1
00:00:03,501 --> 00:00:04,321
Hey, how's it going?

2
00:00:04,661 --> 00:00:05,381
I'm Will Brown.

3
00:00:05,621 --> 00:00:07,881
I lead applied research at Prime Intellect,

4
00:00:07,921 --> 00:00:10,601
and I hope you are all having a great AI engineer conference.

5
00:00:10,721 --> 00:00:11,681
This is always one of my favorites.

6
00:00:11,981 --> 00:00:14,901
It's always just so cool to see what everyone's working on and building,

7
00:00:14,941 --> 00:00:18,981
and I feel like it is a place where people get to talk about what the future looks like,

8
00:00:18,981 --> 00:00:23,301
and start seeing the trends in the industry

9
00:00:23,301 --> 00:00:26,221
and seeing the progress and what works and what doesn't.

10
00:00:26,801 --> 00:00:30,361
And that's kind of the theme I want to think about for this talk today

11
00:00:30,361 --> 00:00:34,581
is what comes next from what we're doing right now

12
00:00:34,581 --> 00:00:37,621
and what are some of the things that we've been working on

13
00:00:37,621 --> 00:00:40,021
and also just been seeing in research that's unfolding

14
00:00:40,021 --> 00:00:44,521
that kind of points at what we think the future of AI might look like

15
00:00:45,121 --> 00:00:46,501
at the organization level.

16
00:00:46,501 --> 00:00:56,141
We have seen a lot of very cool progress in agents and interfaces and bots and things

17
00:00:56,141 --> 00:00:58,881
like that over the past several months, years.

18
00:01:00,181 --> 00:01:02,021
But there's always more coming.

19
00:01:02,161 --> 00:01:03,161
Things are moving so fast.

20
00:01:03,321 --> 00:01:10,701
I think it's a very exciting time to just imagine and speculate and plan for what comes

21
00:01:10,701 --> 00:01:10,981
next.

22
00:01:11,721 --> 00:01:17,341
I don't know if any of you remember back in the early chat GPT days, open AI had kind

23
00:01:17,341 --> 00:01:24,261
of spelled out their framing of what are the different stages of AI progress on the road

24
00:01:24,261 --> 00:01:25,241
to AGI.

25
00:01:26,621 --> 00:01:31,441
And I think we moved along this faster than we thought we would.

26
00:01:32,221 --> 00:01:38,881
We like, I remember 18 months ago I was here in New York at AI engineer giving a talk where

27
00:01:38,881 --> 00:01:43,901
where we had gotten reasoners and chatbots had been around for a while and reasoners

28
00:01:43,901 --> 00:01:46,901
were the new thing, but we didn't really have agents yet.

29
00:01:47,441 --> 00:01:54,581
And now we've climbed pretty quickly through agents and innovators, and the question next

30
00:01:54,581 --> 00:01:57,321
seems to be, okay, what happens for organizations?

31
00:01:58,401 --> 00:02:05,601
does it mean for an organization to truly be AI native, not just as a startup that uses

32
00:02:05,601 --> 00:02:10,201
a lot of agents and builds everything with Claude Code, but at the large enterprise scale,

33
00:02:11,561 --> 00:02:17,481
do companies of the future look like that involve deep integration of agents throughout

34
00:02:17,481 --> 00:02:19,161
the life cycle of the organisation?

35
00:02:19,861 --> 00:02:22,661
And what do we need to build on the path there?

36
00:02:23,981 --> 00:02:26,621
So a bit about Prime Intellect, just for some backdrop.

37
00:02:27,281 --> 00:02:33,621
We build a lot of different things around AI infrastructure for training, deploying,

38
00:02:33,621 --> 00:02:38,641
and continuously improving models, and especially open source models and open source tooling.

39
00:02:38,921 --> 00:02:44,101
And so we do a lot of things, and we roughly do the sorts of things that you would do inside

40
00:02:44,101 --> 00:02:47,021
of a big lab, and these are our products.

