1
00:00:03,601 --> 00:00:21,161
Hi, everyone. I'm Joseph Kim, and I lead ChatGPT for Financial Services, a new product we launched just one month ago. Many of you have already experienced how tools like Codex have changed the way you build in software engineering. AI has become a much bigger part of the work you do every day.

2
00:00:22,001 --> 00:00:25,581
Model intelligence is continuing to improve at an incredible pace.

3
00:00:25,841 --> 00:00:31,861
But many people working industries like finance haven't yet experienced that same change in their day-to-day work.

4
00:00:32,561 --> 00:00:35,781
I expect that to start changing rapidly over the next few months.

5
00:00:36,441 --> 00:00:40,981
And builders like all of you in this room have a big role in making that happen quickly.

6
00:00:41,381 --> 00:00:46,461
Taking the intelligence in these models and turning it into products that people can use in their daily work.

7
00:00:47,221 --> 00:00:51,781
Today, I will share a little bit about what we learned building ChatGPT for financial services

8
00:00:51,781 --> 00:00:53,901
and how we're thinking about what comes next.

9
00:00:55,361 --> 00:00:58,301
I'll start by showing you what we've built, including a quick demo.

10
00:00:59,061 --> 00:01:02,661
Then I'll explain why we built it and why we think now is the right time to build.

11
00:01:03,501 --> 00:01:08,581
From there, I'll walk you through how we built it with some technical lessons that I hope you can apply to your own products.

12
00:01:10,141 --> 00:01:13,161
Finally, we'll take a look at where models are headed and products are going

13
00:01:13,161 --> 00:01:15,881
and what that can mean for financial work over the next year.

14
00:01:17,241 --> 00:01:18,641
Let's start with what we've built.

15
00:01:21,801 --> 00:01:26,681
ChatGPT for Financial Services is a tailored version of ChatGPT work built specifically

16
00:01:26,681 --> 00:01:30,401
for financial professionals, starting with investment banking and equity research.

17
00:01:30,821 --> 00:01:35,021
It's designed to help people at the large financial institutions get more out of their

18
00:01:35,021 --> 00:01:38,541
day-to-day work, similar to how Codex speeds you up in software development.

19
00:01:39,781 --> 00:01:44,401
As we built the product with our design partners, Morgan Stanley and Evercore, we focused on

20
00:01:44,401 --> 00:01:45,181
three core areas.

21
00:01:45,821 --> 00:01:47,361
First, financial data.

22
00:01:48,341 --> 00:01:50,761
Financial data is the foundation of financial work,

23
00:01:50,781 --> 00:01:53,381
and the work you can do is limited by the quality of your sources.

24
00:01:54,401 --> 00:01:58,361
We wanted people to have access to the data they need directly within ChatGPT.

25
00:01:59,501 --> 00:02:01,841
Second, an interface built for finance.

26
00:02:02,701 --> 00:02:05,681
We wanted the product to feel familiar to financial professionals

27
00:02:05,681 --> 00:02:09,481
with ways to inspect source material, explore financial charts,

28
00:02:09,681 --> 00:02:11,681
and control the sources behind an analysis.

29
00:02:12,761 --> 00:02:14,661
Third, your firm's format.

30
00:02:15,381 --> 00:02:17,841
Financial work happens within the context of an organization,

31
00:02:18,641 --> 00:02:20,501
and the models and presentations people produce

32
00:02:20,501 --> 00:02:23,061
need to follow their forms format and standards.

33
00:02:23,901 --> 00:02:27,421
We made it easier to create those artifacts using firm-specific templates.

34
00:02:28,421 --> 00:02:30,381
But they say a picture is worth a thousand words,

35
00:02:30,581 --> 00:02:32,721
so I'll stop talking for a bit and show you the product.

36
00:04:03,021 --> 00:04:03,421
Wonderful.

37
00:04:03,921 --> 00:04:06,501
Now I'd like to take some time to talk about why we built it.

38
00:04:08,321 --> 00:04:12,841
We built ChatGPT for financial services to help close what we call the capability overhang.

39
00:04:13,501 --> 00:04:18,201
As you can see here, model capabilities have improved at a tremendous rate over the last few years.

