Goodfire kills a model's German by tuning one scalar on a single weight subcomponent

A sharp interpretability flex from Goodfire: in a one-day hackathon with their "AI neuroscientist" Silico, they removed a 67M-parameter model's ability to predict German by tuning a single scalar on one subcomponent of the weights — using just 4 German tokens. CEO Eric Ho's read: "we're getting closer to being able to hand-edit a weight and know exactly what it will do."

@ericho_goodfire on X →