Nick Cammarata argues neural networks are already clean, and Neel Nanda disputes it

Nick Cammarata, a former OpenAI interpretability researcher, reverses his 2020 conclusion that the InceptionV1 neuron 4e:55 is uninterpretable. He reports that with better tools, 99% to 100% of the last-layer neurons in modern vision models have clear explanations. Google DeepMind's Neel Nanda says the work offers no real breakthrough and questions its applicability to language models.