YC's Garry Tan: Let U.S. Open-Weight AI Labs Distill Frontier Models

Y Combinator CEO Garry Tan is breaking with leading AI companies over distillation—the practice of extensively prompting one AI model to learn how another works and reasons. While frontier labs call for regulatory intervention against unauthorized distillation, Tan argues the U.S. should embrace it.

"We could argue that there should be an American distillation regime," Tan told CNBC this week, according to TechCrunch. He wants smaller U.S. open-weight AI labs to use the same training techniques on American frontier models, creating a robust domestic alternative to Chinese open-weight options.

Why it matters

Tan's position represents a significant counterpoint from one of Silicon Valley's most influential startup gatekeepers. His stance challenges the emerging consensus among major AI labs that distillation should be restricted, potentially shaping how regulators approach the practice as they craft AI policy. The debate also highlights a fundamental tension: whether knowledge extracted from AI systems trained on public data should itself be treated as a public good or proprietary asset.

The distillation debate heats up

The controversy intensified after Anthropic released its second report alleging Chinese labs conduct "illicit distillation attacks," using fraud and stolen credentials to extract knowledge without permission. Anthropic CEO Dario Amodei has publicly urged U.S. regulators to crack down on the practice.

Distillation itself is a legitimate and common technique AI labs use to train new models. The dispute centers on whether it should require permission from the model being studied.

Tan draws a sharp distinction: he's not advocating for stolen credentials or fraud. Instead, he argues American labs should be free to distill openly through normal API access.

Two arguments for open distillation

Tan's reasoning rests on two pillars. First, he believes AI labs shouldn't dictate what customers do with information their models provide. "Controlling what users and customers do with API calls to closed weight models feels constraining," he told TechCrunch.

Second, he points to how proprietary labs trained their own models—by ingesting vast amounts of copyrighted material without permission from intellectual property holders. "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," Tan said.

Balancing frontier innovation and open access

Tan, who has described himself as an avid AI user, wants equilibrium between frontier labs and open-weight alternatives. He acknowledges frontier labs "are at the frontier and driving it forward" and deserve sustainable business models. But he also believes "open weight models give people freedom and access."

His true concern is concentration. "The nightmare scenario, the doomer scenario for AI is that there's just one company," he told CNBC. "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."

The details were first reported by TechCrunch.