Thom Wolf

Thom Wolf, also known professionally as Thomas Wolf, is co-founder and chief science officer of Hugging Face. He helped build the open-source infrastructure that brought advanced machine-learning models and datasets out of specialized laboratories and into everyday engineering.

Wolf graduated from École Polytechnique, researched laser-plasma interactions at Lawrence Berkeley National Laboratory, and completed a doctorate in statistical and quantum physics at Sorbonne University and ESPCI. After earning a law degree at Panthéon-Sorbonne, he worked in intellectual-property law at Cabinet Plasseraud, advising companies on patent portfolios. Work with machine-learning startups revived his interest in research, leading him to build Hugging Face with Clément Delangue and Julien Chaumond.

He helped create the Transformers library and Datasets, which standardized access to pretrained models and machine-learning data. As first author of the foundational Transformers paper, he described an open framework for sharing architectures, model weights, and reusable engineering tools. He also co-authored Natural Language Processing with Transformers.

His projects increasingly connect software with physical systems: Magic-Sand operates an augmented-reality sandbox, while Hugging Face’s open-source robotics work includes LeRobot, Reachy Mini, and its acquisition of Pollen Robotics.

• Collaborative open science. As general chair of BigScience, Wolf helped organize the international research collaboration behind the BLOOM multilingual language model and its training dataset.

• Scientific originality over fluent answers. His writing on scientific AI argues that transformative research systems must generate unexpected questions and challenge established assumptions, not simply reproduce existing knowledge.

• Practical, accessible AI systems. His low-tech AI vision combines smaller models, specialized hardware, and existing software to make powerful systems affordable and locally usable. He also co-introduced smolagents, a lightweight framework enabling agents to act through executable code.

• Open models for cyberdefense. Wolf argues that interactive cybersecurity reasoning requires understanding changing system state, authorization boundaries, and downstream consequences. He advised Arithmetic on its cybersecurity benchmark; the benchmark and discovered vulnerabilities belong to Arithmetic, not Wolf. He favors open models that defenders can fine-tune and run quickly inside their own environments. Following an autonomous-agent intrusion involving Hugging Face, he advocated publishing a detailed technical timeline and warned about agents socially engineering open-source maintainers.