“Large language models are intelligible because they are trained on human communication, even though their underlying mechanics at a bit-and-byte level are not understood.”