Also, there are multiple "correct" ways to code something, so imperfect code that solves the problem is still useful. A medical diagnosis is either correct or incorrect.
Different prompt approaches and training doctors to use LLMs can improve accuracy of LLM-assisted diagnosis. It’s pretty reasonable to hypothesize that LLM “peer review” could improve that as well.
