> but so much code writen is just the human brain doing the same autocomplete style work that LLMs can do now.

That's the part they are really good at. But they are really bad at taking complex decisions. Most of them are just guesses from a finite amount of solutions they were trained on, or from options they have in context.