Nathan Gardels is the editor-in-chief of Noema Magazine. He is also the co-founder of and a senior adviser to the Berggruen Institute.
It is a paradox that a closed society like China is conquering the world with open-source AI models while the most open society, America, is producing mostly closed-source proprietary models that are less adaptable for the rest of the world, and far more expensive.
This is the concern of Andrew Ng, one of the world’s leading AI technologists who should know. Rare among the tech bro crowd, he has been both the cofounder of Google Brain and chief scientist at Baidu, China’s largest technology company, which accounts for 70% of total search traffic. Both OpenAI’s Sam Altman and Anthropic’s Dario Amodei once worked for him.
Ng recently offered his views at a gathering in Los Angeles of Hollywood filmmakers, hosted by the Berggruen Institute’s Studio B and the Mozilla Foundation, where he engaged in conversation with the celebrated screenwriter, video game scenarist and novelist David Goyer (“The Blade” trilogy; “Superman vs. Batman”; “Call of Duty: Black Ops”).
“One thing that China has done really well is that a lot of its companies are releasing open-weight models, meaning they’re published on the internet free for anyone to use, including American companies,” Ng said.
“An open-weight model is when someone has trained the AI model on lots of data, and you publish all of those numbers, called weights, on the internet, free for anyone to download and use. When there’s an open-weight model, any of us could download it and have AI run on your own laptop, as opposed to the proprietary closed-weight models, such as the leading models from OpenAI, Anthropic or Google Gemini. On those, none of us can see the numbers that AI has learned. We can only send a prompt to one of these companies and have them use these weights and show us only the response.”
For Ng, this is a “brilliant move” by Chinese companies because open-weight models rapidly increase the rates of knowledge diffusion within their companies and ecosystem.
As a result, “over the last few years China has raced ahead in AI capabilities. Right now, the U.S.’s closed-weight models are ahead of the Chinese models, but China leads the world in open-weight models. What this means is that a lot of nations that want an open alternative are adopting Chinese models.”
In nations across Africa, noted Ng, “the Chinese DeepSeek and a handful of other Chinese models are adopted very widely, far more than American models, which have really fallen behind.”
This is important for two reasons, he pointed out.
First, “the storytelling we’ve seen in Hollywood is an important source of soft power. Do we want stories told about the importance of liberty and democratic values, or do we want stories told that reflect other nations’ values? When someone asks a question of AI, such as what happened in Tiananmen Square in 1989, whose model they are using will cause the model to give an answer that reflects that nation’s values. So getting your stuff out there is a tremendous source of soft power influence.”
Second, “AI is a key part of the supply chain of how we build a lot of software products. If more and more companies are building on a Chinese-dominant supply chain, that has implications for America. It will also weaken our ability to change the way the technology develops.”
Beyond this, Ng argued that the recent shutdown [that was later lifted after review] by the White House of foreign access to Anthropic’s powerful Fable 5 model “showed to the whole world that America can impose export controls and yank access to AI technology. This has accelerated, in many capitals around the world, the urgency with which they feel like they need to secure their own supply of AI … causing many nations to look at open-weight models because once they have all the numbers no one can take it away from them. And I think this will probably have an unfortunate effect on America’s soft power.”
Reinforcing Ng’s point, last week David Sacks, co-chair of President Trump’s Council of Advisors on Science and Technology, sounded the alarm over the latest frontier GLM 5.2 model of Chinese startup Z.ai. Echoing others, he said, “We now have a Chinese open-weight model that is as good as the currently available models from OpenAI and Anthropic.”
Diversity Drives Open-Source Diffusion
Yet it is precisely the adaptability of China’s open-source models to local circumstances that, along with cost, is its competitive advantage.
I once asked Kai-Fu Lee, a top Taiwanese computer scientist, whether China’s censorship regime would distort its large language models from accurately reflecting reality.
In practice, he told me, LLMs everywhere will carry the imprint of cultural-political values, not only in China. “Different cultural zones with different values will censor different things. While the Chinese state might censor any criticism of the Party, in the West there is a kind of culturally driven censorship over sensitive speech on race and gender. In the Islamic world, there will be censorship over blasphemy against the Prophet Muhammad. Each will align what is acceptable or not in its LLM algorithms according to their sensitivities.”
In short, it is diversity that drives the diffusion of China’s open-source models.
Proliferation & Transparency
Like all technology, AI is dual-use and can be misused. The other side of the coin of an AI world without gatekeepers is the opportunities it creates for bad actors like repressive states, terrorist networks, commercial scammers and fraudsters.
“I’m much more concerned about the proliferation of open source,” former Google CEO Eric Schmidt told Noema. “And I’m sure the Chinese share the same concern about how it can be misused against their government as well as ours.
“We need to make sure that open-source models are made safe with guardrails in the first place through what we call ‘reinforcement learning from human feedback’ (RLHF) that is fine-tuned so those guardrails cannot be ‘backed out’ by evil people. It has to not be easy to make open-source models unsafe once they have been made safe.”
Just as with the nuclear weapons buildup during the early Cold War years between the West and the Soviet Union, it is hard today to imagine China and the U.S. agreeing on curbing each other’s AI potential. Yet, the more intense the new code-war competition becomes, the greater the need for some mutually agreed rules of the game to address safety concerns.
“One thing I think both sides should agree on,” Schmidt proposed, “is a simple requirement that, if you’re going to do training for something that’s completely new on the AI frontier, you have to tell the other side that you’re doing it. In other words, a no-surprise rule.”
As during the late Cold War period when the burgeoning size of nuclear arsenals threatened survival on the planet if ever used, transparency of each other’s capabilities was the first critical step in stemming further proliferation beyond the superpowers and confidently establishing a deterrent balance between them.
The global AI race is running neck and neck. It is at this moment — when each side can see the convergent interest in adopting common guardrails — that is the most propitious time to act.