Emmanuel Ameisen researches how language models work internally on Anthropic’s interpretability team. He helped develop circuit-tracing methods that turn parts of a model’s computation into graphs researchers can inspect and test.[1][2][3][4][5]
Start with Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic) →
GitHub 
X Career
Before interpretability, Ameisen worked on Anthropic’s fine-tuning team with a focus on tool use. His earlier work at Stripe concerned fraud models and automating model training and deployment. He was previously Head of AI at Insight, helping fellows build applied machine-learning projects, and a data scientist at Zipcar and LocalMotion. This path connects deploying models in working products with the harder question of explaining what their internal computations do.[1]
Contributions
Tracing a computation
Circuit Tracing replaces selected model components with more interpretable approximations, then traces connections between the resulting features. A feature is an internal pattern that researchers can relate to a concept or behavior. The graph shows candidate paths through which those patterns affect a response. Ameisen helped develop the replacement-model and graph methodology, built infrastructure and a stable graph-generation pipeline, and worked on interventions and evaluation.[2]
Testing an explanation rather than trusting a diagram
A graph is useful only if it captures the model’s behavior well enough to support an explanation. The research therefore tests proposed mechanisms by intervening on internal features and checking the effect on outputs. Its companion study applies this approach to Claude 3.5 Haiku, examining behaviors including multilingual reasoning and arithmetic. The papers also describe limitations from imperfect replacement models, complex graphs and features that do not align neatly with human concepts.[2][3]
Ideas
In his Lenny’s Podcast interview, Ameisen explains the distinction between observing an AI system’s answers and investigating how it produces them. The conversation connects language-model training, internal representations and safety research. It provides a general introduction before the mathematical and experimental detail of the circuit-tracing papers.[4]
Tools, writing & talks
Writing, research & talks
Start here
Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)lennysnewsletter.comA conversation explaining language-model representations, internal computation and the purpose of interpretability research.
Start with the interview for a plain-language account of what interpretability can reveal. Then read the circuit-tracing introduction; the full paper benefits from familiarity with neural networks.
Interview guest[4]Original source- Circuit Tracingtransformer-circuits.pub
A method and validation tools for tracing language-model computations through interpretable replacement components.
Core contributor and co-author[2] - On the Biology of a Large Language Modeltransformer-circuits.pub
Case studies of internal mechanisms in Claude 3.5 Haiku using circuit tracing.
Core contributor and co-author[3]