Y Combinator CEO Urges US Labs to Use Distillation Techniques

Garry Tan, the chief executive of Y Combinator, is urging U.S. regulators to allow American open-weight AI labs to use distillation techniques against major frontier models. In a recent interview with CNBC, Tan stated that he would do nothing to stop this practice and suggested that the United States should establish its own rules for such knowledge transfer. His position stands in direct contrast to recent moves by large proprietary AI companies, who have called for government crackdowns on what they describe as illicit extraction of model capabilities.

The controversy centers on how smaller laboratories can leverage the reasoning skills of larger, closed systems. Tan believes that preventing this flow of information stifles innovation and creates an unhealthy imbalance in the ecosystem. He argues that if the goal is to foster a competitive and diverse AI landscape, then the tools used by smaller labs to catch up must remain accessible. This debate highlights a growing tension between protecting proprietary intellectual property and ensuring that foundational AI capabilities do not become the exclusive domain of a few giants.

Distillation defined as a learning method

For readers unfamiliar with the term, distillation is a standard technique in machine learning where a smaller model is trained to mimic the outputs of a larger, more complex model. It is akin to a student learning from a professor by observing their work and reasoning patterns. While often used legitimately within the industry to improve efficiency, the practice has become contentious when applied across company boundaries. Critics argue that using a competitor’s API to train a rival model amounts to stealing trade secrets, while proponents view it as a natural way to disseminate knowledge.

According to reporting from GN technics/ai (en-US), Tan emphasizes that this method allows open-weight labs to create robust alternatives to closed systems. By learning from the best available models, these smaller labs can offer the public more freedom and access to AI capabilities. Tan contends that restricting this process effectively locks intelligence behind restrictive terms of service, which he believes goes against the spirit of open innovation in the technology sector.

Argument against regulatory overreach

Tan’s stance is not an endorsement of illegal activity. He explicitly states that he does not support the use of stolen credentials or fraudulent access to distill models. Instead, he argues that companies should be free to use publicly available APIs and outputs to learn. He points out a perceived double standard in the industry, noting that frontier labs trained their models on vast amounts of public and copyrighted data without seeking permission from the original creators. Tan suggests that the same freedom should apply when those models are used to train new, open systems.

He describes the current situation as an overreach by proprietary labs, who are attempting to dictate what customers can do with the information their models generate. Tan believes that access to intelligence trained on broad public data should be treated as a public good rather than a locked-down asset. By allowing this flow of knowledge, the government can help normalize a more balanced ecosystem where innovation is not stifled by excessive legal barriers.

Avoiding a single provider monopoly

At the heart of Tan’s argument is a fear of consolidation. He describes the worst-case scenario for the industry as a world where a single monolithic company controls all advanced AI capabilities. Such a provider would have the best capital, the best researchers, and the ability to run away with the technology, leaving no alternatives for users or developers. Tan argues that this outcome is dangerous for both the tech sector and the broader economy.

To prevent this, he advocates for a balanced approach where frontier labs are funded and successful, but open-weight models remain viable and accessible. This duality ensures that there is competition and that the benefits of AI are distributed more widely. By permitting distillation, the U.S. can maintain a robust set of non-Chinese open-weight options, preventing a single entity from holding a monopoly on the future of artificial intelligence.