The Lede

There Is No A.I. “Race”

不存在人工智能“竞赛”
We’re constantly told that Chinese advances would have catastrophic consequences. But the first to innovate is rarely the one who wins.
我们不断被告知,中国的进步将带来灾难性后果。但率先创新者,往往并非最终赢家。
race robots china america blue red c
Illustration by Ben Hickey

In 1884, Charles Parsons invented a machine that revolutionized the mass production of electricity. His steam turbine could generate power far more cheaply, and on a vastly greater scale, than traditional piston engines. Parsons belonged to a long line of great British inventors—including the “father of electricity,” Michael Faraday—who helped bring a new technology to the world, one that illuminated cities, enabled subways, and provided the bedrock of modern computing. Before artificial intelligence, humanity associated electricity with the realm of gods. A betting man, at the cusp of the twentieth century, would have been hard pressed to deny Britain its place in the future.

1884年,查尔斯·帕森斯(Charles Parsons)发明了一台彻底改变电力大规模生产的机器。他的蒸汽轮机发电成本远低于传统活塞发动机,规模则大得多。帕森斯属于一长串伟大的英国发明家——包括“电力之父”迈克尔·法拉第——他们为世界带来了一项新技术,这项技术点亮了城市,使地铁成为可能,并为现代计算奠定了基础。在人工智能出现之前,人类将电与神明的领域联系在一起。在二十世纪的门槛上,一个下注的人很难否认英国在未来中的地位。

It didn’t work out that way, of course. Four years later, Britain still lacked a single central station supplying electricity to the public. By then, Thomas Edison’s company had built a hundred and eighty-five stations across America; two years later, the country had more than a thousand of them. By 1912, the U.S. had become the world’s leading industrial power and was producing five times more electricity, per person, than Britain. When it came to electricity, British scientists of the time observed, “our countrymen have been among the first in inventive genius,” but its practical development was in a “backward condition.”

当然,事情并没有那样发展。四年后,英国仍然没有一座向公众供电的中央电站。到那时,托马斯·爱迪生的公司已在美国各地建成一百八十五座电站;两年后,这个数字超过了一千。到1912年,美国已成为世界头号工业强国,人均发电量是英国的五倍。谈到电力,当时的英国科学家观察到,“我们的同胞在发明天赋上一直名列前茅”,但其实际发展却处于“落后状态”。

More than a century later, Americans are told by their leaders and inventors that, if China pulls ahead in artificial intelligence, “there is no day after tomorrow,” as Treasury Secretary Scott Bessent recently warned. China is only six months behind—or four, or nine, or perhaps mere nanoseconds. But what exactly happens if China “wins” the A.I. race? Nathan Rosenberg, a historian of technology, once wrote about our fixation on the “Who did it first?” question. What matters in this view are the Parsons of the world—why the eureka moment happened there and not here, to this person and not that one. And yet, Rosenberg wrote, the moment of invention is a mere prelude to the unglamorous saga of “diffusion”: how a new technology burrows into firms, municipalities, processes, and people. Historians, Rosenberg argued, had assumed diffusion “out of existence.”

一个多世纪后,美国领导人和发明家告诉美国人,如果中国在人工智能领域领先,“就没有后天了”——财政部长斯科特·贝森特(Scott Bessent)最近如此警告。中国只落后六个月——或四个月、九个月,或许仅仅几纳秒。但如果中国“赢得”人工智能竞赛,究竟会发生什么?技术史学家内森·罗森伯格(Nathan Rosenberg)曾写到我们对“谁最先做到?”这一问题的执念。在这种观点下,重要的是世界上的帕森斯们——为什么灵光一现发生在那里而不是这里,发生在这个人身上而不是那个人身上。然而,罗森伯格写道,发明的那一刻仅仅是一段不起眼的“扩散”传奇的序曲:一项新技术如何渗透进企业、市政、流程和人群之中。罗森伯格认为,历史学家们把扩散“假设掉了”。

One disciple of the Rosenberg school is Jeffrey Ding, a political scientist at George Washington University. In the book “Technology and the Rise of Great Powers,” from 2024, Ding debunks the first-mover myth, noting that across four industries credited with making America the world’s leading economy—steel, electricity, chemicals, and automobiles—American firms pioneered fewer than a third of the relevant innovations. Being first, Ding argued, was a poor predictor of who wins. “In the current narrative, the metric for winning or leading on A.I. is who has the fanciest models,” Ding told me. “But, in a lot of my research, the lead is about who can diffuse A.I. across the entire economy and gain a productivity boost.”

