mechanistic-interpretability · 1 topics

goodfires silico finds human cognitive biases geometrically encoded inside llms