# Building the Next Chapter of AI in Finance

Joseph Kim · AI Engineer NYC 2026 · October 14

[Watch recording](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4203s)

[01:10:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4203s) 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.

[01:10:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4221s) Model intelligence is continuing to improve at an incredible pace.

[01:10:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4225s) But many people working industries like finance haven't yet experienced that same change in their day-to-day work.

[01:10:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4232s) I expect that to start changing rapidly over the next few months.

[01:10:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4236s) And builders like all of you in this room have a big role in making that happen quickly.

[01:10:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4241s) Taking the intelligence in these models and turning it into products that people can use in their daily work.

[01:10:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4247s) Today, I will share a little bit about what we learned building ChatGPT for financial services

[01:10:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4251s) and how we're thinking about what comes next.

[01:10:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4255s) I'll start by showing you what we've built, including a quick demo.

[01:10:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4259s) Then I'll explain why we built it and why we think now is the right time to build.

[01:11:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4263s) 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.

[01:11:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4270s) Finally, we'll take a look at where models are headed and products are going

[01:11:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4273s) and what that can mean for financial work over the next year.

[01:11:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4277s) Let's start with what we've built.

[01:11:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4281s) ChatGPT for Financial Services is a tailored version of ChatGPT work built specifically

[01:11:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4286s) for financial professionals, starting with investment banking and equity research.

[01:11:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4290s) It's designed to help people at the large financial institutions get more out of their

[01:11:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4295s) day-to-day work, similar to how Codex speeds you up in software development.

[01:11:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4299s) As we built the product with our design partners, Morgan Stanley and Evercore, we focused on

[01:11:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4304s) three core areas.

[01:11:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4305s) First, financial data.

[01:11:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4308s) Financial data is the foundation of financial work,

[01:11:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4310s) and the work you can do is limited by the quality of your sources.

[01:11:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4314s) We wanted people to have access to the data they need directly within ChatGPT.

[01:11:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4319s) Second, an interface built for finance.

[01:12:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4322s) We wanted the product to feel familiar to financial professionals

[01:12:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4325s) with ways to inspect source material, explore financial charts,

[01:12:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4329s) and control the sources behind an analysis.

[01:12:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4332s) Third, your firm's format.

[01:12:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4335s) Financial work happens within the context of an organization,

[01:12:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4338s) and the models and presentations people produce

[01:12:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4340s) need to follow their forms format and standards.

[01:12:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4343s) We made it easier to create those artifacts using firm-specific templates.

[01:12:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4348s) But they say a picture is worth a thousand words,

[01:12:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4350s) so I'll stop talking for a bit and show you the product.

[01:14:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4443s) Wonderful.

[01:14:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4443s) Now I'd like to take some time to talk about why we built it.

[01:14:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4448s) We built ChatGPT for financial services to help close what we call the capability overhang.

[01:14:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4453s) As you can see here, model capabilities have improved at a tremendous rate over the last few years.

[01:14:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4459s) In just a few years, we've gone from models solving high school math problems

[01:14:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4462s) to systems capable of solving millennial-prized math problems.

[01:14:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4465s) But the value people get from these models in their day-to-day work hasn't kept pace.

[01:14:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4469s) You read about the breakthroughs in the news, yet for many financial professionals,

[01:14:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4473s) the way they work hasn't changed nearly as quickly.

[01:14:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4476s) We call that gap the capability overhang, the distance between what the model is capable of

[01:14:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4480s) and the value people can actually get from it.

[01:14:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4483s) Closing that gap takes more than building a capable model.

[01:14:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4486s) People need the right data and tools and they need an experience that fits the way they

[01:14:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4490s) work.

[01:14:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4491s) That's where builders like you come in.

[01:14:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4493s) I want to show you an example of how much of a difference the right product experience

[01:14:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4496s) can make.

[01:14:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4498s) This is a chart that shows how codex usage grew within OpenAI across different job functions.

[01:15:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4504s) As you can see, engineers started using codex early and that share grew steadily.

[01:15:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4509s) But look at finance, legal, and recruiting.

[01:15:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4512s) At the beginning of 2026, almost none of their AI output was generated in Codex.