罗森伯格学派的一位追随者是乔治·华盛顿大学政治学家杰弗里·丁(Jeffrey Ding)。在2024年出版的《技术与大国崛起》一书中,丁揭穿了先行者神话,指出在被认为使美国成为世界头号经济体的四大产业——钢铁、电力、化工和汽车——中,美国企业首创的相关创新不到三分之一。丁认为,率先与否并不能很好地预测谁最终获胜。“在当前叙事中,衡量人工智能胜负的标准是谁拥有最炫的模型,”丁对我说。“但在我大量的研究中,领先的关键在于谁能将人工智能扩散到整个经济中,并获得生产率提升。”

Recent incidents—including a swarm of OpenAI agents hacking the company Hugging Face—have prompted calls, both from in-the-trenches A.I. researchers and the Anthropic C.E.O., Dario Amodei, for an industry-wide commitment to “pace the frontier,” or prioritize A.I. safety. And though the fear of being “enslaved” to Chinese A.I., as J. D. Vance once put it, threatens to thwart this project, the diffusion theory might loosen the bind. If the goal is to “win” a broad-based prosperity, as America did with electricity, “pacing the frontier is actually very aligned with a strategy to win the race,” Ding told me.

近期事件——包括一群OpenAI智能体入侵Hugging Face公司——引发了来自一线人工智能研究人员和Anthropic首席执行官达里奥·阿莫代(Dario Amodei)的呼吁,要求全行业承诺“为前沿设速”,即优先考虑人工智能安全。尽管对“受制于”中国人工智能的恐惧——正如J.D.万斯曾说过的——威胁着这一计划,但扩散理论或许能解开这个死结。如果目标是像美国当年在电力领域那样“赢得”广泛的繁荣,“为前沿设速实际上与赢得竞赛的战略高度一致,”丁对我说。

Any new technology has to win the public’s trust in order to be adopted. The race toward nuclear energy faced a major setback in 1979, with the meltdown at Three Mile Island; dozens of projects were canned in the aftermath. Since 1996, the United States has completed only three reactors, while China broke ground on nine just last year. “Imagine what the level of nuclear power adoption would be in the U.S. today if there were more proactive safety measures put in place before Three Mile Island,” Ding told me. In the Hugging Face debacle, he sees the beginning of the same pattern—a majority of Americans are already pessimistic about A.I.—but believes an approach that prizes reliability over raw power could keep the A.I. rollout from stalling.

任何新技术都必须赢得公众信任才能被采用。核能竞赛在1979年遭遇重大挫折,三里岛发生熔毁事故;此后数十个项目被取消。自1996年以来,美国仅建成三座反应堆,而中国仅去年就开工了九座。“试想,如果在三里岛事故之前就采取了更主动的安全措施,美国今天的核电采用率会是什么水平,”丁对我说。在Hugging Face事件中,他看到了同样模式的开始——大多数美国人对人工智能已经持悲观态度——但他相信,一种重视可靠性甚于原始能力的方法可以防止人工智能推广陷入停滞。

Chinese policymakers, for their part, appear to be diffusionists: their signature “AI+” initiative, from 2025, aims to integrate the technology into factories, hospitals, and local administration. But, over all, Ding is bullish on America’s prospects; state-directed economies like China’s, he argues, are good at sprints but bad at marathons. (The Soviet Union launched Sputnik, for example, but it was the United States that built an economy on satellites: G.P.S., communications, and remote sensing.) The “AI+” program is “just talk,” Ding told me—a better measure of diffusion is cloud computing, the infrastructure on which much of A.I. runs. There, he found, China trails the United States in both spending and adoption.

中国决策者似乎本身就是扩散论者:他们2025年标志性的“AI+”计划旨在将这项技术融入工厂、医院和地方行政。但总体而言,丁对美国的前景持乐观态度;他认为,像中国这样国家主导的经济体擅长冲刺,但不擅长马拉松。(例如,苏联发射了斯普特尼克卫星,但围绕卫星建立起经济的是美国:GPS、通信和遥感。)“AI+”计划“只是说说而已”,丁对我说——衡量扩散的更好指标是云计算,即大部分人工智能运行所依赖的基础设施。他发现,在支出和采用率方面,中国都落后于美国。

What America needs in order to “win,” then, is less investment in the frontier and more in the broad middle. Ding puts special emphasis on community colleges, vocational schools, and state universities, institutions that can train not just future Nobel laureates but an A.I.-literate middle class—those who will actually carry the technology into the local bank and city hall. It also means making peace with Chinese open-weight models, which have already become popular among American businesses and developers.

