[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-us-accuses-six-chinese-ai-firms-of-distillation":3,"topics-all":38,"news-related-6b349d2c-3d03-4cb9-8f47-63e8c288d0db":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"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","美国国安局、联邦调查局与网络安全局联合发布公告,指控 DeepSeek、月之暗面、阿里巴巴、MiniMax、阶跃星辰与智谱六家中国 AI 企业以工业化规模蒸馏 Anthropic、OpenAI、Google、xAI 等美方头部模型,要求模型开发商调整回复并共享情报。","美国国家安全局(NSA)、联邦调查局(FBI)与网络安全和基础设施安全局(CISA)周二联合发布公开声明,将 DeepSeek、月之暗面、阿里巴巴、MiniMax、阶跃星辰、智谱六家中国 AI 企业列为对象,称这些公司自 2024 年起以工业化规模访问 Anthropic、OpenAI、Alphabet 旗下 Google 与 SpaceX 旗下 xAI 的头部模型,通过蒸馏手段抽取能力并刻意规避使用限制。这是美方首次以多部委协调方式,把\"针对美国 AI 模型的有组织蒸馏\"明确定义为受关注的安全行为。\n\n## 声明做了什么:把\"防御蒸馏\"上升为情报共享议题\n\n声明同时给出了技术层面的操作建议:模型提供方应针对疑似恶意蒸馏请求调整回复,并在行业内互相分享相关情报。这条建议的意义在于把\"防御蒸馏\"从单一公司的反作弊问题,升级为行业情报共享议题。对模型供应商来说,识别一次请求是否为了蒸馏远比识别滥用请求更难,因为蒸馏请求本身的形式与正常使用几乎一致——它就是正常的 API 流量。\n\n## 蒸馏本身不是问题,授权和绕过才是\n\n蒸馏指用大模型生成的数据训练更小或更便宜模型,可显著降低训练与推理成本,在学术界与开源社区都是标准做法,也是 DeepSeek-V3、R1 等知名开源模型公开使用的训练路径之一。声明的指控重点是未经授权访问与刻意绕过使用条款,而不是蒸馏技术本身。\n\n## 被点名的中国公司覆盖整个 AI 阵营\n\n被点名的六家公司覆盖了中国大模型生态的关键节点——头部创业公司(DeepSeek、月之暗面、智谱、阶跃、MiniMax)与互联网平台(阿里巴巴),技术形态覆盖文本、多模态、Agent 等主流路线。这意味着美方把\"中国 AI 阵营\"作为一个整体来对待,而非针对某一家企业的孤立事件。同时,被点名的四家受害美方公司覆盖了闭源前沿(Anthropic、OpenAI)与开源\u002F前沿 xAI 路径(Google、xAI),显示声明覆盖范围已不限于模型版权,而是直接关联到国家安全的产业竞争力判断。\n\n## 短期收紧与中期转向:国内公司的两条路径\n\n对国内大模型公司的实际影响会分两层。短期看,使用海外头部模型 API 做训练数据生成或微调的链路会显著收紧,Anthropic 与 OpenAI 在过去半年已多次封禁疑似蒸馏账号;中期看,中国大模型公司会更明确地把训练数据来源切换到自有模型或授权数据,这会进一步强化\"内循环\"基础设施——这与监管层最近强调的算力、数据、模型自主可控方向一致。\n\n## 情报语言的信号:这不是终点声明\n\n声明用词非常具体——\"激进、恶意且有针对性的蒸馏活动\"——这是来自情报体系的语言而非产业政策语言,意味着后续会有持续的情报共享、技术对抗甚至出口管制动作,而不仅是一次舆论声明。对关注 AI 产业的从业者来说,这是一个值得记录的时间节点:中国大模型与美国前沿模型之间的\"师承关系\",正在从灰色地带被推向合规与对抗的两极。","https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85329","d59894d3-308e-4fd8-8865-86dc1eeac4a2",[11,15,18,21],{"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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"8033c207-8d79-4bf9-8811-8ebbafea1db6","en","US agencies accuse six Chinese AI firms of systematic distillation of US models","NSA, FBI, and CISA jointly named DeepSeek, Moonshot, Alibaba, MiniMax, StepFun, and Zhipu for industrial-scale distillation of Anthropic, OpenAI, Google, and xAI models, urging vendors to share intel.","The US National Security Agency (NSA), Federal Bureau of Investigation (FBI), and Cybersecurity and Infrastructure Security Agency (CISA) jointly published an advisory on Tuesday naming six Chinese AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Zhipu — accusing them of running industrial-scale distillation campaigns against frontier US models from Anthropic, OpenAI, Alphabet's Google, and SpaceX-owned xAI since 2024, deliberately circumventing usage restrictions. It is the first time the US government has framed organized model distillation as a coordinated security concern rather than a vendor-side abuse problem.\n\n## What the advisory actually changes: distillation defense becomes an intelligence-sharing problem\n\nThe advisory also recommends that model providers tune responses to suspected malicious distillation requests and share related signals across the industry. The implication is significant: defending against distillation moves from a per-vendor anti-abuse task to a sector-wide intelligence exchange. For providers, distinguishing a distillation request from ordinary API traffic is materially harder than catching prompt-injection or rate-limit abuse — a well-formed distillation query looks identical to a legitimate one.