Mechanistic Interpretability Hackathon 5 runs Nov 20-22, theme Autointerp Evals

burny_tech · x · 2026-11-19

The 5th Mechanistic Interpretability Hackathon runs Nov 20-22 (48 hours) on the Mechinterp Discord, with submissions to itch.io. Official theme: Autointerp Evals — building ways to automatically score interpretability explanations, similar to OpenAI's autointerp scoring explanations by how well they predict activations. Other suggested directions: testing Nick Cammarata's 'neurons are already clean' claim, steering neurons with clear meanings, or exploring J-lens. Toy-size projects welcome; kickoff, discussion, and closing sessions daily at 5pm UTC.

Original post →

More from Companies & People

Companies & People channel →