One argument I heard from a college that the experimental llm written code (not production ofc) should be "write only" - you specify the requirements and constraints and let the machine figure it out - you specifically avoid "fixing it" or "cleaning it up" or reviewing the code because it disrupts the state of this discrete optimizer (over code). Not saying that it is right, or that it should be applied in prod, just noting that this is one potential way to go for non-critical code with clear "win conditions".