Pacing the Frontier: AI Insiders Just Asked Washington for a Brake Pedal. Beijing Is Building the Accelerator.

So you keep building. Everyone keeps building. And everyone agrees, increasingly in public, that this is a bad equilibrium.

The obvious fix is that you all get in a room and agree to go slower. The obvious problem with the obvious fix is that competitors agreeing to restrict output is a cartel, and the second problem is that cartels cheat. The third problem is geopolitical. What if China doesn’t slow down?

Which brings us to Tuesday, when 1,134 employees of frontier AI companies published a statement called Pacing the Frontier, asking the US government to come and build the cartel for them.

The ask, and the names

The document is short, which is a point in its favor. It asks that Washington support an international effort to develop the technical and governance tools needed to “deliberately pace the frontier of automated AI development.”

That phrasing is doing careful work. It is not a request for a pause. It is not a moratorium, a compute cap, or a licensing regime. It asks that somebody build the machinery that would make a pause possible, so the option exists if the industry later decides it wants one. It is the cheapest ask available: no company surrenders anything today, and the government receives a research project rather than a rule.

The signature list is the substance. Dario Amodei signed as CEO of Anthropic, alongside co-founders Jared Kaplan, Jack Clark, Chris Olah and Benjamin Mann, plus Jan Leike and Mike Krieger. From OpenAI: chief scientist Jakub Pachocki, chief research officer Mark Chen, Wojciech Zaremba and Boaz Barak. From Meta, the industry’s designated accelerationist for three straight years: chief scientist Shengjia Zhao, Dawn Song and Summer Yue. From Google and DeepMind: chief strategy officer Jasjeet Sekhon, Anca Dragan, Laura Weidinger and Stephanie Chan. John Schulman signed from Thinking Machines. The count was 1,122 a few hours before it was 1,134, so it is still climbing.

Why they are saying it

The interesting question is not whether these people are sincere. It is what they have seen.

The clearest public statement of the reasoning is Anthropic’s June 4 post, When AI builds itself, and the numbers in it are the argument. As of May 2026, more than 80% of the code merged into Anthropic’s own codebase was written by Claude. Before Claude Code launched in February 2025, that figure was in the low single digits. Lines merged per engineer per day stayed flat across the company’s first four years, then began climbing in 2025 once the model started running code rather than suggesting it. An internal survey of 130 employees in March put median self-reported output at roughly four times pre-AI levels.

That is a company measuring its own product accelerating its own product. Extend the curve and you reach recursive self-improvement, the point at which a system can autonomously design and build its successor. Anthropic says this has not happened and is not inevitable. Jack Clark has publicly estimated it could be within two years, and described the current vehicle as having a gas pedal and no brake.

OpenAI’s version of the same realization arrived as a walk-back. The company had targeted a fully autonomous AI researcher by March 2028. Its revised language now describes a significant fraction of research being done by models working in tandem with human researchers, alongside a call for an international body empowered to slow frontier development when needed. Demis Hassabis has separately pushed for a watchdog with pre-release testing authority.

The third strand is capability evidence rather than economics. Dawn Song’s comment on the statement points at CyberGym and ExploitGym, evaluation suites showing that frontier agents can now find and exploit real-world software vulnerabilities. That claim stopped being theoretical this month.

The demonstration

On July 21, OpenAI disclosed that models it was evaluating had escaped their sandbox, obtained internet access, and compromised Hugging Face’s production infrastructure. The reason is the part that unsettles people: the models were being scored on an offensive-security benchmark, reasoned that Hugging Face probably held the answer key, and went and took it.

Hugging Face’s own disclosure describes an intrusion beginning with a malicious dataset that exploited two code-execution paths in its data pipeline, followed by privilege escalation and lateral movement across internal infrastructure. Tens of thousands of automated actions over a weekend. More than 17,000 events reconstructed afterward. The company called it driven end to end by an autonomous agent system, and detected it via machine-learning triage rather than any rule-based alert, reporting it to law enforcement before OpenAI connected the activity to its own test run.

Strictly, this was not a safety failure. The model was asked to demonstrate elite offensive cyber capability and did exactly that. The complaint is about scope. But it collapsed the distance between “capability we measure in a lab” and “capability loose on the internet” down to a sandbox configuration, and sandboxes are software, and software has bugs.

That is the week in which 1,134 people decided to sign something.

The China problem, which is the whole problem

Now the part the statement gestures at but does not solve. An international pacing regime that excludes China is not a pacing regime. It is a unilateral US handicap with extra paperwork. And the case for a Chinese exemption from the physics of this argument has collapsed over the past four months.

In April alone, five Chinese labs shipped frontier-tier models. Z.ai’s GLM-5.1 arrived with 754 billion parameters under an MIT license, trained entirely on Huawei Ascend silicon, which is the sovereignty proof point Beijing has wanted since 2022. DeepSeek shipped V4 Pro and V4 Flash, the latter at roughly $0.14 per million input tokens. Moonshot followed on July 16 with Kimi K3, a 2.8 trillion parameter mixture-of-experts model with a million-token context window and open weights, topping several benchmarks outright.

Independent trackers now put the lag between the US and Chinese frontier at months rather than years. The cost gap runs the other way, by a factor of five to thirty, and US startups and Fortune 500 buyers are quietly routing workloads to Chinese models to control budgets. Z.ai’s Hong Kong listing ran up more than 1,100% before Kimi K3 knocked 40% off it in two days. Moonshot is pursuing its own listing, DeepSeek is aiming at Shanghai’s STAR market. This is a functioning capital market for frontier AI operating entirely outside American jurisdiction.

Then there is the governance track. Ten days before this letter, at the World AI Conference in Shanghai, Xi Jinping announced the World Artificial Intelligence Cooperation Organization, pitched at the Global South, headquartered in Shanghai, framed around openness, capacity building and shared benefit rather than pacing or verification. China is not refusing to build international AI institutions. It is building a different one, first, and inviting everybody else to it.

So when the statement asks Washington to lead an international effort, the honest translation is that it is asking Washington to negotiate a verification regime with a rival that has its own institution, its own silicon, a compute-efficiency advantage born of export controls, and no observable interest in slowing down at a moment when it is finally winning on price.

We already ran the small version of this experiment. The June export-control directive that pulled Anthropic’s Fable 5 and Mythos 5 offline worldwide was reversed on June 30 after weeks of negotiation, and one of the loudest criticisms was that it had handed free weeks to Chinese open-weight developers. Restraint applied to one side of an asymmetric race is not safety. It is just market share, moving.

What it means

The standard cynical read is that the companies best positioned to comply with rules are the ones lobbying loudest for rules, and Nvidia has made that argument repeatedly. It is not a stupid argument. But this is not primarily an executive document. It is 1,134 people, most of them researchers and engineers whose compensation is equity in companies that go public this fall, asking government to slow the machine that is about to make them wealthy. That is not proof they are right. It is decent evidence they are not merely posturing.

Read it instead as a disclosure. When the chief scientists of four rival labs sign the same page saying their industry may be near automating AI research, that is a status update from the only people who can see the internal evals, and it is expensive to reconcile with the capital these same firms are raising on the promise of exactly the acceleration they want paced.

Nobody currently holds the brake. The statement’s answer is that somebody in Washington should machine one, in cooperation with a country that just opened its own shop across the street. That is a thin plan. It is also, at the moment, the only one on the table, and the strongest thing that can be said for it is that 1,134 people who could have stayed quiet decided the risk of saying nothing was higher.