Y Combinator's Tan Advocates for Open Distillation Access Across US AI Labs

A Public Stance Against Restrictions

Garry Tan, chief executive of Y Combinator, has taken a position that sets him apart from several prominent voices in American artificial intelligence development. While some frontier model makers have called for regulatory intervention to prevent unauthorised distillation of their systems, Tan argues the opposite: smaller US labs working on open-weight models should have broad freedom to learn from proprietary systems through the same techniques.

In remarks this week, Tan framed his view around what he termed an "American distillation regime". The concept centres on allowing domestic open-weight developers to train their models by extensively querying closed systems operated by frontier labs, a process that extracts knowledge about reasoning patterns and capabilities. Tan is explicit that he does not endorse fraud or credential theft, but he wants the front door open for legitimate use.

The Y Combinator leader's comments arrive as Anthropic released its second report alleging that Chinese laboratories have engaged in what it labels "illicit distillation attacks", using concealed identities and stolen credentials to train models without authorisation. Dario Amodei, Anthropic's chief executive, has publicly urged US regulators to impose restrictions on distillation activity.

Tan's disagreement is notable precisely because Y Combinator sits at the centre of Silicon Valley's startup ecosystem, funding and advising hundreds of early-stage companies each year. His willingness to diverge from the frontier labs' preferred policy direction signals a tension within the American AI industry over how openly knowledge should flow.

Two Arguments for Open Access

Tan's reasoning rests on a pair of foundations. The first is a question of customer rights. He contends that once a user or business pays to access a model through an API, the information returned in those interactions should not be subject to heavy-handed usage restrictions. Attempting to control what customers do with outputs, in his view, represents overreach by the labs.

The second pillar is a matter of consistency. Frontier model developers trained their systems by ingesting vast quantities of publicly available data, including copyrighted text, images and code, often without explicit permission from rights holders. Tan draws a parallel: if those labs did not seek consent when building their models on the work of millions, why should smaller labs be required to seek permission when learning from models trained on that same corpus of human knowledge?

"Controlling what users and customers do with API calls to closed weight models feels constraining," Tan said. He argues that intelligence trained on broadly accessible public data should itself function more as a public good than as something locked behind restrictive terms of service, and that government has a role in normalising that principle.

The analogy is imperfect - frontier labs argue that their models represent original engineering and that distillation amounts to copying their proprietary work product - but it reflects a broader debate about whether AI capabilities, once developed, should be treated as intellectual property to be defended or as infrastructure to be shared.

The Concentration Risk

Underlying Tan's policy preference is a concern about market structure. He describes himself as a heavy user of AI tools and has spoken in the past about the intensity of his engagement with these systems. Yet he worries that if all advanced AI capability consolidates within a single proprietary provider, the result will be a concentration of power that stifles innovation and limits access.

"The nightmare scenario, the doomer scenario for AI is that there's just one company," Tan said. A single entity with the best access to capital, the top researchers and a runaway lead would, in his assessment, be detrimental to the ecosystem.

He is careful to note that he values the work of frontier labs. They are pushing the state of the art forward, and Tan wants that research to remain fundable and commercially viable. But he also wants a parallel track: a robust set of open-weight models that provide developers and users with freedom and choice, particularly models developed and controlled within the United States rather than by Chinese entities.

This is where his call for an American distillation regime fits. If smaller US labs can train open-weight models by learning from domestic frontier systems, the result would be a more diverse and competitive landscape, reducing dependence on any single provider and offering an alternative to open-weight models emerging from China.

Policy Implications and Industry Fractures

Tan's position places him in direct opposition to the policy preferences articulated by some of the most influential figures in AI. Amodei has been vocal in calling for regulatory action, framing unauthorised distillation as a threat to the security and economic interests of American labs. Other frontier developers have similarly tightened their terms of service and invested in detection mechanisms to identify and block distillation attempts.

The divide reflects a deeper fracture in how the US AI industry conceives of its own interests. Frontier labs, which have raised billions in capital and employ large research teams, see their models as proprietary assets that warrant legal and regulatory protection. Open-weight advocates, by contrast, argue that concentration of capability in a few hands creates systemic risk and that widespread access to capable models is essential for innovation and competition.

Regulators in Washington have so far taken a cautious approach. Export controls on advanced chips and model weights have been enacted, aimed primarily at limiting Chinese access to cutting-edge AI capabilities. But domestic policy on distillation and model access remains largely unwritten. Tan's intervention is an attempt to shape that conversation before rules are set.

At Y Combinator, where Tan oversees a portfolio that includes numerous AI startups, the stakes are tangible. Many of those companies rely on access to frontier models through APIs, and some are working on open-weight alternatives. Restrictions on distillation could limit the pathways available to them, forcing reliance on a small number of proprietary providers or on open-weight models developed outside the United States.

What Distillation Actually Entails

Distillation, in technical terms, involves querying a model extensively to understand its behaviour and using those interactions to train a smaller or different model that mimics aspects of the original's reasoning. The technique is widely used within the AI research community for legitimate purposes: creating more efficient models, adapting general-purpose systems to specific tasks, or enabling deployment in resource-constrained environments.

The controversy arises when distillation is conducted without the knowledge or consent of the model provider, particularly when it involves misrepresenting identity or circumventing access controls. Anthropic's reports detail instances of Chinese labs allegedly using networks of accounts, sometimes with stolen credentials, to conduct large-scale distillation operations on American models.

Tan does not defend that conduct. His argument is that distillation itself, conducted openly and through legitimate API access, should not be restricted. The distinction matters: he is advocating for a policy environment in which a US lab can openly purchase API access, run extensive queries and use the results to train an open-weight model, all without fear of legal or regulatory penalty.

Whether that distinction will hold in practice is uncertain. Frontier labs have an interest in preventing any form of distillation that undermines their competitive position, regardless of whether it is conducted openly or covertly. And regulators may find it difficult to draw clear lines between legitimate use and exploitation, particularly when the same techniques can serve both purposes.

The Road Ahead

Tan's call for an American distillation regime is unlikely to be the final word. The debate over model access, intellectual property and competition in AI is still taking shape, and the positions of key players remain fluid. Frontier labs will continue to press for protections, while open-weight advocates will push for access.

What is clear is that the question of who controls access to advanced AI capabilities, and under what terms, will be one of the defining policy battles of the next several years. Tan's willingness to stake out a position that diverges from some of the industry's most powerful voices reflects the high stakes involved, both for the startups in Y Combinator's portfolio and for the broader structure of the AI industry.

For now, the regulatory landscape remains open. How Washington responds to competing pressures - from frontier labs seeking protection, from open-weight advocates demanding access, and from national security officials concerned about Chinese competition - will shape the trajectory of American AI development for years to come. Tan has made his preference clear: he wants the door open, not closed.