Several correlated but not strictly identical changes happened over the same few months: scaling inference compute, heavier use of RL in post-training, and models producing reasoning tokens.
大多数人可能将AI能力加速归因于单一因素(如模型规模增大),但作者指出这是多种因素共同作用的结果,包括推理计算扩展、强化学习在训练后阶段的使用增加以及模型生成推理标记的能力。这一多元归因挑战了单一因素决定论。