41
00:02:47,301 --> 00:02:52,721
And the reason we do this is because we are imagining a future that is coming pretty quickly,

42
00:02:52,801 --> 00:02:59,001
it seems, where open source models are readily available to train and deploy on your own

43
00:02:59,001 --> 00:03:03,441
workflows, on your own data, in your own harnesses, and can be the sorts of things that organizations

44
00:03:03,441 --> 00:03:07,361
organizations can have grow with them over time that allow compounding their institutional

45
00:03:07,361 --> 00:03:10,981
knowledge, their institutional advantage from their workflows, from their data, from their

46
00:03:10,981 --> 00:03:16,421
people and their processes in addition to using the models from the frontier labs.

47
00:03:16,821 --> 00:03:21,621
This talk will be a bit about how we see that evolving and unfolding, where what is the

48
00:03:21,621 --> 00:03:27,021
role of training in an AI native enterprise, how does this coexist with the frontier models,

49
00:03:27,021 --> 00:03:31,621
and what do we think is going to be the version of this

50
00:03:31,621 --> 00:03:34,061
that really unlocks the most value for people

51
00:03:34,061 --> 00:03:35,181
and what does it look like?

52
00:03:36,521 --> 00:03:38,561
And so organizations have lots of knowledge

53
00:03:38,561 --> 00:03:40,541
and the longer organizations have been around

54
00:03:40,541 --> 00:03:42,741
and the more established it is,

55
00:03:42,881 --> 00:03:44,841
the more implicit knowledge there is

56
00:03:44,841 --> 00:03:46,501
that kind of doesn't ever get written down.

57
00:03:47,361 --> 00:03:48,841
There are things that do get written down.

58
00:03:48,981 --> 00:03:51,881
There's code bases, there's emails, there's processes.

59
00:03:52,141 --> 00:03:53,261
Some of these are very explicit.

60
00:03:53,461 --> 00:03:55,441
There are rules of you must do this,

61
00:03:55,501 --> 00:03:56,581
this thing works like that.

62
00:03:57,021 --> 00:04:02,301
But a lot of things live in people's heads and they live in between the lines of emails,

63
00:04:02,741 --> 00:04:05,661
of the meeting notes, of the kind of who talks to who.

64
00:04:05,781 --> 00:04:11,601
A lot of companies have an on paper org chart and then a real org chart of who actually makes

65
00:04:11,601 --> 00:04:16,481
decisions and how does change get adopted throughout the company, what sorts of mandates

66
00:04:16,481 --> 00:04:17,701
are likely to take effect.

67
00:04:18,721 --> 00:04:22,861
People have built up this knowledge from being at companies for a long time of, oh, this

68
00:04:22,861 --> 00:04:23,641
is kind of how things work.

69
00:04:23,841 --> 00:04:25,461
This is where the friction points are.

70
00:04:25,461 --> 00:04:35,021
And I think in many cases, and I think this is really one of the big opportunities for AI at the organizational level is, is that many individual people are kind of not

71
00:04:35,021 --> 00:04:42,761
in a position where they can change the things that are broken in an organization, because no one is actually spelling it out.

72
00:04:42,941 --> 00:04:57,461
No one is, like, taking a bird's-eye view of looking at everything that happens in an organization and saying, hey, this over here is inefficient, this could be better, this thing keeps happening across all these different teams, but nobody's figured out that the solution has actually been discovered by that team.

73
00:04:57,461 --> 00:05:00,541
And these are the sorts of things that agents are great for.

74
00:05:00,981 --> 00:05:06,781
Agents are very capable at combing through large amounts of textual data that's unstructured

75
00:05:06,781 --> 00:05:12,321
and finding these kinds of patterns and turning them into something useful and surfacing them

76
00:05:12,321 --> 00:05:18,221
as, okay, this workflow has a problem because it is wasting time and it's inefficient and

77
00:05:18,221 --> 00:05:19,161
it could be different instead.

78
00:05:20,121 --> 00:05:21,861
And we don't want to just do this once, though.

79
00:05:21,861 --> 00:05:29,341
We want this to become the way that organizations adopt AI broadly, where agents are able to

80
00:05:29,341 --> 00:05:33,261
get the context they need, but also they're able to improve both the agents themselves

81
00:05:33,261 --> 00:05:34,501
as well as the organizations.

