Raw LLM Responses

@togowack Don't be so sure. I'm writing a DirectX 12 race car simulation engine and have found ChatGPT next to useless in that process. For DX12 code, more often than not it hallucinates methods and often entire classes that don't exist in the API, meaning almost none of it even works, let alone is optimal. Correcting it with different prompting doesn't make it go away in my experience. About all I've found it useful for is writing hlsl pixel shaders and so forth. It's pretty good at that, but not as good as a human. I had it write an hlsl Gauss-Jordan compute shader solver once. It worked but was a terrible design. Within five minutes of manual rewriting, I had a version that ran 70 times faster which went up from there as I threw it out and rewrote my own versions. Often it's good as a starting point to get the gears turning, but for me it's quicker just to write stuff myself than to try to get ChatGPT to do it most of the time. On the C++ side I don't even bother outside of very simple questions about specific functions which can be found in tutorials anyway. I treat it as a "quick Google" more than anything. I've found it helpful at times, mostly in debugging small sections of code when I'm tired and making dumb mistakes, but that's about it. I think what we'll see more than anything in the near term is a whole new industry akin to life coaching who teach prompting, endlessly responding to criticism with statements like "you are asking it the wrong questions, for only $x/month I'll show you how to get the results you want." Courses were popping up only a couple of weeks after GPT 3.5 was released promising exactly that. People sure became experts fast. 🙄