How can AI possibly be better than “the best” when the corpus of training data now includes its own slop in addition to all the code by new devs/lazy devs/bad devs scattered all over the internet? Law of averages applies here.

Because LLM models are obviously much more than the sum of their parts.