Thomas Wolf

Thomas Wolf is co-founder and Chief Science Officer of Hugging Face. He helped create the Transformers and Datasets libraries and co-authored research on DistilBERT, a smaller pretrained language model. His work combines making models and datasets accessible with a research interest in whether AI can generate new scientific knowledge.[2][4][6][5]

Wolf identifies Transformers as one of the libraries he created. He is the first-listed author of its research paper, first submitted in October 2019, alongside Lysandre Debut, Victor Sanh, Julien Chaumond, Clément Delangue and other collaborators. The library places Transformer architectures behind a unified API and pairs them with a curated collection of community pretrained models. Its design aims to accommodate researchers extending architectures, practitioners using them and industrial deployments requiring speed and robustness.[2][4]

With Victor Sanh, Lysandre Debut and Julien Chaumond, Wolf co-authored DistilBERT, also first submitted in October 2019. The team addressed the difficulty of running large pretrained language models under constrained computing budgets. Rather than distilling a model only for a particular task, they applied knowledge distillation during pretraining to produce a smaller general-purpose model that could subsequently be fine-tuned. Their training objective combines language-modeling, distillation and cosine-distance losses; the work also includes proof-of-concept and comparative on-device experiments.[5]

Wolf also identifies Datasets as a library he created and co-authored its November 2021 paper, whose author list begins with Quentin Lhoest and includes numerous collaborators. Datasets standardizes interfaces, versioning and documentation, with a lightweight front end intended to behave similarly for small collections and internet-scale corpora. Dataset additions and usage documentation follow a distributed, community-driven approach.[2][6]

Beyond libraries, Wolf has worked on open-science efforts including the BigScience Workshop, which led to the BLOOM model and dataset. He also co-authored the O’Reilly book Natural Language Processing with Transformers and produced educational material including The Ultra-Scale Playbook.[2]

In “The Einstein AI model”, Wolf argues that answering difficult questions with known answers does not establish the ability to make scientific breakthroughs. He contrasts his own ease with school examinations with his recollection of struggling to invent ideas and challenge established knowledge during his doctorate: “I was no Einstein, I was just very good at school.”[7]

Wolf warns that continuing current training trends could yield “a country of yes-men on servers.” His concern is conditional: models need incentives to question learned knowledge and propose ideas that may conflict with their training data. He proposes evaluating whether they can take counterfactual approaches, make general proposals from small hints and ask non-obvious questions that open research paths. He presents how to evaluate these abilities as an open question, rather than a completed benchmark.[7]

Wolf’s work also includes building Hugging Face’s robotics effort, including projects around Reachy Mini. On his Hugging Face profile, he advocates AI and robots that are “open-source, transparent, and affordable; community-built and safe; hackable and fun.” Describing collaboration with Pollen Robotics, he writes that the teams “decided to join forces.”[2][3]

At the 2026 AI Engineer World’s Fair, Wolf interviewed MiniMax’s Olive Song.[1]

Before Hugging Face, Wolf’s career passed through experimental physics and patent law. In his autobiographical account, he describes graduating from École Polytechnique and working on laser-plasma interactions at Lawrence Berkeley National Laboratory’s BELLA Center. Although accepted for a doctorate at MIT, he completed his doctorate in statistical/quantum physics at Sorbonne University and ESPCI, working on superconducting materials for the DGA and Thales.[2]

Wolf recalls wanting a change from the long timescale of physics experiments after his doctorate. He joined the intellectual-property firm Cabinet Plasseraud, obtained a law degree from Panthéon-Sorbonne University and worked as a patent attorney for five years, helping startups and larger companies build and defend intellectual-property assets.[2]

He dates his consulting for deep-learning and AI startups to 2015. Wolf says their work introduced him to machine-learning mathematics, in which he recognized approaches familiar from statistical physics. He studied through books and online courses; about a year later, a friend invited him to start something with Hugging Face. He describes that transition as returning to science and coding.[2]