SHEERA FRENKEL, 孟建国, ADAM SATARIANO2026年4月13日Anduril的自主飞行器Fury近期已在该公司位于俄亥俄州哥伦布市郊外的新厂投产。 Kristian Thacker for The New York TimesAt a military parade in Beijing in September, President Xi Jinping and his special guests, President Vladimir V. Putin of Russia and the North Korean leader, Kim Jong-un, watched as Chinese forces showed off several models of drones that could autonomously fly alongside fighter jets into battle.去年9月在北京举行的一场阅兵式上，中国国家主席习近平与他的贵宾——俄罗斯总统弗拉基米尔·普京和朝鲜领导人金正恩——一同观看了中国军队展示的多款无人机，这些无人机能够自主伴飞战斗机投入作战。The demonstration of technological might immediately set off alarm bells in the United States. Pentagon officials concluded that America’s program for unmanned combat drones was lagging China’s, according to three U.S. defense and intelligence officials. Russia, too, was thought to be ahead in building facilities that could produce advanced drones, said the officials, who were not authorized to speak publicly on military capabilities.这番技术实力展示立即引起了美国的警觉。据三位美国国防和情报官员透露，五角大楼得出结论，美国的作战无人机计划已落后于中国。此外，这些官员表示，俄罗斯在建造生产先进无人机的设施方面也被认为领先于美国。由于未获授权公开谈论军事能力，这些官员要求匿名。U.S. officials pushed domestic defense companies to step up. Last month, Anduril, a defense technology start-up in California, began manufacturing A.I.-backed, self-flying drones that appeared similar to the ones shown in China. Production at a factory outside Columbus, Ohio, started three months ahead of schedule, part of an effort to close the gap with China, one defense official said.美国官员随即敦促国内的国防企业加快步伐。上个月，加州国防科技初创企业Anduril开始制造配备人工智能、能够自主飞行的无人机，其外观与在中国展示的型号相似。一名国防官员表示，位于俄亥俄州哥伦布市外的工厂提前三个月启动生产，这是为了缩小与中国的差距所做的努力之一。China’s military display and the U.S. countermove were part of an escalating global arms race over A.I.-backed autonomous weapons and defense systems. Designed to operate by themselves using A.I., the technology reduces the need for human intervention in decisions like when to hit a moving target or defend against an attack.中国的军事展示以及美国的应对举措是围绕人工智能驱动的自主武器和防御系统不断升级的全球军备竞赛的一部分。这类技术旨在通过人工智能实现自主运行作，从而减少人类在“何时打击移动目标”或“如何防御攻击”等决策中的干预。In recent years, many nations have quietly engaged in a contest of one-upmanship over these arsenals, including drones that identify and strike targets without human command, self-flying fighter jets that coordinate attacks at speeds and altitudes that few human pilots can reach, and central systems run by A.I. that analyze intelligence to recommend airstrike targets quickly.近年来，许多国家在这一领域展开了悄无声息的较量，包括能够在没有人工指令下识别并打击目标的无人机、能够在人类飞行员几乎难以企及的速度和高度协同攻击的自主战斗机，以及由人工智能驱动的中央系统，用于分析情报并快速推荐空袭目标。The United States and China, the world’s largest military powers, are at the center of the competition. But the race has widened. Russia and Ukraine, now in their fifth year of war, are looking for every technological advantage. India, Israel, Iran and others are investing in military A.I., while France, Germany, Britain and Poland are rearming amid doubts about the Trump administration’s commitment to NATO.作为世界上最大的两个军事强国，美中处于这场竞争的中心。但这场竞赛的范围正在扩大。已进入战争第五年的俄罗斯和乌克兰正在寻求一切技术优势。印度、以色列、伊朗等国也在投资军事人工智能，法国、德国、英国和波兰也在重新武装自己——因对特朗普政府在北约承诺上的疑虑。Each nation is aiming to amass the most advanced technological stockpile in case they need to fight drone against drone and algorithm against algorithm in ways that people cannot match, defense and intelligence officials said.