What's this about?
A widely-shared essay arguing you shouldn't treat LLMs like compilers — deterministic machines that turn a precise spec into correct code every time. AI coding is probabilistic and collaborative, not a build step.
What exactly is a "compiler"?
A program that translates human-written source code into machine instructions the exact same way every run — same input, same output, guaranteed. The essay's point is that a language model isn't that: give it the same spec twice and you can get different code, some of it wrong, so the workflow around it must assume variability.
What's the community saying?
HN 147 points, 156 comments
- "Who actually claims this?" — some argue nobody serious treats LLMs as compilers, so the essay attacks a straw man.
- A reframe: "it's an interpreter, actually" — running loosely rather than compiling exactly is a better mental model.
- A defense of human-in-the-loop collaboration as the real engine behind large AI-assisted builds, not the model alone.