In the high-stakes arena of artificial intelligence development, a divergence of opinion is emerging among key industry players and national security strategists. While some prominent Silicon Valley firms and U.S. intelligence agencies are sounding the alarm over the practice of AI model distillation by Chinese companies, Garry Tan, CEO of the influential startup accelerator Y Combinator, advocates a more measured approach.
“I would do nothing” regarding AI distillation, Tan told CNBC at Y Combinator’s recent annual Demo Day. He suggested that instead of implementing punitive measures, the focus should be on fostering a balanced ecosystem for AI development. “We could argue that there should be an American distillation regime,” he posited, implying a need for domestic consideration rather than external prohibition.
Model distillation, a technique where a more sophisticated AI model’s outputs are used to train a smaller, less complex one, has become a significant point of contention. Frontier AI developers, including OpenAI and Anthropic, have raised concerns about potential misuse. Anthropic has publicly accused Chinese entities like Moonshot AI, DeepSeek, and MiniMax of engaging in this practice. OpenAI, meanwhile, has stated its belief that certain model architectures from DeepSeek were derived from its own advanced GPT-4 and GPT-4o models.
These concerns have culminated in official advisories. The U.S. National Security Agency, Cybersecurity and Infrastructure Security Agency, and Federal Bureau of Investigation jointly issued a cybersecurity advisory warning about the activities of China-based AI companies involved in distilling U.S. frontier AI models. This advisory highlights the perceived national security implications of such practices.
However, the narrative surrounding distillation is not without its complexities. Critics of the complaints from Anthropic and OpenAI point to the intricate nature of AI training data. Much of this data may be protected under copyright law, a point that Garry Tan emphasized. This legal gray area is already the subject of significant U.S. litigation. For instance, The New York Times initiated a lawsuit against OpenAI and Microsoft in 2023, alleging unauthorized use of its copyrighted articles for AI model training. Similarly, a consortium of authors reached a settlement with Anthropic in 2025 concerning similar copyright infringement claims.
Tan’s perspective shifts the regulatory focus from actively curtailing distillation to cultivating a stable equilibrium between open-weight models and proprietary frontier models. He believes this balance is achievable as long as frontier models command a premium, thereby sustaining their underlying business models. “This is actually the ideal case. You want open weight models to give people freedom and access,” Tan explained. “If I were a regulator, that’s what I would go after.”
Achieving this equilibrium, Tan acknowledges, is akin to walking a “tightrope.” Yet, he contends that the pursuit is worthwhile, potentially leading to “the best possible outcome” for the AI landscape. He also champions a more pragmatic approach to AI safety, urging a departure from alarmist, doomsday scenarios that have been amplified by recent public discourse, including concerns raised by the resignation of an Anthropic researcher and subsequent media reports on AI extinction risks.
“We need to be focused on science fact, not science fiction,” Tan asserted. “We need to be responding to what is happening right now. If there was a breach and a coordinated attempt by agents to take over our infrastructure, what do we do about it?”
Tan views cybersecurity as an immediate and tangible risk. He also points to other pressing concerns, such as the potential for AI misuse in developing dangerous technologies. For example, Anthropic recently reported blocking access for several scientists in unspecified foreign countries who were using its Claude models for research that the company suspected could be directed towards creating bioweapons. This incident underscores the dual-use nature of advanced AI capabilities.
Looking further ahead, Tan anticipates that widespread AI-driven job displacement and economic transformation will unfold over a longer timeframe. He envisions a societal shift where individuals increasingly offload routine tasks to AI, dedicating more of their professional lives to creative and complex endeavors. “It will take decades for this to actually percolate into society, and that’s not a bad thing,” he remarked.
Despite this long-term outlook on societal impact, Y Combinator is actively investing in AI innovation. At its recent Demo Day, a significant majority of the 196 presenting startups, specifically 149, were identified as ventures in the machine learning and AI sectors, signaling the sustained momentum and enthusiasm within the startup ecosystem for this transformative technology.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/25626.html
