Except that is demonstrably true.

Two things can be true at the same time: I get value and a measurable performance boost from LLMs, and their output can be so stupid/stubborn sometimes that I want to throw my computer out the window.

I don't see what is new, programming has always been like this for me.