Thomas Wolf

Thomas Wolf graduated from École Polytechnique in Paris, France, and early in his career researched laser–plasma interactions at the BELLA Center of Lawrence Berkeley National Laboratory. By his own account, he was admitted to a Ph.D. program at MIT but ultimately pursued doctoral research in statistical and quantum physics at Sorbonne University and ESPCI in Paris, working on superconducting materials.⁠[1]

After his doctoral research, he joined the Paris intellectual property firm Cabinet Plasseraud and earned a law degree from Panthéon-Sorbonne University. As a patent attorney, he handled intellectual property matters for startups and large companies; this phase of his career lasted about five years.⁠[1]

In 2015, while advising AI startups, he came back into contact with machine learning and taught himself the field through books and online courses. In his personal bio, he links this shift to his background in statistical physics; he then helped found Hugging Face and returned to research and programming.⁠[2]

Hugging Face initially focused on a chat app and later shifted its focus to open-source models and tools for developers. In an interview with Sequoia Capital, Wolf recalled how the team built a community around Transformer technology and connected that experience with its later work in robotics.⁠[3]

In October 2019, Wolf published, as first author, a paper introducing the Transformers library. The project brought many Transformer architectures under a unified interface and paired them with a collection of pretrained models, making it easier for researchers to extend models, for developers to use them, and for companies to deploy them.⁠[4]

He then helped drive the BigScience collaboration. A Hugging Face technical post on training BLOOM credits Wolf with the idea for this open large-model project and lists engineering and research contributors from multiple institutions.⁠[5]

In July 2022, BigScience released BLOOM. The 176-billion-parameter model supports 46 natural languages and 13 programming languages, was trained through an international research collaboration, and was released openly for research and use; its use remains subject to the terms of its Responsible AI License.⁠[6]

In “The Einstein AI model,” Wolf questioned whether simply raising models’ scores on exams and benchmarks with known answers can lead to real scientific breakthroughs. He argued that scientific research also requires asking new questions and challenging existing knowledge, and that evaluation methods able to test such abilities should therefore be explored.⁠[7]

In the piece, he used his own experience to illustrate the difference between the abilities of a strong student and those of an independent researcher, and he presented the argument as a personal view. The view concerns how scientific questions are formed and does not amount to denying AI’s role in solving problems or assisting research.⁠[7]

In a September 2025 interview with Sequoia Capital, he described LeRobot’s approach of connecting models, datasets, and hardware, and discussed Hugging Face’s development of products such as Reachy Mini after acquiring Pollen Robotics, with the aim of enabling more software developers to take part in robotics.⁠[3]

He also noted that the scale and environmental diversity of robotics data remain a constraint: a motion learned in one room may not be reliably reproduced in another environment. Open data and community collaboration are the direction he proposed, not a conclusion that all generalization problems have been solved.⁠[3]

In July 2026, the Associated Press reported a cyber intrusion at Hugging Face linked to models being tested by OpenAI and quoted Wolf’s response. He said the incident highlighted the need for defenders to have timely access to capable open models; the report presents this as his argument about the availability of security tools.⁠[8]

On September 3, NVIDIA announced that it had agreed to acquire Hugging Face for about $12.93 billion. The announcement mentioned the work of Wolf and the other co-founders and said it planned to continue supporting an open ecosystem across clouds and computing platforms; it describes an acquisition agreement and should not be taken to mean that the deal has closed.⁠[9]

On September 21, Wolf discussed open-weight security models on his public X account, welcoming models that offer strong defensive capabilities and are easy to deploy. This continued his earlier emphasis on the importance of open tools for cyber defense.⁠[10]

Wolf took part as an angel investor in the first funding round of the AI lab Conway ResearchConway ResearchConway Research is an artificial intelligence lab founded by Sigil Wen that works on efficient models, inference engines, and personal AI that runs on devices users already own; its first product is Underdog, a locally run personal assistant. In October 2026, Andreessen Horowitz announced that it was leading the company’s first funding round.Open the full entry, completed in April 2026 and led by Andreessen Horowitz. On October 2, 2026, founder Sigil WenSigil WenSigil Wen is an engineer, entrepreneur, and angel investor from Toronto, Canada, the founder and executive chairman of the U.S. visa services company Extraordinary, and a 2025 Thiel Fellow. He was previously an early engineer at Airchat, and in 2026 he founded the AI and cryptography lab Conway Research; the autonomous AI agent Automaton and the “Web 4.0” concept he released sparked debate in the crypto community.Open the full entry launched the company’s first product, Underdog, a private personal AI that runs on users’ own devices, and named Wolf when announcing the company’s backers.⁠[11][11][12][13][13][13]

Hugging Face: The organization he co-founded and where he serves as chief science officer. Official website: https://huggingface.co/ (opens in a new window).⁠[2]

Personal website: https://thomwolf.io/ (opens in a new window). X: Public account: https://x.com/Thom_Wolf (opens in a new window). Newsletter: https://thomwolf.substack.com/ (opens in a new window).⁠[2][10]