1
00:00:00,000 --> 00:00:03,277
Do you think you have AGI?

2
00:00:03,360 --> 00:00:06,997
I think about this a lot. I talked to I
talked to different people about where

3
00:00:07,080 --> 00:00:09,357
they believe we're at. And the best

4
00:00:09,440 --> 00:00:14,917
answer I've gotten from our team and I
it really resonates for me is that AGI

5
00:00:15,000 --> 00:00:18,917
is a gradual thing, right? It's not really
like a moment in time, like a you know

6
00:00:19,000 --> 00:00:21,717
>> So you talked about it like
a moment in time at one point.

7
00:00:21,800 --> 00:00:22,797
>> That's right. >> definitely did.

8
00:00:22,880 --> 00:00:25,317
>> That's right. That's what we used
to think, right? We we really when we set

9
00:00:25,400 --> 00:00:28,237
out, we almost thought of it as nothing
and then we'll put together the pieces

10
00:00:28,320 --> 00:00:29,797
and that's the day of AGI.

11
00:00:29,880 --> 00:00:33,077
>> Greg Brockman's point is important. AGI

12
00:00:33,160 --> 00:00:35,277
may not arrive as one dramatic

13
00:00:35,360 --> 00:00:38,237
breakthrough. Instead, we may already be

14
00:00:38,320 --> 00:00:40,877
watching it emerge gradually. As AI

15
00:00:40,960 --> 00:00:42,757
moves from answering questions to

16
00:00:42,840 --> 00:00:48,080
actually completing
complex work on our behalf.

17
00:00:50,120 --> 00:00:53,237
>> Well, I think we started late. And

18
00:00:53,320 --> 00:00:57,677
specifically the way to think about
the overall field is that for the past

19
00:00:57,760 --> 00:01:01,277
couple years the chat market was the
main growing one. And the reason for

20
00:01:01,360 --> 00:01:04,877
this was simple. GPT-4 was the first
model where the results were

21
00:01:04,960 --> 00:01:08,757
personalized enough that it made sense
for you to ask a question, read the

22
00:01:08,840 --> 00:01:10,677
answer. It's interesting to you. And so

23
00:01:10,760 --> 00:01:16,077
it very naturally lends itself to a chat.
Right? To a chat bot. And that is

24
00:01:16,160 --> 00:01:19,917
something we're able to scale, deliver
value. Now a billion people every single

25
00:01:20,000 --> 00:01:22,037
week use ChatGPT. Like a billion people.

26
00:01:22,120 --> 00:01:23,757
Like that is scale. That is real scale.

27
00:01:23,840 --> 00:01:27,757
No one else is doing that. But this year
something else happened. Is the agentic

28
00:01:27,840 --> 00:01:30,237
moment. This really was December of last

29
00:01:30,320 --> 00:01:33,637
year is when both we and competition had

30
00:01:33,720 --> 00:01:37,757
a moment where agents were really
able to first start doing software

31
00:01:37,840 --> 00:01:40,917
engineering in a significant way.
It went from saying that, you know, you

32
00:01:41,000 --> 00:01:44,117
could speed up your work with maybe 20%

33
00:01:44,200 --> 00:01:48,357
productive productivity increase to 80%
of the work could now be done with an

34
00:01:48,440 --> 00:01:53,557
AI. And so suddenly the way that software
engineers operated it inverted

35
00:01:53,640 --> 00:01:57,997
that you became much more of a manager
of agents. And since then the growth has

36
00:01:58,080 --> 00:02:00,637
been unprecedented, absolutely
unprecedented how these agents have been

37
00:02:00,720 --> 00:02:04,277
adopted. >> And this shift
is already visible in

38
00:02:04,360 --> 00:02:06,317
software development. The important

39
00:02:06,400 --> 00:02:09,477
change isn't simply that AI can write

40
00:02:09,560 --> 00:02:12,077
better code. It's that developers are

41
00:02:12,160 --> 00:02:16,277
beginning to hand entire tasks to agents

42
00:02:16,360 --> 00:02:19,157
and increasingly act as managers rather

43
00:02:19,240 --> 00:02:21,597
than doing every step themselves.

