Sébastien Bubeck

Research

Bubeck's early work was in learning theory and optimization. With Nicolò Cesa-Bianchi, Bubeck wrote "Regret Analysis of Stochastic and Nonstochastic Multi-armed Bandit Problems" (2012), a survey of bandit algorithms, and as sole author wrote "Convex Optimization: Algorithms and Complexity" (2014), a monograph on optimization methods and their complexity. "A Universal Law of Robustness via Isoperimetry" (2021), with Mark Sellke, offered a theoretical account of why deep networks need many more parameters than required to fit data smoothly, and received a NeurIPS 2021 Outstanding Paper Award.^10111213

Bubeck's later work turned to large language models. In 2023 Bubeck was first author of "Sparks of Artificial General Intelligence", a study of an early version of GPT-4, and co-authored the "Textbooks Are All You Need" reports on phi-1 and phi-1.5 and the later Phi-3 and Phi-4 reports on Microsoft's small Phi models, which emphasize training on carefully selected web data and synthetic data.⁹⁵⁴⁶⁷

Career

As a vice president in Microsoft GenAI, Bubeck co-authored the technical reports for phi-1, phi-1.5, Phi-3 and Phi-4 and contributed to Phi-2. With an OpenAI affiliation, Bubeck is among the authors of the 2025 model card for the open-weight gpt-oss models and first author of "Early science acceleration experiments with GPT-5" (2025), a collection of case studies in which GPT-5 contributed to research in mathematics, physics, astronomy, computer science, biology and materials science.³⁷²¹

Work and contributions

• Phi small language models Open weights

  Co-author of the phi-1 (2023), phi-1.5 (2023), Phi-3 (2024), and Phi-4 (2024) technical reports, and a listed contributor and author of the Microsoft Research blog post announcing Phi-2. The Phi-3 report describes phi-3-mini as small enough to be deployed on a phone, and the Phi-4 model card on Hugging Face lists the MIT license.⁴⁵⁶⁷³⁸

• gpt-oss Open weights

  Listed among the authors of the gpt-oss-120b and gpt-oss-20b model card (2025), which presents two open-weight reasoning models; the gpt-oss-20b model card on Hugging Face lists the Apache 2.0 license.²¹⁴

• Early science acceleration experiments with GPT-5 Published research

  First author of the 2025 collection of case studies in which GPT-5 contributed concrete steps to ongoing research in mathematics, physics, astronomy, computer science, biology, and materials science.¹

• Sparks of Artificial General Intelligence Published research

  First author of the 2023 paper reporting experiments with an early version of GPT-4 while it was still in development by OpenAI.⁹

Timeline

1. Apr 2012Posted a survey of regret analysis for multi-armed bandit problems with Nicolò Cesa-Bianchi.¹⁰
2. May 2014Posted the monograph "Convex Optimization: Algorithms and Complexity" as sole author.¹¹
3. May 2021Posted "A Universal Law of Robustness via Isoperimetry" with Mark Sellke.¹²
4. Mar 2023Posted "Sparks of Artificial General Intelligence: Early experiments with GPT-4" as first author.⁹
5. Jun 2023Co-authored "Textbooks Are All You Need", the phi-1 technical report.⁴
6. Apr 2024Co-authored the Phi-3 technical report on small language models that run on a phone.⁶
7. Dec 2024Co-authored the Phi-4 technical report.⁷
8. Aug 2025Listed among the authors of the gpt-oss-120b and gpt-oss-20b model card.²
9. Nov 2025Posted "Early science acceleration experiments with GPT-5" as first author.¹

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Listed with an OpenAI affiliation on the November 2025 paper "Early science acceleration experiments with GPT-5", OpenAI¹²
• Vice President, Microsoft GenAI, Microsoft³⁷

Honors and recognition

• NeurIPS 2021 Outstanding Paper Award ("A Universal Law of Robustness via Isoperimetry") · NeurIPS, 2021¹³

Companies in their history

• OpenAI 63 people in the wiki
• Microsoft 75 people in the wiki

Related people

• John SchulmanConnected to OpenAI
• Mark ChenConnected to OpenAI
• Shengjia ZhaoConnected to OpenAI
• Tom BrownConnected to OpenAI
• Diederik P. KingmaConnected to OpenAI
• Michael CollinsLanguage models · Machine-learning theory
• Beren MillidgeLanguage models · Machine-learning theory
• Fabian PedregosaMachine-learning theory · Language models
• Albert GuLanguage models · Machine-learning theory
• Stefano ErmonLanguage models · Machine-learning theory