ChatGPT Is a Blurry JPEG of the Web

If you’re annoyed by the current hype around large language models (the kind of artificial intelligence that can write and chat with humans), or if you still haven’t quite understood how it works, this article is for you. Science fiction writer Ted Chiang offers a lens through which to understand what’s problematic about these models. And he offers a helpful proxy: We will know a large language model has become reliably good in quality when the output it generates will in return be used as training material for new models (which currently isn’t the case).