Garry Tan on AI Model Distillation and the Case for Open-Weight Ecosystems

Y Combinator CEO Garry Tan has called for a regulatory environment that permits U.S. open-weight AI labs to "distill" frontier models. This position stands in direct opposition to frontier labs like Anthropic, which have characterized such practices as "illicit distillation attacks" and urged regulators to crack down on the practice.

The Case for an "American Distillation Regime"

Tan argues that allowing smaller American labs to distill knowledge from frontier models is essential for maintaining a competitive and robust AI ecosystem. His primary objective is to prevent a "nightmare scenario" where a single, monolithic proprietary provider controls the most advanced AI capabilities due to superior access to capital and research talent.

Tan's argument rests on two primary pillars:

Public Good vs. Proprietary Control: Tan contends that because frontier models were trained on broad, public-access data—often without the explicit permission of intellectual property holders—the resulting intelligence should be treated more as a public good than as a proprietary asset locked behind restrictive terms of service.
Strategic Competition: By enabling U.S. open-weight labs to use distillation techniques, the U.S. can develop a more diverse set of high-performing open-weight options, reducing reliance on non-U.S. (specifically Chinese) open-weight models.

Distillation Defined

In the context of AI training, distillation is the process where a model maker uses extensive prompting of a larger, more capable "teacher" model to train a smaller, more efficient "student" model. While this is a a standard technique used legitimately by many AI labs, the conflict arises when this is done in violation of a provider's Terms of Service (ToS) or through fraudulent means.

Industry Conflict: Frontier Labs vs. Open-Weight Advocates

Anthropic recently released a threat intelligence report alleging that Chinese labs are engaging in "illicit distillation attacks" using stolen credentials and fraud to bypass API restrictions. Anthropic CEO Dario Amodei has publicly called for regulatory intervention to stop these practices.

In contrast, Tan views the attempt by frontier labs to dictate what customers do with API outputs as an overreach. He believes that once a user has paid for an API call, they should be free to use that information as they wish.

Community Perspectives and Counterpoints

Discussion among technical practitioners and observers highlights several key tensions regarding the legality and ethics of distillation:

The "Fair Use" and Moral Argument

Many argue that distillation is a form of "transformative work" and that frontier labs have no moral high ground to claim IP theft because they "strip-mined the commons" to build their models.

"The frontier models are based on training data from tons of copyrighted work... The labs have no moral or ethical ownership to the end result, and others should feel free to treat any company-imposed restrictions on their use as invalid."

Economic Sustainability

Some critics question the economic viability of a world where distillation is normalized. They argue that if the value of billions of dollars in research and data acquisition is immediately cloned into open models, the incentive for frontier labs to continue investing in new data creation will vanish.

Trust and Transparency

Concerns have been raised that "open weights" are not the same as "open source." Distilling a black-box proprietary model may simply bake in the undisclosed biases, ideological leanings, or even hidden vulnerabilities of the original provider into the open-weight ecosystem.

Technical Efficiency

Some practitioners suggest that simple distillation is a brute-force approach with diminishing returns. They argue that a more effective method involves using "teacher" models to generate targeted, high-quality synthetic datasets specifically tailored for the student model, rather than simple prompt-response distillation.

Summary of the Strategic Divide

Perspective	Goal	View on Distillation
Frontier Labs	Protect IP and R&D investment	Illicit attack / Regulatory capture
Open-Weight Labs	Democratize access and efficiency	Essential tool for innovation
YC / Garry Tan	Prevent monopoly	Public good / Strategic necessity