Researchers got deep learning models to train inside their own hallucinated ideas of the world.