Superposition is the idea that when a network runs short of neurons, it crams unrelated concepts into the same ones, producing a tangle that only a second model, such as a sparse autoencoder, can untangle. Cammarata now argues the famous evidence for this was misread. The InceptionV1 neuron that seemed to fire on both cat faces and car fronts is, in his view, really a V-shape detector: pointed shapes like fox snouts activate it, while round shapes push its activation negative. He similarly reinterprets the first superposition circuit from the paper "Zoom In", which he co-wrote with Chris Olah, as ordinary generalisation grouping cars and dogs by the direction they face.

As preliminary evidence, he offers numbers for the language model Qwen3-32B: roughly 90 percent of neuron poles have one main job, about 7 percent have two related ones, and only 2 to 4 percent remain unexplained, with Goodfire reporting similar results in models up to trillions of parameters. Most of the analysis was done by AI agents examining neurons in bulk. The reception is divided: OpenAI's Roon called it a moment that "may be remembered as a major moment in scientific history", while Google DeepMind researcher Neel Nanda warned against overhyping it. Roon himself noted the results still need confirmation by others in the field.