As for China’s AI research lab’s use of distillation techniques to extract knowledge from cutting-edge model makers, Y Combinator CEO Garry Tan hopes regulators won’t get involved in it. In fact, he thinks U.S. AI labs should probably play the same game.
“I wouldn’t do anything,” he said in an interview with CNBC earlier this week. “You could argue that there should be an American distillation regime.”
He elaborated to TechCrunch that this means he wants small U.S. promiscuous AI labs to use the same types of training techniques as U.S. Frontier AI labs, giving the U.S. a more robust set of promiscuous options outside of China.
Distillation is when a model author prompts another model extensively to learn how it works and why it works. It is commonly and legitimately used by AI labs to help train new models.
Anthropic released a second report this week alleging that Chinese laboratories are engaging in “illegal distillation attacks,” distilling without authorization and concealing their identities, and relying on fraud and stolen credentials to do so. Anthropic CEO Dario Amodei has previously publicly called on U.S. regulators to crack down on distillation.
Notably, the commanders of Silicon Valley’s prestigious and prolific startup accelerators disagree.
To be clear, Tan is not advocating that American AI research institutes distill using stolen credentials. He wants them to come through the front door freely. In fact, there are two elements to his argument. He feels it’s a step too far for AI labs to dictate what customers can do with the information their models share.
He also points out that its own AI lab did not ask for permission to collect as much human knowledge as possible to train its models. They are notorious for ingesting large amounts of copyrighted content without the permission of these intellectual property owners.
“It feels restrictive to control what users and customers do with API calls to closed-weight models, and there is a role that the government can play here to normalize the fact that access to intelligence trained on extensive public access data should itself be a type of public good, not locked behind restrictive terms of use,” he told TechCrunch when asked why American labs should also be free to distill.
Tan, himself an avid AI user who once described himself as suffering from cyberpsychosis, wants to see a balance between Openweight AI Labs and Frontier Labs.
“They’re on the frontier and they’re pushing it forward. We hope it’s fundable and a great business model that continues,” he told CNBC. “We need an open weight model that gives people freedom and access.”
To him, the true AI doom scenario is for all of the vast power of frontier AI to end up in the hands of a single powerful proprietary provider. “The nightmare scenario, the doomsday scenario for AI, is that there’s only one company,” he said. “They have the best access to capital. They have the best AI researchers. You take the capital and run away, and all of a sudden you have one monolithic company. That’s not good.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
