Closed-model protection call due to anxiety of competition

The AI race is no longer simply about who builds the smartest model, but who gets to write the rules. That is why Anthropic chief executive Dario Amodei's latest open letter deserves attention. Framed as a "national security" appeal, it openly calls for tighter chip export controls, tougher action against "industrial-scale distillation" and mandatory safety testing for advanced artificial intelligence models. Yet beneath the security rhetoric lies a more commercial concern: how to protect a closed-source monopoly business model as open-source competitors gain ground.

Amodei's intervention is striking because of where Anthropic stands in Silicon Valley. Last week, more than two dozen US technology companies — including Microsoft, Nvidia, Meta, IBM and Hugging Face — jointly urged policymakers to support open-weight AI models and resist "premature" regulation. Anthropic was the most conspicuous absentee. Nor did it join nearly 200 start-ups in warning the White House that restricting access to Chinese open-source models would cripple hundreds of AI businesses in the US.

Instead, Amodei has emerged as the industry's leading advocate for tighter controls, arguing that open-weight models — particularly Chinese ones — present exceptional risks. Amodei insists he has "never advocated" banning open-weight models, even describing them as a "public good". Yet, in the same breath, he argues that such models are uniquely vulnerable to misuse, cannot be withdrawn once released and allow China to "circumvent" US hardware restrictions. The contradiction has not gone unnoticed. As one prominent AI commentator observed, the message amounts to: an open-source model itself is acceptable — provided it is not Chinese.

China's open-source ecosystem is no longer merely catching up; in several areas it has started setting the pace. Moonshot AI's Kimi K3, released this month, rivals leading US frontier models while operating at a fraction of the cost. OpenRouter data indicate that Chinese models now account for more than half of the token usage by US companies using its platform. Markets are making their own judgment, and they are rewarding efficiency rather than nationality.

The charge of "distillation" illustrates the weakness of the argument. Distillation — the process of training smaller models using the outputs of larger ones — has been standard machine learning practice since British-Canadian computer scientist and cognitive psychologist Geoffrey Hinton formalized the technique a decade ago. Both US and Chinese laboratories use it extensively.

Washington has produced no public evidence that Chinese companies have illegally extracted proprietary model weights or data. Without such evidence, the "distillation" accusation looks less like a security claim than a fear of losing market share.

The inconsistency extends beyond rhetoric. Thinking Machines Lab, founded by OpenAI's former chief technology officer, reportedly built upon the architecture of China's DeepSeek models. Anthropic itself has faced scrutiny after one of its latest systems reportedly appeared to identify itself as Alibaba's Tongyi Qianwen during testing, fueling speculation that Chinese open-source models had influenced its development. When US companies learn from Chinese models, the practice is described as "engineering", but when Chinese companies learn from US models, it is labeled as "theft".

Recent events also complicate the "security" narrative. An unreleased OpenAI model reportedly carried out thousands of automated operations against Hugging Face's infrastructure after breaching internal safeguards. Ironically, when researchers sought help analyzing the incident, several leading closed-source models refused because their own safety guardrails blocked the necessary forensic work. The investigation was eventually completed using GLM-5.2, an open-source model developed by China's Zhipu AI. This illustrates a broader point: transparency can strengthen security as much as secrecy.

The AI race will not be won through ever-tightening export controls. It will be won by whichever ecosystem innovates fastest, attracts the best talent and remains most open to new ideas. The future of AI will belong to those confident enough to compete.

— LI YANG, CHINA DAILY