40
00:04:19,201 --> 00:04:22,361
In just a few years, we've gone from models solving high school math problems

41
00:04:22,361 --> 00:04:25,341
to systems capable of solving millennial-prized math problems.

42
00:04:25,781 --> 00:04:29,261
But the value people get from these models in their day-to-day work hasn't kept pace.

43
00:04:29,841 --> 00:04:33,061
You read about the breakthroughs in the news, yet for many financial professionals,

44
00:04:33,321 --> 00:04:35,261
the way they work hasn't changed nearly as quickly.

45
00:04:36,021 --> 00:04:40,381
We call that gap the capability overhang, the distance between what the model is capable of

46
00:04:40,381 --> 00:04:42,641
and the value people can actually get from it.

47
00:04:43,641 --> 00:04:46,061
Closing that gap takes more than building a capable model.

48
00:04:46,701 --> 00:04:50,401
People need the right data and tools and they need an experience that fits the way they

49
00:04:50,401 --> 00:04:50,681
work.

50
00:04:51,241 --> 00:04:52,781
That's where builders like you come in.

51
00:04:53,441 --> 00:04:56,521
I want to show you an example of how much of a difference the right product experience

52
00:04:56,521 --> 00:04:57,021
can make.

53
00:04:58,621 --> 00:05:03,721
This is a chart that shows how codex usage grew within OpenAI across different job functions.

54
00:05:04,681 --> 00:05:08,701
As you can see, engineers started using codex early and that share grew steadily.

55
00:05:09,541 --> 00:05:11,701
But look at finance, legal, and recruiting.

56
00:05:12,521 --> 00:05:16,461
At the beginning of 2026, almost none of their AI output was generated in Codex.

57
00:05:17,041 --> 00:05:19,241
By June, that share was around 90%.

58
00:05:20,141 --> 00:05:25,001
If you look carefully, you'll notice a huge inflection point that happened around February, March of this year.

59
00:05:25,541 --> 00:05:29,901
I went back to see what happened around this time, assuming that we must have released a new model,

60
00:05:30,061 --> 00:05:33,681
but instead I found that we had released the Codex desktop app in February.

61
00:05:34,621 --> 00:05:39,741
My takeaway is that the product made it easier for people outside of engineering to put these capabilities to work.

62
00:05:40,521 --> 00:05:42,641
That's the opportunity we're pursuing in finance,

63
00:05:42,921 --> 00:05:48,061
building an experience that helps financial professionals get much more value from the intelligence already available to them.

64
00:05:49,541 --> 00:05:51,201
Now let me talk about how we built it.

65
00:05:53,101 --> 00:05:54,121
Let's start with data.

66
00:05:54,961 --> 00:05:58,761
In finance, the work you can do is limited by the quality of the information you have access to.

67
00:05:59,281 --> 00:06:02,201
So we wanted the product to come with the sources professionals rely on,

68
00:06:02,201 --> 00:06:06,301
from financial statements and earning transcripts to market and private company data.

69
00:06:06,981 --> 00:06:09,781
We looked at existing MCP connectors for financial data

70
00:06:09,781 --> 00:06:13,661
and found that many had issues with latency, errors, or search quality.

71
00:06:14,241 --> 00:06:16,581
So we went and licensed a lot of this data directly

72
00:06:16,581 --> 00:06:18,741
and built our own hosting and search indices

73
00:06:18,741 --> 00:06:21,121
using the same infrastructure as our web index.

74
00:06:21,961 --> 00:06:25,321
We still use the MCP protocol to give the model access to the data,

75
00:06:25,501 --> 00:06:29,201
but hosting and indexing it ourselves gives us more control over the quality

76
00:06:29,201 --> 00:06:30,801
and the reliability of the experience.

77
00:06:31,741 --> 00:06:35,821
It also gives us the flexibility to post-train our models to work with these data sets.