82
00:05:35,201 --> 00:05:41,361
We want agents to be more like our coworkers, but human coworkers learn on the job.

83
00:05:41,921 --> 00:05:47,361
Agents currently learn on the job in relatively primitive ways, and they do this via in-context

84
00:05:47,361 --> 00:05:47,641
learning.

85
00:05:48,321 --> 00:05:49,861
and, oops, sorry.

86
00:05:55,141 --> 00:05:56,881
I'm going to go back a few seconds

87
00:05:56,881 --> 00:05:59,161
so we can edit this on the recording.

88
00:06:00,401 --> 00:06:02,961
Agents learn in relatively primitive ways

89
00:06:02,961 --> 00:06:04,461
through, like, in-context learning,

90
00:06:04,521 --> 00:06:06,821
and this typically looks like them

91
00:06:06,821 --> 00:06:09,121
kind of having a skill file written by somebody

92
00:06:09,121 --> 00:06:10,821
or, like, looking around a code base,

93
00:06:10,901 --> 00:06:13,281
but they don't get the institutional-level

94
00:06:13,281 --> 00:06:16,521
latent knowledge that human employees really do.

95
00:06:17,741 --> 00:06:21,721
They relearn a lot, they have to rediscover things, they spend a lot of input tokens.

96
00:06:21,881 --> 00:06:27,181
I think nowadays people realize that cached input is the bottleneck in many agent costs.

97
00:06:28,061 --> 00:06:32,841
And even if you do get great cache rates and cache costs go down, you still have to reprocess

98
00:06:32,841 --> 00:06:34,901
a lot of input to redo a workflow every single time.

99
00:06:35,941 --> 00:06:41,081
And so to get around this, you kind of want to bake in more into the model, especially

100
00:06:41,081 --> 00:06:43,101
for things that are harder to write down as rules.

101
00:06:43,101 --> 00:06:49,221
I like thinking about the spectrum from rules to patterns, as well as conceptually code

102
00:06:49,221 --> 00:06:49,901
versus data.

103
00:06:50,501 --> 00:06:55,201
And so lots of things live as code, and skills are kind of like this, rules are kind of like

104
00:06:55,201 --> 00:07:00,321
this, but things like preferences and friction, heuristics and trends, these are all data.

105
00:07:00,841 --> 00:07:07,141
And institutions have so much data that they are running on, and the employees of the organization

106
00:07:07,141 --> 00:07:10,881
implicitly learn that models never really get a chance to learn.

107
00:07:10,881 --> 00:07:15,341
And this is the sort of thing that we see as blocking them from becoming true coworkers.

108
00:07:16,461 --> 00:07:19,841
And so there is a question of what goes in context, what goes in the weights.

109
00:07:20,201 --> 00:07:21,961
And part of it is you kind of need to earn it.

110
00:07:22,301 --> 00:07:24,481
Things don't just go in the weights right away.

111
00:07:24,841 --> 00:07:31,181
But if you see every new generation of models, the models get better not because the prompt

112
00:07:31,181 --> 00:07:33,041
is better, but because they train them more.

113
00:07:33,301 --> 00:07:34,281
Training is useful.

114
00:07:34,481 --> 00:07:35,021
Training works.

115
00:07:35,181 --> 00:07:37,481
Training is why we see the progress we do.

116
00:07:37,481 --> 00:07:41,401
and the model is roughly trained on lots of data

117
00:07:41,401 --> 00:07:44,001
that is aiming to approximate things

118
00:07:44,001 --> 00:07:47,201
that the labs expect will be high-value,

119
00:07:47,301 --> 00:07:49,041
repeatable workflows at organizations.

120
00:07:49,941 --> 00:07:51,401
And I like this.

121
00:07:51,601 --> 00:07:53,821
I think there's a process that currently it works.

122
00:07:54,001 --> 00:07:54,801
It's how the models get better.

123
00:07:54,981 --> 00:07:57,601
But if you spell it out, it's relatively inefficient.

