SEBASTIAN MALLABY2026年4月15日 Igor BastidasIn 2022, the Biden administration tried to arrest China’s development of artificial intelligence by denying it cutting-edge semiconductors. President Trump has relaxed that policy a bit without a clear plan to replace it.2022年，拜登政府试图通过切断先进半导体供应来遏制中国的人工智能发展。特朗普总统则在没有明确替代方案的情况下，稍微放松了这一政策。But the chip export controls have failed. China’s tech sector is too sophisticated to be stopped from building powerful A.I. In pursuing an impossible objective, the United States is missing an opportunity to try for one that sounds fanciful but which, after a recent reporting trip to China, I believe is more realistic: America should negotiate with China a global pact on A.I. safety, which would impose universal limits on a technology that can do much good — but, in the wrong hands, would do much harm.但芯片出口管制已经失败了。中国的科技行业自成体系，无法阻止其构建强大的人工智能系统。在追求一个不可能实现的目标时，美国错失了尝试争取的机会：虽然听起来有些理想化，但在最近一次中国采访之行后，我认为它更为现实：美国应与中国就人工智能安全问题达成一项全球性协议，对这项能够带来巨大好处——但若落入坏人之手将造成巨大危害的技术施加普遍限制。The premise of the export restrictions was that the United States would be able to successfully block China’s access to powerful A.I. chips. The premium chip sets used in A.I. data centers are the size of skateboards and can’t be smuggled in a simple suitcase, and it’s hard to put them to use without hands-on support from the chipmakers’ engineering teams. But Chinese developers circumvented controls by training their A.I. models on chips located in other countries. A Chinese model builder needs only to rent capacity on an A.I. data center in one of China’s Southeast Asian neighbors. Concealing the model’s Chinese origin is straightforward.出口限制的前提是，美国能够成功阻止中国获得强大的人工智能芯片。用于人工智能数据中心的高端芯片组件体积如滑板般大小，不可能用手提箱轻易走私，而且在没有芯片制造商工程团队现场支持的情况下，也很难投入使用。但中国的开发人员通过位于其他国家的芯片上训练人工智能模型来规避管制。中国的模型开发者只需在某个东南亚邻国的人工智能数据中心租用算力即可。而且掩盖模型的中国来源并不难。Partly thanks to this loophole, China has rolled out a series of excellent A.I. models. China’s ability to skirt U.S. controls will not change, even if the Senate follows the House in passing a bill to restrict China’s access to outside data centers. China is learning how to do without cutting-edge chips by stacking less powerful chips together. Its model builders also take full advantage of a process known as distillation. Every time a U.S. lab produces a cutting-edge model, Chinese rivals quickly reverse-engineer its capabilities and build a copycat version. The follower has the advantage.部分得益于这一漏洞，中国已经推出了一系列出色的人工智能模型。即便参议院跟进众议院，通过一项限制中国使用境外数据中心的法案，中国规避美国管控的能力也不会改变。中国正在学会通过将较弱的芯片叠加使用来替代先进芯片，其模型开发者还充分利用一种被称为“蒸馏”的过程。每当美国实验室推出先进的模型，中国竞争者就会迅速对其能力进行逆向工程，并构建出模仿版本。后来者反而具备优势。American A.I. scientists used to say that competitors’ being able to follow fast would not matter. An “intelligence explosion” was approaching, the argument went. A.I. systems would soon become capable enough to write upgrades into their own code: A.I. would create better A.I.; better A.I. would create even better A.I.; recursive self-improvement would drive performance skyward. The nation that reached this so-called singularity first would be the winner of the A.I. race, even if the fast follower were just a few months behind the leader. Three and a half years after the Biden administration’s chip controls, A.I. is generating code to upgrade itself. The promised feedback loop has started.美国的人工智能科学家过去常说，竞争对手能够快速跟进并不重要。他们认为，一场“智能爆炸”正在逼近。人工智能系统将很快变得足够强大，能够自己写代码来升级自己：人工智能将创造更好的人工智能；而更好的人工智能又将创造更更好的人工智能；递归式的自我改进将使性能迅速飙升。率先达到所谓“奇点”的国家，即便领先优势只有几个月，也将赢得人工智能竞赛。在拜登政府实施芯片管制三年半后，人工智能已经开始生成用于自我升级的代码。人们所预言的那个“反馈循环”已经启动了。But the accelerating power of the leading models won’t determine who wins the A.I. race. It’s A.I. deployment that will matter. To transform economies and armies, A.I. must be embedded in business processes and weapons systems. The raw power of the cutting-edge models must be turned into applications.但领先模型能力的加速提升并不会决定谁赢得人工智能竞赛。真正重要的是人工智能的落地应用。要改变经济和军队，人工智能必须嵌入到业务流程和武器系统中。最先进模型的原始算力必须转化为实际应用。The upshot is that China and the United States are roughly level in the A.I. contest. Top Chinese models may be a few months behind American ones, and the relative position on military applications is difficult to ascertain as so much is classified. But on industrial applications, China seems to be leading. U.S.-sanctioned companies such as Huawei and Hikvision are rolling out A.I. systems that perform maintenance checks on high-speed trains, manage mining operations and scan water samples to assess pollution. At Huawei’s campus near Shenzhen, I recently took a ride in an autonomous car. A device in the passenger seat massaged my back, and the steering was immaculate.