WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=opHmX7qClzE

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/opHmX7qClzE.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.560 --> 00:00:13.400
Hi everyone.

s3
00:00:13.400 --> 00:00:20.160
I'm Joseph Kim and I lead ChatGPT for financial services, a new product we launched just 1 month ago.

s4
00:00:20.160 --> 00:00:26.360
Many of you have already experienced how tools like Codex have changed the way you build in software engineering.

s5
00:00:26.360 --> 00:00:30.800
AI has become a much bigger part of the work you do every day.

s6
00:00:30.800 --> 00:00:34.720
Model is continuing to improve at an incredible pace.

s7
00:00:34.720 --> 00:00:41.480
But many people working in industries like finance haven't yet experienced that same change in their day-to-day work.

s8
00:00:41.480 --> 00:00:45.280
I expect that to start changing rapidly over the next few months.

s9
00:00:45.280 --> 00:00:50.200
And builders like all of you in this room have a big role in making that happen quickly.

s10
00:00:50.200 --> 00:00:55.880
Taking the intelligence in these models and turning that into products that people can use in their daily work.

s11
00:00:55.880 --> 00:01:04.080
Today, I will share a little bit about what we learned building ChatGPT for financial services and how we're thinking about what comes next.

s12
00:01:04.080 --> 00:01:07.840
I'll start by showing you what we've built including a quick demo.

s13
00:01:07.840 --> 00:01:12.240
Then I'll explain why we built it and why we think now is the right time to build.

s14
00:01:12.240 --> 00:01:18.200
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.

s15
00:01:18.200 --> 00:01:26.080
Finally, we'll take a look at where models are headed and products are going and what that could mean for financial work over the next year.

s16
00:01:26.080 --> 00:01:28.840
Let's start with what we built.

s17
00:01:30.320 --> 00:01:39.680
ChatGPT for financial services is a tailored version of ChatGPT work built specifically for financial professionals starting with investment banking and equity research.

s18
00:01:39.680 --> 00:01:48.640
It's designed to help people at the large financial institutions get more out of their day-to-day work similar to how Codex speeds you up in software development.

s19
00:01:48.640 --> 00:01:54.640
As we built the product with our design partners Morgan Stanley and Evercore, we focused on three core areas.

s20
00:01:54.640 --> 00:01:57.120
First, financial data.

s21
00:01:57.120 --> 00:02:03.320
Financial data is a foundation of financial work, and the work you can do is limited by the quality of your sources.

s22
00:02:03.320 --> 00:02:08.200
We wanted people to have access to the data they need directly within Chat GPT.

s23
00:02:08.200 --> 00:02:11.560
Second, an interface built for finance.

s24
00:02:11.560 --> 00:02:21.480
We wanted the product to feel familiar to financial professionals with ways to inspect source material, explore financial charts, and control the sources behind an analysis.

s25
00:02:21.480 --> 00:02:23.959
Third, your firm's format.

s26
00:02:23.959 --> 00:02:32.680
Financial work happens within the context of an organization, and the models and presentations people produce need to follow their firm's format and standards.

s27
00:02:32.680 --> 00:02:37.200
We made it easier to create those artifacts using firm-specific templates.

s28
00:02:37.200 --> 00:02:42.920
But they say a picture's worth a thousand words, so I'll stop talking for a bit and show you the product.

s29
00:02:51.611 --> 00:02:53.611
[music]

s30
00:03:07.987 --> 00:03:09.987
[music]

s31
00:04:11.680 --> 00:04:12.640
Wonderful.

s32
00:04:12.640 --> 00:04:16.840
Now, I'd like to take some time to talk about why we built it.

s33
00:04:17.120 --> 00:04:22.280
We built ChatGPT for financial services to help close what we call the capability overhang.

s34
00:04:22.280 --> 00:04:28.040
As you can see here, model capabilities have improved at a tremendous rate over the last few years.

s35
00:04:28.040 --> 00:04:34.640
In just a few years, we've gone from models solving high school math problems to systems capable of solving millennial prize math problems.