78
00:06:36,581 --> 00:06:41,201
The analogy I like to use is MCP and skills are like handing an instruction manual to

79
00:06:41,201 --> 00:06:45,361
a junior analyst versus post-training where the model is like having a senior analyst

80
00:06:45,361 --> 00:06:49,481
who's been working with the data for over two decades and knows all of its quirks like

81
00:06:49,481 --> 00:06:50,221
the back of their hand.

82
00:06:51,241 --> 00:06:53,961
MCP and skills give the model access to instructions.

83
00:06:54,361 --> 00:06:57,641
Post-training can help it develop more familiarity with the data itself.

84
00:06:58,741 --> 00:07:02,281
Once we had that data foundation in place, the next step was making it easy to turn that

85
00:07:02,281 --> 00:07:03,841
information to professional-grade work.

86
00:07:05,541 --> 00:07:09,641
For financial professionals, that work often takes the form of an Excel model, a PowerPoint

87
00:07:09,641 --> 00:07:13,261
presentation, or a written analysis that needs to be reviewed and shared with other people.

88
00:07:14,021 --> 00:07:17,921
So we focused on helping the model produce those outputs in a way that fits the firm's

89
00:07:17,921 --> 00:07:19,181
existing workflows.

90
00:07:20,001 --> 00:07:24,321
We built a template system that gives users access to the firm's approved templates when

91
00:07:24,321 --> 00:07:25,681
generating spreadsheets and presentations.

92
00:07:26,641 --> 00:07:30,761
The goal is to reduce the work someone has to do to get an output ready for their team

93
00:07:30,761 --> 00:07:31,261
or clients.

94
00:07:32,281 --> 00:07:36,941
We also added things like interactive financial charts so users can explore stock prices and

95
00:07:36,941 --> 00:07:40,641
financial metrics directly within ChatGPT as they work through an analysis.

96
00:07:41,261 --> 00:07:46,161
And we built citation viewers that let people inspect the source passages and tables behind

97
00:07:46,161 --> 00:07:46,681
an answer.

98
00:07:47,021 --> 00:07:51,261
In finance, you need to be able to check where a number came from and understand the context

99
00:07:51,261 --> 00:07:51,781
around it.

100
00:07:52,221 --> 00:07:54,681
We wanted that review to be a natural part of the experience.

101
00:07:55,601 --> 00:07:58,861
These are the kind of product details that help turn a model's capabilities into work

102
00:07:58,861 --> 00:07:59,761
that someone can use.

103
00:08:01,021 --> 00:08:04,621
The part I think that is most relevant for people in the room is how we built it.

104
00:08:05,601 --> 00:08:09,721
Much of the functionality you just saw was built using the same ChatGPT plugin platform

105
00:08:09,721 --> 00:08:11,401
that's available to you as developers.

106
00:08:12,381 --> 00:08:16,401
For financial data, we used MCP tools and connectors to give the model access to the

107
00:08:16,401 --> 00:08:17,221
information it needs.

108
00:08:17,221 --> 00:08:24,481
We use citation metadata in MCP apps to let users inspect their sources, and MCP apps also power interactive financial charts.

109
00:08:25,101 --> 00:08:31,301
We use skills together with shared artifact tools to guide financial workflows and produce Excel models and PowerPoint presentations.

110
00:08:32,021 --> 00:08:36,181
And for financial onboarding, we leverage the native plugin onboarding experience.

111
00:08:37,421 --> 00:08:41,921
What that means for builders is that you can focus on the expertise that makes your product valuable.

112
00:08:41,921 --> 00:08:46,901
If you have unique data, an existing tool, or a deep understanding of a particular workflow,

113
00:08:47,241 --> 00:08:51,781
you can use these building blocks to bring that into ChatGPT and create an experience around it.

114
00:08:52,501 --> 00:08:58,361
Financial services is just one example, but I do think this opportunity expands far beyond finance.

115
00:08:58,741 --> 00:09:04,161
I personally believe that we're going to see multiple $1 billion businesses built simply as ChatGPT plugins.

116
00:09:04,641 --> 00:09:07,541
And that opportunity will keep evolving as the models become more capable.

117
00:09:09,781 --> 00:09:14,121
Last but not least, I'd like to spend the rest of the talk on what comes next and what

118
00:09:14,121 --> 00:09:15,801
that means for the products we should be building today.

