Y Combinator's Garry Tan wants US open-weight AI labs to 'distill' frontier models, too

The global landscape of artificial intelligence is experiencing a significant strategic pivot, with Y Combinator CEO Garry Tan stepping forward with a bold proposal that challenges conventional thinking around AI model development and geopolitical competition. Tan suggests that rather than restricting the practice of "distillation," American open-weight AI labs should actively engage in it, using techniques to learn from and build upon existing frontier models. This initiative, he argues, is crucial for establishing a strong, independent American presence in the open-weight AI space, creating alternatives to non-American, particularly Chinese, offerings.

Tan's views come at a time of heightened scrutiny and geopolitical tension concerning AI. US federal agencies, including the NSA, FBI, and Cybersecurity and Infrastructure Security Agency (CISA), have recently issued warnings about alleged industrial-scale distillation campaigns by Chinese firms targeting leading American AI models. This backdrop makes Tan's "do nothing" approach to curbing distillation, and instead encouraging an "American distillation regime," particularly noteworthy and a point of contention within the AI community.

Understanding the Core Idea: Distillation and Open-Weight AI

To fully grasp Garry Tan's proposal, it's important to understand two key concepts: AI model distillation and open-weight AI models.

What is AI Model Distillation?

In the context of artificial intelligence, "distillation" refers to a technique where a smaller, simpler AI model, often called the "student" model, is trained to replicate the behavior and outputs of a larger, more complex "teacher" model. The student model learns from the teacher's predictions, probabilities, or intermediate representations, effectively "distilling" the knowledge of the larger model into a more compact and efficient form. This process can significantly reduce the computational resources needed to run the model, making it faster and more cost-effective. Distillation is a common research method, but when it involves using the outputs of a proprietary model without authorization, it can raise concerns about intellectual property and terms of service.

What are Open-Weight AI Models?

Open-weight AI models are those where the trained parameters—the billions of numbers that encode the model's intelligence—are made publicly available. This means users can download these "weights," run the model on their own infrastructure, inspect its inner workings, and even modify or fine-tune it for specific tasks. It's crucial to distinguish open-weight from truly open-source AI models. While open-weight provides the trained model parameters, a truly open-source model also includes the underlying code and training data, allowing for complete reproduction of the model from scratch.

The benefits of open-weight models are substantial for innovation and accessibility. They democratize access to advanced AI capabilities, allowing startups, researchers, and smaller businesses to build on sophisticated models without the immense cost and resources required to train one from scratch. This fosters competition, drives down costs, and prevents vendor lock-in, giving organizations greater control over their data and AI deployments. Furthermore, open-weight models can often be run locally on "edge" devices, reducing latency and increasing reliability for critical applications.

The Geopolitical Chessboard: US vs. China in AI

The debate around AI model distillation is not merely technical; it's deeply intertwined with geopolitical competition, particularly between the United States and China. Both nations view AI as a strategic asset for national security and economic power.

Recently, US federal agencies have issued a joint advisory accusing six China-based AI companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—of engaging in "industrial-scale distillation campaigns" against leading American frontier models such as Claude, GPT, Gemini, and Grok. Anthropic, a prominent US AI lab, specifically alleged that operators linked to Alibaba conducted a massive distillation campaign, generating over 151 million Claude exchanges between May and July 2026, purportedly to improve Alibaba's Qwen models. OpenAI also suspects that DeepSeek's V3 and R1 architectures were derived from its GPT-4 and GPT-4o models.

These allegations highlight a significant concern for American AI developers: the potential for foreign entities to leverage their costly and resource-intensive innovations without proper authorization, potentially undermining their competitive edge and national security. The availability of model weights could enable countries of concern to enhance their military and intelligence capabilities, posing risks to US national security and public safety. China, for its part, has shown a strong lead in open-source AI models and a focus on deployment-ready technologies, making it an attractive partner for emerging economies and reinforcing its push for "algorithmic sovereignty."

Garry Tan's Counter-Narrative: An "American Distillation Regime"

Against this backdrop of accusations and national security concerns, Garry Tan's perspective stands out. He publicly stated at Y Combinator's annual Demo Day that he would "do nothing" to stop Chinese labs from distilling American frontier models. Instead, he advocates for a policy that allows and even encourages American open-weight labs to engage in similar distillation practices on US frontier models.

