Interpretability Agents: Automating and Scaling Model Interpretation

Explore a lecture on Automated Interpretability Agents (AIAs) for scaling model interpretation. Discover how AIAs, built from language models with tools, can design and perform experiments to answer questions about models of interest. Learn about their ability to operationalize hypotheses as code, update based on observed model behavior, and reach human-level performance on various model understanding tasks. Gain insights into the potential of AIAs to automate and scale model interpretation, making intensive explanatory auditing more accessible to model users and providers. Understand how this research aims to create a richer, iterative, and modular interface for interpretability that can scale to large and complex models.

Interpretability Agents