By Evan Vega
For most of a decade, interpretability researchers have worked on the assumption that you can’t study an AI model one neuron at a time. Networks run out of room, the theory goes, so they pack unrelated concepts into the same neuron because those concepts rarely collide. That’s superposition, and it’s why so much of the field trains a second model, usually a sparse autoencoder, to untangle the first one.
On 15 October 2026 Nick Cammarata, a co-author of the 2020 paper that made superposition part of the canon, published Neural Networks Are Already Clean with the interpretability lab Goodfire. His argument: the canonical evidence was misread, the famous cat-and-car neuron detects a V shape, and about 90% of neuron poles in the open Qwen3-32B model have one main job. Novel Cognition read all 27,000 words, checked the quotations it rests on, and went through the reaction.
Nick Cammarata, who co-authored the 2020 paper that made superposition part of interpretability’s canon, argues in a Goodfire essay that the evidence was misread: the famous cat-and-car neuron detects
The reach was enormous: the launch post on X passed 2.7 million views in a day, and the Hacker News thread reached 743 points. Working interpretability researchers outside Goodfire were cooler. Neel Nanda, who leads mechanistic interpretability at Google DeepMind, called it interesting work but not a significant update and said he still thinks superposition is real. Liv Gorton argued that superposition is a claim about geometry, more features than dimensions, so recursive generalization doesn’t compete with it. Leo Gao said that if you can’t simulate a neuron you haven’t explained it, and Stephen Casper noted that understanding 90% of the behaviour of 90% of neurons doesn’t add up to 81% of the network.
Cammarata answered critics on X, conceded that his definition of monosemantic was loose, and said the technical paper is maybe a few weeks away. Some amplifiers went further than the essay does; it never calls superposition a hoax, and it labels its biggest conclusion highly speculative.
InceptionV1 neuron 4e:55 has been the field’s standard example of a polysemantic neuron since 2017. It responds to cat faces and to car fronts, and the 2020 paper Zoom In said it “isn’t responding to some commonality of cars and cat faces.” Cammarata, a co-author of that paper, now reads it as a V-shape detector: its strongest inputs include V-shaped flames, fireboat water jets and fox faces, and its negative pole is an O.
The circuit Zoom In used to introduce superposition, a car feature spreading into unrelated dog neurons, was one Cammarata found in 2019. He now says the cars and dogs were grouped by shared pose, and that the visualization tool was reward hacked toward dogs. InceptionV1’s unreadable last layer, he estimates, goes from about 5% interpretable with old tools to about 99% in a ConvNeXt of the same size trained on the same data. When he starved layers of neurons, they got coarser instead of packing unrelated concepts together.
What remains of the mess, he argues, is recursive generalization: a neuron that detects “iy” grows into “diy” and then “do it yourself”, one related offshoot at a time.
The headline figure comes from preliminary audits of Qwen3-32B: about 90% of neuron poles with one main job, about 7% with two related jobs, and about 2-4% not yet understood. The essay itself says the share moves with thresholds and counts each neuron’s positive and negative sides separately. Agents did most of the labelling, Kimi K3 appears as a single attribution graph, and the technical companion paper with the language-model detail had not been published as of 16 October.
The comparison point is Anthropic’s 2023 Towards Monosemanticity, where the median neuron in a one-layer model scored 0 on an interpretability rubric. The same paragraph scored the median dictionary feature at 12, a figure the essay doesn’t mention. Goodfire, which raised a $150 million Series B in February 2026, has a commercial interest in models being readable; many of the image-model results can be checked in its public microscope.
The re-reading of the old image-model evidence is the strongest part, because it comes with pictures and a public microscope. The language-model numbers wait on the companion paper. Cammarata names his own bar: one convincing superposition circuit in a real network. Qwen3-32B is open weights, so anyone can go looking.
The full file, with every number, every quote check and the reaction by venue, is at alreadyclean.novcog.us.com: the cat-and-car neuron, the numbers, the reaction and what the essay doesn’t show yet. Primary source: Neural Networks Are Already Clean (Goodfire, 15 October 2026).
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily Colorado News