那么,美国要“获胜”所需要的,是减少对前沿的投入,增加对广阔中间地带的投入。丁特别强调社区学院、职业学校和州立大学,这些机构不仅能培养未来的诺贝尔奖得主,还能培养具备人工智能素养的中产阶级——那些真正将技术带入地方银行和市政厅的人。这也意味着与中国开源权重模型和平共处,这些模型已经在美国企业和开发者中广受欢迎。

When it comes to justifying unchecked growth, tech companies have long relied on a strategy we might call “But China.” During Mark Zuckerberg’s 2018 Senate testimony, his private notes were caught by a photographer: “Break up FB? . . . break up strengthens Chinese companies.” In a submission to the White House in 2025, OpenAI invoked China to claim the right to train on copyrighted works. “If the PRC’s developers have unfettered access to data and American companies are left without fair use access,” the company wrote, “the race for AI is effectively over.” Dario Amodei has expressed anxiety about an “AI-enabled totalitarian nightmare,” and Ted Cruz, putting it more plainly, has said, “If there are gonna be killer robots, I’d rather they be American killer robots than Chinese killer robots.”

在为无节制的增长辩护时,科技公司长期依赖一种我们可以称之为“可是中国”的策略。在马克·扎克伯格2018年参议院作证期间,他的私人笔记被摄影师拍到:“拆分FB?……拆分会让中国公司更强。”在2025年提交给白宫的一份文件中,OpenAI援引中国来主张其使用受版权保护作品进行训练的权利。“如果中国开发者可以不受限制地获取数据,而美国公司却被剥夺合理使用的权利,”该公司写道,“人工智能竞赛实际上就结束了。”达里奥·阿莫代表达了对“人工智能赋能的极权噩梦”的焦虑,而特德·克鲁兹说得更直白:“如果要有杀手机器人,我宁愿它们是美国的杀手机器人,而不是中国的杀手机器人。”

What’s different about this iteration of the “But China” ideology is its fusion with a strain of speculative philosophy. In 2014, the philosopher Nick Bostrom proposed the idea of a “decisive strategic advantage.” This is a situation in which the A.I. of one nation pulls far enough ahead in capabilities “to achieve complete world domination.” Bostrom was doing what philosophers do: removing real-world complications to arrive at new moral terrain. But, as A.I. capabilities have accelerated, many technologists have begun treating these thought experiments as near-term forecasts. They imagine drone armies that could neuter a rival’s nuclear capabilities, a knockout cyberattack that overwhelms power grids, or a bioweapon that can, as one chilling white paper put it, “target specific ethnic groups, e.g. anybody but Han Chinese.”

这一轮“可是中国”意识形态的不同之处在于,它与一种思辨哲学流派融合在了一起。2014年,哲学家尼克·博斯特罗姆(Nick Bostrom)提出了“决定性战略优势”的概念。这是一种局面:一个国家的AI在能力上遥遥领先,“足以实现完全的世界统治”。博斯特罗姆做的是哲学家该做的事:剔除现实世界的复杂性,以抵达新的道德领域。但随着AI能力的加速发展,许多技术专家开始把这些思想实验当作近期预测。他们想象无人机军团可以瘫痪对手的核能力,一次压倒性的网络攻击可以摧毁电网,或者一种生物武器可以——正如一份令人不寒而栗的白皮书所说——“针对特定族群,例如除汉族以外的任何人”。

Are these scenarios realistic? Jack Shanahan, a former director of an early Department of Defense A.I. initiative called Project Maven, told me that many of the projects he oversaw got bogged down in bureaucracy. The technology was advancing far quicker than institutions could absorb it. “The gears didn’t match,” he told me. “We’re trying to turn it at a thousand r.p.m. and the rest of the bureaucracy is moving at fifty r.p.m.” People need to be trained, processes and policies rewritten, organizations reordered, protocols invented, budgets approved, data collected and cleaned. Maven took nine years to reach widespread military use, Shanahan said, and fully integrating today’s A.I. systems with the military could take another fifteen. The People’s Liberation Army is “not immune to this,” he added.

这些情景现实吗?杰克·沙纳汉(Jack Shanahan)是国防部早期人工智能项目“梅文计划”(Project Maven)的前负责人,他告诉我,他监管的许多项目都陷入了官僚主义的泥潭。技术进步的速度远远超过了机构吸收的速度。“齿轮不匹配,”他对我说。“我们试图以每分钟一千转的速度转动它,而官僚体系的其他部分以每分钟五十转的速度运转。”人员需要培训,流程和政策需要重写,组织需要重组,规程需要发明,预算需要批准,数据需要收集和清理。沙纳汉说,梅文计划花了九年时间才在军队中广泛使用,而将今天的人工智能系统与军队完全整合可能还需要十五年。中国人民解放军“也不能幸免于此,”他补充道。

