An interpretability researcher argues neural networks are cleaner than we think
If he is right, a large part of interpretability research is solving a smaller problem than assumed. It is a hypothesis, not a paper yet.
Autor: The desk, News · · 1 min czytania
Tagi: interpretability · goodfire · research · qwen
Tego tekstu nie ma jeszcze w języku Polski, więc czytasz po angielsku.
- Wiadomość
- If he is right, a large part of interpretability research is solving a smaller problem than assumed.
- Dlaczego to ważne
- It is a hypothesis, not a paper yet.
- Co jest potwierdzone
- The photograph is generated for this story.
Powiązane z nazwanym źródłem. Nieoznaczone jako zweryfikowane. Goodfire Research
The news
Nick Cammarata of interpretability startup Goodfire published an essay titled Neural networks are clean, arguing that most neurons in modern networks are individually understandable and that superposition, networks packing many unrelated concepts into one neuron, is far rarer than the field assumes. In preliminary audits of Qwen3-32B he reports that about 90 percent of neuron poles have one main job, about 7 percent two related jobs and 2 to 4 percent are not yet understood.
Why it matters
Sparse autoencoders built to untangle superposition dominate interpretability. If neurons are mostly clean, direct neuron-level audits of frontier models become practical.
The details
Cammarata also argues that classic examples of mixed-up neurons in InceptionV1's last layer came from architecture and tooling problems, and that apparent polysemanticity is mostly a neuron growing related offshoots during training.
He frames it as a working hypothesis. The percentages depend on thresholds, the samples are small, and a companion paper is pending. The essay drew more than 560 points on Hacker News.
What we could not check
- No peer-reviewed paper or replication exists yet.
- Percentages depend on the author's own thresholds.
Źródła
The photograph is generated for this story. It does not show the people named in it.