This is the same reasoning that Ilya Sutskever uses. That next token prediction objective causes the model to learn world models. Because being able to predict

This is the same reasoning that Ilya Sutskever uses. That next token prediction objective causes the model to learn world models. Because being able to predict the next token requires a lot of the knowledge/model of the world.