Essay asks whether recursive self-improvement makes AI humanity's last invention
A 36Kr reposted essay by the WeChat account Lonely Brain (author Lao Yu) examines recursive self-improvement, defined as an AI system helping design and build a stronger successor system, with each generation depending less on humans. The author argues the important shift is not a higher intelligence level but a faster rate of change, and distinguishes this from simpler loops such as a model checking its own typos or an agent patching a few lines of code. The piece also argues that Amdahl's Law and physical constraints — GPUs, fabs, clinical trials — mean recursion does not automatically produce unlimited exponential growth, and it questions whether full AGI may face theoretical limits.
The essay frames RSI as turning research capability from an exogenous into an endogenous variable, while holding that verification and physical bottlenecks still cap any acceleration.
概述
这篇文章探讨了人工智能领域正在发生的递归自我改进现象,即AI开始深度参与制造比自身更强的下一代系统。
文章指出,这不仅会带来技术进化的指数级加速,也将彻底改变人类社会的研发范式和组织形态。
同时,作者深入分析了递归系统面临的验证瓶颈、局部最优和过拟合等真实约束。在人机协作的未来,人类的核心价值将从单纯的“执行”转向设定目标、保持审美判断力以及承担最终责任。
最后,文章强调无论技术如何演进,个人和组织都需要建立诚实的反馈闭环,跨越自我保护的自尊心门槛,保持持续更新的进化能力。
01
真正可怕的不是“AI 变聪明”,而是“AI 开始决定自己变聪明的速度”
最近 AI 领域最性感、也最让人类产生强烈自我怀疑的一个词,叫Recursive Self-Improvement(递归自我改进,简称 RSI)。
Some source content is omitted. Continue reading on the original site.