Jordan Taylor comments on SAE regularization produces more interpretable models