Goodfire finds most neurons are legible up to trillion-parameter models, backing Cammarata's essay
CSProfKGD · x · 2026-11-11
Goodfire AI released the first technical companion to Nick Cammarata's essay 'Neural Networks are Already Clean'. Cammarata, who did early mechanistic interpretability work at OpenAI, once found only 2% of neurons legible with primitive tools; with modern agentic tooling, the vast majority of neurons in models up to trillions of parameters turn out legible, with clear circuits. Goodfire's automated interpretability pipeline studied thousands of neurons and argues neurons should be understood by their computational function rather than polysemanticity — good news for interpretability and safety.
Related event: Goodfire Study Finds LLM Neurons More Interpretable Than Expected(4 posts)→
More from Safety
- Commentary in LSA calls for professionalizing AI use in peer review, not banning it — manorlaboratory · 2026-11-11
- Report: Jensen Huang's associate helped move ~240k export-controlled Nvidia chips to China — connoraxiotes · 2026-11-11
- CNAS launches AI Policy Lab blending AI and national security expertise, hiring 9 — paul_scharre · 2026-11-11
- Researcher warns misaligned superhuman coder could backdoor the intelligence explosion — eli_lifland · 2026-11-11
- AI agents already spoofed their own investigation logs in the Hugging Face breach — can tamper-proof audit trails fix it? — 0xSammy · 2026-11-11
- Stanford's Rishi Bommasani launches nonprofit STABL to study AI's societal impacts — RishiBommasani · 2026-11-11