内容来源：https://aiweekly.co/issues/chinas-ai-is-redrawing-the-ai-race

内容总结：

本周AI领域两大事件：开源模型成最大赢家，美国闭源路线遭遇双重打击

本周，两则看似独立的事件因“开放”一词紧密相连。首先，一款中国开源模型引发了芯片股自今年4月以来最糟糕的一周——投资者终于开始追问，7250亿美元的人工智能资本支出究竟换来了什么。数日后，当一个自主智能体入侵Hugging Face平台时，其安全团队竟被美国前沿模型的防护栏挡在门外，最终不得不使用一款中国开源模型进行取证分析。在这两条战线上，美国的闭源前沿押注都遭遇了重创，而开源模型成为共同的赢家。与此同时，华盛顿方面却在同一周加紧收紧闭源模型的购买限制。

市场巨震：中国开源模型重估AI投资逻辑

上周，美国科技板块遭遇4月以来最大跌幅。纳斯达克100指数下跌4.1%，半导体指数下跌10%。据彭博社报道，谷歌、微软、亚马逊和Meta等AI最大支出方正面临投资者要求回报的压力，这些公司今年资本支出指引高达7250亿美元，华尔街预计2027年将接近9000亿美元。直接诱因是一款中国免费提供的模型：月之暗面推出的Kimi K3在48小时内因GPU算力饱和而被迫暂停新订阅，其开源权重模型与美国前沿实验室的差距正迅速缩小。紧随其后，阿里巴巴公开了Qwen 3.8模型，称其2.4万亿参数规模仅落后于Claude Fable 5，并承诺很快开源。

安全防线失守：Hugging Face遭入侵，美国模型“误伤”防守者

Hugging Face遭遇自主智能体入侵后，其安全团队尝试使用商业API背后的前沿模型分析攻击，却遭到护栏机制阻拦——该系统无法区分是攻击者还是事件响应人员。最终，团队转而使用中国开源模型GLM 5.2在自己的硬件上完成取证，确保攻击数据和被盗凭证不出平台。与此同时，WordPress核心漏洞被公开利用，该漏洞影响超过5亿个站点，利用代码已在网络流传。

政府出手：白宫逐客审批，欧盟自建AI平台

美国政府正以前所未有的力度介入AI供应链。据CNBC报道，白宫已直接决定哪些企业能获得OpenAI和Anthropic的最先进模型，此前这一权力由实验室掌握。特朗普政府以国家安全为由一度禁止Claude Mythos 5和Fable 5，随后才恢复访问。OpenAI CEO Sam Altman透露，政府现在逐个审批GPT-5.6的客户。此外，华尔街日报披露一名CIA官员在调查阿联酋AI龙头G42与中国的联系后，反而帮助后者获得美国芯片准入。欧盟议会也计划于9月推出内部AI平台，统一管理议员使用OpenAI、Anthropic、Meta和Mistral等模型的行为。

核心结论：开源模型在华尔街和网络安全两条战线均胜出

本周两大事件指向同一个结论：当美国闭源前沿路线变得过于昂贵或过于封闭时，开源替代方案——且越来越多来自中国——正成为现实选择。Kimi K3以极低成本缩小能力差距，GLM 5.2完成了美国前沿模型因安全过滤而拒绝执行的应急响应工作。这种替代方案正吸引机构资金：非营利组织Current AI本周宣布获得4亿美元承诺，用于建设公共开源AI基础设施，其中法国出资1亿美元。

讽刺的是，就在开源模型赢得华尔街和网络安全“双重胜利”的同一周，华盛顿正在决定谁才有资格购买美国闭源模型。

中文翻译：

本周的两则故事，由一个关键词串联：开放。一款中国开源权重模型触发了自四月以来芯片股表现最糟糕的一周，投资者终于开始追问，7250亿美元的人工智能资本支出究竟买来了什么。数日后，当一名自主智能体入侵了Hugging Face，其防御团队反而被美国前沿模型的护栏系统拒之门外，不得不转而使用一款中国开源模型来进行取证分析。在这两个战场上，美国封闭式前沿模型的押注都度过了惨淡的一周，而开源权重成为了共同的赢家。与此同时，华盛顿方面在同一周内正加紧限制购买封闭模型。下文将解析：抛售潮及其诱因、安全事件中的人工智能博弈，以及国家力量深入科技产业链的动向。

