Garry Tan: US Should Distill AI Models Too, Don't Stop China!

Y Combinator CEO Garry Tan is sparking debate with his unconventional stance on AI model distillation, a practice that has recently become a hot topic in the tech world. While some, including AI firm Anthropic, are sounding alarms about "illicit distillation attacks" by Chinese labs and advocating for U.S. regulators to step in, Tan has a surprisingly different view: he suggests the U.S. should not only "do nothing" to stop it but also encourage American labs to employ similar techniques.

Distillation is a method where one AI model learns from another, typically through extensive prompting. It's a widely used and legitimate technique. However, Anthropic recently published a report alleging that certain Chinese labs are using deceptive methods, including fraudulent credentials, to distill information from advanced frontier models without authorization.

This has led to calls for stricter regulations to safeguard intellectual property in the rapidly evolving AI sector.

However, Tan, a prominent figure in Silicon Valley, believes the U.S. could benefit from adopting a similar strategy, albeit through legitimate means. He advocates for smaller, American open-weight AI labs to be free to use distillation techniques on American frontier AI labs.

His primary objective is to cultivate a more robust and diverse ecosystem of open-weight AI options within the U.S., distinct from those developed in other nations. Tan's argument rests on two main pillars.

First, he questions the premise of AI labs dictating how users can utilize information obtained from models via API calls, labeling such controls as "constraining." Second, he points out that proprietary AI labs themselves amassed vast quantities of human knowledge, including copyrighted material, without explicit permission from intellectual property holders to train their foundational models. He proposes that government intervention should aim to normalize legitimate distillation practices rather than suppress them.

He emphasizes that he is not endorsing illegal distillation, but rather advocating for open and lawful application of these techniques.

This perspective from the leader of a top startup accelerator provides a compelling counter-narrative to the growing calls for increased regulation and control in the dynamic field of artificial intelligence.

#GarryTan #TechPolicy