Replying to
@thomasfuchs@hachyderm.io
@thomasfuchs@hachyderm.io it's interesting to consider *why* AlphaFold is/was good. My take is that 1) protein folding is basically impossible to predict without stochastic methods anyway, 2) close enough to correct is probably fine in a lot of cases because, 3) the outputs still have to be tested against reality. If you synthesise a protein for some practical use and AlphaFold was wrong about how it will fold such that it doesn't work, then it doesn't really matter.
AlphaFold being wrong sometimes isn't a dealbreaker because reality is the ultimate arbiter here. Unlike LLMs whose outputs can, in effect, define reality. "Computer says no" and all that.