“The recent accelerated rate of AI progress is expected to slow down and revert to the long-run average as Reinforcement Learning (RL) compute becomes the majority of total training compute.”