[01:15:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4517s) By June, that share was around 90%.

[01:15:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4520s) If you look carefully, you'll notice a huge inflection point that happened around February, March of this year.

[01:15:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4525s) I went back to see what happened around this time, assuming that we must have released a new model,

[01:15:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4530s) but instead I found that we had released the Codex desktop app in February.

[01:15:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4534s) My takeaway is that the product made it easier for people outside of engineering to put these capabilities to work.

[01:15:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4540s) That's the opportunity we're pursuing in finance,

[01:15:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4542s) building an experience that helps financial professionals get much more value from the intelligence already available to them.

[01:15:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4549s) Now let me talk about how we built it.

[01:15:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4553s) Let's start with data.

[01:15:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4554s) In finance, the work you can do is limited by the quality of the information you have access to.

[01:15:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4559s) So we wanted the product to come with the sources professionals rely on,

[01:16:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4562s) from financial statements and earning transcripts to market and private company data.

[01:16:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4566s) We looked at existing MCP connectors for financial data

[01:16:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4569s) and found that many had issues with latency, errors, or search quality.

[01:16:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4574s) So we went and licensed a lot of this data directly

[01:16:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4576s) and built our own hosting and search indices

[01:16:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4578s) using the same infrastructure as our web index.

[01:16:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4581s) We still use the MCP protocol to give the model access to the data,

[01:16:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4585s) but hosting and indexing it ourselves gives us more control over the quality

[01:16:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4589s) and the reliability of the experience.

[01:16:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4591s) It also gives us the flexibility to post-train our models to work with these data sets.

[01:16:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4596s) The analogy I like to use is MCP and skills are like handing an instruction manual to

[01:16:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4601s) a junior analyst versus post-training where the model is like having a senior analyst

[01:16:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4605s) who's been working with the data for over two decades and knows all of its quirks like

[01:16:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4609s) the back of their hand.

[01:16:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4611s) MCP and skills give the model access to instructions.

[01:16:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4614s) Post-training can help it develop more familiarity with the data itself.

[01:16:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4618s) Once we had that data foundation in place, the next step was making it easy to turn that

[01:17:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4622s) information to professional-grade work.

[01:17:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4625s) For financial professionals, that work often takes the form of an Excel model, a PowerPoint

[01:17:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4629s) presentation, or a written analysis that needs to be reviewed and shared with other people.

[01:17:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4634s) So we focused on helping the model produce those outputs in a way that fits the firm's

[01:17:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4637s) existing workflows.

[01:17:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4639s) We built a template system that gives users access to the firm's approved templates when

[01:17:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4644s) generating spreadsheets and presentations.

[01:17:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4646s) The goal is to reduce the work someone has to do to get an output ready for their team

[01:17:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4650s) or clients.

[01:17:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4652s) We also added things like interactive financial charts so users can explore stock prices and

[01:17:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4656s) financial metrics directly within ChatGPT as they work through an analysis.

[01:17:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4661s) And we built citation viewers that let people inspect the source passages and tables behind

[01:17:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4666s) an answer.

[01:17:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4667s) In finance, you need to be able to check where a number came from and understand the context

[01:17:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4671s) around it.

[01:17:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4672s) We wanted that review to be a natural part of the experience.

[01:17:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4675s) These are the kind of product details that help turn a model's capabilities into work

[01:17:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4678s) that someone can use.

[01:18:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4681s) The part I think that is most relevant for people in the room is how we built it.

[01:18:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4685s) Much of the functionality you just saw was built using the same ChatGPT plugin platform

[01:18:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4689s) that's available to you as developers.

[01:18:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4692s) For financial data, we used MCP tools and connectors to give the model access to the

[01:18:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4696s) information it needs.

[01:18:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4697s) We use citation metadata in MCP apps to let users inspect their sources, and MCP apps also power interactive financial charts.

[01:18:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4705s) We use skills together with shared artifact tools to guide financial workflows and produce Excel models and PowerPoint presentations.

[01:18:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4712s) And for financial onboarding, we leverage the native plugin onboarding experience.

