Y Combinator CEO Garry Tan argues U.S. open-weight AI labs should distill knowledge from frontier models, similar to Chinese labs, to expand American options. He told CNBC that American labs should be allowed to distill without using stolen credentials, emphasizing the need for a more open approach.
Tan explained that distillation involves prompting models to learn how they work, a common practice used to train new models. He noted that proprietary labs have historically ingested vast amounts of human knowledge without permission, highlighting the need for a balanced approach.
Anthropic recently released a report alleging Chinese labs are engaging in illicit distillation, using fraud and stolen credentials. Tan, however, believes the U.S. should allow open-weight labs to distill freely, arguing that access to intelligence trained on public data should be a public good.
"Controlling what users and customers do with API calls to closed weight models feels constraining," Tan said. He believes the government should normalize the idea that such access should be more open, rather than restricted by terms of service.
Tan's stance contrasts with Anthropic CEO Dario Amodei, who has called for stricter regulations on distillation. He argues that the current system allows proprietary labs to dominate, creating a monolithic AI provider that would be detrimental to innovation.
Tan warns that the worst-case scenario for AI is a single dominant company controlling all frontier models. He advocates for a balance between open-weight and proprietary labs, ensuring that both can thrive and contribute to the field.
Source: techcrunch
