Anthropic 的 Dario Amodei 回应：不反对开放权重模型，但担忧中国 AI

Anthropic 创始人兼 CEO Dario Amodei 周一下午回应了业界的议论：有人认为该公司以某种方式支持美国政府禁止中国开放权重模型的行动，甚至支持禁止开放权重模型本身。

“任何读过我过去文章的人都应该知道，我并不认为这类禁令是有用的措施；但让我明确说明，以免存在任何疑问：Anthropic 从未倡议禁止开放权重模型，”他写道，强调为其本人所加。

此前，Nvidia 创始人兼 CEO Jensen Huang 在 X 平台发布了他的第一条帖子，分享了一封周五发布的公开信。Nvidia 及包括 Hugging Face、Meta、Microsoft、Mistral 和 Nvidia 在内的一长串其他 AI 公司在信中敦促政策制定者，不要对开放权重 AI 模型施加范围广泛的“过早限制”。

尽管该公开信没有特别提及中国，但业界的辩论一直围绕这样的指控展开：中国 AI 实验室的能力不断提升，往往是通过窃取美国同行的知识产权实现的。其中一种方式是蒸馏，即一个 AI 用大量提示词“轰炸”另一个模型，以了解其运作方式。

Amodei 在周一发布的博客文章中表示，他认为开放权重模型与中国构成的威胁属于不同范畴。“不具备危险能力的开放权重模型是一种公共产品：除了运行它们所需的计算资源外，它们不需要任何成本，并能为企业、开发者和研究人员创造价值。”

Amodei 在文章中称，他长期以来一直对 AI 抱有担忧，但企业使用开放权重模型——即使是来自中国的模型——并不在他的担忧之列。

他担心的是，“威权政府”会构建比美国模型更强大的模型，以实现“永久性的军事优势”和／或利用 AI 压迫本国人民。他表示，中国共产党（CCP）并不是唯一令他担忧的威权政府，但它是“能力最强的”。

Amodei 还表示，他担心 AI 不仅会促成网络安全攻击，也会促成“生物攻击”。而在他看来，开放权重模型在这些情境下更危险，因为很难为其施加安全护栏或监控其使用情况。他还援引英国 AI 安全研究所的一份报告指出，一旦开放权重模型被发布，就无法再撤回。

这与开源倡导者的观点相反：他们认为，获得不受单一实体控制的强大开放模型，有助于防御方保护自身安全。

Amodei 列出了他认为能够遏制中国的行动，包括限制其获得高性能芯片——这长期以来一直是美国的政策——以及呼吁正式打击蒸馏行为。如果美国认定涉及美国模型的知识产权盗窃行为确实存在，美国曾威胁将对中国实施制裁。

值得注意的是，Amodei 还表示，他支持日益增多的相关努力，其中部分由美国推动，以建立一个模型安全测试组织，尤其是在包括中国在内的全世界都同意接受其测试的情况下。

“我认为这个想法实际上已接近共识：令我感到鼓舞的是，Trump 政府近几个月来已朝这个方向推进；近期行业提案也提出，应当对最具能力的模型实施此类测试，无论其原产国，也无论其是开放还是封闭模型（同时完全豁免能力较弱的模型，例如来自初创公司和学术界的模型），”Amodei 写道。

“请注意，要想有效，”他补充道，“测试必须是全球性的，这意味着即使是中共也需要参与其中。我认为这实际上可能实现……围绕防止 AI 生物武器的有限合作或许是可能的，因为这同样符合中国的利益。”

Anthropic founder and CEO Dario Amodei responded on Monday afternoon to murmuring in the industry that his company somehow supports efforts by the U.S. government to ban open-weight Chinese models, or possibly open-weight models generally.

“Anyone who has read my past writing should know that I don’t regard such bans as a useful measure, but let me state it clearly so that there is no doubt: Anthropic has never advocated for a ban on open-weights models ,” he wrote, emphasis his own.

His response came after Nvidia founder and CEO Jensen Huang took to X (his first post on the platform) to share an open letter on Friday in which Nvidia and a long list of other AI companies including Hugging Face, Meta, Microsoft, Mistral, and Nvidia urged policymakers not to impose broad “premature restrictions” on open-weight AI models.

While that letter did not mention China specifically, the debate in the industry has centered on allegations that Chinese AI labs are growing in capability often by stealing intellectual property from their American counterparts. One method is distillation, where an AI bombards another model with prompts to learn how it works.

Amodei said in the blog post published Monday that he views open-weight models as falling into a different bucket than the threat China poses. “Open-weights models that don’t have dangerous capabilities are a public good: they don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.”

Amodei said in his post that he has longstanding fears about AI, but businesses using open-weight models, even ones from China, are not among them.

His fear is that “authoritarian governments” will build models that are more powerful than those in the U.S. to achieve “permanent military superiority” and/or use AI to repress their own people. He said the Chinese Communist Party (CCP) isn’t the only authoritarian government he’s concerned about, but it is the “most capable.”

Amodei also said he fears AI will enable “biological attacks” not just cybersecurity ones. And, in his view, open-weight models are more dangerous in those scenarios because it’s difficult to apply guardrails to them or monitor their usage. He also noted, citing a U.K. AI Security Institute report, that once open-weights are released they cannot be withdrawn.

This runs counter to what open source advocates argue: that access to powerful open models not controlled by a single entity helps defenders protect themselves.

Amodei listed actions he believes would thwart China, including restricting its access to powerful chips (which has been a longstanding U.S. policy) and calling for a formal crackdown on distillation. The U.S. has threatened sanctions against China if it determined there was IP theft involving U.S. models.

Interestingly, Amodei also said he supports growing efforts, some led by the U.S., to create a model safety testing organization, especially if the entire world, including China, agreed to submit to it.

“I think this idea is actually close to a consensus: I have been heartened both that the Trump administration has moved in this direction in recent months, and by recent industry proposals that would apply such testing to the most capable models regardless of their country of origin or whether they are open or closed (while exempting less capable models, such as those from startups and academia, entirely),” Amodei wrote.

“Note that to be effective,” he added, “testing would need to be global, which means even the CCP would need to be on board. I think this may actually be possible … limited cooperation around preventing AI biological weapons may be possible because it is in China’s interest too.”