“Large language models can read 256 times faster than they can write.”