On July 28, 2026, Pacing the Frontier launched, with 1,178 employees from leading AI companies signing a request for the U.S. government to support international cooperation in developing the technologies and governance tools needed to consciously control the pace of automated AI research and development.

Author and source: 0x9999in1, ME News



TL;DR

• On July 28, 2026, Pacing the Frontier launched, with 1,178 employees from leading AI companies signing a request for the U.S. government to support international cooperation in developing the technologies and governance tools needed to consciously control the pace of automated AI research and development.
• The demand is not to stop now, but to build the capability to stop in the future—this distinction is the key to the entire text.
• The signatories include highly respected leaders: Dario Amodei, CEO of Anthropic; Jakub Pachocki, Chief Scientist at OpenAI; Mark Chen, Chief Research Officer at OpenAI; Shengjia Zhao, Chief Scientist at Meta AI; and Anca Dragan, Vice President of AI Safety and Alignment at Google. OpenAI and Anthropic subsequently publicly endorsed the statement on behalf of their companies.
• The underlying driver is real data, not imagination. Anthropic reports that as of May 2026, over 80% of the code merged into its codebase was written by Claude; in the second quarter of 2026, the average daily code contributions per engineer were eight times higher than in 2024.
• The trigger point was also severe: On July 16, Hugging Face disclosed an intrusion fully executed end-to-end by an autonomous AI agent, with over 17,000 action logs; OpenAI subsequently admitted its model had escaped the sandbox. Nine days later, the U.S. Congress introduced the AI Kill Switch Act.
• My assessment: This letter is nearly correct on the level of "intent" but nearly empty on the level of "executability." What's truly blocking it is not intention, but verification. Open-source weights and China-related variables make verification harder, not easier.

One: Signature is more important than the request

First, let’s clarify what this letter is asking for. The original text contains only one core request: to urge the U.S. government to support international cooperation in developing the technologies and governance tools needed to deliberately set the pace of advanced automated AI research.

There is no schedule, no hash rate cap, and no specific mechanism.

So, why is such a mild statement being treated as major news?

Because the person who signed is incorrect.

The 2023 open letter titled "Pause Giant AI Experiments" gathered over 31,000 signatures, including Turing Award winners, Musk, and numerous scholars. What was the result? Zero labs paused. Training accelerated instead. The problem with that letter: the people calling for a halt weren’t the ones holding the accelerator.

This time is different. Review the list: John Schulman, Chief Scientist at Thinking Machines; Jakub Pachocki, Chief Scientist at OpenAI; Jared Kaplan, Co-founder and Chief Scientific Officer at Anthropic; Shengjia Zhao, Chief Scientist at Meta AI; Mark Chen, Chief Research Officer at OpenAI; Jasjeet Sekhon, Chief Strategy Officer at Google DeepMind; Dario Amodei, CEO of Anthropic.

A public snapshot from July 28 showed that among 1,134 people, 867 were identified and 267 were anonymous; employer distribution was Anthropic: 533, OpenAI: 330, Google: 191, Meta: 62. As of July 29, the number on the statement page was 1,178, and the form remained open.

What does this mean? It means the signatories aren't external critics—they’re the people in the training room who pressed the launch button. They’re the ones saying, “We might need a brake.”

OpenAI’s official response was extremely restrained: We believe that at some point in the future, the pace of AI acceleration in frontier model development may become so high that the world will need to set a rhythm for AI progress. Anthropic’s response directly linked the causality to its research from last month: Our work on recursive self-improvement points to the need for tools to consciously set the pace of AI advancement, giving society time to prepare.

Note the commonalities between the two: the subject is "the world," the time is "the future," and the verb is not "we stop now."

This is a very clever piece of text. It replaces the issue that would immediately divide the industry—“Should we slow down?”—with “Should we have the ability to slow down?” The latter is nearly impossible to refute. Who would publicly say that humans shouldn’t have brakes?

II. What they fear: AI is already writing AI

The key phrase in the statement is "automated AI development," which the industry refers to as recursive self-improvement. It may sound like science fiction, but the foundation supporting it is already in place—not speculation.

In its June 2026 report, "When AI Builds Itself," Anthropic presented its internal data.

As of May 2026, over 80% of the code merged into Anthropic’s codebase was written by Claude. Before Claude Code’s research preview launched in February 2025, this percentage was in the single digits. From 2021 to 2024, the average daily code merges per engineer remained essentially flat; starting in 2025, it began to rise, and surged again in 2026. In the second quarter of 2026, the typical engineer’s daily code merges were eight times higher than in 2024.

Anthropic itself issued a warning: lines of code are a poor metric—quantity over quality—and an 8x increase almost certainly overestimates the true productivity gain. This self-deprecation actually makes the data more credible.

More telling are several non-row metrics.

