Neural Networks Are Already Clean: What the Essay Says

"Neural networks are already clean" is the title and the thesis of a research essay Nick Cammarata of Goodfire published on October 15, 2026. It argues that the field's belief that neural networks pack unrelated concepts into the same neurons rested on misread evidence, and that most neurons in modern models can be described once the tools are good enough. This post reports what the essay says, keeps its numbers apart from each other, and lists what its promised companion paper would have to show. It is the first cluster of our guide to AI interpretability and whether the black box is opening.

What kind of document this is

The page header calls it "An informal research essay and working hypothesis about neural networks." That framing is the author's, and it should travel with every number quoted from it. The essay is a 65-minute read with interactive figures, written in the first person, and it ends by promising more: "A technical companion paper will offer a more detailed account of language model neuron experiments." Until that paper exists, the language-model results are one lab's preliminary audit, described by the person who ran it.

Cammarata co-authored the 2020 Distill article "Zoom In," which introduced the circuits framing and the superposition explanation for messy neurons. He later started the automated interpretability project at OpenAI. So when he writes "So there were two explanations for polysemanticity, and I helped propose both," he is describing a field he helped build, and reversing a position he held.

The claim, in the essay's own words

Here is the argument compressed to its load-bearing sentences. A polysemantic neuron is one that "appears to have multiple unrelated jobs or that looks impossible to understand." The famous example "responded to both cat faces and the fronts of cars." The leading explanation was superposition, which the essay defines as a new learning principle: "when a neural network runs out of neurons, it stops grouping related things together and instead begins folding unrelated things together, simply because they don’t interfere." If that is what networks do, you cannot study neurons directly. You have to train a second model, such as a sparse autoencoder, to approximate the untangled version and study that instead.

The essay's counter-claim: "The canonical evidence of polysemanticity and superposition that convinced much of the field, including me, that networks pack concepts arbitrarily was misread." What looked like arbitrary packing "was mostly a motif I’ll call recursive generalization." The motif is simple. A neuron starts on one thing and sprouts related offshoots over training, the way a branch sprouts sub-branches: "one neuron that detected “iy” grew to detect “diy,” and then “do it yourself,” which doesn’t contain “iy” at all."

On why early models looked so tangled, the essay names ordinary causes instead of a deep one. Architectural problems made later layers overflow. Weak tools amplified the mess. The 2023 explainer models "couldn’t run the loop of conjecture and test." Nobody knew what language-neuron motifs to expect, so people looked for the wrong things. Most studies used small models. And both humans and automated tools are, in his words, "strongly biased toward describing neurons as detectors for groups of nouns," when the neuron is often tracking a shared pattern instead.

The conclusion he draws is the one that made headlines: "This means we can study networks directly, instead of first untangling them."

The numbers, kept separate

Most coverage merged the essay's figures into one rising curve. They are not one curve. They come from different models, different tools, and different criteria, and the essay itself presents them as separate statements.

Two things follow. First, the vision-model progression is a story about tools and architecture on one small 2014 network, and it says nothing by itself about language models. Second, the Qwen3-32B figure is the one that matters for the models people use, and it is the one with the least published method behind it. The essay is candid about that in a footnote: a single cleanliness score "should be taken as a broad pointer to the regime a neural network is in, rather than something precise to either optimize directly or hold too tightly." In a smaller, closer look at strong activations, he reports a figure "closer to 97% non-arbitrary mixtures," then adds that those "are relatively small samples."

For contrast, the essay also quotes the two 2023 baselines it is arguing against: Anthropic's Towards Monosemanticity, which "gave the median neuron a 0 on its rubric out of a possible 14," and the OpenAI automated interpretability project, which "found about 1.7% of neurons interpretable by our most generous measure." The gap between 1.7% and 90% is the whole claim. The essay's explanation of the gap is better tools, bigger models, and knowing what to look for.

The four-step argument

Near the end of the summary, the essay states its case as four steps, and labels the last one. Networks use one learning principle, not two, so there is no blocker to ordinary science on them. With modern models and tools, that science goes better than expected. Understanding compounds, so it gets easier as the map fills in. Then: "At minimum, then, a network’s own mechanisms should be part of the toolkit." The next sentence carries its own warning label: "Highly speculative: this might be most of what we need, and interpretability might be far more tractable, possibly in the near term, than we’ve assumed."

