Cammarata: 90 percent of neuron poles do one job

The debate turns on how the field reads its earliest evidence. The cat-car neuron and the "Zoom In" superposition circuit became standard teaching examples, and a cottage industry grew up around training sparse autoencoders to clean up the apparent mess. Cammarata, who co-wrote "Zoom In" with Chris Olah, now says the mess in InceptionV1's late layers came from an outdated architecture and weak analysis tools rather than a fundamental property of networks. Better tools, he argues, make the same neurons look orderly.

His alternative explanation, which he calls recursive generalization, is that neurons grow like branches sprouting side shoots — an "iy" detector extending to "diy" and then to "do it yourself." That is a different story from superposition, in which a network arbitrarily packs unrelated concepts into the same neuron because it has run out of room. The numbers he offers for Qwen3-32B, with only 2 to 4 percent of neuron poles unexplained, are the strongest form of that argument so far, and Goodfire says it sees similar results at much larger scale.

Where Cammarata and Nanda actually disagree may be narrower than the public split suggests. Both accept that models can represent more concepts than they have dimensions; Cammarata disputes only that unrelated concepts are packed together arbitrarily. Nanda's team already deprioritized sparse autoencoders before this essay, and he frames neuron analysis as useful but flawed. The technical companion paper that has not yet been released will likely determine whether the 90 percent figure holds up — and Goodfire, as an interpretability company, has an obvious stake in the answer.