124
00:07:57,861 --> 00:08:00,061
So the way that a model would learn how to get really good

125
00:08:00,061 --> 00:08:04,341
at using some combination of Salesforce tools, for example,

126
00:08:04,341 --> 00:08:10,001
Let's say someone who knows how those tools get used has to be looking for work.

127
00:08:11,221 --> 00:08:14,981
And maybe they previously worked at a place that were using these tools and now they are

128
00:08:15,521 --> 00:08:18,701
browsing expert network sites to make money in their free time.

129
00:08:19,241 --> 00:08:23,181
And they are sharing their knowledge to someone who works at an expert network company.

130
00:08:23,461 --> 00:08:29,561
This person at this network company will then take this data from the human and pass it

131
00:08:29,561 --> 00:08:31,981
to someone else whose job is to create RL environments.

132
00:08:31,981 --> 00:08:37,021
This RL environment might get sold then by someone who doesn't have this expert knowledge

133
00:08:37,021 --> 00:08:41,801
to a researcher at a lab who also really isn't the expert in the domain, who then will be

134
00:08:41,801 --> 00:08:48,201
refining it to put it into a training pipeline to go into the next frontier model run to

135
00:08:48,201 --> 00:08:53,661
make a model that has kind of through this game of telephone gotten some knowledge of

136
00:08:53,661 --> 00:08:55,001
that human originally.

137
00:08:55,321 --> 00:08:58,641
But there's so many steps in this pipeline and it's a relatively secret pipeline.

138
00:08:58,641 --> 00:09:04,801
People talk across two stages of it, but the human who knows this thing is not supervising

139
00:09:04,801 --> 00:09:09,241
the training run, is not directly inputting how the model gets trained in the way that

140
00:09:09,241 --> 00:09:12,461
a human could if they were doing it for their coding agent that they're working with day

141
00:09:12,461 --> 00:09:12,721
to day.

142
00:09:13,701 --> 00:09:19,261
And so what we see as an opportunity, and also where we would hope to go, is to close

143
00:09:19,261 --> 00:09:23,281
this gap and to allow the person who, the same way that when we work with our coding

144
00:09:23,281 --> 00:09:27,361
agents, we're prompting them the same things every time, telling them to not repeat the

145
00:09:27,361 --> 00:09:27,801
same mistakes.

146
00:09:29,301 --> 00:09:30,701
This is how you could train models.

147
00:09:31,221 --> 00:09:37,561
You can tell the models things and they will learn these things durably, not just as a

148
00:09:37,561 --> 00:09:41,221
line in your agent's MD which accumulates and you have to refactor it all the time.

149
00:09:41,561 --> 00:09:45,461
You can do some of that, but a lot of times what you really want is you're giving contextual

150
00:09:45,461 --> 00:09:45,941
preference.

151
00:09:46,141 --> 00:09:50,461
In this setting, this thing that you did was not exactly what I wanted and I would have

152
00:09:50,461 --> 00:09:51,321
preferred this instead.

153
00:09:51,601 --> 00:09:55,541
And we want systems that can compound this into durable learning.

154
00:09:55,541 --> 00:09:59,021
And training is a very effective way of doing this, because the shape of data you're giving

155
00:09:59,021 --> 00:10:04,821
there is quite close to what you would want from an actual training pipeline.

156
00:10:08,761 --> 00:10:10,581
And so there's several different types of learning.

157
00:10:10,781 --> 00:10:13,741
There's things that people do, there's things that you can do at the harness layer with

158
00:10:13,741 --> 00:10:14,701
skills and memory.

159
00:10:15,161 --> 00:10:19,361
There's knowledge that just lives throughout an organization in file systems and databases

160
00:10:19,821 --> 00:10:23,321
that you retrieve with tools and with code.

161
00:10:23,321 --> 00:10:27,201
Then there's the models themselves which have this really broad world knowledge.

162
00:10:27,381 --> 00:10:34,161
But we all have, I'm sure, experienced frustration when using a model whose cut off of knowledge

163
00:10:34,161 --> 00:10:35,661
is a while back.

164
00:10:36,581 --> 00:10:39,981
You can't really talk to the models about the news of the past six months nowadays.

165
00:10:40,241 --> 00:10:45,421
They have to go relearn a world model in context every single time.

