The claim
Nick Cammarata, a researcher at the interpretability startup Goodfire who previously worked at OpenAI, has published a long essay titled Neural Networks Are Already Clean. Its argument is that most of what the field calls polysemanticity — a single neuron appearing to do several unrelated jobs — comes from ordinary, fixable problems with tooling and metrics, plus a pattern he calls recursive generalization, rather than from superposition.
Superposition is the idea that a network packs more features than it has dimensions, storing them in overlapping directions, which would mean interpretability has to untangle a network before it can study it. Cammarata’s position is blunt: “I now think superposition should be far down the list,” he writes, and, of the canonical example, “superposition was the wrong explanation for InceptionV1.”
The evidence he offers
The central reinterpretation concerns a neuron in InceptionV1, a 2020-era image model, that seemed to respond to both cats and cars. Cammarata argues it responds to a V shape, which appears in both cat faces and the front of cars — the feature was coherent and the feature visualisation was misleading.
Around that sit several measurements. In preliminary audits of Qwen3-32B, he reports about 90% of neuron poles having one main job, about 7% having two related jobs, and 2–4% not yet understood, while noting the numbers depend heavily on thresholds. For InceptionV1’s last layer he traces estimated interpretability from about 5% with 2017-era tools, to about 50% with better tools, to about 70% with modern regularisation, and about 99% in a same-size ConvNeXt. Squeezing one layer from 480 neurons down to 10 made the neurons coarser, he says, not superposed.
The essay is explicit about its own status. It calls itself a working hypothesis, labels the final step of its four-step argument “highly speculative”, says the cleanliness figure for strong activations rests on relatively small samples, and defers the question of whether networks hold more features than dimensions to a companion paper.
Four researchers, four objections
MIT Technology Review, which put the post at 26,000 words and quoted its opening line — “I think I have good news about the nature of neural networks” — collected the pushback.
Chris Potts of Stanford University called it little more than a hunch: “As far as I can tell, nothing about this blog post makes the problems any easier.” Jacob Hilton of the Alignment Research Center said researchers understand less than 0.01% of a top-tier language model’s inner workings, that the post gives the public an unrealistic picture, and that full understanding requires explaining how neurons form circuits — not just what individual neurons do.
Oren Etzioni, formerly technical director of the Allen Institute for AI, pointed out that the original InceptionV1 work had rejected a neuron-level description put forward by Geoffrey Hinton and others in 1986, and said Cammarata’s new view “takes us full circle” — “What’s new in 2026 is that he now accepts it.” Neel Nanda of Google DeepMind called the work interesting but not a significant update to interpretability as a whole.
What is actually unresolved
The disputed points are narrow and checkable. Whether findings from a small 2020 vision model generalise to current language models; whether examining many examples is rigorous enough without formal evaluation; and how much neuron-level cleanliness buys you when circuits remain the open problem.
Cammarata has conceded ground on the framing himself. “It’s not infinitely good news; it’s not everything solved,” he told MIT Technology Review. The companion paper on features and dimensions, and Goodfire’s unpublished technical results on language models, are what would turn a working hypothesis into a result.
