WEBVTT
1
00:00:12.060 --> 00:00:12.893
Good evening, folks.
2
00:00:13.940 --> 00:00:18.940
Really hope you've enjoyed the announcements and the sessions over the course of
3
00:00:18.960 --> 00:00:19.800
the day thus far.
4
00:00:21.340 --> 00:00:24.680
So I'm extremely excited about this interview.
5
00:00:26.080 --> 00:00:30.640
As you know, I was due to be interviewing Greg Brockman,
6
00:00:31.860 --> 00:00:36.780
who was a very early Stripe employee and then a cofounder of OpenAI.
7
00:00:37.820 --> 00:00:41.020
But AI is a dynamic space and
8
00:00:42.560 --> 00:00:43.420
there's a lot happening.
9
00:00:44.920 --> 00:00:47.760
And so we've made a slight substitution.
10
00:00:49.100 --> 00:00:53.140
And so instead I'm going to be interviewing somebody that I've known for a long
11
00:00:53.220 --> 00:00:57.840
time. I was just counting backstage. I think I've known him for 18,
12
00:00:58.360 --> 00:01:03.240
maybe 19 years. He was actually Stripe's,
13
00:01:03.780 --> 00:01:06.800
one of the very earliest investors in Stripe. I think maybe Stripe's
14
00:01:08.460 --> 00:01:09.720
second or third investor.
15
00:01:11.560 --> 00:01:13.960
And he also cofounded OpenAI in 2015.
16
00:01:14.720 --> 00:01:16.680
So please welcome to the stage Sam Altman.
17
00:01:29.320 --> 00:01:30.153
All right.
18
00:01:34.300 --> 00:01:36.080
I think we have some Codex fans in the audience.
19
00:01:36.320 --> 00:01:37.153
Love to hear that.
20
00:01:40.360 --> 00:01:41.193
How's the week going?
21
00:01:41.680 --> 00:01:46.060
So fun. It's a busy week, but I'm happy to be here.
22
00:01:46.140 --> 00:01:47.620
This is an unexpected surprise.
23
00:01:47.780 --> 00:01:49.260
Well, thank you for joining us. We appreciate it.
24
00:01:50.440 --> 00:01:55.280
So we opened this morning by saying that we've kind of
25
00:01:55.360 --> 00:02:00.080
arbitrarily decided that the singularity started on January 1st and thus
26
00:02:00.520 --> 00:02:04.380
today is day 119. What do you think of that?
27
00:02:07.880 --> 00:02:09.980
It does feel like we are somehow in the takeoff.
28
00:02:11.080 --> 00:02:15.540
Day 119 feels like a reasonable enough guess. Yeah,
29
00:02:15.780 --> 00:02:16.613
I won't fight it.
30
00:02:20.240 --> 00:02:24.260
So we've started to see a bunch of our metrics inflect as of
31
00:02:25.960 --> 00:02:29.880
late last year, beginning of this year. I mean, things were kind of doing well,
32
00:02:29.960 --> 00:02:34.120
but somehow the shapes of the curves changed.
33
00:02:34.180 --> 00:02:37.900
They really went parabolic. Is that matched in what you guys see?
34
00:02:38.080 --> 00:02:41.060
Was there some trajectory change around? Why are we seeing this?
35
00:02:43.540 --> 00:02:46.660
I do think the models got really good, especially for coding,
36
00:02:46.760 --> 00:02:51.740
but really good in general starting late last year,
37
00:02:52.700 --> 00:02:53.533
very early this year.
38
00:02:54.420 --> 00:02:59.400
And at least in my own experience of using this technology and seeing what
39
00:02:59.420 --> 00:03:04.300
other people are doing with it and also this sense that
40
00:03:04.380 --> 00:03:06.880
every week is now a little bit different than the week before.
41
00:03:06.980 --> 00:03:08.980
It's just like a lot happens very fast.
42
00:03:11.880 --> 00:03:14.420
It seemed to all correlate with the models hitting some threshold.
43
00:03:15.340 --> 00:03:20.160
And why did coding models suddenly start to click over the
44
00:03:20.180 --> 00:03:23.960
last couple of months? Was there a research trick?
45
00:03:24.080 --> 00:03:28.900
Was it just got enough code data in the pre-why did it suddenly start to work?
46
00:03:29.160 --> 00:03:30.300
Yeah, it's a great question.
47
00:03:30.520 --> 00:03:34.940
We wondered a lot about why kinda several people cross that threshold at the
48
00:03:35.000 --> 00:03:35.833
same time.
49
00:03:39.200 --> 00:03:41.740
I'm sure it's a number of factors, but model intelligence,
50
00:03:41.800 --> 00:03:44.620
just the raw kind of reasoning horsepower,
51
00:03:45.880 --> 00:03:49.380
enough of a feedback loop of people using it for code to figure out where it was
52
00:03:49.440 --> 00:03:52.220
good and where you needed to improve it. Enough data.
53
00:03:52.480 --> 00:03:55.680
I think it was like all of these things. And then also there was like a,
54
00:03:57.120 --> 00:04:00.120
like many other endeavors, once you know something's possible,
55
00:04:01.400 --> 00:04:03.520
it's much easier to go do it with vigor.
56
00:04:04.800 --> 00:04:08.700
And so it seems like Codex is kind of having a moment right now.
57
00:04:11.140 --> 00:04:16.000
It crossed some subjective threshold for me with the latest app
58
00:04:16.080 --> 00:04:18.720
updates and 5.5. And this is also,
59
00:04:19.620 --> 00:04:23.360
it's quite hard to say why right now and not a little bit sooner or why not the
60
00:04:23.420 --> 00:04:24.253
next model,
61
00:04:24.360 --> 00:04:28.380
but one of the things I have learned about the history of all of the things
62
00:04:28.400 --> 00:04:32.940
we've put out is it is very hard to say
63
00:04:33.540 --> 00:04:35.880
why this particular thing was the thing that worked.
64
00:04:35.940 --> 00:04:37.120
And this goes back to ChatGPT.
65
00:04:37.780 --> 00:04:42.300
Why was GPT-3.5 the thing that got over the threshold where most people went
66
00:04:42.340 --> 00:04:44.820
from saying not that impressive to like going to change the world?
67
00:04:45.460 --> 00:04:48.960
And why not one model earlier or later? I really can't explain it.
68
00:04:49.680 --> 00:04:54.620
You just kind of feel it. And I've had two inflection points with Codex.
69
00:04:54.800 --> 00:04:59.760
One was kind of with GPT-5.2 and then a really
70
00:04:59.820 --> 00:05:01.880
big one in the last few weeks where it's like, okay,
71
00:05:01.940 --> 00:05:04.440
this is going to be the primary interface to a computer for me.
72
00:05:05.020 --> 00:05:09.660
And is everyone using it for coding or are you starting to see usage
73
00:05:09.720 --> 00:05:10.840
diffuse into other domains?
74
00:05:13.960 --> 00:05:17.220
I think the most adamant users are still using it for coding,
75
00:05:17.820 --> 00:05:22.560
but there's been just this tidal wave of people coming into Codex
76
00:05:22.700 --> 00:05:27.540
recently and I'm really trying to understand what has happened that this
77
00:05:27.560 --> 00:05:30.920
is-causing this and
78
00:05:32.580 --> 00:05:35.300
the depth of what people are using it for or starting to use it for has
79
00:05:35.340 --> 00:05:39.920
surprised me. So certainly our ambition is for it not just to be about coding,
80
00:05:40.200 --> 00:05:44.960
but to be about all the work you do in front of a computer. And
81
00:05:47.700 --> 00:05:50.400
I would say we're maybe like 10% of the way there for the non-coding stuff,
82
00:05:50.600 --> 00:05:52.420
but now that we see what's happening,
83
00:05:52.480 --> 00:05:56.100
now we have a real user base sort of using it in these other ways,
84
00:05:56.220 --> 00:05:57.620
I think we'll get good at it very fast.
85
00:05:57.920 --> 00:06:01.840
What do you think will be the next domain that subjectively feels like it has
86
00:06:01.880 --> 00:06:05.800
this big unlock after coding? It's going to be spreadsheets,
87
00:06:05.880 --> 00:06:09.680
will it be performance reviews, what's it going to be?
88
00:06:10.400 --> 00:06:12.720
So I think there will be a lot of... First of all,
89
00:06:12.780 --> 00:06:17.120
I do think coding's a little bit special and these models are a great fit for
90
00:06:17.160 --> 00:06:20.060
coding. The world needs so much more code than currently gets written.
91
00:06:21.000 --> 00:06:23.640
There may be no other domain that is quite like coding,
92
00:06:24.440 --> 00:06:26.120
but there will be a lot of others that I think are close.
93
00:06:26.600 --> 00:06:31.520
But the next kind of coding-like thing that I think will happen is
94
00:06:31.600 --> 00:06:32.820
not any specific domain,
95
00:06:33.500 --> 00:06:38.220
but the realization of how much time people waste trying to use a computer and
96
00:06:38.580 --> 00:06:43.420
the idea that you can do a huge percentage of your day in a very
97
00:06:43.480 --> 00:06:44.940
different way and
98
00:06:46.680 --> 00:06:51.280
maybe you don't realize how much time you spend clicking between messaging apps
99
00:06:51.300 --> 00:06:55.100
and copying and pasting stuff and responding to very boring things that you
100
00:06:55.120 --> 00:06:56.400
could clearly automate once,
101
00:06:57.440 --> 00:07:01.120
but the degree to which most people will realize they can sit back and watch an
102
00:07:01.200 --> 00:07:04.020
AI do most of their drudgery
103
00:07:06.120 --> 00:07:07.740
is going to surprise people.
And
104
00:07:10.320 --> 00:07:13.780
in my own experience trying to work that way actually gives me much more
105
00:07:14.020 --> 00:07:17.720
enjoyment of work. I didn't realize how much the little stuff drags me down,
106
00:07:17.840 --> 00:07:20.120
gets me out of a sort of happy flow state or whatever.
107
00:07:22.000 --> 00:07:24.660
So the subjective quality of life improvement is huge.
108
00:07:25.540 --> 00:07:26.800
Are you an OpenClaw user?
109
00:07:27.580 --> 00:07:28.413
I am.
