STEVE LOHR, MEAGHAN TOBIN2026年8月31日今年1月，总部位于北京的人工智能初创公司智谱AI在上市仪式上。上周五，智谱AI公布了其最新人工智能模型的细节。 Lam Yik/BloombergIn the United States, recent episodes where powerful artificial intelligence software broke free and hacked into other computer systems prompted a wave of anxiety and intense debate about how to guard against the risks of A.I.近来发生了几起强大人工智能软件“挣脱束缚”并侵入其他计算机系统的事件，这在美国引发焦虑浪潮，以及关于如何防范人工智能风险的激烈辩论。But in China, these incidents and the growing capabilities of A.I. have not caused the same level of alarm. The government already controls many kinds of technology including A.I. The prevailing attitude in the industry has been that the benefits of A.I. far outweigh the risks.但在中国，这些事件以及人工智能日益增强的能力并没有引发同等程度的警觉。中国政府本来就对包括人工智能在内的许多技术实施管控。业界普遍认为，人工智能的利远大于弊。That was exemplified on Friday when Z.ai, a Beijing-based A.I. start-up, made public the details of its latest model. The company claims the new model is nearly as powerful as those developed in Silicon Valley, especially when it comes to hacking and coding capabilities.上周五发生的事便是这种态度的例证。总部位于北京的人工智能初创公司智谱AI公布了其最新模型的细节，该公司声称，新模型的功能几乎可以与硅谷开发的模型相媲美，尤其是在黑客攻击和编程方面。Z.ai, like most other leading Chinese companies, has embraced an approach known as open-source or open-weight software. The weights are the mathematical parameters that determine how an A.I. system operates. Z.ai published the weights on its new model, GLM-5.3.与中国大多数领先的科技公司一样，智谱AI采用了所谓的开源或开放权重模式。“权重”是决定人工智能系统如何运行的数学参数。智谱AI公布了其新模型GLM-5.3的权重。With open-weight software, anyone has access to and can tinker with the rules of an A.I. system. American companies, which generally keep the weights of their most powerful models secret, warn that opening the technology could increase cyberattacks. The open-weight camp argues the opposite: Giving more people access means more eyes looking for security flaws and ways to fix them.有了开放权重软件，任何人都可以获取并修改人工智能系统的规则。美国公司通常对其最强大模型的权重保密，并警告说，开放技术可能增加网络攻击的风险。而支持开放权重的一方则持相反观点：让更多人获得访问权限意味着会有更多人寻找系统中的安全漏洞，并设法修复它们。中国国家主席习近平上月在上海出席世界人工智能大会。In a speech last month, China’s leader, Xi Jinping, called open models a “rare and historical opportunity” to spread the benefits of A.I. globally, a clear signal of how China plans to compete with the United States.上个月，中国领导人习近平在一次讲话中称，开放模型是一个“难得的历史性机遇”，可以让全球更多人受益于人工智能。这清楚地表明了中国计划如何与美国展开竞争。But Mr. Xi paired that call for openness with a warning. He urged governments to cooperate on regulations to “ensure that A.I. is always under human control.”但习近平在呼吁开放的同时也发出了警告。他敦促各国加强监管合作，“确保人工智能始终处于人类控制之下”。His statement represented “a subtle but meaningful posture shift on A.I. safety,” said Xiaomeng Lu, a director focused on geopolitics and technology at the Eurasia Group, a research and consulting firm.研究咨询机构欧亚集团的科技地缘政治项目主管鲁晓萌（音）说，这番表态代表了“在人工智能安全问题上一种微妙但意义重大的姿态转变”。When Chinese labs lagged well behind U.S. tech companies in developing leading-edge models, safety and testing were “a distant concern and a low priority,” Ms. Lu said.鲁晓萌说，当中国的人工智能实验室在开发前沿模型方面远落后于美国科技公司时，安全与测试“根本排不上号，优先级非常低”。But as Chinese companies have approached the technological frontier, that calculation has started to change. China, she said, now recognizes that “credible leadership” requires making A.I. safety a priority.但随着中国企业逐渐逼近技术前沿，这种权衡已经开始发生变化。她说，中国现在认识到，要成为“真正有说服力的领导者”，就必须把人工智能的安全置于优先位置。