s36
00:04:34.640 --> 00:04:38.680
But, the value people get from these models in their day-to-day work hasn't kept pace.

s37
00:04:38.680 --> 00:04:44.840
You read about the breakthroughs in the news, yet for many financial professionals, the way they work hasn't changed nearly as quickly.

s38
00:04:44.840 --> 00:04:52.320
We call that gap the capability overhang, the distance between what the model is capable of, and the value people can actually get from it.

s39
00:04:52.320 --> 00:04:55.240
Closing that gap takes more than building a capable model.

s40
00:04:55.240 --> 00:05:00.040
People need the right data and tools, and and they need an experience that fits the way they work.

s41
00:05:00.040 --> 00:05:02.040
That's where builders like you come in.

s42
00:05:02.040 --> 00:05:07.280
I want to show you an example of how much of a difference a right product experience can make.

s43
00:05:07.440 --> 00:05:13.480
This is a chart that shows how Codex usage grew within OpenAI across different job functions.

s44
00:05:13.480 --> 00:05:18.360
As you can see, engineers started using Codex early, and that share grew steadily.

s45
00:05:18.360 --> 00:05:21.280
But, look at finance, legal, and recruiting.

s46
00:05:21.280 --> 00:05:25.840
At the beginning of 2026, almost none of their AI output was generated in Codex.

s47
00:05:25.840 --> 00:05:28.919
By June, that share was around 90%.

s48
00:05:28.919 --> 00:05:34.360
If you look carefully, you'll notice a huge inflection point that happened around February-March of this year.

s49
00:05:34.360 --> 00:05:43.440
I went back to see what happened around this time, assuming that we must have released a new model, but instead I found that we had released the Codex desktop app in February.

s50
00:05:43.440 --> 00:05:49.200
My takeaway is that the product made it easier for people outside of engineering to put these capabilities to work.

s51
00:05:49.200 --> 00:05:51.680
That's the opportunity we're pursuing in finance.

s52
00:05:51.680 --> 00:05:58.320
Building an experience that helps financial professionals get much more value from the intelligence already available to them.

s53
00:05:58.320 --> 00:06:01.520
Now, let me talk about how we built it.

s54
00:06:01.800 --> 00:06:03.680
Let's start with data.

s55
00:06:03.680 --> 00:06:08.120
In finance, the work you can do is limited by the quality of the information you have access to.

s56
00:06:08.120 --> 00:06:15.880
So, we wanted the product to come with the sources professionals rely on, from financial statements and earning transcripts to market and private company data.

s57
00:06:15.880 --> 00:06:23.120
We looked at existing MCP connectors for financial data and found that many had issues with latency, errors, or search quality.

s58
00:06:23.120 --> 00:06:30.880
So, we went and licensed a lot of this data directly and built our own hosting and search indices using the same infrastructure as our web index.

s59
00:06:30.880 --> 00:06:40.600
We still use the MCP protocol to give the model access to the data, but hosting and indexing it ourselves gives us more control over the quality and the reliability of the experience.

s60
00:06:40.600 --> 00:06:45.360
It also gives us the flexibility to post train our models to work with these data sets.

s61
00:06:45.360 --> 00:06:56.760
The analogy I like to use is MCP and skills are like handing an instruction manual to a junior analyst versus post training, where the model's like having a senior analyst who's been working with the data for over two decades

s62
00:06:56.760 --> 00:06:59.760
and knows all of its quirks like the back of their hand.

s63
00:06:59.760 --> 00:07:02.880
MCP and skills give the model access and instructions.

s64
00:07:02.880 --> 00:07:07.480
Post training can help it develop more familiarity with the data itself.

s65
00:07:07.480 --> 00:07:14.040
Once we had that data foundation in place, the next step was making it easy to turn that information to professional grade work.

s66
00:07:14.280 --> 00:07:22.800
For financial professionals, that work often takes the form of an Excel model, a PowerPoint presentation, or written analysis that needs to be reviewed and shared with other people.

s67
00:07:22.800 --> 00:07:28.760
So, we focus on helping the model produce those outputs in a way that fits the firm's existing workflows.

