In a blog post titled 'Our position on open-weights models,' Anthropic CEO Dario Amodei addressed recent debates over open-weights AI models, particularly those from China. He firmly denied that Anthropic has ever advocated for a ban on open-weights models, countering claims that the company seeks protectionist measures.
Amodei outlined two primary national security concerns: first, the risk that authoritarian governments, especially China, could develop more powerful AI models than the US for military or repressive purposes; second, the risk that powerful AI models could be misused for cyber or biological attacks. He argued that banning US businesses from using Chinese open-weights models would not address these threats, as bad actors are unlikely to be legitimate US companies.
Instead, Amodei supports three measures: restricting the sale of powerful chips and chipmaking equipment to China; cracking down on industrial-scale distillation, which allows China to build better models despite chip bans; and requiring mandatory safety testing for all sufficiently capable models, both open and closed. He emphasized that open-weights models without dangerous capabilities are a public good, but that risks should be empirically assessed through testing.
Amodei also commented on an open letter supporting open-weights models, agreeing with its points about expanding access and competition, but disagreeing that open-weights models necessarily favor defenders over attackers. He called for rigorous pre-release testing to settle such questions.