AI's chief scientists asked for a brake. Not to use it yet

On July 28 a single-page website went up with three paragraphs and one request on it. By the next morning it carried more than 1,200 signatures from people who build frontier AI models, including the chief scientist of OpenAI, the chief scientist of Meta, and the CEO of Anthropic.

One sentence, 1,200 names

The statement is called Pacing the Frontier. The ask: that the US government support an international effort to develop the technical and governance tools needed to "deliberately pace the frontier of automated AI development."

That is all of it. No pause, no compute cap, no licensing scheme. The signers say directly that they are not asking anyone to slow down right now.

The roster is why it made news. Dario Amodei signed, along with Anthropic co-founders Jack Clark, Jared Kaplan, Benjamin Mann and Chris Olah. So did OpenAI chief scientist Jakub Pachocki, Google DeepMind's VP of AI safety Anca Dragan, Meta chief scientist Shengjia Zhao, and John Schulman, now chief scientist at Thinking Machines. It opened with over 1,100 names and passed 1,200 inside a day.

Then the companies themselves backed it, which rarely happens with employee letters. @OpenAI posted that at some future point acceleration may run high enough that "the world will need to pace the rate of AI advancement." @AnthropicAI said it supports the petition, "signed by our CEO, several co-founders, and senior staff," and pointed back at its own June research on recursive self-improvement.

The specific fear

The statement names one worry. The leading companies think they may be close to automating AI research itself. Once models do most of the work of building better models, the pace stops being set by how fast humans can think, and the letter warns that capability development could then run "beyond our ability to understand or control the resulting systems."

Anthropic put numbers behind that in June. Its report said Claude now writes roughly 80% of the company's new production code, and that the length of task the model handles reliably has been doubling about every four months. You can argue with the extrapolation. The input is already real: the tools are a large part of how the next tools get built.

The rest of the letter is game theory. No lab will ease off on its own, because easing off on its own just hands over the lead. So the request is that somebody build the coordination machinery, verification included, so that the option exists at all.

The pushback on X

Not everyone on X read this as courage. Nate Soares of the Machine Intelligence Research Institute, who has argued for years that this risk is underrated, called the statement "decent by the standards of 2024, but ultimately it's still softpeddling." What bothered him most was the opening line about AI helping create a better future. He said he understood why they wrote it that way and was sorry it was not stronger.

The structural complaint is harder to shrug off. A request for the ability to slow down someday, signed by the people who could slow down today, costs its signers nothing this quarter.

The field also does not agree with itself. That same day Mark Zuckerberg published a Wall Street Journal op-ed arguing that the real danger is who ends up holding this capability, and that the answer is spreading it widely instead of concentrating it in a few labs. Meta's own chief scientist signed the pacing letter. Two of the largest companies in AI now hold opposite theories of safety, and one of them is arguing with its own research staff.

My read

The policy fight will run for years and will not reach your business. The admission sitting inside it already has.

These people just said, on the record and with their names attached, that their own systems are a major input into their own progress, and that they cannot forecast the result well enough to plan around it. That is not a claim about 2030. It describes this year, and it explains why the tool you standardized on in March is a different product in July.

What it means if you run a shop here

You are not going to help coordinate the frontier. You can coordinate your own operation, and the failure the letter describes has a small, dull local version: software doing things faster than anyone is checking them.

It looks like a follow-up sequence still sending an expired offer three weeks later. Or a sync that quietly stopped writing records to your CRM, which nobody notices until someone goes looking for a customer's history and finds a gap.

The fix is not clever. Every automation you run should have a name attached to it, a log somebody actually reads, and a way to switch it off that does not require calling a developer. If nobody can say who owns it, then nobody is checking it.

That is most of what we do in a free process audit at New Face Design: walk through the work your team repeats, decide what is worth handing to software, and confirm that whatever you already turned on is still doing what you think it does.