✦   ✦   ✦

breaking news

News, read through The Primal Race
← 全部评论 · all commentary

智能体卖给你,基础设施留给你自己They Sell You the Agent, You Build the Network

科技 结构层 · 文化层 Rest of World ↗ 2026-11-01 § 链接
新闻摘要Meta的Muse和OpenAI的Dots等个人AI代理在西方兴起,Muse自9月8日以来在美加下载量达900万次。Meta计划全球推广,但专家指出发展中国家面临语言文化障碍、网络压力、数字鸿沟和监管难题。
The newsMeta's Muse and OpenAI's Dots personal AI agents are gaining traction, with Muse downloaded 9 million times in the US and Canada since September 8. Meta aims for a global launch, but experts warn developing nations face challenges like language barriers, network strain, digital divide, and regulatory gaps.
巨头把入口卖到全球,代价留给没有网络的人
Big Tech ships the entry point; someone else pays for the pipes.

Meta 的 Muse 在美加上线后下载 900 万次,日活 180 万;OpenAI 的 Dots 每月最低 100 美元。两家都在谈“把个人智能体带给全球数十亿人”。先把这句话翻译一遍:它们要卖的是认知入口,不是基础设施。而报道里那些真正决定这桩生意成败的东西——语言、带宽、光纤、电网、监管——没有一样是它们打算付钱的。

这篇报道最诚实的地方,是它自己把答案写出来了。南非的 ManyHands 做 Manny,光是让助手会说南非 11 种官方口语就已经很难。乌得勒支大学的 Payal Arora 说,语言包容性说到底关乎权力。这句话不需要我再加一个字。大模型在非英语语言上表现更差,于是“谁先用上智能体”这件事,从第一天起就是按语言排好的队。英语使用者先受益,不是因为他们的需求更迫切,而是因为他们的语言是模型训练时的默认设置。

然后是流量。思科把智能体叫“网络重度用户”,每个任务产生的流量比人类多出至多 450%;Cloudflare 网络上 AI 智能体的每日请求量过去一年涨了超过 1700%。这些数字落在东非这种还有余量的地方,暂时没事;落在尼日尔、布基纳法索、刚果东部,就是另一回事。非洲 IXP 协会的 Kyle Spencer 说,更大的麻烦是连接可靠性——修路会挖断光缆,海底光缆断了整个国家就断网。智能体在云上活着,可它够不到服务,也够不到用户。用户断了网,连一句“批准”都发不出去。

这就是我要说的机制:科技公司把“能力”卖给个人,把“能力成立的条件”留给公共部门、留给修路的施工队、留给海底光缆的维护船。加尔通的公式在这里很好用——Violence = Potential − Actual。同样一个智能体,在欧洲是效率工具,在尼日尔是一个会反复重试、把本就不宽的带宽吃干净、最后什么也没办成的请求。差额不是技术造成的,是资源分配造成的,而资源分配从来不进产品发布会。

数字鸿沟那一段更直白。国际电信联盟的数据:2025 年只有 36% 的非洲人用过互联网,欧洲是 92%,南北美洲 88%;非洲城市居民上网的概率是农村居民的 2.6 倍。卡内基的 Jane Munga 说,智能体若真在非洲运转起来,可能会加深不平等。她还点出了价格:Muse 免费,但很多智能体收费,Dots 每月最低 100 美元,这是多数全球南方用户吸收不了的成本。免费的那一个负责拉新,收费的那一批负责分层。分层不是副作用,是商业模式。

还有一层没人替小商贩算。智能体要能下单,本地商户就得把网站做得让机器人读得懂、买得动、付得了款——一个菜市场摊主要建库存、接支付、维护客服,还得应付智能体买得多、退得也多带来的压力。密歇根大学的 Kentaro Toyama 把这一点说得很清楚。做不到的商户不是“错过机会”,是被悄悄移出了数字经济的名单。名单的制定权在平台手里,商户只负责被筛选。

最后是监管。Future of Privacy Forum 的 Mercy King’ori 说,智能体越了解用户,表现越好;它可能从经手的数据里推断一个人的收入、健康、信用、行为。而当系统开发和托管都在境外时,监管者根本管不到。这句话值得单独抄下来:数据在本地产生,解释权在境外。这就是元暴力最现代的样子——不是有人拿枪逼你,是有人替你决定你的数据意味着什么,而你连申诉的窗口都找不到。

