“Jürgen Schmidhuber's paper "World Models" was the first to successfully demonstrate a reinforcement learning agent trained on imagined rollouts from a world model that performed well in a real environment.”