Goodfire's Nick Cammarata argues most neurons are legible, reinterpreting the superposition circuit from "Zoom In"

Superposition made a specific prediction. A network short of neurons packs unrelated concepts into the same ones, because they rarely get in each other's way ^[10]. Cammarata's essay, "Neural Networks Are Already Clean," goes after the evidence for that, case by case ^[1].

Take the InceptionV1 neuron. Pointed shapes such as fox snouts or flames activate it, and round shapes such as the letter O push its activation negative ^[11]. On that reading the neuron has one job. He interprets the "Zoom In" circuit, where a car seemed to hide among dog neurons, as ordinary generalization that groups cars and dogs by which way they face the camera ^[12].

The ambiguity that remains gets a different name, recursive generalization. A neuron that recognized the letter sequence "iy" later responded to "diy" and then to "do it yourself" ^[13]. That last string does not contain "iy", so a literal string match would have lost the neuron at the final step. Unlike superposition, Cammarata says, these groupings can teach you something about the network ^[14].

The numbers need care before they transfer. The 2023 figure counts neurons, and Cammarata's own summary calls the tools behind it "primitive" ^[3]. The Qwen3-32B figure counts neuron poles, and about 90 percent with one main job plus about 7 percent with two related ones comes to about 97 percent ^[17]. The unit and the tools both changed, so we would not treat the move from 1.7 to 90 percent as a like-for-like gain. For the 97 percent to carry to another model, the agents' judgment of "one main job" would have to hold under the same criterion there.

The method itself is old. Cammarata does not introduce a new technique. He goes back to studying neurons directly, with better tools, larger models and automated agents ^[4]. His summary frames it as a bitter lesson: "the simple thing, at absurd scale, works" ^[5]. The Decoder's report does not describe an independent replication or how the agents' labels were checked.

Cammarata's forecast goes past the data. He wrote: "I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving." ^[19]

Roon, an OpenAI employee who posts under that pseudonym on X ^[7], quote-tweeted the essay: "i think this may be remembered as a major moment in scientific history." ^[6] The Decoder summarizes Nanda's warning in a single sentence ^[8]. The LessWrong linkpost asks the skeptic's question: "How much of a real advance is this, vs things we already knew?" It adds: "How is this likely to be useful in practice for the purpose of alignment?" ^[9]