s68
00:07:28.760 --> 00:07:35.600
We built a template system that gives users access to the firm's approved templates when generating spreadsheets and presentations.

s69
00:07:35.600 --> 00:07:40.960
The goal is to reduce the work someone has to do to get an output ready for their team or clients.

s70
00:07:40.960 --> 00:07:50.040
We also added things like interactive financial charts, so users can explore stock prices and financial metrics directly within ChatGPT as they work through an analysis.

s71
00:07:50.040 --> 00:07:55.720
And we built citation viewers that let people inspect the source passages and tables in behind an answer.

s72
00:07:55.720 --> 00:08:01.040
In finance, you need to be able to check where a number came from and understand the context around it.

s73
00:08:01.040 --> 00:08:04.360
We wanted that review to be a natural part of the experience.

s74
00:08:04.360 --> 00:08:09.800
These are the kind of product details that help turn a model's capabilities into work that someone can use.

s75
00:08:09.800 --> 00:08:14.280
The part I think that is most relevant for people in the room is how we built it.

s76
00:08:14.280 --> 00:08:21.120
Much of the functionality you just saw was built using the same ChatGPT plugin platform that's available to you as developers.

s77
00:08:21.120 --> 00:08:26.520
For financial data, we used MCP tools and connectors to give the model access to the information it needs.

s78
00:08:26.520 --> 00:08:34.080
We use citation metadata and MCP apps let users inspect the sources, and MCP apps also power interactive financial charts.

s79
00:08:34.080 --> 00:08:40.840
We use skills together with shared artifact tools to guide what financial workflows and produce Excel models and PowerPoint presentations.

s80
00:08:40.840 --> 00:08:46.200
And for financial onboarding, we leveraged the native plugin onboarding experience.

s81
00:08:46.360 --> 00:08:51.000
What that means for builders is you could is that you can focus on the expertise that makes your product valuable.

s82
00:08:51.000 --> 00:09:01.200
If you have unique data, an existing tool, or a deep understanding of a particular workflow, you can use these building blocks to bring that into ChatGPT and create an experience around it.

s83
00:09:01.200 --> 00:09:07.560
Financial services is just one example, but I do think this this opportunity expands far beyond finance.

s84
00:09:07.560 --> 00:09:13.520
I personally believe that we're going to see multiple $1 billion businesses built simply as ChatGPT plugins.

s85
00:09:13.520 --> 00:09:17.680
And that opportunity will keep evolving as the models become more capable.

s86
00:09:18.480 --> 00:09:26.040
Last but not least, I'd like to spend the rest of the talk on what comes next and what that means for the products we should be building today.

s87
00:09:26.960 --> 00:09:31.560
I'd like to start by talking about model capabilities and expectations for the future.

s88
00:09:31.560 --> 00:09:37.160
Models are improving so quickly that it can be difficult to get an intuitive sense of what comes next.

s89
00:09:37.160 --> 00:09:41.160
One useful way to build that intuition is to look backwards.

s90
00:09:41.160 --> 00:09:46.160
Something remarkable about the last couple of years is how predictable parts of that progress have been.

s91
00:09:46.160 --> 00:09:50.520
Model improvements have come from three broad, sometimes overlapping drivers.

s92
00:09:50.520 --> 00:09:58.280
Scaling compute, improving algorithms, and something we call unhobbling, which is about finding ways to get more useful work out of the models.

s93
00:09:58.280 --> 00:10:01.320
We'll come back to unhobbling on a on a later slide.

s94
00:10:01.320 --> 00:10:09.400
In 2002 and 2003, the most visible driver was pre-training, which meant scaling compute by making the models bigger and adding more data.

s95
00:10:09.400 --> 00:10:14.640
By leveraging scaling laws, researchers were able to predict model at performance with high accuracy.

s96
00:10:14.640 --> 00:10:23.240
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.

s97
00:10:23.240 --> 00:10:27.440
Improvements on the algorithm side have also been surprisingly steady.

s98
00:10:27.440 --> 00:10:35.000
One study found that the compute needed to reach a given level of language model performance from scratch fell by roughly half every eight months.

