1
00:00:00,030 --> 00:00:05,557
The way that OpenAI is framing Astra
is basically an early step toward AGI, but

2
00:00:05,640 --> 00:00:09,427
I thought we could start a conversation
just with basically what's fundamentally

3
00:00:09,510 --> 00:00:13,350
different with Astra relative
to prior generations of model.

4
00:00:14,580 --> 00:00:16,867
It feels, first of all,
thank you for having me.

5
00:00:16,950 --> 00:00:22,860
It feels like a new step in this process
towards models that can really help us

6
00:00:23,310 --> 00:00:28,747
create value, do work, uh, discover new
science, help us create help, start new

7
00:00:28,830 --> 00:00:33,307
companies or create new products.
Using it subjectively feels to me very

8
00:00:33,390 --> 00:00:38,167
different than any model before.
What's the kind of principal use case

9
00:00:38,250 --> 00:00:41,677
that's different this time around?
You know, I think we can get into the

10
00:00:41,760 --> 00:00:43,817
generations of model that
we're very coding focused.

11
00:00:43,900 --> 00:00:47,887
I know a lot of the development was very
cybersecurity focused, but taking it

12
00:00:47,970 --> 00:00:52,717
away from the software engineer, what
is the everyday person unlocking AI that

13
00:00:52,800 --> 00:00:56,947
they weren't able to do previously?
One thing that I think people will

14
00:00:57,030 --> 00:01:01,537
immediately notice is if you have
an idea and you want to work interactively

15
00:01:01,620 --> 00:01:05,467
with an AI to get, uh, you know,
a complex piece of a whole piece of

16
00:01:05,550 --> 00:01:08,977
software built.
This is the first model to me where I

17
00:01:09,060 --> 00:01:11,137
could sort of tell someone,
like, just give it a try.

18
00:01:11,220 --> 00:01:13,207
There's. There's a good chance it'll work.

19
00:01:13,290 --> 00:01:17,407
And, you know, I've watched
people make computer games.

20
00:01:17,490 --> 00:01:23,700
I watch people do sort of like home sort
of DIY electrical engineering projects.

21
00:01:24,060 --> 00:01:28,087
Uh, I've watched people do very complex
simulations for some piece of science

22
00:01:28,170 --> 00:01:30,667
they're working on. Uh,
certainly a lot of the work where

23
00:01:30,750 --> 00:01:33,877
you would sort of normally sit down
and have to, you know, build a financial

24
00:01:33,960 --> 00:01:37,447
model here and then a PowerPoint
presentation around it and then figure

25
00:01:37,530 --> 00:01:42,007
out how to, like, make a little
interactive piece of code to sort of try

26
00:01:42,090 --> 00:01:45,330
different simulations.
That stuff is it's all so doable now.

27
00:01:45,630 --> 00:01:51,217
Um, what I hope will happen is people
I think people will be surprised at the

28
00:01:51,300 --> 00:01:54,067
beginning, but then as they build up
more trust in the model and more like,

29
00:01:54,150 --> 00:01:58,027
wow, I can really do this.
Just start throwing harder and harder

30
00:01:58,110 --> 00:02:01,987
tasks and more creative ideas at it.
And uh, they will find out what the

31
00:02:02,070 --> 00:02:06,207
model can really do for them.
So the specifics of how Astra's being

32
00:02:06,290 --> 00:02:09,207
released. I think it's
important so that a version

33
00:02:09,290 --> 00:02:14,217
of it with specific guardrails.
Um, what was the thinking behind that

34
00:02:14,300 --> 00:02:17,900
and what are those specific guardrails,
uh, in this early release of it?

35
00:02:18,380 --> 00:02:22,347
Yeah. So the the challenge
of our industry is

36
00:02:22,430 --> 00:02:25,767
that we have these models that are
getting incredibly capable and

37
00:02:25,850 --> 00:02:30,500
incredibly useful and that people want
to use to everything from, you know,

38
00:02:30,920 --> 00:02:33,870
make their lives a little
easier to starting new companies.

39
00:02:34,070 --> 00:02:37,977
And then, on the other hand, as these
models get more capable, um, the risks

40
00:02:38,060 --> 00:02:40,737
that we have to mitigate,
uh, also become more serious.