国防和情报官员表示，每个国家都致力于积累最先进的技术储备，以备在需要时能够以人类无法企及的方式进行无人机对无人机、算法对算法的作战。Russia, China and the United States are all building A.I. arms as a deterrent and for “mutually assured destruction,” Palmer Luckey, Anduril’s founder, said in an interview in February.Anduril创始人帕尔默·勒基在2月接受采访时表示，俄罗斯、中国和美国都在制造人工智能武器，既作为威慑手段，也为了实现“相互确保摧毁”。The buildup has been compared to the dawn of the nuclear age in the 1940s, when the atomic bomb’s destructive power forced rival nations into an uneasy standoff, leading to more than four decades of nuclear weapons brinkmanship.这种军备建设被拿来与20世纪40年代核时代的开端做比较。当时，原子弹的巨大破坏力迫使敌对国家进入一种不稳定的对峙状态，并导致了持续四十多年的核武器边缘政策。But while the implications of nuclear weapons are well understood, A.I.’s military capabilities are just beginning to be known. The technology — which does not need to pause, eat, drink or sleep — is set to upend warfare by making battles faster and more unpredictable, officials said.核武器的影响已为人们所熟知，而人工智能的军事能力才刚刚开始显现。官员们表示，这项技术无需暂停、无需进食、无需饮水、无需睡眠，它将以更快的速度和更不可预测的方式颠覆战争形态。Exactly which nation is furthest ahead is unclear. Many programs are in a research and development phase, and budgets are classified. Operatives from China, the United States and Russia watch one another’s factory lines, military displays and weapons deals to deduce what the other is doing, intelligence officials said.目前尚不清楚哪个国家处于领先位置。许多项目仍处于研发阶段，相关预算也属于机密。情报官员表示，中国、美国和俄罗斯的谍报人员都在密切留意彼此的工厂生产线、军事展示和武器交易，以推断对方的动向。China and Russia are experimenting with letting A.I. make battlefield decisions on its own, two U.S. officials said. China is developing systems for dozens of autonomous drones to coordinate attacks without human input, while Russia is building Lancet drones that can circle in the sky and autonomously pick targets, they said.两名美国官员表示，中国和俄罗斯正在试验让人工智能自主做出战场决策。中国正在开发能够让数十架自主无人机在没有人工干预的情况下协同发动攻击的系统，俄罗斯则在建造“柳叶刀”无人机，这种无人机可以在空中盘旋并自主选择目标。Even as the specifics of the technologies remain veiled, the intentions are clear. In 2017, Mr. Putin declared that whoever leads in A.I. “will become the ruler of the world.” Mr. Xi said in 2024 that technology would be the “main battleground” of geopolitical competition. In January, Defense Secretary Pete Hegseth directed all branches of the U.S. military to adopt A.I., saying they needed to “accelerate like hell.”即使这些技术的具体细节仍处于保密状态，但其意图已十分明确。普京在2017年曾宣称，谁在人工智能领域领先，“谁就将成为世界的统治者”。习近平在2024年表示，科技创新将成为地缘政治竞争的“主战场”。今年1月，国防部长皮特·海格塞斯要求美国各军种采用人工智能，称他们需要“拼命加速”。上个月，马里兰州安德鲁斯联合基地，国防部长海格塞斯和特朗普总统。海格塞斯已要求军队各军种采用人工智能。Billions of dollars are being poured into the efforts. The Pentagon requested more than $13 billion for autonomous systems in its latest budget, and has spent billions more over the past decade, though the total is difficult to track because A.I. funding has been spread across many programs.数十亿美元正被投入到这些努力中。五角大楼在其最新预算中为自主系统申请了超过130亿美元，且在过去十年中已耗资数百亿，不过由于人工智能的资金分散在许多项目中，总金额难以统计。China, which some researchers said was spending amounts comparable to those of the United States, has used financial incentives to spur private industry to build A.I. capabilities. Russia has invested in drone and autonomy-related programs, analysts said, using the war in Ukraine to test and refine them on the battlefield.一些研究人员表示，中国的投入规模与美国相当，并通过财政激励措施推动私营企业发展人工智能能力。分析人士称，俄罗斯也在无人机和自主系统方面进行投资，并利用俄乌战争的战场测试和完善这些技术。