Goodfire · 2026-10-15 · notable
Nick Cammarata — neural networks are already clean
Goodfire researcher Nick Cammarata argues that neural networks are already mostly interpretable. In a first audit of Qwen3-32B, he finds about 90% of neurons have one main job and only 2–4% are not yet understood.
A Goodfire essay says most 'messy' neurons were misread, and that models are cleaner than superposition theory suggests.
What is it?
"Neural Networks Are Already Clean" is an informal research essay by Nick Cammarata of the interpretability lab Goodfire, published on 15 October 2026. Cammarata argues that canonical examples of polysemantic neurons, such as InceptionV1's "cat and car" neuron, were misread. He says most apparent mixing comes from neurons growing related offshoots during training, plus weak tools and limited data.
How does it work?
The essay re-examines classic circuits, retrains InceptionV1 and ConvNeXt models across a roughly 300x size range to watch neurons form, and audits neurons in Qwen3-32B. Cammarata also squeezed an InceptionV1 layer from 480 to 10 neurons; a sparse autoencoder recovered features that only partly matched the original curve detectors, which he reads as evidence against a hidden high-dimensional code.
Why does it matter?
Much interpretability work assumes neurons are tangled and must be unpacked with sparse autoencoders first. If Cammarata is right, researchers could study many neurons directly. He calls the essay a working hypothesis; a companion technical paper is still to come, and the post reached 438 points on Hacker News.
Who is it for?
interpretability researchers and ML engineers