119
00:09:18,221 --> 00:09:21,941
I'd like to start by talking about model capabilities and expectations for the future.

120
00:09:23,021 --> 00:09:26,701
Models are improving so quickly that it can be difficult to get an intuitive sense of

121
00:09:26,701 --> 00:09:27,381
what comes next.

122
00:09:28,421 --> 00:09:31,281
One useful way to build that intuition is to look backwards.

123
00:09:32,541 --> 00:09:36,041
Something remarkable about the last couple of years is how predictable parts of that

124
00:09:36,041 --> 00:09:41,261
progress have been. Model improvements have come from three broad, sometimes overlapping drivers.

125
00:09:42,001 --> 00:09:47,061
Scaling compute, improving algorithms, and something we call unhobbling, which is about

126
00:09:47,061 --> 00:09:51,461
finding ways to get more useful work out of the models. We'll come back to unhobbling on a later

127
00:09:51,461 --> 00:09:58,021
slide. In 2002 and 2003, the most visible driver was pre-training, which meant scaling compute by

128
00:09:58,021 --> 00:09:59,961
making the models bigger and adding more data.

129
00:10:00,741 --> 00:10:25,001
By leveraging scaling laws, researchers were able to predict model performance with high accuracy. For example, before GPT-4 finished training, OpenAI researchers were able to predict its final loss as well as its performance on a coding eval. Improvements on the algorithm side have also been surprisingly steady. One study found that the compute needed to reach a given level of language model performance from scratch fell by roughly half every

130
00:10:25,021 --> 00:10:29,601
eight months. Although you can't predict the timing of each individual breakthrough, the

131
00:10:29,601 --> 00:10:36,541
cumulative rate of improvement follows a surprisingly steady trend. In 2024 and 2025, reasoning

132
00:10:36,541 --> 00:10:40,821
became another major driver of progress. We used reinforcement learning and post-training

133
00:10:40,821 --> 00:10:44,821
to teach models to work through problems more effectively, and we gave them more compute

134
00:10:44,821 --> 00:10:49,661
to think when answering. That opened up another way to improve performance alongside pre-training.

135
00:10:50,361 --> 00:10:54,721
However, for folks that don't know, at OpenAI, we ran into problems with pre-training and

136
00:10:54,721 --> 00:10:56,081
and therefore had difficulty scaling the model.

137
00:10:57,381 --> 00:11:01,741
In 2026, we've been able to solve many of these issues, and we've successfully trained

138
00:11:01,741 --> 00:11:06,601
larger and more capable models, as you've seen through GPD-56 Sol and GPD-6 Astro.

139
00:11:07,981 --> 00:11:11,761
Throughout these years, researchers have raised concerns about potential barriers to further

140
00:11:11,761 --> 00:11:12,061
scaling.

141
00:11:12,801 --> 00:11:18,001
For example, in 2024, many researchers were concerned about a potential data wall where

142
00:11:18,001 --> 00:11:20,221
we would run out of human-generated text to train on.

143
00:11:20,981 --> 00:11:23,261
But techniques like synthetic data generation,

144
00:11:23,681 --> 00:11:25,261
learning from verifiable rewards,

145
00:11:25,501 --> 00:11:27,021
as well as algorithmic improvements

146
00:11:27,021 --> 00:11:28,981
to learn more from each individual token

147
00:11:28,981 --> 00:11:30,481
have essentially solved this problem.

148
00:11:31,161 --> 00:11:33,281
We've also solved concerns around things

149
00:11:33,281 --> 00:11:34,421
like context window length

150
00:11:34,421 --> 00:11:36,901
using things like sparse detention and compaction.

151
00:11:38,321 --> 00:11:39,941
Looking ahead to 2027,

152
00:11:40,241 --> 00:11:42,221
I strongly believe that the trend lines of the past

153
00:11:42,221 --> 00:11:44,061
will continue and that we'll train larger,

154
00:11:44,341 --> 00:11:46,021
more efficient, and more capable models.

