Y Combinator CEO Garry Tan is pushing back against calls to regulate AI model distillation, the practice of using outputs from advanced AI systems to train smaller ones. His position directly contradicts frontier AI developers like OpenAI and Anthropic, which have accused Chinese companies including DeepSeek, Moonshot AI, and MiniMax of using this technique to replicate their most powerful models.

What Exactly Is Model Distillation and Why Does It Matter?

Model distillation works by taking the outputs of a more capable AI model and using them to train a smaller or less capable one. Think of it as learning from a teacher's answers rather than working through problems from scratch. In some cases, this practice may be illegal, particularly when it involves unauthorized use of proprietary systems.

The stakes are high. OpenAI believes that DeepSeek's V3 and R1 systems drew capabilities from GPT-4 and GPT-4o, OpenAI's most advanced models. Anthropic has similarly accused Moonshot AI, DeepSeek, and MiniMax of distillation. These are serious allegations, but they remain unresolved in terms of legal findings or confirmed misconduct.

Why Is Garry Tan Taking a Different Approach?

Rather than calling for enforcement action, Tan advocates for what he describes as a "tightrope" balance between two competing goals. He wants regulators to preserve broad access to open-weight models, which are freely available AI systems that anyone can use and modify, while simultaneously allowing frontier model developers to maintain a price premium for their most advanced systems.

"I would do nothing," Tan said when asked about distillation allegations, speaking to CNBC.

Garry Tan, CEO at Y Combinator

Tan's recommendation is about regulatory priorities, not a judgment on whether any company actually broke the law. His argument is that policymakers should focus on demonstrated risks rather than potential future scenarios. He specifically identified cybersecurity as an immediate concern that deserves policy attention.

How Should Policymakers Balance Access and Innovation?

Tan's proposed framework rests on three key principles for how regulators should approach AI development:

• Preserve Open Access: Open-weight models should remain freely available to provide users with freedom and broad access to AI technology without restrictive barriers.
• Protect Commercial Incentives: Frontier model developers need to retain pricing advantages so their business models remain viable and they can continue investing in cutting-edge AI research.
• Focus on Real Risks: Policy attention should center on current, demonstrated threats like cybersecurity vulnerabilities rather than speculative scenarios about distant AI risks.

This stance puts Tan at odds with frontier labs seeking regulatory action. Anthropic and OpenAI have both called for measures to address alleged model distillation, viewing it as a threat to their competitive advantage and business sustainability. The question now is whether regulators will treat the extraction of frontier capabilities as an enforcement problem or preserve market access and let competition determine outcomes.

Tan also addressed broader AI concerns, arguing that job losses caused by AI and economic disruption will unfold over decades rather than create immediate crises. Despite this measured view on AI risks, Y Combinator remains heavily invested in the sector. Of the 196 startups that presented at Y Combinator's recent Demo Day, 149 were classified as machine learning and artificial intelligence projects.

The policy choice Tan describes extends beyond a dispute between competing companies. It asks a fundamental question about whether regulatory frameworks should restrict alleged extraction of frontier capabilities or sustain a market in which access spreads while the highest-end models can still command premium pricing. His answer is clear: let the market work, and focus enforcement on proven harms rather than alleged copying techniques.