Skip to content
ainet·Indexed by this site · Originally published· AI updates →· Its place in the scan record →

Reuters: China is increasingly using U.S. AI models to train its own models through distillation

[中, 美 칩 규제 뚫으려 美 AI 모델

AI summary

According to Reuters' review of more than 300 Chinese academic papers, patents, publicly available corporate materials, job postings and government documents, as the United States tightens controls on advanced chips, China is increasingly using AI models developed by U.S. companies to train its own AI systems through distillation, and this practice has expanded from academic research to mainstream industrial applications. Distillation refers to training a smaller student model on answers generated by a powerful teacher model; a paper published in November 2024 by a Shanghai Jiao Tong University team said it selected about 5,000 answers from OpenAI o1, translated them via GPT-4o mini, and produced 10,750 bilingual training samples for further training of Alibaba Qwen2.5-72B. U.S. authorities have warned that collecting U.S. model outputs on an industrial scale may constitute intellectual property theft; a U.S. Commerce Department official expressed concerns to Chinese regulators at a meeting in Chengdu in July, and the Chinese side defended distillation as a standard AI development technique but made no specific commitments; OpenAI's terms restrict using its outputs to develop competing models, and Anthropic also restricts training competing systems without permission.

Read originalainet.link