Ali Shehper comments on Sparse Autoencoders Work on Attention Layer Outputs