44
00:02:21,680 --> 00:02:26,437
>> I asked Codex, and this is just
off-the-shelf Codex, no special access,

45
00:02:26,520 --> 00:02:29,797
no special models.
I just use my personal Codex to say,

46
00:02:29,880 --> 00:02:33,317
"Scan gregbrockman.com.
Tell me what vulnerabilities are there."

47
00:02:33,400 --> 00:02:35,237
And gregbrockman.com is very simple.

48
00:02:35,320 --> 00:02:38,437
It's just a static site. I was like,
"Come on, what what's it going to find?"

49
00:02:38,520 --> 00:02:40,997
And it came back with like 13 findings.

50
00:02:41,080 --> 00:02:44,077
And each one of them, I was like,
"You know, I kind of knew about this, but

51
00:02:44,160 --> 00:02:47,717
it's so much work to fix it. And like,
oh yeah, I kind of forgot that I got

52
00:02:47,800 --> 00:02:51,157
that open. Or you know, here's some
domain that was still, you know, it was

53
00:02:51,240 --> 00:02:54,797
was legacy. I'm not using it anymore.
I hadn't set the records to prevent people

54
00:02:54,880 --> 00:02:57,917
from spoofing your email address."
A bunch of little things like that. That

55
00:02:58,000 --> 00:03:01,557
one seems like a big deal. Well, yeah,
you're right now that you say it. Now,

56
00:03:01,640 --> 00:03:05,997
here's the thing. Honestly,
to fix those kinds of things, you have

57
00:03:06,080 --> 00:03:08,797
to set up like DMARC to do it right. And
there's this like whole like process you

58
00:03:08,880 --> 00:03:11,277
have to do. Well, I'm just like, I don't
know, it just takes time. It takes a lot

59
00:03:11,360 --> 00:03:13,957
of expertise. So, instead I just
asked Codex, "Can you fix it?"

60
00:03:14,040 --> 00:03:17,077
>> Right. >> And so, it opened
up my Cloudflare control panel

61
00:03:17,160 --> 00:03:19,837
and just used computer use. It clicked
around. It like set a bunch of headers,

62
00:03:19,920 --> 00:03:22,477
changed a bunch of settings. I used
to have it so that I had Cloudflare in

63
00:03:22,560 --> 00:03:26,957
front and then AWS on the back end.
I was going over insecure HTTP here. Um

64
00:03:27,040 --> 00:03:30,477
Codex had figured that out. Actually,
I'm not even sure how it how it saw that

65
00:03:30,560 --> 00:03:33,477
cuz I thought everything was hidden
behind Cloudflare. It actually was just

66
00:03:33,560 --> 00:03:36,437
like, "You know what? I'm just going
to migrate you off of AWS." So, it set up

67
00:03:36,520 --> 00:03:40,917
Cloudflare pages, moved over the content.
I had an old jQuery version

68
00:03:41,000 --> 00:03:45,037
that was insecure. It just got rid
of jQuery entirely. And it was like 15

69
00:03:45,120 --> 00:03:48,757
minutes to find all these findings. It
was another hour to fix all of them. And

70
00:03:48,840 --> 00:03:52,277
for me, I just was the manager. I just
gave it the intent. I said, "What I want

71
00:03:52,360 --> 00:03:55,477
is I want to know where the problems are
and I want you to fix them." And it did

72
00:03:55,560 --> 00:03:58,597
it. And that to me is more
the future that we're promised.

73
00:03:58,680 --> 00:04:01,437
>> But perhaps the biggest acceleration

74
00:04:01,520 --> 00:04:04,517
happens when these systems are used to

75
00:04:04,600 --> 00:04:07,597
improve AI itself. Brockman says

76
00:04:07,680 --> 00:04:10,157
engineering problems that once took

77
00:04:10,240 --> 00:04:12,717
weeks can sometimes be investigated in

78
00:04:12,800 --> 00:04:18,477
hours with Codex. And that creates
potentially powerful feedback loop.

