Recent successes in transformers and other architectures suggest that predictive learning may be superior to RL in creating rich representations of the world.

Great post! By the way, is there any material to support this sentence?