随著人工智慧竞赛升温，美国科技公司对被称为「模型蒸馏」的技术日益感到担忧。蒸馏是一种透过大量查询现有先进人工智慧模型并记录其回应，进而训练出新模型的高效低成本方法。尽管在学术界很常见，但美国政策制定者和企业现在将其视为中国企业在人工智慧领域快速追赶竞争对手的主要手段。

包括 Anthropic 与 OpenAI 在内的美国实验室指控多家中国企业违反服务条款，进行工业级规模的「对抗性蒸馏」以复制其技术成果。矽谷方面已向美国政府警告，这种大规模的技术剽窃行为不仅构成严重的商业损失，更可能演变成对美国人工智慧领导地位的生存威胁。

然而，人工智慧业界对于如何应对此威胁存在分歧。一方面，政府官员与大型科技公司寻求透过资讯共享与潜在制裁来打击恶意蒸馏；另一方面，开源支持者与批评者则质疑这些美国巨头本身也是透过未经授权抓取网路资料来训练模型，认为他们只是借由国家安全的名义来阻碍市场竞争。

As the artificial intelligence arms race intensifies, US tech companies are growing increasingly alarmed by a technique known as model distillation. Distillation is a highly cost-effective method of training a new model by heavily querying an existing advanced AI and logging its responses. While common in academia, US policymakers and businesses now view it as a primary tool used by Chinese companies to quickly catch up to their rivals in the AI sector.

US labs, including Anthropic and OpenAI, have accused several Chinese companies of violating terms of service by conducting adversarial distillation on an industrial scale to copy their technological breakthroughs. Silicon Valley has warned the US government that this massive extraction of intellectual property represents not only severe commercial losses but also a potential existential threat to American leadership in AI.

However, the AI industry is divided on how to address this perceived threat. While government officials and major tech companies are seeking to combat malicious distillation through information sharing and potential sanctions, open-source advocates and critics remain skeptical. They point out that these American giants trained their own models by scraping uncompensated internet data, suggesting they might be leveraging national security concerns merely to stifle market competition.