Comment by kiitos
2 days ago
I'm not using "hallucinate" to mean "makes mistakes". I'm using it to mean "code that is syntactically correct and passes tests but is semantically incoherent". Which is the same thing that "hallucination" normally means in the context of a typical user LLM chat session.
Why would you merge code that was "semantically incoherent"? And how does the answer to that question, about "hallucinations" that matter in practice, allow you to then distinguish between "hallucinations" and "bugs"?