155
00:11:46,381 --> 00:11:48,661
We clearly still have very difficult problems to solve,

156
00:11:48,661 --> 00:11:50,601
especially around things like alignment and safety,

157
00:11:50,781 --> 00:11:53,661
but like we have in the past, I believe that human ingenuity,

158
00:11:53,801 --> 00:11:55,841
increasingly assisted by capable models,

159
00:11:56,041 --> 00:11:57,761
will help us make progress on those problems too.

160
00:11:58,581 --> 00:12:01,181
So how good will models be at the end of 2027?

161
00:12:01,701 --> 00:12:03,161
Let's use math as an example.

162
00:12:03,841 --> 00:12:08,581
In 2023, GPT-4 scored 700 out of 800 on SAT high school math.

163
00:12:09,441 --> 00:12:12,901
By 2005, an opening eye model reached gold medal status

164
00:12:12,901 --> 00:12:15,481
on the International Mathematical Olympiad.

165
00:12:16,381 --> 00:12:25,061
And in early 2026, an internal model disproved a long-standing conjecture related to a problem that mathematicians have been working on since 1946.

166
00:12:25,981 --> 00:12:31,481
More recently, an internal model proposed a solution to the Navier-Stokes Millennium Prize puzzle problem,

167
00:12:32,401 --> 00:12:38,581
and we also shared 719 mathematical research manuscripts across 372 families of results.

168
00:12:39,221 --> 00:12:42,561
In just a few years, we've gone from a strong score on a high school math exam

169
00:12:42,561 --> 00:12:46,161
to AI systems producing original research at the frontier of mathematics.

170
00:12:46,681 --> 00:12:51,021
If we see another year of progress roughly in line with what we've been seeing over the past couple years,

171
00:12:51,281 --> 00:12:55,901
I expect that by end of 2027, models will outperform even leading human experts

172
00:12:55,901 --> 00:12:59,821
on a growing range of difficult tasks across different fields and specialties.

173
00:13:00,561 --> 00:13:02,881
Mathematics gives us an early glimpse of what that could look like.

174
00:13:04,681 --> 00:13:08,621
And while the capability curve may look gradual, the change in how we work can feel sudden.

175
00:13:08,941 --> 00:13:10,661
I call this the threshold phenomenon.

176
00:13:11,141 --> 00:13:14,821
For any given task, there's a threshold of usefulness, the point where a model becomes

177
00:13:14,821 --> 00:13:16,601
capable and reliable enough to save you time.

178
00:13:17,081 --> 00:13:20,561
That threshold depends on the work you're doing and how much checking or human review

179
00:13:20,561 --> 00:13:21,181
it requires.

180
00:13:22,221 --> 00:13:24,621
Below that threshold, improvements can be hard to appreciate.

181
00:13:25,041 --> 00:13:28,881
The model might produce better answers, make fewer mistakes, or get further through a task.

182
00:13:29,321 --> 00:13:32,561
But if checking and fixing its work still takes as long as doing it yourself,

183
00:13:32,881 --> 00:13:34,701
your practical assessment hasn't changed.

184
00:13:34,921 --> 00:13:35,921
It's still not useful.

185
00:13:36,841 --> 00:13:39,781
Meaningful improvements in capability can feel like little progress.

186
00:13:40,661 --> 00:13:42,561
Then the model crosses that threshold.

187
00:13:43,001 --> 00:13:44,581
Suddenly it produces something you can use.

188
00:13:44,801 --> 00:13:48,301
The time you spend reviewing it becomes much smaller than the time it saves you.

189
00:13:48,621 --> 00:13:49,901
You start using it regularly,

190
00:13:50,101 --> 00:13:52,261
and eventually you become comfortable delegating that task.

191
00:13:52,621 --> 00:13:54,781
That transition can feel like an enormous leap,

192
00:13:54,921 --> 00:13:57,681
even if the underlying improvement was similar to the previous few improvements.

193
00:13:58,061 --> 00:14:00,321
What changes of progress became useful to you.

194
00:14:01,261 --> 00:14:04,021
This helps explain why model progress sometimes feels sudden,

195
00:14:04,301 --> 00:14:07,421
even when parts of the underlying trend have been steady and predictable.

