中国AI几乎追平美国，Kimi K3开源模型引发市场震荡

三周前，我在这里发出警告：https://open.substack.com/pub/garymarcus/p/china-catches-up?r=8tdk6&utm_medium=ios

“自2023年夏天以来，我在这里经常提出的‘无护城河 => 竞争者增多 => 价格战 => 利润稀缺’论点的最终结果已经到来——并可能毁掉美国的人工智能产业。”

上周，中国公司Moonshot.AI宣布了一个名为Kimi K3的模型，其整体水平与最好的美国模型相当，而且与那些模型不同，它是一个“开放权重”模型，消费者可以免费下载在本地运行（如果他们有支持的大规模硬件）。美国股市周五部分受此消息影响下跌，其影响可能巨大。这也不是偶然。中国的GLM 5.2模型由Z.ai推出，不久前也引起了巨大震动。阿里巴巴的新Qwen模型可能会进一步扰乱市场。

这一切对OpenAI和Anthropic的商业模式提出了严重质疑，可能会扼杀或严重削弱它们的首次公开募股（IPO）。这也让美国能够“赢得”人工智能竞赛这一本就不太可能的想法充满疑问。

Ryan Fedasuak（后来上CNN讨论此事）对此表示有些惊讶：

但没人应该感到惊讶。我在DeepSeek发布后于2025年1月写的一篇文章——上周有读者称其“可怕地准确”——几乎把一切都讲清楚了。https://open.substack.com/pub/garymarcus/p/the-race-for-ai-supremacy-is-over?r=8tdk6&utm_medium=ios 当时我警告说

如果美国继续主要关注大语言模型（LLM），他们不会在人工智能上取得决定性胜利。

我们最终会趋于平局。

缺乏技术护城河意味着像OpenAI这样的公司将保持不盈利状态。

CHIPS法案将适得其反，而无法有效遏制中国。

模型会继续变得更高效、成本更低，但幻觉和可靠性问题将继续存在。

不断围绕LLM竞赛会耗费原本可以投入开发更多原创想法的资源。

这基本上就是已经发生的情况。我希望更多人能够听从这些警告。

他们没有这么做，在缺乏与硅谷保持距离、且很少征求外部意见的政府中，这几乎是不可避免的。通过把硅谷的幻想当作现实对待，美国政府犯下了巨大的错误；经济现在岌岌可危，他们支持的公司似乎边缘化，而中国似乎有可能进一步削弱这些公司。我们本不应该一开始就全面押注生成式人工智能（GenAI），因为它似乎在技术护城河方面几乎没有太大优势。

一堆廉价的排外式对中国的轻率否定根本没带来任何结果。

问题是我们现在该怎么办？

首先，国会应该进行调查。美国是如何浪费了领先优势的？是否存在战略或战术上的失误？全面押注于一个可能被复制的单一技术是否愚蠢？美国公司在保护其知识产权方面做得够吗？中国政府是否在补贴中国公司，补贴程度如何？中国是否涉及间谍行为？美国公司做出的选择是否明智？投资美国公司可能赚钱还是亏钱？移民限制是否适得其反，导致人才流失？（Kimi的创始人去了卡内基梅隆大学后又回到了中国。为什么？）对于下一代人工智能或未来技术（如量子计算或核聚变），是否有可借鉴的经验？

但调查无法带我们走得太远。真正的问题是战略性的。

这里有七个特朗普政府可能考虑的选项，最后还有一个可能比前七个更有利于世界的“变数”。

什么都不做。让OpenAI和Anthropic自立更生。没有政府补贴，也没有秘密救助。如果OpenAI和Anthropic能找到生存之道，例如专注于狭窄的垂直领域，或通过“前向部署”关注个别大客户，那很好。如果它们倒闭，也无妨。谷歌、微软和亚马逊不会消失，并且可以满足美国的需求，尤其是那些中国模型可能不被允许涉及的机密需求。然而，这可能会将大量人工智能市场拱手让给中国，所以这并不是理想选择。

禁止开源。祝你好运；至少对于大型语言模型（LLM）来说，猫已经逃出袋子了。（即使它是非法的，到目前为止，它仍然会在暗网上以及那些没有禁止它的国家继续存在。

建立监管护城河，作为保护美国公司的后门手段。上周末，刚刚加入OpenAI的Dean Ball似乎认为事情会朝这个方向发展。1：#footnote-1 我希望不是这样。正如Will Manidis在这里所说：https://x.com/willmanidis/status/2078500818127315290?s=61，这将是保护主义，“……提议利用终端期、后期官僚国家的非正式、强制性权力，将美国市场上的更便宜的前沿竞争对手从OpenAI或Anthropic清除。”这也很可能会压垮美国的初创公司。用Matt Stoller的话说，这就是以“公司共产主义”代替开放和竞争的市场。当然，这会减缓进步并提高价格，从而以牺牲成千上万其他人的利益来帮助两家美国公司。