41
00:02:40,820 --> 00:02:44,307
The models could do more damage if we
don't, um, if we don't do a good job at

42
00:02:44,390 --> 00:02:47,037
that. And so we have spent, uh, you know,

43
00:02:47,120 --> 00:02:49,737
obviously this model took us a little
longer to release than we were hoping.

44
00:02:49,820 --> 00:02:53,367
I think it'll be worth the wait, but we
really wanted to spend the time on the

45
00:02:53,450 --> 00:02:55,437
safety and security
alignment of this model.

46
00:02:55,520 --> 00:02:59,277
Yes. Um, and I think
as people understand the

47
00:02:59,360 --> 00:03:02,937
power and impact and capabilities, they
will be, they will be happy that we did.

48
00:03:03,020 --> 00:03:08,427
Um, we will, uh, we'll have different
tiers of cyber access for this model,

49
00:03:08,510 --> 00:03:11,907
for cyber in particular.
Today we're rolling it out, uh, to

50
00:03:11,990 --> 00:03:15,207
Trusted Access Partners.
And then in the coming days, assuming

51
00:03:15,290 --> 00:03:17,667
everything goes well,
we'll roll it out more broadly.

52
00:03:17,750 --> 00:03:19,797
Um, you couldn't say.

53
00:03:19,880 --> 00:03:23,727
Well, you know, if this model is,
you know, has a cyber problem, why let

54
00:03:23,810 --> 00:03:27,920
anyone use it for cyber?
And I think it's very important to note,

55
00:03:28,100 --> 00:03:32,877
to note that the world is very close
to a complete change in the landscape of

56
00:03:32,960 --> 00:03:38,007
cyber attacks, and the only way that we
see for society to collectively defend

57
00:03:38,090 --> 00:03:41,667
itself against this coming wave of models,
um, from around the world and

58
00:03:41,750 --> 00:03:47,277
from other companies, is to use tools
like Astra to rapidly defend against

59
00:03:47,360 --> 00:03:52,497
these new kind of cyber threats.
Um, so we will have multiple programs,

60
00:03:52,580 --> 00:03:56,367
uh, for people with sort of different
levels of verification and trust.

61
00:03:56,450 --> 00:04:00,447
But we do think it's quite important
that the world use these models to, uh,

62
00:04:00,530 --> 00:04:04,317
collectively defend ourselves.
What happened prior to release is

63
00:04:04,400 --> 00:04:07,857
important, right?
You determined that Astra can find and

64
00:04:07,940 --> 00:04:12,317
develop zero day exploits without
any human intervention, essentially.

65
00:04:12,400 --> 00:04:17,800
And so you paused some of the work
to strengthen the safeguards.

66
00:04:18,460 --> 00:04:22,907
What specifically was it that you saw
that made you hit pause the behavior, I

67
00:04:22,990 --> 00:04:26,927
suppose, of the model.
So it's worth pointing out that, um, the

68
00:04:27,010 --> 00:04:29,087
model that we're
launching today is GPT six.

69
00:04:29,170 --> 00:04:31,127
Astra has been done training for a while.

70
00:04:31,210 --> 00:04:34,960
Um, the model that we recently talked
about pausing, uh, as a future model.

71
00:04:35,170 --> 00:04:38,830
Um, but with Astra,
we we did hit cyber critical.

72
00:04:39,130 --> 00:04:42,677
And there that required, under
our preparedness framework, a new set of

73
00:04:42,760 --> 00:04:46,217
safeguards just to be able to release
Astra, um, because of those cyber

74
00:04:46,300 --> 00:04:51,047
capabilities you talked about.
I got a lot of questions for you, Sam,

75
00:04:51,130 --> 00:04:55,307
about Astra and chain of thought.
Um, I don't know if that comes as a

76
00:04:55,390 --> 00:04:58,727
surprise to you or not, but as Astra
becomes more autonomous for future

77
00:04:58,810 --> 00:05:02,537
generations of the model,
become more autonomous, can open.

78
00:05:02,620 --> 00:05:08,087
I sort of say with confidence that you
see and understand what the model is

79
00:05:08,170 --> 00:05:12,227
planning before it acts.
In other words, you have visibility into

80
00:05:12,310 --> 00:05:16,157
the reasoning that basically
uses of the model can't see open.

