Your last paragraph doesn’t hold water.

Yesterday’s models produced okay small results, but unmaintainable large results. People bet that models would continue to improve, so that tomorrow’s model can fix today’s slop. This is a gamble, and it’s not clear yet whether it will pay off. If that progress stops for any reason, you end up stuck with a model that’s good at starting things, but not good at maintaining them, and a bunch of code that gets more and more expensive to maintain. https://maevyn.codes/decoherence is worth a read.

And that’s quite apart from the inherent decay due to information/reality drift.