Ali Shehper comments on Sparse Autoencoders Work on Attention Layer Outputs

[Ali Shehper]

This could also be the reason behind the issue mentioned in footnote 5.