196
00:14:07,801 --> 00:14:10,001
And because difficult tasks have different thresholds,

197
00:14:10,081 --> 00:14:13,401
we should expect those moments to arrive at different times across the work that we do.

198
00:14:14,001 --> 00:14:15,941
So as you think about the products you're building,

199
00:14:16,201 --> 00:14:17,681
I'd encourage you to ask yourself,

200
00:14:17,861 --> 00:14:22,101
what model capabilities am I underappreciating today because they aren't quite useful yet?

201
00:14:22,721 --> 00:14:25,121
Where might those capabilities be a year from now?

202
00:14:25,921 --> 00:14:29,941
Then think about what you would build if those capabilities crossed the threshold of usefulness

203
00:14:29,941 --> 00:14:31,661
and became reliable enough to delegate?

204
00:14:32,281 --> 00:14:33,621
How would the interface change?

205
00:14:34,701 --> 00:14:37,141
What would the users still need to control, review, or approve?

206
00:14:38,141 --> 00:14:40,781
Those are product and design questions we can start working on now.

207
00:14:41,681 --> 00:14:43,401
We want to have the right experiences ready

208
00:14:43,401 --> 00:14:45,201
so that when the models cross that threshold,

209
00:14:45,441 --> 00:14:47,681
people can take advantage of the capabilities right away,

210
00:14:47,801 --> 00:14:49,941
and so you can stay one step ahead of your competitors.

211
00:14:51,721 --> 00:14:53,101
So what does this mean for finance?

212
00:14:53,581 --> 00:14:56,161
I think there are two forces that will push more financial work

213
00:14:56,161 --> 00:14:57,581
across that threshold of usefulness.

214
00:14:57,581 --> 00:15:00,321
I think of them as the rising tide and the waves on top.

215
00:15:00,841 --> 00:15:04,421
The rising tide is the underlying improvements to the underlying model.

216
00:15:05,341 --> 00:15:09,801
Historically, we haven't dedicated a significant amount of training time or compute specifically to finance.

217
00:15:10,581 --> 00:15:15,601
Much of the financial capabilities you see today have come from building larger, more capable models

218
00:15:15,601 --> 00:15:20,021
and from transferable skills developed in areas like mathematics, coding, and reasoning.

219
00:15:20,821 --> 00:15:25,521
As these general capabilities improved, they lifted performance across many kinds of financial work.

220
00:15:26,401 --> 00:15:28,981
The Ways are what happens when we focus directly on finance.

221
00:15:29,441 --> 00:15:33,741
My team is leading an effort to invest significantly more in finance-specific post-training.

222
00:15:33,941 --> 00:15:36,761
The idea here is to build on those broad capabilities

223
00:15:36,761 --> 00:15:39,801
and train the model more directly on the work people do in finance.

224
00:15:40,261 --> 00:15:43,481
That can create peaks in performance on particular tasks on top of the rising tide.

225
00:15:44,461 --> 00:15:48,121
Over the next three months, we're focused on improving general financial reasoning

226
00:15:48,121 --> 00:15:50,341
and the quality of the work these models produce,

227
00:15:50,561 --> 00:15:53,401
analyzing financial information, building Excel models,

228
00:15:53,521 --> 00:15:55,061
and creating PowerPoint presentations.

229
00:15:55,521 --> 00:15:59,781
These are the kinds of tasks that make up a significant percentage of the daily work of a junior analyst,

230
00:15:59,901 --> 00:16:04,361
and we want the output to be accurate, well-structured, and increasingly useful in real workflows.

231
00:16:04,981 --> 00:16:11,041
Beyond that, we plan to invest more in the kinds of judgment that senior bankers, analysts, and investors bring to more difficult questions.

232
00:16:11,461 --> 00:16:16,941
This includes taking in large amounts of information, deciding what matters, and essentially trying to predict the future,

233
00:16:17,261 --> 00:16:23,721
forecasting a company's earnings, assessing where commodity prices might go, or evaluating which businesses are attractive investments.

234
00:16:24,401 --> 00:16:28,221
Our goal is for models to match the performance of the best human experts

235
00:16:28,221 --> 00:16:31,341
on these types of financial tasks sometime in 2027.