166
00:10:45,821 --> 00:10:50,661
This does not allow them to really operate as fluently as they can discuss things that

167
00:10:50,661 --> 00:10:51,821
happened longer ago.

168
00:10:51,821 --> 00:10:56,901
So, organizations, to get the most value out of models, we imagine they will want to increasingly

169
00:10:57,521 --> 00:11:02,381
have models that really understand what is going on in the organization at an intuitive

170
00:11:02,381 --> 00:11:02,761
level.

171
00:11:03,081 --> 00:11:05,001
This is the sort of thing that you want to train for.

172
00:11:06,621 --> 00:11:07,661
So how do you get there?

173
00:11:07,961 --> 00:11:14,941
What is the road for an organization to go from having this idea in the abstract of we're

174
00:11:14,941 --> 00:11:17,121
going to train models on our data to actually getting there?

175
00:11:17,121 --> 00:11:21,681
And I think one really useful starting point is evals as essentially test cases.

176
00:11:21,841 --> 00:11:27,401
I think, like, with code, we've gotten to a point of maturity around, like, testing and life cycles

177
00:11:27,401 --> 00:11:34,341
and software development life cycles for ensuring that code bases grow healthily over time, usually.

178
00:11:34,941 --> 00:11:37,841
And agents are doing so much of this work for us.

179
00:11:38,001 --> 00:11:42,841
And the version of continual learning that I think is, like, most practical for most organizations

180
00:11:42,841 --> 00:11:45,581
is just treating a model kind of like a code base,

181
00:11:45,581 --> 00:11:48,761
where it is the sort of thing that you can have a life cycle around.

182
00:11:49,221 --> 00:11:50,661
You can add test cases to.

183
00:11:50,821 --> 00:11:51,741
You can add features to.

184
00:11:51,861 --> 00:11:53,821
You can add assets to.

185
00:11:54,141 --> 00:11:57,801
And this is the sort of thing that the process of using agents,

186
00:11:58,241 --> 00:11:59,681
like initially you collect traces.

187
00:11:59,841 --> 00:12:01,721
These traces have your prompts

188
00:12:01,721 --> 00:12:03,701
and they have your feedback to things that models did.

189
00:12:03,981 --> 00:12:07,101
And this, just from the process of using these models,

190
00:12:07,321 --> 00:12:10,981
you are accumulating the assets you need to create evals.

191
00:12:11,081 --> 00:12:15,001
It's just that making evals historically has been a little bit painful

192
00:12:15,001 --> 00:12:18,481
and a lot of organizations have not necessarily wanted to do it manually,

193
00:12:18,761 --> 00:12:21,701
but the agents are quite quickly getting to the point where,

194
00:12:22,041 --> 00:12:26,381
thankfully, we have robots that can do a lot of the eval curation for us,

195
00:12:26,461 --> 00:12:30,441
especially as we train them more on the best practices of making evals

196
00:12:30,441 --> 00:12:32,981
in ways that are robust and amenable to learning.

197
00:12:34,121 --> 00:12:38,401
And so I think another useful example here is the decision model craze

198
00:12:38,401 --> 00:12:42,361
that has kind of been taking the world by storm in the past couple weeks with Jev,

199
00:12:42,361 --> 00:12:44,721
and also the parallels around routing

200
00:12:44,721 --> 00:12:46,981
and people being very excited about cost optimization

201
00:12:46,981 --> 00:12:49,561
because we see that it's possible to spend

202
00:12:49,561 --> 00:12:51,861
very large amounts of money on Claude every month

203
00:12:51,861 --> 00:12:55,401
and we figure out when should we use a cheaper model.

204
00:12:56,001 --> 00:12:58,501
And this is the sort of thing that I think people are recognizing

205
00:12:58,501 --> 00:12:59,961
needs to kind of live in the harness.

206
00:13:00,361 --> 00:13:03,181
But it also is an attribute that looks relatively similar

207
00:13:03,181 --> 00:13:04,941
to decision models.

