Garry Tan of Y Combinator Advocates for US Open-Weight AI Labs to Refine Advanced Models
In discussions about Chinese AI labs and their distillation methods for acquiring knowledge from leading model creators, Y Combinator CEO Garry Tan is advocating for minimal regulatory interference. He suggests that U.S. AI labs should adopt similar methods.
“I would do nothing,” Tan stated in a recent CNBC interview. “We could debate the need for an American distillation framework.”
In a follow-up with TechCrunch, he clarified that he supports smaller American open-weight AI labs employing equivalent training strategies on leading U.S. AI labs, aimed at creating a stronger array of open-weight alternatives independent of Chinese influence.
Distillation involves training one model using prompts from another model to understand its reasoning processes. This method is officially accepted among AI labs for training new systems.
Recently, Anthropic released a report claiming that Chinese labs are participating in “illicit distillation attacks,” masquerading their identities to engage in unauthorized distillation, often using deceit and stolen credentials. Anthropic CEO Dario Amodei has previously called for stricter regulations concerning distillation practices.
It is significant that the head of Silicon Valley’s renowned startup accelerator holds a different view.
Tan is not endorsing the use of stolen credentials for distillation. He advocates for transparency and access, arguing that AI labs should not overly restrict what their customers can do with the insights derived from their models.
He points out that proprietary AI labs didn’t seek permission while utilizing vast amounts of human knowledge to develop their models, often consuming copyrighted content without clearance from the respective copyright holders.
“Limiting user actions based on API interactions with closed-weight models feels restrictive; the government can help recognize that access to models trained on publicly available data should be a public good rather than confined by strict service guidelines,” he explained to TechCrunch regarding the need for American labs to have the right to distill as well.
Tan, an enthusiastic AI user who once referred to himself as experiencing cyber psychosis, emphasizes the need for a balance between open-weight AI labs and advanced frontier labs.
“They are leading the charge and pushing the boundaries of the field. We want that to remain viable and profitable,” he told CNBC. “Open-weight models should empower users and grant them access.”
For him, the real risk in the AI landscape is if all the significant capabilities of frontier AI get concentrated in a single dominant, proprietary firm. “The worst-case scenario for AI is when there’s just one company that controls everything,” he remarked. “It could have the most funding and the best researchers, taking the lead alone and creating a monolithic system. That would not be ideal.”