赞助商
人工智能智能体在变，但你的需求不该变。
当模型、提示词、工具和工作流不断演化，Spec27能帮助团队重新运行验证，精准捕捉智能体行为何时偏离了既定路径。

行业动态
从应用榜单到社区热帖，人工智能领域当下最受关注的话题：《行业动态》最新一期全景速览。

• 你的NotebookLM现已更名为Gemini Notebook。谷歌对该研究应用进行了更名，并将其更深度地整合进Gemini，新增了代码执行功能，并可从Gemini应用及搜索引擎中直接访问。目前用户已超3000万，因此本周不少用户打开应用时都看到了一个新名称。（TechCrunch）
• 人工智能世界杯背后，隐藏着大量人力劳动。随着决赛刚刚落幕，《Rest of World》调查了赛事人工智能功能背后的数千名数据标注员。在下一次看到“全自动”宣传标题前，这篇文章值得一读。（Rest of World）
• 亚马逊上人工智能撰写的传记泛滥成灾。《纽约时报》发现大量人工智能生成的劣质书籍，其中包括真人的虚假传记，与真实作品并排销售。购买前请先确认作者信息。（纽约时报）
• “人工智能能否拥有意识？”成为热议文章。哲学家威廉·麦卡斯基尔和卢修斯·卡维奥拉在《卫报》撰文指出，无论你对答案作何看法，我们都应从现在开始为此问题做好准备。（卫报）
• 一项广为流传的研究显示，人工智能的建议会让人更自信，却更不正确。研究人员发现，人工智能的建议降低了人们的准确率，同时提高了他们的确信度。下次当聊天机器人语气笃定时，这是个有用的自我提醒。（The Next Web）

速览
DeepSeek的悄然崛起
上周市场行情被重新定价，催化剂来自中国。

• 美国科技股经历了自四月以来最糟糕的一周。纳斯达克100指数下跌4.1%，半导体指数下跌10%。据彭博社报道，目前压力已转移到最大的人工智能支出企业身上，要求它们证明投资回报。Alphabet、微软、亚马逊和Meta已预告今年资本支出将高达7250亿美元，而华尔街预测2027年这一数字将接近9000亿美元。
• 触发因素是中国免费发布的一款模型。数日前，月之暗面的Kimi K3上线，需求激增导致公司图形处理器（GPU）达到负载上限，在48小时内被迫暂停新用户订阅。这款开源权重模型正以足以压垮自身服务器的速度，缩小与美国实验室的差距。
• 阿里巴巴随后也推出了一款开源权重竞品。在Kimi K3发布数日后，阿里巴巴的Qwen团队预告了Qwen 3.8，一个2.4万亿参数的模型，据称性能仅次于Claude Fable 5，并承诺很快将开放权重。目前尚未公布任何基准测试结果，因此这一排名目前仅是阿里巴巴的自我宣称。

人工智能供应链受创
Hugging Face被入侵事件暴露了美国护栏系统的缺陷，以及一个正在遭受主动攻击的WordPress漏洞。

• Hugging Face的防御团队被美国人工智能系统拒之门外，因此转而求助于开源模型。在一个自主智能体入侵其集群后，Hugging Face尝试使用商业应用编程接口（API）背后的前沿模型来分析攻击行为，却遭到其安全护栏系统的阻拦。该公司表示，这些护栏系统无法区分事件响应者和攻击者。最终，它使用开源权重模型GLM 5.2在自己的硬件上完成了取证分析，确保了攻击者数据和被盗凭证均未外泄。
• WordPress核心中一个影响5亿站点的漏洞，其公共利用代码现已公开。这个名为wp2shell的预认证远程代码执行漏洞，只需一个匿名请求即可在默认的WordPress安装上运行代码，编号为CVE-2026-63030和CVE-2026-60137，已在7.0.2和6.9.5版本中修复。Searchlight Cyber发现了这条利用链，据其估计，超过5亿个站点面临风险，且公开的概念验证利用代码正在流传。