208
00:13:05,181 --> 00:13:06,681
And the reason I bring both these up together

209
00:13:06,681 --> 00:13:08,781
is because there's things with very clear labels

210
00:13:08,781 --> 00:13:09,921
where there's a right answer

211
00:13:09,921 --> 00:13:13,281
and there's an opportunity for a human or another agent to give feedback.

212
00:13:13,501 --> 00:13:16,201
You can have the big model and the small model do the same task,

213
00:13:16,541 --> 00:13:18,941
judge whether the small model was as good as the big model.

214
00:13:19,361 --> 00:13:22,861
If so, then this is now data of, like, for this kind of thing,

215
00:13:22,941 --> 00:13:24,361
we can start routing to the small model more.

216
00:13:24,741 --> 00:13:27,781
Same for a decision model that's learning to kind of mimic the way

217
00:13:27,781 --> 00:13:30,181
that a human in a certain situation would make a judgment.

218
00:13:30,721 --> 00:13:32,741
This is a great labelable action,

219
00:13:33,361 --> 00:13:35,541
where the human might be doing the approval

220
00:13:35,541 --> 00:13:38,761
the same way that, like, we approve our tool calls and Claude Code sometimes,

221
00:13:38,761 --> 00:13:44,921
times, but I think most people have turned on auto tool acceptance, but every organization

222
00:13:44,921 --> 00:13:48,821
for more bespoke things that aren't just about tool security are going to have their own

223
00:13:48,821 --> 00:13:52,361
rules and preferences, and these are the sorts of things that you could then compound into

224
00:13:52,361 --> 00:13:52,921
a decision model.

225
00:13:53,441 --> 00:13:57,081
And so this is a very believable form of continual learning that a lot of people could just plug

226
00:13:57,081 --> 00:13:59,821
in tomorrow and see immediate value from.

227
00:14:01,121 --> 00:14:05,041
But more broadly, what we see people wanting to do, and this is a lot of the work that

228
00:14:05,041 --> 00:14:09,101
we do at Prime Intellect is working with organizations that are looking to do relatively

229
00:14:09,101 --> 00:14:15,141
ambitious training projects around turning their traces from deploying a model in production

230
00:14:15,141 --> 00:14:21,881
for a high-volume use case into rich, robust environments that are blended, real, and synthetic

231
00:14:21,881 --> 00:14:26,741
that kind of capture the complexity of the real tasks they want to be using long-running

232
00:14:26,741 --> 00:14:27,421
agents for.

233
00:14:28,041 --> 00:14:32,481
And agents are very useful with sufficient orchestration around them, sufficient tooling

234
00:14:32,481 --> 00:14:34,881
and humans in the loop at turning these into

235
00:14:35,021 --> 00:14:37,101
environments that are rich enough to train in.

236
00:14:37,581 --> 00:14:42,721
And we see that this is getting more and more automated with every new generation of open and closed models.

237
00:14:43,021 --> 00:14:57,461
And this means that we have an opportunity to start having this whole thing of taking the stuff that we're already doing and turning it into a self-improvement loop just kind of fall out magically from applying agents to curation of the data

238
00:14:57,461 --> 00:15:01,361
and creating interfaces for humans to give signal about what they want.

239
00:15:02,261 --> 00:15:05,141
This is an idea that I've been, for my own setup,

240
00:15:05,301 --> 00:15:07,121
been playing around with a lot

241
00:15:07,121 --> 00:15:10,141
that I think will become a popular interface

242
00:15:10,141 --> 00:15:12,661
for interacting with models more in the future.

243
00:15:12,841 --> 00:15:14,501
You're seeing it a little bit in some of the bots,

244
00:15:14,681 --> 00:15:18,141
but I think, especially now that models are good at writing code on the fly,

245
00:15:18,781 --> 00:15:20,801
these kind of widget card features

246
00:15:20,801 --> 00:15:24,061
that are, I think, a very practical form of dynamic UI

247
00:15:24,061 --> 00:15:27,061
that are also really useful for eliciting human decisions.

248
00:15:27,061 --> 00:15:31,641
So they're kind of like the tool call approval thing, but much more personalized, much more

249
00:15:31,641 --> 00:15:37,581
specific, and also kind of get around the problem of, like, okay, you have 20 different

250
00:15:37,581 --> 00:15:41,581
bots that are going, and you have to remember which ones are where, and who is curating

251
00:15:41,581 --> 00:15:44,261
all this for you, and then people are like, do you want one bot?

