Goodfire reports better neuron explanations from scaling agent investigations

AIGoodfire reports that scaling agent-driven investigations reliably improves its neuron explanations, as measured by simulation scores. It says each neuron described gives clues to the next, so its understanding of the network compounds.

AIGoodfire says it can generate short natural-language descriptions of neurons' roles. Given to a judge model, the descriptions predict when a neuron will fire and simulate its behavior. Goodfire says the method activates neurons more strongly than anything in its corpus.