@cocoaphony Those are applications of the Bitter Lesson as I read it, with the general point being that the methods that win make most effective use of increasing computation, not the ones that bake in human-crafted domain knowledge; better to exploit more compute than encode more human understanding.

If each new model is throwing new compute at the “problem,” then one may assume what requires encoding human understanding today—feedback, tests, Markdown—gets beat by more compute tomorrow.