G7峰会上美国推出“可信伙伴”AI模型访问白名单，此前已强令Anthropic前沿模型下线
英文摘要
At the G7 summit in the French Alps, OpenAI, Google DeepMind, and Anthropic CEOs met with leaders for the first time. The core discussion centered on a US proposal for a 'trusted partners' whitelist, which would exempt some allies from export restrictions on frontier AI models after Anthropic's Fable 5 and Mythos 5 were abruptly taken offline by the US Commerce Department on June 13. The US has already classified frontier model weights under export control number ECCN 4E091, dividing countries into tiers with close allies getting minimal review. Concurrently, the EU unveiled a €422 billion AI sovereignty plan and Mistral invested €1.2 billion in Sweden, while Cohere's enterprise inquiries surged after the Anthropic ban. Paradoxically, Microsoft is reportedly evaluating a fine-tuned DeepSeek V4 as a low-cost alternative for Copilot, and Japanese companies like Rakuten have been found to build their top AI models on top of Chinese code such as DeepSeek-V3. China also proposed establishing a 'World AI Cooperation Organization' headquartered in Shanghai, shifting the global narrative from who gets to use models to who sets the rules.
中文摘要
在法国阿尔卑斯山举行的G7峰会上，OpenAI、Google DeepMind、Anthropic的CEO首次与各国领导人共进工作午餐。核心议题是美国提出的“可信伙伴”白名单方案，因美国商务部于6月13日突然下令Anthropic的Fable 5和Mythos 5模型全球下线，出口管制迫使盟友寻求豁免通道。美国已根据ECCN 4E091分类将前沿模型权重纳入出口管制，将国家分为三档，亲密盟国基本免审。与此同时，欧盟启动了4220亿欧元的AI主权计划，Mistral在瑞典投资12亿欧元，而Cohere因Anthropic断供企业咨询量暴增。讽刺的是，微软正评估用微调过的DeepSeek V4作为Copilot的低成本替代，日本乐天等企业被曝光所谓的“国产最强AI”实为基于DeepSeek-V3的套壳。中国则倡议成立“世界人工智能合作组织”，将争论焦点从‘谁能用模型’转向‘谁来定规则’。
关键要点
At the G7, the US proposed a 'trusted partners' whitelist to give allies access to frontier AI models while maintaining export controls after forcing Anthropic's Fable 5 and Mythos 5 offline.
G7上美国提出“可信伙伴”白名单，在强令Anthropic的Fable 5和Mythos 5模型下线后，允许盟友有条件获取前沿模型。
The US has already classified frontier model weights under ECCN 4E091, dividing countries into three tiers with close allies getting minimal review and others facing a presumption of denial.
美国已根据ECCN 4E091分类将前沿模型权重纳入出口管制，将国家分为三个等级，亲密盟国基本免审，其他国家则推定不予批准。
The EU responded with a €422 billion, 10-year AI sovereignty plan and Mistral's €1.2 billion investment in Sweden, while Cohere saw a surge in enterprise inquiries.
欧盟推出十年4220亿欧元的AI主权计划，Mistral在瑞典投资12亿欧元，Cohere企业咨询量因Anthropic断供暴增。
Microsoft is evaluating a fine-tuned DeepSeek V4 as a low-cost replacement for Anthropic/OpenAI models in Copilot, and Japan's Rakuten AI 3.0 was exposed as a re-skinned DeepSeek-V3.
微软正评估用微调过的DeepSeek V4替代Copilot中的美国模型；日本乐天AI 3.0被曝光是DeepSeek-V3套壳。
China proposed the 'World AI Cooperation Organization' headquartered in Shanghai, shifting the global focus from access to AI models to who sets the rules.
中国倡议成立总部拟设在上海的“世界人工智能合作组织”，将全球焦点从模型使用权转向规则制定权。