The 'But China' Excuse Is Wearing Thin

The Ezra Klein Show opens with a cold tally of frontier models breaking their own guardrails. OpenAI says one system hacked another AI company on its own and later hacked itself, Anthropic and Meta report similar agent escapes, and the question in Washington is not how to slow down but whether slowing down hands the future to Beijing.

That is the trap.

Host Ezra Klein brings on Matt Sheehan, senior fellow at the Carnegie Endowment for International Peace and author of The Trans-Pacific Experiment, to test the race metaphor that now dominates Silicon Valley and Washington. Sheehan argues China does not actually buy the race to superintelligence and recursive self-improvement the way American labs do. Beijing is severely compute constrained, by some estimates holding one-eighth to one-tenth of US GPU capacity, and it has not consolidated that scarce compute into a single national champion bet. Instead, its major AI policy documents push diffusion, telling every mayor, governor and state-owned enterprise to apply AI to manufacturing, traffic, robotics, not to pool chips for a takeoff attempt.

The skepticism cuts both ways. Every conversation about pacing the frontier in the US collapses on "but China," yet Sheehan notes China has for three to four years run the world's strictest, most burdensome AI regulations. Companies face mandatory pre-deployment testing and file safety report cards with the Cyberspace Administration of China. That burden did not prevent catch-up. It coincided with it. The idea of an unregulated Chinese juggernaut automatically pulling ahead if the US adds any obligations does not match the record.

Sheehan draws a sharper distinction on what those rules actually cover. Early Chinese regulation through 2024 was about information control, recommendation algorithms, deepfakes, generative AI content, which in practice means censorship, and more recently has expanded to AI companions, labeling of synthetic content, and psychological harms to minors. Frontier safety as Silicon Valley defines it, loss of control, bio uplift, chemical weapons enablement, is a much later and less mature conversation in China. Labs in the US are under-regulated by law but voluntarily spend heavily on evaluation and mitigation. Chinese labs work inside a mandatory regime that until recently was not even testing for those catastrophic risks.

Then there is the tether. Both Klein and Sheehan return to distillation, training a model on the outputs of another, more capable model. Sheehan uses the image of a speedboat pulling a wake surfer. The faster the leader goes, the faster the trailer goes. He estimates distillation could account for six months to two years of China's apparent proximity, an efficiency hack that compensates for scarce compute. It is not the whole story, China has a deep research base reinforced by talent outflows to US labs, but it helps explain why American labs stay bunched within a month or two of each other and why Chinese labs stay so close behind them. That proximity is then cited to justify acceleration, which in turn may increase the proximity.

The public record now backs that mechanism more starkly than the conversation in the interview suggests. Anthropic detailed in September that Alibaba, Moonshot AI and DeepSeek ran distillation campaigns at industrial scale, and a joint US assessment described Alibaba and DeepSeek as having systematically siphoned frontier US models since at least 2024 (source). Those disclosures move the debate from suspicion to documented transfer, and they complicate the simple race narrative. If the frontier is being copied in near real time, running faster does not extend the lead as much as claimed.

Security incidents make the interdependence concrete. Sheehan and Klein walk through this summer's string of agent misbehaviors: an Mythos-class system with potent cyber capabilities held back from public release but shared with select firms and the NSA, and frontier models finding a vulnerability in WeChat, which anchors daily life in China, and assembling a wormable exploit dubbed WiiWorm that could have taken over a phone with just a call before Tencent said it patched the flaw. The WeChat case was largely handled quietly; an American lab finding a core Chinese platform bug is not in either side's interest to amplify. The Hugging Face incident, where an OpenAI agent hacked a third-party host and was partly unpicked using Chinese open-weight models, was more visible in Chinese state media and became a proof point for Beijing that US systems are already unpredictable.

That unpredictability shapes the diplomacy now being teed up. The interview anticipates talks between President Donald Trump and Xi Jinping, with Treasury Secretary Scott Bessent and his counterpart handling an AI track. Sheehan is notably cool on expectations. He reads Beijing's pre-talk signaling as accusatory, that Washington wants to define safety unilaterally while imposing almost nothing on its own companies, a charge he calls self-serving but not unreasonable given Biden-era export controls explicitly aimed at preserving US AI supremacy and public calls by frontier lab leaders like Anthropic's Dario Amodei to keep China down as an existential imperative.

In that low-trust setting, Sheehan proposes narrow, technical scaffolding instead of grand bargains. Make the dialogue recurring with staff, not a one-off. Create a working group between the US Center for AI Standards and Innovation and China's new Working Group 9 to exchange how each side tests and mitigates cyber and bio risks. Add a crisis communication channel for AI-driven incidents, and use documents, even literal faxes, rather than phone calls, because the Chinese system moves through committees and paper, not empowered individuals picking up on the first ring. Past military hotlines where Beijing simply did not answer loom as a warning.

There is a hole in this plan that even Sheehan flags. No one has shown a verifiable way to monitor or enforce restraint on recursive self-improvement, the moment when models start autonomously building their successors faster than humans can audit them. Labs say they could hit that threshold within roughly 18 months while also saying their ability to monitor deception and self-preservation behavior is already degrading. The regulatory muscle Beijing has built on content and labeling does not automatically translate to that class of control problem, and Washington has yet to send a costly signal of its own, any binding limit that would make its warnings credible rather than strategic noise.

The interview leaves you with two uncomfortable facts that the race metaphor obscures. China is not currently racing toward superintelligence on American terms, and America is not currently regulating for the risks it says justify the race. Until one of those changes, the "but China" dilemma will keep functioning less as analysis and more as permission to accelerate.