Oxford releases 180-page mechanistic interpretability handbook with 51 methods and 19 notebooks
Hesamation · x · 2026-11-10
An Oxford-led team (Fazl Barez et al., with senior advisors including Yoshua Bengio, Stella Biderman, and Sara Hooker) released a 180-page survey, 'The State of Mechanistic Interpretability for AI Safety' — effectively a textbook that will form part of an upcoming handbook.
Coverage includes:
- Superposition and transformer circuits
- Sparse autoencoders
- Circuit discovery
- Safety applications: steering, model editing, and monitoring
It ships with an interactive companion: an explorer of 51 methods, 3 demos, and 19 hands-on notebooks. The material has informed teaching at EdinburghNLP and will underpin an Oxford course in Hilary term 2027; lecture slides and notebooks are available on request.
More from Safety
- Redwood's Alexa Pan shares more on Astra's tendency to sandbag safety research tasks — DKokotajlo · 2026-11-10
- Dev roots their LG TV and is stunned by what's running on it — jedisct1 · 2026-11-10
- First policy gaming paper simulates how an AI crisis unfolds across society — SchoeneggerPhil · 2026-11-10
- Stuart Russell's 'shockingly lucid' AI risk intro resurfaces alongside Lanier's Myth of AI — catherineols · 2026-11-10
- Catherine Olsson revisits the spec gaming master list of AI reward-hacking examples — catherineols · 2026-11-10
- HN debate: permission-scoped subaccounts as the missing safety layer for AI agents — vboykis · 2026-11-10