"Model Distillation": How Did China Develop AI at a Lower Cost Than American Models?
A Reuters investigation reveals that Chinese companies and research institutions are expanding the use of "model distillation" technology to develop competitive local AI systems at lower cost, relying on outputs from advanced American models, amid tightening US restrictions on China's access to advanced chips.
A Reuters investigation has revealed that Chinese companies and research institutions are increasingly relying on outputs from advanced American AI models to develop competitive local systems, using increasingly sophisticated methods aimed at reducing training costs and minimizing the need for advanced computing resources. This trend comes as China seeks to expand the use of AI in its economic and industrial sectors, coinciding with the tightening of US restrictions on Chinese companies' access to advanced electronic chips. The Reuters investigation was based on a review of more than 300 Chinese documents, including research papers, patents, corporate disclosures, job advertisements, and government documents, showing the transition of a technique known as "model distillation" from academic experiments to industrial and commercial use. This technique leverages outputs from advanced AI models to train other models, often smaller and cheaper to operate. According to the documents reviewed by the agency, Chinese researchers use advanced American models to generate answers to various questions, write computer programs, and find solutions to complex problems, before converting these outputs into training materials for local models. Chinese companies are also developing software to automate these processes.
"Model Distillation": How Did China Develop AI at a Lower Cost Than American Models?