s99
00:10:35.000 --> 00:10:42.320
Although you can't predict the timing of each individual breakthrough, the cumulative rate of of improvement follows a surprisingly steady trend.

s100
00:10:42.320 --> 00:10:47.320
In 2024 and 2025, reasoning became another major driver of progress.

s101
00:10:47.320 --> 00:10:54.960
We used reinforcement learning and post-training to teach models to work through problems more effectively, and we gave them more compute to think when answering.

s102
00:10:54.960 --> 00:10:59.080
That opened up another way to improve performance alongside pre-training.

s103
00:10:59.080 --> 00:11:06.000
However, for folks that don't know, at OpenAI, we ran into problems with pre-training and therefore had difficulty scaling the model.

s104
00:11:06.000 --> 00:11:16.680
In 2026, we've been able to solve many of these issues, and we've successfully trained larger and more capable models, as you've seen through GPT-5 Soul and GPT-6 Astra.

s105
00:11:16.680 --> 00:11:21.560
Throughout these years, researchers have raised concerns about potential barriers to further scaling.

s106
00:11:21.560 --> 00:11:29.800
For example, in 2024, many researchers were concerned about a potential data wall, where we would run out of human-generated text to train on.

s107
00:11:29.800 --> 00:11:40.040
But techniques like synthetic data generation, learning from verifiable rewards, as well as algorithmic improvements to learn more from each individual token, have essentially solved this problem.

s108
00:11:40.040 --> 00:11:46.920
We've also solved concerns around things like context window length using things like sparse attention and compaction.

s109
00:11:47.080 --> 00:11:55.280
Looking ahead to 2027, I strongly believe that the trend lines of the past will continue and that we'll train larger, more efficient, and more capable models.

s110
00:11:55.280 --> 00:12:07.400
We clearly still have very difficult problems to solve, especially around things like alignment and safety, but like we have in the past, I believe that human ingenuity, increasingly assisted by capable models, will help us make progress on those problems, too.

s111
00:12:07.400 --> 00:12:10.440
So, how good will models be at the end of 2027?

s112
00:12:10.440 --> 00:12:12.480
Let's use math as an example.

s113
00:12:12.480 --> 00:12:18.200
In 2023, GPT-4 scored 700 out of 800 on SAT high school math.

s114
00:12:18.200 --> 00:12:25.320
By 2005, an OpenAI model reached gold medal status on the International Math Mathematical Olympiad.

s115
00:12:25.320 --> 00:12:34.880
And in 20 In early 2026, an internal model disproved a long-standing conjecture related to a problem that mathematicians have been working on since 1946.

s116
00:12:34.880 --> 00:12:48.000
More recently, an internal model proposed a solution to the Navier-Stokes Millennium Prize puzzle um problem, and we also saw shared 719 mathematical research manuscripts across 372 families of results.

s117
00:12:48.000 --> 00:12:55.440
In just a few years, we've gone from a strong score in a high school math exam to AI systems producing original research at the frontier of mathematics.

s118
00:12:55.440 --> 00:13:04.680
If we see another year of progress roughly in line with what we've been seeing over the past couple years, I expect that by end of 2027, models will outperform even leading human experts

s119
00:13:04.680 --> 00:13:08.920
on a growing range of difficult tasks across different fields and specialties.

s120
00:13:08.920 --> 00:13:13.120
Mathematics gives us an early glimpse of what that could look like.

s121
00:13:13.440 --> 00:13:17.680
And while the capability curve may look gradual, the change in how we work can feel sudden.

s122
00:13:17.680 --> 00:13:19.880
I call this the threshold phenomenon.

s123
00:13:19.880 --> 00:13:25.760
For any given task, there's a threshold of usefulness, the point where a model becomes capable and reliable enough to save you time.

s124
00:13:25.760 --> 00:13:30.920
That threshold depends on the work you're doing and how much checking or human review it requires.

s125
00:13:30.920 --> 00:13:33.760
Below that threshold, improvements can be hard to appreciate.

s126
00:13:33.760 --> 00:13:38.080
The model might produce better answers, make fewer mistakes, or get further through a task.

