Large language models generate text one token at a time.