tl;dr: A former OpenAI/Anthropic interpretability researcher argues that neural networks are far more directly interpretable than the field believed, and that the canonical evidence for "superposition" (networks packing unrelated concepts into shared neurons) was misread. The real phenomenon is "recursive generalization"—neurons branching into related sub-concepts as they grow—meaning researchers can study networks directly rather than first decomposing them with tools like sparse autoencoders. With modern architectures and better tools, roughly 90%+ of neurons in models up to trillions of parameters appear to have one main interpretable job.