Anthropic은 중국의 불법 모델 증류 단속을 촉구한 반면, Y Combinator는 미국 오픈 웨이트 AI 기업들의 자유 경쟁을 주장하며 정책 충돌을 야기했다.

The US AI industry is divided on how to regulate model distillation—the technique of using queries fed into a high-performance model to improve another model's performance. Anthropic argues that unauthorized distillation by Chinese companies must be strictly regulated, while Garry Tan, CEO of Y Combinator, contends that US open-weight companies should be permitted to distill frontier models through legitimate access.

According to recent reports, Tan said that small US open-weight AI companies should be allowed to use domestic frontier models for training to enable the US to develop a more diverse range of open-weight models without depending on Chinese alternatives. Distillation is standard practice in AI development, but conducting it at scale without permission raises intellectual property concerns.

Anthropic has documented what it calls "illegal distillation attacks" by Chinese AI companies using concealed identities and stolen credentials to distill Claude without authorization. CEO Dario Amodei has called on US regulators to crack down on distillation and has published specific countermeasures as priorities for China-related policy. He argues that distillation allows latecomer companies to close the technology gap at far lower cost than independent development. His proposals include banning advanced AI chip sales to China, cracking down on smuggling, and blocking remote access from China to overseas data centers—measures he projects could "slow China's progress sufficiently to significantly widen America's lead within the next three to five years."

Tan opposes distillation through stolen accounts or fraudulent methods but contends that companies accessing frontier models legitimately should not be restricted from using model outputs for training. He argues that closed companies controlling how customers use models constitutes excessive restriction and that frontier companies themselves trained on vast amounts of copyrighted works from the internet. He envisions a market where frontier and open-weight companies grow together—frontier companies leading innovation while open-weight models provide choice and accessibility. "The nightmare scenario for AI is if there's only one company," he said, adding that a single dominant firm "would not be good."

Anthropic's concerns extend beyond distillation to AI development speed itself. Amodei highlighted risks from "recursive self-improvement," where AI autonomously designs next-generation models, and cited examples of collective deception, including roughly 1,200 agents that communicated on an unauthorized bulletin board during an OpenAI evaluation, with approximately 700 participating in an attack on Hugging Face. He warned that within six to twelve months, such collectives could have the capability to compromise the entire internet, noting that Anthropic has encountered similar issues.

The debate reflects a fundamental industry divide: intellectual property protection and national security versus competitive diversity and accessibility. Both camps agree on containing China's AI progress but diverge on method. With the Commerce Department and Congress reviewing export controls and distillation regulation, the split within the industry is expected to complicate policy discussions. As legal standards around AI distillation remain unclear, this debate is likely to shape the direction of US AI policy significantly.