I’ve actually found that ChatGPT 4 has some sense of realistic expectations about what it can accomplish in the context of a conversation. For example, I asked

I’ve actually found that ChatGPT 4 has some sense of realistic expectations about what it can accomplish in the context of a conversation. For example, I asked it to implement polygon triangulation supporting polygons with holes in it, and it was like “Whoa now, I can tell you all about ear clipping, but that’s too big and complex to implement here”. ChatGPT 3.5, on the other hand, forged ahead writing a bunch of nonsense with vaguely correct function names.