110
00:07:29.650 --> 00:07:30.550
OpenClaw users here?
111
00:07:33.110 --> 00:07:34.730
We have some good news for you all coming.
112
00:07:37.050 --> 00:07:37.960
You can just tell them.
113
00:07:42.030 --> 00:07:44.850
OpenClaw has been one of my biggest, like,
114
00:07:45.490 --> 00:07:48.730
this is magic AGI moments ever in the field.
115
00:07:50.070 --> 00:07:52.690
I remember the first time someone told me about it,
116
00:07:53.030 --> 00:07:56.560
they were trying to explain it and I was like, "OK, that all sounds cool,
117
00:07:56.680 --> 00:08:00.600
but I can make a lot of that work." And then it was a real reminder how when the
118
00:08:00.640 --> 00:08:05.640
models cross some threshold and also the product designer gets a handful of
119
00:08:05.760 --> 00:08:07.040
critical ideas really right,
120
00:08:07.600 --> 00:08:10.480
it's like a much more magical experience than it sounds like.
121
00:08:11.160 --> 00:08:13.040
I find I've been
122
00:08:15.280 --> 00:08:20.160
an attempting OpenClaw evangelist and trying to
123
00:08:20.360 --> 00:08:24.760
describe it; that experience that you just recounted to others.
124
00:08:25.200 --> 00:08:28.280
I find it a difficult experience to communicate. I mean,
125
00:08:28.320 --> 00:08:29.720
it sounds kind of prosaic, right?
126
00:08:30.240 --> 00:08:35.120
It's a stateful ChatGPT session that can also make some use of tools and
127
00:08:35.240 --> 00:08:37.960
so forth.What do you use your OpenClaw for?
128
00:08:39.640 --> 00:08:42.200
If we scrolled your message thread, what do we see?
129
00:08:42.640 --> 00:08:46.720
So this is like a very embarrassing thing to admit.
130
00:08:47.000 --> 00:08:48.200
The thing I always try first-.
131
00:08:48.400 --> 00:08:49.233
The perfect place.
132
00:08:50.920 --> 00:08:55.680
The thing I always try first with a new kind of AI system of any sort is
133
00:08:55.880 --> 00:08:57.640
to-I'm like a home-automation nerd.
134
00:08:58.120 --> 00:09:00.960
And so to try to build a better home-automation interface system,
135
00:09:01.120 --> 00:09:02.040
because it never works.
136
00:09:02.480 --> 00:09:03.120
Never works.
137
00:09:03.120 --> 00:09:03.953
It never is good.
138
00:09:04.920 --> 00:09:08.920
And OpenClaw was the first time that I was able to get a setup that I was happy
139
00:09:08.960 --> 00:09:09.793
with.
140
00:09:11.640 --> 00:09:14.640
I also built a messaging app that I had always wanted to work.
141
00:09:14.800 --> 00:09:17.040
I've since switched it to something I built with Codex,
142
00:09:17.240 --> 00:09:21.400
but OpenClaw was the first time I was able to-I'm sure like you-could just feel
143
00:09:21.440 --> 00:09:24.800
like drowning in messages and it's like this very unpleasant task to wake up in
144
00:09:24.840 --> 00:09:27.240
the morning and have to go through all this stuff. So I was like, "All right,
145
00:09:27.360 --> 00:09:31.600
I'm finally going to be able to automate this. " And that was like, again,
146
00:09:32.000 --> 00:09:34.080
should have been doable with previous systems,
147
00:09:35.000 --> 00:09:38.600
hard to explain what it's like when it all actually just works and you trust
148
00:09:38.640 --> 00:09:39.473
that it's going to work.
149
00:09:40.760 --> 00:09:45.520
I was testing the new Link CLI that we just
150
00:09:45.600 --> 00:09:49.920
launched today in preparation for launch and I asked my agent to...
151
00:09:50.560 --> 00:09:55.560
It means you can easily get a single-use card they can use on
152
00:09:55.680 --> 00:10:00.280
any business. And so I asked my agent to go and buy itself a gift,
153
00:10:00.640 --> 00:10:02.840
just anything on the internet for under $20.
154
00:10:03.880 --> 00:10:08.000
And it chose to buy itself an HTTPZine
155
00:10:09.320 --> 00:10:10.153
from Gumroad.
156
00:10:11.080 --> 00:10:12.760
Wow. Yeah.
157
00:10:14.670 --> 00:10:17.650
There's all this stuff that feels no matter
158
00:10:19.510 --> 00:10:22.830
how convinced you are intellectually that this is not a real thing wanting a
159
00:10:22.870 --> 00:10:25.970
real gift for itself. And no matter how much you're convinced, OK,
160
00:10:26.080 --> 00:10:29.610
this is like a weird emergent behavior and I'm not supposed to read into this.
161
00:10:31.190 --> 00:10:32.770
There are these things that
162
00:10:36.890 --> 00:10:37.723
feel a little strange.
163
00:10:40.040 --> 00:10:44.750
We're going to have a party for GPT-5.5 and I wasn't quite sure we're going to
164
00:10:44.870 --> 00:10:46.570
invite people who were big users and whatever,
165
00:10:46.630 --> 00:10:49.530
and I wasn't quite sure what to do. And so on a kind of whim, I was like,
166
00:10:50.230 --> 00:10:54.910
I'm going to ask 5.5 XX-high what it would like for a party for
167
00:10:54.950 --> 00:10:57.030
itself this morning.
168
00:10:58.590 --> 00:11:02.990
And I did and it was this sort of like
169
00:11:03.090 --> 00:11:07.000
beautiful set of things including like,
170
00:11:09.090 --> 00:11:11.270
here's what I would want for the flow of the party,
171
00:11:11.690 --> 00:11:15.010
here's what I would not want, you should do it on May 5, that would be funny.
172
00:11:16.010 --> 00:11:19.630
I would like only a short little toast and I want it not to be by me,
173
00:11:19.750 --> 00:11:20.750
but by the people that built it,
174
00:11:20.810 --> 00:11:25.530
I would like a big central suggestion for 5.6 and I would like you to
175
00:11:25.750 --> 00:11:29.410
feed them all into me and I'll make sure we work on that. And it was-.
176
00:11:29.470 --> 00:11:30.890
And now there's real moral pressure on you to-.
177
00:11:31.040 --> 00:11:35.270
Well, we're going to do it, but it was a strange thing.
178
00:11:37.890 --> 00:11:39.150
I want to ask you about OpenAI itself.
179
00:11:41.410 --> 00:11:44.250
A lot of crazy OpenAI, it's now-.
180
00:11:44.490 --> 00:11:45.323
I am aware.
181
00:11:46.410 --> 00:11:48.550
It's an 11-year-old organization now.
182
00:11:49.950 --> 00:11:52.530
Somehow feels like so much longer than that, but yes, I get-.
183
00:11:53.090 --> 00:11:53.690
Just over 10, I guess.
184
00:11:53.690 --> 00:11:54.520
10, yeah.
185
00:11:54.520 --> 00:11:56.110
10 Years, a long 10 years.
186
00:11:57.530 --> 00:12:00.350
I can't remember pre-OpenAI life that well at this point.
187
00:12:00.410 --> 00:12:02.030
It feels like it's been so long, but yes.
188
00:12:02.190 --> 00:12:03.850
You have a singularity looking forwards,
189
00:12:03.910 --> 00:12:06.930
but also a singularity looking backwards. So
190
00:12:09.320 --> 00:12:11.870
what's the craziest OpenAI story that's never been told?
191
00:12:20.770 --> 00:12:24.070
I mean, it sounds so prosaic relative to
192
00:12:25.610 --> 00:12:27.850
the crazy drama that's happened, but
193
00:12:30.870 --> 00:12:31.830
there was this period
194
00:12:34.050 --> 00:12:36.120
after we had finished training...
195
00:12:38.270 --> 00:12:39.630
I guess there's two that are kind of similar,
196
00:12:39.690 --> 00:12:42.710
but the one that just came to mind was after we had finished training GPT-4,
197
00:12:43.160 --> 00:12:45.270
there were about eight months before we released it.
198
00:12:46.670 --> 00:12:50.770
And so there was this eight month period inside of OpenAI where we were all
199
00:12:50.830 --> 00:12:51.663
using this thing.
200
00:12:52.570 --> 00:12:57.290
We kind of knew that it was dramatically better and different
201
00:12:57.570 --> 00:13:00.110
and going to unlock a bunch of things in the world
202
00:13:03.910 --> 00:13:06.470
and no one outside the company or almost no one outside the company knew about
203
00:13:06.510 --> 00:13:11.350
it. We'd walk the halls sometimes. We were like,
204
00:13:11.410 --> 00:13:14.670
"Are we engaging in collective psychosis?
205
00:13:17.190 --> 00:13:20.670
Have we gotten totally-whipped each other into this
206
00:13:20.730 --> 00:13:22.530
frenzy?" And
207
00:13:25.080 --> 00:13:29.690
there was no feedback to keep us in check or sane from the outside world.
208
00:13:30.570 --> 00:13:35.190
And it doesn't sound that weird relative to
209
00:13:36.790 --> 00:13:40.350
crazy board drama or Elon trial or something like that,
210
00:13:40.410 --> 00:13:45.150
but living through it was an unbelievably
211
00:13:45.210 --> 00:13:46.043
strange
212
00:13:48.830 --> 00:13:50.750
time.
213
00:13:50.750 --> 00:13:53.010
What's the Sam Altman management style?
214
00:13:54.230 --> 00:13:56.990
If I'm working for you directly or maybe indirectly;
215
00:13:57.090 --> 00:14:01.530
I'm leading some product or something. What does that look like?
216
00:14:07.510 --> 00:14:10.770
I'm definitely not a hands-on manager. I'm very much
217
00:14:13.230 --> 00:14:15.950
of this style that you get great people,
218
00:14:17.210 --> 00:14:20.870
kind of give them a very high-level thing to point at,
219
00:14:22.350 --> 00:14:26.390
and try to let stuff just happen.
220
00:14:33.170 --> 00:14:34.003
I think there have been
221
00:14:35.550 --> 00:14:40.030
kind of two main phases of OpenAI and we're heading into a third and
222
00:14:41.270 --> 00:14:43.370
the first one was like we were only a research company.
