On October 15, 2026, Goodfire published “Neural Networks Are Already Clean”, a long essay by Nick Cammarata, with contributions from Curt Tigges and Tarun Gogineni, labeled “a working hypothesis”. It argues that individual neurons in trained networks are already clean enough to study directly, and that apparent polysemanticity mostly comes from ordinary, fixable problems and a motif the author calls recursive generalization, rather than from superposition, which he says became the field’s default explanation on the basis of misread evidence. Cammarata notes he previously held and promoted the superposition view.
The evidence revisits vision models: classic examples such as the InceptionV1 “cat and car” neuron from the Zoom In circuits work are reinterpreted (for example as a V-shape detector and a shared-pose circuit), feature visualization is shown to “reward-hack” toward dogs in deeper layers, and the author estimates InceptionV1’s last layer goes from about 5% interpretable with 2020-era tools to about 50% with better tools and 70% with modern regularization, and about 99% in a ConvNeXt of the same size. In one experiment, compressing InceptionV1’s mixed3b layer from 480 to 10 neurons and retraining, a sparse autoencoder recovered the original curve detectors only loosely.
For language models, preliminary automated audits of Qwen3-32B, with agents running explanation-and-test loops, find roughly 90% of neuron poles have one main job, about 7% two related jobs and 2 to 4% not yet understood. The essay contrasts this with earlier baselines it cites: a 2023 Anthropic one-layer study scoring the median neuron 0 of 14, and OpenAI’s automated interpretability finding about 1.7% of neurons interpretable.
Why it matters: superposition and the sparse-autoencoder program built on it are central to current interpretability. A credible argument that raw neurons are mostly interpretable would change where the field spends effort.
What it does not show: the author stresses the percentages depend heavily on thresholds and on what counts as “one thing”, several samples are small, the step to “this may be most of what interpretability needs” is labeled highly speculative, and the larger-model results are deferred to a companion paper. It is an essay, not a peer-reviewed study.