“Large Language Models learn from text written by humans about the world, rather than learning directly from the world itself.”