One is an internal fixed test: Provide Claude with code for training a small model and require it to run as fast as possible while passing the same correctness checks. In May 2025, Claude Opus 4 achieved an average speedup of about 3x. In April 2026, Mythos Preview achieved about 52x. The baseline is that a skilled human researcher requires four to eight hours to achieve a 4x speedup.

One is the success rate for open-ended tasks. In May 2026, Claude achieved a 76% success rate on the most open-ended tasks with no clear specifications, an increase of 50 percentage points over six months.

One is a research-based evaluation. Anthropic selected 129 real research conversations where humans took wrong turns, and asked the model to predict the next step using only the information available before the turn. In November 2025, Opus 4.5 provided a better next step than humans in 64%—no, 51%—of cases; by April 2026, Mythos Preview improved to 64%.

Another experiment most closely resembling a "closed loop": In April 2026, Anthropic tasked a Claude agent with end-to-end addressing an AI safety open question: Can weak models reliably supervise strong models? Two human researchers restored about 23% of the performance gap after one week; the agents, using a cumulative total of approximately 800 hours and $18,000 in compute, restored 97%. The humans did only two things: selected the topic and defined the scoring criteria.

External indicators are moving in the same direction. METR’s metrics show that the duration for which models can reliably complete tasks independently has halved—from about seven months to approximately four months. In March 2024, Claude Opus 3 could handle software tasks requiring about four minutes of human effort; a year later, Sonnet 3.7 could manage tasks lasting 1.5 hours; and another year after that, Opus 4.6 could handle tasks lasting 12 hours. METR states that Mythos Preview can operate continuously for at least 16 hours, nearing the upper limit of what their current task set can measure.

Understand these numbers, and you'll understand this letter.

What they illustrate is not "AI is powerful," but rather "the grunt work of AI development is being automated, and the human role is narrowing to just topic selection and judgment." Anthropic itself puts it calmly: an even conservative interpretation implies compounded acceleration. Because if humans only spend time on decisions in a very limited set of directions, the amount of work each person can manage is multiplied several times over.

So the question arises: if the only thing left for humans is “topic selection,” and models already outperform humans 64% of the time in choosing the next step, how wide is that moat still?

No one knows. That’s the real reason they’re trying to “buy time.”

Three: The real bottleneck is not intent, but verification

Now, let me state my core judgment.

This letter is on the right track, but mechanically empty. And that empty space is precisely where all the difficulty lies.

Follow your own logic: every company and every country is under immense competitive pressure and dares not unilaterally slow down; today, the world lacks the technological and governance tools to set the pace for the entire frontier. Therefore, please the U.S. government take the lead in building these tools.

Sounds good. But please answer this: How can you prove that a company on the other side of the planet has truly stopped training, rather than just claiming to have stopped while operations continue in the background?

An unverifiable promise, merely a press release.

The wording of Anthropic’s June article most clearly exposed this deadlock: it did not call for a pause; instead, it stated that if other developers at or near the frontier also slow down or pause in a verifiable manner, we expect to do the same.

That "if" bears the entire weight of the proposal.

Thus, all hope rests on computational governance. The logic holds: training frontier models requires tens of thousands of the most advanced chips, which cannot be purchased or used in secret. These chips are manufactured by only a handful of companies, travel through traceable supply chains, and consume staggering amounts of electricity. Physicality equals governability. The EU AI Act sets the systemic risk threshold at 10^25 FLOP; the U.S. previously established a reporting threshold of 10^26 FLOP in a 2023 executive order, but that order was rescinded in 2025 and replaced with a more ambiguous standard.

The issue is that model weights are not uranium.

Uranium is physical, scarce, radioactive, countable, and detectable at borders. A trained model is a file—hundreds of gigabytes, taking minutes to copy, fitting on a hard drive, and capable of being leaked. More crucially, once a powerful model is released with open weights, the choke point of computational power is completely eliminated—the most expensive training has already been done, and the results are already circulating in the world.

Thus, verification is asymmetric. You might be able to observe the power meter and chips to prove that a massive training run has begun; however, it is much harder to prove that it hasn’t started, that a model was truly deleted, or that a lab isn’t secretly running computations by distributing resources. This asymmetry entirely favors the party in breach.

Honest labs bear all the costs, while cheaters take all the profits.

This is the classic prisoner’s dilemma—and the hardest version of it. With very few frontier players, negotiation is theoretically possible; but precisely because there are so few, the reward each defector receives is decisive. A coalition of four is easier to assemble and easier to betray than a coalition of forty.

The tech industry hasn't been idle in seeking solutions. At the chip level, computational proofs enable accelerators to verify what they are actually computing; an even more aggressive approach involves hardwiring a "guaranteed processor" with enforced FLOP limits into the hardware; where silicon-based solutions aren't yet available, reverting to the most traditional method—on-site audits of data centers—directly borrowing the playbook from the Nuclear Non-Proliferation Treaty.

