🤔 AI/LLM is not "new compiler"

I'm stumbling upon "AI is the new compiler, so there's no need to review generated code" too much recently.

That's basically false and harmful analogy. Compilers are deterministic; LLMs aren't.

I know speckit exists and I know that verification may be attempted at an interface boundary.

But!

First, spec always has gaps while compiler's input doesn't. In case of spec, any gaps are filled by LLM with something that sounds good.

Second, zero review is as safe as your test suite is complete, which for most real systems it isn't.

Non-functional requirements such as scope creep, security, performance, memory leaks, race conditions aren't caught by tests at all.

Compilers are a good exception, there is a mathematical verification possible and these are super-well-tested against all possible code.

LLMs are neither.

So no, AI is not the new compiler. A fast and capable collaborator but not a proven, deterministic translation layer.

Review the output. Test it. Check the constraints.

#ai #llm #codereview