252
00:15:44,441 --> 00:15:46,721
And then you're like, oh, one's great, but I actually want two.

253
00:15:46,901 --> 00:15:47,501
Then you have two.

254
00:15:47,641 --> 00:15:48,181
Then you have 20.

255
00:15:48,441 --> 00:15:52,381
And I think we kind of are oscillating back and forth between this single stream and the

256
00:15:52,381 --> 00:15:53,981
multi-stream agent threads.

257
00:15:54,061 --> 00:15:55,801
But what we really want to do are make decisions.

258
00:15:55,801 --> 00:16:00,721
We want to give ideas to the agents, let them run with the ideas, and then come back to

259
00:16:00,721 --> 00:16:01,861
us when we need to make decisions.

260
00:16:02,441 --> 00:16:05,981
And the agents hopefully should make good choices about when to bring us decisions.

261
00:16:06,421 --> 00:16:11,581
And if agents can create interfaces on the fly for us that allow us to very directly

262
00:16:11,581 --> 00:16:16,901
give them the information they need, one, this is much more ergonomic for us as humans

263
00:16:18,121 --> 00:16:22,561
than reading, trying to parse through a long thread message to see where was the question.

264
00:16:22,561 --> 00:16:29,161
It's also kind of like the human dual side of the decision model, Jev style, but it also

265
00:16:29,161 --> 00:16:30,441
is perfect data.

266
00:16:30,921 --> 00:16:36,321
And so it's a setting where we've been experimenting with this as a way of training models to kind

267
00:16:36,321 --> 00:16:41,761
of get this feedback loop going from having human input with these kinds of natural interfaces

268
00:16:41,761 --> 00:16:46,661
be the sort of thing that directly acts as a label and is the sort of thing that a model

269
00:16:46,661 --> 00:16:49,641
can directly improve on because there is a very explicit target.

270
00:16:49,641 --> 00:16:52,061
and it gets around a lot of the problems

271
00:16:52,061 --> 00:16:54,581
that we kind of hear about in reward hacking and whatnot.

272
00:16:55,241 --> 00:16:56,621
Those things are worth addressing too,

273
00:16:56,681 --> 00:16:59,801
but this is where we see a very useful starting point

274
00:16:59,801 --> 00:17:01,701
being for a lot of these problems.

275
00:17:02,001 --> 00:17:03,061
And this can be done,

276
00:17:03,141 --> 00:17:04,501
it doesn't have to be with every model in your stack,

277
00:17:04,661 --> 00:17:05,901
it can be the orchestration layer models,

278
00:17:06,021 --> 00:17:06,801
it can be the decision models,

279
00:17:07,141 --> 00:17:08,801
but it is an entry point for

280
00:17:08,801 --> 00:17:10,961
kind of starting to have this compounding

281
00:17:10,961 --> 00:17:12,901
of preferences and latent knowledge

282
00:17:12,901 --> 00:17:13,981
happen in the models.

283
00:17:18,161 --> 00:17:20,221
And so I think there's a recent example today

284
00:17:20,221 --> 00:17:21,561
I just added this earlier

285
00:17:21,561 --> 00:17:24,141
From a company called Paper Instruments

286
00:17:24,141 --> 00:17:26,381
They use some of our tooling to do this

287
00:17:26,381 --> 00:17:27,441
They're friends of ours

288
00:17:28,021 --> 00:17:31,261
But they trained a model for 300k

289
00:17:31,261 --> 00:17:33,861
That is kind of Pareto Frontier optimal

290
00:17:33,861 --> 00:17:37,281
For knowledge work for things like spreadsheets and slides

291
00:17:37,281 --> 00:17:39,801
And it's like an employee scale investment

292
00:17:39,801 --> 00:17:40,841
For an org to do

293
00:17:41,501 --> 00:17:44,221
where you could go hire a software engineer

294
00:17:44,221 --> 00:17:45,541
or you could train a model.

