Risk and reward: China’s AI play is different from America’s

Early in the Cold War, even as the United States and the Soviet Union engaged in a relentless nuclear arms race, they treated nuclear technology as a tool of soft power and diplomacy — something to share and export, a seal on alliances and a token of friendship.

This history suggests a framework for understanding the difference between the American and Chinese approaches to artificial intelligence. The United States is behaving as if frontier AI models could be the equivalent of nuclear weapons, offering an immensely destructive form of power that requires a careful and jealous watch. American AI tycoons constantly talk as if they were engaged in an Oppenheimer-style race, the American military-industrial complex is intent on keeping China from closing the current AI gap, and both industry and government increasingly treat frontier models as grave security threats.

Meanwhile, China treats the models that its companies produce as the equivalent of nuclear energy, a lower-risk technology to be shared and commercialized to win friends and influence the world and undermine the advantages of the American AI giants.

For a while, I’ve wondered about the sincerity of the Chinese position. Since its companies and models have consistently lagged behind Anthropic and OpenAI, China has obvious incentives to use open source to undermine Silicon Valley’s profits and reach, all while letting Washington worry about the risks that the most cutting-edge models may create.

We do not yet have a Chinese model that surpasses our own, but this past week brought a model that may come close, Kimi K3 from the Chinese company Moonshot AI. And so far Beijing is treating Kimi K3 like all its prior models: It’s going to be open source for maximum adoption, without any American-style fretting over risks and guardrails and export controls.

This decision should make us a little more confident that Beijing’s overall approach is sincere. It suggests that the Chinese don’t just reject the utopian and dystopian scenarios common in Silicon Valley. They may not even be especially “risk-pilled,” in the sense of worrying about the kinds of perils that today’s AI models might already create — through the power to hack and bring down complex systems, the power to help terrorists create bioweapons, the uncertain risks of rogue AI behavior and more.

American AI companies and the national security state seem to assume that we’re already at a point where you simply can’t let the world use the best AI models without controls and guardrails any more than you’d ship atomic bombs or ICBMs to random homes or private companies. Whereas the Chinese are willing to shrug and push the envelope and let America fret about safety and sovereignty, on the assumption that AI is a normal technology and the safety concerns will get figured out along the way.

Or to return to a more cynical reading, maybe they’re making a double bet. If AI creates only modest risks and isn’t actually some kind of strategic superpower, their opensource models are the best way to undermine the American advantage and make the world like China more than us. Alternatively, if we do get some kind AI Chernobyl from someone misusing a Chinese model — well, the repercussions and backlash will still probably hurt US companies the most, because the American public is already skeptical of AI and primed for data center moratoria.

I’m not sure where this leaves the AI safety movement. The cause of AI safety has advanced along several lines in the United States, benefiting from the progressive left’s hostility to Silicon Valley, the Pentagon’s anxiety about the power of privately controlled models and the AI industry’s internal doomer culture.