Stanislas Polu

Stanislas Polu is co-founder of Dust, a Paris, France, company that lets businesses build AI agents connected to their own data and tools. He started it with Gabriel Hubert, a friend since they were students in France.

Nine Months at Oracle

Polu studied at École Polytechnique and then at Stanford, where Hubert also studied, at a time when Facebook had only a few hundred people. Instead of joining a fast-growing startup, he went to Oracle for nine months. "It was awesome, experiencing what the big kind of computer science company looks like," he told the MAD Podcast. He then moved back to France.

An Instagram Analytics Startup Sold to Stripe

With Hubert he started a company "for the exercise of starting a company," with no plan and no vision, he said. After several pivots it became Totems, an Instagram analytics platform for brands, which grew to "between $1 and $2 million ARR with a small team of eight or 10." Struggling to raise a Series A, the founders sold it to Stripe in 2014, when Stripe had about 80 people.

Polu stayed at Stripe for five years, worked with Greg Brockman, and built infrastructure to add non-card payment methods, which became the first version of the Sources API.

Mathematical Reasoning at OpenAI

Through Brockman, by then OpenAI's CTO, and Ilya Sutskever, he joined OpenAI to work on mathematical reasoning. He calls that period "the PhD I always wanted to do," he told Latent Space. By the end of summer 2022 he had reached a simple conclusion, he said on the MAD Podcast: "those models are already extremely useful. They're extremely powerful, and they are not deployed in the world at all." On Latent Space he said timing mattered too, calling that moment "the last train leaving the station to start a company."

Competing With LangChain in 2022

Dust began in 2022 as an open-source developer framework that competed with LangChain, then tried XP1, a browser extension, in early 2023, before settling on its enterprise product for building custom agents on company data. Polu has argued that the product layer helps explain why companies were slow to use large language models.