Latent Space: The AI Engineer Podcast · Oct 30, 2026 · 28m · 19m saved
With Joseph Kim, Head of Financial Services at OpenAI
OpenAI tested well-known data connectors for ChatGPT and found some returned errors more than 40% of the time, Joseph Kim, Head of Financial Services at OpenAI, told swyx on Latent Space: The AI Engineer Podcast.
Most companies building on a model provider assume the provider holds an advantage they cannot copy. Joseph said nearly everything his team shipped could be built by an outside startup on the same public tools, and the part that decides success is the last 20%.
"I really do think that there's going to be a billion-dollar business built on top of a ChatGPT plugin. But 95% of what we built was something that absolutely a startup could do, another company could do." — Joseph Kim
Joseph runs the team that launched ChatGPT for Financial Services the month before this interview, with Morgan Stanley and Evercore as design partners. The conversation was recorded on the sidelines of his keynote at AI Engineer New York.
The full interview is covered here so you can skip it. 28 minutes of audio, 9 minutes of reading.
Here are the 10 insights that matter.
This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.
Key Takeaways
A startup could build 95% of ChatGPT for Financial Services on the same public plugin tools
The base model gets about 80% of the way; the last 20% decides whether it fails
Some well-known data connectors showed error rates above 40%, so OpenAI licensed the data itself
Financial reasoning goes into OpenAI's main model training run, not a separate finance model
Factual errors that change a conclusion are the career-ending mistakes, and models are making fewer of them
What models still lack is taste, like knowing when a slide looks ugly
Humans voting on pairs of slides trained ChatGPT's slide taste, and humans beat models at it
OpenAI's own financial evals were broken, not the model
Flat GPT-6 scores traced back to wrong answer keys
1. What OpenAI Launched
ChatGPT for Financial Services adds firm-specific templates and licensed data on top of the tools ChatGPT already offers consumers. Joseph named S&P and Daloopa among the data sources.
Morgan Stanley and Evercore worked with OpenAI as design partners. Joseph said customer reception has been strong, and licensed data is the feature customers mention most.
