Thomas Wolf

Physics and patent law

Thomas Wolf graduated from Ecole Polytechnique in Paris and then worked on laser-plasma interactions at the BELLA Center of the Lawrence Berkeley National Laboratory^[1]. He was accepted for a PhD at MIT but took his PhD in statistical and quantum physics at Sorbonne University and ESPCI in Paris, working on superconducting materials for the French defence procurement agency DGA and for Thales^[1].

After the PhD he joined the intellectual property law firm Cabinet Plasseraud in Paris, earned a law degree from Pantheon Sorbonne University and worked as a patent attorney for five years^[1]. In 2015, while consulting for deep learning and machine learning startups, he became interested in the mathematics behind the methods, which he saw as re-branded statistical physics approaches, and began teaching himself machine learning through books and online courses^[1]. About a year later he joined Hugging Face^[1].

Hugging Face and its open-source libraries

Wolf is co-founder and Chief Science Officer of Hugging Face, where he has worked on the company's open-source, open-science and robotics efforts^[1]. He has created or advised the Transformers, Datasets, Diffusers, Accelerate, DataTrove, smolagents and LeRobot libraries^[1].

He is the first author of Transformers: State-of-the-Art Natural Language Processing, the paper describing the Transformers library, published in the system demonstrations track of EMNLP 2020^[4]. He co-authored DistilBERT, a 2019 paper by Victor Sanh, Lysandre Debut, Julien Chaumond and Wolf, which showed that a BERT model could be reduced in size by 40% while keeping 97% of its language understanding capabilities and running 60% faster^[5]. His site also lists a Datasets library paper at EMNLP 2021 and the TransferTransfo paper on conversational agents from a NeurIPS 2018 workshop^[1].

Wolf took part in the BigScience Workshop on large language models, which led to the BLOOM model and dataset^[1]. The BLOOM paper describes a 176-billion-parameter open-access multilingual language model and lists 392 authors under the BigScience Workshop^[6]. He has also promoted open science through the FineWeb dataset and the Ultra-Scale Playbook, and co-authored the O'Reilly book Natural Language Processing with Transformers^[1].

Robotics and AI for science

Wolf has built Hugging Face's robotics effort, including projects around the Reachy Mini robot^[1]. In April 2025 Hugging Face acquired Pollen Robotics, a French humanoid robotics startup founded in 2016, which had worked with Hugging Face on the open-source LeRobot project^[7]. In July 2025 Wolf co-authored the Hugging Face blog post announcing Reachy Mini, an open-source robot programmable in Python and priced from $399^[8].

His stated research interest is whether AI systems can take part in generating new scientific knowledge rather than only speeding up known workflows, and he supports AI-for-science collaborations between academia and industry through huggingscience.co^[1]. He spoke at the HumanX conference in San Francisco in April 2026^[2].

Writing on AI and science

In a 2025 blog post titled The Einstein AI Model, Wolf argued that he did not expect AI to compress the 21st century's scientific progress into a few years^[1]. Speaking to Fortune at VivaTech in Paris in June 2025, he said current models are very bad at asking great questions, and that asking the question is the hard part of science^[3]. Fortune reported that he warned current models behave like agreeable yes-men on servers, unlikely to challenge existing assumptions^[3]. His site also lists a post responding to Anthropic CEO Dario Amodei's essay on DeepSeek and export controls, in which Wolf says he mostly disagrees^[1].

In a December 2025 Substack essay, What Jobs Are Made Of, he discussed research showing falling employment for workers aged 22 to 25 in AI-exposed occupations^[9]. He argued that judgment and agency are hard to measure in AI benchmarks but central to how workers create value, and that AI may place more weight on judgment, taste and agency in work^[9].

Known for

• Co-founding Hugging Face
• The Transformers and Datasets libraries
• The DistilBERT paper
• Hugging Face's open-source robotics work, including Reachy Mini

Sources

1. Thomas Wolf: personal website and short bio — checked 6 Nov 2026
2. HumanX: Thomas Wolf speaker page, San Francisco, April 2026 — checked 6 Nov 2026
3. Fortune: Hugging Face cofounder says AI more likely to create yes-men on servers, June 2025 — checked 6 Nov 2026
4. ACL Anthology: Transformers: State-of-the-Art Natural Language Processing, EMNLP 2020 — checked 6 Nov 2026
5. arXiv: DistilBERT, a distilled version of BERT, 2019 — checked 6 Nov 2026
6. arXiv: BLOOM: A 176B-Parameter Open-Access Multilingual Language Model, 2022 — checked 6 Nov 2026
7. TechCrunch: Hugging Face buys a humanoid robotics startup, April 2025 — checked 6 Nov 2026
8. Hugging Face blog: Reachy Mini, The Open-Source Robot, July 2025 — checked 6 Nov 2026
9. Thomas Wolf, Substack: What Jobs Are Made Of, December 2025 — checked 6 Nov 2026