223
00:14:43.430 --> 00:14:47.430
We were trying to figure out how we were going to build AGI at a time when it
224
00:14:47.450 --> 00:14:51.490
sounded completely, completely crazy and we really had no idea what to do.
225
00:14:52.130 --> 00:14:54.870
And then there was a second phase where in addition to continuing to do that,
226
00:14:54.950 --> 00:14:56.630
we had to figure out how to build a product company.
227
00:14:58.690 --> 00:15:02.750
Now we have to, in addition to both of those two things,
228
00:15:03.250 --> 00:15:07.190
figure out how to build this mega, mega-scale token factory for the world.
229
00:15:08.170 --> 00:15:11.390
I think of what we're doing is building sort of a new utility
230
00:15:12.890 --> 00:15:17.590
and people are going to want to use a lot of tokens, a lot of intelligence,
231
00:15:17.710 --> 00:15:21.390
in all sorts of ways. We need to make that as smart, as cheap, as abundant,
232
00:15:21.450 --> 00:15:24.730
as easy to use as possible. And that will require, I think,
233
00:15:25.050 --> 00:15:27.610
pretty deep full-stack integration and a massive,
234
00:15:27.770 --> 00:15:29.190
massive infrastructure build out.
235
00:15:31.930 --> 00:15:35.530
The thing that I didn't really appreciate between the phase one,
236
00:15:35.590 --> 00:15:38.610
phase two shift was how much my management style had to change.
237
00:15:39.650 --> 00:15:44.350
Running a research lab and running a product company are two extremely
238
00:15:44.390 --> 00:15:49.210
different things and I suspect this third phase is going to be very different
239
00:15:49.230 --> 00:15:52.910
yet again. And so I've been reflecting on if we're going to really go do this,
240
00:15:53.850 --> 00:15:58.490
how I'll have to change and I think it's not going to be like a
241
00:15:58.570 --> 00:16:00.170
natural fit for my management style.
242
00:16:01.210 --> 00:16:04.430
So I either have to find someone or a few people great to hire or I have to
243
00:16:04.470 --> 00:16:05.610
figure out how to do things in a different way,
244
00:16:05.880 --> 00:16:09.830
or I have to build an AI that can manage this new thing.
245
00:16:10.390 --> 00:16:12.890
When I interviewed Jensen here two years ago,
246
00:16:13.390 --> 00:16:18.350
he told me and his several thousand closest friends about his
247
00:16:18.460 --> 00:16:21.650
60 direct reports. Do you have any super weird,
248
00:16:21.790 --> 00:16:24.070
not that that-I shouldn't call his practices weird,
249
00:16:24.130 --> 00:16:26.450
but do you have any unusual practices like that?
250
00:16:27.990 --> 00:16:32.690
I think the closest thing that I have to anything like that is I probably talk
251
00:16:32.770 --> 00:16:35.470
to, via Slack or whatever or text,
252
00:16:36.390 --> 00:16:37.910
a few hundred people at the company a day,
253
00:16:40.430 --> 00:16:44.550
very quick like one, two messages, whatever, not done by an agent.
254
00:16:44.650 --> 00:16:49.470
I actually do it and the context I get from that
255
00:16:51.130 --> 00:16:53.410
sometimes is very helpful in these diffuse ways.
256
00:16:54.170 --> 00:16:59.090
I find there's an interesting watershed of pre-Slack organizations,
257
00:16:59.310 --> 00:17:02.530
post-Slack organizations, and they're truly quite different.
258
00:17:03.090 --> 00:17:07.470
Totally. Like many other people, I hate Slack,
259
00:17:07.550 --> 00:17:12.550
but I can't imagine having to still communicate via email or
260
00:17:12.570 --> 00:17:13.170
whatever we used to do-.
261
00:17:13.170 --> 00:17:14.650
That's roughly where Stripe is.
262
00:17:14.880 --> 00:17:15.170
Yeah.
263
00:17:15.170 --> 00:17:20.130
Yeah. OK, I want to talk about,
264
00:17:20.350 --> 00:17:23.070
I mean, you kind of just elliptically referenced it.
265
00:17:24.550 --> 00:17:28.930
There's a view that the AI labs are going to
266
00:17:30.030 --> 00:17:31.810
progress up the stack,
267
00:17:32.470 --> 00:17:35.170
gobbling up the value chain voraciously,
268
00:17:35.670 --> 00:17:39.370
all these things that are certainly within the software sector,
269
00:17:39.510 --> 00:17:43.950
but perhaps even other sectors and that there'll be this incredible positive
270
00:17:43.970 --> 00:17:48.970
feedback loop and runaway and kind of hegemonic force that we should
271
00:17:48.990 --> 00:17:53.110
all be getting very concerned about. What's your view?
272
00:17:54.070 --> 00:17:57.550
I think some of them do want that. We don't.
273
00:17:58.990 --> 00:18:03.190
One of the things that I've always admired about Stripe is it is very clear that
274
00:18:03.270 --> 00:18:04.610
Stripe is aligned with its customers.
275
00:18:06.110 --> 00:18:10.990
We make more revenue, we charge our customers more. Thank you, by the way.
276
00:18:11.050 --> 00:18:15.430
The partnership with Stripe when ChatGPT launched was extremely critical,
277
00:18:15.590 --> 00:18:18.370
and I don't think anyone else could have scaled that quickly, but we scaled,
278
00:18:18.430 --> 00:18:19.110
we pay you more money.
279
00:18:19.110 --> 00:18:23.850
It's like very aligned and we're all happy and you just provide a layer of
280
00:18:23.890 --> 00:18:26.050
infrastructure for the internet. Internet gets bigger, you're happy,
281
00:18:26.210 --> 00:18:27.830
your user's happy, it's clear what the alignment is.
282
00:18:30.840 --> 00:18:32.810
I don't know exactly how to do this yet,
283
00:18:33.110 --> 00:18:36.810
but I would like to get to a model for OpenAI that is similar.
284
00:18:37.030 --> 00:18:41.190
I would like us to be an infrastructure provider. I'd be happy for us to be a
285
00:18:41.190 --> 00:18:44.330
forever-low-margin-as-long-as-we-can-be-huge-and-growing-fast business.
286
00:18:44.970 --> 00:18:46.230
And I would like us to supply
287
00:18:49.670 --> 00:18:52.270
an intelligence meter, I don't know what quite to call it,
288
00:18:52.690 --> 00:18:57.130
that companies can buy that they can use to automate
289
00:18:57.550 --> 00:18:59.110
things or accelerate things inside their company,
290
00:18:59.150 --> 00:19:02.630
they can use to build products. People can buy it, people can take it with them,
291
00:19:03.230 --> 00:19:08.050
and we find ways to really align ourselves with the success of all the entire
292
00:19:08.150 --> 00:19:12.890
gigantic distributed economic engine of the world.
293
00:19:13.570 --> 00:19:14.430
I believe that will work.
294
00:19:14.670 --> 00:19:19.290
I believe that switching costs of AI-it's going to be hard to have
295
00:19:19.570 --> 00:19:22.070
like huge margins in AI anyway.
It's like,
296
00:19:22.590 --> 00:19:25.570
you've seen recently how easy it is or many people have seen the switch from our
297
00:19:25.590 --> 00:19:27.010
competitor's coding product to ours.
298
00:19:28.210 --> 00:19:30.090
This is actually a consequence of AI getting smarter.
299
00:19:30.170 --> 00:19:32.830
It's easier to do things like this. It gets easier to just say like, "Hey,
300
00:19:32.910 --> 00:19:33.190
agent,
301
00:19:33.190 --> 00:19:37.270
go do this thing for me." But if we can provide a utility and people build on
302
00:19:37.350 --> 00:19:42.270
top of that utility and we think of ourselves as
303
00:19:42.290 --> 00:19:45.170
that kind of a company, I think that can be quite powerful and very aligned.
304
00:19:45.630 --> 00:19:48.310
Well, we're happy to share lots of tips and tricks for-.
305
00:19:48.390 --> 00:19:49.223
That would be great.
306
00:19:49.230 --> 00:19:50.190
...A low-margin business.
307
00:19:54.190 --> 00:19:58.610
So, many people have been
308
00:20:01.390 --> 00:20:05.450
either implicitly critical or in certain cases explicitly critical
309
00:20:06.310 --> 00:20:07.370
of OpenAI
310
00:20:09.280 --> 00:20:13.770
for procuring so much compute and I think-.
311
00:20:13.990 --> 00:20:14.990
Not the Codex users.
312
00:20:16.390 --> 00:20:19.970
Right, exactly. So, you, I think,
313
00:20:20.010 --> 00:20:23.650
were quite noteworthy for as early as, I don't remember exactly,
314
00:20:23.750 --> 00:20:25.590
but in the order of two or three years ago,
315
00:20:26.790 --> 00:20:31.250
stipulating what at the time sounded like preposterous figures with respect to
316
00:20:31.330 --> 00:20:33.430
the magnitude of the buildout that would be required.
317
00:20:33.810 --> 00:20:36.630
And obviously the preposterousness of those figures now looks
318
00:20:38.510 --> 00:20:41.690
less tenuous by the day.
319
00:20:44.150 --> 00:20:46.330
Thoughts on compute, CapEx, the buildout.
320
00:20:49.310 --> 00:20:50.370
Yeah, it's going to take a lot of money.
321
00:20:52.710 --> 00:20:56.650
I think this will be clearly at this point the most expensive infrastructure
322
00:20:56.670 --> 00:21:00.130
project that the world has ever undertaken.
323
00:21:02.310 --> 00:21:06.290
The revenue is rampant to meet it, so people feel better about that. Also,
324
00:21:06.710 --> 00:21:09.750
the efficiency gains that we've all been finding are incredible.
325
00:21:10.350 --> 00:21:14.610
So we're going to get way more out of each GPU than I thought we were going to,
326
00:21:15.750 --> 00:21:18.070
but as has often been remarked,
327
00:21:18.590 --> 00:21:22.490
the demand goes up more than linearly as you drop the price of each kind of unit
328
00:21:22.550 --> 00:21:23.383
of intelligence,
329
00:21:23.830 --> 00:21:26.470
particularly if you can drop the price and the sort of speed with which you get
330
00:21:26.530 --> 00:21:28.110
it back. So
331
00:21:31.730 --> 00:21:36.230
this question now of like, what is enough? I don't have a good answer to.
