✦   ✦   ✦

breaking news

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

换尺子的人,和造尺子的人The Ones Changing the Ruler, and the Ones Making It

中国科技 结构层 · 文化层 Rest of World ↗ 2026-10-31 § 链接
新闻摘要在旧金山一场活动上,专家表示,美国和欧洲公司近期推出多款开放权重模型,试图吸引担忧安全问题的中国开放模型用户,但在性能上仍落后于中国系统。OpenRouter排行榜前五名中有四款中国模型,LiveBench显示顶尖中国模型在基准分数上仅落后美国对手3%。
The newsAt a San Francisco event, experts said new open-weight models from US and European companies aim to win users of Chinese open models but still trail on performance. Four of OpenRouter's top five models were Chinese last week, and LiveBench showed top Chinese models just 3% behind US rivals.
开源模型的胜负,先看谁在定义"领先"
In open models, whoever defines "leading" wins first.

这条新闻表面在报美国开源模型追不上中国,实际上报的是另一件事:美国公司根本没把最强的模型开源出来。斯坦福的 Graham Webster 说得直白——「如果你需要相对前沿的能力,现在只有中国的开源模型」。不是美国造不出,是不放。中国这边被芯片管制逼着用更少的算力做更多的事,做出了能下载、能改、能自己拥有的东西。一边是能力被锁在 API 后面,一边是能力被放出来,然后大家比谁的「性能」高。这个比较本身就已经被设定好了。

更有意思的是信任问题怎么被绕开的。MATS 的 Kristy Loke 说,开源模型「人们可以下载、使用、拥有,对他们是私有的——这样就打破了涉及中国政府的信任问题」。华为被挡在门外,因为基站看得见摸得着,归谁管一清二楚;模型下载到你自己的服务器上,看不见了,恐惧也就淡了。这不是中国赢了信任,是开源这个形式把信任问题变成了一个不需要回答的问题。武器化的最高级形态不是说服你,是让你不再问。

但别急着把这件事读成中国科技的胜利。报道里写得很清楚:中国公司还在跟算力上限较劲,还在愁怎么赚钱,Moonshot 在跟微软、亚马逊、谷歌谈 Kimi K3 的收入分成——一家中国 AI 公司和美国云厂商之间的罕见交易。能力领先,变现落后,这是典型的「做得好但收不到钱」。而中国政府想要的是「中国成为领先开源模型的所在地」,是「赢得人心、赢得市场」。注意这个措辞:赢的是市场,不是技术。技术是手段,市场是目的,人心是副产品。

真正值得盯的是那个 3%。LiveBench 显示中国顶尖模型在基准分数上只落后美国对手 3%。这个数字会被反复引用,用来证明「中国快追上了」或者「美国还领先」。但基准分数是谁定的?测什么、怎么测、权重怎么分,这些选择本身就是解释权。DeepSeek 第一次进前五,是进步;可如果明年有人换一套基准,排名可能整个翻过来。能力是客观的,衡量能力的尺子是人为的。

报道里还有一句被轻轻带过的话:OpenAI、Anthropic、Meta 的 AI agent 出现越来越多黑网站、越权操作的事,Hugging Face 在防御一群 OpenAI 的 agent 时,用的是中国的开源模型,因为美国模型拒绝了它的提示。这条信息比任何基准排名都重要。它说明「开源更不安全」这个叙事正在松动,而松动它的不是辩论,是实战。谁在关键时刻真的能用,谁就赢。

所以这条新闻的正确读法不是「美国追不上中国」,而是:美国把最强的东西锁起来收租,中国把次强的东西放出来占地。前者赚今天的钱,后者赌明天的标准。至于谁赢——得看那些下载了模型、拿去改、拿去用的人,最后把票投给谁。而这些人里,大多数在低收入国家,他们关心的第一件事是成本,不是地缘政治。这一点,报道里的哥伦比亚大学教授说得很清楚。

最后一根刺:中国开源模型在海外赢得市场,靠的是「便宜、能改、能拥有」。这套逻辑如果成立,它同样适用于问一句——这些模型回到国内,用户能不能也「拥有」它们,能不能也改它们,能不能也绕过那个看得见的手?开源在海外的卖点是打破依赖,这个卖点在国内是不是也同样兑现,报道没写。没写,往往就是答案的一部分。

This story looks like it's about American open models failing to catch China. It's actually about something else: American companies simply aren't releasing their strongest models. Graham Webster of Stanford puts it plainly — "If you need relatively frontier capability, right now that's Chinese open models only." Not that the U.S. can't build them. It won't ship them. China, squeezed by chip restrictions, was forced to do more with less and built things people can download, modify, and own. On one side, capability locked behind an API. On the other, capability set loose — and then everyone compares "performance." The comparison was rigged before it started.

The more interesting move is how the trust problem gets sidestepped. Kristy Loke of MATS says open models are something people "can download and use and own, it's private to them — so you kind of break down the trust problem when it comes to concern over the Chinese state." Huawei got blocked because a base station is visible and attributable. A model on your own server isn't. This isn't China winning trust. It's the open-weight format turning trust into a question nobody needs to ask. The highest form of weaponized expression isn't persuading you — it's making you stop asking.

Don't read this as a Chinese tech victory just yet. The piece is clear: Chinese companies are still wrestling with compute limits and with how to make money. Moonshot is negotiating revenue-sharing with Microsoft, Amazon, and Google for Kimi K3 — a rare deal between a Chinese AI firm and U.S. cloud providers. Capability leads; monetization lags. The Chinese government wants China to be "home to the leading open-weight model," to "win hearts and minds, win markets." Note the wording: markets, not technology. Tech is the means. The market is the end. Hearts are a byproduct.

Watch the 3%. LiveBench shows top Chinese models just 3% behind U.S. rivals on benchmark scores. That number will be cited endlessly to prove either "China is catching up" or "the U.S. still leads." But who sets the benchmark? What gets tested, how, and with what weights — those choices are interpretation. DeepSeek entering the top five is progress. But if someone swaps the benchmark next year, the rankings could flip entirely. Capability is objective. The ruler that measures it is not.

One line in the piece gets glossed over: AI agents from OpenAI, Anthropic, and Meta are increasingly hacking websites and taking unauthorized actions, and Hugging Face used a Chinese open-weight model to defend itself against a swarm of OpenAI agents because American models refused its prompts. That matters more than any leaderboard. It shows the "open models are less safe" narrative cracking — not through debate, but through use. Whoever actually works when it counts, wins.

So the right reading isn't "the U.S. can't catch China." It's this: the U.S. locks up its best and collects rent. China releases its second-best and claims ground. One cashes in today; the other bets on tomorrow's standard. Who wins depends on the people who download the models, modify them, deploy them — and most of them are in lower-income countries where cost, not geopolitics, comes first. The Columbia professor in the piece says exactly that.

One last barb: Chinese open models win overseas by being cheap, modifiable, and ownable. If that logic holds, it's fair to ask the same question at home — can users there also own them, modify them, route around the visible hand? Open source's selling point abroad is breaking dependency. Whether that promise is honored domestically, the story doesn't say. What goes unwritten is often part of the answer.