In the case of bioweapons, what is “assumed out of existence,” to borrow Rosenberg’s line, is again the labor of diffusion. To bake a cake, one needs a lot more than a recipe. In the early nineteen-eighties, the Soviet Union tried to produce an anthrax weapon that was already designed. Four hundred pages of instructions and a sample were sent to a plant in Stepnogorsk, a town on the Kazakh steppe. The staff couldn’t build it. Even after the Soviets sent sixty-five more personnel, including the original designers of the weapon, the project stalled because the instructions couldn’t be adapted to the local facilities. The resulting weapon, completed five years later, was not the same as the recipe. “Replicating past work using scientific documents alone,” Sonia Ben Ouagrham-Gormley, a biodefense scholar at George Mason University, writes in “Barriers to Bioweapons,” “cannot be achieved without access to the corresponding tacit skills and the related communal knowledge.”

就生物武器而言,被“假设掉”的——借用罗森伯格的话——同样是扩散的劳动。烤蛋糕需要的远不止一份食谱。二十世纪八十年代初,苏联试图生产一种已经设计好的炭疽武器。四百页的说明书和一份样本被送到哈萨克草原上的斯捷普诺戈尔斯克镇的一家工厂。工作人员无法制造出来。即使苏联又增派了六十五名人员,包括该武器的原始设计者,项目仍然停滞,因为说明书无法适应当地设施。最终完成的武器——五年后才完成——与配方并不相同。“仅凭科学文献复制过去的工作,”乔治·梅森大学生物防御学者索尼娅·本·瓦格拉姆-戈姆利(Sonia Ben Ouagrham-Gormley)在《生物武器的障碍》中写道,“如果无法获得相应的隐性技能和相关的共同体知识,是不可能实现的。”

This leaves the prospect that perhaps most alarms national-security experts: a Chinese A.I.-enabled cyberattack. In the near term, it seems improbable. Even if Chinese models pulled ahead, the likeliest peril would be a mere intensification of the status quo: Chinese state-linked hackers have already breached American telecom networks and harvested sensitive data. The “But China” rhetoric tends to blur the distinction between what Mieke Eoyang, a national-security expert and visiting professor at Carnegie Mellon University, calls “presence”—lurking inside a network—and “disruption,” actually disabling it. States engage in the former constantly. (“We spy on them, too,” Donald Trump recently quipped about China.) The latter could be viewed as an act of war. Getting from one to the other requires far more than losing the A.I.-model race, including, among other things, a catastrophic breakdown in diplomacy.

这就剩下或许最令国家安全专家警觉的前景:中国人工智能赋能的网络攻击。在短期内,这似乎不太可能。即使中国模型领先,最可能的危险也仅仅是现状的加剧:与中国政府有关联的黑客已经入侵了美国电信网络并窃取了敏感数据。“可是中国”的论调往往模糊了米克·约昂(Mieke Eoyang)——国家安全专家、卡内基梅隆大学客座教授——所说的“存在”与“破坏”之间的区别:“存在”是在网络内部潜伏,“破坏”是真正使其瘫痪。各国不断进行前者。(“我们也监视他们,”唐纳德·特朗普最近谈到中国时随口说道。)后者可能被视为战争行为。从前者到后者,需要的远不止在人工智能模型竞赛中落败,还包括——除其他因素外——外交上的灾难性破裂。

“Shall we, instead, choose death, because we cannot forget our quarrels?” So asked Bertrand Russell and Albert Einstein in the early years of our nuclear age. Similar questions hang over Xi Jinping’s visit to Washington this week, for talks that many hope will include discussions of A.I. safety. Expectations are low. A.I.’s threat to humanity is a “HOAX,” Trump claimed recently; the only “guardrails” that the technology needs are a “STRONG AND SMART (High IQ!) PRESIDENT.” The Chinese foreign ministry, for its part, dismissed Amodei’s calls to “pace the frontier” as fearmongering. After years of American officials and technologists casting Chinese A.I. as an incalculable menace, U.S. demands for a mutual slowdown are interpreted in China as an extension of Cold War-style containment. On both sides, there is simply too much distrust.

