Is AI interpretability close to a breakthrough? A new essay has researchers divided
Nick Cammarata argues famous polysemantic neurons were misread, claiming most Qwen3-32B neurons have one job.
Nick Cammarata argues that classic mechanistic-interpretability examples were misread and that superposition is less dominant than researchers thought. He says the InceptionV1 “cat-car” neuron is a V-shape detector, and that in Qwen3-32B about 90 percent of neuron poles have one main job, 7 percent have two related jobs, and only 2–4 percent are unexplained. Goodfire reportedly sees similar patterns in models up to trillions of parameters, with bulk analysis done by AI agents. Neel Nanda of Google DeepMind welcomes the evidence as a useful tool but not a breakthrough, while Cammarata says readable networks would materially help AI safety.