“The majority of AI training compute will eventually be dedicated to post-training reinforcement learning, a field that is still in its infancy.”