[01:18:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4717s) What that means for builders is that you can focus on the expertise that makes your product valuable.

[01:18:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4721s) If you have unique data, an existing tool, or a deep understanding of a particular workflow,

[01:18:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4727s) you can use these building blocks to bring that into ChatGPT and create an experience around it.

[01:18:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4732s) Financial services is just one example, but I do think this opportunity expands far beyond finance.

[01:18:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4738s) I personally believe that we're going to see multiple $1 billion businesses built simply as ChatGPT plugins.

[01:19:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4744s) And that opportunity will keep evolving as the models become more capable.

[01:19:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4749s) Last but not least, I'd like to spend the rest of the talk on what comes next and what

[01:19:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4754s) that means for the products we should be building today.

[01:19:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4758s) I'd like to start by talking about model capabilities and expectations for the future.

[01:19:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4763s) Models are improving so quickly that it can be difficult to get an intuitive sense of

[01:19:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4766s) what comes next.

[01:19:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4768s) One useful way to build that intuition is to look backwards.

[01:19:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4772s) Something remarkable about the last couple of years is how predictable parts of that

[01:19:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4776s) progress have been. Model improvements have come from three broad, sometimes overlapping drivers.

[01:19:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4781s) Scaling compute, improving algorithms, and something we call unhobbling, which is about

[01:19:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4787s) finding ways to get more useful work out of the models. We'll come back to unhobbling on a later

[01:19:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4791s) slide. In 2002 and 2003, the most visible driver was pre-training, which meant scaling compute by

[01:19:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4798s) making the models bigger and adding more data.

[01:20:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4800s) 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

[01:20:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4825s) eight months. Although you can't predict the timing of each individual breakthrough, the

[01:20:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4829s) cumulative rate of improvement follows a surprisingly steady trend. In 2024 and 2025, reasoning

[01:20:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4836s) became another major driver of progress. We used reinforcement learning and post-training

[01:20:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4840s) to teach models to work through problems more effectively, and we gave them more compute

[01:20:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4844s) to think when answering. That opened up another way to improve performance alongside pre-training.

[01:20:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4850s) However, for folks that don't know, at OpenAI, we ran into problems with pre-training and

[01:20:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4854s) and therefore had difficulty scaling the model.

[01:20:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4857s) In 2026, we've been able to solve many of these issues, and we've successfully trained

[01:21:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4861s) larger and more capable models, as you've seen through GPD-56 Sol and GPD-6 Astro.

[01:21:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4867s) Throughout these years, researchers have raised concerns about potential barriers to further

[01:21:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4871s) scaling.

[01:21:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4872s) For example, in 2024, many researchers were concerned about a potential data wall where

[01:21:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4877s) we would run out of human-generated text to train on.

[01:21:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4880s) But techniques like synthetic data generation,

[01:21:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4883s) learning from verifiable rewards,

[01:21:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4885s) as well as algorithmic improvements

[01:21:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4887s) to learn more from each individual token

[01:21:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4888s) have essentially solved this problem.

[01:21:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4891s) We've also solved concerns around things

[01:21:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4893s) like context window length

[01:21:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4894s) using things like sparse detention and compaction.

[01:21:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4898s) Looking ahead to 2027,

[01:21:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4900s) I strongly believe that the trend lines of the past

[01:21:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4902s) will continue and that we'll train larger,

[01:21:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4904s) more efficient, and more capable models.

[01:21:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4906s) We clearly still have very difficult problems to solve,

[01:21:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4908s) especially around things like alignment and safety,

[01:21:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4910s) but like we have in the past, I believe that human ingenuity,

[01:21:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4913s) increasingly assisted by capable models,

[01:21:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4916s) will help us make progress on those problems too.

[01:21:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4918s) So how good will models be at the end of 2027?

[01:22:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4921s) Let's use math as an example.

[01:22:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4923s) In 2023, GPT-4 scored 700 out of 800 on SAT high school math.

[01:22:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4929s) By 2005, an opening eye model reached gold medal status

[01:22:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4932s) on the International Mathematical Olympiad.

[01:22:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4936s) And in early 2026, an internal model disproved a long-standing conjecture related to a problem that mathematicians have been working on since 1946.