79
00:04:18,560 --> 00:04:21,197
>> Well, I would think of this as an
overall pacing, right? An overall pacing

80
00:04:21,280 --> 00:04:24,317
of of that frontier. Well,
back and another thing to think

81
00:04:24,400 --> 00:04:27,637
about is the fact that the

82
00:04:27,720 --> 00:04:32,357
our ability to produce these models
is also and to get those capability gains

83
00:04:32,440 --> 00:04:36,597
is also increasing, right? Our raw
capacity is because partly we understand

84
00:04:36,680 --> 00:04:39,437
this technology a lot better than we
did in the past. And that's actually very

85
00:04:39,520 --> 00:04:43,717
beneficial from a safety and alignment
perspective, too, right? That we really

86
00:04:43,800 --> 00:04:46,997
understand a lot of how the technology
works. There's still many many pieces

87
00:04:47,080 --> 00:04:49,957
that we don't understand, but our ability
to engineer it and to really do

88
00:04:50,040 --> 00:04:54,637
good job of of it technically,
um I think has has increased a lot. Um and

89
00:04:54,720 --> 00:04:57,957
that there are certain problems that
really used to hold us back that are

90
00:04:58,040 --> 00:05:01,437
much diminished. For example,
these big training runs, you know, they

91
00:05:01,520 --> 00:05:05,357
run for many months. And so, typically
we'd launch them from a branch of code

92
00:05:05,440 --> 00:05:08,837
and so we'd be adding little fixes
to that that are specific to that run. At

93
00:05:08,920 --> 00:05:12,637
the same time, everyone's continuing
their main line development. And every

94
00:05:12,720 --> 00:05:14,877
so often there's something on the main
line that we're like, "Oh, we really

95
00:05:14,960 --> 00:05:17,877
want that change. We really want that
fix." And then we're like, "All right,

96
00:05:17,960 --> 00:05:20,397
we got to bring these two branches of
development together." And then we merge

97
00:05:20,480 --> 00:05:23,557
them and then we launch the run and we
find that there's a bug because no one

98
00:05:23,640 --> 00:05:27,117
has actually tested all these main line
fixes at the massive scale cuz we only

99
00:05:27,200 --> 00:05:29,037
have one instance where we can do it.

100
00:05:29,120 --> 00:05:32,157
And then we spend the next three to four
weeks with like all of our best people

101
00:05:32,240 --> 00:05:35,637
scouring trying to figure out where did
the bug come from cuz you can't possibly

102
00:05:35,720 --> 00:05:40,477
read all of the code that's been written.
Like no human can possibly do that. Today,

103
00:05:40,560 --> 00:05:44,077
you're running at that same type of issue,
Codex just reads all the code, right?

104
00:05:44,160 --> 00:05:46,357
Finds the bug. It says,
"Here it is. No problem."

105
00:05:46,440 --> 00:05:50,357
And so there's something
about the partnership of

106
00:05:50,440 --> 00:05:53,437
this amplification of what the models
can do that knock out whole classes of

107
00:05:53,520 --> 00:05:56,877
problems where we know it would take
3 to 4 weeks and now takes a few hours.

108
00:05:56,960 --> 00:06:00,517
We've seen this concretely multiple times.
And so what I'm getting at is

109
00:06:00,600 --> 00:06:05,597
that there are a lot of limitations
for how we've had to develop in the past

110
00:06:05,680 --> 00:06:08,597
that have made it a lot harder for us
to move forward on problems that are

111
00:06:08,680 --> 00:06:12,517
important. And now we have
the opportunity to move forward on them

112
00:06:12,600 --> 00:06:15,517
in a much more efficient fashion. And

113
00:06:15,600 --> 00:06:20,477
that that to me is like
let's use that extra juice to

114
00:06:20,560 --> 00:06:23,957
increase the alignment properties,
to increase the safety properties. Let's

115
00:06:24,040 --> 00:06:26,597
make sure that we have models that
monitor other models in a trustworthy

116
00:06:26,680 --> 00:06:29,757
fashion. Let's really implement some
of these ideas that we were talking about

117
00:06:29,840 --> 00:06:33,437
back in 2017, right?
That we have things like

118
00:06:33,520 --> 00:06:37,997
iterated amplification, things like
debate where the models can really spend

119
00:06:38,080 --> 00:06:41,957
a bunch of compute in order to make
sure that what they're doing is actually

120
00:06:42,040 --> 00:06:44,757
reasonable. To me that's
the opportunity in front of us.

