LLMs are nothing like compilers. This sort of analogy based verbal reasoning is flimsy, and I understand why it correlates with projecting intelligence onto LLM output.