dkirmani comments on [missing post]

[dkirmani]

Yep, but:

given only model weights

Also, the cossim-based approach should work on arbitrary submodules of a given model! Also it’s fast!