Tangentially, yes, let's imagine LLMs as compilers.
How insane is it to advocate the usage of these non-deterministic compilers, where each time you compile may produce different semantics?
And then people resort to saving and hand-editing the compiled output.
But when they want to change the source, they recompile and have to start over hand-editing the output again.
More than non deterministic : LLMs don't have a specification to obey to in the first place, while compilers (rather, programming languages) do.
Actually, in professional usage in a technical setting this is my prime objection to heavily LLM driven development. Were the tools in usage deterministic then I'd be a lot less objecting to the mandating of their incorporation into workflows.
I want to be reading, writing, testing and maintaining the software at the same layer. Right now extreme AI usage leads to reading, testing and maintaining happening in a less expressive language than writing and guess which of those four activities developers enjoy the least and find the hardest - it sure isn't writing.