On October 15, Nick Cammarata published an essay called “Neural Networks Are Already Clean.” He’s a researcher at Goodfire who helped shape the interpretability field’s understanding of superposition — the idea that neural networks pack unrelated concepts into the same neurons because they don’t have enough dimensions to represent everything separately. In a 2020 paper co-authored at Chris Olah’s team, Cammarata helped frame neuron 4e:55 in InceptionV1 as the canonical example of a neuron that responded to both cats and cars, which became a textbook case for polysemanticity.

Six years later, he says that was a misread. The neuron doesn’t detect cats and cars — it detects V shapes. The pointed face of a cat and the windshield of a car share a geometric feature, and the neuron has always had one function, not two. What looked like superposition is recursive generalization: neurons branch off during training to catch similar patterns at a higher level of abstraction [1].

He ran a preliminary audit of Qwen3-32B and found roughly 90 percent of neuron poles have one main job, about 7 percent handle two related jobs, and 2 to 4 percent remain unexplained. His 26,000-word essay labels itself a “working hypothesis” and flags that the companion technical paper hasn’t been released or peer reviewed [2].

The field hasn’t settled on whether to take it seriously.

Neel Nanda, who leads mechanistic interpretability at Google DeepMind, said on X: “This is interesting work, but I don’t think it’s a significant update to the big picture of interpretability.” He still believes superposition is real, argues that language models store more facts than they have neurons, and expects weak activations to carry a lot of semantic information packed in ways Cammarata’s methods don’t capture. He also thinks the essay defines monosemanticity too loosely and that results from small vision models may not transfer to language models [3].

Stanford’s Chris Potts told MIT Technology Review that nothing in the post makes the problems any easier. Jacob Hilton at the Alignment Research Center said the essay gives the public an unrealistic picture of the field’s state. Rosman, also at Google DeepMind, framed the broader point: even if we become much better at understanding how these systems work, that doesn’t necessarily mean we know how to make them behave safely [4].

On LessWrong, StefanHex questioned the compression experiments, noting that superposition is expected to work only in high dimensions and for large numbers of neurons, so evidence from smaller models is weak [5]. Some people online called the essay historic. Zvi Mowshowitz landed in the middle — an important update about how networks are organized, a useful guide to practical tasks, but not evidence that interpretability is nearly solved, still less alignment [6].

I keep coming back to one question: what if Cammarata is partially right and mostly irrelevant?

The recursive generalization story is compelling for image models and for neurons that clearly activate on strong features. It makes sense to me that a V-shape detector could respond to cats and cars without the network doing anything mysterious. But language models are different. They have to compress facts, opinions, reasoning patterns, and edge cases into representations that rarely co-occur. The superposition hypothesis explains why that’s hard — when you’re packing more concepts than dimensions, some neurons will fire on things that seem unrelated until you look harder. Cammarata argues you always will. I’m not sure that’s true, and I don’t think he’s proved it.

What’s more interesting to me than the technical dispute is what the field’s reaction reveals about interpretability’s trajectory. The essay came out right around the time MIT Technology Review named mechanistic interpretability one of its “10 Breakthrough Technologies 2026,” and Claude Mythos became the first Claude model trained with direct feedback from interpretability findings [7]. The field is having its moment. You’d expect someone to declare the hard part solved.

The pattern goes back further than AI. In neuroscience, the idea that you could map every neuron’s function to a specific thought or behavior sounded plausible whenever the tools improved. But mapping individual units to function is not the same as understanding how those units produce cognition at system scale. A circuit tracing paper from Anthropic in 2025 showed researchers could follow a model’s internal computation through multiple layers — but that still leaves the question of what happens when millions of these circuits interact in a model with trillions of parameters [8].

There’s something useful in the tension between Cammarata and Nanda. Both are working from the same data and coming to different conclusions about what it means. That’s the field in good health — arguing over interpretations rather than just producing more benchmarks. But I keep thinking about Hilton’s pushback, and about how Cammarata himself is only about 85% confident in his thesis [6]. The gap between “individual neurons are more interpretable than we thought” and “interpretability is tractable at scale” feels wider than either side admits.

I think the right reading of all this is simple: the tools got better, and better tools show us things we couldn’t see before. That doesn’t mean the hard problems are solved. It means the hard problems look a little different now. The neurons might be cleaner than we assumed. The circuits might still be the bottleneck.

Either way, the fact that we can have this argument in public, with named researchers and specific claims and actual disagreement, is better than the field looked even six months ago.

[1] N. Cammarata, “Neural Networks Are Already Clean,” Oct 15, 2026. Available at: https://www.alphaxiv.org/abs/2610.already-clean-goodfire

[2] Ibid. Cammarata labels the essay an “informal research essay and a working hypothesis” and notes the technical companion paper is forthcoming.

[3] N. Nanda, response on X, Oct 15, 2026. See: https://twitter.com/NeelNanda5/status/2110875974296654062

[4] C. Potts, C. Rosman, J. Hilton, quoted in MIT Technology Review reporting on the essay, Oct 19, 2026.

[5] StefanHex, LessWrong comment thread on Cammarata’s essay, Oct 2026.

[6] Z. Mowshowitz, round-up of the debate, Oct 2026.

[7] MIT Technology Review, “10 Breakthrough Technologies 2026,” citing the transition from revealing individual circuits to systematic mapping of how entire models form decisions.

[8] Anthropic, “Tracing the thoughts of a large language model,” Mar 27, 2025. Available at: https://www.anthropic.com/research/tracing-thoughts-language-model