I'm not sure what "10% performance gain" is supposed to mean here; but moving from "It does a decent job 95% of the time but screws it up 5%" to "It does a decent job 98% of the time and screws it up 2%" to "It does a decent job 99.5% of the time and only screws it up 0.5%" are major qualitative improvements.

Yeah I think that throwing more and more compute at the same training data produces smaller and smaller gains.

Maybe quantum compute would be significant enough of a computing leap to meaningfully move the needle again.