LLMs can produce optimal software but they really hate to. They will always produce software that is the average. So if you are a 75th percentile engineer, it will push you towards the median. But if you know what's possible, you can make it produce that

this is why i think the posttraining chat models are bad, even though they are unfortunately what i have to work with. i want a lot more control over it, and they are too collapsed still