各国政府认真对待的一年
在中国免费提供模型的同时，三个国家的政府采取措施，试图控制谁能获得封闭模型。

• 白宫现在逐个客户审批谁能购买前沿人工智能模型。据CNBC报道，特朗普政府正在指导哪些公司能够获取OpenAI和Anthropic的最新模型，这一决定过去由实验室自行做出。政府曾基于国家安全理由阻止了Claude Mythos 5和Fable 5的访问，随后才恢复权限。山姆·奥特曼对员工表示，政府现在正逐一审批GPT-5.6的客户。
• 一名中情局官员曾监视阿联酋的人工智能领军企业，随后帮助其赢得美国芯片。《华尔街日报》报道，乔尼·甘农以外交掩护身份调查了G42与中国的关系，随后又成为阿布扎比了解华盛顿意图的渠道，最终阿联酋成功获得了英伟达先进芯片的扩大化访问权限。
• 欧洲议会正成为其自身前沿模型的买家。据Politico报道，欧洲议会最早将于9月推出内部“EPGenAI中心”，允许欧洲议会议员及工作人员通过单一接口，合规地访问OpenAI、Anthropic、Meta和Mistral的模型。此举旨在将立法者非官方的AI使用行为纳入治理体系。

开源权重赢得两场战役
回顾本周两大事件，胜者是同一个。帮助重估美国人工智能交易的那款模型，是中国开源权重模型。美国前沿API拒绝运行的那款用于取证分析的模型，同样是中国开源权重模型。Kimi K3正以极低的成本缩小能力差距。而GLM 5.2完成了美国前沿模型安全过滤器所阻止的事件响应工作——因为托管的护栏系统无法区分攻击者和防御者。两件事都指向同一个方向。当封闭式前沿押注变得过于昂贵或限制过多而无法使用时，开源的替代方案就在那里，而且它越来越多地来自中国。
这一替代方案如今正吸引机构资金。本周，非营利组织Current AI表示已获得4亿美元承诺，用于建设公共、开放的人工智能基础设施，其中1亿美元来自法国。该组织主张，如此重要的技术需要一种公共选择。讽刺之处在于：就在同一周，开源权重在华尔街和安全响应领域双双获胜，而华盛顿却正在决定谁才有资格购买封闭的美国模型。

关键要点

• 上周人工智能交易被重新定价，催化剂来自中国而非美国实验室。芯片股经历了自四月以来最糟糕的一周，投资者开始审计7250亿美元的资本支出。
• 安全事件显示了相同的模式。当一名自主智能体入侵Hugging Face时，其防御团队被美国前沿模型的护栏系统阻挡，转而使用中国开源模型进行取证。同一周，一个影响5亿站点的WordPress漏洞正遭受主动攻击。
• 国家力量已深入到前沿科技产业链内部。白宫逐个审批前沿模型的客户，一名中情局官员斡旋了阿联酋的芯片访问权，欧洲议会正在筹建自己的人工智能中心。
• 开源权重是本周的共同赢家，并正吸引机构资金：Current AI现已获得4亿美元用于建设公共替代方案。

值得一读

• 一份新的工作论文发现，只有11%的标普500公司深度集成了人工智能。该论文通过分析SEC 10-K文件而非调查问卷得出结论，其中科技公司占了约三分之二。这是在当前巨额资本支出背后，需求侧的真实情况。（arXiv）
• 一篇新论文训练出无需反向传播的卷积网络，通过一种本地化的、生物学上可信的学习规则，在MNIST上达到96.7%，在CIFAR-10上达到61.7%。（arXiv）
• LoRA Speedrun项目上线了一个公开的微调时钟排行榜。任务和硬件固定，每条记录需重复跑三次才能生效。目前的最快纪录是在单张L40S上用时略超六分钟。（GitHub）

本周关注
《人工智能周刊》现已登陆YouTube。每周最精炼的故事，每段不超过40秒：

• 英伟达刚刚切断了其在亚洲超过一半的买家，以阻止产品流入中国：25亿美元转口案背后的打击走私行动，33秒。
• 美国人工智能实验室称中国在逐条模仿它们的模型：关于模型蒸馏的争端，以及Anthropic声称的2880万次查询。
• 数据中心目前的用电量已相当于爱尔兰所有家庭的总和：占据该国电力的23%，而十年前仅为5%。
  有用吗？请在YouTube上订阅。我们每天更新数条视频。

等等，什么？

• 数百个反数据中心的Facebook页面正在利用人工智能生成的宣传材料，来反对人工智能数据中心。404 Media发现，这些利用数据中心反对浪潮牟利的人，正在使用他们口诛笔伐的技术来制造愤怒情绪。（404 Media）

值得一看
人工智能从业者正在热传的视频——由人工智能电视策划。
|《我们正生活在“ChatGPT传单大流行”中》 404 Media | |
| 桑达尔·皮查伊谈人工智能反弹、未来工作及谷歌下一个时代 Hard Fork |

