By Titik Puspa 
In a bold stance that directly challenges prevailing concerns about intellectual property and AI development, Garry Tan, the chief executive of the prestigious Y Combinator startup accelerator, has voiced his opposition to regulatory intervention concerning the practice of "distillation" by Chinese AI labs. Tan’s perspective, articulated in recent interviews with CNBC and TechCrunch, goes further, suggesting that U.S. AI laboratories should consider adopting similar strategies to foster a more robust ecosystem of open-weight models within the United States. This viewpoint emerges against a backdrop of escalating accusations, notably from Anthropic, regarding alleged illicit distillation attacks originating from China.
Distillation, in the context of artificial intelligence, refers to a sophisticated training technique where a smaller, or "student," model learns from a larger, more advanced, or "teacher" model. This is achieved by extensively prompting the teacher model to understand its reasoning processes and knowledge base, subsequently using this acquired information to train the student model. This method is widely recognized and legitimately employed within the AI community to develop more efficient and specialized models, often at a lower computational cost than training from scratch. It allows developers to leverage the immense capabilities of large, frontier models and translate them into more accessible and practical applications.
The controversy intensified this week with the release of Anthropic’s second report detailing allegations of "illicit distillation attacks" orchestrated by Chinese laboratories. According to Anthropic, these operations involve entities attempting to distill knowledge from frontier models while deliberately obscuring their identities, often resorting to fraudulent practices and the misuse of stolen credentials. This report builds upon previous statements by Anthropic CEO Dario Amodei, who has publicly urged U.S. regulators to implement stricter controls on distillation practices. Amodei’s calls for regulatory action highlight a growing apprehension within some AI development circles regarding the potential for intellectual property theft and unfair competitive advantages.
Tan’s divergence from this prevailing sentiment is particularly noteworthy, given his leadership position at Y Combinator, an organization instrumental in nurturing countless successful technology startups. His "do nothing" approach, as stated to CNBC, is not a call for inaction in terms of ethical AI development, but rather a plea against governmental overreach in dictating how AI models are trained and how their outputs can be utilized. He elaborated to TechCrunch that his vision entails empowering smaller, American open-weight AI laboratories to employ distillation techniques on models developed by leading American AI research institutions. The ultimate goal, according to Tan, is to cultivate a stronger and more diverse landscape of American-developed open-weight AI models, thereby reducing reliance on models that might be perceived as having origins or development practices that raise concerns.
Tan’s argument rests on two primary pillars. Firstly, he contends that it constitutes an overreach for AI laboratories to attempt to dictate the permissible uses of the information their models generate when interacting with users or other systems. He believes that once an API call is made and the model responds, the subsequent use of that output should largely be at the discretion of the user, especially when dealing with models trained on broadly accessible data.
Secondly, Tan draws a parallel to the development of the very frontier models that are now the subject of these distillation concerns. He points out that these proprietary AI labs, in their formative stages, did not seek explicit permission from intellectual property holders when they amassed vast quantities of human knowledge to train their foundational models. This includes a significant volume of copyrighted material, a practice that has already led to legal challenges and settlements, such as Anthropic’s landmark $1.5 billion copyright settlement approved in July 2026. Tan suggests a degree of hypocrisy in the current calls for stringent controls on distillation when the training data for many leading models was itself acquired without explicit consent from creators.
"Controlling what users and customers do with API calls to closed weight models feels constraining," Tan articulated to TechCrunch, "and there’s a role government can play here to normalize the fact that access to intelligence that was trained on broad public access data should itself also be more a form of a public good than something locked away behind restrictive terms of service." This statement underscores his belief that information derived from publicly available data should, in principle, be more accessible and usable for the development of AI, rather than being hoarded behind proprietary barriers.
Tan’s personal engagement with AI is well-documented. He has previously described himself as an avid AI user, even to the point of experiencing what he termed "cyber psychosis" due to his deep immersion in the technology. This firsthand experience likely informs his nuanced perspective on the balance between innovation and control in the rapidly evolving AI landscape.
He envisions a symbiotic relationship between open-weight AI initiatives and the frontier labs driving the cutting edge of research. "They are at the frontier and driving it forward," Tan told CNBC. "We want that to be fundable, and be a great business model ongoing. You want open weight models to give people freedom and access." This suggests a desire for a pluralistic AI ecosystem where both large, proprietary models and more accessible, open-weight alternatives can thrive, each serving different but complementary roles.
Tan’s most significant concern appears to be the potential for a monopolistic future in AI development. He articulated a "doomer scenario" where the immense power of frontier AI becomes concentrated in the hands of a single, dominant proprietary provider. "The nightmare scenario, the doomer scenario for AI is that there’s just one company," he stated. "It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly there’s one company that’s monolithic. And that would be bad." This fear of a singular AI hegemon underscores his advocacy for fostering a diverse and competitive landscape, where open-weight models play a crucial role in democratizing access and innovation.
The implications of Tan’s stance are far-reaching. If adopted by other influential figures in the venture capital and startup communities, it could shift the discourse around AI regulation away from strict controls on information flow and towards policies that encourage broader access and innovation, even if it means embracing techniques that some deem controversial. His argument for an "American distillation regime" is essentially a call to level the playing field, allowing domestic AI developers to utilize the same powerful tools and techniques that he believes are being employed by international competitors, thereby strengthening the U.S. position in the global AI race.
The debate over distillation and the open-weight versus proprietary model dichotomy is likely to intensify as AI capabilities continue to advance. Tan’s perspective offers a counterpoint to the more cautionary narratives, emphasizing the potential benefits of open access and the risks of unchecked consolidation in this transformative field. His advocacy for a more permissive approach to distillation, coupled with his concerns about monopolistic control, paints a picture of a desired future for AI where innovation is fostered through a combination of cutting-edge research and widespread accessibility. The coming months and years will likely see significant developments in how governments and industry leaders navigate these complex issues.
