A Goodfire researcher argues neural networks are far more interpretable than the field believed

Nick Cammarata, an interpretability researcher at Goodfire who previously worked on the subject at OpenAI, published a long essay arguing that neural networks are much easier to understand than the field had come to believe. He described it as “an informal research essay and working hypothesis”, not a peer-reviewed paper.

Since about 2022 the field’s dominant explanation for “polysemantic” neurons, which seem to respond to several unrelated things, has been superposition: the idea that a network short of space packs unrelated concepts into the same neurons (Anthropic’s 2022 “Toy Models of Superposition”). If so, neurons cannot be studied directly, and researchers first train a second model such as a sparse autoencoder to untangle them, the approach behind Towards Monosemanticity and its successors. Cammarata co-authored some of the original evidence for this, with Christopher Olah, and now says it was misread.

The canonical evidence of polysemanticity and superposition that convinced much of the field, including me, that networks pack concepts arbitrarily was misread.

— Nick Cammarata, “Neural Networks Are Already Clean”

His evidence is strongest for image models. He reinterprets the best-known “cat and car” neuron as a detector of V shapes. He argues that the messy late layers of InceptionV1, the early field’s standard test model, came from architectural problems and weak tools: with modern architecture changes, at the same size and on the same data, he reports 99–100% of last-layer neurons as easily interpretable. When he shrank a layer 48-fold, its neurons became coarser instead of mixing unrelated concepts, the opposite of what superposition predicts. For language models the claims are labelled preliminary. In audits of Qwen3-32B he estimates about 90% of neurons have one main job and 2–4% are not yet understood, and he reports similar results in trillion-parameter models such as Kimi K3, with details deferred to a companion paper. He gives his overall confidence as “maybe 85%” and says progress on attention has been slower.

The reaction online was unusually enthusiastic. OpenAI’s head of strategic futures, Dean Ball, called it possibly “one of the most important single days in deep learning history”. The OpenAI researcher known as roon wrote that it “may be remembered as a major moment in scientific history”. Dylan Field, Figma’s chief executive, said it “might be the most important essay of 2026”. Nathan Calvin, general counsel at the advocacy group Encode AI, was more cautious. He said he was “less sure about how much the paper should change overall optimism”, and posted an assessment he had asked Claude to write, which called the language-model results early and noted that labelling neurons is not the same as explaining how millions of them produce behaviour such as deception. Cammarata’s own closing section, on what it could mean for alignment, is framed as speculation.