"Another critical lesson is that humans are distinctly bad at monitoring automated processes".
Humans are also distinctly bad at noticing certain kinds of bugs in software. Think off-by-one errors, deadlocks, or any sort of bug you've stared at for days and not noticed the one missing or extra semicolon. But LLMs can generate a tsunami of subtly wrong code in the time a reviewer will notice one typo and miss all the rest.