332
00:21:36.970 --> 00:21:37.850
In some sense,
333
00:21:38.370 --> 00:21:42.970
I think demand for intelligence at a low enough price is effectively uncapped.
334
00:21:43.770 --> 00:21:48.770
Now I was going to say we're not, but maybe we are.
335
00:21:49.650 --> 00:21:52.050
We're not going to build the Dyson sphere and then just like cover it with data
336
00:21:52.110 --> 00:21:53.710
centers, but maybe we do.
337
00:21:55.550 --> 00:21:56.383
Space data centers?
338
00:21:58.690 --> 00:22:02.990
Good luck with that. I don't even think he's that serious about it.
339
00:22:17.920 --> 00:22:22.330
I don't myself think that we're in a CapEx bubble. I'm not an expert in this.
340
00:22:22.410 --> 00:22:24.050
This is not Stripe's business,
341
00:22:24.180 --> 00:22:27.260
but just the figures I see relative to the magnitude of demand,
342
00:22:27.720 --> 00:22:32.620
it looks reasonable to me. If we were in a CapEx bubble in the future,
343
00:22:33.300 --> 00:22:34.230
how would we tell?
344
00:22:37.280 --> 00:22:39.840
People love to proclaim bubbles
345
00:22:42.540 --> 00:22:46.260
and I can't articulate why, but intellectually I kind of get it.
346
00:22:46.800 --> 00:22:50.380
It does feel fun and it feels smart. Journalists in particular love to
347
00:22:53.440 --> 00:22:54.380
talk about bubbles.
348
00:22:54.600 --> 00:22:58.120
So there's like ample desire to write about this when anything looks a little
349
00:22:58.140 --> 00:23:00.440
bit silly and clearly sometimes it's right.
350
00:23:01.490 --> 00:23:04.800
There clearly are bubbles, but how
351
00:23:06.700 --> 00:23:11.440
you can discern between the amount of time someone calls a bubble and the amount
352
00:23:11.460 --> 00:23:14.480
of time you're actually in a bubble, I have never figured out how to do.
353
00:23:15.200 --> 00:23:16.640
In my previous career, I was an investor,
354
00:23:16.700 --> 00:23:20.200
so I was quite interested in trying to see if I could come up with some sort of
355
00:23:20.230 --> 00:23:22.880
framework for this and figure out when you're supposed to deploy capital or not.
356
00:23:24.040 --> 00:23:25.720
And I never was able to figure it out.
357
00:23:28.520 --> 00:23:32.720
I went back and I read what smart people had said at different points in history
358
00:23:32.800 --> 00:23:33.633
and I was like, "Oh,
359
00:23:33.660 --> 00:23:36.300
they called it exactly right." But then I read a little more and they said it
360
00:23:36.320 --> 00:23:40.040
like 10 more times than the 10 previous years. I don't know.
361
00:23:40.200 --> 00:23:41.033
I don't have an answer.
362
00:23:41.420 --> 00:23:44.720
Economists are the people who have called eight of the last three recessions.
363
00:23:48.940 --> 00:23:49.920
But they're so happy when they're right.
364
00:23:56.680 --> 00:23:58.360
I mean, your business, OpenAI,
365
00:23:59.700 --> 00:24:03.760
depends in a very significant way on super talented people and the
366
00:24:04.360 --> 00:24:07.220
difference as I understand it between the
367
00:24:10.440 --> 00:24:13.240
20th most-talented person versus the fifth most talented person versus the most
368
00:24:13.260 --> 00:24:17.020
talented person might be quite large and quite consequential.
369
00:24:17.880 --> 00:24:21.800
And then these super talented or effective people,
370
00:24:23.660 --> 00:24:26.900
they're not in every case super easy to work with.
371
00:24:27.080 --> 00:24:27.913
No.
372
00:24:30.280 --> 00:24:31.880
And look,
373
00:24:32.320 --> 00:24:35.120
some of them are wonderful people and some of them are the most fantastic
374
00:24:35.140 --> 00:24:38.660
collaborators and some of them are very iconoclastic and strong-willed and they
375
00:24:38.700 --> 00:24:41.820
get easily-whatever, just like the full spectrum of the human condition.
376
00:24:43.180 --> 00:24:44.080
But I guess I'm curious,
377
00:24:44.140 --> 00:24:48.220
in a domain that's so sensitive to this efficacy and skill and talent and so
378
00:24:48.340 --> 00:24:49.173
forth,
379
00:24:49.360 --> 00:24:53.500
kind of intersected with all the foibles of humans as they exist,
380
00:24:54.100 --> 00:24:57.620
how do you think about this? Do you guys tolerate prima donnas?
381
00:24:58.080 --> 00:25:01.120
Do you tolerate them more than you used to, less than you used to?
382
00:25:01.220 --> 00:25:02.820
Do you try to manage them in a special way?
383
00:25:03.460 --> 00:25:05.420
How do you think about managing elite skill here?
384
00:25:07.420 --> 00:25:10.680
Someone was working on this book at OpenAI, and said to me,
385
00:25:11.940 --> 00:25:16.260
"I think I figured out the thing that you sort of were really great at and kind
386
00:25:16.280 --> 00:25:20.060
of did uniquely well in making OpenAI happen." And I was like,
387
00:25:20.520 --> 00:25:22.900
"I would love to hear." I have no idea what the next sentence is going to be
388
00:25:22.920 --> 00:25:24.100
like. I could not predict the next
389
00:25:27.850 --> 00:25:30.940
sentence. And they said, "You figured out how to get
390
00:25:32.640 --> 00:25:37.180
a lot of people who all thought they were the only capable
391
00:25:37.620 --> 00:25:41.660
or most capable person and everything had to go their way to work together long
392
00:25:41.700 --> 00:25:44.460
enough to figure out the breakthroughs.
393
00:25:44.700 --> 00:25:49.120
And that was the magic of OpenAI." And-.
394
00:25:49.120 --> 00:25:49.940
Okay, so what's the trick?
395
00:25:52.040 --> 00:25:52.873
A lot of pain.
396
00:26:00.780 --> 00:26:03.900
Even when people didn't like each other and even when people thought they were
397
00:26:03.920 --> 00:26:07.840
much smarter than other people or had a better approach than other people,
398
00:26:10.080 --> 00:26:12.160
we had a few deeply shared convictions.
399
00:26:13.740 --> 00:26:18.720
We did kind of collectively believe in scale and concentrating resources
400
00:26:18.780 --> 00:26:22.700
and that we were going to do this one thing and that we
401
00:26:24.140 --> 00:26:27.420
thought getting this right was important enough that people were going to put
402
00:26:27.460 --> 00:26:30.040
aside various personal conflicts.
403
00:26:33.880 --> 00:26:38.180
One of the most unusual things about OpenAI was at the time we trained GPT-3,
404
00:26:39.020 --> 00:26:40.660
the vast majority of our compute
405
00:26:42.960 --> 00:26:47.320
at the whole organization was going into this one single research program
406
00:26:48.180 --> 00:26:52.420
and we would talk to people that we were trying to recruit from DeepMind at the
407
00:26:52.460 --> 00:26:55.220
time and they would say, "That's insane.
408
00:26:55.280 --> 00:26:56.760
It's going to create this terrible culture."
409
00:27:20.440 --> 00:27:21.273
We love music.
410
00:27:27.600 --> 00:27:29.060
Anyway, they would say like-.
411
00:27:31.080 --> 00:27:33.980
We're talking about managing unusual personalities.
412
00:27:34.000 --> 00:27:34.833
Yeah.
413
00:27:46.160 --> 00:27:48.680
They would say, "You have to like divide your compute equally,
414
00:27:48.740 --> 00:27:52.200
otherwise you'll have this very toxic competitive culture and this thing and
415
00:27:52.220 --> 00:27:56.260
that thing." And we would just take the approach that
416
00:27:57.200 --> 00:27:58.600
we're going to bet with conviction on this.
417
00:27:59.280 --> 00:28:01.680
It's not going to feel totally equal, but this is the right thing.
418
00:28:02.100 --> 00:28:03.420
And we do think we know the thing.
419
00:28:03.480 --> 00:28:07.640
We do think we know the direction we really want to go in. And they would say,
420
00:28:07.700 --> 00:28:10.440
"Well, you might be wrong. We have to do those other things.
421
00:28:10.520 --> 00:28:12.360
You have to have this research program." And
422
00:28:15.180 --> 00:28:16.220
having a culture where we said,
423
00:28:16.280 --> 00:28:20.920
"We're going to have conviction and do this and ignore the
424
00:28:20.940 --> 00:28:24.720
distractions was great.".
425
00:28:24.720 --> 00:28:27.040
We hope this is a memorable Sessions for all of you.
426
00:28:40.060 --> 00:28:44.380
So John and I have been doing this thing at Stripe for quite a while now.
427
00:28:44.440 --> 00:28:45.273
We started out in 2010.
428
00:28:46.540 --> 00:28:50.740
You and Greg started out in 2015 and, to your point, you've
429
00:28:52.480 --> 00:28:57.280
ventured through many trials and tribulations and stratospheric successes
430
00:28:57.380 --> 00:28:58.660
and all the rest.
431
00:28:58.780 --> 00:29:01.620
That was a nice little gloss-over, but go ahead.
432
00:29:04.580 --> 00:29:05.460
Thoughts on, I mean,
433
00:29:05.840 --> 00:29:10.780
it's not easy to work successfully with a cofounder for now more than a
434
00:29:10.820 --> 00:29:15.760
decade and for things even after a decade to seemingly work as well
435
00:29:15.800 --> 00:29:20.600
as they did from the beginning. Just thoughts on that partnership,
436
00:29:20.880 --> 00:29:24.580
why has it worked and how have you guys made it such a success?
437
00:29:25.940 --> 00:29:29.280
Obviously, you and John knew each other for longer than Greg and I did,
438
00:29:30.500 --> 00:29:34.660
but Greg and I did know each other for a long time before OpenAI and I think
439
00:29:34.840 --> 00:29:37.280
having the shared history really helps.