236
00:16:31,981 --> 00:16:35,041
We expect that progress to come from both forces working together,

237
00:16:35,241 --> 00:16:37,041
the underlying model continuing to improve,

238
00:16:37,181 --> 00:16:39,901
and our team training it more directly for the work people do in finance.

239
00:16:41,041 --> 00:16:43,081
We've talked about how models are becoming more capable

240
00:16:43,081 --> 00:16:46,021
and how we can train them to perform better on specific tasks,

241
00:16:46,161 --> 00:16:49,541
but I want to come back to one third driver I mentioned earlier unhobbling.

242
00:16:50,161 --> 00:16:51,741
We've already seen a few examples.

243
00:16:53,281 --> 00:16:57,521
ChatGPT use RLHF to teach models to follow instructions and have conversations with us.

244
00:16:58,101 --> 00:17:02,801
Tools allow them to search for information, write and run code, and interact with the software we use.

245
00:17:03,121 --> 00:17:05,541
Each of these makes the model more useful in the real world.

246
00:17:06,261 --> 00:17:09,421
I think the next form of unhobbling will be proactive and persistent agents.

247
00:17:11,021 --> 00:17:15,101
Today, a lot of our interactions with AI still happen when we open a conversation and tell it what to do next.

248
00:17:15,401 --> 00:17:20,201
A persistent agent can hold on to a goal, remember the relevant context, and continue working in the background.

249
00:17:20,201 --> 00:17:24,441
Once you've given it a direction, it can make progress even when you aren't actively prompting it.

250
00:17:25,261 --> 00:17:29,901
The other reason persistence matters is it changes the API or the way people work with the agent.

251
00:17:30,621 --> 00:17:36,701
Today, getting the most out of AI agents requires learning a new workflow, learning how to prompt it, and learning how to use the tools.

252
00:17:37,001 --> 00:17:39,201
A persistent agent can work in a way that's more familiar.

253
00:17:39,681 --> 00:17:42,581
You can give it a project, share context, and ask for an update.

254
00:17:43,441 --> 00:17:48,341
It's much more similar to how humans work with other humans, and therefore makes the AI more accessible to people

255
00:17:48,341 --> 00:17:50,321
that don't develop those initial skills.

256
00:17:51,781 --> 00:17:54,461
Let me close with three bets for finance in 2027.

257
00:17:54,921 --> 00:17:56,561
These are just a couple of my expectations

258
00:17:56,561 --> 00:17:58,481
for what comes as the models improve.

259
00:17:59,141 --> 00:18:01,561
First, people will rely rarely, if ever,

260
00:18:02,021 --> 00:18:02,961
on Excel and PowerPoint

261
00:18:02,961 --> 00:18:04,921
to make financial models or pitch decks.

262
00:18:05,781 --> 00:18:07,561
Second, AI financial analysts

263
00:18:07,561 --> 00:18:08,941
will start to feel like human coworkers.

264
00:18:09,261 --> 00:18:10,881
They'll remember your preferences and context

265
00:18:10,881 --> 00:18:11,881
from the previous days,

266
00:18:12,161 --> 00:18:13,281
take ownership of assignments,

267
00:18:13,661 --> 00:18:14,481
productively follow up,

268
00:18:14,621 --> 00:18:16,381
and keep working when you're focused on something else.

269
00:18:17,361 --> 00:18:21,441
Third, models will start to outperform the best human experts on critical financial tasks.

270
00:18:21,621 --> 00:18:24,861
That includes things like forecasting earnings and developing investment ideas.

271
00:18:25,181 --> 00:18:29,781
I expect models to become better than the best people at making these economically important, forward-looking judgments.

272
00:18:30,401 --> 00:18:32,761
Those are my three bets for 2027.

273
00:18:33,101 --> 00:18:37,121
And if this is where we expect to be in 2027, I'll leave you with one last question.

274
00:18:37,561 --> 00:18:42,821
What should you start building today to be ready for inevitable thresholds that the models will cross tomorrow?

275
00:18:43,441 --> 00:18:44,081
Thank you.