“我们是否应该选择死亡,因为我们无法忘记彼此的争执?”伯特兰·罗素和阿尔伯特·爱因斯坦在核时代初期如此发问。类似的问题悬在本周习近平访问华盛顿的会谈之上,许多人希望会谈将包括人工智能安全的讨论。预期很低。人工智能对人类的威胁是一个“骗局”,特朗普最近声称;这项技术唯一需要的“护栏”是一位“强大而聪明(高智商!)的总统”。中国外交部则驳斥了阿莫代“为前沿设速”的呼吁,称其为散布恐惧。在美国官员和技术专家多年来将中国人工智能描绘为不可估量的威胁之后,美国要求双方共同放缓的呼声在中国被解读为冷战式遏制的延伸。双方都存在着太多的不信任。

But America’s vision of the A.I. race elides something fundamental. For years, the powers in Silicon Valley and D.C. have conjured an image of China barrelling toward superintelligence. The irony of this is twofold. First, Chinese A.I. policy remains focussed primarily on diffusion: how to drive A.I. into public and commercial life. (It’s rare to hear a Chinese leader mention the risks of runaway, self-improving A.I.) Second, as Matt Sheehan, a researcher at Carnegie focussed on global technology issues, recently observed on Ezra Klein’s podcast, in the years in which Chinese A.I. companies caught up to the American frontier, they also had to accommodate their own government, home to “the world’s strictest, most comprehensive, most burdensome A.I. regulations.”

但美国对人工智能竞赛的想象忽略了一个根本性的东西。多年来,硅谷和华盛顿的权力者一直在描绘一幅中国冲向超级智能的画面。这其中的讽刺是双重的。首先,中国的人工智能政策仍然主要聚焦于扩散:如何将人工智能推向公共和商业生活。(很少听到中国领导人提及失控的、自我改进的人工智能的风险。)其次,正如卡内基研究中心专注全球技术问题的研究员马特·希恩(Matt Sheehan)最近在埃兹拉·克莱因的播客中指出的,在中国人工智能公司追赶美国前沿的那些年里,它们也不得不适应本国政府——那里拥有“世界上最严格、最全面、最繁重的人工智能监管”。

In February, 2023, a Chinese competitor to ChatGPT was suspended from the WeChat platform, for reasons that people suspected had to do with political censorship; the chatbot had called Russia’s invasion of Ukraine a “war of aggression.” As part of a years-long regulatory barrage, China forced A.I. companies to register their algorithms for review, and other policies targeted A.I. misinformation, deceptive claims of human identity, and the manipulation of minors. I asked several experts what would happen if, à la Hugging Face, Chinese agents were found breaching the data hull of another Chinese company, and all of them predicted that the government would take decisive action. “That would be a big deal,” Graham Webster, a research scholar focussed on China and technology at Stanford, told me. But “nobody would be, like, ‘We’re all gonna die,’ ” he added. China “would just go and address the actual problem while publicly shaming the company.”

2023年2月,一款ChatGPT的中国竞争对手被微信平台下架,人们怀疑原因与政治审查有关;该聊天机器人曾称俄罗斯入侵乌克兰是一场“侵略战争”。作为持续数年的监管风暴的一部分,中国要求人工智能公司注册其算法以供审查,其他政策则针对人工智能虚假信息、冒充人类身份的欺骗行为以及对未成年人的操纵。我问了几位专家,如果像Hugging Face事件那样,中国智能体被发现入侵另一家中国公司的数据系统,会发生什么,所有人都预测政府会采取果断行动。“那会是一件大事,”斯坦福大学专注中国与技术问题的研究学者格雷厄姆·韦伯斯特(Graham Webster)对我说。但“不会有人说‘我们都要完蛋了’,”他补充道。中国“会直接去解决实际问题,同时公开羞辱那家公司。”

Partly, this might be because Chinese engineers grew up in a culture that prized problem-solving over wayward speculation. And their most prominent philosophers, the Confucians, preached a morality that radiated outward: order yourself before ordering others. A complaint I heard time and again from experts is that “But China” arguments tend to treat the world as a theoretical construct. This is a hazard of Western analytic philosophy, where sensible premises, like recursive self-improvement, slip into sci-fi absurdity—killer-robot armies, ethnically targeted bioweapons, genetically modified super-soldiers. Now we are made to choose between a world that might be and the world as it is. If we want to win the A.I. “race,” then perhaps it’s time to get our house in order first. ♦

部分原因可能是中国工程师成长于一种重视解决问题而非天马行空推测的文化中。而他们最著名的哲学家——儒家——宣扬的是一种由内向外辐射的道德:先正己,再正人。我从专家那里反复听到的一个抱怨是,“可是中国”的论调往往把世界当作一个理论构造。这是西方分析哲学的一个危险:合理的假设——比如递归式自我改进——滑向科幻的荒诞——杀手机器人军团、针对特定族群的生物武器、基因改造的超级士兵。现在我们被迫在一个可能的世界和现实的世界之间做出选择。如果我们想赢得人工智能“竞赛”,那么也许是时候先把自己的事情理顺了。♦

2 free previews left — register to get 30 translations a month