Twenty-five major U.S. tech companies issued a statement opposing regulations on AI model "distillation" technology, but leaders like OpenAI and Anthropic did not sign. As the U.S. AI industry divides over openness and regulation in response to China's distillation push, South Korea finds itself unable to align with either side due to a growing atmosphere of self-censorship regarding the adoption of distillation.
According to the IT industry on July 26, companies including Nvidia, Microsoft, Meta, Palantir, Dell Technologies, IBM, Hugging Face, Mistral, and Andreessen Horowitz released a joint letter on July 24, stating that "hasty restrictions on open-weight AI models could stifle competition or drive innovation overseas."
Distillation is a technique that trains smaller, lighter models using the output of larger models. The companies described distillation as "a widely used technique that plays a crucial role in AI innovation," arguing that concerns about illegal distillation should be addressed through targeted legal and commercial frameworks rather than comprehensive regulations. Nvidia CEO Jensen Huang and Microsoft CEO Satya Nadella also shared the letter on their personal social media accounts to lend support.
The letter makes three main points: 1) open-weight models are essential for maintaining U.S. AI leadership; 2) they are safer than closed models because they can be verified by multiple researchers; and 3) distillation has been a standard practice since the era of open-source software.
Box CEO Aaron Levie, a signatory, stated that "the more innovation there is, whether in the U.S. or China, the more affordable and efficient AI will become." In contrast, OpenAI, Anthropic, Google, and xAI did not sign the letter, highlighting a divide between leading companies focused on closed models and those with vested interests in the open-weight ecosystem. Notably, OpenAI and Anthropic have prohibited the distillation of their models in their service agreements.
The backdrop to this statement is the shock from China's Moonshot AI's Kimi K3. After Kimi K3 demonstrated performance comparable to top U.S. models when unveiled on July 16, Michael Kratsios, head of the White House Office of Science and Technology Policy, claimed on July 22 that Moonshot AI had distilled Anthropic's Claude model to develop Kimi K3.
Kratsios pointed out that Moonshot AI has built a sophisticated internal platform to evade detection by switching access routes and raised suspicions that it obtained banned Nvidia GB300 servers via Thailand. U.S. Treasury Secretary Scott Vessen also suggested that sanctions could be imposed if illegal distillation is confirmed.
This is not the first controversy surrounding distillation from China. On February 24, Anthropic revealed that three Chinese AI companies—DeepSeek, Moonshot AI, and MiniMax—used approximately 24,000 fake accounts to collect over 16 million outputs from Claude. MiniMax was responsible for about 13 million of those outputs, and there were indications that the new Claude model was altering its response patterns within 24 hours of its release. The U.S. industry views this as evidence that China is leveraging distillation at a national level to circumvent its overwhelming capital gap.
In South Korea, the atmosphere is markedly different. Following controversies over theft and derivation in the Ministry of Science and ICT's independent AI foundation model (Dopamo) project, there is a growing trend of self-censorship regarding the adoption of distillation.
On January 15, the Ministry announced that the Naver Cloud and NC AI consortium was eliminated in the first phase evaluation of Dopamo. A key factor was the high similarity between the vision encoder weights of Naver Cloud's "HyperCLOVA X Seed 32B Sync" model and Alibaba's "Qwen 2.4" series. Upsstage also faced allegations that parts of its "Solar Open 100B" model were similar to China's Zhipu AI's "GLM-4.5-Air," although the accuser later acknowledged a lack of evidence and apologized three days later. Both cases are more about allegations of theft and derivation due to weight similarity rather than distillation itself, but they have contributed to a growing reluctance to adopt distillation domestically, according to industry insiders.
The issue is that South Korea's AI development capital significantly lags behind that of the U.S. and China. Even combined government budgets and corporate investments fall short of the scale of individual investments by global tech giants, making efficiency through distillation necessary. However, within the competitive landscape among domestic companies, prioritizing relative advantage over objective performance complicates the introduction of distillation, which could lead to reporting among competitors.
Unlike the U.S., which invests astronomical capital, and China, which fosters corporate collaboration through state leadership, competition among domestic companies is reportedly dragging down the level of AI models.
Meanwhile, the second phase evaluation of Dopamo is set to begin in early August, with four consortiums including SK Telecom, LG AI Research, Upsstage, and Motif Technology participating.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.