Y Combinator CEO Calls for U.S. Open Distillation Regime

Garry Tan, Chief Executive Officer of Y Combinator, has called for freedom for U.S. open-weight AI developers to distill. He argues that U.S. companies should also actively engage in learning from the outputs of frontier models. According to reporting by Julie Bort of TechCrunch AI, his remarks to CNBC and additional explanations provided to TechCrunch were conveyed.

He expressed opposition to intervention by regulators. On the controversy over distillation by Chinese developers, he calls for tolerance rather than tougher regulation. The aim is to preserve a robust, non-Chinese open-weight option in the United States.

Overview of the U.S.

Distillation Regime Proposal and Background of the Remarks

In an interview with CNBC earlier this week, Tan was asked about regulation and opposed intervention. He made remarks to the effect that an American distillation regime could be debated. In his explanation to TechCrunch, he said it refers to a system in which small U.S. open-weight developers learn from the models of U.S. frontier developers.

I would do nothing. We could argue that there should be an American distillation regime.

He envisions using the same kind of learning technique within the United States. It is a method of widely querying frontier developers’ inference results to learn their behavior and reasoning processes. The idea is to channel the results into expanding U.S. open weights.

Distillation refers to a technique of obtaining large volumes of responses from other models to aid training. It is widely used in development practice in training new models. The boundary between legitimate use and illicit acquisition is the current point of contention.

Proposal to Strengthen Open Weights

Tan’s argument centers on thickening the layer of U.S. open weights. He intends to secure options that do not depend on Chinese open models. The aim is to keep a foundation in the country that entrepreneurs can use freely.

As head of Y Combinator, he emphasizes the development environment for founders. Access to high-performance foundation models has become the basis for new businesses. If terms of use become stricter, small developers will have less freedom.

He positions himself as an enthusiastic AI user. He has in the past described his own usage habits in self-deprecating terms. Drawing on that experience, he calls for both advancing the frontier and spreading open models.

They are at the frontier and driving it forward. We want that to be fundable, and be a great business model ongoing.

He does not deny the role of frontier developers. He says businesses that pioneer the cutting edge should remain fundable. At the same time, he says open models should give people freedom and means of access.

Differences with Anthropic’s

Illicit-Distillation Report

Anthropic this week published its second report on illicit distillation by Chinese developers. It alleges they distilled without authorization while concealing their identities, relying on fraud and stolen credentials. The company’s CEO, Dario Amodei, had previously called for a crackdown by U.S. regulators.

Tan does not share this pro-regulation argument. Attention has focused on the fact that the head of a prominent startup incubator has objected. His position is that the response should be to nurture open models, not to regulate.

The dividing line between the two is how far use of API outputs should be permitted. Anthropic takes issue with unauthorized acquisition and misuse of credentials. Tan opposes restricting even learning via legitimate access.

He does not defend distillation using stolen credentials. He assumes straightforward access as a legitimate user. His distinction is that illicit intrusion and legitimate use should be told apart.

Argument Contrasting the History of

Training-Data Collection

Tan’s argument rests on two pillars. First, he argues it goes too far for developers to excessively restrict what customers do. Second, he argues similar issues existed in the training process on the closed-model side.

Closed models were trained by collecting human knowledge at massive scale. There is a history of ingesting copyrighted materials without rights holders’ permission. Citing that, he argues it is unbalanced to strictly restrict only the use of outputs.

He told TechCrunch he was concerned about restrictions on the use of API call results. He takes the view that access to intelligence trained on broadly public data should have a character close to a public good. He said it should not be locked away behind restrictive terms of service.

Controlling what users and customers do with API calls to closed weight models feels constraining, 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

He also mentioned a role for government. He said there is room to ease lock-in through terms of service and normalize open access. It is an idea of shaping the terms of access to intelligence from a public perspective.

This argument also overlaps with disputes over rights clearance for training data. Lawsuits and settlements continue between publishers and rights holders and AI developers. The question of who decides the terms for using outputs has resurfaced.

Wariness of Single-Company Dominance and

Logic of Seeking Balance

What Tan warns against is concentration of power in a single powerful company. He called a scenario in which one company monopolizing capital and research talent runs away with the lead a disastrous script. He rejects a future in which massive frontier capabilities are held by a single monopoly provider.

The nightmare scenario, the doomer scenario for AI is that there’s just one company. It has the best access to capital. It has the best AI researchers.

He seeks a balance between frontier development and open models. The idea is to maintain cutting-edge momentum while preserving user choice. Reconciling a fundable business model with openness will be the challenge.

Also in the background are U.S. authorities’ findings of systematic distillation by Chinese AI companies. Model exfiltration is being debated from national security and technology-competition perspectives. Amid that, the pro-distillation argument has emerged as an alternative of nurturing domestic open models.

The future focus will be alignment of terms of service and legal frameworks. Developers are moving to strengthen fraud detection and access management. Meanwhile, the startup-support side seeks rules that protect freedom of legitimate use.

Editorial Opinion

In the short term, the clash over interpretation of API terms of use is likely to intensify. Frontier developers may strengthen detection and blocking of misuse. The open-weight camp is likely to advocate distillation within the scope of legitimate use, and the policy debate looks set to continue.

In the long term, the framework of whether to treat access to intelligence as a public good or as private property will be questioned. Institutional design that reconciles recouping development funding with ensuring openness will be needed. Mechanisms to avoid concentration in a single company are likely to become the focus of competition policy.

The remaining issue is whether legitimate distillation and illicit acquisition can be technically distinguished. The criteria for how far use of API outputs is permitted remain unclear. The question is whether the government can present neutral guidelines.

References

• “Y Combinator’s Garry Tan wants US open-weight AI labs to ‘distill’ frontier models, too”, by Julie Bort — TechCrunch AI, 2026-09-11T20:59:47.000Z (ARR)
• Source URL: https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/