81
00:05:16,240 --> 00:05:19,450
I can for obvious reasons.
Could you go into that a little bit?

82
00:05:19,570 --> 00:05:22,127
Yeah. So we have talked
for a long time about

83
00:05:22,210 --> 00:05:24,490
the importance of monitor
ability of models.

84
00:05:24,670 --> 00:05:28,067
We've also said that we believe
in defense in depth and this is only one

85
00:05:28,150 --> 00:05:30,557
part of it. There is sandboxing.

86
00:05:30,640 --> 00:05:32,940
Uh, maybe most importantly
there is alignment.

87
00:05:33,250 --> 00:05:36,587
There are a number of things that come
together to be able to make the safety

88
00:05:36,670 --> 00:05:38,987
guarantees that we want to be
able to make with a model.

89
00:05:39,070 --> 00:05:41,237
But monitor ability is an important thing.

90
00:05:41,320 --> 00:05:45,407
And we've been talking, I think,
for well over a year now about the

91
00:05:45,490 --> 00:05:49,637
importance of chain of thought
monitoring and how we have made various

92
00:05:49,720 --> 00:05:53,440
decisions there that help us preserve
chain of thought vulnerability.

93
00:05:53,770 --> 00:05:57,437
Um, even if it means we don't maximize
the capabilities we could otherwise get.

94
00:05:57,520 --> 00:06:00,677
And we think that is important.
Um, it's an important component.

95
00:06:00,760 --> 00:06:03,827
We also, uh, we want to be clear.
We don't think it's the only component

96
00:06:03,910 --> 00:06:07,697
that matters.
We live on Bloomberg Television and on

97
00:06:07,780 --> 00:06:11,710
Bloomberg Radio, and we're speaking
to Sam Altman, the CEO of OpenAI.

98
00:06:12,040 --> 00:06:16,427
Um, you know, there's a lot of bad PR in
the world around AI at the moment, every

99
00:06:16,510 --> 00:06:19,510
day citizen sort
of worry about the impact.

100
00:06:19,810 --> 00:06:21,047
One way I kind of wanted
to put that to you.

101
00:06:21,130 --> 00:06:25,037
Is it whether you can say sort
of unequivocally there, if Astra or any

102
00:06:25,120 --> 00:06:29,297
future generation of all behaved in a
way that was dangerous, that you would

103
00:06:29,380 --> 00:06:31,990
be able to detect it
and then shut it down.

104
00:06:34,510 --> 00:06:37,667
The obvious hard part of that, uh,
question is that people have different

105
00:06:37,750 --> 00:06:40,477
opinions about what model should be
able to do and shouldn't be able to do.

106
00:06:40,560 --> 00:06:43,937
Yes, we think there are some things that
very clearly model should not be able to

107
00:06:44,020 --> 00:06:47,417
do. Uh, there are some
clear red lines, and

108
00:06:47,500 --> 00:06:50,327
we talk a lot about being
able to detect those.

109
00:06:50,410 --> 00:06:55,157
But people also have strong opinions
about things within the broad bounds of

110
00:06:55,240 --> 00:06:59,027
what is possible about how they use AI
and or how they want to be able to use

111
00:06:59,110 --> 00:07:01,277
AI. And this is going to be a difficult

112
00:07:01,360 --> 00:07:06,670
question for society, because there
is not going to be agreement on, you know,

113
00:07:06,760 --> 00:07:11,747
what is and isn't acceptable use of I.
I think it's very important in the same

114
00:07:11,830 --> 00:07:16,667
way that, you know, people do things
electricity that I may not always like,

115
00:07:16,750 --> 00:07:20,227
but I think everybody has a right to
electricity that we say, you know, there

116
00:07:20,310 --> 00:07:22,897
going to be some rules for society.
You can't go like electrocute somebody

117
00:07:22,980 --> 00:07:26,197
else on the street.
But this is an important platform that

118
00:07:26,280 --> 00:07:29,167
belongs to all of us and that we
need to be able to, to use.

119
00:07:29,250 --> 00:07:33,817
So there will be, uh, ongoing
conversation, negotiation in society.

120
00:07:33,900 --> 00:07:36,397
There will, of course, be some things
that we agree on that people just should

121
00:07:36,480 --> 00:07:37,380
not do with. I.

