Ex-OpenAI interpretability lead: ~90% of neurons are monosemantic, superposition narrative was wrong
inductionheads · x · 2026-10-16
Nick Cammarata, who started OpenAI's first large-scale automated interpretability project, publishes a striking reassessment: preliminary audits of Qwen3-32B suggest 90% of neuron poles have one main job, 7% two related jobs, and only 2-4% remain unexplained—with similar results in dense and MoE models up to trillions of parameters like Kimi K3, to be detailed in a companion paper. He argues the canonical superposition examples (including the 2019 Zoom In car/dog circuit) were misread as ordinary generalization, that larger models have cleaner representations, and that interpretability may be the most important scientific problem today since alignment is impossible without knowing what we're building. Commenters call it a major dunk on Anthropic's hyped interpretability science and a vindication of Goodfire.
More from AGI Musings
- AI circles debate whether a recent PhD signals poor judgment as credentials lose meaning — repligate · 2026-10-16
- Debate: relational preferences — AI that terminally wants to obey — as an alignment bet — teortaxesTex · 2026-10-16
- Sabine Hossenfelder fires back: academia's impact metric turned physics into a paper mill — skdh · 2026-10-16
- Researcher credits GPT for multiple accepted papers, incl. disproving Wegner's 1965 conjecture — ctjlewis · 2026-10-16
- dlwh: We optimized the legible version of our goals — and forgot the computation we meant to speed up — dlwh · 2026-10-16
- Ideas no longer have a waiting period: agents can build them as soon as they're articulable — akbirthko · 2026-10-16