So, there's a big qualitative difference in whether you can trust the output.

You can either "just believe", and prepare for inevitable, nasty surprises down the line. Or you can verify in ways you don't have to with stable compilers, eating up most of, if not more than all the efficiency gains you felt you had by using the LLM.

The two aren't comparable even remotely. One is a tool, the other is a slot machine. One allows for a new layer of abstraction, the other allows for a new layer of imprecision and hoping for the best.