Annah comments on Classifying representations of sparse autoencoders (SAEs)