Building AI we can understand, shape, and study in the open
I’m Stella Biderman, an AI researcher and the Executive Director of EleutherAI.
I study how large models acquire their capabilities, how we can make open-weight systems safer, and how to ensure that researchers outside the world’s largest technology companies can participate in shaping the future of AI.
Research agenda
Three connected lines of work guide my research.
Open-Weight Safety
Safety techniques for models whose users can inspect, modify, and independently deploy the weights—without relying on API controls, surveillance, or removable refusal behavior.
Read about open-weight safetyInterpretability Over Time
Studying models as learning processes, not just finished artifacts, so we can explain when capabilities emerge, predict training outcomes, and intervene before problems are fixed in place.
Read about training dynamicsOpen Science
Building public data, models, code, and research infrastructure so academics, auditors, nonprofits, and small organizations can conduct meaningful AI research independently.
Read about open scienceRecent news
I speak on a panel at the National Academies meeting Organizing Mathematical Knowledge in the Age of AI and Formalization in Washington, D.C.
We release Bergson, an open-source library that makes modern data-attribution methods practical at the scale of large language models and pretraining datasets. Read the paper.
Where Does Social Reasoning Come From? Capability Provenance in Language Models is accepted to COLM 2026.
I give an oral presentation at ICML 2026 on Don’t Just “Fix It in Post”: A Science of AI Must Study Training Dynamics, a position paper arguing that AI research must study the processes that produce models, not only their final behavior.
Education
M.S. in Computer Science, Georgia Institute of Technology, 2022
B.S. in Mathematics with honors, University of Chicago, 2016
B.A. in Philosophy, University of Chicago, 2016