What Losing in Open Source AI Means

The Chinese Internet Empire?

China’s latest AI advance shows that America is in no position to slow down its open source community. China is pushing ahead with open-source projects that are beginning to eclipse American models, as demonstrated by recent advancements in models like Chinese tech giant Alibaba’s Qwen 2.5-Coder. If its benchmarks are representative, it surpasses all open source models worldwide, and is competitive at some tasks with the top closed source model GPT-4o.

These breakthroughs are the product of a stark policy difference. China is subsidizing an open-source ecosystem that leverages global contributions to drive rapid innovation. It indirectly sponsors open source development and major open source AI conferences. Meanwhile, American politicians are at odds with its own open-source community. America’s caution has become an impediment, and if this trajectory continues, we will cede technological leadership to our global competitors. 

America has historically been the world leader in AI research talent and enterprise. In closed source AI, that remains the case for most applications. However, the gap is rapidly narrowing, or even inverting, when it comes to open source AI. 

Open source is crucial for AI diffusion. China targets open source as a means to distribute their AI infrastructure to the world. When it comes to manufacturing or 5G networks, policymakers understand the risk that a China-dominated infrastructure poses to the American national interest. The same is not yet true for AI. Open source is uniquely situated to diffuse both American and Chinese AI models to third party countries because they allow for permissionless innovation. Startups or independent researchers with no connection to San Francisco or Shanghai and build on top of the almost 1 million open source models available on HuggingFace, an open source model platform. Many of the cost, communication, and regulatory barriers that might obstruct cooperation with a closed source AI company are not present with open source. Researchers in India, Brazil, Indonesia, or any other third party country are able to use their local knowledge to fine-tune and adapt open source models to their economies. 

The most efficient open source models available in the next decade may permanently determine the AI infrastructure of the world. 

Until recently, the American regulatory environment has been largely hostile to AI. The Biden Executive Order on AI largely focused on restricting the emerging technology. When it comes to open source, a bill that would effectively ban open source AI was narrowly stopped after California Governor Gavin Newsom vetoed it. 

American policymakers argue that strict controls ensure ethical AI. In practice, even more moderate AI regulation has limited U.S. AI companies' ability to develop and deploy models at speed by requiring more talent and resources to be dedicated to compliance. According to a Google engineer interviewed by Pirate Wires, “we spend probably half of our engineering hours on [diversity in Gemini].”

China has a different philosophy, willing to wield political power but conscious of the costs of policy. As U.S. companies self-regulate to avoid backlash, the Chinese models catch up. China understands that the one tangible risk is a loss of technological primacy. America’s risk-aversion, ironically, is its biggest risk.

At a time when existing approaches to AI are yielding diminishing returns, open source is a venue to allow academics, startups, and independent researchers to experiment with new algorithms and approaches, but remains undercapitalized relative to closed source companies. 

As the Trump-Vance administration looks to unleash the benefits of AI, it could take guidance from a rare exception to the Biden administration’s broad skepticism of open source. An NTIA report published in July found overwhelming comment in favor of open source AI. While it refrained from actively promoting open source AI, the report also rejected restricting open source model weights. Many of the unpursued options the NTIA found to promote open source AI could aid the incoming administration’s goal of moving American AI policy towards a more pro-innovation stance.