“Researchers at Anthropic, including Chris Ola, used a sparse autoencoder to identify individual neurons in a large language model with specific functions, such as a "Golden Gate Bridge neuron".”