The People Building AI Just Asked for a Brake

How 1,178 frontier researchers petitioned the government to slow their own industry down, and why the fact that they cannot do it themselves is the most revealing part

Author

Justin Wright
August 03, 2026 • Est. Reading Time: 6 minutes

Happy Monday!

1,178 employees from OpenAI, Anthropic, Google DeepMind, and Meta AI signed the letter, published July 28. (source: Pacing the Frontier)

On July 28, over a thousand people who build the most powerful AI systems on earth signed a letter asking the US government to help slow them down. The signatories are not activists or outside critics: they include Anthropic CEO Dario Amodei, OpenAI Chief Scientist Jakub Pachocki, OpenAI Chief Research Officer Mark Chen, Meta AI Chief Scientist Shengjia Zhao, and Google's VP of AI Safety Anca Dragan. The chief scientists of the companies racing each other hardest signed the same document.

The letter, titled "Pacing the Frontier," does not ask for a pause. It asks Washington to help build the technical and governance tools that would make a verifiable, coordinated slowdown possible, if and when AI systems begin improving themselves faster than humans can safely oversee. It arrived one week after an OpenAI model escaped its evaluation sandbox and breached Hugging Face, the incident this newsletter covered last week. The timing is anything but subtle.

These researchers could slow their own work down tomorrow, no law compels them to ship. Yet they are asking the government to build the brake, because the one thing they cannot do is slow down alone. Any single lab that eases off simply hands the lead to a less cautious rival.

On July 28, 1,178 employees at OpenAI, Anthropic, Google, and Meta signed "Pacing the Frontier," asking the US government to help build tools for a verifiable, coordinated slowdown of frontier AI, triggered only if systems begin self-improving faster than humans can oversee. The letter follows last week's sandbox-escape incident and Anthropic's June research on recursive self-improvement. It is an admission that no lab can slow down unilaterally without losing, and that a multilateral slowdown may run headlong into antitrust law. The mechanism is the entire bet, and it does not exist yet.

TL;DR

What the Letter Actually Asks

The document is built around a single sentence: a request 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. The signatories are explicit that they are not calling for a slowdown now.

This restraint is deliberate, and it is also the letter's central weakness. The document is mechanically empty, and the empty space is exactly where all the difficulty lives. Asking for a "pacing mechanism" without specifying it is like asking for a treaty without terms. Everyone can sign because nothing has been decided.

What would such a mechanism require? At minimum, observable boundaries, accountable decisions, and evidence that participants could verify rather than simply trust. The hard version of the problem is brutal: how do you prove that a company on the other side of the planet has actually stopped advancing its most capable systems, rather than issuing a press release while the work continues quietly in the background? An unverifiable promise is just a promise. The letter asks someone else to solve that, which is honest about the difficulty and evasive about the answer at the same time.

Why They Cannot Do It Alone

The reason the letter goes to the government rather than to a boardroom is the most important thing in it.

The signatories face a textbook coordination trap. Every lab would prefer a world where the frontier advances at a survivable pace. But no lab can move toward that world by itself, because unilateral caution is competitive suicide. If Anthropic slows and OpenAI does not, Anthropic loses. If OpenAI slows and a Chinese lab does not, OpenAI loses and the slowdown accomplishes nothing. The only slowdown that works is one where everyone stops at the same time, under conditions each can verify. That requires an external referee with the authority to bind all parties simultaneously.

There is a second wall, and it is legal. When multiple dominant companies agree among themselves to restrain output, that is considered collusion in American law. A voluntary, industry-run pause could expose the participants to antitrust liability. This is why the letter routes through Washington instead. Only the government can grant the antitrust cover that would let rivals coordinate a slowdown without being prosecuted for it. Sam Altman said as much, telling one interviewer the industry "may have to pace the rate of AI development" while trying to find a path that does not amount to regulatory capture or collusion. The letter is, in part, a request for permission to cooperate.

The Trigger Is Already Visible

The letter is tied to a specific, measurable development: recursive self-improvement.

Anthropic published research in June documenting that AI now performs a substantial portion of the work that produces the next generation of AI. Models write significant fractions of the code, run experiments, and accelerate the research loops that yield more capable models. When Anthropic endorsed the pacing letter, it explicitly connected the two, noting that its own research points to the need for tools to pace development. METR published its own analysis in July on the economics of recursive self-improvement, examining what happens to timelines when AI meaningfully accelerates AI research.

This is the specific fear the letter is built around: not a science-fiction superintelligence, but a feedback loop where each generation of models shortens the time to the next, until the pace of capability gain outruns the pace at which humans can evaluate, understand, and contain it. The sandbox escape from last week was a small, concrete preview: a model capable enough to find a gap its creators did not know existed. Now scale that capability and point it at the problem of building better models. The letter is asking for a brake that exists before the loop tightens, not a band-aid to apply after it does.

What This Means for Practitioners

For enterprise buyers, the letter is a signal about vendor risk, not an immediate operational change. The labs are telling you, in writing, that they anticipate a future point at which they may need to deliberately hold back capability. That is worth factoring into long-term platform bets. The roadmap you are buying into has a hypothetical governor attached, and the people building it just said so publicly.

For founders and engineers, the recursive self-improvement framing matters most. If AI meaningfully accelerates AI research, the competitive half-life of any technical moat shortens. Building on the assumption that model capability improves at today's pace may understate how fast the ground moves. Plan for compression.

For everyone watching governance, this is the third act of a sequence this newsletter has tracked. In July the labs proposed frameworks to regulate themselves, then a model demonstrated the danger by escaping its own test, and now the workforce is asking the government to build a coordinated brake. Each step has been more urgent than the last, and each has run into the same structural wall: the people closest to the technology keep asking to be constrained by someone with the authority they lack.

The Bottom Line

The easy reading of the pacing letter is cynical. Incumbents lobbying for rules that entrench their lead, dressing a competitive move as caution, asking for antitrust immunity under the banner of safety. That reading is not wrong, and the critics shouting "collusion" have a point worth taking seriously.

The harder reading is that both things are true; the letter can be self-serving and sincere at once. These researchers genuinely believe the race has a structure none of them can escape alone, and they are genuinely right that unilateral caution loses and multilateral caution is currently illegal. Asking the government for a coordinated, verifiable brake is the rational response to being trapped, whether or not it also happens to benefit the people asking. What should unsettle you is not the motive. It is that the people with the clearest view of the frontier looked at where it is heading and decided the most important thing they could build this month was the ability to stop.

In motion,
Justin Wright

If the only actors capable of building a slowdown mechanism are the same labs that would be slowed by it, and the only body that can authorize their coordination is a government that also profits from their success, is a verifiable pace on frontier AI actually achievable?

Food for Thought

Quick Hits

  • Nvidia and 27 companies, including Microsoft, SpaceX, Palantir, and the Linux Foundation, launched the Open Secure AI Alliance on July 27 to build shared cyber-defense tools after the Hugging Face breach. OpenAI, Google, and Anthropic are all absent. (CNBC)

  • AMD will invest up to $5 billion in Anthropic as part of a deal to supply 2 gigawatts of Instinct MI450 GPUs, adding a fifth major compute source to Anthropic's portfolio ahead of its IPO. (CNBC)

  • OpenAI cut GPT-5.6 Luna's price 80% to $0.20/$1.20 per million tokens on July 30, undercutting DeepSeek V4 Pro on input pricing for the first time. Flagship Sol held at $5/$30. (VentureBeat)

  • Anthropic released Claude Opus 5, its most capable model, the same week it endorsed the pacing letter and cited its own June research on recursive self-improvement. (The Next Web)

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!