s127
00:13:38.080 --> 00:13:43.520
But if checking and fixing its work still takes as long as doing it yourself, your practical assessment hasn't changed.

s128
00:13:43.520 --> 00:13:45.280
It's still not useful.

s129
00:13:45.280 --> 00:13:49.480
Meaningful improvements in capability can feel like like little progress.

s130
00:13:49.480 --> 00:13:51.400
Then the model crosses that threshold.

s131
00:13:51.400 --> 00:13:53.400
Suddenly, it produces something you can use.

s132
00:13:53.400 --> 00:13:57.200
The time you spend reviewing it becomes much smaller than the time it saves you.

s133
00:13:57.200 --> 00:14:01.400
You start using it regularly, and eventually you become comfortable delegating that task.

s134
00:14:01.400 --> 00:14:06.880
That transition can feel like an enormous leap, even if the underlying improvement was similar to the previous few improvements.

s135
00:14:06.880 --> 00:14:10.080
What changed is the progress became useful to you.

s136
00:14:10.080 --> 00:14:16.600
This helps explain why model progress sometimes feels sudden, even when parts of the underlying trend have been steady and predictable.

s137
00:14:16.600 --> 00:14:22.840
And because difficult tasks have different thresholds, we should expect those moments to arrive at different times across the work that we do.

s138
00:14:22.840 --> 00:14:31.520
So, as you think about the products you're building, I'd encourage you to ask yourself, what model capabilities am I under appreciating today because they aren't quite useful yet?

s139
00:14:31.520 --> 00:14:34.720
Where might those capabilities be a year from now?

s140
00:14:34.720 --> 00:14:41.000
Then think about what you would build if those capabilities crossed the threshold of usefulness and became reliable enough to delegate.

s141
00:14:41.000 --> 00:14:43.440
How would the interface change?

s142
00:14:43.440 --> 00:14:46.839
What would the user still need to control, review, or approve?

s143
00:14:46.839 --> 00:14:50.440
Those are product and design questions we can start working on now.

s144
00:14:50.440 --> 00:15:00.000
We want to have the right experiences ready so that when the models cross that threshold, people can take advantage of the capabilities right away, and so you can stay one step ahead of your competitors.

s145
00:15:00.480 --> 00:15:02.360
So, what does this mean for finance?

s146
00:15:02.360 --> 00:15:06.400
I think there are two forces that will push more financial work across that threshold of usefulness.

s147
00:15:06.400 --> 00:15:09.760
I think of them as a rising tide and the waves on top.

s148
00:15:09.760 --> 00:15:13.880
The rising tide is the underlying improvements to the underlying model.

s149
00:15:13.880 --> 00:15:19.400
Historically, we haven't dedicated a significant amount of training time or compute specifically to finance.

s150
00:15:19.400 --> 00:15:29.680
Much of the financial capabilities you see today have come from building larger, more capable models, and from transferable skills developed in areas like mathematics, coding, and reasoning.

s151
00:15:29.680 --> 00:15:35.280
As these general capabilities improved, they lift the performance across many kinds of financial work.

s152
00:15:35.280 --> 00:15:38.280
The ways are what happen when we focus directly on finance.

s153
00:15:38.280 --> 00:15:42.800
My team is leading an effort to invest significantly more in finance specific post training.

s154
00:15:42.800 --> 00:15:48.880
The idea here is to build on those broad capabilities and train the model more directly on the work people do in finance.

s155
00:15:48.880 --> 00:15:53.240
That can create peaks in performance on particular tasks on top of the rising tide.

s156
00:15:53.240 --> 00:16:04.320
Over the next 3 months, we'll focus on improving general financial reasoning and the quality of the work these models produce, analyzing financial information, building Excel models, and creating PowerPoint presentations.

s157
00:16:04.320 --> 00:16:13.720
These are the kinds of tasks that make up a significant percentage of the daily work of a junior analyst, and we want the output to be accurate, well-structured, and increasingly useful in real workflows.

