I noticed the commonly used words list. Is this perhaps a failing of Diceware? I feel like people choose a passphrase, but then they tinker with it because they don't think they'll remember it, which leaves some words much more likely to be chosen and some words much less likely to be chosen.

So, instead of the 7776 words you end up with (well, still very many words) fewer.

I'm keen to see further work in optimisations (such as yours) and also into the psychology of choosing passphrases.

I agree that there comes a point where more GPUs is just "Meh, so what?"