Liu Pengyu, a spokesman for the Chinese Embassy in Washington, said China had proposed international frameworks for governing military A.I. and called for “a prudent and responsible attitude” toward its development.中国驻华盛顿大使馆发言人刘鹏宇表示，中国已提出规范军事人工智能的国际框架，并呼吁对其发展采取“审慎和负责任的态度”。The Pentagon and Russia’s Ministry of Defense did not respond to requests for comment.五角大楼和俄罗斯国防部均未回应置评请求。The dynamics may resemble the Cold War, but experts cautioned that the A.I. era was different. Start-ups and investors now play a role in the military and are as critical as universities and governments. A.I. technology is becoming widely available, opening the door for countries from Turkey to Pakistan to develop new capabilities. What’s emerging is a grinding innovation race without any obvious endpoint.这种态势可能类似于冷战，但专家警告称，人工智能时代有所不同。初创公司和投资者如今在军事领域发挥着作用，其重要性不亚于大学和政府。人工智能技术正变得广泛应用，为从土耳其到巴基斯坦等国家开发新能力打开了大门。一场没有明确终点、不断推进的创新竞赛正在形成。Ethical questions about ceding life-or-death choices to machines are being overtaken by the rush to build. The only major accord on A.I. weaponry between China and the United States was reached in 2024, a nonbinding pledge to maintain human control over the decision to use nuclear weapons. Other countries, like Russia, have made no commitments.将生死抉择交给机器所引发的伦理问题正在被各国竞相研发的浪潮所淹没。中美之间唯一关于人工智能武器的重要协议是在2024年达成的，这是一项不具约束力的承诺，旨在保持人类对核武器使用决策的控制。其他国家，如俄罗斯，则未做出任何承诺。Some argued that A.I.’s impact would be bigger than any arms race.有人认为，人工智能的影响将超过任何一场军备竞赛。“A.I. is a general-purpose technology like electricity. And we don’t talk about an electricity arms race,” said Michael Horowitz, a former Pentagon official involved in autonomous weapons development. “To the extent A.I. is transforming our military, it’s the way that electricity or computers or the airplane did.”“人工智能是一种通用技术，就像电力一样。我们不会说电力军备竞赛，”曾参与自主武器开发的五角大楼前官员迈克尔·霍洛维茨说。“如果说人工智能正在改变我们的军队，其方式与电力、计算机或飞机当年如出一辙。”去年9月，在北京天安门广场举行的中国军队阅兵式上，一架作战无人飞行器。The Buildup Begins军备建设拉开序幕In 2016 at an air show in the southern Chinese city of Zhuhai, a Chinese supplier flew 67 drones in unison. An animated film separately showed the drones destroying a missile launcher, a demonstration of their capabilities.2016年，在中国南方城市珠海的一个航展上，一家中国供应商让67架无人机编队飞行。为了展现它们的能力，还用一段动画呈现出这些无人机摧毁导弹发射器的场景。Russia, too, was building its drone arsenal. In 2014, its military planners set a goal of making 30 percent of its combat power autonomous by 2025. By 2018, the Russian military was testing an unmanned armed vehicle in Syria. While the tank failed, losing its signal and missing targets, it underscored Moscow’s ambitions.俄罗斯也在建设无人机武器库。2014年，该国的军事规划部门设定了到2025年实现30%作战力量自主化的目标。到2018年，俄罗斯军方在叙利亚测试了一款无人战车。虽然该战车失败了——信号丢失、未能击中目标，但它凸显了莫斯科的雄心。In Washington, Lt. Gen. Jack Shanahan, who had previously worked in intelligence at the Defense Department, was assessing whether A.I. could solve a more immediate problem. The U.S. military was collecting so much data — drone footage, satellite imagery, intercepted signals — that nobody could make sense of it all.在华盛顿，此前曾在国防部从事情报工作的中将杰克·沙纳汉正在评估人工智能能否解决一个更紧迫的问题。美军收集的数据太多了——无人机拍摄的画面、卫星图像、截获的信号，以至于没人能全部理清。“There was nothing in any of the research labs in the military that were capable of generating results in less than a couple of years,” General Shanahan said. “We had a problem we could not solve without A.I.”“军方任何研究实验室都无法在短短几年内产生成果，”沙纳汉将军说。“我们有一个不靠人工智能就无法解决的问题。”