How curious that these very different applications have all become similar tasks in the world of AI, approached through similar pieces of engineering. Comprehension questions are treated as translation questions, and translation models are further simplified into language models — essentially, predict the next word, given previous words. How could this interchangeability possibly work? How could next-word prediction approximate moral judgement? -- Nan Da, "Literary Critism in the Age of AI"