Anthropic CEO Dario Amodei stated his company does not support banning open-weight AI models. He denied claims that his company backs U.S. government efforts to ban Chinese or open-weight models. Amodei said bans are not effective. He made it clear that Anthropic has never called for a ban on open-weight models.
These comments followed a post by Nvidia CEO Jensen Huang. It was an open letter signed by major AI companies. These included Meta, Microsoft, Hugging Face, Mistral, and Nvidia. The letter urged policymakers to avoid early, broad restrictions on open-weight AI models. The open-weight AI models are models whose underlying code is publicly available.
The discussion in the sector focuses on the growing capabilities of Chinese AI companies. Some accuse them of using U.S. companies' intellectual property. They use a technique called distillation. This involves sending many queries to an AI model to understand its workings. Then, they recreate its abilities.
Open Models Are Not the Problem
Amodei explained that open-weight models without dangerous abilities offer public benefit. They need only the computing power to run. They provide great value to companies, developers, and researchers. His old concerns about AI do not include companies using these models. This is true even if they are Chinese. The real worry is authoritarian governments developing models. These models could surpass U.S. counterparts. This could give them lasting military superiority or enhance tools for control.
He noted the Chinese Communist Party is not the only authoritarian system he worries about. However, he sees it as most capable of achieving this scenario. He believes open-weight models, even from China, are not the main issue. Instead, the potential misuse by powerful governments is the primary concern.
Risks Beyond Cyber Attacks
Amodei also mentioned that the most dangerous AI scenarios go beyond cyber attacks. They include the potential use of AI in biological attacks. He pointed out that open-weight models increase these risks. This is because they lack controls that can be put on closed models. It is also hard to track how they are used after release.
Amodei cited a report from the UK's AI Security Institute. The report states open-weight models cannot be recalled once public. This view differs from open-source software advocates. They believe sharing powerful models with everyone helps researchers and security experts. It gives them better tools for defense and risk discovery.
Stricter Controls on China and Global Safety Tests
Amodei suggested actions to counter China's AI progress. This includes continuing to restrict China's access to advanced computer chips. The U.S. has followed this policy for years. He also proposed strict measures against using distillation if it involves intellectual property theft. This means copying someone else's ideas or creations.
At the same time, he supports initiatives for an international body. This body would test the safety of the most advanced AI models. He believes all nations must participate for this to succeed. This includes China.
Amodei noted growing agreement on this approach. He pointed to actions by the Trump administration. He also mentioned industry proposals. These suggest standard safety tests for the most powerful models. This would apply regardless of origin or if they are open or closed. Less advanced models, like those from startups or academic institutions, would be exempt.
Amodei concluded that international cooperation on limiting AI risks is possible. This applies to the development of biological weapons. He stated such cooperation also benefits China.