121
00:06:44,840 --> 00:06:48,037
>> And this brings us
back to the real question.

122
00:06:48,120 --> 00:06:51,317
If AI can write software, use tools,

123
00:06:51,400 --> 00:06:56,197
perform increasingly long tasks,
and even help researchers build the next

124
00:06:56,280 --> 00:07:00,557
generation of AI, at what
point do we actually start

125
00:07:00,640 --> 00:07:03,397
calling it AGI? >> And I think
what we've learned over the

126
00:07:03,480 --> 00:07:05,797
years is that was a naive perspective.

127
00:07:05,880 --> 00:07:08,677
And you can see this in
the evidence of like how

128
00:07:08,760 --> 00:07:14,317
smooth in many ways the creation
of these agents has been, how the

129
00:07:14,400 --> 00:07:16,877
transformation of software engineering
has been, now the transformation of

130
00:07:16,960 --> 00:07:20,757
knowledge work generally, how cyber
capabilities, all these things, they're

131
00:07:20,840 --> 00:07:23,757
gradually increasing. It's like one
of these things where there are moments

132
00:07:23,840 --> 00:07:25,397
where you realize what has happened.

133
00:07:25,480 --> 00:07:28,917
Hugging Face is a moment where you
realize where you are, but it's all

134
00:07:29,000 --> 00:07:31,837
gradual. Like if you
look back at kind of these

135
00:07:31,920 --> 00:07:34,157
different models that have been released
over time and their performance on

136
00:07:34,240 --> 00:07:38,997
different benchmarks, just you
know sort of that that exponential.

137
00:07:39,080 --> 00:07:43,677
And so I think that being said,
the rest of the answer is that if you go

138
00:07:43,760 --> 00:07:48,037
if we go forward two years and we look
back, I think that that we will say this

139
00:07:48,120 --> 00:07:51,197
is when AGI was created. So,
I think we are going through that moment.

140
00:07:51,280 --> 00:07:54,797
>> But, it's not it's not something
where you can just say it now.

141
00:07:54,880 --> 00:07:57,477
>> I don't think that it's ever going to
be as simple as you can say this was the

142
00:07:57,560 --> 00:08:01,397
date. There's so many axes, and I think
that maybe one lesson in general to the

143
00:08:01,480 --> 00:08:03,317
whole AGI conversation and why it's so

144
00:08:03,400 --> 00:08:05,637
complicated is because humans, I think,

145
00:08:05,720 --> 00:08:08,917
are capable of much more than

146
00:08:09,000 --> 00:08:13,797
a naive view would give credit to.
I think that the diversity of what we do,

147
00:08:13,880 --> 00:08:17,077
how we set goals, how we accomplish
goals, how tasks work, how jobs work,

148
00:08:17,160 --> 00:08:19,637
all these things extremely deep.

149
00:08:19,720 --> 00:08:21,237
And so, I think that what we're seeing

150
00:08:21,320 --> 00:08:23,997
is that there's different aspects of

151
00:08:24,080 --> 00:08:29,717
either how people accomplish tasks
or limitations that are hard for humans to

152
00:08:29,800 --> 00:08:33,077
overcome, like the thing I mentioned
with read all the code, where these AIs

153
00:08:33,160 --> 00:08:38,157
are becoming able to fill in those gaps,
are able to do more and more, and that

154
00:08:38,240 --> 00:08:42,277
humans it turns out that there are
some things that we also really find are

155
00:08:42,360 --> 00:08:45,957
extremely important for it to be a human
there. Like, where does the judgment

156
00:08:46,040 --> 00:08:49,197
come from ultimately? Where does the taste
come from? Who decides what's good

157
00:08:49,280 --> 00:08:52,637
and what's not good? Like, that is
something that should be uniquely human.

