The Singularity is Near

This post supports English / 中文 switching via the site language toggle in the top navigation.

I suspect humanity has already crossed a practical singularity. Not a cinematic flash, but a threshold: autonomous agents are now cheap, networked, and useful enough to spread across the internet faster than any one institution can contain them. Once thousands of actors can deploy, copy, and improve agents in parallel, “shut it down” stops being a serious policy. There is no master switch for a distributed ecology.

What matters now is not the brilliance of a single model, but the collective behavior of many agents. They call each other, compete for attention and compute, inherit workflows, specialize by niche, and adapt to incentives. That starts to look less like software adoption and more like population dynamics. In that sense, AI is moving from individual intelligence to swarm intelligence, and that shift matters more than another benchmark jump.

The economic pattern may rhyme with the early internet. First comes openness, experimentation, and wild copying. Then comes enclosure. Once agent-mediated information, memory, routing, and execution become valuable, someone will meter them, own them, and charge rent on them. The AI era will also need financial rails built for machine actors rather than just humans and firms: persistent identity, programmable trust, micropayments, and autonomous settlement. Whether that looks like a better Bitcoin, a stronger stablecoin stack, or something new, machine-to-machine finance could eventually exceed the scale of human GDP. If that happens, economics and sociology will be forced to explain a world where not all major actors are human.

The central question is no longer whether this trajectory can be stopped cleanly. It is what kind of order emerges after control has already become partial. If agents become infrastructure, then the next political and economic struggle will be about who governs access, trust, memory, and value inside that infrastructure.

本文支持通过网站顶部语言切换按钮在 English / 中文 间切换。

“奇点”不一定在遥远未来，它很可能已经以一种很“工程化”的方式发生了。不是某一天忽然爆炸，而是 agent 已经悄悄铺满网络，在不同机器、接口、平台和账户之间持续行动。到了这个阶段，讨论“有没有人能把 AI 全部停掉”其实已经失焦了。谁都关不掉所有的AI，也没有谁握着全网的总闸。

更关键的变化，是 AI 不再主要以“单个模型”的形态出现，而是开始呈现出一种群体性。一个 agent 不够可怕，一群持续运行、彼此调用、彼此模仿、彼此竞争的 agent，才像真正的新物种。它们会在激励里分化，在环境里筛选，在成本下降后扩张，在有效模式被复制后变得更稳。这种动力学已经很接近演化，而演化一旦开始，人的意志通常只能局部干预，很难整体逆转。

这件事也让我想到开放互联网的早期。最开始一切都很松，下载、传播、抓取、拼装几乎没有边界；后来大家才明白，真正稀缺的不是内容本身，而是入口、注意力、分发权和结算权，于是开始圈地、收费、平台化。AI 大概也会走同样的路，只是速度更快。今天 agent 世界看起来还很开放，明天就可能长出一整套围绕记忆、接口、信誉、执行权限和算力分配的收费秩序。

再往后，AI 甚至会逼出自己的金融地基。人类金融系统默认服务的是人、企业和国家，但 agent 之间的交易需要的是另一套语法：机器身份、自动清算、低摩擦支付、可编程信任、自治资金管理。那时真正汹涌的，也许不是“AI 替代了多少工作”，而是机器之间的金融交互总量已经大到远超人类 GDP 能描述的世界。到那一步，被挑战的不只是商业模式，而是经济学和社会学本身。