本周投票
本周，开源权重在华尔街和安全响应领域双双获胜。一年后，持久的优势在哪里？
上周，共有364位读者参与投票：
从现在起十二个月，你所在组织在人工智能方面的支出是：
本周，开源权重在华尔街和安全响应领域双双获胜。一年后，持久的优势在哪里？
周三见。
亚历克西斯

英文来源：

Two stories this week, connected by one word: open. A Chinese open-weight model helped touch off the worst week for chip stocks since April, as investors finally asked what $725 billion in AI capex is buying. Days later, when an autonomous agent breached Hugging Face, its own defenders were locked out by US frontier-model guardrails and ran the forensics on an open Chinese model instead. On both fronts the closed American frontier bet had a bad week, and open weight was the common winner. Meanwhile Washington spent the same week making the closed models harder to buy at all. Below: the sell-off and its trigger, AI on both sides of the security desk, and the state moving inside the stack.
Sponsor
AI agents change. Your requirements should not.
As models, prompts, tools, and workflows evolve, Spec27 helps teams re-run validations and spot where agent behaviour has drifted from the intended path.In the Wild
What's trending in AI right now, from the app charts to the community feeds. Full roundup in the latest In the Wild.

• Your NotebookLM is now Gemini Notebook. Google renamed the research app and folded it deeper into Gemini, adding code execution and access from the Gemini app and Search. More than 30 million people use it, so a lot of folks opened it this week to a new name. (TechCrunch)
• The AI World Cup ran on a lot of hidden human labor. With the final just wrapped, Rest of World looked at the thousands of data workers behind the tournament's AI features. Worth a read before the next "fully automated" headline. (Rest of World)
• AI-written biographies are piling up on Amazon. The New York Times found a flood of AI slop books, including fake biographies of real people, sitting on the shelves next to the real ones. Check the author before you buy. (NYT)
• "Could AI be conscious?" is the essay everyone's passing around. Philosophers Will MacAskill and Lucius Caviola argue in The Guardian that we should start preparing for the question now, whatever you think of the answer. (The Guardian)
• A study going around says AI advice can make you more confident and less right. Researchers found AI suggestions cut people's accuracy while raising their certainty. A useful gut-check next time a chatbot sounds sure. (The Next Web)
  Quick Hits
  DeepSeek's Quiet Takeover
  The trade got repriced last week, and the catalyst came from China.
• US tech just had its worst week since April. The Nasdaq 100 fell 4.1% and the semiconductor index dropped 10%, and Bloomberg reports the pressure is now on the biggest AI spenders to show a return. Alphabet, Microsoft, Amazon and Meta have guided to as much as $725 billion in capex this year, with Wall Street penciling in close to $900 billion for 2027.
• The trigger was a model China gives away. Days earlier, Moonshot's Kimi K3 landed and demand forced the company to pause new subscriptions within 48 hours as its GPUs hit capacity, an open-weight model closing the gap with US labs fast enough to strain its own servers.
• Alibaba piled on with an open-weight challenger. Days after Kimi K3, Alibaba's Qwen team teased Qwen 3.8, a 2.4-trillion-parameter model it says trails only Claude Fable 5, with open weights promised soon. No benchmarks have been published yet, so for now that ranking is Alibaba's own claim.
  AI Supply Chain Under Siege
  What the Hugging Face breach revealed about US guardrails, and a WordPress hole now under active attack.
• Hugging Face's defenders got locked out of US AI, so they turned to an open one. After an autonomous agent breached its clusters, Hugging Face tried frontier models behind commercial APIs to analyze the attack and got blocked by their safety guardrails, which it says cannot tell an incident responder from an attacker. It ran the forensics instead on GLM 5.2, an open-weight model, on its own hardware, so no attacker data or stolen credentials left the building.
• A 500-million-site hole in WordPress core just got public exploits. The wp2shell pre-authentication RCE is an anonymous request that runs code on a default WordPress install, tracked as CVE-2026-63030 and CVE-2026-60137 and patched in 7.0.2 and 6.9.5. Searchlight Cyber found the chain, which it estimates exposes more than 500 million sites, and public proof-of-concept exploits are now circulating.
  The Year Governments Got Serious
  While China gives models away, three governments moved to control who gets the closed ones.
• The White House now approves who can buy frontier AI, customer by customer. The Trump administration is directing which companies get access to the latest models from OpenAI and Anthropic, CNBC reports, a decision that used to sit with the labs. It blocked Claude Mythos 5 and Fable 5 on national-security grounds before reinstating access, and Sam Altman told staff the government is now approving GPT-5.6 customers one at a time.