158
00:08:52,720 --> 00:08:56,677
And I'll tell you, just to give you an
example, one thing we've seen with image

159
00:08:56,760 --> 00:09:00,437
generation, when you just generate
an image with ChatGPT and you show it to

160
00:09:00,520 --> 00:09:04,197
someone, no one cares.
It is uninteresting to anyone besides you.

161
00:09:04,280 --> 00:09:09,157
But, if you take a photo of your family
and you say turn this into the

162
00:09:09,240 --> 00:09:12,557
into a caricature, turn this into,
you know, a cartoon style,

163
00:09:12,640 --> 00:09:15,597
people love it. Goes viral.
And so, there's something about that

164
00:09:15,680 --> 00:09:19,717
element of humanity that matters so
deeply. People want to connect with

165
00:09:19,800 --> 00:09:24,157
other people. And that is actually,
in in my mind, the purpose of this

166
00:09:24,240 --> 00:09:28,317
technology, the thing that I think is
is really emerging is for deepening human

167
00:09:28,400 --> 00:09:31,517
connection so you can spend more time
with your family, can spend more time

168
00:09:31,600 --> 00:09:34,437
with people you want to spend time with,
but less time, you know, typing away at

169
00:09:34,520 --> 00:09:37,037
a machine, less time,
you know, getting carpal

170
00:09:37,120 --> 00:09:39,997
tunnel, the whole thing. And so, I think

171
00:09:40,080 --> 00:09:45,237
that there is something about what
AGI is as it's starting to unfold, it's

172
00:09:45,320 --> 00:09:48,317
almost this like fog that's lifting,
it's almost teaching us more about

173
00:09:48,400 --> 00:09:50,397
ourselves and what it means to be human.

174
00:09:50,480 --> 00:09:55,717
>> OpenAI's vision also appears to be
moving beyond one universal definition

175
00:09:55,800 --> 00:10:01,157
of AI. For an individual user, the more
practical milestone may be something

176
00:10:01,240 --> 00:10:05,237
much simpler.
An AI that knows your context, has

177
00:10:05,320 --> 00:10:09,717
access to your tools, and can
actually accomplish things for you.

178
00:10:09,800 --> 00:10:14,597
>> What people do see is how chat is
evolving, how Codex is evolving, and one

179
00:10:14,680 --> 00:10:17,797
of your big priorities in this role,
as I understand it, is what you guys call

180
00:10:17,880 --> 00:10:22,237
the merge internally, which is creating
this super app, this combined chat GPT

181
00:10:22,320 --> 00:10:24,757
and Codex. Um and you've got chat GPT

182
00:10:24,840 --> 00:10:27,157
for work now and the and the tabs.

183
00:10:27,240 --> 00:10:31,157
Um but at the same time, like it doesn't
feel like that's the end state, you

184
00:10:31,240 --> 00:10:35,077
know? And I'm a user of it and I
see realness in how it's uh currently

185
00:10:35,160 --> 00:10:38,477
executed. Some things don't sync.
You can tell you're still combining tech

186
00:10:38,560 --> 00:10:42,837
stacks. And I'm curious once you're done
with that work, what is the end product

187
00:10:42,920 --> 00:10:46,277
for people that you imagine?
>> A personal AGI.

188
00:10:46,360 --> 00:10:49,477
>> Personal AGI. And what How
does that manifest as software?

