Goodfire puts neuron explanations to a 76%-versus-11% response test

Goodfire's 2026 research essay argues that neural networks may already be legible without first untangling them: many apparent mixtures reflect generalization across related patterns. In Neural Networ