救助大型人工智能实验室。绝对不行：

全面禁止中国模型。Axios认为这可能正在酝酿中。这是上述措施的一种变体；肯定会提高消费者价格，并抑制创新。

在我看到上述传闻之前，我就在X上开了个玩笑：

正如投资人兼CEO Phil Libin今天早上在给我的消息中所说，“当资本、创意和人才自由流动时，美国体系会受益。我们不需要在零和游戏中‘获胜’。我们通过将游戏转变为非零和游戏来获胜。这就是我们赢得20世纪的方法，也是我们赢得21世纪的最佳机会。”

以每股两美分的价格收购大型人工智能实验室，如OpenAI和Anthropic——这已经算是慷慨的了，因为它们在一个尚未建立起盈利能力的行业中，基本都是亏损企业。将它们转变为国家实验室，通过竞争程序对所有具有资质的科学家开放。由于人工智能建立在大量未经补偿获取的人类知识产权之上，这里也存在一种诗意的正义。缺点是政府可能会用这些数据做一些不光彩的事情，比如大规模监控。考虑到Hegseth曾批评Anthropic拒绝这样做，并且考虑到现任政府倾向于以薄弱理由起诉批评者和敌人，我们应该担心政府滥用权力。

放弃“赢得”人工智能战争的零和思维，转而致力于将人工智能作为全球公共产品。在2017年《纽约时报》中（基于2016年在“AI for Good”会议上的一次演讲），我提出了“人工智能的CERN：https://www.nytimes.com/2017/07/29/opinion/sunday/artificial-intelligence-is-stuck-heres-how-to-move-it-forward.html?smid=nytcore-ios-share”，我写道

我羡慕我的高能物理同行，尤其是欧洲核研究组织CERN，这是一个巨大的国际合作组织，拥有数千名科学家和数十亿美元的资金。他们从事雄心勃勃、定义明确的项目（例如使用大型强子对撞机发现希格斯玻色子），并与全世界分享他们的成果，而不是仅限于某一国家或公司……一个国际人工智能计划……可以真正改善世界——如果它使人工智能成为公共产品，而不是少数特权人士的财产，这种改变将更加显著。

也许现在终于是时机？长期以来，实现这一目标的可能性似乎很低，但也许现在时机成熟。中国本身似乎对这一想法持热情态度，据习近平刚刚在一次人工智能会议上发表的讲话所示，这是他首次在人工智能会议上担任主角。

我不知道是否应该完全相信习近平——你可以在这里观看他的完整讲话：https://www.youtube.com/watch?v=ApCmqmhE1rg）——但是他这么说无疑为讨论提供了巨大的空间。我仍然认为将人工智能置于公共领域，并通过国际合作推动医学和科学的发展，是一个好主意。在我讨论过的各种选项中，这似乎是最有前景的，是将建立在大量互联网成果基础上的人工智能——这些内容主要来自公众的辛勤工作和思考，却几乎未获得报酬——转化为真正的全球公共财富的最后机会。

如果特朗普真想获得诺贝尔和平奖，这可能是他的机会。决定禁止闭源人工智能——要求所有制造商与认证的科学家和研究机构共享他们的权重、架构和训练数据——并与中国达成一项致力于将人工智能用于公共利益的协议——将是给予人类的一份巨大礼物。

加里·马库斯在2016年提出过一个“人工智能的CERN”计划，并在2023年的参议院会议上轻描淡写地提及过，看到这种想法的某种变体得以实现，他一定会感到非常高兴。

在后来的推文中，迪安声称他并不是建议美国采取监管壁垒将中国拒之门外，只是陈述其不可避免性。（X上的少数人相信他；大多数人认为他是在巨大的反对面前退缩。我对他的意图保持开放态度。）

加里，你完全正确。在大型语言模型/人工智能的热潮开始之前，没有人谈论机器学习军备竞赛，因为普遍认为机器学习是造福人类的工具。你用CERN做类比是合适的，原因相同。

感谢你的见解以及你不知疲倦地传递信息的努力。

除了理论上发财的机会之外，我实在无法理解为什么有人想要开发人工智能。我无忧无虑地度过了人生的前45年，以为我们都看过《终结者》和《2001太空漫游》，也以为真正聪明的人会明白风险远大于潜在的回报。人们到底怎么了，竟然会把一生投入到开发本质上是数字奴隶的东西上？真的有那么多人在过去的150年里就一直垂涎着拥有自己的奴隶吗？我意识到现有的人工智能并不具备意识，比喻并不完全贴切，但人工智能开发者和奴隶主的动机确实类似：他们都想要廉价劳动力，而不顾道德代价。或者说，他们只是想扮演上帝，通过创造“智能”来满足自己？还是说，这只是因为他们小时候经常被关进储物柜，现在不得不面对这种事情？

