In a recent discussion about the practices of Chinese AI laboratories, Y Combinator's CEO, Garry Tan, expressed a desire for minimal regulatory interference. He suggests that American AI labs could also benefit from employing similar distillation methods currently utilized by their Chinese counterparts.
In an interview with CNBC, Tan stated, “I would do nothing,” emphasizing the need for an American distillation framework to allow domestic open-weight AI labs to learn from leading U.S. AI models. This strategy aims to enhance the range of open-weight options available in the U.S., making them distinct from those developed in China.
Distillation refers to the process where a model leverages prompts to understand the workings and reasoning of another model. This technique is widely accepted within the AI community to facilitate the training of new models.
This week, Anthropic released a report claiming that certain Chinese labs are conducting “illicit distillation attacks” by concealing their identities and using fraudulent means to gain access to models without proper authorization. Dario Amodei, CEO of Anthropic, has previously urged U.S. regulators to take action against such practices.
Interestingly, Tan holds a different perspective from many in the industry regarding these practices.
It is important to clarify that Tan does not support unethical behaviors like using stolen credentials for distillation. Instead, he advocates for a more transparent approach. He believes that it is excessive for AI labs to control how users interact with the data obtained from their models. Furthermore, he points out that these proprietary AI systems did not seek permission when they aggregated vast amounts of knowledge and content from various sources, often including copyrighted materials.
“Regulating what customers can do with API calls to closed-weight models seems overly restrictive. The government has a role in ensuring that access to intelligence trained on widely available public data leans towards being a public good rather than being tightly controlled by restrictive terms,” Tan explained to TechCrunch.
Tan, an enthusiastic AI user who has even described himself as experiencing cyber psychosis, advocates for a balance between open-weight and frontier AI labs. He believes it is essential for frontier labs to thrive financially while allowing open-weight models to promote accessibility and freedom.
He highlights a critical concern regarding the future of AI, warning against the potential monopolization of this powerful technology. “The worst-case scenario for AI is a single company dominating the landscape,” he asserted. “If one entity holds a monopoly, backed by extensive capital and top-tier researchers, it would undermine innovation and diversity in the field.”