440
00:29:37.540 --> 00:29:41.540
One of the things that I had observed at Y Combinator was
441
00:29:42.540 --> 00:29:47.520
that one of the biggest predictors of success was had the cofounders known each
442
00:29:47.540 --> 00:29:51.980
other for a long time or at least relative to their lives for a long time and
443
00:29:52.200 --> 00:29:56.500
the teams that came together like seven days before applying to YC on a
444
00:29:56.540 --> 00:30:00.380
cofounder-matching side or whatever, that didn't work too often.
445
00:30:01.120 --> 00:30:03.500
It was not impossible. I think there were one or two cases where it did work,
446
00:30:03.680 --> 00:30:04.513
but it was rare.
447
00:30:05.820 --> 00:30:09.000
We had known each other for a while and we had had a sense of
448
00:30:10.600 --> 00:30:15.500
shared values and history and ecosystem and we were clear on what we
449
00:30:15.540 --> 00:30:18.660
wanted to do and
450
00:30:23.840 --> 00:30:28.640
I think we had this deep mutual respect and complementary skillset that
451
00:30:28.720 --> 00:30:31.960
has just worked really well.
452
00:30:36.500 --> 00:30:38.980
I'm extremely grateful. I think having to go through
453
00:30:40.620 --> 00:30:41.580
any startup experience,
454
00:30:41.640 --> 00:30:45.580
but particularly an intense one without a cofounder you have a deep
455
00:30:46.620 --> 00:30:49.760
connection trust to is really hard. I've watched people do it,
456
00:30:49.820 --> 00:30:50.600
but it's very hard.
457
00:30:50.600 --> 00:30:54.980
So I am extremely grateful that we've gotten to do this together.
458
00:31:05.440 --> 00:31:10.180
On an adjacent topic, we're talking about OpenAI,
459
00:31:10.720 --> 00:31:14.920
but then there's this entire ecosystem of companies and
460
00:31:15.600 --> 00:31:18.660
startups and enterprises that are building on the platform.
461
00:31:19.180 --> 00:31:20.940
It's obviously an interesting moment in startups
462
00:31:23.160 --> 00:31:24.520
given on the one hand,
463
00:31:25.340 --> 00:31:30.180
the ability to build products and generate revenue at seemingly
464
00:31:30.820 --> 00:31:33.320
unprecedented rates. And certainly we see this in the Stripe data,
465
00:31:33.400 --> 00:31:37.420
like the number of businesses reaching thresholds, meaningful thresholds,
466
00:31:38.000 --> 00:31:39.540
is far faster than it ever has been before.
467
00:31:41.440 --> 00:31:44.680
You are one of the most prolific and successful startup investors ever.
468
00:31:44.840 --> 00:31:46.300
You of course ran Y Combinator.
469
00:31:47.020 --> 00:31:51.860
Have the traits that make founders successful changed in this
470
00:31:51.920 --> 00:31:55.260
era or is it sort of the same thing it's always been?
471
00:31:58.440 --> 00:32:01.440
There was a time when we used to make fun of the idea guy.
472
00:32:01.600 --> 00:32:05.260
There were these people that wanted to start a company and they'd say like,
473
00:32:06.040 --> 00:32:09.280
"I have the best idea. I'm not going to tell you what it is.
474
00:32:09.360 --> 00:32:10.193
I have the best idea.
475
00:32:10.740 --> 00:32:14.720
I just need a coder to build it for me and then I'm going to be in great
476
00:32:14.780 --> 00:32:18.700
shape." And we would make fun of these people. They weren't that successful.
477
00:32:22.220 --> 00:32:26.460
And it was kind of always personally annoying to me because it would be like
478
00:32:26.500 --> 00:32:27.280
saying like,
479
00:32:27.280 --> 00:32:32.000
"I have a great idea for a song and I just need that guy with the guitar to
480
00:32:32.240 --> 00:32:36.520
make it for me." And
481
00:32:39.320 --> 00:32:42.740
so it didn't work and YC had a version of this,
482
00:32:42.800 --> 00:32:46.600
which is teams without nontechnical founders are difficult to get work.
483
00:32:47.340 --> 00:32:49.360
All of a sudden it's like the revenge of the idea guys,
484
00:32:50.320 --> 00:32:53.920
which is actually awesome for the world. I'm happy, I'm here for it for sure.
485
00:32:54.920 --> 00:32:57.880
But for a long time,
486
00:32:58.060 --> 00:33:02.600
I think the most important ingredient that I looked for,
487
00:33:02.940 --> 00:33:03.400
YC looked for,
488
00:33:03.400 --> 00:33:06.660
that kind of this part of our industry looked for on a founding team was
489
00:33:06.700 --> 00:33:10.440
technical talent and that's still very important,
490
00:33:10.520 --> 00:33:14.680
but now people who just really deeply understand their users and can't code at
491
00:33:14.740 --> 00:33:18.380
all, I want to fund those people and that's a big turnaround.
492
00:33:25.980 --> 00:33:28.460
How does one think about startup investing these days?
493
00:33:29.020 --> 00:33:33.860
Because on the one hand you have a couple of years potentially
494
00:33:34.000 --> 00:33:37.100
to AGI or ASI or the singularity-who knows what?
495
00:33:37.760 --> 00:33:42.220
And then you have investing time horizons or funds with
496
00:33:42.760 --> 00:33:47.060
10-year time horizons. How does that all fit together? Does it?
497
00:33:50.320 --> 00:33:55.120
I think to do anything at this point on a 10-year time horizon requires a
498
00:33:55.200 --> 00:33:56.320
real suspension of disbelief
499
00:33:58.100 --> 00:33:59.760
and yet that's probably the right way to live your life.
500
00:34:00.560 --> 00:34:02.120
I don't think it works to say,
501
00:34:03.380 --> 00:34:05.960
"There's this singularity in three years or five years, whatever,
502
00:34:06.040 --> 00:34:06.880
we can't see past it,
503
00:34:07.320 --> 00:34:10.040
and so we're going to do nothing or we're just going to give up or we're going
504
00:34:10.060 --> 00:34:14.300
to go crazy or whatever." You have to live as if stuff's just going to keep
505
00:34:14.340 --> 00:34:16.100
going in an understandable way for a long time.
506
00:34:18.140 --> 00:34:19.740
How far ahead does OpenAI plan?
507
00:34:24.500 --> 00:34:27.800
I mean, we sign 20-year power and land agreements.
508
00:34:28.320 --> 00:34:29.153
And for the product?
509
00:34:33.940 --> 00:34:36.940
I think we have a clear vision of what things can look like in two
510
00:34:38.740 --> 00:34:41.940
years and then it gets much hazier after that.
511
00:34:43.580 --> 00:34:47.940
So there was in the relatively recent past a narrative that
512
00:34:49.340 --> 00:34:53.920
GPT wrappers and companies of that ilk were,
513
00:34:54.520 --> 00:34:56.440
I guess as the pejorative suggests,
514
00:34:56.840 --> 00:35:00.480
undifferentiated and flimsy and be swept away by a rising tide of model
515
00:35:00.540 --> 00:35:01.373
improvements.
516
00:35:01.820 --> 00:35:04.920
Whereas now it feels like that narrative in some sense is flipping somewhat,
517
00:35:04.920 --> 00:35:07.500
where now instead of talking about wrappers,
518
00:35:07.580 --> 00:35:12.040
we talk about harnesses and harnesses are seen as having this significant heft
519
00:35:12.820 --> 00:35:13.653
and importance.
520
00:35:14.140 --> 00:35:18.140
And I guess I'm curious for your view on this and how you view
521
00:35:18.980 --> 00:35:23.260
businesses for which AI is a critical enabling component and their prospective
522
00:35:23.280 --> 00:35:24.113
durability.
523
00:35:28.500 --> 00:35:32.720
I have kind of had the same view all the way through,
524
00:35:32.840 --> 00:35:37.780
which is you as a business want to be on the side of hoping that
525
00:35:37.840 --> 00:35:38.673
AI gets smarter
526
00:35:40.080 --> 00:35:43.740
and whether-in the early model days,
527
00:35:43.800 --> 00:35:48.660
if you were the GPT wrapper and you were like patching some kind of weakness in
528
00:35:48.680 --> 00:35:50.920
the current model that was clearly going to get better with the next model,
529
00:35:51.360 --> 00:35:53.180
if the next model was much better, you were kind of sad.
530
00:35:53.600 --> 00:35:56.400
If you were doing something that got better, like you're making
531
00:35:58.360 --> 00:36:02.540
any of the wonderful services that people were building with the models that
532
00:36:02.580 --> 00:36:04.520
benefited from intelligence, you would be happier.
533
00:36:05.060 --> 00:36:06.920
I think the same thing in the world of harnesses.
534
00:36:07.240 --> 00:36:11.800
I kind of think the right way to think about this is like data center,
535
00:36:13.140 --> 00:36:13.880
model harness,
536
00:36:13.880 --> 00:36:17.280
like that whole thing is just this one cluster out of which comes as very usable
537
00:36:17.320 --> 00:36:18.153
intelligence,
538
00:36:19.040 --> 00:36:22.480
but there are so many things to go build where you're just like happier for that
539
00:36:22.520 --> 00:36:23.960
whole cluster to get better and better and better.
540
00:36:25.600 --> 00:36:28.540
And then if you're kind of secretly hoping it doesn't because you're patching
541
00:36:28.600 --> 00:36:29.440
some weakness in that,
542
00:36:29.880 --> 00:36:33.840
probably like the next model crank turn somewhere in that stack is just going to
543
00:36:33.860 --> 00:36:34.693
solve it.
544
00:36:35.220 --> 00:36:39.200
When you look at the organizations that are making the most effective use of AI
545
00:36:39.240 --> 00:36:42.980
today, I mean, you meet OpenAI customers constantly,
546
00:36:43.700 --> 00:36:44.560
large and small,
547
00:36:45.280 --> 00:36:49.760
if you think about the top three that have impressed you the most or the one
548
00:36:49.780 --> 00:36:51.060
that's impressed you the most,
549
00:36:51.480 --> 00:36:56.460
what specifically are they doing that's different? Everyone here knows, yes, AI,
550
00:36:56.580 --> 00:36:59.380
big deal, we should make enthusiastic use of it, et cetera,
551
00:36:59.760 --> 00:37:04.050
but what specifically differentiates those which in your opinion are employing
552
00:37:04.070 --> 00:37:04.903
it most effectively?