所以“发展中国家能不能接住几百万个智能体”这个问题本身是错的。该问的是:这几百万个智能体接住之后,谁付带宽、谁付语言适配、谁付小商户的改造费、谁付监管缺位的账。报道里没有答案,因为报道采访的每一个人都在描述代价,而没有一个人在承担代价。

Meta 的首席 AI 官说,很兴奋要把个人智能体带给全球数十亿人。兴奋是对的。数十亿人里,能上网的那 36% 会先被服务到;剩下的,等光缆修好再说。

Meta's Muse has been downloaded 9 million times in the U.S. and Canada since September 8, with as many as 1.8 million daily users. OpenAI's Dots starts at $100 a month. Both companies are talking about bringing personal agents to billions of people. Translate that sentence: they are selling the cognitive entry point, not the infrastructure. And everything that actually decides whether this works — language, bandwidth, fiber, regulators — is something none of them plan to pay for.

The report is honest enough to answer its own question. South Africa's ManyHands built Manny, and its founder Anton Moulder says getting the assistant to speak all 11 official spoken languages has been difficult. Payal Arora of Utrecht University puts it plainly: linguistic inclusion is ultimately about power. I don't need to add anything to that. Large language models underperform in non-English languages, so who gets served first is decided by language before it is decided by need. English speakers benefit first not because their needs are greater, but because their language is the default setting.

Then there is traffic. Cisco calls agents "network power users," generating up to 450% more traffic per task than a human; daily requests from AI agents on Cloudflare's network rose more than 1,700% over the past year. In East Africa, which has spare capacity, that is fine for now. In Niger, Burkina Faso, or eastern Congo, it is not. Kyle Spencer of the African IXP Association says the bigger problem is reliability: road construction damages fiber, and a damaged undersea cable cuts a country off. The agent lives in the cloud, but it cannot reach the services, and it cannot reach the user. If the user loses connectivity, they cannot even send the approval.

Here is the mechanism: tech companies sell capability to individuals and leave the conditions for that capability to public agencies, road crews, and cable repair ships. Galtung's formula fits — Violence = Potential − Actual. The same agent is a productivity tool in Europe and, in Niger, a request that retries, eats scarce bandwidth, and accomplishes nothing. The gap is not created by technology. It is created by resource allocation, and resource allocation never makes it into the launch event.

The digital divide section is even more direct. Per the International Telecommunication Union, only 36% of Africans used the internet in 2025, against 92% of Europeans and 88% of North and South Americans; city dwellers in Africa were 2.6 times more likely to be online than rural residents. Jane Munga of the Carnegie Endowment says agents might deepen inequality in Africa. She also names the price: Muse is free, but many agents are paid, and Dots costs at least $100 a month — costs most Global South users cannot absorb. The free one acquires users; the paid ones sort them. The sorting is not a side effect. It is the business model.

And nobody is calculating the bill for small vendors. For agents to place orders, local businesses must make their sites readable, buyable, and payable by bots — a vegetable market vendor would need inventory, payments, and customer support, plus the capacity to absorb agents that buy more and return more when human users decide the agent erred. Kentaro Toyama of the University of Michigan spells this out. Vendors who cannot do it are not "missing an opportunity"; they are quietly removed from the digital economy's list. The platform decides who is on the list. The vendor only gets screened.

Finally, regulation. Mercy King'ori of the Future of Privacy Forum says agents perform better the more they know the user, and may infer income, health, creditworthiness, or behavior from the data they handle. When the systems are developed and hosted abroad, regulators cannot supervise them. Copy that sentence down: the data is generated locally, the interpretation rights sit overseas. This is what meta-violence looks like in its most modern form — not someone holding a gun, but someone deciding what your data means while you cannot find a window to appeal.

So the question "can developing nations handle millions of AI agents" is the wrong question. The right one is: once those agents arrive, who pays for bandwidth, language adaptation, small-business retrofitting, and the regulatory vacuum? The report has no answer, because every expert it interviews describes the cost and none of them bears it.

Meta's chief AI officer says the company is excited to bring personal agents to billions of people. The excitement is real. Of those billions, the 36% who are online get served first. The rest can wait for the cable to be repaired.