Large language models can do jaw-dropping things. But nobody knows exactly why. (MIT Technology Review, 03/04/2024)