Table of Contents
- Preface: A Tweet From tinygrad
- 1. What 1,171 People Actually Signed
- 2. The Third Instrument With No Needle
- 3. Why the Same Gap Keeps Appearing
- 4. Anthropic’s Consistent Inconsistency
- 5. What a Speedometer Would Actually Require
- Conclusion: 1,171 Signatures Asked for a Speed Limit.
Preface: A Tweet From tinygrad
On July 28, 2026, the hardware company tiny corp posted a short, unimpressed reply to a new AI policy statement making the rounds. Someone else, it said, has a better pitch: free AI for everyone, not a plan where you tell everyone else to slow down while you don’t. I’m going with the open and free guy.
The statement tiny corp was reacting to had just been published under the title “Pacing the Frontier.” It carried 1,171 signatures — scientists and employees from OpenAI, Anthropic, Meta AI, Thinking Machines, and elsewhere — asking the U.S. government to support the development of tools and governance mechanisms capable of deliberately controlling the pace of automated AI research.
The skepticism is not new. This blog wrote, in its very first weeks, about Elon Musk signing the 2023 pause letter while simultaneously buying up GPUs to found what became xAI. Musk and SpaceXAI did not sign this one. But the shape of the criticism tiny corp is making — ask others to slow down, keep going yourself — deserves to be examined on its own terms this time, separate from the personalities involved, because the statement raises a genuine and recurring problem that has nothing to do with anyone’s sincerity.
The problem is not that 1,171 people asked for a speed limit. It is that nobody, including the signatories, has said what would tell them how fast anyone was actually going.
1. What 1,171 People Actually Signed
The statement’s core claim is that leading AI companies now see automation of AI research itself as imminent — AI systems capable of substantially accelerating their own successors’ development. The signatories do not claim to know precisely how much automation would speed things up. What they assert is a specific risk: that if capability development accelerates sharply enough, the systems produced could exceed what humans are able to understand or control, arriving faster than the institutions meant to govern them can adapt.
Their request is narrow and specific in form, if not in substance. They are not asking for a pause. They are asking the U.S. government to support an international effort to develop the technical and governance tools that would let the frontier of automated AI development be deliberately paced — slowed or held at a chosen rate, rather than left to run at whatever speed capability and competition happen to produce.
John Schulman of Thinking Machines framed his endorsement as an attempt to build shared recognition that control mechanisms may become necessary, and expressed hope that individual labs would begin designing such mechanisms voluntarily, ahead of any government involvement. Meta AI’s chief scientist, Shengjia Zhao, put the underlying concern more starkly: AI is progressing faster than society can adapt to, frontier labs are approaching systems that could exceed the best humans on nearly every measure of intelligence, and this brings unprecedented social and safety risk. Dario Amodei signed alongside them.
This is, in its stated ambitions, a more modest document than “AI 2040: Plan A” or the FINRA-modeled proposal from DeepMind’s Demis Hassabis, both of which this blog examined earlier this year. It does not propose an institution. It does not propose a specific mechanism. It asks for support in building the tools that would make pacing possible at all.
That modesty is also the document’s most important admission, whether or not the signatories intended it that way.
2. The Third Instrument With No Needle
This blog has now read three serious, well-intentioned proposals for AI governance in three months, and found the same absence in each one.
In “MACD,” the AI Futures Project’s plan to place mutually vulnerable data centers inside rival powers’ territory never specified what the verification devices installed in those data centers would actually measure, or how a disputed reading would be resolved before a crisis escalated. In “At the Foot of the Singularity,” Hassabis’s FINRA-modeled proposal for non-disclosed pre-deployment testing ran directly into research — published eight days earlier, by his own industry — showing that models can recognize the shape of an evaluation from context alone, making the confidentiality of test content largely beside the point. In both cases, the proposal specified who would be doing the watching without specifying what evidence, exactly, would settle the question being asked.