[01:22:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4945s) More recently, an internal model proposed a solution to the Navier-Stokes Millennium Prize puzzle problem,

[01:22:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4952s) and we also shared 719 mathematical research manuscripts across 372 families of results.

[01:22:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4959s) In just a few years, we've gone from a strong score on a high school math exam

[01:22:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4962s) to AI systems producing original research at the frontier of mathematics.

[01:22:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4966s) If we see another year of progress roughly in line with what we've been seeing over the past couple years,

[01:22:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4971s) I expect that by end of 2027, models will outperform even leading human experts

[01:22:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4975s) on a growing range of difficult tasks across different fields and specialties.

[01:23:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4980s) Mathematics gives us an early glimpse of what that could look like.

[01:23:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4984s) And while the capability curve may look gradual, the change in how we work can feel sudden.

[01:23:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4988s) I call this the threshold phenomenon.

[01:23:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4991s) For any given task, there's a threshold of usefulness, the point where a model becomes

[01:23:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4994s) capable and reliable enough to save you time.

[01:23:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4997s) That threshold depends on the work you're doing and how much checking or human review

[01:23:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5000s) it requires.

[01:23:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5002s) Below that threshold, improvements can be hard to appreciate.

[01:23:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5005s) The model might produce better answers, make fewer mistakes, or get further through a task.

[01:23:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5009s) But if checking and fixing its work still takes as long as doing it yourself,

[01:23:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5012s) your practical assessment hasn't changed.

[01:23:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5014s) It's still not useful.

[01:23:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5016s) Meaningful improvements in capability can feel like little progress.

[01:23:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5020s) Then the model crosses that threshold.

[01:23:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5022s) Suddenly it produces something you can use.

[01:23:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5024s) The time you spend reviewing it becomes much smaller than the time it saves you.

[01:23:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5028s) You start using it regularly,

[01:23:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5030s) and eventually you become comfortable delegating that task.

[01:23:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5032s) That transition can feel like an enormous leap,

[01:23:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5034s) even if the underlying improvement was similar to the previous few improvements.

[01:23:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5038s) What changes of progress became useful to you.

[01:24:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5041s) This helps explain why model progress sometimes feels sudden,

[01:24:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5044s) even when parts of the underlying trend have been steady and predictable.

[01:24:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5047s) And because difficult tasks have different thresholds,

[01:24:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5050s) we should expect those moments to arrive at different times across the work that we do.

[01:24:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5053s) So as you think about the products you're building,

[01:24:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5056s) I'd encourage you to ask yourself,

[01:24:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5057s) what model capabilities am I underappreciating today because they aren't quite useful yet?

[01:24:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5062s) Where might those capabilities be a year from now?

[01:24:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5065s) Then think about what you would build if those capabilities crossed the threshold of usefulness

[01:24:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5069s) and became reliable enough to delegate?

[01:24:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5072s) How would the interface change?

[01:24:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5074s) What would the users still need to control, review, or approve?

[01:24:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5078s) Those are product and design questions we can start working on now.

[01:24:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5081s) We want to have the right experiences ready

[01:24:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5083s) so that when the models cross that threshold,

[01:24:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5085s) people can take advantage of the capabilities right away,

[01:24:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5087s) and so you can stay one step ahead of your competitors.

[01:24:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5091s) So what does this mean for finance?

[01:24:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5093s) I think there are two forces that will push more financial work

[01:24:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5096s) across that threshold of usefulness.

[01:24:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5097s) I think of them as the rising tide and the waves on top.

[01:25:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5100s) The rising tide is the underlying improvements to the underlying model.

[01:25:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5105s) Historically, we haven't dedicated a significant amount of training time or compute specifically to finance.

[01:25:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5110s) Much of the financial capabilities you see today have come from building larger, more capable models

[01:25:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5115s) and from transferable skills developed in areas like mathematics, coding, and reasoning.

[01:25:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5120s) As these general capabilities improved, they lifted performance across many kinds of financial work.

[01:25:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5126s) The Ways are what happens when we focus directly on finance.