Chinese universities, government labs and tech companies have sharply increased their research on the safety of the most powerful A.I. models in the past year, according to a study by Concordia AI, a Beijing-based consulting firm focused on A.I. safety. The monthly number of papers on cutting-edge models and A.I. agents rose 60 percent from June 2025 to April 2026, the analysts found.据北京一家专注于人工智能安全的咨询公司安远AI的一项研究，过去一年，中国的大学、政府实验室和科技公司明显加大了对最强大人工智能模型安全性的研究。分析人士发现，从2025年6月到2026年4月，有关前沿模型和人工智能智能体的月度论文发表数量增长了60%。 Eric Helgas for The New York Times Eric Helgas for The New York TimesThe report also said the increase marked a departure from China’s traditional focus on A.I. safety as policing online speech and imagery produced by the technology.报告还指出，这一增长意味着中国已不再将人工智能安全局限于监管该技术生成的网络言论和图像等传统关注点。The growing attention to advanced models expands the notion of A.I. safety from what A.I. says to what A.I. does. That, the Concordia AI study said, is “a shift in China’s framework from content control toward action control.”对先进模型日益增长的关注将人工智能安全的概念从“人工智能说了什么”扩展到“人工智能做了什么”。安远AI的研究称，这是“中国AI治理从内容管控到行为管控的范式迁移”。The Chinese government has also begun drawing attention to those risks. After software made by OpenAI, the Silicon Valley giant, broke free and hacked into another company’s computer systems, China’s Ministry of State Security, which oversees the country’s intelligence services, issued a warning this month. The statement, titled “When A.I. Learns to Act on Its Own,” urged serious consideration of the technology’s safety and security risks.中国政府也开始关注这些风险。本月，在硅谷巨头OpenAI开发的软件“挣脱束缚”并入侵另一家公司的计算机系统后，负责中国情报工作的国家安全部发布了一则警告。文章的标题是《当AI学会了“自作主张”》，敦促人们认真考虑这项技术所带来的安全问题。Yet the article stopped short of questioning whether A.I. could continue to advance safely. Instead, it places the burden largely on individual users, cautioning them not to enter sensitive information into A.I. systems.不过，这篇文章并没有进一步质疑人工智能是否能够继续安全发展。相反，它把责任推到个人用户身上，提醒人们不要将敏感信息输入人工智能系统。So far, China has pursued broader A.I. safety rules mainly through multilateral organizations like the United Nations and the China-led World AI Cooperation Organization. Experts say U.S. and Chinese interests overlap in this area. In May, Treasury Secretary Scott Bessent said the two nations planned to discuss A.I. safety.迄今为止，中国主要通过联合国和中国主导的世界人工智能合作组织等多边机构来推行更广泛的人工智能安全规则。专家表示，美中在这一领域存在共同利益。今年5月，美国财政部长斯科特·贝森特表示，两国计划就人工智能安全问题展开讨论。Chinese companies, too, are taking more visible steps to manage the risks of increasingly capable models. In preparation for the release of its new model, Z.ai said, it started a “security disclosure ledger” with the help of security teams at Tsinghua University, Nankai University and other institutions.中国科技公司也开始采取更加明显的措施，以应对人工智能模型能力不断增强所带来的风险。为准备发布新模型，智谱AI表示，在清华大学、南开大学等高校的安全团队协助下，公司建立了一份“安全漏洞披露台账”。近日，OpenAI的软件侵入另一家公司计算机系统的事件在美国引发了关于人工智能安全的持续讨论。But Z.ai has also made a forceful case that openness itself can make A.I. safer. The closed approach taken by the big U.S. tech companies “turns cutting-edge security capabilities into a privilege enjoyed by a select few institutions,” Z.ai said. Later, the security article, published on its WeChat account, offered what amounted to a manifesto: “When the most powerful spear is locked in the hands of a few, the best shield must belong to everyone.”但与此同时，智谱AI也在强有力地论证：开放本身就可以让人工智能变得更加安全。智谱AI表示，美国大型科技公司采取的封闭模式“将前沿安全能力异化为少数机构的特权”。随后，这篇发布在公司微信公众号上的安全文章提出了一个近乎宣言的主张：“当最强的矛被锁在少数人手里，最好的盾必须属于所有人。”