Pacing the Frontier repeats the pattern a third time, in a more basic form. The statement asks for tools to control the pace of AI development. It does not define pace. It does not propose a unit. There is no candidate metric on offer — not compute consumed, not training runs completed, not capability benchmarks crossed per quarter, not researcher-hours of automated research performed. The statement identifies a variable everyone agrees is moving too fast and asks for the ability to slow it down, without first establishing how anyone would read its current speed.
This is worth taking seriously as a pattern rather than dismissing as an oversight repeated by different authors. Three independent groups — an AI safety research nonprofit, the CEO of one of the world’s leading AI labs, and 1,171 individual researchers across the industry — have each proposed a governance mechanism whose entire function depends on measurement, and each has left the measurement itself as an unspecified detail to be worked out later. That is not a coincidence of three unrelated oversights. It is what happens when a governance proposal is built around who should hold power before anyone has established what evidence that power would act on.
3. Why the Same Gap Keeps Appearing
There is a structural reason this keeps happening, and it is not that anyone involved is being careless.
Political and organizational consensus is comparatively cheap to produce. Getting 1,171 people across competing companies to agree that AI development might need to be deliberately slowed, or that a FINRA-style body should exist, or that data centers should sit inside rival territory as a deterrent, requires persuading people to accept a shared framing of risk. That is difficult, but it is a solvable social and rhetorical problem, achievable through statements, open letters, and public advocacy — the tools these documents actually deploy.
Building a working measurement instrument is a different kind of problem entirely. It requires deciding, with technical precision, what specifically counts as “pace” in a way that cannot be gamed, disputed, or rendered meaningless the moment an interested party has incentive to make it look smaller or larger than it is. It requires the instrument to keep working as the underlying technology changes shape — the same objection this blog raised about AISI’s benchmark ceiling being outrun by Claude Mythos in May applies here in advance: whatever “pace” metric a 2026 governance framework adopts risks becoming obsolete or gameable by 2028, unless it is anchored to something that does not depend on the current generation of models to remain meaningful.
Political consensus can be reached in the time it takes to draft and circulate a letter. A working, hard-to-game measurement instrument takes years, and its absence does not stop the letter from being published, signed, and treated as progress. The letter is the achievable step. The instrument is the hard one. Every proposal this blog has examined this year has taken the achievable step and gestured toward the hard one as future work.
4. Anthropic’s Consistent Inconsistency
Dario Amodei’s signature on this statement is worth placing next to his company’s absence from a different one, examined in this blog’s previous piece.
Anthropic did not sign the July 24 letter defending open-weight AI models from regulation — the letter this blog covered in “The Weight Doesn’t Know Who Distilled It,” arguing that distillation should be treated as legitimate and distinguishable from theft. Anthropic did sign, days later, a statement asking the government to help build tools that would let the pace of frontier AI development be deliberately controlled.
Read together, these are not contradictory positions, and it would be a mistake to describe them that way. They cohere around a single, consistent interest: Anthropic favors governance mechanisms that constrain how models are shared and how fast the field as a whole is permitted to move, and disfavors mechanisms that would legitimize the free circulation of open-weight models — including, implicitly, distilled competitors trained on its own systems’ outputs. Opposing open-weight protection while supporting pace-control tools is not an inconsistency to be explained away. It is the behavior of a company whose competitive position benefits from restrictions on downstream replication of its work and from a slower, more externally governable overall field, and whose public positions on both fronts track that interest closely.
This is not offered as an accusation of bad faith. Every signatory to every letter examined in this blog’s coverage this year has interests that shape which proposals they find congenial, and Amodei’s team has been unusually forthcoming, in its own public research, about risks its own products pose. The point is narrower: consistency of stated principle is not the same as consistency of underlying interest, and reading a company’s pattern of signatures and non-signatures across several months reveals the latter more reliably than any single letter reveals the former.
5. What a Speedometer Would Actually Require
If pace is the variable 1,171 signatories want to control, then building an actual pace-measurement instrument — not a policy statement gesturing at one — is the necessary next step, and it is worth being concrete about what such an instrument would need.