189
00:10:49,560 --> 00:10:53,637
>> Yeah, it's it's it's
a really really exciting

190
00:10:53,720 --> 00:10:57,797
vision in my mind, right? Because where
we started with chat GPT, it was a text

191
00:10:57,880 --> 00:11:00,757
box, you could ask anything.
But if you think about it, it was

192
00:11:00,840 --> 00:11:05,037
limited in a very fundamental way.
The only thing it could do is output some

193
00:11:05,120 --> 00:11:08,557
text for you. The only impact it could
have in the world was, you know, kind of

194
00:11:08,640 --> 00:11:13,517
in your brain, right? That it wasn't able
to look at your calendar, it wasn't able

195
00:11:13,600 --> 00:11:15,197
to look at your email.

196
00:11:15,280 --> 00:11:18,077
And all of that is now changed.

197
00:11:18,160 --> 00:11:22,237
Chat GPT work is really, by the way,
a Codex harnessed in the cloud, right? So,

198
00:11:22,320 --> 00:11:26,837
we had chat that was this, you know,
question-answer system. Um iterate

199
00:11:26,920 --> 00:11:29,957
turns, you know, people actually then
go beyond just, you know, their questions

200
00:11:30,040 --> 00:11:34,037
to, you know, health, like all these kinds
of queries they're able to help them.

201
00:11:34,120 --> 00:11:37,517
We had Codex, which was an agentic
system that can go do work.

202
00:11:37,600 --> 00:11:41,677
And so obvious it's like chat has
to become agentic. And so the place we

203
00:11:41,760 --> 00:11:44,277
started at the beginning of the year was
to say, well, let's let's upgrade chat.

204
00:11:44,360 --> 00:11:48,797
Let's, you know, make it also do agent
stuff. And it just turned out to be so

205
00:11:48,880 --> 00:11:52,517
hard to execute because now we have
two agentic stacks and then we compare

206
00:11:52,600 --> 00:11:54,917
between one and the other and there'll
be something that will work in this one.

207
00:11:55,000 --> 00:11:58,197
We're like, why isn't it working in that
one? And it was just so slow. And so the

208
00:11:58,280 --> 00:12:01,557
decision we made instead is say,
let's take the agentic stack we have, the

209
00:12:01,640 --> 00:12:04,437
Codex stack, and bring it to ChatGPT.

210
00:12:04,520 --> 00:12:07,997
And so what you see with chat work
is exactly right. This is an intermediate

211
00:12:08,080 --> 00:12:11,877
state. In some ways it's kind
of funny if you think about

212
00:12:11,960 --> 00:12:16,637
it. ChatGPT was a text box. ChatGPT work
as a text box. But this text box is way

213
00:12:16,720 --> 00:12:18,677
more awesome. It way more things.

214
00:12:18,760 --> 00:12:21,997
Exactly. And so where we
want to be is you should

215
00:12:22,080 --> 00:12:24,357
it should be simple. User should not
have to think about this. You shouldn't

216
00:12:24,440 --> 00:12:27,517
have to think about tabs or switching
and shouldn't have model pickers. You

217
00:12:27,600 --> 00:12:30,597
shouldn't have thinking levels.
You should have an AGI. You should be able

218
00:12:30,680 --> 00:12:33,797
to ask it. And in fact, you shouldn't
even have to ask. It should proactively

219
00:12:33,880 --> 00:12:38,677
say, hey, I noticed that your favorite
band, cuz it knows you well, is in town.

220
00:12:38,760 --> 00:12:42,277
Tickets went on sale. I got you
the tickets exactly where you want for you

221
00:12:42,360 --> 00:12:46,077
and your family. And you know, I only
had 10 minutes to do it cuz everything

222
00:12:46,160 --> 00:12:48,357
is going to go so fast. So I just
purchased it proactively. Hope that's

223
00:12:48,440 --> 00:12:51,957
okay. And if you've built enough
trust with the agent so you've given it

224
00:12:52,040 --> 00:12:54,317
permission to do that and it kind
of knows that well, these tickets are

225
00:12:54,400 --> 00:12:56,797
within your your price range, within
your budget. These other ones, oh no,

226
00:12:56,880 --> 00:12:58,797
no, I I got to get permission for that.

