Kids outlearn AI—and we still don't know why 孩子的学习能力超越AI——而我们仍不知其原因

Human children achieve fluent language mastery from roughly 10–30 million words, while modern LLMs require trillions of tokens—creating a massive "data efficiency gap" Current LLM scaling strategies face an impending data ceiling, with easily available internet text potentially exhausted by the 2030s Reverse-engineering child language acquisition could yield more data-efficient AI models and serve underrepresented language communities The decades-old Chomsky-Skinner debate on innate grammar vers 人类儿童仅需约1-3亿词即可掌握母语，而现代LLM需训练15万亿token以上，存在巨大"数据效率差距" 互联网可训练数据可能在本世纪30年代前耗尽，数据效率成为AI发展的关键瓶颈 认知科学家正通过研究儿童语言习得机制，探索提升AI数据效率的新路径 语言习得理论争议（先天语法论vs环境学习论）可通过在机器模型中测试假设来推进