While some Silicon Valley titans and national security experts are calling for action against Chinese companies involved in model distillation, Garry Tan, chief executive of Y Combinator, a well-known startup accelerator, is less concerned.
“I wouldn’t do anything” about distillation, he told CNBC at Y Combinator’s annual Demo Day. “You could argue that there should be an American distillation regime.”
Distillation is the process of using the output of a more capable AI model to train a smaller or less capable AI model, sometimes illegally. This was a major concern for the developers of the Frontier model, namely OpenAI and Anthropic. Anthropic has accused Chinese companies such as Moonshot AI, DeepSeek, and MiniMax of this behavior, but OpenAI believes that DeepSeek’s V3 and R1 model architectures were extracted from its own GPT-4 and GPT-4o models.
The National Security Agency, Cybersecurity and Infrastructure Security Agency, and Federal Bureau of Investigation issued an official cybersecurity advisory warning on the topic on Tuesday.
Critics have pushed back on complaints about Anthropic and OpenAI’s distillation because much of the data used to train these AI models can be subject to copyright laws, Tan emphasized. The dispute is currently the subject of a US legal battle, with the New York Times suing OpenAI and Microsoft for misusing an article on model training data in 2023, and a consortium of book authors settling a lawsuit with Anthropic in 2025 over similar claims.
However, Tan believes that rather than curbing distillation, regulators should focus on creating an equilibrium between the indiscriminate weight model and the frontier model, as long as the frontier model maintains a price premium that maintains the viability of the business model.
“This is actually the ideal case. We need an open weight model that gives people freedom and access,” he explained. “If I were a regulator, I would pursue that.”
Mr Tan acknowledged that this was a difficult balance to strike, calling it a “tightrope walk”. Nevertheless, he says it’s a balance worth pursuing, noting that it “may yield the best possible outcome.”
Tan has also touted a more restrained approach to AI safety, although recent doomsday headlines related to the resignation of anthropologist Jacob Coxon have sparked record public anxiety over AI’s existential risks.
“We need to focus on scientific facts, not science fiction,” he says. “We need to respond to what’s happening now. If there’s a breach and there’s a coordinated attempt by an agent to take over our infrastructure, how do we respond to that?”
Tan believes cybersecurity is a pressing risk. And there are other short-term, existential risks. For example, on Thursday, Anthropic reported that it had blocked Claude’s access to five unspecified foreign scientists who were using the model to conduct research on dangerous pathogens. Antropic was concerned that these researchers were secretly using Claude to create biological weapons.
However, other risks have longer timelines. Tan believes that AI-related job losses and economic transformation will not have an immediate impact. Rather, he predicts that over time, people will automate rote tasks and spend more of their working hours on creative pursuits.
“It’s going to take decades for this to really permeate society, but that’s not a bad thing,” he said.
Despite this long timeline, Y Combinator’s AI investments aren’t stopping at this point. Of the 196 startups announced at Demo Day, 149 were categorized as machine learning and AI ventures.
