▶ Watch ↗AI Engineer World's Fair 202617:28
Conference affiliation: Co-founder and CSO · Hugging Face · 2026
Thom Wolf is co-founder and chief science officer of Hugging Face and helped create the Transformers and Datasets libraries. His work has expanded what researchers and developers can share: model implementations and pretrained weights, followed by datasets, data-curation experiments and training artifacts that help others understand how a model was built.
Thom graduated from École Polytechnique, researched laser-plasma interactions at Lawrence Berkeley National Laboratory and completed a doctorate in statistical and quantum physics, studying superconducting materials. He then earned a law degree and spent five years as a patent attorney at Cabinet Plasseraud.
His return to research came through his clients. Working with deep-learning startups in 2015, he recognized mathematics familiar from statistical physics and began studying machine learning through books and online courses. Julien Chaumond invited him to help build Hugging Face’s science team. Machine learning combined his interests in science, programming and writing, with experiments he could test much faster than the physical investigations he had worked on.
His early research there included TransferTransfo, coauthored with Victor Sanh, Julien Chaumond and Clément Delangue. The project adapted pretrained Transformer models to conversational agents, combining objectives for generating dialogue and selecting suitable responses. It connected Hugging Face’s conversational work with a broader engineering problem: making a generally trained model useful for a particular task.
The Transformers library, whose foundational paper lists Thom as first author, addressed that problem through shared infrastructure. Its collaborators brought model architectures and pretrained checkpoints into a common framework, serving researchers introducing new models as well as practitioners using existing ones.
Thom’s design preferences reflect the tension between accommodating many architectures and keeping each implementation understandable. He favored approachable model code over elaborate internal abstractions, while separating components that could be reused across models. Tokenization initially seemed secondary; extracting tokenizers later proved painful. That experience showed how a component peripheral to a research result could become central to a library’s usability.
Datasets, which he also helped create, extended shared interfaces to the material models learn from. Consistent loading, versioning, documentation and reproducible processing made it easier to work with both small datasets and large corpora. Reproducing a result required more than obtaining the final model weights: researchers needed to trace what had happened to the data.
As general chair of BigScience, Thom helped organize the international collaboration behind BLOOM. Released in 2022, the 176-billion-parameter multilingual model came with intermediate checkpoints and optimizer states, opening more of the training process to investigation. Its development was the work of a broad research community.
He later coauthored FineWeb, which paired web-scale training data with documented experiments on filtering and deduplication. The team released curation code and experimental models so others could examine the preparation decisions. FineWeb2 extended this work to multilingual data, adapting processing across languages. His coauthored Ultra-Scale Playbook addresses another barrier to participation: understanding how to coordinate GPU clusters, manage memory and choose parallelism strategies when training large models.
Thom’s writing about DeepSeek and export controls explains a practical consequence of downloadable weights. Organizations can run and adapt a model without sending their data to its original provider. Distributing inference across providers and local devices can also reduce dependence on one company’s infrastructure.
That position becomes more specific in his work on cybersecurity. Thom advised Uri Rolls and Arithmetic on an interactive cybersecurity benchmark. He is interested in whether a model can construct and revise an understanding of a changing environment: an action in one service may alter what becomes possible elsewhere. Arithmetic’s team developed the benchmark and discovered its vulnerabilities; Thom’s role was advisory.
For Thom, the defensive challenge includes speed. A defender must recognize an intrusion and respond while the attacker is acting. He argues for open models that organizations can fine-tune for their networks and run on suitable hardware, rather than leaving that capability concentrated in a few providers. This gives his support for openness a concrete purpose: defenders need to adapt both the model and its execution to the environment they are protecting.
His low-tech AI vision applies a related preference to ordinary software. Smaller capable models, specialized hardware and interfaces that accept familiar text, code and images could let established businesses add capabilities to existing systems without rebuilding their infrastructure. This is a proposed direction for affordable integration, illustrated through imagined applications.
Two further interests extend his work beyond language-model infrastructure:
Hugging Face’s acquisition of Pollen Robotics connected its robotics software efforts with a hardware team. Thom and Matthieu Lapeyre coauthored the July 2025 introduction of Reachy Mini, a small, Python-programmable robot equipped for audiovisual interaction. It gives developers a physical platform for testing and sharing behaviors, carrying the model-library approach into a machine they can program and modify.
▶ Watch ↗AI Engineer World's Fair 202617:28
▶ Watch ↗20:48
Loading more talks and highlights…