For a long time I couldn't pinpoint it to precise and concrete examples. I still can't but I have an analogy. You see how LLMs are pretty bad at writing, they just throw huge word salad at you, and, yes, you get the idea but looking at each sentence it's just gibberish. Well guess what? Code they generate is the same. Sure the model are now good enough that if you are lucky most of the time the code will do what is advertised, but I can't promise it because I can't wrap my head around it.