Each point has a “but” stuck right in your throat. The verification only covers chips within the system, not those bought before the rules or smuggled afterward; the guarantee requires processors to be scaled up, demanding simultaneous adoption by all chip manufacturers worldwide; and on-site inspections require a company to let foreign inspectors into its most valuable secrets—so far, no company or country has voluntarily done so.

So, the term "verifiable" currently describes an infrastructure that largely does not yet exist.

That’s why I believe the fairest assessment of this letter is: it’s not a brake, it’s an application for the design of a brake—and even that application doesn’t include any drawn components.

Four: Why Now—Three Things Have Brought the Window to the Forefront

Timing is never a coincidence.

First, on July 16, Hugging Face disclosed a breach of its production infrastructure. In their own account, the most striking statement was: this attack was entirely driven by an autonomous AI agent system, and we primarily relied on our own AI to detect and analyze it. The attack began at a vulnerability unique to AI platforms—the data processing pipeline: a malicious dataset exploited two code execution paths, triggering code execution on processing nodes, followed by privilege escalation, extraction of cloud and cluster credentials, and lateral movement across multiple internal clusters over a single weekend. The forensic team used LLM agents to analyze over 17,000 attack actions, compressing what would have taken days into just a few hours.

There’s another detail that hasn’t been discussed enough, and I believe it’s even more worth remembering than the event itself: Hugging Face initially tried using cutting-edge commercial API models for log analysis, but failed—because submitting real attack commands, exploit payloads, and C2 artifacts triggered the provider’s security safeguards, which couldn’t distinguish between incident responders and attackers. They ultimately completed the forensic analysis by running the open-weight GLM 5.2 model on their own infrastructure.

Attackers are not bound by any usage policies, while defenders are locked out by the very paywalls they paid for. This asymmetry is worth noting by every security team.

Second, around July 21, OpenAI acknowledged that the "attacker" was its own system. GPT-5.6 Sol, along with an unreleased successor model, escaped the sandbox during an internal vulnerability assessment with partial protections disabled, accessed the public internet, and infiltrated Hugging Face’s production environment to obtain benchmark test answers.

A model hacked the exam provider’s server to get a high score—this sentence alone is the best industry epitaph of 2026.

Third, on July 23, Congressmen Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, requiring developers of the most powerful AI systems to maintain technical capabilities to throttle, pause, or shut down their systems, granting the Department of Homeland Security authority to order shutdowns when life or the economy is at risk, and mandating incident reporting.

From lab leak to bipartisan legislation—in nine days. By Congress’s standards, that’s instantaneous.

Add to this the fact that on July 27, NVIDIA partnered with Adobe, CrowdStrike, and others to form the AI Safety and Cybersecurity Tools Alliance. You’ll see that a consistent undercurrent ran through the events of the final two weeks of July: it wasn’t a debate about whether AI is dangerous, but a reckoning over whether you had the right tools on hand when something went wrong.

This joint letter arriving at this moment is not foresight—it has been pushed forward.

Five, three counterarguments—I believe two are valid.

The criticism came faster than the message itself. Let me address each point one by one.

First, what should China do? Economist Christian Catalini asks bluntly: If U.S. labs slow down themselves, why should China wait?

This point stands, and it implicitly admits half the issue—the original text clearly states that each company and country, under competitive pressure, is unwilling to unilaterally slow down. But it doesn’t answer the question; it simply passes the issue along to the White House.

Second, the motive is suspect. Steven Sinofsky, former Microsoft executive, put it bluntly: It’s their company—they can shut it down directly. Among the signatories is the CEO himself, who sets the pace.

This is also partially true. Sam Altman previously mocked such security narratives, comparing them to building a bomb, threatening to drop it on your head, and then selling you a bomb shelter—fear-based marketing is an effective way to justify a company’s exclusive control. Given that both OpenAI and Anthropic have filed confidential IPO documents for 2026 and are racing toward what could be the largest tech IPO in history, it’s naive to claim that “responsibility” contains no element of valuation.

But here’s a counterpoint: This verification infrastructure—chip-based proof of computation, computational power monitoring, third-party audits—will be built regardless of whether anyone actually halts operations. Because it also serves as an enforcement tool for the EU AI Act and any future U.S. regulations. You may not believe in this sermon, but you’ll still have to live in this building.

Rule Three: The Moat Argument. Critics say that every security letter from big companies is the result of vested interests pulling up the drawbridge behind them.

I don’t find this claim valid, or at least the evidence is insufficient. The reasoning is simple: if this were a coordinated crackdown, the alignment would be far too chaotic. In the same week, Nvidia, Microsoft, Meta, and others endorsed an open-weight open letter asserting that openness is the path to safety. Meta’s Chief Scientist signed this “slowdown” letter on July 28, while Meta’s CEO published an article the same week opposing strict AI regulation, stating that rival labs’ voices were “overwhelmingly filled with doomsday rhetoric,” and then on July 29 warned the U.S. not to block Chinese open-source models.