In 2017, General Shanahan helped create Project Maven, a Defense Department effort for the military to incorporate A.I. into its systems. One aim was to work with Silicon Valley to build software to swiftly process images like drone footage for intelligence purposes. Google was tapped to help.2017年，沙纳汉将军协助创建了马文项目(Project Maven)，这是国防部为将人工智能纳入军事系统而做出的努力。其目标之一是与硅谷合作，构建能够快速处理图像（如无人机拍摄的画面）用于情报目的的软件。谷歌曾被选中提供帮助。But the project quickly ran into hurdles. The Pentagon’s procurement system, built around legacy contractors and long timelines, slowed things down.但该项目很快遇到障碍。五角大楼的采购体系建立在传统承包商和漫长的交付周期之上，拖慢了进度。When word spread inside Google about Project Maven, employees also protested, saying a company that had once pledged “Don’t be evil” should not help identify targets for drone strikes. Google eventually backed away from the project.当马文项目的消息在谷歌内部传开后，员工们也提出了抗议，称一家曾承诺“不作恶”的公司不应帮助识别无人机打击的目标。谷歌最终退出了该项目。In 2019, Palantir, a data analytics company co-founded by the tech investor Peter Thiel, took over Maven. New defense tech start-ups like Anduril also emerged, supplying the federal government with A.I.-backed sensor towers along the southern U.S. border.2019年，由科技投资者彼得·蒂尔共同创立的数据分析公司Palantir接管了马文项目。Anduril等新的国防科技初创公司也出现了，它们在美国南部边境沿线向联邦政府提供由人工智能支持的传感器塔。帕尔默·勒基是Anduril公司的创始人，该公司开发了基于人工智能的传感器塔，用于部署在边境地区。 Philip Cheung for The New York TimesIn China, Beijing pushed commercial tech companies toward defense partnerships in a strategy called “civil-military fusion.” Private firms were drawn into military procurement, joint research and other work with defense institutions. Companies working on drones and unmanned boats found growing military demand for their technologies.在中国，北京通过军民融合战略推动商业科技公司与国防部门合作。私营企业被纳入军事采购、联合研究以及与国防机构的其他合作项目中。从事无人机和无人艇的公司发现，其技术在军事领域的需求日益增长。Russia’s invasion of Ukraine in 2022 turned theory into reality.2022年俄罗斯入侵乌克兰，将理论变成了现实。Outgunned, outspent and outnumbered, Ukraine held off Russia with an improvised arsenal of cheap technology. Hobbyist racing drones were used to attack Russian positions on the front lines, eventually becoming more lethal than artillery and, in some cases, gaining autonomous capabilities. Remote-controlled boats kept Russia’s Black Sea fleet pinned down.在武器、经费和人数均处于劣势的情况下，乌克兰依靠廉价技术拼凑而成的武器库挡住了俄罗斯的进攻。业余爱好者使用的竞速无人机被用来攻击前线的俄罗斯阵地，其杀伤力最终超过了火炮，在某些情况下甚至具备自主作战能力。遥控艇则让俄罗斯黑海舰队陷入困境。Russia adapted as well. Its Lancet drone, which was initially piloted by humans, has incorporated autonomous targeting features.俄罗斯也进行了调整。其最初由人工操控的“柳叶刀”无人机已集成自主目标锁定功能。俄罗斯士兵在乌克兰某未公开地点搬运一架“柳叶刀”无人机。这张照片截取自俄罗斯国防部新闻处8月发布的视频。“The four years of brutality on the battlefield in Ukraine has served as a laboratory for the world,” said Mr. Horowitz, the former Pentagon official.“乌克兰战场四年的残酷厮杀已成为全世界的实验场，”前五角大楼官员霍洛维茨说。In recent months, Ukraine began sharing its troves of battlefield data with Palantir and other firms so A.I. systems can better learn to fight wars.最近几个月，乌克兰开始将其海量战场数据与Palantir等公司共享，以便人工智能系统更好地学习如何作战。Across Europe, where governments are aiming to diminish their reliance on the American military, the lessons from Ukraine resounded. In February, Germany, France, Italy, Britain and Poland said they would develop a joint air defense system to guard against drones.在欧洲，各国政府正努力减少对美国军队的依赖，乌克兰的教训产生了强烈共鸣。今年2月，德国、法国、意大利、英国和波兰表示，他们将共同开发一套防空系统，以防御无人机威胁。