Foreign media reports that Y Combinator's Chief Executive Officer, Garry Tan, recently publicly opposed the United States' tightening of restrictions on AI distillation. He believes that regulatory authorities should not prevent local American open-weight laboratories from using similar training methods to extract capabilities from cutting-edge models, in order to expand the supply of open models in the United States itself.
The dispute stems from accusations by Anthropic.
This statement came after Anthropic released a new report. This week, Anthropic claimed that some Chinese AI laboratories are launching “illegal distillation attacks” on their models by hiding their identities, using fraudulent methods, and employing stolen credentials.
Anthropic The CEO, Dario Amodei, has also previously publicly called on US regulatory authorities to crack down on distillation activities. In contrast to this stance, Tan stated in an interview with CNBC that if it were up to him, 'he wouldn't do anything at all'.
Tan argues that American laboratories can also distill
Tan then further explained to TechCrunch that what he meant was not to support account theft or bypassing permissions, but rather to hope that smaller, more open-weighted AI laboratories in the United States could access cutting-edge American models in a compliant manner and use similar training techniques.
Distillation usually refers to a process where one model learns the output methods and reasoning patterns of another model by making numerous calls to it, and then uses this knowledge to train a new model. This approach is not uncommon in the AI industry and is also widely used in model development.
There are two core points to Tan: First, closed-model companies should not overly restrict how customers use the information output by the models; second, many proprietary model companies absorb a large amount of public internet content during the training phase, which includes unauthorized copyrighted material. Therefore, imposing strict restrictions on downstream usage methods now is not well-founded.
The focus is on the balance between openness and closure.
Tan believes that it is very important for frontier model companies to continue to push forward the boundaries of their capabilities and form sustainable business models; however, at the same time, the United States also needs a strong open-weight model system that provides developers and enterprises with more choices, rather than concentrating these capabilities in the hands of a few closed platforms.
In his view, the worst scenario is not that models spread too quickly, but rather that in the end, only one company remains in the industry that possesses the strongest capital, research personnel, and computing power resources, and it will dominate the supply of AI capabilities for a long time. Following this logic, allowing more domestic American models to participate in competition is in itself a way to reduce concentration.