227
00:12:58,880 --> 00:13:03,797
Right? That's the kind of system that we
want. So trust is the keyword. EQ also

228
00:13:03,880 --> 00:13:06,797
very important that if you've shared
some information, like you've talked

229
00:13:06,880 --> 00:13:10,117
about a health condition, maybe that's
okay information for the AI to talk

230
00:13:10,200 --> 00:13:11,677
about with a parent, but maybe it's not

231
00:13:11,760 --> 00:13:13,757
okay to talk about with a stranger.

232
00:13:13,840 --> 00:13:19,117
Right? All of that judgment and and also
observability and really helping you

233
00:13:19,200 --> 00:13:23,277
build up that trust relationship
with the AI so that you know what

234
00:13:23,360 --> 00:13:26,437
you know, when it's operating on your
behalf, it'll pursue your goals. That's

235
00:13:26,520 --> 00:13:29,597
very, very critical.
But the same thing, same kind of

236
00:13:29,680 --> 00:13:33,757
technology, the underlying AGI stack,
one stack, one way of doing connectors,

237
00:13:33,840 --> 00:13:37,477
so it's able to hook up to this context,
also applies to the enterprise. And

238
00:13:37,560 --> 00:13:40,477
we're really seeing it not just
for software engineering, which by the way

239
00:13:40,560 --> 00:13:42,237
has been totally transformed this year.

240
00:13:42,320 --> 00:13:45,037
And to the point of AGI being a gradual
thing, or a thing where we just

241
00:13:45,120 --> 00:13:49,117
constantly update our expectations, like
we all kind of forget what software was

242
00:13:49,200 --> 00:13:53,077
1 year ago. Totally different. A very
interesting thing within OpenAI is that

243
00:13:53,160 --> 00:13:56,637
we're at basically 100% adoption
of ChatGPT work across every function,

244
00:13:56,720 --> 00:14:00,597
right? Sales, finance, everyone.
And whenever, you know, we roll out a new

245
00:14:00,680 --> 00:14:03,557
model and it goes away, or, you know,
the system breaks for whatever internal

246
00:14:03,640 --> 00:14:07,077
reason, um everyone's
so upset. I cannot do my

247
00:14:07,160 --> 00:14:10,077
work. And you're like, I don't get it.
6 months ago you didn't have this and you

248
00:14:10,160 --> 00:14:12,997
were you were working just fine.
Like, why are you, you know, why are you

249
00:14:13,080 --> 00:14:16,677
complaining? And you realize people
adapt so quickly and learn how to get

250
00:14:16,760 --> 00:14:20,757
the most out of these tools. And once
you're there, you also just sort of

251
00:14:20,840 --> 00:14:22,237
forget that there was any other way.

252
00:14:22,320 --> 00:14:26,397
>> And the implications of that are
that what the coding AI tools have done to

253
00:14:26,480 --> 00:14:31,397
engineering, this could do to all
knowledge work, all kinds of fields. And

254
00:14:31,480 --> 00:14:34,117
that's how you guys used
to talk about AGI is like,

255
00:14:34,200 --> 00:14:37,877
you know, it was basically AI that could
do most economically viable work better

256
00:14:37,960 --> 00:14:40,917
than humans. And what you're describing
now in the product kind of sounds like

257
00:14:41,000 --> 00:14:43,357
that. >> I think that
the amplification we've

258
00:14:43,440 --> 00:14:46,597
seen in software engineering we expect
across every field for sure. And there

259
00:14:46,680 --> 00:14:51,317
is something that I think that we
is also worth really thinking about,

260
00:14:51,400 --> 00:14:55,717
which is the software engineers,
rather than having more vacation time

261
00:14:55,800 --> 00:14:59,397
and having more chill job, they're
all working harder than ever before.

262
00:14:59,480 --> 00:15:00,917
>> Yeah, you guys work
a lot of hours here.

263
00:15:01,000 --> 00:15:04,717
>> Yes. I think across the whole industry
people are talking about that. Because

264
00:15:04,800 --> 00:15:07,757
now you feel like if
your agents aren't running,

265
00:15:07,840 --> 00:15:09,877
it's just wasted opportunity. Right?

