Fengyuan Hu comments on Normalizing Sparse Autoencoders