More than a thousand people who actually train and ship frontier models just asked Washington for something specific: not “stop AI,” but tools that could pace it if automated research starts compounding faster than oversight.
On July 28–29, 2026, Pacing the Frontier published a statement from 1,178 employees of frontier AI companies. Within hours, OpenAI and Anthropic endorsed it as companies — Anthropic’s post (reposted by co-founder Jack Clark) says the petition is signed by its CEO, several co-founders, and senior staff, and ties the ask to Anthropic’s own recursive self-improvement research. Meta’s chief scientist signed as an individual while Meta’s CEO published a same-week essay arguing the opposite instinct on access (Zuckerberg’s WSJ piece).
TL;DR
| Question | Answer |
|---|---|
| What do they want? | US-backed international work on technical + governance tools to pace automated AI R&D |
| What don’t they demand? | An immediate, unilateral pause on all training |
| Why now? | Labs say they may be close to AI that automates AI research |
| Who signed? | 1,178 staff — including chiefs of science at OpenAI, Anthropic, Meta; Anthropic’s CEO; Google safety VPs |
| Company endorse? | OpenAI + Anthropic yes (Anthropic: CEO + co-founders + senior staff); staff site comments still personal capacity |
| Anthropic’s why | Points to June 2026 RSI / “AI builds itself” research |
| Read it | pacingthefrontier.com |
The ask in one paragraph
The letter’s logic chain:
- AI could be dramatically good — not guaranteed.
- Leading companies believe they may be close to automating AI research.
- That could accelerate capability past understanding/control.
- Society may need to buy time for security and oversight.
- No firm/country will slow alone under competition.
- Today the world lacks tools to deliberately pace frontier-wide progress.
- Therefore: US government should support an international effort to build those tools.
Exact request line:
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.
What “tools to pace” could mean (concrete)
The letter is deliberately non-prescriptive on mechanisms, but the discourse around it points at families of capability:
| Tool family | Example job |
|---|---|
| Compute / training transparency | Know when frontier runs start; verify claims |
| Eval + incident sharing | Common cyber/bio evals; rapid disclosure norms |
| Verification tech | Prove a lab paused a class of runs |
| Export / access controls | Gate the most automatable R&D stacks |
| International forums | Make unilateral pause less suicidal |
Without those, “please slow down” is a speech act that loses to markets. With them, pacing becomes a coordination technology problem — closer to arms-control instrumentation than to vibes-based ethics blogs.
Who is on the list (notable names)
From the public signatory preview and reporting (NBC, Verge):
| Person | Role (as listed) |
|---|---|
| Jakub Pachocki | Chief Scientist, OpenAI |
| Mark Chen | CRO, OpenAI |
| Dario Amodei | CEO, Anthropic |
| Jared Kaplan | CSO / co-founder, Anthropic |
| Jack Clark | Co-founder, Anthropic |
| Chris Olah | Interpretability lead, Anthropic |
| Shengjia Zhao | Chief Scientist, Meta AI |
| Dawn Song | VP AI Research, Meta |
| Anca Dragan | VP AI Safety & Alignment, Google |
| Jasjeet Sekhon | CSO, Google DeepMind |
| John Schulman | Chief Scientist, Thinking Machines |
Personal comments on the site lean hard on coordination failure (Leo Gao’s “deadly race” framing) and eval shock (cyber agents discovering real vulns — Dawn Song cites CyberGym / ExploitGym).
Why the cross-lab roster matters
OpenAI + Anthropic + Google/DeepMind + Meta names on one letter is the story for reporters — rivals agreeing that unilateral restraint fails. It is also the attack surface for skeptics: “if you wanted to slow down, you would ship fewer agents.” The letter’s reply is structural: competition makes unilateral slowdown irrational; therefore build shared tools so slowing is verifyable and mutual.
Same-week polarity on explainx.ai:
- Zuckerberg WSJ — AI for everyone (diffuse power)
- NVIDIA × SSI scale-up (quiet lab + compute)
- Mythos cryptanalysis and Codex Security (capability and defense both shipping)
Pacing is not “stop research.” It is “instrument the frontier so society can choose speed.”
What OpenAI said publicly
OpenAI’s corporate post (mirrored on X): at some future point, frontier acceleration may be high enough that the world needs to pace advancement; OpenAI hopes to contribute to US-government-led work with other labs and the open-source community on mechanisms that make that possible. Link hub: pacingthefrontier.com.
That is carefully not “we will slow tomorrow.” It is “build the steering wheel before the engine hits recursive gear.”
