Anthropic founder and CEO Dario Amodei on Monday afternoon responded to industry tweets suggesting the company is in some way supporting the U.S. government’s efforts to ban Chinese openweight models, or openweight models in general.
“Anyone who has read my past writings should know that I do not consider such a ban to be a beneficial measure, but let me state it clearly for the avoidance of doubt: Anthropic has never advocated for a ban on indiscriminate weight models,” he wrote, emphasizing his opinion.
His response came after Nvidia founder and CEO Jensen Huang shared an open letter to X (first post on the platform) 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.
Although the letter did not specifically mention China, industry discussions have centered on claims that Chinese AI labs often improve their capabilities by stealing intellectual property from U.S. AI labs. One method is distillation. In this method, the AI prompts another model to learn how it works.
Amodei said in a blog post published Monday that he believes the open-weight model falls into a different bucket than the threat posed by China. “An open-weight model with no dangerous features is a public good; it costs nothing beyond the compute needed to run, and it provides value to businesses, developers, and researchers.”
Amodei said in his post that he has long-standing concerns about AI, but that companies using indiscriminate weight models, even those made in China, are not among them.
His concern is that “authoritarian governments” will build a more powerful model than the United States to use AI to achieve “enduring military superiority” and/or to repress their own citizens. He said the Chinese Communist Party (CCP) is not the only authoritarian government he is concerned about, but it is the “most capable” one.
Amodei also said he was concerned that AI could enable “biological attacks” as well as cybersecurity attacks. And, in his view, open-weight models are more dangerous in such scenarios because it is difficult to apply guardrails and monitor their usage. He also cited a report from the UK Institute for AI Security, which said that once open weight is released, it cannot be withdrawn.
This goes against the argument of open source proponents that access to a powerful open model that is not controlled by a single entity helps defenders protect themselves.
Amodei cited actions he believes would hinder China, including restricting access to powerful chips (a longstanding U.S. policy) and calling for a formal crackdown on distillation. The United States has threatened to impose sanctions if China determines that intellectual property related to American models has been stolen.
Interestingly, Amodei also said he supports expanding the US-led effort to create a model safety testing laboratory, especially if the entire world, including China, agrees to follow suit.
“I think this idea is actually close to consensus, and I’m encouraged both by the Trump administration’s moves in this direction in recent months and recent industry proposals to apply such tests to the most capable models, regardless of country of origin or whether they are open or closed (and completely exempting less capable models, such as those from startups or academia),” Amodei wrote.
“Please note that to be effective, testing would need to be done globally, which means the Chinese Communist Party would need to be involved as well. I think this may actually be possible… Limited cooperation on preventing AI biological weapons may be possible, because that is also in China’s interest.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