我是真心在问这些问题。如果Gary或者其他在评论中提到的人工智能研究人员能给我一些见解，我将非常感激。作为一名学术哲学家，研究重点是伦理学，我觉得所有这些事情都深刻地充满虚无主义且反人类。在一个体面的世界里，人工智能早就应该在摇篮中被扼杀了。

Three weeks ago, I warned here that：https://open.substack.com/pub/garymarcus/p/china-catches-up?r=8tdk6&utm_medium=ios

“The ultimate culmination of the “no moat => more competitors => price wars => profits are scarce” argument that I have been making here regularly since the summer of 2023, has arrived — and may wreck the U.S. AI industry”

Last week, the Chinese company Moonshot.AI announced a model called Kimi K3 that is largely on a par with the best American models, and unlike those models, it is an “open weight” model that consumers will be able to download to run locally (if they have the large-scale hardware to support it) for free. The US stock market dropped partly on this news Friday, and the repercussion are likely to be immense. It’s no fluke, either. Chinese model GLM 5.2 from Z.ai made shockwaves not long before. Alibaba’s new Qwen model may add to the disruption.

All this calls the business models of OpenAI and Anthropic into serious question, and may kill or greatly undermine their IPOs. It also casts serious doubt on the never-too-plausible idea that America might “win” the AI race.

Ryan Fedasuak (who later went on CNN to discuss) expressed some surprise at this:

But it’s not like anyone should be surprised. The essay I wrote in January 2025 after DeepSeek came out — which one reader last week called “scarily accurate” — pretty much laid it all out.：https://open.substack.com/pub/garymarcus/p/the-race-for-ai-supremacy-is-over?r=8tdk6&utm_medium=ios At the time I warned that

the US would not achieve a decisive victory in AI over China if they continued to focus largely on LLMs

that we would instead converge on a tie

that the lack of a technical moat would mean that companies like OpenAI would remain unprofitable.

that the CHIPS act would backfire, without doing much to hold China back

that models will continue to get more efficient and less expensive, but that hallucinations and reliability problems will persist.

that racing endlessly around LLMs would sap resources that could instead go into developing more original ideas.

That’s basically what has happened. I wish more people had listened.

That they didn’t is probably inevitable in government that lacked distance from Silicon Valley, and rarely consulted anyone for an outside opinion. By treating Silicon Valley fantasies as real, US government has massively blundered; the economy is now precarious, the companies they supported seem marginal, and China seems poised to undercut those companies further. We should never have gone all-in on GenAI, which never seemed to offer much in the way of a technical moat, in the first place.

A bunch of cheap xenophobic dismissals of China got us nowhere.

The question is what should we do now?

Well, to begin with, Congress ought to investigate. How did the US squander its lead? Were there strategical or tactical errors made? Was going all in on a single technology that could be replicated foolish? Did US companies do enough to protect their IP? Is the Chinese government subsidizing the Chinese companies, and to what degree? Is espionage from China involved? Are the US companies making good choices? Would investments in US companies likely make money or lose money? Have restrictions on immigration backfired, leading to a talent drain? (Kimi’s founder went to Carnegie Mellon and returned to China. Why?) Are there lessons to be learned for the next iteration of AI, or future technologies such as quantum computation or fusion?

But investigation won’t take us far. The real questions are strategic.

Here are seven options the Trump administration might consider, ending with a wild card that long-term might do more for the world than any of the first seven.

Do nothing . Let OpenAI and Anthropic stand on their own feet. No government subsidies or backdoor bailouts. If OpenAI and Anthropic can find a way to thrive, e.g., by focusing on narrow verticals, or “forward deployment” focusing on individual large customers, great. If they go under, so be it. Google and Microsoft and Amazon aren’t going anywhere, and can provide for US needs, especially classified ones where Chinese models presumably won’t be welcome. This may however concede a lot of the AI market to China, so it’s not ideal.

Outlaw open source . Good luck with that; at least with LLMs, the cat is already out of the bag. (And even if it were illegal, at this point it would continue to thrive on the dark web, and in countries that didn’t ban it.

Build a regulatory moat, as a backdoor way of protecting American companies. Over the weekend Dean Ball, who just moved to OpenAI seemed to argue that this is where things are headed. 1：#footnote-1 I hope not. As Will Manidis puts i：https://x.com/willmanidis/status/2078500818127315290?s=61 t, this would be protectionism, “… a proposal to use the informal, coercive power of the terminal, late-stage bureaucratic state to clear the American market of a cheaper frontier competitor to OpenAI or Anthropic.” This would also likely crush US startups. In Matt Stoller’s words, it is “corporate communism” in lieu of an open and competitive market. Certainly it would slow down progress and raise prices, helping two US companies at the expense of hundreds of thousands of others.

