“Anthropic's interpretability research identified a specific "Golden Gate Bridge neuron" that activates when the model processes text about the Golden Gate Bridge.”