295
00:17:46,241 --> 00:17:47,761
And you can also spread this out

296
00:17:47,761 --> 00:17:48,681
over a longer period of time

297
00:17:48,681 --> 00:17:50,401
if you're kind of compounding the knowledge over time.

298
00:17:50,621 --> 00:17:51,581
But it's the sort of thing

299
00:17:51,581 --> 00:17:53,821
that is increasingly becoming quite practical

300
00:17:54,201 --> 00:17:55,161
and it's the sort of thing

301
00:17:55,161 --> 00:17:57,801
that a lot of organizations we see wanting to do now

302
00:17:57,801 --> 00:17:59,621
and we kind of see the trend

303
00:17:59,621 --> 00:18:01,441
of how much more the demand for this

304
00:18:01,441 --> 00:18:03,421
is growing month after month

305
00:18:04,661 --> 00:18:06,161
as people start recognizing,

306
00:18:06,501 --> 00:18:07,441
one, the value of this

307
00:18:07,441 --> 00:18:09,721
and how kind of the open models

308
00:18:09,721 --> 00:18:14,761
are good enough that you can get them to be much better and also much more personalized

309
00:18:14,761 --> 00:18:20,721
by turning your knowledge for your domain, your harness, your environment into a training

310
00:18:20,721 --> 00:18:21,281
setting for them.

311
00:18:22,121 --> 00:18:26,301
But also it's the sort of thing that naturally lends itself to a learning loop.

312
00:18:26,661 --> 00:18:30,741
And so this is a prototype of something we're doing internally around kind of how we're

313
00:18:30,741 --> 00:18:35,001
imagining people can really deploy continual learning in the world, which is that you can

314
00:18:35,001 --> 00:18:41,021
have traces, you can start collecting these, and you can then, like, every night, see the

315
00:18:41,021 --> 00:18:46,161
proposed experiment, look at the things that have been flagged as issues, and kick off

316
00:18:46,161 --> 00:18:46,841
a training run for the night.

317
00:18:47,001 --> 00:18:48,921
Maybe your inference capacity in the day is serving your agents.

318
00:18:49,261 --> 00:18:54,121
At night, employees are not using them as much, and you can use that capacity for training

319
00:18:54,121 --> 00:18:54,441
instead.

320
00:18:54,661 --> 00:18:59,341
And so this is a very natural way that we see organizations able to, like, quite soon,

321
00:18:59,981 --> 00:19:04,601
actually do for real continual learning in the weights of the models and start compounding

322
00:19:04,601 --> 00:19:08,981
their knowledge into better and better capabilities for their agents.

323
00:19:10,521 --> 00:19:15,001
And so this is kind of where we see a lot of this culminating towards, there's so many

324
00:19:15,001 --> 00:19:15,661
things you do first.

325
00:19:16,161 --> 00:19:22,141
You discover your workflow, you create skills, you make evals, you do routing, and then eventually

326
00:19:22,141 --> 00:19:27,401
you start to want to get the economies of scale, get the compounding benefits that come

327
00:19:27,401 --> 00:19:32,621
from the data inside of your companies, your organizations, and have the organizations become

328
00:19:32,621 --> 00:19:36,901
increasingly self-improving in ways that get better and better over time.

329
00:19:38,761 --> 00:19:39,441
That's what I got.

330
00:19:41,181 --> 00:19:43,001
I can maybe take questions if there's any time.

331
00:19:43,581 --> 00:19:43,801
I don't know.

332
00:19:43,901 --> 00:19:44,321
No questions.

333
00:19:45,661 --> 00:19:47,501
And, yeah, we're at Prime Intellect.

334
00:19:47,661 --> 00:19:49,061
We are working on a lot of this stuff.

335
00:19:49,241 --> 00:19:53,381
A lot of it is open source and public, and we blog about it and tweet about it, but also

336
00:19:53,381 --> 00:19:56,921
we work with companies and we're hiring and growing, and so get in touch if you want to

337
00:19:56,921 --> 00:19:57,261
get in touch.

338
00:19:57,621 --> 00:19:57,981
Thanks.

339
00:19:57,981 --> 00:19:59,801
Thank you.