Sparse Autoencoders Work on Attention Layer Outputs — LessWrong