China also advanced. At the 2024 Zhuhai Airshow, Norinco, one of the country’s main defense manufacturers, revealed multiple weapons with A.I. capabilities. One of its systems showed an entire brigade, including armored vehicles and drones, which were controlled and operated by A.I.中国也在推进。2024年珠海航展上，中国主要国防制造商之一的北方工业展示了多款具备人工智能能力的武器。其中一个系统展示了一整个由人工智能控制和操作的旅级单位，包括装甲车辆和无人机。Another craft, unveiled by the state-run Aviation Industry Corporation of China, was a 16-ton jet-powered drone designed to serve as a flying aircraft carrier that could deploy dozens of smaller drones midflight.另一款由中国航空工业集团推出的飞行器是一架16吨重的喷气动力无人机，设计用于充当飞行航母，可在飞行中部署数十架小型无人机。2024年珠海航展上亮相的一款能够发射小型无人机群的无人机母舰。‘Left Click, Right Click’“点击左键、点击右键”A week after American and Israeli forces struck Iran in February, a senior Pentagon official gave a glimpse into what computerized warfare now looks like at a conference livestreamed by Palantir.今年2月美以联军对伊朗发动打击一周后，一位五角大楼高级官员在Palantir直播的一场会议上展示了当今计算机化战争的真实面貌。A satellite feed showed a warehouse. With the click of a mouse, an officer selected a row of white trucks parked outside to target in real time. In seconds, the A.I. software suggested a weapon, calculated fuel and ammunition needs, weighed the cost and generated a strike plan.卫星画面显示了一个仓库。一名军官点击鼠标，实时选中了停在外面的一排白色卡车作为打击目标。几秒钟内，人工智能软件就推荐了武器、计算了燃料和弹药需求、权衡了行动成本，并生成了打击计划。It was the present-day version of Project Maven, which General Shanahan had started and was now run by Palantir and powered by commercial A.I. The system analyzed intelligence from various sources, generated target lists ranked by priority and recommended weapons, all but eliminating the lag between identifying a target and destroying it.这就是沙纳汉将军当年启动、如今由Palantir运营并由商用人工智能驱动的马文项目的现代版本。该系统分析来自不同来源的情报，生成按优先级排序的目标清单并推荐武器，几乎消除了从发现目标到摧毁目标之间的延迟。Embedded with a military version of Claude, the chatbot made by the A.I. firm Anthropic, Maven helped generate thousands of targets in the opening weeks of the Iran campaign, a pace that Adm. Brad Cooper, the head of U.S. Central Command, attributed in part to “advanced A.I. tools.”马文项目嵌入了人工智能公司Anthropic开发的军用版聊天机器人Claude，在伊朗行动的最初几周，该系统帮助生成了数以千计的目标，美国中央司令部司令布拉德·库珀上将将这一速度部分归功于“先进的人工智能工具”。Cameron Stanley, the Defense Department’s chief digital and artificial intelligence officer, who spoke at Palantir’s conference, said that what Maven was doing was “revolutionary.” Human involvement amounted to “left click, right click, left click,” he said.出席Palantir会议并演讲的国防部首席数字与人工智能官卡梅伦·斯坦利表示，马文项目正在做的工作是“革命性的”。他说，人类的参与仅限于“点击左键、点击右键、点击左键”。The claims about Maven’s abilities might be overstated and much of the American advantage came from the scale of data flowing in and the skills of the people using it, said Emelia Probasco, a senior fellow at Georgetown University’s Center for Security and Emerging Technology.乔治城大学安全与新兴技术中心高级研究员埃梅利亚·普罗巴斯科表示，关于马文项目能力的说法可能被夸大了，美国的优势很大程度上来自数据流的规模，以及使用这些工具的人员技能。“It’s not rocket science,” she said. “I suspect that China already has something like it.”“这不是火箭科学，”她说，“我怀疑中国已经有了类似的东西。”In a recent report analyzing thousands of People’s Liberation Army procurement documents, Ms. Probasco found that China was building systems that mirrored American ones. In one case, China was trying to replicate the Joint Fires Network, an American program set up to link sensors and weapons globally so a drone on one side of the world could cue a strike from the other.