Garry Tan wants US open-weight AI labs to 'distill' frontier models, too
summary
Y Combinator CEO Garry Tan argues that U.S. regulators should allow American open-weight AI labs to use distillation techniques on frontier models. He believes that restricting this practice creates a dangerous monopoly and ignores the fact that frontier models were built using public data without permission. By encouraging a diverse ecosystem of open-weight models, the U.S. can prevent a single company from controlling the most powerful AI intelligence. This approach ensures that AI capabilities remain a public good rather than a proprietary secret.
key points
- Distillation allows smaller labs to extract reasoning and knowledge from larger models, which Tan views as a necessary tool for maintaining competitive open-weight options.
- Tan contends that it is hypocritical for closed-weight labs to restrict API usage when they previously trained their models on massive amounts of copyrighted public data without consent.
- The primary risk to the AI industry is the emergence of a monolithic company that controls all top-tier research and capital, making widespread access and freedom essential.
who should read it
AI developers and policymakers interested in the tension between proprietary and open-source AI development. It is also relevant for venture capitalists and startup founders navigating the regulatory landscape of machine learning.
discussion skeptical
consensus
Commenters are divided between a trust-based view and a cynical view regarding AI data privacy. One group argues that enterprise agreements and ZDR (Zero Data Retention) policies, particularly via providers like AWS Bedrock, provide sufficient legal and technical guarantees. The opposing group maintains that these are merely 'pinky promises' and that the history of scraping copyrighted data proves AI labs lack the ethical standing to be trusted with proprietary IP.
top insight
The discussion reveals a critical distinction between 'user data' and 'transformed data,' suggesting that companies might bypass ZDR promises by claiming that once data passes through a transformer, it is no longer 'user data' but a new entity. This insight highlights a potential legal loophole where models could 'learn' from prompts without technically storing the original inputs, rendering traditional privacy contracts ineffective.
Generated by an LLM from the article and thread. Check the sources before relying on it.