s158
00:16:13.720 --> 00:16:20.280
Beyond that, we plan to invest more in the kinds of judgment that senior bankers, analysts, and investors bring to more difficult questions.

s159
00:16:20.280 --> 00:16:25.880
This includes taking in large amounts of information, deciding what matters, and essentially trying to predict the future.

s160
00:16:25.880 --> 00:16:33.200
Forecasting a company's earnings, assessing where commodity prices might go, or evaluating which businesses are attractive investments.

s161
00:16:33.200 --> 00:16:40.800
Our goal is for models to match the the performance of the best human experts on these types of financial tasks sometime in 2027.

s162
00:16:40.800 --> 00:16:43.839
We expect that progress to come from both forces working together.

s163
00:16:43.839 --> 00:16:49.800
The underlying model continuing to improve, and our team training it more directly for the work people do in finance.

s164
00:16:49.800 --> 00:16:58.920
We talked about how models are becoming more capable and how we can train them to perform better on specific tasks, but I want to come back to one third driver I mentioned earlier, unhobbling.

s165
00:16:58.920 --> 00:17:00.720
We've already seen few examples.

s166
00:17:00.720 --> 00:17:06.760
Chat GPT used RLHF to teach models to follow instructions and have conversations with us.

s167
00:17:06.760 --> 00:17:11.760
Tools allow them to search for information, write and run code, and interact with the software we use.

s168
00:17:11.760 --> 00:17:14.839
Each of these makes the model more useful in the real world.

s169
00:17:14.839 --> 00:17:19.640
I think of the next form of unhollowing will be proactive and persistent agents.

s170
00:17:19.720 --> 00:17:24.120
Today a lot of our interactions with AI still happen when we open a conversation and tell it what to do next.

s171
00:17:24.120 --> 00:17:29.200
A persistent agent can hold on to a goal, remember the relevant context, and continue working in the background.

s172
00:17:29.200 --> 00:17:34.040
Once you've given it a direction, it can make progress even when you aren't actively prompting it.

s173
00:17:34.040 --> 00:17:39.160
The other reason persistence matters is it changes the API or the way people work with the agent.

s174
00:17:39.160 --> 00:17:45.640
Today getting the most out of AI agents requires learning a new workflow, learning how to prompt it, and learning how to use it to tools.

s175
00:17:45.640 --> 00:17:48.440
A persistent agent can work in a way that's more familiar.

s176
00:17:48.440 --> 00:17:52.120
You can give it a project, share context, and ask for an update.

s177
00:17:52.120 --> 00:18:00.520
It's much more similar to how humans work with other humans, and therefore makes the AI more accessible to people um that don't develop those initial skills.

s178
00:18:00.680 --> 00:18:03.800
Let me close with three bets for finance in 2027.

s179
00:18:03.800 --> 00:18:07.880
These are just a couple of my expectations for what comes as the models improve.

s180
00:18:07.880 --> 00:18:14.480
First, people will rely rarely, if ever, on Excel and PowerPoint to make financial models or pitch decks.

s181
00:18:14.480 --> 00:18:17.960
Second, AI financial analysts will start to feel like human coworkers.

s182
00:18:17.960 --> 00:18:25.760
They'll remember your preferences and context from the previous days, take ownership of assignments, proactively follow up, and keep working when you're focused on something else.

s183
00:18:25.760 --> 00:18:30.360
Third, models will start to outperform the best human experts on critical financial tasks.

s184
00:18:30.360 --> 00:18:33.880
That includes things like forecasting earnings and developing investment ideas.

s185
00:18:33.880 --> 00:18:39.200
I expect models to become better than the best people at making these economically important forward-looking judgments.

s186
00:18:39.200 --> 00:18:41.960
Those are my three bets for 2027.

s187
00:18:41.960 --> 00:18:46.400
And if this is where we expect to be in 2027, I'll leave you with one last question.

s188
00:18:46.400 --> 00:18:52.280
What would What should you start building today to be ready for inevitable thresholds that the models will cross tomorrow?

s189
00:18:52.280 --> 00:18:54.440
Thank you.

s190
00:18:54.473 --> 00:18:56.473
[applause]