Stanislas Polu
Last reviewed
Jun 8, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,471 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,471 words
Add missing citations, update stale details, or suggest a clearer explanation.
Stanislas Polu is a French software engineer, AI researcher, and entrepreneur. He is the co-founder and chief technology officer of Dust, a Paris-based company that builds an enterprise platform for AI agents. Before founding Dust in 2022, Polu spent roughly three years as a research engineer at OpenAI, where he led work on applying large language models to automated theorem proving and formal mathematics, including the GPT-f prover and the miniF2F benchmark. Earlier in his career he was an early engineer at the payments company Stripe, which he joined when it acquired his first startup, Totems.[1][2][3]
Polu grew up in France and passed through the country's selective preparatory-class system, completing intensive coursework in advanced mathematics and physics from 2002 to 2004. He then attended École Polytechnique, one of France's leading engineering schools, from 2004 to 2007 with a specialization in computer science, and went on to earn a master's degree in computer science from Stanford University, which he completed in early 2009. His early roles included a software engineering internship at Apple and a position as a member of technical staff at Oracle, where he worked on the company's Clusterware software.[4]
By 2011 Polu had moved into entrepreneurship, co-founding a startup called Totems with Gabriel Hubert, whom he had met at Stanford. Totems built a marketing and analytics suite for brands on Instagram, helping companies measure and grow their social-media presence.[5][6] Stripe acquired the company in an acqui-hire announced in early 2015, though some sources date the deal to 2014, and its Instagram tools were wound down by the end of March 2015 as the team moved onto payments.[6] Polu then worked as an engineer at Stripe from 2015 to 2019, helping scale the company's products and engineering teams before turning to artificial intelligence research.[1][5]
Polu joined OpenAI around 2019 as a research engineer, working on the laboratory's effort to use large language models for mathematical reasoning. His focus was formal mathematics: the practice of writing proofs in machine-checkable languages so that a computer can verify every step.[1][7]
In September 2020, Polu and OpenAI co-founder Ilya Sutskever published "Generative Language Modeling for Automated Theorem Proving," which introduced GPT-f, a prover built on the GPT-3 architecture for the Metamath formal system. A 700-million-parameter model proved about 31.6 percent of a held-out test set, improving on the prior state of the art, and several of the short proofs it discovered were accepted into Metamath's main library. OpenAI described this as the first time a deep-learning system contributed proofs that were adopted by a formal-mathematics community.[3][7]
Polu went on to work with the Lean proof assistant. He co-authored "Proof Artifact Co-Training" (2021) and miniF2F (2021), a cross-system benchmark of 488 Olympiad-level problems spanning Metamath, Lean, Isabelle, and HOL Light that became a standard yardstick for neural theorem provers.[8][9] In February 2022 he was first author of "Formal Mathematics Statement Curriculum Learning," with Jesse Michael Han, Kunhao Zheng, Mantas Baksys, Igor Babuschkin, and Sutskever. The paper showed that expert iteration, alternating proof search with model training, let a model bootstrap from easier to harder problems and reach a new state of the art on miniF2F, solving several problems adapted from the AMC, AIME, and International Mathematical Olympiad competitions.[10][11]
| Year | Paper | System | Known for |
|---|---|---|---|
| 2020 | Generative Language Modeling for Automated Theorem Proving | Metamath | Introduced GPT-f; proofs adopted into the Metamath library |
| 2021 | Proof Artifact Co-Training | Lean | Co-training on proof artifacts to improve theorem proving |
| 2021 | miniF2F | Multi-system | Cross-system Olympiad benchmark of 488 problems |
| 2022 | Formal Mathematics Statement Curriculum Learning | Lean | Expert iteration; state of the art on miniF2F |
His theorem-proving papers are among his most cited; as of 2026 his Google Scholar profile listed more than 2,200 citations.[12] After leaving OpenAI he stayed involved in AI for mathematics, co-authoring the 2024 NuminaMath dataset, one of the largest public collections of mathematical problem and solution pairs (around 860,000) used to train reasoning models.[12]
Polu left OpenAI in September 2022, convinced that frontier models were already capable enough to be economically useful but were held back by a missing product layer between the models and real company work.[2] He reunited with Hubert, by then a former head of product at the French insurance company Alan, to build that layer, and the two founded Dust. The company launched its first products by the end of 2022 and was incorporated in early 2023, with Hubert as chief executive officer and Polu as chief technology officer.[1][5]
Dust deliberately does not train its own models. Instead it builds a model-agnostic platform on top of frontier models from providers such as OpenAI and Cohere, connecting them to a company's internal knowledge through connectors for tools like Notion, Slack, GitHub, and Google Drive.[1] Early products included a developer framework for chaining model calls, positioned as an alternative to LangChain, and XP1, a browser-extension assistant with access to a user's open tabs.[5][13] The company later refocused on enterprise "agents," configurable assistants that draw on internal data and tools, and described its goal as a "multiplayer" operating system in which teams and software agents share a workspace rather than each employee using an isolated assistant.[2][14]
| Round | Date | Amount | Lead investors |
|---|---|---|---|
| Seed | June 2023 | 5 million euros (about 5.5 million dollars) | Sequoia Capital |
| Series A | June 2024 | 16 million dollars | Sequoia Capital |
| Series B | May 2026 | 40 million dollars | Sequoia Capital and Abstract |
Dust raised an initial seed round of 5 million euros led by Sequoia Capital in June 2023, followed by a 16-million-dollar Series A in June 2024 and a 40-million-dollar Series B co-led by Sequoia and Abstract in May 2026, with participation from Datadog and Snowflake Ventures, bringing total funding to more than 60 million dollars.[1][14][15] At the time of the Series B the company said it was used by more than 3,000 organizations, had crossed 20 million dollars in annual recurring revenue earlier in 2026, and counted customers including Qonto, Alan, Pennylane, and Datadog.[2][14]
Polu has argued that the main bottleneck to useful enterprise AI is not raw model capability but deployment: connecting general-purpose models to the specific data, tools, and workflows of a given company. Consistent with this, Dust is model-agnostic by design, on the premise that no single provider's model will dominate every task and that much of the value lies in the orchestration layer above the models.[2][14] Polu has framed the shift to "multiplayer" AI, where agents collaborate with whole teams rather than serving individuals in isolation, as the change most likely to transform how organizations work. His research on formal mathematics reflects a longer-standing interest in machine reasoning, and he has spoken publicly about the evolution of reasoning and agency in large language models, including at industry conferences.[16] Much of Dust's early tooling was released as open source.[13]