• Sources: primary, discussion
  • Summary: Nick Cammarata of Goodfire argues that canonical InceptionV1 polysemanticity came from architectural problems and weak tools, and that retrained versions have 99 to 100 percent interpretable last-layer neurons. In preliminary audits of Qwen3-32B, the team estimates about 90 percent of neuron poles have one main job, and only 2 to 4 percent are not yet understood. The essay is a working hypothesis. A technical companion paper is promised.
  • Why it matters: If the claim holds, interpretability work can study a model's own neurons directly instead of first training sparse autoencoders to decompose it.
  • Follow-up: Publication of the technical companion paper on language model neurons.
  • Comments: The HN thread debates whether the rubric for one main job is too generous and how the claim squares with sparse autoencoder results.