Abstract:
While U.S. artificial intelligence companies such as OpenAI and Anthropic continue to accuse Chinese companies of copying advanced technologies through "model distillation," Garry Tan, CEO of Y Combinator, a well-known U.S. startup accelerator, has publicly expressed a completely different view. He said that if it were up to him, "nothing would be done" on the issue of distillation, and believed that the current regulatory focus should not be on combating distillation.
Garry Tan made the above statement during an interview during Y Combinator’s annual Demo Day event. Over the past few months, model distillation has become one of the most controversial topics in the global AI industry. Distillation refers to using the output generated by a more powerful artificial intelligence model to train a smaller or less capable model. In some cases, this process may not be authorized by the original model developer, so it is regarded by some leading AI companies as a threat to intellectual property rights and commercial interests.
At present, Anthropic has publicly accused Chinese companies such as Dark Side of the Moon, DeepSeek and MiniMax of allegedly using its models for distillation training. OpenAI believes that part of DeepSeek's model architecture may borrow from or originate from the output results of its GPT-4 and GPT-4o series models. At the same time, the U.S. National Security Agency, the Cybersecurity and Infrastructure Security Agency, and the FBI recently jointly issued a cybersecurity warning, listing relevant behaviors as security risks that require focus.
However, Garry Tan is not in favor of tougher restrictions on distillation. He believes that what regulators should really focus on is how to establish a healthy balance between open weight models and cutting-edge closed-source models, rather than simply banning or cracking down on distillation practices. In his view, as long as the leading model can still maintain sufficient performance advantages and price premiums to enable its business model to continue to operate, then the existence of an open model will be conducive to the popularization of technology and the diffusion of innovation.
He said that the ideal state should be for the open weight model to provide freedom and usage opportunities to more developers and enterprises, while the cutting-edge model maintains commercial value by virtue of its technological leadership. Although this balance is very fragile and as difficult as "walking a tightrope", it is still the most worthy development path for the industry to pursue.
Another background to the distillation controversy comes from the long-standing copyright issues in the artificial intelligence industry. Many critics point out that companies such as OpenAI and Anthropic have also faced questions about the unauthorized use of news reports, books and other copyrighted content when training their own models. The New York Times previously sued OpenAI and Microsoft, accusing them of using news content to train models without permission; many writers have also filed lawsuits against Anthropic over similar issues. Garry Tan believes this complicates the industry's ethical stance on distillation.
In addition to the issue of model competition, Garry Tan also remains cautious about the current rising "AI doomsday theory." Recently, as some artificial intelligence researchers have publicly expressed concerns about the risk of advanced AI running out of control, society's discussion of the existential risks of artificial intelligence has increased significantly. But Tan believes that the industry should pay more attention to problems that have already occurred or are about to occur in reality, rather than indulging in science fiction disaster scenarios.
In his view, risks in the field of network security are the most pressing challenge at present. For example, if AI systems are used for cyber attacks, infrastructure penetration, or automated crimes, they are far more worthy of prioritizing resources to deal with than the distant scenario of AI exterminating humanity. Regarding issues such as employment impacts and changes in economic structure, he believes that although these impacts are real, they are more likely to appear gradually in the next few decades rather than bursting out in a short period of time.
It is worth noting that despite the escalating debate around artificial intelligence safety and model competition, Y Combinator continues to bet heavily on the AI industry. Among the 196 startups displayed on this Demo Day, 149 were related to machine learning and artificial intelligence, accounting for more than three-quarters of all projects, reflecting the agency's high optimism for the future development of artificial intelligence.
Analysts believe that Garry Tan’s statement represents another voice within Silicon Valley. Unlike cutting-edge model developers who advocate strengthening protection and limiting technology proliferation, this faction emphasizes open ecology, full competition, and technology democratization. As the AI industry continues to develop at a rapid pace, debates surrounding “distillation”, open source models, national security and intellectual property are expected to continue to become important issues in the global artificial intelligence field in the next few years.
Comments