Bail out the big AI labs . Just no:

Ban Chinese models altogether. Axios thinks that might be in the works. Variant on the above; sure to raise prices to consumers, and stifle innovation.

Even before I read the above rumors, I made a joke about this on X:

As investor and CEO Phil Libin put it to me in a message this morning, “The American system benefits when there’s free flow of capital, ideas, and people. We don’t need to “win” a zero-sum game. We win by transforming games to non-zero-sum. This is how we won the 20th century and it’s our best shot of winning the 21st.”

Buy out the big AI labs, OpenAI and Anthropic, altogether , for two cents on the dollar — which is generous, since they are both mostly money-losing ventures in an industry that has yet to convincingly establish profitability. Turn them into national laboratories, available to all accredited scientists through a competitive process. Since AI is bult on human IP that has largely been taken without compensation, there is poetic justice here, too. The downside is what unsavory things the government might do with the data, e..g, in terms of mass surveillance. Given that Hegseth blasted Anthropic for refusing to do just this, and given the current administration’s tendency to prosecute critics and enemies on thin grounds, we should worry about government abuse.

Give up on the zero-sum effort to “win” the AI war, and instead move towards making AI a global public good . In 2017 in the New York Times (based on a talk in 2016 at the conference AI for Good) I proposed a “ CERN for AI：https://www.nytimes.com/2017/07/29/opinion/sunday/artificial-intelligence-is-stuck-heres-how-to-move-it-forward.html?smid=nytcore-ios-share ”, writing

I look with envy at my peers in high-energy physics, and in particular at CERN, the European Organization for Nuclear Research, a huge, international collaboration, with thousands of scientists and billions of dollars of funding. They pursue ambitious, tightly defined projects (like using the Large Hadron Collider to discover the Higgs boson) and share their results with the world, rather than restricting them to a single country or corporation… An international A.I. mission .. could genuinely change the world for the better — the more so if it made A.I. a public good, rather than the property of a privileged few.

Maybe now is finally the moment? For a long time the odds of pulling this off seemed long, but maybe the time is right. China itself seems warm to something of the sort, per a speech that Xi Jinping just made, headlining an AI conference for the first time:

I don’t know whether to take Xi at face value — you can view his full talk here：https://www.youtube.com/watch?v=ApCmqmhE1rg ) – but the fact that he said this gives a tremendous opening for discussion. I still think putting AI in the public domain, with an international effort towards medicine and science, would be a good idea. Of the options I have discussed, it seems to me to be the most promising, a last-ditch way to turn AI, built on so much of the internet that was sourced from hard work and thinking of the public, largely without compensation, into a true public good for the world.

If Trump actually wants to earn a Nobel Peace Prize, this could be his chance. A decision to outlaw closed source AI — requiring all manufacturers to share their weights, architectures and training data with accredited scientists and research organizations — combined with a deal with China to dedicate AI to the public good – would be a great gift to humanity.

Gary Marcus proposed a CERN for AI in 2016, echoing it gently at the Senate in 2023, and would be thrilled to see some variation on that idea see the light of day.

In a later tweet, Dean claimed he was not proposing that the US adopt a regulatory moat to keep China out it, just stating it as inevitable. (Few people on X believe him; most seem him as retrenching in the face of huge opposition. I leave his intent open.)

You are exactly right, Gary. Before the LLM/AI hype cycle began, nobody talked about a machine learning arms race, because it was widely recognized that ML is a toolset for the benefit of humanity. Your CERN analogy is appropriate for the same reason.

Thank you for your insights and your tireless efforts to inform.

Other than the theoretical chance of getting rich, I simply cannot understand why anyone wants to build AI at all. I blissfully went through the first 45 years of life thinking that we had all seen "The Terminator" and "2001" and that genuinely smart people would understand that the risks aren't worth the potential pay off. What the hell is wrong with people that makes them devote their lives to developing what are, in effect, digital slaves? Are there really that many people out there who have just been salivating for the last 150 years for the chance to get their own slave? I realize that existing AI isn't sentient and that the analogy isn't perfect, but the motivations of AI developers and slave owners are certainly similar: they both want a free labor supply, no matter the moral costs. Or, is it more that they just want to play God by creating "intelligence"? Are we just being made to deal with the fact that all these people constantly got stuffed in lockers when they were kids?

I mean these questions sincerely. If Gary or any other AI researchers in the comments could give me some insight, I would certainly appreciate it. As an academic philosopher with a research emphasis in ethics, I find all of this garbage deeply nihilistic and anti-human. In a decent world, AI would have been strangled in the crib.