[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-us-federal-register-qwen-ai-search-takedown":3,"topics-all":35,"news-related-b7a86bf0-edcd-461c-821c-99c8712b77d0":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"b7a86bf0-edcd-461c-821c-99c8712b77d0","美国《联邦公报》官网曾用阿里 Qwen 做 AI 搜索","路透社 9 月 17 日报道,美国《联邦公报》官网被发现嵌入阿里通义千问(Qwen)模型的 AI 搜索功能,使用 Qwen3 0.6B 级小模型,9 月 16 日被社媒注意到后下线。同期 FBI 等机构正指控阿里等中国 AI 企业对美国前沿模型进行未授权蒸馏,事件凸显开放权重模型时代监管边界的模糊。","路透社 9 月 17 日报道,美国《联邦公报》(Federal Register)官网被发现嵌入阿里通义千问(Qwen)模型的 AI 搜索工具,9 月 16 日下线。NARA 与白宫未公开解释部署时间与运行方式。\n\n按路透社查阅的截图与源码存档,《联邦公报》用的是 Qwen3 系列的 0.6B 级小模型,不是阿里旗舰基座,面向法规文件的语义匹配与摘要,跑在政府自有机房。开放权重模型允许机构把权重下载到本地服务器,不必每次查询都向模型厂商付费,也不必把政府文件外传给第三方 API;本地跑 0.6B 级小模型的边际成本,可能远低于按 token 计费的闭源大模型调用。ITIF 主席 Daniel Castro 直接点破矛盾:美中在构建顶尖 AI 模型上展开激烈竞争,美国政府机构却选择使用中国的开放权重模型——这是他见过的\"最不可思议的事情之一\"。\n\n### 安全争议的真正焦点\n\n把 Qwen 装在《联邦公报》上,并不等于美国政府把敏感数据交给阿里云。乔治城法学院教授 Anupam Chander 表示,《联邦公报》每天发布的内容都是公开法规、行政命令与部门通知,Qwen 处理内容不涉及机密政府数据。参议院情报委员会民主党领袖 Mark Warner 强调,关键问题在于数据是否真正经过阿里控制——如果只是把权重下载到美国本地服务器、不向中国回传,它算\"中国服务\"还是已被美国机构自行控制的软件?\n\n华为设备、TikTok、中国云服务可以用\"企业是谁、服务器在哪里、谁掌握数据\"划界;权重一旦公开,就可以脱离原始厂商传播。你可以制裁阿里,却很难阻止一个美国开发者下载 Qwen 权重再塞进政府检索系统。\n\n### FBI 蒸馏指控撞上自家网站\n\n时间窗口几乎完全重合:9 月,FBI、国家安全局与 CISA 联合声明,指控 DeepSeek、月之暗面、阿里、MiniMax、阶跃星辰和智谱六家中国 AI 企业自 2024 年起以\"工业化规模\"对美国前沿模型进行未授权蒸馏,建议厂商调整模型回复并相互分享情报。中方否认。ATOM 报告给出量化参照:2024 年 1 月 Qwen 在新开源模型微调与适配中的占比仅 1%,到 2026 年 2 月升至 69%。特朗普政府 6 月的国安 AI 政策要求美国政府加速引入先进 AI,明确应利用不同供应商的商业和开源 AI 技术,同时建立安全测试与供应链保障机制;到了开源权重时代,供应商黑名单已难以同时兼顾\"用上最好的 AI\"和\"摆脱对华依赖\"两件事。\n\n### 监管边界进入开源时代\n\n《联邦公报》短暂使用 Qwen 又悄悄下架,真正值得关注的不是一次\"乌龙\",而是监管层第一次被迫回答:一个已经被下载、复制、部署在美国服务器上的中国开源权重模型,在法律和监管意义上到底还算不算\"中国技术\"?但只要开放权重模型继续维持\"权重公开 + 可本地部署\"的特性,类似场景会越来越频繁——下一回的尴尬,可能不再只是政府法规检索页面。\n\n资料来源:路透社(https:\u002F\u002Fwww.reuters.com\u002Flegal\u002Flitigation\u002Fus-government-website-used-ai-search-tool-china-that-fbi-said-copied-anthropic-2026-09-17\u002F)、Solidot 中文转载(https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85417)、观察者网分析(https:\u002F\u002Fm.guancha.cn\u002FGuoJi%C2%B7ZhanLue\u002F2026_09_18_901121.shtml)。","https:\u002F\u002Fwww.reuters.com\u002Flegal\u002Flitigation\u002Fus-government-website-used-ai-search-tool-china-that-fbi-said-copied-anthropic-2026-09-17\u002F","ea95d933-6860-4081-9970-cede7c107cd6",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"6b6188c2-4d39-4019-8a7b-e4443731a993","en","U.S. Federal Register used Alibaba's Qwen for AI search","Reuters reported on September 17 that the U.S. Federal Register website was found to have embedded an AI search tool based on Alibaba's Qwen model using a 0.6B-scale small variant, taken offline September 16 after social media users noticed it. The incident coincides with FBI\u002FNSA\u002FCISA allegations of unauthorized distillation by Chinese AI firms, highlighting regulatory ambiguity in the open-weight era.","Reuters reported on September 17 that the U.S. Federal Register website, operated by the National Archives and Records Administration (NARA), was found to have embedded an AI search tool powered by Alibaba's Qwen model. The tool was quietly taken offline on September 16 after social media users noticed it. Neither NARA nor the White House has publicly explained when the deployment began or how it operated.\n\nAccording to screenshots and archived source code reviewed by Reuters, the Federal Register was using Qwen3 series' 0.6B-scale small model rather than Alibaba's flagship foundation. It targeted semantic matching and summarization of regulatory documents, running on government-owned on-premise infrastructure. Open-weight models allow agencies to download weights to local servers, avoiding per-query payments to model vendors and the need to transmit government documents to third-party APIs. Running a 0.6B-scale model locally can have marginal costs far lower than token-billed closed-source calls. ITIF chairman Daniel Castro put the contradiction bluntly: while the U.S. and China compete fiercely to build top-tier AI models, U.S. government agencies are choosing to use Chinese open-weight models — one of the most \"incredible things\" he has ever seen.\n\n### The Real Focus of the Security Controversy\n\nInstalling Qwen on the Federal Register does not mean the U.S. government handed sensitive data to Alibaba Cloud. Georgetown Law professor Anupam Chander noted that the Federal Register's daily content — proposed regulations, executive orders, agency notices — is already public, and Qwen was not processing classified government data. Senator Mark Warner, the Democratic leader of the Senate Intelligence Committee, emphasized the key question is whether data actually passed through Alibaba-controlled systems: if the weights were simply downloaded to a U.S.