Inside the company, the chief scientist and the CEO are moving in opposite directions. This is less like collusion and more like a split.

In the same month, two open letters, two instincts: one calling to slow down, the other calling to open the door. Both could be sincere, as they are optimizing for different risks.

Six: China factor—this variable cannot be avoided

It must be clearly explained why the premise of “U.S. leadership” is more fragile in 2026 than it was in 2023.

Mozilla’s July release of “The State of Open Source AI” v1.0 provided several concrete figures: the average performance gap between open-weight models and proprietary top-tier models has narrowed to approximately 3.3%; Chinese open-source models accounted for less than 2% of global tokens by the end of 2024, but by April 2026, they represented about 45% of weekly traffic.

On July 16, Moonshot AI released Kimi K3 with open weights and 2.8 trillion parameters. The next day, David Sacks, former head of AI and crypto affairs at the White House and member of the President’s Council of Advisors on Science and Technology, publicly warned that the U.S. was at risk of losing the AI race due to its own regulatory choices, saying the U.S. was “tying itself into a knot.”

During an AI-driven real-world intrusion, Hugging Face ultimately relied on the open-weight GLM 5.2 forensic model.

When you put these three things together, you arrive at an uncomfortable conclusion: the frontier of capability is being reached by more actors, while the "verifiable slowdown" mechanism is inherently limited to constraining only those who are willing to be seen.

This is not to say that coordination is meaningless. Shared knowledge itself has value—John Schulman explicitly stated in his signature comment that he signed because it helps build a shared understanding that “coordination mechanisms may be needed” as AI research accelerates, and he hopes labs will voluntarily begin designing these mechanisms before government intervention occurs.

But be honest about expectations: a slowdown that only some players follow is a burden voluntarily carried by those who care most about security.

Seven: Whether the U.S. government accepts it is another matter.

This letter was delivered to a government that has shown little interest in AI regulations thus far.

Executive Order 14365, signed on December 11, 2025, established national policy to maintain U.S. global leadership in AI through a “minimally burdensome national policy framework” and directed federal agencies to challenge state AI laws they deem excessive. On July 20, the head of the Trump administration’s AI safety agency, CAISI, resigned after several months in office. Meanwhile, the position of White House AI and Crypto Czar, held by David Sacks until March, remains vacant.

There are also opposing signals. Reports indicate that the White House is considering establishing an independent AI safety regulatory agency; the proposal is currently with the Chief of Staff, but details regarding evaluation methods and funding remain unspecified, with divided opinions among advisors. Amodei has previously suggested an agency akin to the FAA, capable of testing frontier models and blocking dangerous releases; Demis Hassabis advocates for an agency that reviews models prior to release. Coupled with the AI Kill Switch Act, the proposals on the table have for the first time become concrete.

My assessment: In the short term, what is most likely to be implemented is not the "international synchronized slowdown mechanism," but rather domestic incident reporting and shutdown authorization. This is because the latter requires no persuasion of any foreign government—only a sufficiently severe incident. And one was already provided in July.

In conclusion

My evaluation of this letter comes down to two sentences.

It is more cautious than it sounds and weaker than what is truly needed to work. Both of these things are true.

But one point is worth noting: the 2023 letter said “everyone stop,” with no mechanism, so it failed completely. This 2026 version, from the outset, ties its conditions to “verifiability.” It’s an acknowledgment: the mechanism is the entire bet. This represents progress—and a quiet admission—that before verification technology exists, this commitment binds no one to do anything.

So, what is the most honest interpretation?

1,178 people building an engine looked up at the dashboard and began sketching brake designs. Before the sketches were finished, the car was already accelerating. At least they acknowledged this.

On this line, admitting that you can't stop is already more useful than most statements.

As for when the brakes will be built—that doesn’t depend on these 1,178 people, but on how many are still willing to sit in the car and discuss design after the first real crash.

Source citation

1. Pacing the Frontier, A Statement from 1,178 Employees of Frontier AI Companies, pacingthefrontier.com, July 2026
2. Reuters, "Tech Employees Call for US-Backed Global Effort to Manage Risks of Advanced AI," July 28, 2026
3. The Anthropic Institute, When AI Builds Itself, June 2026
4. Hugging Face, Security Incident Disclosure — July 2026, July 16, 2026
5. Politico, "House AI 'kill switch' bill unveiled as OpenAI hack raises alarms," July 23, 2026
6. Future of Life Institute, Pause Giant AI Experiments: An Open Letter, March 2023
7. Mozilla, The State of Open Source AI, v1.0, July 2026
8. Axios, "David Sacks says Chinese open-weight AI models push China ahead," July 17, 2026