Is AI interpretability close to a breakthrough? A new article divides researchers

Mechanistic interpretability researcher Cammarata, in a new article, reinterprets classic findings such as polysemantic neurons and superposition, arguing that these most famous examples have been misread, sparking disagreement in the field. He reinterprets the neuron in InceptionV1 that responds to both cat faces and car fronts as a V-shape detector, treats the car hidden in the dog neuron in the paper Zoom In as ordinary generalization, and proposes a pattern of recursive generalization. He gives preliminary numbers for Qwen3-32B: about 90% of neuron poles have only one primary function, about 7% have two related functions, and 2% to 4% are unexplainable; Goodfire says it sees similar results on models with trillions of parameters. OpenAI employee Roon says this could be an important moment in the history of science, while Google DeepMind interpretability team lead Neel Nanda says it does not change the big picture and insists that superposition still exists. Cammarata believes interpretability is crucial to AI safety, otherwise it is hard to judge whether models are deceiving.