Read the labels. The strong version is marked speculative by its author. The minimum version, neurons in the toolkit alongside decompositions, is close to what one critic, Google DeepMind's Neel Nanda, allowed when he told BankInfoSecurity (opens in a new tab) the approach "has some advantages over the standard alternatives."

Where AI helped look inside AI

One under-reported part of the essay is how the work was done: "Agents can map out neurons and circuits, build explorable microscopes and explore them, test explanations thoroughly, and track hundreds of thousands of attribution graphs." That cuts both ways, and the essay knows it. The 2023 tools failed, on its own account, because they rewarded "explanations that could simulate activations rather than ones that were right, which often come apart." An AI-generated description of a neuron is a hypothesis. It becomes evidence when a test could have broken it and did not, and the companion paper is where those tests would be laid out so someone else can run them.

What the essay does not claim

It does not claim the models are solved. On language models: "Those aren’t close to solved due to sheer size, but I’m optimistic about scaling the process." It does not claim every neuron has a tidy story, and it says forcing one is a mistake: "If a neuron might do a couple unrelated things and you’re not sure, it’s fine to say so." It does not claim the superposition question is closed everywhere, only that it should be "far down the list" of explanations to reach for, rather than the default. And in an interview with MIT Technology Review (opens in a new tab), the author said: "It’s not infinitely good news; it’s not everything solved."

The sharpest published pushback so far, from Liv Gorton, concerns the vision examples and the definition of superposition. We lay her tests next to the essay's account in the cat and car neuron: solved, or still open? and do not pick a side. Her broader point applies here: "Much of the essay’s other evidence comes from models and data readers can’t inspect, where the same method has to be taken on trust."

What the companion paper must show

This section is the update trigger for this post. When the companion paper is released, these are the questions a reader should bring to it. If it answers them, the Qwen3-32B figure moves from one lab's audit to a checkable result. If it does not, the figure stays where it is.

1. The rubric. What counts as "one main job," and who or what applied the label. The essay says the share moves with thresholds for weak activations, with depth, and with how you group related things. The paper should fix those settings and report the figure at more than one of them.
2. The sample. How many neuron poles were audited, how they were chosen, and whether the 97% figure from "relatively small samples" and the 90% headline figure come from the same procedure.
3. The tests, not just the descriptions. For named neurons, the synthetic-stimulus results, the ablations, and the steering outputs. A description that fits the top examples is rung one of the ladder in our pillar. A prediction that held on unseen inputs is rung two. The paper needs rung two to support the headline.
4. Something others can run. Qwen3-32B is an open model. Released neuron lists, prompts, and scoring code would let an independent group reproduce a sample of the audit. That is the difference between "we found" and "it is so."
5. The Kimi K3 results. The essay defers them entirely. "Similar results" up to trillions of parameters is the most consequential sentence in the piece and currently has no numbers attached.
6. The parts the essay flags as novel. The "double-negative regime" in SwiGLU neurons and the circuit motifs at different depths. These are presented as discoveries made possible by seeing most of a network. They should come with the same tests.
7. A response to the published disputes. The author has said he plans more studies of neuron 4e:55. Whether the paper revises, defends, or sets aside the V-shape reading tells a reader how the method handles a failed prediction.

None of this is hostile. It is the standard the essay sets for itself when it says the field's old tools rewarded explanations that simulated activations rather than ones that were right. The way to know which kind the new explanations are is to publish the tests.

The outside view: an unscored percentage is not a score

There is a smaller lesson in this for anyone who buys AI instead of researching it. "Roughly 90% of neuron poles have one main job" is a preliminary estimate with a stated method behind it that has not yet been published, from a source that says so plainly. Most numbers in AI vendor copy are worse: no method, no caveat, no source. The RedHub Standard has a rule for RedHub's own instruments called "no invented benchmarks." A graded instrument compares the inputs you supply to rules you can read, and the result is based solely on what you entered. That is the outside view: the number is yours, the rule is visible, and nothing is borrowed from a study you cannot open.

Apply the same test to the next interpretability headline you read. Is the rubric stated. Is the sample stated. Could anyone else run it. The essay passes the first two partially and promises the third. Hold it to the promise, and hold everyone else to the same one.