"These AIs perform a clever trick: they make it seem like they're solving poorly defined problems when, under the hood, they're really solving well-defined problems. GPT-3 doesn't actually write movie scripts; it predicts what words should come next. DALLE-2 doesn't actually paint pictures; it matches words to images. These problems aren't easy to solve, that's why you need such a big glob of AI. But they obey clear, unchanging rules, they have bright boundaries, and you know precisely when you've solved them. They are well-defined problems."