在最近一份分析数以千计中国人民解放军采购文件的报告中，普罗巴斯科发现，中国正在建设与美国类似的系统。其中一个例子显示，中国正试图复制“联合火力网络”，这是一个旨在全球连接传感器和武器的美国项目，让一架无人机可以在世界一侧为另一侧的打击提供指引。In some areas, China clearly leads. Its manufacturing dominance means it can produce autonomous weapons at a scale the Pentagon cannot match.在某些领域，中国显然处于领先地位。其制造业优势意味着它能够以五角大楼无法匹敌的规模生产自主武器。Inside the Trump administration, the push for A.I. weapons has taken on an almost evangelical fervor. Last month, the Pentagon labeled Anthropic a security risk, partly because the company wanted to limit its technology’s use for automated weapons.在特朗普政府内部，推动人工智能武器的热情已近乎狂热。上个月，五角大楼将Anthropic列为安全风险，部分原因在于该公司希望限制其技术在自动武器中的应用。“We will win the A.I. race,” Jacob Helberg, the under secretary of state for economic affairs, said last month at the Hill & Valley Forum, an annual conference in Washington, which he co-founded to bridge Silicon Valley and the government.“我们将赢得人工智能竞赛，”负责经济事务的副国务卿雅各布·赫尔伯格上个月在由他共同创立的“国会山与硅谷论坛”上表示，该年度会议旨在连接硅谷与政府。2024年“国会山与硅谷论坛”上，美国国务院负责经济事务的副国务卿雅各布·赫尔伯格与Palantir首席执行官亚历克斯·卡普。At the conference, tech executives, investors and government officials cheered speakers who called for tech companies to give the military unfettered access to A.I.在会议上，科技高管、投资者和政府官员对那些呼吁科技公司向军方无限制开放人工智能的演讲者报以热烈掌声。Anduril’s Mr. Luckey argued that the A.I. arms buildup might prevent major wars. The logic mirrored the Cold War: If both sides knew what the machines could do, neither would risk finding out.Anduril的勒基认为，人工智能军备建设可能防止大规模战争。其逻辑与冷战类似：如果双方都清楚这种机器的能力，就没有人愿意冒险去验证。“Conflicts between superpowers will similarly deteriorate if you can build the things that deter warfare effectively enough,” he said.“如果能制造出足以有效遏制战争的武器，超级大国之间的冲突也会随之缓和，道理是类似的，”他说。Yet deterrence assumes rationality, while A.I. weapons are designed to move faster than human reason. In exercises dating to 2020, researchers explored how autonomous systems could accelerate escalation and erode human control — with some alarming results.然而，威慑的前提是理性，而人工智能武器的设计初衷却是超越人类理性的反应速度。在2020年以来的演习中，研究人员探讨了自主系统如何加速升级并削弱人类控制——结果令人担忧。In one scenario, a system operated by the United States and Japan responded to a missile launch from North Korea by autonomously firing an unexpected counterattack.在一个场景中，由美国和日本操作的系统在应对朝鲜导弹发射时，自主发动了出人意料的反击。“The speed of autonomous systems led to inadvertent escalation,” said the report by analysts at RAND Corporation, a nonprofit research organization that works with the military.兰德公司分析师的报告称，“自主系统的速度导致了无意的升级。”General Shanahan, who retired from the military in 2020 and is now a fellow at the Center for a New American Security, a think tank, said the race he had helped start kept him up at night. Governments must set clear boundaries before the technology outruns their control, he said.2020年从军方退役、现为智库新美国安全中心研究员的沙纳汉将军表示，自己曾参与推动的这场竞赛让他夜不能寐。他说，各国政府必须在技术发展超出控制之前设定明确的界限。退役中将、前联合人工智能中心主任杰克·沙纳汉表示，人工智能将“进一步强调越来越快地行动”。“There is a risk of an escalatory spiral where we’re in danger of fielding untested, unsafe and unproven systems if we’re not careful, because we each feel like the other side is hiding something from us,” he said.“存在升级螺旋的风险——如果我们不小心，我们就有可能部署未经测试、不安全、未经证实的系统，因为双方都觉得对方在隐瞒什么，”他说。Sheera Frenkel常驻旧金山湾区，主要报道科技如何影响日常生活，尤其关注社交媒体公司，包括 Facebook、Instagram、Twitter、TikTok、YouTube、Telegram 和 WhatsApp。孟建国(Paul Mozur)是时报全球科技记者，常驻台北。此前，他自香港、上海和首尔报道亚洲科技与政治的交叉议题。Adam Satariano是时报科技记者，常驻伦敦。翻译：纽约时报中文网点击查看本文英文版。