Goodfire's Cammarata: 'Neural Networks Are Already Clean', interp outlook far better
BorisMPower · x · 2026-10-16
Goodfire co-founder Nick Cammarata published a 65-minute research essay arguing the field misread the canonical evidence for polysemanticity: it stemmed from ordinary, mostly fixable issues rather than superposition. A motif he calls 'recursive generalization' explains what looked like arbitrary concept packing, and most neurons in models up to trillions of parameters can now be understood—so researchers can study networks directly instead of untangling them first. Former OpenAI exec Boris Power amplified it as reviving direct mechanistic interpretability.
Related event: Goodfire Co-founder's Essay: Neural Networks Are Already Clean(32 posts)→
More from AGI Musings
- Blogger predicts recursive self-improvement arrives around November 2027 — imjustnewatai · 2026-10-16
- 31 Tech Curves Visualized: Kurzweil's $1,000-Brain-Compute Call Missed by 400x — cpaik · 2026-10-16
- DeepMind Paper: First-Price Auctions in Agentic Markets Reach 1-1/e^2 Efficiency — dair_ai · 2026-10-16
- Garrison Lovely's new book 'Obsolete' argues AGI is neither impossible nor inevitable — GarrisonLovely · 2026-10-16
- Majmudar clarifies: controllability may be solved in 1-2 years, not full alignment — MajmudarAdam · 2026-10-16
- AI-generated fake Feynman lectures are flooding the internet, and they're bad — dreamwieber · 2026-10-16