is it possible to pretrain a language model using pure reinforcement learning from scratch? random weights, no cross-entropy loss pretraining. you may have many questions in your head.