Goodfire argues neural networks are more legible
Goodfire’s Nick Cammarata argues that apparent neuron “mixtures” often reflect ordinary, traceable generalization, not arbitrary feature packing. He reports preliminary audits finding roughly 90% of Qwen3-32B neuron poles have one main job, and says improved tools and architecture made 99–100% of a vision model’s final-layer neurons interpretable. These are his research claims, not settled results; he says a technical paper is forthcoming and notes uncertainty about applying the approach to real-world language-model behavior.
Treat this as a research hypothesis, not evidence that models are safe; watch for the technical paper and independent validation.