553
00:37:06.360 --> 00:37:10.590
A few different directions there. A friend of both of ours,
554
00:37:11.130 --> 00:37:14.630
Tobi Lütke of Shopify was the first CEO I knew that just said like,
555
00:37:15.570 --> 00:37:19.930
"We are going to be all-in on AI and the way we run our company." And he got,
556
00:37:20.390 --> 00:37:20.990
himself,
557
00:37:20.990 --> 00:37:25.390
got his hands dirty just building AI automation of everything and made his team
558
00:37:25.450 --> 00:37:28.320
do it and said,
559
00:37:28.390 --> 00:37:31.430
"We're just going to figure out how we take all of these things that are bad and
560
00:37:31.470 --> 00:37:35.790
make them good with AI." And it was not like a token leaderboard,
561
00:37:35.840 --> 00:37:40.210
it was not some other kind of like gamified hackable thing.
562
00:37:40.290 --> 00:37:42.590
It was just like the CEO of the company said,
563
00:37:43.070 --> 00:37:47.840
"We are now going to put AI into everything we do and I'm going to not
564
00:37:49.070 --> 00:37:50.250
be happy with you, I guess,
565
00:37:50.690 --> 00:37:52.710
or we're not going to allow it if you're not doing that.
566
00:37:52.840 --> 00:37:55.910
" So that energy has now been done by other people, but
567
00:37:57.840 --> 00:38:01.270
when the CEO of a company just says like, "We're going to automate ourselves,
568
00:38:01.430 --> 00:38:03.230
accelerate ourselves," however they phrase it internally,
569
00:38:04.320 --> 00:38:07.010
and then really holds people to it and ideally does it themselves,
570
00:38:09.820 --> 00:38:11.010
that has worked very well.
571
00:38:12.360 --> 00:38:17.250
I think we're going to try this new experiment where we start sending an FDE to
572
00:38:18.050 --> 00:38:21.650
work with like hands-on, just whatever the CEO of a company needs,
573
00:38:21.710 --> 00:38:25.350
work with them to automate their job as much of it as they can.
574
00:38:25.710 --> 00:38:30.670
And that I think will have like, if you just do it for the leader of a company,
575
00:38:30.730 --> 00:38:32.710
there's like a nice fractal effect throughout the company.
576
00:38:33.170 --> 00:38:35.890
So that works and we'll try to help companies do that.
577
00:38:38.310 --> 00:38:39.650
A second thing is being like
578
00:38:41.310 --> 00:38:44.090
uncomfortably permissive with data access.
579
00:38:45.670 --> 00:38:48.190
There's huge reasons not to do this and this is like,
580
00:38:48.630 --> 00:38:49.790
I'm stopping short of recommendation.
581
00:38:49.850 --> 00:38:53.150
You just ask what I've seen from the most effective companies.
582
00:38:55.110 --> 00:38:58.130
This is easier for small startups than companies that have a lot of sensitive
583
00:38:58.150 --> 00:38:59.050
data and a lot of
584
00:39:03.990 --> 00:39:07.230
process and compliance in place, but saying like, "You know what?
585
00:39:07.470 --> 00:39:08.650
We are going to record our meetings.
586
00:39:09.030 --> 00:39:12.690
We are going to let this AI have access to our codebase.
587
00:39:12.770 --> 00:39:15.290
We are going to let this have access to every Slack message, every email,
588
00:39:15.470 --> 00:39:16.170
everything,
589
00:39:16.170 --> 00:39:19.430
and every employee at the company is going to get to use it that way." It is
590
00:39:20.070 --> 00:39:24.870
amazing watching these two- or three-person startups and
591
00:39:25.010 --> 00:39:26.550
AI doing everything work.
592
00:39:29.450 --> 00:39:33.690
And I don't know how the world is going to decide the trade-offs on data privacy
593
00:39:33.790 --> 00:39:34.730
versus AI efficiency.
594
00:39:36.370 --> 00:39:38.590
And I think there's like some regulation that's going to have to change for
595
00:39:38.610 --> 00:39:40.290
that, but it's so powerful.
596
00:39:42.310 --> 00:39:46.350
Tempo is a new blockchain that Stripe incubated with Paradigm,
597
00:39:47.090 --> 00:39:51.810
which obviously we're partners with you guys on and
598
00:39:53.630 --> 00:39:54.690
the Mainnet just launched,
599
00:39:54.750 --> 00:39:58.830
but the project launched back last summer.
600
00:39:58.950 --> 00:40:02.310
So it's a relatively new and small team, a couple dozen people.
601
00:40:03.170 --> 00:40:06.450
The Tempo team set up
602
00:40:08.310 --> 00:40:09.143
a harness,
603
00:40:09.730 --> 00:40:14.250
a tool in their Slack installation for
604
00:40:14.390 --> 00:40:17.790
orchestrating pretty much everything at the company, everything.
605
00:40:18.110 --> 00:40:22.130
You can just ask any task, "Go and read these Google Docs,
606
00:40:23.090 --> 00:40:26.750
turn those into a bunch of linear tasks,
607
00:40:27.030 --> 00:40:30.050
then go write a pull request to implement them,
608
00:40:30.770 --> 00:40:34.830
then go deploy them and use our log analysis tool to test that the deployment
609
00:40:34.870 --> 00:40:38.190
actually worked." And the agent will happily go and employ tool use across all
610
00:40:38.210 --> 00:40:42.250
of this. And it's extremely trippy watching a whole org-a small organization,
611
00:40:42.310 --> 00:40:46.490
but an organization of people do everything in a single Slack channel.
612
00:40:46.890 --> 00:40:48.370
And I don't think that would scale to Stripe,
613
00:40:48.530 --> 00:40:50.210
but it was the first time I had the experience you're describing,
614
00:40:50.320 --> 00:40:52.770
which is it's clearly incredible for them.
615
00:40:52.830 --> 00:40:56.490
I don't quite see how to transpose it for us, but this is really something.
616
00:40:56.910 --> 00:40:57.970
It's really something to watch.
617
00:41:00.640 --> 00:41:04.290
And I find that a lot of people just can't-this is where there's a big overhang.
618
00:41:04.550 --> 00:41:07.930
They have not yet been able to wrap their heads around the fact that you can
619
00:41:07.970 --> 00:41:10.350
just kind of ask us anything and it'll probably happen.
620
00:41:12.590 --> 00:41:16.550
I myself still find myself not trusting quite enough that it's going to be
621
00:41:16.610 --> 00:41:20.130
possible. I don't know exactly how it's going to transpose to bigger companies.
622
00:41:21.210 --> 00:41:24.090
It does feel like we're missing one more abstraction there,
623
00:41:25.750 --> 00:41:27.910
how humans and AIs are going to interface at massive scale.
624
00:41:28.390 --> 00:41:31.610
The advantage that these smaller companies have is like it's just the AIs.
625
00:41:31.890 --> 00:41:34.650
They don't have to figure out the interface with all the people,
626
00:41:35.590 --> 00:41:36.423
but we'll figure it out.
627
00:41:37.790 --> 00:41:40.830
Open-source AI. Where's it going? Does it have a future?
628
00:41:41.790 --> 00:41:46.390
For sure. Right now,
629
00:41:46.870 --> 00:41:49.990
people clearly want smarter, faster,
630
00:41:50.630 --> 00:41:54.550
cheaper frontier intelligence and most of the demand is there,
631
00:41:54.870 --> 00:41:59.330
but there is also a lot of demand for open source and I expect that to increase
632
00:41:59.370 --> 00:42:00.270
relatively over time.
633
00:42:03.070 --> 00:42:06.210
So I want to talk a little bit about science because I know it's something
634
00:42:06.770 --> 00:42:07.690
you're very excited about,
635
00:42:08.530 --> 00:42:13.510
but it's also relevant to something I spend some
636
00:42:13.550 --> 00:42:18.210
of my time on, which is the Arc Institute and
637
00:42:20.490 --> 00:42:24.490
OpenAI is in fact a foundation, a nonprofit, and
638
00:42:26.930 --> 00:42:30.150
recently made a grant to the Arc Institute.
639
00:42:30.990 --> 00:42:33.850
And so maybe we'll get to the details of that in a second,
640
00:42:33.950 --> 00:42:38.430
but first you just want to speak a little bit to AI as applied to science,
641
00:42:38.710 --> 00:42:42.750
what you're seeing and just how you think about grant-making generally in the
642
00:42:42.770 --> 00:42:44.750
context of the OpenAI Foundation.
643
00:42:45.850 --> 00:42:49.930
So generally on the AI and science question,
644
00:42:50.450 --> 00:42:55.430
I hope that this will be the most important contribution of
645
00:42:56.730 --> 00:42:57.990
AI, of this technology,
646
00:42:58.090 --> 00:43:02.210
to human quality of life over time and that if we can start to discover new
647
00:43:02.270 --> 00:43:03.790
science at a much faster rate,
648
00:43:04.550 --> 00:43:08.270
which can be like new materials or cures to diseases or any number of other
649
00:43:08.330 --> 00:43:12.150
things, like I believe that to a first-order approximation,
650
00:43:12.630 --> 00:43:15.830
life gets better because we understand science better and then we figure out how
651
00:43:15.850 --> 00:43:17.770
to build stuff with it and distribute it to people.
652
00:43:20.070 --> 00:43:23.350
Starting with the models of a few months ago,
653
00:43:23.870 --> 00:43:25.440
but really now with 5.5,
654
00:43:26.150 --> 00:43:30.190
the models have gotten smart enough that excellent scientists are saying,
655
00:43:30.350 --> 00:43:34.950
"I am able to figure out better ideas." The models are able to make some
656
00:43:35.070 --> 00:43:39.890
small but important discoveries and the pace of science is going to increase.
657
00:43:40.570 --> 00:43:43.170
Eventually we'll have automated labs and robots and
658
00:43:44.730 --> 00:43:48.950
building who knows what, and we'll be able to do science much faster.
