The researcher who found superposition says he misread it
Two sentences in today's briefs is the wrong weight for what Nick Cammarata published on October 15. "Neural Networks Are Already Clean" is a working researcher walking back the founding evidence of the subfield he helped create, in public, with an owner's manual of numbers attached. If he is right, a decade of interpretability effort was aimed at the wrong target. If he is wrong, the field loses an afternoon.
What he is actually claiming
Cammarata co-wrote "Zoom In," the 2020 Distill paper that introduced the first superposition circuit, and he started OpenAI's automated interpretability project in 2023, which found about 1.7 percent of neurons legible. Superposition was the explanation he and colleagues reached for when a network made no sense: when a model runs out of neurons it stops grouping related concepts and starts folding unrelated ones together, because they rarely collide. That made the knot real, and it justified the field's retreat into training a second model — a sparse autoencoder — to approximate the network as it would look untangled, then studying the approximation.
His essay rejects that retreat. The canonical polysemantic neuron of InceptionV1, famous for firing on both cat faces and car fronts, is a V-shape detector in his reading: pointed things like fox snouts, flames and water jets push it positive, round things like the letter O push it negative. The 2019 circuit that seemed to show a car hiding among dog neurons is ordinary generalization, he writes — cars and dogs grouped by which way they face the camera — and nothing downstream ever reads that combination back out.
What replaces superposition is "recursive generalization": ordinary generalization extended further than anyone expected. A neuron that detects "iy" grows to cover "diy," then "do it yourself," which contains no "iy" at all. The groupings are coherent at every step; they branch like a tree. Look at the final shape without watching it grow and it reads as noise.
The numbers that carry the argument
The claim stands on audits, not on two re-readings. In preliminary work on Qwen3-32B, Cammarata estimates roughly 90 percent of neuron poles do one main job, about 7 percent do two closely related jobs, and only 2 to 4 percent remain unexplained — similar results in models up to trillions of parameters, including Kimi K3, which the forthcoming companion paper will cover in more detail. Those percentages compress a high-dimensional picture, and he says so: the count moves with the threshold for weak activations, which is why he calls the figure a pointer to a regime rather than a measurement.
The controlled experiments are harder to wave away. Across roughly a 300-fold size range of ConvNeXt models, the largest one's last layer is about 99 percent interpretable; the same-size InceptionV1 sat near 5 percent by his own estimate and near zero by others'. Better tools alone took the original 2014 network to about 50 percent, and retraining that architecture with modern regularization took it to about 70 percent, with model size and training data unchanged. On GPT-2 XL, the model behind the 2023 project, his team scored roughly 70 percent of 500 neurons as doing about one thing — though only about 5 percent came out pristine.
Two results aim at the field's actual project. When he cut InceptionV1's layer 3b from 480 neurons to 10 and retrained, the layer did not start packing unrelated concepts — its neurons got coarser, one general "curves and loops" detector where ten curve detectors had been. And when he trained sparse autoencoders on those layers to see whether they found building blocks the neurons lacked, 88 percent of the last layer's dictionary activity turned out to be near-copies of single neurons, with the remainder faint. He is careful: SAEs still help with "what is in this image," especially in the smallest model. But he never found one that grasped something the neurons did not.
Why this is a product story, not just a paper
Goodfire is an interpretability company with a commercial interest in this thesis, and it should be read that way. It also happens to sell the shortest path from "neurons are readable" to something a buyer will pay for. On October 8 it launched activation-probe monitors on Baseten — small classifiers that read a model's internal activations mid-generation, escalating only suspicious sessions to an LLM judge. On Kimi K3, monitoring about 1,500 sessions cost roughly $51, against $233 for a cheap judge and about $10,000 for a top-tier one, catching 94 percent of malicious hacking sessions while flagging 8.7 percent of harmless ones. Goodfire's own published cascade reports 93 percent recall at a 5.5 percent false-positive rate for under $200 per million exchanges.
That is the essay's argument already running against hostile traffic. Probes work because a model's internal state is an interpretable signal. If Cammarata's picture holds, the answer to "how do we know what a model is doing" stops being a research program and becomes a monitoring line item — the same shift we saw coming when chain-of-thought monitorability started degrading, which we covered in Astra's hidden reasoning loop is the real story, not Astra. For what mechanistic interpretability even is, our explainer AI 101 — What is mechanistic interpretability? is the starting point.
There is a quieter casualty. Anthropic's dictionary-learning program, OpenAI's scaled sparse autoencoders and a long tail of academic work assume features must be extracted before they can be studied. Cammarata's position is that dictionaries belong at the end of an investigation, not the start — that an autoencoder trained on the cat-car neuron splits the V into a cat feature and a car feature, moving you further from what the neuron does. That is a direct argument that a large share of the last three years of interpretability engineering was scaffolding around a tooling failure.
What the skeptics say
Neel Nanda, who leads interpretability at Google DeepMind and is one of the field's founding figures, is the most substantive counterweight. He calls the essay interesting but not a change to the big picture. He expects neuron-level analysis to have serious flaws and miss a lot; he still thinks superposition is real, on the argument that language models store far more facts than they have neurons, with weak activations carrying much of that information; he doubts findings from small image models carry over; and he thinks the essay defines monosemanticity too loosely. His team already deprioritized sparse autoencoders for lack of payoff, so he is not simply defending an incumbent method.
The structural objections matter more than the tone of the reaction. The companion paper is unreleased, no third party has replicated the audits, and none of this is peer reviewed — the 90 percent figure is a company's own count on a threshold it chose, from a company with an obvious stake. And Roon, the OpenAI researcher who called it a possible major moment in scientific history, flagged the same caveats: replication is missing, and interpretability progress will likely push capabilities forward too.
The reasonable reading is narrower than either camp. Cammarata has shown that a specific set of famous examples was misread and a specific set of tools was broken — feature visualization reward-hacking into dog neurons, a double-negative regime in SwiGLU neurons distorting the top activations of about 4 percent of neurons, a residual-stream dimension that made every vision map look like a single arc. Those are findings about instruments. The leap from better instruments to "networks are already clean" is the part still owed to the field, and it is exactly the leap that decides whether the safety argument lands.
What to watch
Whether the companion paper appears with thresholds, negative results and the Kimi K3 detail it promises — and whether anyone outside Goodfire reruns the Qwen3-32B audit with their own rubric. Whether DeepMind changes course, since Nanda's team deprioritized autoencoders for effort reasons rather than philosophical ones. Whether a frontier lab publishes a neuron-level attribution graph for a model it actually ships, rather than the one Goodfire published for Kimi K3. And whether the probe monitors keep their numbers on traffic Goodfire never saw.
One thing worth holding onto: the essay's own standard is specific, and testable. An agent read only a neuron's written explanation and wrote five examples meant to drive it plus five near misses meant to fool it, before any measurement; across 200 random neurons the hits fired the target 76 percent of the time and the near misses did 11 percent. The explanation won 164 times, tied 15, lost 4. An explanation that cannot lose is not an explanation — whatever happens to the grand thesis, that is the right shape for the argument.
A decade of interpretability research was built on evidence one of its authors now says was misread. Should the field rerun its foundational examples before building the next layer of tooling on them? Tell us in the comments.