[01:25:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5129s) My team is leading an effort to invest significantly more in finance-specific post-training.

[01:25:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5133s) The idea here is to build on those broad capabilities

[01:25:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5136s) and train the model more directly on the work people do in finance.

[01:25:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5140s) That can create peaks in performance on particular tasks on top of the rising tide.

[01:25:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5144s) Over the next three months, we're focused on improving general financial reasoning

[01:25:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5148s) and the quality of the work these models produce,

[01:25:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5150s) analyzing financial information, building Excel models,

[01:25:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5153s) and creating PowerPoint presentations.

[01:25:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5155s) These are the kinds of tasks that make up a significant percentage of the daily work of a junior analyst,

[01:25:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5159s) and we want the output to be accurate, well-structured, and increasingly useful in real workflows.

[01:26:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5164s) Beyond that, we plan to invest more in the kinds of judgment that senior bankers, analysts, and investors bring to more difficult questions.

[01:26:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5171s) This includes taking in large amounts of information, deciding what matters, and essentially trying to predict the future,

[01:26:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5177s) forecasting a company's earnings, assessing where commodity prices might go, or evaluating which businesses are attractive investments.

[01:26:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5184s) Our goal is for models to match the performance of the best human experts

[01:26:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5188s) on these types of financial tasks sometime in 2027.

[01:26:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5191s) We expect that progress to come from both forces working together,

[01:26:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5195s) the underlying model continuing to improve,

[01:26:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5197s) and our team training it more directly for the work people do in finance.

[01:26:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5201s) We've talked about how models are becoming more capable

[01:26:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5203s) and how we can train them to perform better on specific tasks,

[01:26:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5206s) but I want to come back to one third driver I mentioned earlier unhobbling.

[01:26:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5210s) We've already seen a few examples.

[01:26:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5213s) ChatGPT use RLHF to teach models to follow instructions and have conversations with us.

[01:26:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5218s) Tools allow them to search for information, write and run code, and interact with the software we use.

[01:27:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5223s) Each of these makes the model more useful in the real world.

[01:27:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5226s) I think the next form of unhobbling will be proactive and persistent agents.

[01:27:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5231s) Today, a lot of our interactions with AI still happen when we open a conversation and tell it what to do next.

[01:27:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5235s) A persistent agent can hold on to a goal, remember the relevant context, and continue working in the background.

[01:27:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5240s) Once you've given it a direction, it can make progress even when you aren't actively prompting it.

[01:27:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5245s) The other reason persistence matters is it changes the API or the way people work with the agent.

[01:27:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5250s) 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.

[01:27:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5256s) A persistent agent can work in a way that's more familiar.

[01:27:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5259s) You can give it a project, share context, and ask for an update.

[01:27:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5263s) It's much more similar to how humans work with other humans, and therefore makes the AI more accessible to people

[01:27:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5268s) that don't develop those initial skills.

[01:27:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5271s) Let me close with three bets for finance in 2027.

[01:27:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5274s) These are just a couple of my expectations

[01:27:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5276s) for what comes as the models improve.

[01:27:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5279s) First, people will rely rarely, if ever,

[01:28:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5282s) on Excel and PowerPoint

[01:28:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5282s) to make financial models or pitch decks.

[01:28:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5285s) Second, AI financial analysts

[01:28:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5287s) will start to feel like human coworkers.

[01:28:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5289s) They'll remember your preferences and context

[01:28:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5290s) from the previous days,

[01:28:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5292s) take ownership of assignments,

[01:28:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5293s) productively follow up,

[01:28:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5294s) and keep working when you're focused on something else.

[01:28:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5297s) Third, models will start to outperform the best human experts on critical financial tasks.

[01:28:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5301s) That includes things like forecasting earnings and developing investment ideas.

[01:28:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5305s) I expect models to become better than the best people at making these economically important, forward-looking judgments.

[01:28:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5310s) Those are my three bets for 2027.

[01:28:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5313s) And if this is where we expect to be in 2027, I'll leave you with one last question.

[01:28:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5317s) What should you start building today to be ready for inevitable thresholds that the models will cross tomorrow?

[01:28:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=5323s) Thank you.