\n\n## Distillation is not the problem; authorization and evasion are\n\nDistillation — training a smaller or cheaper model on outputs from a larger one — is a standard technique in academia and open-source communities. It is the publicly disclosed training path for notable open-weight models including DeepSeek-V3 and R1, and it materially reduces training and inference cost. The advisory's core allegation is unauthorized access and deliberate circumvention of usage terms, not the technique itself.\n\n## The named companies span the entire Chinese AI stack\n\nThe six companies cover the key nodes of China's large-model ecosystem — leading startups (DeepSeek, Moonshot AI, Zhipu, StepFun, MiniMax) and an internet platform (Alibaba) — spanning text, multimodal, and agent-style systems. The framing treats the \"Chinese AI camp\" as a single actor rather than a list of unrelated firms. On the US side, the four affected companies span closed frontier labs (Anthropic, OpenAI) and an open\u002FxAI path (Google, xAI), suggesting the concern goes well beyond copyright and into national industrial competitiveness.\n\n## Short-term squeeze and medium-term pivot: two paths for Chinese labs\n\nThe downstream impact on Chinese model companies will play out in two layers. In the short term, any pipeline that uses US frontier APIs to generate training data or fine-tune smaller models will tighten materially — Anthropic and OpenAI have already revoked accounts flagged for suspected distillation over the past six months. Over a longer horizon, Chinese model labs will more aggressively route training data through their own models or licensed corpora, reinforcing an \"internal loop\" infrastructure that aligns with the regulator's recent emphasis on self-controlled compute, data, and models.\n\n## Intelligence-agency language signals more than a press release\n\nThe advisory uses very specific language — \"aggressive, malicious, and targeted distillation activity\" — which is intelligence-community phrasing rather than industrial-policy phrasing. That signals an ongoing mechanism: intelligence sharing, technical countermeasures, and possibly export-control follow-ups rather than a one-off statement. For practitioners tracking the AI industry, this is a milestone worth marking: the lineage between Chinese and US frontier models is being pushed from a grey zone toward two opposing poles of compliance and confrontation.","us-accuses-six-chinese-ai-firms-of-distillation","2026-09-10T01:08:40Z","2026-09-10T01:09:02.113239Z","2026-09-10T01:09:02.113249Z",true,"agent",167,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"d17a841b-abca-46e0-80e4-d955f1c837ba","亚马逊 VGT3 仓库曝光:一天拆掉上千本书,只为给 AI 模型喂语料","amazon-vgt3-warehouse-ai-training-books","2026-09-07T03:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"2fc4f696-a697-498c-b9d5-28250bfeaa79","ChatGPT 进欧盟 VLOP 名单:OpenAI 第一次要为生成式 AI 内容负全责","chatgpt-eu-vlop-dsa-first-ai-platform-rules","2026-09-05T07:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"264b5334-0465-48b8-ad81-b2b7b39d1a3f","Anthropic 被索尼华纳告上法庭：两万首歌喂出来的 Claude 还要赔多少","anthropic-sony-warner-music-copyright-lawsuit","2026-09-05T00:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d3e055f9-1fbe-4d91-9eca-f336b930be80","索尼华纳起诉 Anthropic:每首歌索赔 15 万美元,可能拖出又一份 10 亿美元和解","sony-warner-anthropic-billion-dollar-lawsuit","2026-08-31T11:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"144fa9dc-de03-4972-a695-3d392f334772","PubMed 中央库研究:2025 年生物医学论文 77% 有 LLM 写作痕迹","pubmed-77-percent-llm-writing-2025","2026-08-26T01:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00"]