Pacing the
Frontier
“AI could help create a dramatically better future, but that outcome is not guaranteed. The world's leading AI companies believe they could be close to automating AI research. It is hard to predict exactly how much this will accelerate AI progress, but there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems.
To realize AI's potential, industry, government, and society at large may need the option to buy time to address emerging risks, develop security measures, and strengthen oversight. But each company—and country—is under intense competitive pressure not to unilaterally slow that acceleration. And today, the world lacks the technical and governance tools to deliberately pace frontier-wide progress.
Building on work already underway to monitor frontier model releases:
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.”
— the statement in full, as published, signed by 1,134 employees of frontier AI companies
Every mark below is one signatory. They are grouped by employer, so the shape of the field is the composition of the letter. Filled marks signed with a name; hollow marks signed anonymously; the cyan marks are the few who left a comment. Click any mark to read them.
1,134 signatures, four labs
Search a name or a title — “chief scientist”, “pretraining”, “alignment” — or narrow the field with the filters. Matches are listed in full and ringed in the grid below; click either to load that person into the panel at the bottom.
The site publishes signatories in a fixed, non-alphabetical order. Positions 1–12 are Schulman, Pachocki, Kaplan, Zhao, Chen, Sekhon, Amodei, Clark, Dragan, Zaremba, Song, Olah. Filter to Exec / VP / founder to see all 21 marks that carry a C-suite-level title.
Composition by employer
Signatories self-report an employer. Anthropic accounts for 533 of the 1,134 signatures (47.0%), OpenAI 331 (29.2%), Google 192 (16.9%) and Meta 63 (5.6%); the remaining 15 come from Thinking Machines, Inherent and six single-signatory companies.
A concentration this lopsided can be read two ways, and the data does not settle which is right. It may indicate unusual depth of internal consensus at one company; it may also make the statement easier to characterise as one company's position paper. Both readings are compatible with the same numbers.
Anthropic titles nearly everyone “Member of Technical Staff,” so its function mix is largely unknowable from titles — the 2.6% is a naming convention, not a finding about who signed. The comparison is only safe between employers that use descriptive titles.
| Employer | Signed | Est. headcount | Est. rate |
|---|---|---|---|
| Anthropic | 533 | 3,800 – 5,000 | ≈ 11 – 14% |
| OpenAI | 331 | ≈ 7,000 | ≈ 4.7% |
| Google DeepMind | 192 | ≈ 5,600 | ≤ 3.4% |
| Meta | 63 | unclear | — |
Only 4 of the 192 Google-affiliated signatories actually wrote “Google DeepMind”; 188 wrote plain “Google.” If the denominator is the parent company rather than the lab, the true rate is far lower than 3.4%.
Position in the published order
The statement site presents signatories in a fixed sequence that is not alphabetical. Because the entire ordering ships in one payload it can be laid out end to end, which makes its structure measurable.
What is measurable. The boundaries in the sequence are sharp rather than gradual. No signatory in the first 456 positions is anonymous. No comment appears after position 474. The final 262 entries are an unbroken run of anonymous signatories, with no named signatory among them. Eighteen of the twenty-one exec-level titles fall in the first 43 positions. By decile, the anonymous share runs 0, 0, 0, 0, 4.4, 0, 0, 30.7, 100, 100 per cent.
What follows from it. Those step changes are very difficult to produce by arrival order alone — signatures accumulating over time would give fluctuating rates, not a clean cutoff at 474 and a pure terminal block. So the sequence is almost certainly the product of a deliberate ordering rather than the order in which people signed.
What does not follow. Why it is ordered this way is not determinable from the data. A presentational choice to lead with the most credentialed names would produce this pattern. So would a mundane implementation detail — a display-priority field, or anonymous entries bucketed last because they cannot be ranked by name or title. This page takes no position on which.
Whatever produced it, the ordering has an effect on readers: someone who scans the first screen of the statement encounters chief scientists and named researchers, and would have to scroll a long way to reach the anonymous quarter of the roster. That is worth knowing when assessing how the signature count reads at a glance — but it is an observation about reception, not evidence of intent.
The 70 comments
Seventy signatories (6.2%) left a comment; 68 of those have retrievable text. The comment box is optional, so a low rate is unremarkable. Because 94% of signatories left no comment, not commenting carries no information about a particular person — including the executives on the list. An earlier draft of this page treated their silence as meaningful; that inference does not survive the base rate, and has been removed.
- Total
- Median
- Longest
- Shortest
Comments containing an explicit reservation
Twelve of the sixty-eight readable comments qualify the endorsement they accompany — noting a doubt, a condition, or a disagreement with the approach while still signing.
These twelve were selected by reading all 68 comments and deciding which contain a reservation. Reasonable people would draw the line differently — Dawn Song's comment, for instance, is broadly supportive while calling pacing “heavy-handed and potentially extreme.” Read the comments and judge for yourself rather than taking the count as a measurement.
Across these twelve, the reservations from OpenAI signers tend toward regulatory capture, scope creep and preserving US competitiveness, while Anthropic's commenters more often use existential framing. That is a pattern in a very small sample — 14 and 22 comments respectively, of which a handful carry reservations — and it should be treated as an impression from reading, not a finding.
Companies with no signatories in the roster
Each count below is a search of all 1,134 self-reported affiliations. A zero means nobody in this snapshot listed that employer — not that the company was approached and declined. Signing requires corporate-email verification, and neither the eligibility list nor any invitation process is public, so an absence has several possible causes: nobody was eligible, nobody chose to sign, or nobody was asked.
