"Did a Chinese Firm Copy ChatGPT?"... What Is 'Distillation'? [Tech Talk]
Chinese AI Growth Accelerates as 'Distillation Controversy' Intensifies
Replicating and Compressing Large Models into Smaller Ones
"U.S. Computers Powering China's AI": Growing Discontent
The open-source large language model (LLM) “Kimi K3” developed by the Chinese AI company Moonshot AI has sent shockwaves through the industry by achieving performance similar to that of major AI research institutions such as OpenAI and Anthropic, but at a far lower cost. This has been likened to the “Deepseek shock” of the past.
Yang Zelin, founder of Moonshot AI, a Chinese open-source AI development company. Screenshot from Xiaohongshu
View original imageThis time, however, the U.S. response has become notably more aggressive. Not only corporations, but even government officials are raising suspicions of “distillation attacks” by Chinese AI. Distillation, which has long been used by AI researchers for model replication and compression, is now emerging as a central controversy as the U.S.-China tech rivalry intensifies.
Did Chinese AI Models Copy from American Ones?
On the 17th (local time), visitors are seen touring the booth of the Chinese AI startup Moonshot at the World Artificial Intelligence Conference (WAIC) held in Shanghai, China. Photo by Yonhap News Agency
View original imageScott Bessent, U.S. Secretary of the Treasury, claimed in a Fox Business interview on the 21st (local time), “We have discovered watermarks of American LLMs in many Chinese AI models.” Secretary Bessent emphasized that “we will closely investigate this issue in the coming days or weeks,” and that “if foreign models steal technology from our great companies, we will impose sanctions on them.”
It is not only the government taking action. Last month, Anthropic suggested to the U.S. Senate Committee on Banking that there is a possibility Alibaba in China engaged in “distillation” of Anthropic’s model. Anthropic called this “the largest distillation attack ever.”
Distillation: A Technique Used in AI Research for Decades
'Intelligent distillation' is a development technique that has been used in the AI industry for decades. It improves performance and reduces size by training a smaller model on the outputs of a large model. The term is derived from mainstream distillators. Pixabay
View original imageDistillation is a model replication and compression technique in AI. In this approach, a pre-trained large model acts as the “teacher,” and another “student” model is created by imitating the teacher through learning from its outputs. Typically, this involves collecting a large number of answers from the teacher AI to generalize intelligence. The resulting student model is much smaller than the teacher but can achieve similar performance.
AI researchers have used distillation techniques for decades to replicate and compress each other’s models, advancing technology along the way. It has evolved significantly through the work of leading AI researchers such as Jürgen Schmidhuber and Geoffrey Hinton. In the open-source community, where the sharing and disclosure of all information about models is common, distillation has become central to development.
“U.S. Computers Are Helping Chinese AI Win” – Grievances from American AI Firms
However, as the performance of open-source models has improved rapidly, a rift has begun emerging between American and Chinese companies. In fact, as early as February, both Google and OpenAI voiced suspicions that Chinese companies might have secretly distilled their models. In May, Anthropic wrote on its company blog that “large-scale distillation attacks are underway in China,” arguing that “what has allowed Chinese AI research institutes to develop models nearly on par with American ones is the large-scale theft of innovative U.S. technologies.”
The reason American AI companies are wary of distillation is clear: if cutting-edge U.S. models—trained at astronomical cost—are immediately distilled by Chinese startups, there is little incentive to remain at the forefront of innovation. Anthropic went so far as to say, “It amounts to victories for authoritarian governments being achieved thanks to U.S. computing resources.”
There are also concerns that cost-effective AI models created through distillation could shake up the business outlook for U.S. AI firms. Microsoft already announced last month that it would consider introducing Deepseek’s open-source model as a low-cost service. Should such cost-effective options gain traction in the AI market, advanced companies’ pricing dominance will likely weaken.
“The U.S. Also Replicated Distillation Research” – A More Skeptical Perspective
Views within the industry about distillation are varied. While major American AI companies strongly oppose large-scale distillation attacks, some warn that the AI industry would not have progressed without distillation.
Hot Picks Today
"Two Spend 500,000 Won in a Flash, Yet Wait Two Hours?" ... 'Dreamland' Shunned by Kids and Foreign Tourists
- The Era of Getting 200 Million Won After 30 Years Is Over... A Japanese Bank's "Perpetual Department Head" Who Just Clocked In Cries Out
- Spain Hit by Its Worst Tragedy... Lee Kangin, Moreno and Real Madrid Join Soccer World in Mourning
- They Said, “Make It Your Home for Life” - Then Suddenly Kicked 200 Passengers Off... What Happened to the Around-the-World Cruise?
- "130 Rice Bowls Sold Out in a Day"... 'Unfiltered Face' Adds 8 Million Won to Night Market Sales: China's Bungeunwoo
Jürgen Schmidhuber, who launched the first chapter in AI distillation techniques with a 1991 paper, stated through his official X (formerly Twitter) account on the 23rd, “I support the free distillation of AI by private companies. In 1991, I released distillation research for free in Europe, and this work was later replicated in the U.S. and China.” The point being that the U.S. AI industry also benefitted significantly from distillation.
© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.