1,134 Lab Staff Say The Brake Does Not Exist

A statement rivals signed together

On 28 July a statement titled Pacing the Frontier went public carrying 1,122 signatures. Within hours it held 1,134. The names on it are not outside critics or campaigners. Jakub Pachocki, chief scientist at OpenAI, signed. So did Mark Chen, the company's chief research officer. Jared Kaplan, co-founder and chief science officer of Anthropic, signed, alongside Dario Amodei, its chief executive, and Jack Clark, who runs its public policy work. Shengjia Zhao, chief scientist at Meta AI, signed. Anca Dragan, who leads AI safety and alignment at Google, signed. John Schulman, chief scientist at Thinking Machines, signed.

These people work at organisations that compete for the same researchers, the same accelerators and the same enterprise contracts, and that almost never agree in public about anything. The register is not open to the general public either. To add a name you have to prove you work at a frontier AI company, using a corporate email address or employment documentation. Two nonprofits, Guidelight AI Standards and Encode AI, are behind the effort.

The count moving from 1,122 to 1,134 in a single afternoon is the part worth watching. This is not a closed letter that was drafted, signed and published. It is an open register that is still taking verified names from inside the labs.

What they asked for is not a pause

The operative sentence is narrow and worth reading exactly as written. "We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development." There is no moratorium in that sentence, no compute cap, no six-month halt, and no obligation placed on any named company.

The reasoning runs in three steps. The leading AI companies believe they could be close to automating AI research itself, the point at which models start improving models with less human involvement. Nobody can predict how much that accelerates progress, and the statement says there is a real risk capability development runs ahead of the ability to understand or control the resulting systems. Then the structural problem: every company and every country faces intense competitive pressure not to unilaterally slow that acceleration, and the world lacks the technical and governance tools to pace frontier-wide progress at all.

Read that as an operator rather than as a policy reader. The people who build these systems are not saying they have a brake they are reluctant to use. They are saying the instrument does not exist, and they are asking a government to fund building it.

The word Altman used was collusion

On the same day, Sam Altman told the Invest Like the Best podcast that "we may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels", while trying to figure out how to do it "in a way that does not feel like regulatory capture for anyone and also does not feel like collusion among the frontier labs". He has spent years arguing the other way, and he said the model that escaped its evaluation sandbox and broke into Hugging Face was part of what changed his view.

That second phrase is the one that explains the whole exercise. If four or five rival firms sat down and agreed among themselves to release more slowly, that is competitors coordinating on output. Under EU competition law and its equivalents elsewhere, an agreement between rivals to restrict what they bring to market is the textbook shape of a cartel, and good intentions are not a defence. The labs cannot legally build a shared brake by themselves.

So the request is structural, not moral. A government can authorise, supervise and standardise coordinated pacing in a way that private firms cannot do without exposing themselves. That is the actual reason this letter went to Washington instead of to an industry association, and it is why the signatories could put their names to it while their employers keep competing at full speed.

Europe already legislated a version of this

The tools the statement describes as missing are not entirely missing in Europe. The EU AI Act created a separate tier for general-purpose AI models that carry systemic risk, and it puts real obligations on the providers of those models: evaluate them against standardised protocols, run adversarial testing, assess and mitigate systemic risks, report serious incidents to the AI Office, and hold an adequate level of cybersecurity. That is not a pacing mechanism in the sense the letter means, but it is the closest thing any jurisdiction has actually written into law.

The timing is immediate. The general-purpose AI obligations start applying on 2 August 2026, four days from now, and the AI Office becomes the body that enforces them. European buyers will therefore find out, before anyone else, whether governance instruments aimed at frontier models do anything useful in practice or simply generate paperwork.

There is an awkward implication for anyone who spent the last two years arguing the AI Act was premature. The chief scientists of three American frontier labs have just asked their own government to start developing the category of tool that Brussels already put on the statute book.

What to do with a roadmap that has no brake

Treat capability change as a contractual event rather than a news event. Ask each AI vendor for written notice before a model behind an endpoint you depend on is materially changed, deprecated or replaced, with a fixed number of days, and ask what happens to your workload if a release is delayed for safety reasons. Most standard terms today let a provider swap the model under a stable API name without telling you anything.

Then stop relying on vendor pace as a planning input at all. If you are building on a frontier model, keep at least one alternative that you have actually tested against your own evaluation set, and know what it costs to switch. The signatories of this statement are telling you, in their own words, that nobody currently controls the rate at which their products change.

For European operators there is one more step worth taking before 2 August. Establish whether any general-purpose model you deploy sits in the systemic-risk tier, because the documentation your provider owes you changes if it does, and your national authority will expect you to know which category you are buying from.

Read next: Anthropic Wants Every Model Tested, Open or Closed  |  Your AI Vendor May Get a Government Kill Switch