结论是，中国和美国在人工智能竞争中大致处于同一水平。中国顶尖模型可能落后美国几个月，而在军事应用方面，由于大量信息属于机密，双方的相对位置难以判断。但在工业应用方面，中国似乎领先。像华为和海康威视这样受到美国制裁的公司，正在推出用于高铁维护检测、矿山运营管理以及水样污染检测的人工智能系统。在深圳附近的华为园区，我最近还体验了一次自动驾驶汽车：副驾驶座上的设备会为我按摩背部，转向操作无可挑剔。Fans of chip controls continue to insist that even a modest slowing of China’s A.I. advance is worth pursuing. If China is a formidable adversary, imagine how much more formidable it might be if the chip controls were lifted. But the controls are failing to deliver the prize of a China with limited A.I., so it is worth considering their cost. My China trip persuaded me that the cost is too high.芯片管制的支持者仍坚持认为，即便只是稍微减缓中国人工智能的发展，也值得去做。如果中国已经是一个强大的对手，那么可以想象，一旦取消芯片管制，它会变得多么可怕。但这些管制并没有实现限制中国人工智能发展的目标，因此有必要重新评估其代价。这次中国之行让我相信，这一代价过于高昂。The Biden administration made a strategic choice to prioritize the slowing of China rather than addressing other worries. The alternative would have been to say to China: You are a tech superpower. We are a tech superpower. Let’s work together to make sure A.I. doesn’t fall into the hands of rogue states and terrorists. The goal would have been an A.I. equivalent of the 1968 Nuclear Nonproliferation Treaty, a regime that would require all countries using A.I. to sign up for safeguards on it.拜登政府做出了一个战略选择：其优先考量是减缓中国的发展，而非着手应对其他方面的担忧。而当时本可以采取的另一种策略是对中国说：你是一个科技超级大国，我们也是一个科技超级大国。让我们合作，确保人工智能不会落入流氓国家和恐怖分子之手。目标应当是达成一项类似于1968年《核不扩散条约》的人工智能协议，建立起一套制度，要求所有使用人工智能的国家都必须遵守相关安全保障。The Biden team did not think China would collaborate on something like that. But over a dozen conversations with A.I. leaders in Beijing, Shanghai, Shenzhen and Hangzhou made it clear to me that China’s elite does care about A.I. safety.拜登团队认为，中国不会在这方面合作。但我在北京、上海、深圳和杭州与人工智能领袖进行的十多次交谈清楚地表明，中国的精英阶层确实关心人工智能安全问题。I visited a prominent tech company that builds and distributes an A.I. foundation model. For now, that model is open source, meaning that users can download and modify it at will. If a user prompts the A.I. to conduct cyberattacks, there’s nothing anyone can do to stop that person. But the chief executive of this tech company made a striking admission: As A.I. becomes more powerful, it would be crazy to continue making it open source, he said. You wouldn’t open-source a nuclear weapon, he added.我访问了一家开发并发布人工智能基础模型的知名科技公司。目前，该模型是开源的，这意味着用户可以自由下载和修改。如果有人要求该人工智能发起网络攻击，几乎没有办法阻止。但这家公司的首席执行官坦率承认：随着人工智能变得越来越强大，继续保持开源将是疯狂的。他补充说，你不会把核武器开源。During my trip, the controversy surrounding the advanced model OpenClaw illustrated the rising concern for A.I. safety. Throngs of ordinary Chinese downloaded the digital assistant, eager to experiment with a capable A.I. agent. The enthusiasm apparently confirmed that China loves innovation more than it fears it. But researchers and industry leaders told me that they were appalled. OpenClaw makes your computer “naked,” an eminent business school professor told me. Soon after he said that, China’s leaders firmly discouraged the use of OpenClaw on government systems and warned citizens that the agent might wreak havoc with their data.在我此次访问期间，先进模型OpenClaw引发的争议凸显了人们对人工智能安全日益增长的担忧。大量普通中国用户下载了这一数字助手，渴望尝试这一强大的智能体。这种热情似乎印证了中国对创新的热爱超过了对风险的担忧。但研究人员和行业领袖告诉我，他们对此感到震惊。一位知名商学院教授表示，OpenClaw会让你的电脑“裸奔”。他发表这一看法后不久，中国的领导人便明确反对在政府系统中使用OpenClaw，并警告公众该智能体可能对他们数据造成巨大破坏。For now, China’s instinct to race for powerful A.I. overwhelms any caution. This is a rational response to a U.S. administration that is equally determined to put speed ahead of safety. But if a U.S. leader went to China and offered to scrap chip controls in exchange for collaboration on A.I. nonproliferation, there would be at least some chance of the proposal succeeding.目前，中国竞逐强大人工智能的本能压倒了所有的谨慎。这是对同样决心将速度置于安全之上的美国政府的理性回应。但是，如果一位美国领导人前往中国，提出取消芯片控制以换取在人工智能防扩散方面的合作，这一提议至少存在一定成功的可能性。This presumes that U.S.-Chinese dialogue is even possible. But the West should not succumb to self-fulfilling fatalism. At times during the Cold War, the United States pursued its interests by switching from confrontation to détente: the Nuclear Nonproliferation Treaty came just six years after the Cuban missile crisis. Now is a good time to recall that history.这一设想的前提是美中对话仍然可能。但西方不应陷入消极的自我预言。在冷战期间，美国有时会通过从对抗转向缓和来推进自身利益：《核不扩散条约》正是在古巴导弹危机发生仅六年后达成的。现在正是回顾这段历史的好时机。Sebastian Mallaby是美国外交关系委员会的高级研究员，著有《The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence》。他还是美国外交关系委员会的播客节目“The Spillover”的联合主持人。翻译：纽约时报中文网点击查看本文英文版。