what they lack in pretraining, they make it up in reinforcement learning