Guillaume Corlouer
Bio
Guillaume Corlouer is an independent AI safety researcher based in Brighton, England. He holds a PhD from the University of Sussex (Sackler Center for Consciousness Science), where he studied information flow in ECoG time series during visual perception under the supervision of Lionel Barnett and Anil Seth, with prior training in mathematics at the University of Paris-Sud. His AI safety work centers on mechanistic interpretability, applying information-theoretic and complex-systems methods to discover latent variables from neural network activations. He was a fellow in the 2023 PIBBSS program, investigating stochastic gradient descent on singular models in the context of Singular Learning Theory, and subsequently a PIBBSS research affiliate, during which he co-authored an information-theoretic study of lying in LLMs presented at the ICML 2024 Workshop on LLMs and Cognition. He also participated in AI Safety Camp 8 (2023), where his team produced the paper "Linearly Structured World Representations in Maze-Solving Transformers," demonstrating that transformers trained on maze-solving learn linear representations of maze structure.
Links
- Personal Website
- https://sites.google.com/view/guillaumecorlouer/home
- Twitter / X
- LessWrong
- guillaume-corlouer
Grants
from Long-Term Future Fund
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.
Details
- Last Updated
- Mar 22, 2026, 4:19 PM UTC
- Created
- Mar 20, 2026, 2:51 AM UTC