122
00:07:37,680 --> 00:07:40,987
But there's a lot of edges that we're
now going to have to feel our way

123
00:07:41,070 --> 00:07:43,530
through. And giving
people autonomy and freedom

124
00:07:43,890 --> 00:07:48,547
to use AI is important to our mission.
Sam at the G20 this week, um, Treasury

125
00:07:48,630 --> 00:07:53,407
Secretary Scott Besson said that the eye
industry had done a terrible job of

126
00:07:53,490 --> 00:07:56,790
explaining how AI benefits
ordinary Americans.

127
00:07:57,480 --> 00:08:02,497
What's your view of that and whether you
would, on reflection, consider OpenAI to

128
00:08:02,580 --> 00:08:07,327
be in that bucket? Uh,
I think the industry has done a

129
00:08:07,410 --> 00:08:11,280
terrible job of this on the whole,
um, I think we have done

130
00:08:12,600 --> 00:08:15,727
a bad job ourselves, uh, you know,
maybe better than some others.

131
00:08:15,810 --> 00:08:19,507
Worse than some others. Um,
it is difficult, of course, to

132
00:08:19,590 --> 00:08:24,307
explain, but this is, uh, we eat the
world, in my opinion, should want us to

133
00:08:24,390 --> 00:08:27,697
worry a lot about this technology.
The world should want us to think about

134
00:08:27,780 --> 00:08:30,907
all the ways that could go wrong.
To talk about what we see coming, but

135
00:08:30,990 --> 00:08:34,957
not to sort of, like, stir up doom
or concentrate power for ourselves.

136
00:08:35,040 --> 00:08:38,167
Just because in the history of technology,
the way you solve the hard

137
00:08:38,250 --> 00:08:43,057
problems is to look at them,
talk clearly about them, have scientists

138
00:08:43,140 --> 00:08:46,440
debate them, uh, work really
hard to prevent and mitigate them.

139
00:08:46,830 --> 00:08:51,150
And that has been our intention.
But clearly, sometimes it ends up

140
00:08:51,390 --> 00:08:53,857
scaring people. Now,
there's another thing that I think

141
00:08:53,940 --> 00:08:56,647
has gone really bad.
Besides talking about the ways this has

142
00:08:56,730 --> 00:09:00,997
got could go wrong, which is, um,
I don't think we've done a great job as an

143
00:09:01,080 --> 00:09:05,497
industry of really talking
about how to empower people with

144
00:09:05,580 --> 00:09:07,357
this technology. Um, right.

145
00:09:07,440 --> 00:09:10,537
You know, you hear some people talk
about, well, we're going to give you a

146
00:09:10,620 --> 00:09:12,007
cure for cancer. And that's so great.

147
00:09:12,090 --> 00:09:14,797
So we're going to, you know, take these
risks and you're going to have to have

148
00:09:14,880 --> 00:09:17,317
less power and autonomy.
We want people to have more power and

149
00:09:17,400 --> 00:09:19,087
autonomy. You know,
I don't think a cure for

150
00:09:19,170 --> 00:09:21,187
cancer is enough. I mean,
wonderful thing to have,

151
00:09:21,270 --> 00:09:23,790
obviously, but we should
set our sights higher.

152
00:09:24,350 --> 00:09:28,827
We want these tool tools to enable
a boom in creativity and entrepreneurship,

153
00:09:28,910 --> 00:09:32,517
in people figuring out new ways to be
useful to each other, and wonderful

154
00:09:32,600 --> 00:09:35,210
products and services
to enable each other with.

155
00:09:35,540 --> 00:09:39,267
And I think that has been
lost in the messaging.

156
00:09:39,350 --> 00:09:43,707
And I also think there's some companies
that, um, have a different opinion

157
00:09:43,790 --> 00:09:46,700
there. But, but it's very,
very important to us

158
00:09:47,000 --> 00:09:50,817
that this technology is something where
people get more power and more autonomy,

159
00:09:50,900 --> 00:09:53,967
not less. And I think
people rightly get afraid

160
00:09:54,050 --> 00:09:57,300
when they feel like I leaders
might not quite want that for them.

161
00:09:58,080 --> 00:10:02,397
So that's in part where I wanted to ask
you about pricing on Astra, but also

162
00:10:02,480 --> 00:10:03,327
just generally. Right.

