“The amount of compute spent on reinforcement learning post-training for large language models is now approaching or surpassing the amount spent on pre-training.”