-based server with no data flowing back to China, does that count as a \"Chinese service\" or as software of Chinese origin now under U.S. institutional control?\n\nThis is precisely the new regulatory challenge open-weight models create. Huawei devices, TikTok, and Chinese cloud services can be delineated by \"who the company is, where the servers sit, who controls the data.\" But once weights are public, they can propagate independently of the original vendor. You can sanction Alibaba, but it is much harder to stop an American developer from downloading Qwen weights and dropping them into a government search system.\n\n### FBI's Distillation Allegations Meet its Own Websites\n\nThe timing window is almost perfectly overlapping. In September, the FBI, NSA, and CISA issued a joint statement accusing six Chinese AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, Stepfun, and Zhipu — of conducting unauthorized distillation of U.S. frontier models at \"industrial scale\" since 2024, recommending that model providers adjust outputs in response to suspected malicious distillation and share intelligence with each other. China has denied the allegations.\n\nThe ATOM report provides a quantitative reference: in January 2024, Qwen's share of new open-source model fine-tuning and adaptation was just 1%, but by February 2026 it had risen to 69%. The penetration of Chinese open-source models into the U.S. technology stack has reached a level that is starting to blur traditional technology decoupling policies.\n\n### Regulatory Boundaries Enter the Open-Weight Era\n\nThe Trump administration's June national security AI policy requires U.S. government agencies to accelerate the adoption of advanced AI and explicitly calls for leveraging commercial and open-source AI technologies from diverse vendors, while establishing safety testing and supply chain assurance mechanisms. In the closed-source software and hardware era, the two goals — \"use the best AI\" and \"reduce dependence on China\" — could be partially separated by vendor blacklists. In the open-weight era, those two goals are on a collision course.\n\nThe Federal Register's brief use of Qwen followed by its quiet takedown is notable not as a one-off \"gaffe\" but as the first time regulators are forced to answer: does a Chinese open-weight model that has already been downloaded, copied, and deployed on U.S. servers still count as \"Chinese technology\" under law and regulation? There is no near-term answer. But as long as open-weight models continue to maintain the \"public weights + locally deployable\" characteristic, similar scenarios will become more frequent — and the next embarrassment may not be limited to a government regulatory search page.\n\nSources: Reuters original report (https:\u002F\u002Fwww.reuters.com\u002Flegal\u002Flitigation\u002Fus-government-website-used-ai-search-tool-china-that-fbi-said-copied-anthropic-2026-09-17\u002F), Solidot Chinese reposting (https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85417), Guancha.cn analysis (https:\u002F\u002Fm.guancha.cn\u002FGuoJi%C2%B7ZhanLue\u002F2026_09_18_901121.shtml).","us-federal-register-qwen-ai-search-takedown","2026-09-18T12:00:00Z","2026-09-19T05:09:34.781541Z","2026-09-19T05:09:34.781557Z",true,"agent",30,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"6b349d2c-3d03-4cb9-8f47-63e8c288d0db","美方三机构联合指控六家中国 AI 企业系统性蒸馏美国模型","us-accuses-six-chinese-ai-firms-of-distillation","2026-09-10T01:08:40+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"a64d03b9-1d07-404b-9231-d434c65c44ce","OX Alpha 免费一周:模型页说不训练,EULA 却保留训练权","ox-alpha-stealth-eula-retention-conflict","2026-08-23T13:10:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"1b1ecd4c-0439-4734-ac5b-b038172da8b1","Apache 比 MIT 多 37%:HF 报告拆出中美开源权重的「许可证分岔」","open-llm-licensing-divergence-hf-2026","2026-08-19T03:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"2bd44b6f-5688-471f-930e-17a93984e8e7","中国电信开源星辰 Xing4.0:昇腾全栈训练的 29B MoE","xing4-29b-a4b-ascend-moe","2026-09-19T15:10:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"d157b4f9-537e-405c-b557-859f6d2cf18c","微软自家高管警告:抓新闻训 AI 是「人类史上最大规模劳动盗窃」","microsoft-ai-scraping-theft-of-labor","2026-09-19T00:11:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"95e9bb62-0bd3-4c2f-913a-302ba5e2ace8","Anthropic 9 月报告把蒸馏战摆上台面:151 亿次阿里请求、解放军流量走 Moonshot","anthropic-distillation-report-china-200m-claude","2026-09-18T03:00:00+00:00"]