Individuals not found in the roster include Sam Altman, Demis Hassabis, Ilya Sutskever, Yann LeCun, Greg Brockman, Mira Murati and Alexandr Wang — subject to the same caveat. Present, and easy to miss at position 7: Dario Amodei, CEO, Anthropic, the only sitting chief executive of a frontier lab in the list.
One reading of the data
Everything in this section is argument rather than measurement. It is one person's reading, separated out so the rest of the page can be used without it. The sections above are built to let you reach a different conclusion.
The ask is close to the lowest-commitment request that still requires government involvement: no pause, no moratorium, no compute threshold, no timeline, no enforcement mechanism, no named institution. It asks for research on a capability. That may be exactly why it could clear internal review at four competing companies at once.
Several commenters describe the statement as an attempt to establish common knowledge — to make it visible that a slowdown option is widely wanted. Jeff Klingner of Google puts it more plainly than the statement does: “Once all of us in China & the US and at the various labs see that all the rest of us also think we need to slow this down — that's when coordination becomes possible.” John Schulman and Micah Carroll make versions of the same point.
The most distinctive material in the corpus is firsthand. Nicholas Joseph, who leads pretraining at Anthropic, describing how much of his team's work the models now do; Alex Cloud saying he no longer writes code; Ethan Perez saying safety teams sprint every few months and expect to hit something they cannot solve in time. That is testimony no outside advocate could produce, and it is the part of the statement least replaceable by argument.
Three limitations are raised by signers themselves rather than by critics. Coverage — the case for coordination assumes broad participation, and the roster is four US companies, with three comments engaging China. Definition — Lauren Deason writes that “much more clarity will be needed” on what pacing means; Joshua Achiam does not know what form the tools should take; Brandon Houghton notes that rules covering only public models could push capability out of view. Concentration — 47% from one company, and four commenters worry that pacing mechanisms could centralise power.
The open question is whether a low-commitment ask is the thing that made broad cross-lab agreement possible, or the thing that makes the agreement less meaningful than the headline count suggests. The data on this page cannot resolve that, and I do not think the statement can either.
John Schulman's line — “I'd also like to see labs start designing these mechanisms voluntarily, even before the USG gets involved” — is the cheapest possible follow-through and needs no legislation. Whether anything voluntary appears in the coming months is observable, and would distinguish a signal from a commitment better than any reading of the text.
Where these numbers come from
Every figure on this page is computed at load time from the statement site's own embedded data payload — 1,134 signatory records and 70 quote records — not from the rendered page, which serves only five comments in HTML and paginates twenty signatories at a time.
Completeness check
Every populated quoteId on a signatory record
resolves to a quote, and every quote maps back to a signatory — zero orphans
in either direction. The client receives the entire roster up front;
“SHOW MORE” only advances an index into an array already in memory, with no
fetch behind it. So 70 comments is the true total, not a first page.
Known gaps
- Two comments are unrecoverable. Jason Wolfe (OpenAI) and Mathieu Lariviere (Meta) have quote records whose text chunks the server references but never sends. Reproduces on a clean re-fetch; appears to be a bug on their end. All comment statistics here are over n = 68.
- Themes are keyword-coded, not hand-labelled — reproducible but crude, and a comment can carry several. The twelve reservations in section 05 were hand-verified by reading.
- Roles are inferred from self-reported titles, 56.9% of which are just a company name. The stated-role facet is a keyword classifier over that text — it describes what people wrote about themselves, not what they do. Anthropic's near-universal “Member of Technical Staff” makes its function mix unrecoverable, which is why the safety-share comparison carries a warning rather than a conclusion.
- The ordering is the site's own. The position of each signatory is the sequence the payload ships in. That it correlates so tightly with seniority, commenting and anonymity is an observation about the published order, not a claim about how or when anyone signed.
- The vocabulary comparison is thin by construction. Terms are ranked by their rate in one employer's comments against their rate across the whole corpus, with a minimum of two uses. With 13–22 comments per employer, a single distinctive sentence can put a word on the list. Counts are shown for exactly this reason; treat it as a reading aid.
- One unverified claim in the corpus. Walker Smith (Meta) asserts “the unintentional model breakout from OpenAI that hacked Huggingface” as established fact. No corroboration found; treat as his characterisation.
- The count is live; this page is not. Press coverage earlier the same day cited 1,122; this snapshot is 1,134. Every percentage here describes 28 July 2026 and will drift from the statement site over time.
Separating measurement from argument
Sections 01–06 and 08 are intended to be usable by someone who disagrees with every conclusion in section 07. Where a reading is contestable it is placed in a marked commentary block, and where a label is a judgement rather than a field in the source — the reservation flag, the keyword themes, the stated-role classes — the page says so at the point of use rather than only here. If you find interpretation presented as measurement anywhere on this page, that is a defect worth reporting.
Names, snapshots and removal
This page reproduces 1,134 names and self-reported titles that the statement publishes openly, and makes them searchable in ways the source does not — the original paginates twenty at a time, while this one filters and lists. That is a real difference in exposure even though the underlying data is public, and it is a deliberate choice made so the roster can be examined as a whole.
Two consequences follow. Signatories who signed anonymously are shown only as anonymous, and no attempt is made here or anywhere in the data to identify them. And because this is a frozen snapshot, a name removed from the live statement would persist here until this page is rebuilt — so anyone who wants their entry removed or corrected should be able to reach the operator of this site directly, and the snapshot should be refreshed periodically against the source.