本文肯定了对Transformer当前学习能力及局限性的分析框架,并指出对抗性世界模型是逼近现实本质的关键功能之一。作者认为,单纯增加参数和算力以扩展一个低效范式,将被能主动假设与验证真理的简洁方案所超越,尽管规模化可能因人类智能本身有限而意外通向AGI。引用推文补充了强化学习(RL)作为从干预中学习的范式,比监督学习更强大,而世界建模与RL的结合有望实现对反事实的学习。
Very well written blog. I think of RL as learning from interventions, and it kinda explains why it's more powerful as a ...