CEO's clarification on open-weight position
Anthropic co-founder and CEO Dario Amodei publicly rejected claims that his company wants Chinese open-weight AI models banned, writing in a Monday blog post that "Anthropic has never advocated for a ban on open-weights models." The statement, titled "Our position on open-weights models," followed days of criticism that Anthropic was the only major frontier AI lab to withhold its signature from an industry letter defending open-weight releases. Amodei framed less-capable open models as a public good but warned that once sufficiently capable model weights are released, they cannot be recalled or fitted with new safeguards, complicating response to misuse.
Three proposed policy measures
Rather than a ban, Amodei outlined three targeted measures he said better address national security risks. First, he urged Washington to tighten chip and chipmaking-equipment export controls to China and crack down on smuggling and workarounds, calling a chip ban "the most efficient and direct way" of preventing Beijing from building frontier AI models for military use and domestic surveillance. Second, he called for action against what he described as industrial-scale distillation, the practice of using a more capable model's outputs to train another system, warning it could shrink the US-China capability gap to months. Third, he proposed mandatory safety testing for cyber, biological and alignment risks on any sufficiently capable model, open or closed, regardless of where it was built, with exemptions for less-capable models from startups and academia.
Industry divide over the Nvidia letter
The clarification came after Nvidia, Microsoft, Meta, Hugging Face and others published an open letter titled "Open Weights and American AI Leadership" on Nvidia's website on July 24. That letter, which had grown to more than 50 signatories by late July including OpenAI, Google and SpaceX, urged US policymakers not to impose broad restrictions on open-weight models. Anthropic and Amazon remained notable holdouts. Amodei said he agreed with parts of the letter, including the case that open weights expand access and strengthen competition, but disputed the suggestion that broad access necessarily helps defenders more than attackers, a position Nvidia CEO Jensen Huang had promoted. Critics including former White House AI adviser David Sacks publicly suggested Anthropic's stance was shaped by competition from lower-cost Chinese models; Amodei rejected that interpretation, noting Anthropic's revenue is concentrated on API access and it has no history of releasing weights.
Chinese AI competition and Kimi K3
The policy debate has intensified following the release of the Kimi K3 model by Chinese startup Moonshot AI, which prompted US officials to weigh restrictions on American businesses' use of Chinese open-weight models, first reported by Axios on July 20. Reporting attributes growing Washington scrutiny in part to concerns that Chinese developers are leveraging large-scale distillation of American models to bypass hardware limits. Amodei argued that Beijing remains the most technologically capable authoritarian state and that the primary concern is frontier models used for military superiority or domestic surveillance. He also pointed to findings from the UK's AI Security Institute on biological-risk thresholds as evidence that capability-based pre-release testing should replace assumptions about open versus closed models.
Open questions about testing thresholds
Coverage notes that Amodei's proposed testing regime leaves unresolved who sets the capability threshold, who runs the evaluations, and how "sufficiently capable" would be measured, whether by benchmark scores, training compute, or risk-category rubrics. Amodei conceded any global testing regime "would need to be global, which means even the CCP would need to be on board," a prospect described as the largest unanswered question around the proposal. Export controls and anti-distillation measures are policies Washington can impose unilaterally, but a binding international testing framework does not yet exist. On Monday, Nvidia separately launched the Open Secure AI Alliance with more than 35 founding members including Microsoft, IBM, Hugging Face, Palantir, Red Hat, CrowdStrike and the Linux Foundation; Anthropic was not listed among them.
Follow-up signals: US policymakers are expected to weigh next steps on potential restrictions affecting Chinese open-weight models, while industry attention turns to any further signatories to the Nvidia letter and to whether Anthropic or Amazon publicly articulate their own positions on capability thresholds and testing regimes.
Share this article