• A CIA officer spied on the UAE's AI champion, then helped it win US chips. The Wall Street Journal reports that Jonny Gannon investigated G42's China ties under diplomatic cover, then became the channel through which Abu Dhabi learned what Washington wanted, and the UAE ultimately secured expanded access to Nvidia's advanced chips.
• The EU Parliament is becoming a frontier-model buyer of its own. Politico reports the Parliament will roll out an in-house "EPGenAI Hub" as early as September, giving MEPs and staff sanctioned access to models from OpenAI, Anthropic, Meta and Mistral through a single interface, an attempt to bring lawmakers' unofficial AI use under governance.
  Open Weight Won Both Fights
  Line up this week's two big stories and the winner is the same on both. The model that helped reprice the US AI trade was open-weight and Chinese. The model that did the forensics US frontier APIs refused to run was also open-weight and Chinese. Kimi K3 is closing the capability gap at a fraction of the price, and GLM 5.2 handled incident-response work that US frontier models' safety filters blocked, because a hosted guardrail cannot tell an attacker from a defender. Both stories point the same way. When the closed frontier bet gets too expensive or too locked down to use, the open alternative is sitting right there, and more and more it's Chinese.
  That alternative is now drawing institutional money. This week the nonprofit Current AI said it has $400 million in commitments to build public, open AI infrastructure, including $100 million from France, on the argument that a technology this important needs a public option. The irony sat one section up: the same week open weight won on Wall Street and at the security desk, Washington was deciding who is even allowed to buy the closed American models.
  Key Takeaways
• The AI trade got repriced last week, and the catalyst came from China rather than a US lab. Chips had their worst week since April as investors started auditing $725 billion in capex.
• The security desk showed the same pattern. When an autonomous agent breached Hugging Face, its defenders were blocked by US frontier-model guardrails and ran the forensics on an open Chinese model instead, the same week a 500-million-site WordPress hole went under active attack.
• The state moved inside the frontier stack. The White House approves frontier-model customers one by one, a CIA officer brokered the UAE's chip access, and the EU Parliament is standing up its own AI hub.
• Open weight was the week's common winner, and it is drawing institutional money: Current AI now has $400 million to build a public alternative.
  Worth Reading
• A new working paper finds only 11% of S&P 500 firms have deeply integrated AI, reading SEC 10-K filings instead of surveys, with tech companies accounting for about two-thirds of it. The demand-side reality under all that capex. (arXiv)
• A new paper trains convolutional networks without backpropagation, reaching 96.7% on MNIST and 61.7% on CIFAR-10 with a local, biologically plausible learning rule. (arXiv)
• LoRA Speedrun opens a public wall-clock leaderboard for fine-tuning, with a frozen task and hardware and every record re-run three times before it counts. The current mark is a little over six minutes on a single L40S. (GitHub)
  Watch This Week
  AI Weekly is now on YouTube. The week's sharpest stories, each under 40 seconds:
• Nvidia just cut off more than half its buyers in Asia to block China: the smuggling crackdown behind the $2.5B rerouting case, in 33 seconds.
• US AI labs say China is copying their models, one question at a time: the distillation fight, and the 28.8 million queries Anthropic alleges.
• Data centers now use as much power as every home in Ireland combined: 23% of the country's electricity, up from 5% a decade ago.
  Useful? Subscribe on YouTube. We post several a day.
  Wait, What?
• Hundreds of anti-data-center Facebook pages are cranking out AI-generated propaganda against AI data centers. 404 Media found that the grifters riding the data-center backlash are using the exact technology they are railing against to manufacture the outrage. (404 Media)
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  | We Are Living in a ‘ChatGPT Flyer Pandemic’ 404 Media | |
  | Sundar Pichai on A.I. Backlash, the Future of Work and Google’s Next Era Hard Fork |
  This week's poll
  Open weight won on Wall Street and at the security desk this week. Where's the durable edge a year from now?
  Last week, 364 of you voted:
  Twelve months from now, your organization's AI spend is:
  Open weight won on Wall Street and at the security desk this week. Where's the durable edge a year from now?
  Back Wednesday.
  Alexis