This is a special post for quick takes by nonveumann. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Kernel of something that might inspire someone else who knows more than I.
Assuming weights that have “grokked” a task are more interpretable, is there use in modifying loss functions to increase grokking likelihood? Perhaps by making it path dependent on the updates of the weights themselves?