163
00:10:03,410 --> 00:10:07,190
I think the direction of travel
has been that you've cut prices.

164
00:10:07,520 --> 00:10:09,987
There is competition, of course,
with the other Frontier labs.

165
00:10:10,070 --> 00:10:12,867
Um, anthropic. But you know,
how much is that the

166
00:10:12,950 --> 00:10:18,207
balance of access to the technology.
You need to move quickly in releasing

167
00:10:18,290 --> 00:10:22,707
the technology and also still trying
to sort of find fair market value of how

168
00:10:22,790 --> 00:10:28,257
you price the technology we want
to create incredibly capable,

169
00:10:28,340 --> 00:10:31,047
incredibly low cost,
incredibly abundant intelligence.

170
00:10:31,130 --> 00:10:33,890
And I think Aster is a
great step forward there.

171
00:10:34,010 --> 00:10:36,920
The token efficiency of Aster
is so impressive to me.

172
00:10:37,280 --> 00:10:40,707
Um, the amount of work you can get
done with a relatively small amount of

173
00:10:40,790 --> 00:10:45,237
tokens, and the kind of the price per
dollar of a task that you might like, I

174
00:10:45,320 --> 00:10:47,967
think is just remarkable. Um,
only a few weeks ago, we cut the

175
00:10:48,050 --> 00:10:53,937
price of Luna, our small model, by 80%.
Um, uh, and our goal is to have the best

176
00:10:54,020 --> 00:10:56,847
price performance at every level
of intelligence all the way along the

177
00:10:56,930 --> 00:10:59,427
curve. Um, we will
continue to keep dropping

178
00:10:59,510 --> 00:11:02,157
prices dramatically. Um,
I'm confident in the shape of the

179
00:11:02,240 --> 00:11:04,947
technology and the work we're
doing to enable that to happen.

180
00:11:05,030 --> 00:11:10,077
And, um, you know, of course, like,
we'll make some money along the way, but

181
00:11:10,160 --> 00:11:15,297
I think the opportunity in front of us
is so huge, we will be able to continue

182
00:11:15,380 --> 00:11:18,530
to bring prices down
and still have huge revenue.

183
00:11:20,330 --> 00:11:22,730
Let's end our conversation here. Sam with

184
00:11:24,020 --> 00:11:28,257
Aster is a big step towards I actually
doing work autonomously, right?

185
00:11:28,340 --> 00:11:31,667
I've come to sort of see
the emphasis on that.

186
00:11:31,750 --> 00:11:36,467
If you staying ahead in the race that
you've just outlined requires OpenAI to

187
00:11:36,550 --> 00:11:43,277
just keep investing more heavily, maybe
than expected, to keep pushing out the

188
00:11:43,360 --> 00:11:47,010
boundaries of what the model is capable
of doing, the cadence of release.

189
00:11:47,650 --> 00:11:51,257
Where do you do you sit right now
on your considerations of doing that as a

190
00:11:51,340 --> 00:11:54,977
private company versus doing that as
a public company, where the sort of

191
00:11:55,060 --> 00:11:59,237
financial incentive and fiduciary
duty of being a public company isn't

192
00:11:59,320 --> 00:12:04,217
necessarily the same as being committed
to developing the technology, getting

193
00:12:04,300 --> 00:12:06,450
the technology released
out into the world.

194
00:12:07,780 --> 00:12:10,787
Um, if we do go public someday,
which I assume we will.

195
00:12:10,870 --> 00:12:14,597
I plan to spend a lot of time reminding
people who might want to buy or might

196
00:12:14,680 --> 00:12:18,767
not want to buy our our stock that we
are mission first and that we're also

197
00:12:18,850 --> 00:12:22,690
extremely long term oriented.
So we are going to make decisions for,

198
00:12:22,960 --> 00:12:26,207
um, what we think is best for society
as a whole with this technology.

199
00:12:26,290 --> 00:12:29,087
And we're going to make decisions
on a multi-decade timeframe.

200
00:12:29,170 --> 00:12:31,547
Um, and we think we'll do
really great as a company.

201
00:12:31,630 --> 00:12:33,780
Uh, but we're gonna
put those things first.