I have a completely different opinion, mostly because I working with 2 different code bases. What I have found is that it tends to use the same methodology as what it has to work with. And one code base is so bad, the ai just stacks bad on bad. And with out knowing underling data handling in things like games where speed count's the part that does the design has completely sucked.
HiTech
I agree. It definely does better with a clean slate where it can build it from scratch. And it does better with common tasks. General business programming. Doing so odd game based scripting code it has a lot more trouble because there is less training data available.
For instance it might have more trouble with a custom graphics engine than Unreal code because it has more training data for the latter.
But code is becoming less precious. It's almost getting to the point where you would tell it to review the code base, generate a inferred spec, then review and refine the spec with stakeholders, then tell it to just reimplement everything in an hour. In what ever language you want. In an hour. We're not there yet, but within 5 years.
I do sometimes suggest a structure I want. But I always ask first. I say what I want ot achive and ask first what we should do. Often he will suggest a better way than I was thinking and if you start with opinion, the sycophancy will drive it to try and satisfy you instead of saying, "Dude, you are approaching this all wrong. Let me tell you how this should be done." That might be while it piles bad code on bad code trying to follow the lead instead of say to scrap it and start over.
