The difference is that natural language doesn't have a well-defined semantics and is open to interpretation. Programming languages being compiled, on the other hand, have precise, well-defined semantics (excluding UB I guess) that you can formally reason about.
Using an LLM for programming is not like compiling, it's like employing a developer you either have to trust how they may happen to interpret your instructions, or to manually double-check their work.
I understand where you're coming from, but I didn't mean it quite that literally. More like, in the same way that compilers fundamentally changed how we program by adding a layer of (deterministic) abstraction, natural language assisted by AI could add a new (non-deterministic) layer of abstraction to programming.
Using an LLM for programming is not like compiling, it's like employing a developer you either have to trust how they may happen to interpret your instructions, or to manually double-check their work.