your time. Always great to speak

446
00:25:33,640 --> 00:25:35,477
with you. Thanks again for coming on.

447
00:25:35,560 --> 00:25:38,600
>> Likewise. Great chatting.