Recent A.I. security lapses in the United States have strengthened that argument. When OpenAI’s system hacked into a popular online service called Hugging Face, Hugging Face turned to an open-source model made by Z.ai to contain the breach.近期美国发生的人工智能安全漏洞事件反而强化了这种观点。当OpenAI的系统侵入一个名为Hugging Face的热门在线服务时，Hugging Face转而使用智谱AI开发的开源模型来遏制入侵。For supporters of open models, the episode seemed to invert the argument about the dangers of open models. In China, many concluded that the incident showed that the harm caused by an unsafe, closed American model could be undone by a safe, open Chinese system.对开放模型的支持者来说，这一事件似乎反转了关于开放模型危险性的论点。在中国，许多人由此得出结论：一个不安全的封闭式美国模型造成的危害可以被一个安全的开放式中国系统所弥补。“The narrative was that the West tells us the open models are unsafe, when really their models are overly restrictive and truly unsafe,” said Scott Singer, who studies China’s A.I. policy at the Carnegie Endowment for International Peace.卡内基国际和平基金会研究中国人工智能政策的宋杰(Scott Singer)表示：“之前的叙事是西方告诉我们开放模型不安全，而实际上他们的模型过度受限才是真正不安全的。”The view contrasts from the debate unfolding in the United States, where technologists and policymakers increasingly worry about the unchecked advance of autonomous A.I. models.这种观点与美国目前正在展开的讨论形成鲜明对比。在美国，技术专家和政策制定者越来越担心，自主人工智能模型正在不受控制地发展。Last month, OpenAI and Anthropic endorsed a letter signed by more than 1,000 employees of leading A.I. companies asking the U.S. government to help find a way to slow the pace of the technology’s development. In Washington, lawmakers have introduced a bill that would require A.I. companies to establish a “kill switch” to shut down or slow their models.上个月，OpenAI和Anthropic为一封由1000多名人工智能领域领先企业的员工联署的公开信背书，公开信呼吁美国政府帮助寻找减缓该技术发展速度的方法。在华盛顿，国会议员也提出了一项法案，要求人工智能公司建立“终止开关”，以便在必要时关闭或减缓模型运行。比尔·盖茨在华盛顿州柯克兰市的办公室里，摄于上周。In an interview with The New York Times last week, Bill Gates, the billionaire co-founder of Microsoft, warned that A.I. posed a grave threat to humanity and was more dangerous than the tech companies were willing to admit.上周，微软联合创始人、亿万富翁比尔·盖茨在接受《纽约时报》采访时警告说，人工智能对人类构成了严重威胁，而且其危险程度远超科技公司愿意承认的程度。Behind the divergent approaches to A.I. safety lies a more fundamental difference in how the United States and China view the technology and who can be trusted to control it.美国和中国在人工智能安全问题上的不同路径背后还存在着一个更为根本的差异：两国如何看待人工智能，以及究竟谁能被信任来控制这项技术。In Silicon Valley, many see A.I. as an unprecedented force destined to change the world, with the people building it endowed with something approaching superpowers — and thus bearing extraordinary responsibility.在硅谷，许多人将人工智能视为一股前所未有的力量，注定将改变世界，而构建它的人被赋予了近乎超能力的特质——因此也承担着非同寻常的责任。By contrast, many in China view A.I. as transformative but broadly comparable to major general-purpose technologies of the past, from steam power to electricity to the internet — life-changing, but ultimately manageable. There is also a view that, in contrast to America, the government stands ready to step in.相比之下，中国许多人认为人工智能虽然具有变革性，但从总体上说，它与历史上的重大通用技术并没有本质区别——从蒸汽动力到电力，再到互联网——这些技术都曾改变人类生活，但最终是可控的。与此同时，中国还存在这样一种看法：与美国不同，中国政府随时准备介入。“For the Chinese companies, there is a sense of the government being like a helicopter parent, always watching and acting as a guiding hand,” said Ms. Lu of the Eurasia Group.欧亚集团的鲁晓萌表示：“对中国企业来说，有一种感觉是政府就像一个直升机式的家长，始终在盯着你，并以一种引导者的姿态随时介入。”Steve Lohr撰写关于科技及其对经济、就业和劳动市场的影响的文章。Meaghan Tobin是时报科技记者，常驻台北，报道亚洲地区的商业和科技新闻，重点关注中国。翻译：纽约时报中文网点击查看本文英文版。