wassname comments on Interpretability with Sparse Autoencoders (Colab exercises)