Y Combinator CEO Garry Tan has advice for OpenAI and Anthropic. This comes as the two companies raise concerns about Chinese rivals allegedly using their models to train competing systems. He has asked the American AI companies to do nothing. Tan believes regulators should instead focus on balancing open-weight and frontier models While some technology companies and US national security officials have called for action against Chinese companies accused of model distillation.
Concerns Over AI Model Distillation Practices
Distillation is a method where the outputs of a more advanced artificial intelligence (AI) model are utilized to train a smaller or less capable model. This technique is prevalent in AI development but raises concerns when companies allegedly use outputs from another developer’s model without authorization or in breach of terms.
Recently, OpenAI and Anthropic have both expressed worries regarding the practices of Chinese AI firms. Anthropic has accused Moonshot AI, DeepSeek, and MiniMax of distilling its models. Meanwhile, OpenAI suspects that DeepSeek’s V3 and R1 model architectures were derived from its GPT-4 and GPT-4o models. However, these accusations remain unverified and are currently based on the companies’ claims.
The issue of model distillation is compounded by ongoing legal challenges faced by AI developers concerning the data used to train their systems. Critics of OpenAI and Anthropic’s allegations have questioned whether these companies have also used copyrighted material without permission in their own model training processes.
Anthropic said it was concerned that the researchers could have been covertly using Claude in connection with the creation of biological weapons. Rather than attempting to prevent distillation, Tan believes regulators should concentrate on finding an equilibrium between open-weight models, whose underlying model weights can be made available for others to use or modify, and frontier models developed by companies such as OpenAI and Anthropic. Cybersecurity is among the risks Tan considers more immediate.
His comments come as Anthropic said on Thursday that it had blocked access to Claude in five cases involving scientists in unspecified foreign countries who were using the models to conduct research on dangerous pathogens.