Exclusive-China taps US AI models to train homegrown rivals as chip curbs bite
Chinese entities are employing increasingly sophisticated methods to utilize output from leading American AI models for training domestic alternatives, according to a Reuters examination of over 300 Chinese academic papers, patents, corporate disclosures, recruitment notices, and government documents. This practice, known as distillation, has evolved from academic experiments into mainstream industrial usage in China.
Chinese researchers are harnessing US models to generate responses to queries, craft computer code, and determine solutions to complex problems, which are then employed by Chinese firms to develop software automating the process and specialized personnel preparing training data for smaller local models.
Although distillation is not unlawful and is prevalent globally, it cannot replicate the full capabilities of leading models or render advanced computing power unnecessary. However, it enables developers to retain certain capabilities in smaller systems, potentially decreasing dependence on American technology. This trend could diminish the impact of US restrictions on chip sales to China, as developers can construct AI capabilities by leveraging their American counterparts' resources.
On November 18-19, US President Donald Trump and Chinese President Xi Jinping are set to meet during the Asia-Pacific Economic Cooperation summit in Shenzhen, with distillation contributing to tensions between the world's two biggest economies. In September, US authorities accused Chinese firms of industrial-scale distillation of American AI models, and during July meetings in Chengdu, US Commerce Department officials raised concerns with Chinese regulators about developers gathering content from leading American AI models, warning that some activities might constitute intellectual-property theft.
Distillation involves a powerful 'teacher' model producing answers or demonstrating task completion, which developers use to train a smaller 'student' system. Chinese developers benefit from distillation by enabling one expensive query to perform repeated work. For instance, an American model generates a response to a difficult question, allowing a Chinese team to retain valuable examples and train a local system around them.
A November 2024 paper from Shanghai Jiao Tong University researchers described selecting around 5,000 answers from OpenAI’s o1 model, translating them using GPT-4o mini, and compiling 10,750 bilingual training examples to further train Alibaba's Qwen2.5-72B model. Researchers cautioned against excessive reliance on distillation, as it could hinder original innovation and leave developers dependent on breakthroughs made by more advanced model providers.
Written by urgent.news from Investing.com's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.