The Interpretability team’s mission is to discover and understand how large language models work internally, as a foundation for AI safety and positive outcomes.

The team argues that it is very challenging to reason about the safety of neural networks without understanding them.

Their goal is to be able to explain large language models’ behaviors in detail, and then use that to solve a variety of problems ranging from bias to misuse to autonomous harmful behavior.

The team is multidisciplinary: some members have deep backgrounds in machine learning, including one member often described as having started mechanistic interpretability and another who was on the famous scaling laws paper.

Other members joined after careers in astronomy, physics, mathematics, biology, data visualization, and more.

The page lists published research including Natural Language Autoencoders (turning Claude’s thoughts into text), Emotion concepts and their function in a large language model, A “diff” tool for AI (finding behavioral differences in new models), The assistant axis (situating and stabilizing the character of LLMs), Signs of introspection in LLMs, Persona vectors (monitoring and controlling character traits), Open-sourcing circuit tracing tools, Tracing the thoughts of a large language model, Auditing language models for hidden objectives, and Insights on Crosscoder Model Diffing.