It’s not straightforward to encode either way, no. The issue with applying brute-force spaghetti code is the same that it usually is with spaghetti code: it’s very difficult to work out whether it’s doing the right thing, or how to fix it if it isn’t. I don’t use them, but my understanding is that, especially for non-standard tasks, LLMs are very good at producing spaghetti code at a large scale, which makes this issue worse.
Yes, having checks / automated testing would help with this, whether you’re using an LLM or not. I’ve mentioned writing unit tests for the edge cases as a pipe dream before, for exactly this reason. That’s still time-consuming, though, especially if you allow multicolour.
just to be clear: i’m not a big fan of the technology either. i fear that the effects on society and environment will be devastating. and i think it’s crazy to throw that much money and resources into the development of genAI. but if there is one use case for the technology, I see it in coding/software development and related fields. i recently used genAI to finally ditch Windows and switch to CachyOS for good - no way I’ll ever go back… so it empowered me in this case to cut the ties to big tech. i felt like a centaur (not like a reverse centaur - see the new book by cory doctorow for further explanation of the reference