Article 311: "Neural Networks Are Already Clean": The Interpretability Argument Splitting the Field

Nick Cammarata published "Neural Networks Are Already Clean" with the interpretability lab Goodfire on 15 October, and it is still the argument running through the field a week later. The claim is a reversal of the premise mechanistic interpretability has worked from for most of a decade: that individual neurons are a bad unit of study because networks pack unrelated concepts into the same one, and that you therefore need a second model — usually a sparse autoencoder — to unfold the first.

Cammarata says the founding evidence for that was misread. The famous neuron responding to cat faces and car fronts detects a V shape. The 2019 circuit that introduced superposition to circuits research — which he found himself — groups cars and dogs by shared pose. InceptionV1's unreadable final layer came from an old architecture and weak tooling; a ConvNeXt of the same size trained on the same data reads at about 99%. What remains of the mess he attributes to recursive generalisation, neurons sprouting related offshoots the way a branch sprouts sub-branches. He reports that roughly 90% of neuron poles in Qwen3-32B have one main job, and calls reading neurons at scale with agents a "bitter lesson" for interpretability after Sutton.

The reaction split predictably. The launch post passed 2.7 million views in a day and the Hacker News thread hit 743 points. Working researchers were cooler: Neel Nanda, who leads mechanistic interpretability at Google DeepMind, called it interesting but not a significant update to the big picture, and said he still thinks superposition is real. Liv Gorton argued the essay mischaracterises what superposition predicts. Leo Gao and Stephen Casper questioned how far labelled neurons add up to an understood network. Cammarata conceded his definition of monosemantic was loose and said the technical paper is a few weeks out.

Key Takeaway: The part of this that will survive scrutiny is the vision work, because it is checkable — the cat-and-car re-reading, the pose circuit, the public microscope. The part that will be argued about is the language-model number, because "90% of Qwen3-32B neuron poles" is a preliminary audit largely scored by agents, the essay says the figure moves with thresholds, and the companion paper has not shipped. That gap between a falsifiable vision claim and a soft language-model claim is the whole controversy in miniature. The reason it matters beyond academic priority is budget: dictionary learning, transcoders and attribution graphs built on replacement models are an enormous share of safety research spend, all justified by the assumption that neurons are unreadable. If that assumption is wrong, the field has been buying telescopes to look at something it could have read directly for years — and the correction is expensive, politically awkward, and being made by the person who installed the assumption. Cammarata named his own falsifier: a single convincing superposition circuit in a real network. Qwen3-32B is open weights. Somebody with a GPU budget will try to clear that bar, and the result will settle more than the essay can.

Sources: Goodfire, "Neural Networks Are Already Clean", 15 October 2026; MIT Technology Review interview, 19 October 2026; The Decoder, 16 October 2026; The Superposition File analysis