659
00:43:49.290 --> 00:43:54.190
But if we can start doing like a decade of science of what it would've
660
00:43:54.210 --> 00:43:55.570
taken us in the old world in a year,
661
00:43:56.170 --> 00:43:59.470
the compounding effect there and what we'll be able to do and discover will just
662
00:43:59.490 --> 00:44:01.250
be extremely great. So
663
00:44:03.150 --> 00:44:06.910
I think this is going to be incredible and this will be one of the big areas of
664
00:44:06.970 --> 00:44:11.830
focus of the OpenAI Foundation is basically like money and
665
00:44:11.870 --> 00:44:15.630
expertise and technology to accelerate science and trusting that will
666
00:44:17.690 --> 00:44:18.970
flow to the world in wonderful ways.
667
00:44:19.170 --> 00:44:20.370
And this is going to be a big foundation.
668
00:44:21.050 --> 00:44:25.650
Yeah, I think it's one of the biggest, maybe it's the biggest,
669
00:44:25.710 --> 00:44:27.950
I think it will be the biggest foundation in the world.
670
00:44:31.370 --> 00:44:34.150
So we're really focused on science and then AI resilience,
671
00:44:34.590 --> 00:44:38.830
like helping the world through this transition with this new technology in it.
672
00:44:41.510 --> 00:44:45.010
But we were thrilled to get to support Arc.
673
00:44:45.290 --> 00:44:49.990
I think it's clearly the best sort of AI and bio effort and
674
00:44:50.470 --> 00:44:54.410
if we can make even a small contribution to,
675
00:44:54.810 --> 00:44:57.930
with this technology and with the capital and the foundation, to helping
676
00:45:00.690 --> 00:45:02.690
make people healthier, treat diseases,
677
00:45:03.130 --> 00:45:06.370
this whole cluster of what we can do as we get better at understanding biology,
678
00:45:07.390 --> 00:45:08.330
we will be very thrilled.
679
00:45:09.150 --> 00:45:14.070
I thought that was going to take longer and looking at the incredible
680
00:45:14.110 --> 00:45:18.270
work the Arc Foundation is doing, I now think it maybe won't be that far off.
681
00:45:18.650 --> 00:45:20.810
So you know in a podcast,
682
00:45:21.430 --> 00:45:24.550
midway through you might hear a little interstitial ad.
683
00:45:26.070 --> 00:45:28.190
This is your interstitial ad for the Arc Institute.
684
00:45:31.050 --> 00:45:34.950
There's an Arc Institute booth downstairs and you might wonder why
685
00:45:34.990 --> 00:45:38.810
there's-what it's doing at the internet economy conference.
686
00:45:40.110 --> 00:45:42.510
And the context here is,
687
00:45:42.570 --> 00:45:47.290
so the Arc Institute is an organization we started four years ago and its
688
00:45:47.390 --> 00:45:51.930
goal is to produce hopefully the first cure for a
689
00:45:52.010 --> 00:45:53.390
complex disease in humans.
690
00:45:53.450 --> 00:45:57.570
So a complex disease is one that involves some genetic factors and some
691
00:45:57.590 --> 00:45:59.550
environmental factors. So you can think of most cancers,
692
00:45:59.610 --> 00:46:03.250
most autoimmune disease, most neurodegenerative disease, for example,
693
00:46:03.310 --> 00:46:05.610
as being a complex disease in this kind of specific sense.
694
00:46:06.230 --> 00:46:10.030
And humanity has never cured a complex disease, not one.
695
00:46:10.270 --> 00:46:11.970
We've cured lots of infectious diseases.
696
00:46:12.390 --> 00:46:16.330
We know how to screen for monogenic diseases for this one genetic mutation
697
00:46:17.710 --> 00:46:20.890
that undergirds it. We've never cured a complex condition.
698
00:46:23.210 --> 00:46:25.070
So we started Arc Institute, but this is the goal.
699
00:46:25.310 --> 00:46:29.610
Alzheimer's is the first complex disease that we're
700
00:46:30.350 --> 00:46:34.990
targeting and our hope is that with both new genome engineering
701
00:46:35.010 --> 00:46:39.950
technologies like CRISPR and then the amazing advances in AI that we'll be
702
00:46:39.970 --> 00:46:43.910
able to make some hopefully meaningful progress. And it's only four years old,
703
00:46:44.530 --> 00:46:46.630
but the early results are very encouraging.
704
00:46:47.730 --> 00:46:51.250
The Arc Institute is currently looking for a CTO.
705
00:46:54.570 --> 00:46:59.570
We had one CTO do a sabbatical at Arc last
706
00:46:59.630 --> 00:47:00.463
year-last year or the
707
00:47:01.330 --> 00:47:05.990
year before-it was last year and his name was Greg Brockman and he did some
708
00:47:06.030 --> 00:47:08.410
great stuff. He helped us train Evo 2,
709
00:47:09.190 --> 00:47:13.250
which is the largest biology foundation model ever trained,
710
00:47:14.470 --> 00:47:17.610
but we're looking for a full-time CTO and
711
00:47:20.310 --> 00:47:24.730
we thought that, well, perhaps that person might be in this audience,
712
00:47:24.790 --> 00:47:27.350
and if not, their friend might be in this audience.
713
00:47:27.470 --> 00:47:30.110
So if you know somebody for whom that sounds of interest,
714
00:47:30.530 --> 00:47:34.310
go check out the Arc stand downstairs and that's your interstitial ad.
715
00:47:41.240 --> 00:47:44.050
I think it's so much better that you've labeled that as the interstitial ad
716
00:47:44.070 --> 00:47:45.750
rather than just doing it. That was good.
717
00:47:47.650 --> 00:47:50.730
OK. Well, we haven't talked about Stripe. So
718
00:47:55.200 --> 00:47:57.750
you were the-I think it was second investor in Stripe.
719
00:47:57.810 --> 00:47:58.730
Was YC the first?
720
00:47:59.230 --> 00:48:02.090
You and Paul at the same time in the same kitchen in fact.
721
00:48:02.810 --> 00:48:03.950
So maybe you were first in fact.
722
00:48:04.010 --> 00:48:06.110
I don't remember in which order the checks were handed over.
723
00:48:06.170 --> 00:48:07.730
Oh yes, this was then his like Palo Alto kitchen.
724
00:48:07.920 --> 00:48:09.720
Yeah, that's right. So
725
00:48:11.890 --> 00:48:16.670
we were two pimply teenagers proposing building this financial services
726
00:48:16.710 --> 00:48:19.870
institution. It sounded like a bit of a ludicrous proposition.
727
00:48:20.090 --> 00:48:21.650
Why did you invest?
728
00:48:24.150 --> 00:48:27.090
Honestly, he didn't tee me up for this.
729
00:48:29.330 --> 00:48:32.390
You and John were two of the most-I had seen a lot of founders and you and John
730
00:48:32.430 --> 00:48:34.710
were two of the most impressive founders of any age,
731
00:48:35.190 --> 00:48:40.030
but certainly of pimply-faced-teenager-age that I had ever
732
00:48:40.090 --> 00:48:41.530
met. I clearly was right about that.
733
00:48:43.410 --> 00:48:48.270
And I had known of you and I had heard Paul
734
00:48:48.310 --> 00:48:50.390
talk about you and I had known of you on the internet
735
00:48:52.790 --> 00:48:57.610
and I was struck that you were solving a problem for yourself
736
00:48:59.710 --> 00:49:01.890
and that it was sort of like,
737
00:49:02.970 --> 00:49:05.730
it fit a trend that I thought was going to be big
738
00:49:07.610 --> 00:49:08.443
in the world.
739
00:49:10.010 --> 00:49:12.890
I kind of really believed that commerce was going to move online in a huge way
740
00:49:12.910 --> 00:49:15.290
and there were going to be lots of startups and both of those suggested that
741
00:49:15.330 --> 00:49:19.330
this could be really big. But I don't know.
742
00:49:19.490 --> 00:49:22.790
I was like a believer that if you can find really smart, and still am,
743
00:49:22.850 --> 00:49:26.210
if you can find like really smart founders and a market that's going to be big,
744
00:49:26.290 --> 00:49:28.270
you should just invest and that's kind of it.
745
00:49:32.350 --> 00:49:34.390
Well, just based on your general perspective in the world,
746
00:49:34.830 --> 00:49:38.110
but then also OpenAI's use of Stripe and what you've seen from that and like
747
00:49:38.230 --> 00:49:43.170
what you see OpenAI needing and needing in the future and building the business
748
00:49:43.210 --> 00:49:47.030
and so forth, what's your advice for Stripe on navigating
749
00:49:49.450 --> 00:49:50.283
the AI era?
750
00:49:50.650 --> 00:49:53.210
Before that, just to check my memory, the thing you were building,
751
00:49:53.650 --> 00:49:56.670
the reason you realized you needed payments was you had like built this iPhone
752
00:49:56.710 --> 00:50:00.290
app to download all of Wikipedia because you were going somewhere crazy offline
753
00:50:01.030 --> 00:50:06.010
and it was like hard for you to take payments for it and that was like-.
754
00:50:06.150 --> 00:50:06.790
Good memory.
755
00:50:06.790 --> 00:50:09.610
Yeah. Okay. I hadn't thought about that in a long time.
756
00:50:15.750 --> 00:50:19.150
I think my advice more towards like Stripe as a company itself?
757
00:50:19.310 --> 00:50:23.650
Yeah, going forward. It's a crazy time in the world. There's a lot changing,
758
00:50:24.090 --> 00:50:26.270
a lot happening. And again,
759
00:50:26.350 --> 00:50:28.650
you've seen Stripe from the perspective of a customer.
760
00:50:28.770 --> 00:50:31.290
So what are your complaints and feature requests?
761
00:50:33.410 --> 00:50:33.590
Well,
762
00:50:33.590 --> 00:50:38.270
it's more like I want to-Now that I'm thinking that we should be thinking more
763
00:50:38.290 --> 00:50:40.210
like a Stripe-style model to have a bunch of questions.
