“The scaling of reinforcement learning, as demonstrated by DeepSeek, involves having a model generate answers and then grading the completion for accuracy, which serves as the reward signal.”