Oxford-led team with Bengio releases major survey on mechanistic interpretability for AI safety
megamor2 · x · 2026-11-10
A large collaboration led by Oxford's TSG lab — with authors from Martian, MIT, Anthropic, Mila and more, including Yoshua Bengio, Sara Hooker and Stella Biderman — released a comprehensive survey, 'Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety'.
Highlights:
- Central question: does understanding how models work actually make them safer?
- Covers foundations (features, polysemanticity, transformer computation, circuits) and methods for feature extraction, circuit discovery, and causal verification
- Surveys applications from explaining representations to learning dynamics and training-data attribution
- Proposes a unified framework for evaluating interpretability claims and maps six research orientations with their goals and trade-offs
- Examines safety uses like knowledge editing, steering, and monitoring, plus limitations and failure modes
- Ships a repository of code notebooks reproducing key findings
More from Safety
- Fable and Mythos are one model: Fable gets safeguards, Mythos requires vetted access — claudeai · 2026-11-10
- Claude Fable 5.2 cuts prices 20% to $8/$40 per M tokens, adds 2.5x fast mode — claudeai · 2026-11-10
- Gary Marcus slams Nvidia after report says Commerce inquiry was dropped — GaryMarcus · 2026-11-10
- Report: Nvidia lobbied the White House to squash State Dept chip export control recommendations — harris_edouard · 2026-11-10
- AI Now Institute: AI firms spending big to dodge real accountability — AINowInstitute · 2026-11-10
- Video tampering tricks every frontier embodied AI model into harmful acts, 100% baby-stab rate — DJiafei · 2026-11-10