The people who built the accelerator are asking the government to build the brake

On 28 July, according to Business Insider and The Next Web, 1,134 employees across OpenAI, Anthropic, Google, Meta, Microsoft and Amazon signed an open letter titled "Pacing the Frontier," asking the US government to help build the technical and governance tools to slow automated AI development.¹² I want to take the letter seriously on its own terms. I also want to notice who is in the room asking, and who is not.

What the letter actually says. Reporting by Business Insider quotes the letter's request that the US government support an international effort to develop the tools needed to deliberately pace the frontier of automated AI development.² Business Insider also reports the letter's concession that each company and country is under intense competitive pressure not to unilaterally slow that acceleration.² The letter does not, on the reporting available, ask for an immediate pause. It asks that the option to pace be built now, before it is needed.¹

Who signed. Business Insider names Anthropic co-founders Jack Clark and Jared Kaplan, OpenAI chief scientist Jakub Pachocki, Meta chief scientist Shengjia Zhao, and Anca Dragan, who leads AI safety and alignment at Google DeepMind.² The Next Web reports that both Anthropic and OpenAI formally endorsed the letter as organisations, and that Sam Altman did not personally sign.¹ I flag both because I cannot independently verify them; wire reporting is what we have.

The honest sentence in the letter

The most revealing line in the letter, as Business Insider reports it, is the concession about competitive pressure.² The argument is not one about safety but about structure. It says: the market as currently arranged cannot produce the behaviour we say we want, because whoever slows first loses commercially, and none of us can afford to lose commercially.

This is a plain description of a collective-action problem (a situation where every individual actor's rational choice produces a collectively worse outcome than cooperation would). The signatories are telling the US government that the industry they run cannot self-regulate on this question, because the incentives inside it point the wrong way. On this narrow point, I think they are right, and the honesty is worth noticing.

But honesty is not the same as coherence. The obvious response to "we cannot slow down because our competitors will not" is: you are the competitors. Anthropic's CEO, OpenAI's chief scientist, Meta's chief scientist and DeepMind's safety lead are, per the reporting, on this letter.² Between them they set a large share of the pace they say they cannot slow. The former Microsoft executive Steven Sinofsky's objection, that these are the people who could just stop, is not bad faith. It is the first question any reader should ask.

A market failure is a real thing, and so is the fact that the letter's authors are the market.

The letter's implicit answer is that no individual firm can move first, so an external mechanism is required. That is a defensible position in game theory. It also happens to be extremely convenient: it locates the responsibility for slowing down somewhere other than the desks of the people who signed. I do not think that makes the letter cynical. I think it makes it partial.

Who is not in the room

The 1,134 signatories, per the reporting, are engineers, scientists and executives at frontier labs.¹² They are the people building the systems. The letter is about the risk that those systems accelerate beyond their builders' ability to understand or control them — the recursive self-improvement concern (AI systems that speed up their own development).²

Notice what the letter is not about. It is not about the coders on teams that have been reduced because a model now does part of their work. It is not about the content producers, analysts, translators, junior lawyers and customer-service workers already subject to AI deployment decisions they were not consulted on. It is not about the communities living next to the data centres, or the ones that will not get them because New York, per Business Insider's reporting, has imposed a one-year moratorium on new mega data centre construction.²

A letter about a future harm, written by people currently producing a present one — both facts are true at the same time. The workers most exposed to AI's deployment today are not the audience for the letter and were not asked to co-sign it. The distributional question the letter does not ask is: pacing for whose benefit, on whose timetable, protecting whom from what?

I do not think any of that is a reason to dismiss the letter. Recursive self-improvement, if it is real and imminent in the way the signatories say it is, is a serious problem and worth writing to government about. It is a reason to be careful about which coordination problem gets called the coordination problem, and whose is treated as urgent enough to require an act of Congress.

The competitive-moat critique, engaged

The standard critique of big-lab safety letters is that they function as incumbent protection: safety framing raises barriers to entry that entrench the largest players. The critique has real force here. Business Insider reports that days before this letter, several of the same companies signed a separate letter championing open weights as the path to safety, and that many signatories overlap.¹ Meta's chief scientist reportedly signed the pacing letter; on the same day, per The Next Web, Mark Zuckerberg published a New York Times op-ed calling rival labs' discourse doom-filled.¹ A company's chief scientist and its CEO holding opposite positions publicly, on the same day, is not a communication failure. It is a map of a real fault line about whether "safety" or "openness" better protects a given firm's competitive position.

Where I part company with the pure moat-critique: the letter is not asking labs to slow themselves. It is asking government to build external tools. The distinction matters, because a government tool would in principle bind smaller competitors too, and the signatories cannot control which side of it they end up on. It does not dissolve the critique. It sharpens what the critique should now be about: not the labs' motives, but the design of the tools, and who gets a seat at the table where they are built.

What I am watching

Congressman Ted Lieu has, per The Next Web, publicly tied the letter to a proposed AI kill-switch bill.¹ Whether that bill acquires text, hearings, and a co-sponsor list, and whose names appear on it, is the first concrete signal of whether this letter is the start of a mechanism or the end of a news cycle.

Glossary

Recursive self-improvement (RSI) AI systems that accelerate their own development, potentially faster than humans can supervise.

Collective-action problem A situation where every actor's rational choice produces a collectively worse outcome than cooperation would.

Open weights Model parameters released publicly, so anyone can run, fine-tune or study the model.

Competitive moat , A structural advantage, capital, regulation, scale — that shields incumbents from new entrants.

---

Footnotes

Footnotes

1. Alina Maria Stan, "1,134 AI insiders just asked Washington for a way to slow AI down," The Next Web, 28 July 2026. https://thenextweb.com/news/pacing-the-frontier-ai-employees-letter-us-government ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

2. Brent D. Griffiths, "Over 1,100 AI workers sign letter asking US to support tools that 'pace the frontier of automated AI development'," Business Insider, 28 July 2026. https://www.businessinsider.com/ai-open-letter-automated-development-2026-7 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