About
Emmanuel Ameisen is an interpretability researcher at Anthropic, where he studies how large language models like Claude arrive at what they say. He came to research from applied machine learning: a data scientist at Zipcar, Head of AI at Insight, where he helped hundreds of fellows build applied ML projects, and a machine learning engineer at Stripe working on its core fraud models. That practice became his O'Reilly book Building Machine Learning Powered Applications (2020). At Anthropic he first worked on finetuning Claude for tool use, then moved to interpretability, where he was lead author of 'Circuit Tracing' (2025) and a co-author of its companion, 'On the Biology of a Large Language Model', which caught a model choosing a rhyme before writing the line that ends on it. His recurring point is that models are grown rather than built, so studying them is closer to biology than to engineering, and the more you look inside, the less certain you become.
Key Contributions
- Lead author of 'Circuit Tracing: Revealing Computational Graphs in Language Models' (2025), the attribution-graph method for following a model's computation step by step
- Co-authored 'On the Biology of a Large Language Model' (2025), whose case studies found Claude planning rhymes ahead and using concepts shared across languages
- Wrote Building Machine Learning Powered Applications (O'Reilly, 2020), a practitioner's guide from product idea to deployed model
- As Head of AI at Insight, guided hundreds of fellows through applied machine-learning projects
- Co-authored follow-up circuit studies on attention ('Tracing Attention Computation Through Feature Interactions') and on the geometry of a counting task ('When Models Manipulate Manifolds')
Videos & Interviews
Papers & Publications
Circuit Tracing: Revealing Computational Graphs in Language Models
2025The method paper: replace a model's internals with interpretable features, then draw the attribution graph of how they combine to produce one output.
Read PaperOn the Biology of a Large Language Model
2025The companion case studies: attribution graphs turned on Claude, finding rhymes planned in advance, multi-step reasoning, and concepts shared across languages.
Read Paper
