Thomas Wolf

Thomas Wolf is a Hugging Face co-founder and chief science officer, and a creator of the Transformers and Datasets libraries that make machine-learning models and data easier to use. His defining contribution is turning research into tools other people can inspect, adapt and build upon—an ambition that now extends into robotics and scientific discovery.

From physics through patent law

Wolf’s first scientific career was in physics. After graduating from École Polytechnique, he worked on laser-plasma interactions at Lawrence Berkeley National Laboratory, then completed a doctorate in statistical and quantum physics in Paris, researching superconducting materials. Programming was already part of his life, but experiments moved at a pace he found frustrating: a single investigation could take years.

He changed direction, earned a law degree and spent five years as a patent attorney at Cabinet Plasseraud. The route back to science came through his clients. In 2015, working with deep-learning startups, he recognized mathematics familiar from statistical physics and began educating himself through books and online courses. Julien Chaumond subsequently invited him to help build Hugging Face’s science team. Machine learning brought together his appetite for science, writing and software that could be tested quickly.

His early research at Hugging Face included TransferTransfo, coauthored with Victor Sanh, Chaumond and Clément Delangue. The project applied pretrained Transformer models to conversational agents, combining transfer learning with training objectives for generating and selecting dialogue responses. It addressed a practical question that would become central to his work: how can a broadly trained model become useful for a particular task?

Making research usable—and reproducible

Transformers supplied an answer at the level of infrastructure. Wolf and his collaborators built a shared interface to model architectures and pretrained checkpoints, allowing developers to use advances without reconstructing every research implementation. The library needed to accommodate researchers introducing new models as well as practitioners deploying existing ones.

Wolf’s design preferences were concrete. Separating tokenizers from models made components easier to reuse; sharing checkpoints helped the community grow. He also favored keeping individual model implementations approachable rather than forcing every architecture through an elaborate internal abstraction. Tokenization had initially seemed secondary, and extracting it later proved painful—a lesson in how apparently peripheral components shape the usability of research software.

Datasets, another library he helped create, extended that approach to the material models learn from. Its common interfaces, versioning and documentation support both small datasets and much larger corpora, with community contributions expanding the collection. Wolf wanted reproducibility to reach beyond model weights: researchers also needed ways to trace what happened to their data.

That concern developed into open science at scale. He helped advance BigScience, the international collaboration that released BLOOM in 2022. Its 176-billion-parameter multilingual model came with intermediate checkpoints and optimizer states, giving researchers more of the training process to investigate. The achievement belonged to a broad collaboration; Wolf’s contribution fits his larger effort to make consequential AI research accessible beyond a handful of industrial laboratories.

More recently, he coauthored FineWeb, which paired a large web-text training dataset with documented experiments on filtering and deduplication. The team released its curation code and experimental models, making the preparation decisions available for others to examine. FineWeb2 extended that work to multilingual data, adapting the processing pipeline across languages. His coauthored Ultra-Scale Playbook tackles another obstacle: teaching how to coordinate GPU clusters, manage memory and choose parallelism strategies. Access includes knowing how to build the model, not simply being able to download it.

The ideas driving his next chapter

• Scientific discovery needs better questions. Wolf’s argument about scientific AI begins with an uncomfortable distinction from his own career: excelling at school did not make him equally good at originating research. He questions whether difficult exams with known answers adequately measure the ability to produce scientific breakthroughs. He wants systems that challenge assumptions, pursue counterfactuals and ask questions that open unexpected research paths. Even an elegant mathematical proof, in his view, differs from identifying a fruitful new field. How to evaluate that capacity remains an open problem.

• Open models provide resilience. His response to the debate over DeepSeek and export controls argues that downloadable weights change the relationship between a model’s country of origin and its users. Organizations can run and adapt a model without sending their data to its original provider. He connects that independence to reliability: distributing inference across providers and local devices can reduce dependence on a single company’s infrastructure as AI becomes more deeply embedded in everyday activity.

• Low-tech AI. Wolf imagines useful intelligence becoming an ordinary component of existing systems. Smaller capable models, specialized hardware and interfaces that understand familiar text, code and images could lower the cost of adoption. His example is an established business adding capabilities to older software rather than rebuilding its infrastructure. This is a proposed direction, with imagined applications, rather than a claim that every such integration is already dependable.

• LeRobot and accessible robotics. Wolf is helping carry Hugging Face’s approach into physical machines: connect policy models, datasets and hardware through software that developers can use and modify. He sees diversity of training environments as a central challenge—a robot that succeeds in one room may struggle in the next. Shared demonstrations from many locations could help address that limitation. He also argues for local execution, because a robot’s dependence on a remote service can have physical consequences when connectivity fails. His vision for robotics gives affordable arms and expressive desktop robots room alongside humanoids, widening the opportunities to experiment.

The hardware effort became tangible with Hugging Face’s acquisition of Pollen Robotics and the July 2025 introduction of Reachy Mini, whose launch Wolf coauthored with Matthieu Lapeyre. Programmable in Python and equipped for audiovisual interaction, the small robot gives developers a physical platform for testing and sharing behaviors. Wolf’s public statement about joining forces with Pollen emphasizes robots that are transparent, affordable and modifiable. His current scientific interests reach further: whether AI can generate genuinely new knowledge. Both ambitions ask people to become participants in developing the technology, with the tools to change what it does.