Garry Tan, CEO of Y Combinator, suggests that American open-weight AI labs should adopt distillation techniques to train their models using frontier AI systems. He argues that this approach would strengthen the domestic open-source ecosystem and prevent the concentration of AI power within a single proprietary provider, which he describes as a potential nightmare scenario.
Tan clarified that he does not support the use of stolen credentials or fraudulent activities to access frontier models. Instead, he advocates for a framework where open-weight labs can legitimately interact with frontier models. He believes that proprietary labs should not dictate how customers utilize the information shared by their systems, especially given that these labs trained their own models on vast amounts of copyrighted data without permission.
The proposal comes amid concerns from companies like Anthropic, which recently released a report alleging that Chinese labs are conducting illicit distillation attacks by hiding their identities and using stolen credentials. Anthropic CEO Dario Amodei has previously urged U.S. regulators to restrict such practices, a stance that Tan explicitly disagrees with, preferring to see a balance between frontier and open-weight AI development.