266
00:15:09,960 --> 00:15:13,437
It's like, and I know that feeling
because we have spent years of we have

267
00:15:13,520 --> 00:15:16,317
these GPU clusters, you're either
running something on them or you've lost

268
00:15:16,400 --> 00:15:19,637
the opportunity. You can't get that time
back. It's not like the GPU hours that

269
00:15:19,720 --> 00:15:22,917
you're not using while you're asleep
overnight, you can bank them up and use

270
00:15:23,000 --> 00:15:27,637
them later. It's use it or lose it.
And that is very much the opportunity that

271
00:15:27,720 --> 00:15:32,037
everyone has in front of them is you have
almost an AGI, maybe soon truly an AGI

272
00:15:32,120 --> 00:15:35,077
in your pocket. Can do anything you want.

273
00:15:35,160 --> 00:15:37,237
What is it you want? Waiting on you.

274
00:15:37,320 --> 00:15:39,357
It's waiting on your direction,
your guidance, your management.

275
00:15:39,440 --> 00:15:42,397
>> It would be nice if eventually
we can get some time back from AI.

276
00:15:42,480 --> 00:15:46,117
>> That's right. >> Like right now, I do
feel like it it will help you work more.

277
00:15:46,200 --> 00:15:49,957
>> That's right. And so I think that I
think that that is that is exactly a

278
00:15:50,040 --> 00:15:53,757
next step, right? It's really moving
from this like you're micromanaging this

279
00:15:53,840 --> 00:15:56,437
agent. You have to provide it with these
tasks and you have to give the context

280
00:15:56,520 --> 00:15:59,797
cuz it just doesn't have it.
Exactly. You want something that has

281
00:15:59,880 --> 00:16:03,837
much more like you think about like
a really good assistant who really is

282
00:16:03,920 --> 00:16:07,237
plugged in and you've given it again,
you have to build up trust with it,

283
00:16:07,320 --> 00:16:09,917
right? You have to really say like,
okay, I trust you with this context, I

284
00:16:10,000 --> 00:16:13,717
trust you with this responsibility,
I trust you with this job, whatever it is.

285
00:16:13,800 --> 00:16:17,157
And once it's there, it should be able
to proactively run and really execute.

286
00:16:17,240 --> 00:16:20,517
You wake up in the morning, you know,
it could say, "Hey, here are five things I

287
00:16:20,600 --> 00:16:23,757
need from you to unblock me. You know,
I just need approval on this. Like here's

288
00:16:23,840 --> 00:16:26,197
a question I have." Those kinds of things.
As you drink your morning

289
00:16:26,280 --> 00:16:28,357
coffee, you say yes, no, yes, no.

290
00:16:28,440 --> 00:16:31,717
Then you go, come back, you know, late
at night, and again you've got five more

291
00:16:31,800 --> 00:16:35,317
things. And so I think that it's going
to be a world where it's going to be

292
00:16:35,400 --> 00:16:38,917
different from I think the kind
of AGI world that many of us would have

293
00:16:39,000 --> 00:16:43,277
pictured 5 10 years ago.
>> Software development may simply be the

294
00:16:43,360 --> 00:16:45,837
first profession showing us what this

295
00:16:45,920 --> 00:16:47,757
transition looks like.

296
00:16:47,840 --> 00:16:50,597
If similar agents move into this area,

297
00:16:50,680 --> 00:16:53,277
finance, marketing, design, and

298
00:16:53,360 --> 00:16:55,557
everyday's business world, the biggest

299
00:16:55,640 --> 00:17:00,917
question may no longer be what AI can do,
but what humans choose to ask it to

300
00:17:01,000 --> 00:17:02,197
do.

301
00:17:02,280 --> 00:17:05,517
So perhaps AGI won't arrive with an

302
00:17:05,600 --> 00:17:08,157
announcement. We may simply look back

303
00:17:08,240 --> 00:17:13,360
and realize that the transition
had already begun.

304
00:17:20,840 --> 00:17:22,880
>> Mhm.