“Will Brown argues that reinforcement learning allows researchers to trade compute for data, enabling them to extract more value from a smaller amount of high-quality human data.”