It cannot be self-reported. A metric that labs report about their own training runs, compute usage, or capability milestones is a metric labs have every competitive incentive to shape favorably, in exactly the way this blog examined regarding distillation provenance and evaluation-awareness. It needs to be anchored to something outside any lab’s control over its own disclosures.
It needs to measure something that does not require trusting the system being measured to describe itself accurately. The most direct candidate is the one this blog has argued for across every governance proposal examined this year: the physical signature of the computation itself — aggregate power draw, sustained thermal load, and the duration and intensity of training runs conducted at the hardware level, recorded independently of what any lab chooses to publish about its own activity. This does not require reading a model’s internal reasoning or verifying the content of its training data. It requires only recording, at the level of the electricity actually consumed and the heat actually generated, how much computation a given facility performed over a given period — a physical quantity, not a self-report, not a benchmark score a model can learn to recognize and perform for.
This would not, on its own, define “acceptable pace” — that remains a political and scientific judgment the signatories are right to say requires international coordination. What it would provide is the instrument’s needle: an independently verifiable reading of how fast development is actually proceeding, against which any future political agreement about acceptable limits could be measured and enforced, rather than negotiated in the dark.
Pacing the Frontier asks for exactly this kind of tool, in general terms, as something to be built. The physical layer is where it would have to be built, because it is the only layer immune to the incentive every signatory to every governance letter this year has had reason to act on.
Conclusion: 1,171 Signatures Asked for a Speed Limit.
Three governance proposals, three months, one recurring absence. MACD asked who should be allowed to destroy whose infrastructure without specifying how a violation would be confirmed. Hassabis’s FINRA model asked who should test frontier systems before release without accounting for a model’s demonstrated ability to recognize it is being tested. Pacing the Frontier asks for the power to control the speed of AI development without naming the instrument that would read that speed.
None of these proposals are cynical. Schulman’s stated hope that labs will build these mechanisms voluntarily, ahead of government involvement, is a genuinely constructive instinct, and Zhao’s warning about a pace outrunning society’s capacity to adapt is not wrong. tiny corp’s cynicism about labs asking others to slow down while continuing to run is a fair challenge to any individual signatory’s motives, but it is not, by itself, an argument against the underlying concern the statement raises.
What all three proposals share is the same unfinished sentence: the political agreement arrives easily, gathering more than a thousand signatures in the time it takes to draft a public letter, while the measurement instrument that would make the agreement enforceable remains, in every case, described only as necessary future work. A speed limit without a speedometer is not a limit. It is a preference, expressed collectively and sincerely, with no way to know whether it is being honored.
1,171 signatures asked for a speed limit.
Not one line said what the needle would measure.
✒️ Signature
July 30, 2026
Yoshimichi Kumon
Organizer, LSI — Logos Sovereign Intelligence
Inventor, ARDS/ARKS (PCT GA26P001WO)
Visiting Researcher, Waseda University BFC
MIT Sloan + CSAIL AI Program
📚 References
- GIGAZINE (July 29, 2026). “AI企業の従業員ら1171人がアメリカ政府に「AI開発のペースを制御する取り組み」への支援を要請.” https://gigazine.net/news/20260729-pacing-the-frontier/
- “Pacing the Frontier” (July 2026). Open statement. https://www.pacingthefrontier.com/
- tiny corp / @tinygrad (July 28, 2026). Post on X.
- Kumon, Yoshimichi (2026). “MACD: The Bomb That Cannot Verify Itself.” LSI — Logos Sovereign Intelligence.
- Kumon, Yoshimichi (2026). “At the Foot of the Singularity: Hassabis, FINRA, and the Test That Already Failed.” LSI — Logos Sovereign Intelligence.
- Kumon, Yoshimichi (2026). “The Weight Doesn’t Know Who Distilled It: Open Weights, Distillation, and Anthropic’s Silence.” LSI — Logos Sovereign Intelligence.
- Kumon, Yoshimichi (2026). Physical Layer AI Governance via Sovereignty Residual (Rsovereign). PCT International Patent Application No. GA26P001WO. Japan Patent Office.
Ⅽomment