The Case for Open-Weight AI: Why Distillation Should Be a Right, Not a Risk

When Innovation Meets Access

As artificial intelligence continues its rapid ascent through the business landscape, a critical debate is emerging around one fundamental question: who should have access to the most advanced AI models? This is not merely an academic discussion—it's about the future of technological innovation and the democratization of intelligence itself.

I've been following this issue closely, particularly after hearing from Y Combinator CEO Garry Tan, who recently made headlines with his views on what he calls 'American distillation regime.' In a recent interview with CNBC, Tan stated categorically: 'I would do nothing,' regarding government intervention in the practice of distillation. His argument is compelling and deserves deeper scrutiny.

"We could argue that there should be an American distillation regime."

Garry Tan, Y Combinator CEO

The Distillation Debate

Distillation in AI refers to the process where one model learns from another by extensively prompting and analyzing its responses. It's a legitimate technique used by researchers and developers to understand how models work, and to train new ones.

This practice is not without controversy. Anthropic recently published a report detailing what they term 'illicit distillation attacks'—Chinese AI labs allegedly engaging in unauthorized distillation using stolen credentials and fraud. This has prompted calls from some quarters for strict regulation.

But Tan's perspective offers a different lens entirely. He doesn't advocate for the use of stolen credentials or any form of illegal activity. Rather, he believes that open-weight AI labs should be free to use the same training techniques on American frontier models—just as Chinese labs are doing with U.S. models.

A Matter of Access and Equity

At its core, Tan's argument is about access and equity in AI development. He sees it not as a threat to innovation but as a pathway for broader participation in the AI ecosystem. This idea resonates with me because it aligns with a fundamental truth: technology should empower people, not restrict them.

Consider this analogy: when public libraries make books available for checkout, they're enabling access to knowledge. Similarly, if we want AI to be a public good rather than a proprietary asset, then open-weight models that can be distilled and built upon represent the logical next step.

Proprietary labs have historically done their part in training frontier models by ingesting vast quantities of public human knowledge—often without explicit permission from intellectual property holders. The question is whether they should now have the right to restrict how others use that same knowledge through distillation techniques.

The Public Good Argument

Tan's stance is grounded in a belief that access to intelligence trained on broadly accessible data should be treated as a public good, not something locked behind restrictive terms of service. This is not just about fairness—it's about the long-term health of AI innovation.

He points out that when we look at the landscape of AI development, there's already a growing tension between the closed models of large tech companies and open-weight alternatives. While proprietary models offer powerful capabilities, they also create barriers to entry for smaller players and startups.

"The nightmare scenario, the doomer scenario for AI is that there's just one company," Tan noted. "It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly there's one company that's monolithic. And that would be bad."

Global Implications

The implications of this debate extend far beyond national borders. As we've seen with Chinese AI labs, the global race for AI supremacy is not just about who has the most powerful models—it's also about how those models are distributed and used.

If American companies want to compete globally while maintaining their edge in innovation, they must ensure that their own open-weight labs have the tools and freedom to innovate. The current approach of strict restrictions could leave U.S. startups at a significant disadvantage.

The Regulatory Challenge

Regulation in this space presents a unique challenge because it must balance protection for intellectual property with the need to foster innovation. Tan's view suggests that instead of creating more barriers, we should focus on creating frameworks that encourage collaboration and open research.

This isn't about ignoring the legitimate concerns around unauthorized access or theft. It's about recognizing that the current system may be too restrictive for those who are trying to build a better future through open-source AI development.

Building the Future Together

Looking ahead, we need to think not just about what is technically possible, but what is ethically and strategically wise. The AI landscape is evolving rapidly, and we must be thoughtful about how we shape it.

As investors and technologists, we have a responsibility to ensure that the future of AI is inclusive, open, and fair. Tan's vision of an 'American distillation regime' may be controversial—but it highlights an important truth: the most innovative solutions often emerge from openness, not isolation.

The question isn't whether we should regulate AI development, but rather how to regulate it in a way that promotes progress rather than stifles it. This requires nuanced thinking, and I believe Garry Tan's perspective offers a valuable contribution to that conversation.

Conclusion

Distillation isn't about theft—it's about access. It's about ensuring that the next generation of AI development is not confined to a select few but is instead open to everyone who wants to contribute to humanity's technological advancement. In an age where AI can reshape industries and transform lives, we must ask ourselves: do we want to build walls or bridges?

The path forward should embrace the principles of openness, innovation, and collaboration. That's not just good business—it's good for society.