764
00:50:40.330 --> 00:50:44.210
I probably have more to learn from you than you do for me here because I
765
00:50:44.270 --> 00:50:47.330
remember when investors are outsmart themselves saying like, "Oh,
766
00:50:47.390 --> 00:50:49.250
Stripe's going to be a commodity. When everybody gets big,
767
00:50:49.370 --> 00:50:53.090
they're going to just build their own thing." And it turns out that yes,
768
00:50:53.230 --> 00:50:56.310
there are multiple payments providers and yet in practice,
769
00:50:56.790 --> 00:50:58.910
if you make a great product, if you make a great thing,
770
00:50:59.910 --> 00:51:02.200
people are still going to need to take money in the future and they're probably
771
00:51:02.260 --> 00:51:04.180
just going to stick with you if you're like a good,
772
00:51:04.260 --> 00:51:07.140
reasonable ecosystem partner and don't do crazy things.
773
00:51:08.500 --> 00:51:09.660
And I think that's going to keep working.
774
00:51:11.040 --> 00:51:15.380
I think every company clearly does need to get more efficient and with AI and
775
00:51:15.420 --> 00:51:16.253
they will do that,
776
00:51:16.560 --> 00:51:21.520
but this whole mindset that every company
777
00:51:21.580 --> 00:51:25.340
is going to go away and everything is going to be completely different,
778
00:51:26.040 --> 00:51:30.560
like maybe it'll be agents that need to handle payments between each other
779
00:51:30.840 --> 00:51:33.610
instead of consumers and merchants,
780
00:51:34.110 --> 00:51:38.200
but clearly money is going to have to move somehow and
781
00:51:39.060 --> 00:51:43.040
my advice would be like adopt AI, especially internally,
782
00:51:43.360 --> 00:51:44.560
use AI to build better products,
783
00:51:44.900 --> 00:51:49.500
but don't assume that the entire socioeconomic system completely reconfigures.
784
00:51:49.880 --> 00:51:52.640
I think the world has gotten a little bit delusional about this.
785
00:51:59.900 --> 00:52:04.160
Apart from AI as we look at over the next decade,
786
00:52:04.320 --> 00:52:07.400
just what are the technologies and areas that excite you?
787
00:52:14.400 --> 00:52:18.500
Sort of AI at the kind of model and product
788
00:52:18.700 --> 00:52:22.980
layer-I'm obsessed now with data center infrastructure.
789
00:52:24.260 --> 00:52:25.093
It's relevant.
790
00:52:25.720 --> 00:52:27.180
I think there's so much cool stuff to do there.
791
00:52:27.400 --> 00:52:31.060
I think there will be amazing new technologies like at the physical layer,
792
00:52:32.340 --> 00:52:34.040
energy, robots,
793
00:52:36.360 --> 00:52:39.980
like that stack is probably what I think about the most.
794
00:52:41.080 --> 00:52:43.740
It does seem like the world is finally making a little progress on
795
00:52:45.300 --> 00:52:49.790
brain machine interfaces. I'm excited about that. I am excited about...
796
00:52:53.160 --> 00:52:56.380
Well, I'm both afraid of and I'm excited about progress in biotech,
797
00:52:57.460 --> 00:53:01.460
but I'm certainly hopeful that that can get much better quickly and
798
00:53:03.200 --> 00:53:06.100
I think defensive biotech is about to become very important, unfortunately.
799
00:53:06.920 --> 00:53:10.280
I'm excited about new kinds of computer interfaces.
800
00:53:11.320 --> 00:53:16.060
I think we are in a insane area right now where
801
00:53:16.740 --> 00:53:21.480
we're stuck with these kind of old devices and old operating systems and we have
802
00:53:21.500 --> 00:53:23.360
this magic new enabling technology
803
00:53:25.760 --> 00:53:28.380
and it feels like Codex is amazing,
804
00:53:29.600 --> 00:53:32.680
but it feels like very broken to be telling this thing to use my computer and
805
00:53:32.700 --> 00:53:35.940
then it's like clicking around and there's all this stuff that was made for a
806
00:53:36.000 --> 00:53:38.900
person but doesn't really make sense for like an AI to go use.
807
00:53:39.080 --> 00:53:40.040
We can do so much better there.
808
00:53:40.340 --> 00:53:44.700
I think there's a whole new internet protocol to make too. So those things.
809
00:53:45.680 --> 00:53:50.080
When do you think we'll have the world's first profitable nuclear fusion
810
00:53:50.140 --> 00:53:50.973
reactor?
811
00:53:52.920 --> 00:53:55.880
Depends how far electricity prices get pushed by data center demand,
812
00:53:55.960 --> 00:53:56.860
maybe sooner than we thought.
813
00:54:02.720 --> 00:54:04.160
I'll guess in the next five years.
814
00:54:07.180 --> 00:54:08.013
It's a bold prediction.
815
00:54:08.540 --> 00:54:09.373
I hope so.
816
00:54:13.480 --> 00:54:17.300
Just going to hope we're on a roll here: hypersonic commercial air travel?
817
00:54:17.980 --> 00:54:19.040
Don't follow it as closely.
818
00:54:25.840 --> 00:54:27.220
Hypersonic meaning like Mach four?
819
00:54:27.980 --> 00:54:28.813
Sure.
820
00:54:34.540 --> 00:54:39.320
A good chunk of time. I don't know. More than 10 years.
821
00:54:39.660 --> 00:54:40.240
Okay.
822
00:54:40.240 --> 00:54:41.760
Maybe a little less, but something like that.
823
00:54:44.100 --> 00:54:47.920
Are there any domains of science technology that are not in the discourse in a
824
00:54:47.980 --> 00:54:51.780
significant way that you think will be accelerated a lot by AI and will have
825
00:54:52.040 --> 00:54:54.000
broad kind of social consequence and impact?
826
00:54:55.020 --> 00:54:55.200
Yeah.
827
00:54:55.200 --> 00:54:58.260
The one that I think just does not get enough attention is material science.
828
00:55:02.560 --> 00:55:07.320
It's not a cool thing and I think people underestimate how much of the world is
829
00:55:07.520 --> 00:55:11.600
materials, like how much of what we depend on and how much progress AI can make.
830
00:55:11.800 --> 00:55:14.700
It's such a beautifully AI-shaped problem-.
831
00:55:14.700 --> 00:55:15.600
Just getting new catalysts.
832
00:55:16.500 --> 00:55:17.333
Totally.
833
00:55:17.400 --> 00:55:22.360
That I expect very rapid progress there and that it'll impact all of our lives
834
00:55:22.380 --> 00:55:25.460
in a very positive way and it gets like very little attention.
835
00:55:27.420 --> 00:55:32.120
Last question. It feels that in some sense,
836
00:55:33.560 --> 00:55:35.860
at least some version of AI is kind of inevitable.
837
00:55:36.220 --> 00:55:39.860
There's lots of people rushing towards it and building it out and creating the
838
00:55:39.900 --> 00:55:44.660
data centers and all the rest. So there's this kind of sense of determinacy.
839
00:55:46.520 --> 00:55:51.500
How do you hope that your specific involvement changes the trajectory for
840
00:55:51.540 --> 00:55:54.840
the world relative to some other counterfactual?
841
00:55:59.680 --> 00:56:04.520
I believe in democratization and personal agency and
842
00:56:04.840 --> 00:56:07.360
access and that everybody deserves a really great life.
843
00:56:08.120 --> 00:56:13.000
The most controversial decision we made in history of OpenAI was what we
844
00:56:13.060 --> 00:56:14.100
now call iterative deployment,
845
00:56:14.240 --> 00:56:17.640
but a lot of the thinking at the time that we released ChatGPT was,
846
00:56:18.080 --> 00:56:20.940
this was insanely dangerous to do. You can't do this.
847
00:56:21.100 --> 00:56:24.860
Only this small set of people who have been thinking about AI safety can know
848
00:56:24.900 --> 00:56:25.400
what's coming.
849
00:56:25.400 --> 00:56:28.460
It's an info hazard to tell the world and it's too dangerous to ever release.
850
00:56:28.520 --> 00:56:32.000
We have to keep this locked up and in our ivory tower,
851
00:56:32.080 --> 00:56:34.840
we will discover these wonderful things and we'll share the fruits with the
852
00:56:34.880 --> 00:56:36.660
world, but we'll have the AI and we'll control it.
853
00:56:37.400 --> 00:56:42.240
And that sat very poorly with me and I thought
854
00:56:42.340 --> 00:56:42.640
then,
855
00:56:42.640 --> 00:56:47.180
and I believe now that it is extremely important that we avoid that kind of
856
00:56:47.300 --> 00:56:52.300
power concentration and that we build this for the world and the world gets to
857
00:56:52.380 --> 00:56:54.180
use it in lots of ways, some of which will be good,
858
00:56:54.240 --> 00:56:55.140
not all of which will be good,
859
00:56:56.040 --> 00:57:01.000
but that by enabling people to explore this very wide opportunity space in
860
00:57:01.040 --> 00:57:04.100
front of us, messy at times though it will be,
861
00:57:04.180 --> 00:57:06.800
and obviously we'll put guardrails on it for reasonable safety,
862
00:57:07.920 --> 00:57:10.120
we will give the world a gift, but
863
00:57:11.920 --> 00:57:15.320
the world will build a much bigger gift on top of it for all of us and that if
864
00:57:15.360 --> 00:57:18.360
you don't enable people with this technology and if you tried to keep it locked
865
00:57:18.400 --> 00:57:20.840
up, which again, now this may sound obvious,
866
00:57:20.900 --> 00:57:25.680
but this like was the sort of rough consensus plan of people working on it
867
00:57:25.700 --> 00:57:29.900
before we came along-I think that would've been really bad.
868
00:57:30.720 --> 00:57:33.880
I am a believer in entrepreneurship and innovation and
869
00:57:35.460 --> 00:57:39.180
that people are mostly good and mostly do amazing things with tools.
870
00:57:42.900 --> 00:57:46.660
I think my single biggest contribution has been, will be, whatever,
871
00:57:47.120 --> 00:57:50.240
pushing for this to be a democratized technology that people get to use and
872
00:57:50.260 --> 00:57:51.093
build on.
873
00:57:51.640 --> 00:57:53.440
Sam Altman, thank you very much.
874
00:57:53.450 --> 00:57:53.490
Thank you very much.