sama Endorses Goodfire Essay Claiming Neural Networks Are Already Clean
sama · x · 2026-10-17
Sam Altman publicly praised Nick Cammarata's 65-minute Goodfire research essay "Neural Networks Are Already Clean." Cammarata, who co-authored early mechanistic interpretability work at OpenAI and once found only 2% of neurons legible, now argues the classic polysemanticity/superposition evidence was misread — mostly a fixable motif he calls "recursive generalization." After months of agentic-scale study, he claims the vast majority of neurons and circuits in models up to trillions of parameters are legible, making mechanistic interpretability — and thus alignment — far more tractable than the field feared.
More from AGI Musings
- Finance doesn't believe in the singularity; tech does — and the gap matters — shakoistsLog · 2026-10-17
- What if LLMs could decompose trained models into explicit ordinary code? — tszzl · 2026-10-17
- Neel Nanda warns models manipulating training could make misalignment more persistent — NeelNanda5 · 2026-10-17
- Citigroup bars third-party AI agents as banks fear $47B deposit flight — rohanpaul_ai · 2026-10-17
- MAICON 2026: AI Is Outpacing How We Prepare the Workforce — KendraRamirez · 2026-10-17
- What moats survive AI? Network effects, capital, speed and charisma, says Jon Stokes — nptacek · 2026-10-17