Christopher Olah comments on Mechanistic Interpretability for the MLP Layers (rough early thoughts)