Dario Amodei, CEO of Anthropic, published a 3,800-word essay on September 12 proposing a three-step plan to regulate AI development. The framework includes mandatory safety testing for all models, regardless of whether they are proprietary or open-weight, and calls for granting independent evaluators like METR access to frontier systems.
The proposal advocates for global coordination on AI regulation and stricter export controls on chips to China. Amodei argues that open-weight models pose a national security risk, as they allow adversaries to replicate advanced capabilities through distillation. In this process, smaller models mimic powerful ones by querying them millions of times.
Anthropic reported that operators linked to Alibaba conducted a distillation campaign against its Claude model, involving nearly 29 million exchanges. The company suggests that such large-scale API usage could be treated as intellectual property theft. These measures could impact the distribution strategies of companies like Meta, Alibaba, and Mistral.
The initiative has received support from Sam Altman, Elon Musk, and Demis Hassabis. Critics note that this regulatory environment could benefit companies like OpenAI and xAI, which operate closed systems, by imposing higher compliance costs on competitors that release open-weight models.