What Anthropic said publicly
Anthropic’s corporate post (Jul 29, 2026; Jack Clark reposted):
We support this petition, signed by our CEO, several co-founders, and senior staff. Our own research on recursive self-improvement, published last month, points to the need for tools to deliberately pace the frontier of AI development so society can prepare. We’re glad to see broad agreement across the field.
Two layers matter:
- Personal vs company. The petition site says comments are personal capacity. Anthropic’s @AnthropicAI post is the company layer — CEO Dario Amodei and co-founders (including Jack Clark, Jared Kaplan, Chris Olah, Benjamin Mann on the signatory list) are not just names on a form; the firm is owning the endorsement.
- Research → petition. Anthropic is explicitly bridging its June Institute work on recursive self-improvement — evidence that AI is already accelerating AI R&D inside labs — to the employee letter’s ask for verifiable pacing tools. Same thesis, public petition edition.
For RSI framing elsewhere on explainx.ai, see Weco’s AIDE² RSI ladder (research-agent Level 1 claims) — different lab, same “are we already in the loop?” question.
How this differs from other July letters
| Document | Core ask |
|---|---|
| Pacing the Frontier | Tools to pace automated R&D internationally |
| Open-weights American leadership letter | Keep US open-weight ecosystem competitive / legal |
| Little Tech / Chinese open-weight ban letter | Trade / export-control framing |
| Zuckerberg WSJ op-ed | Distribute “personal superintelligence,” resist concentration |
Same month, opposite vibes: slow the race vs open the gates. Both can be sincere — they optimize different risks.
What “pacing tools” might even mean
The letter is intentionally thin on mechanism. Plausible buckets people will fight over:
- Compute / training monitoring shared across jurisdictions
- Eval gates before releasing models that automate R&D loops
- Treaty-like commitments that only bind if rivals bind
- Kill-switch / pause protocols that aren’t unilateral suicide
Critics on X already shout collusion, regulatory capture, or “stop China somehow.” Supporters say without a coordinated option, safety work is theater while capabilities compound. explainx.ai’s read: treat the letter as common-knowledge building (Schulman’s comment) more than a finished policy bill.
For the broader export-control and open-weight fight, see the AI policy timeline and Anthropic’s open-weights position.
How builders should respond (without waiting for Congress)
You do not need a treaty to improve pacing hygiene inside your company:
- Log autonomous research/agent runs that modify training or exploit pipelines.
- Separate eval sandboxes from production credentials (the HF intrusion timeline is the cautionary tale).
- Share high-severity agent findings through coordinated disclosure, not only marketing blogs.
- Support internal kill-switches for unbounded self-improvement loops.
- Read the letter at pacingthefrontier.com before quoting it second-hand.
The ask is for government-backed international tooling. The interim ethic is: do not make unilateral restraint impossible by hiding your own accelerants. Pair capability launches (Codex Security, Mythos crypto) with explicit risk notes — that is the culture the letter is trying to bootstrap.
Criticism map (so you can steelman both sides)
Critics say: Labs want regulation that entrench incumbents; signatories keep shipping agentic cyber tools; “pace” is PR.
Supporters say: Without mutual verification, restraint is suicide; eval shocks already show agents finding real vulns; governments must fund coordination tech like they fund standards bodies.
Your job as a reader is not to pick a fandom. It is to notice when a company ships an accelerant without a corresponding control note — and when a safety essay ignores diffusion benefits. Hold Codex Security, Mythos, and this letter in the same browser session.
Closing note for explainx.ai readers
This launch moves fast; verify primary docs before you standardize tooling or brief a client. Re-check availability, pricing, and model IDs on the official pages linked above, then tell us what broke in production so we can update the guide. Follow @explainx_ai for follow-ups when the vendors ship the next patch — and prefer measured evals on your own traffic over screenshot economics. If you only needed a headline, you already have it; if you are implementing, the checklists above are the part that saves a weekend.
Related on explainx.ai
- Copilot for Word document AI worm (XPIA)
- NVIDIA × SSI — Vera Rubin + 10× compute (Jul 2026)
- Mythos cryptanalysis — HAWK & AES
- Weco AIDE² — RSI ladder
- Zuckerberg — The AI Future Is for Everyone
- Open-weights American AI leadership letter
- Little Tech Association Chinese open-weight ban letter
- Anthropic on open weights — Amodei
- AI policy timeline 2026
- OpenAI Daybreak / Codex Security
- Codex Security CLI open-sourced
Sources
- Pacing the Frontier
- Anthropic on X — petition support
- Anthropic Institute — recursive self-improvement
- OpenAI on X
- NBC News
- The Verge
Signatory counts and company positions can change as the form stays open. Snapshot as of July 29, 2026.