Our position on open-weights models
anthropic ·
Anthropic CEO Dario Amodei clarifies that the company has never advocated for banning open-weights AI models, despite recent speculation. He identifies two primary risks: authoritarian governments like China using superior AI for military or repression purposes, and powerful models being misused for cyber or biological attacks. Amodei argues that prohibitions on US businesses using Chinese open-weights models would do little to mitigate these threats, as bad actors are not typically legitimate US firms. Instead, he recommends three targeted measures: restricting chip and chipmaking equipment sales to China, cracking down on industrial-scale distillation that lets China leverage limited chips, and requiring mandatory safety testing for all sufficiently capable models, whether open or closed. He responds to an open letter supporting open-weights models, agreeing that they expand access and competition, but disputing the claim that they inherently aid defenders more than attackers, calling for empirical testing rather than assumptions. Amodei concludes that policy should focus on these practical measures rather than a blanket ban on open-weights models. He also notes that recent research on modular training strategies, cited in the post, was a collaboration between Anthropic and AE Studio.
anthropic