Tan's rationale is multifaceted. He argues that it is an overreach for AI labs to dictate how customers use their models, especially when these proprietary labs themselves ingested copyrighted material to train their initial models. He believes that controlling API calls on closed-weight models feels "constraining" and that fostering an "American distillation regime" would lead to a more robust set of open-weight options for the US, ensuring they aren't solely reliant on Chinese alternatives.

For Tan, the ideal scenario involves a market equilibrium where both open-weight releases and frontier systems coexist. He contends that frontier models should maintain a price premium to sustain their business models, while open-weight models should provide freedom and access to a wider user base. This balance, he suggests, is a "tightrope" that could lead to the best possible outcome for the AI ecosystem.

Furthermore, Tan downplays "doomsday" AI fears, urging regulators to focus on immediate, tangible risks such as cybersecurity breaches and biosecurity threats rather than speculative existential scenarios. His position, coming from the head of a major startup accelerator, is particularly influential as Y Combinator funds numerous AI companies.

Y Combinator's Broader Vision for Open Models

Garry Tan's advocacy for an "American distillation regime" aligns with Y Combinator's broader belief in the transformative power and economic impact of open models. YC has been a vocal proponent of how open models are "collapsing AI costs" and reshaping the economics of AI. The accelerator sees a future where the majority of AI token usage will flow through open models, enabling cost-effective automation at scale, while frontier models continue to handle the most advanced use cases.

This vision emphasizes the importance of accessibility and the ability for a wide range of developers and businesses to innovate with AI. By supporting an environment where American labs can legitimately distill frontier models, Y Combinator aims to accelerate the development of diverse and competitive AI solutions within the US. The organization continues to heavily invest in AI startups, reflecting its commitment to this evolving technological frontier.

Industry Reactions and Future Outlook

Garry Tan's proposal creates a clear divide within the AI industry. On one side are companies like Anthropic and OpenAI, along with US federal agencies, who advocate for stronger measures against unauthorized distillation, viewing it as a national security and intellectual property threat. They argue that such practices undermine the massive investments made in developing cutting-edge frontier models.

On the other side is Tan's perspective, which prioritizes fostering a vibrant, competitive, and accessible open-weight AI ecosystem within the US. He believes that by embracing distillation through legitimate means, the US can counter foreign dominance in open models and ensure its own technological leadership.

The outcome of this debate will significantly influence future AI policy and the trajectory of AI innovation. Regulators face the challenge of balancing the need for intellectual property protection and national security with the desire to promote innovation, competition, and broad access to AI technologies. The discussion will likely shape licensing agreements, API usage terms, and potentially lead to new frameworks for how AI models can be legitimately built upon and shared. The path chosen will determine whether the US AI landscape becomes more centralized with a few powerful proprietary models or a more decentralized, open ecosystem with diverse American-made alternatives.

Frequently Asked Questions

What is Garry Tan's main proposal regarding AI model distillation?

What is Garry Tan's main proposal regarding AI model distillation?

Garry Tan, CEO of Y Combinator, proposes that US open-weight AI labs should be allowed to "distill" American frontier models. He argues against regulatory crackdowns on this practice, suggesting it's crucial for building a robust set of American open-weight AI options to compete with non-American alternatives, particularly from China.

What is AI model distillation?

AI model distillation is a technique where a smaller, simpler AI model (the "student") is trained using the outputs or knowledge of a larger, more complex AI model (the "teacher"). This process aims to create a more efficient and cost-effective model that retains much of the teacher model's capabilities.

Why is this proposal controversial?

The proposal is controversial because it clashes with the concerns of major AI labs like Anthropic and OpenAI, as well as US federal agencies, who have accused Chinese firms of "industrial-scale distillation" of American frontier models, viewing it as a threat to intellectual property and national security.

How do open-weight models differ from open-source models?

Open-weight models make their trained parameters (the "weights" that define their intelligence) publicly available, allowing users to download and run them. Open-source models, on the other hand, provide not only the trained weights but also the underlying code and training data, enabling users to reproduce the model from scratch.