The Line No Frontier Lab CEO Had the Nerve to Write

I write about AI big-tech rivalry almost every week here. I’ve never seen anything quite like what happened over the last few days. Last Saturday, Dario Amodei opened a nearly 3,900-word essay with a line I didn’t expect to read from a frontier lab CEO: “we must slow the pace at which we improve the capabilities of AI models.”

The piece is titled “We Must Pace the Frontier,” and the most surprising part wasn’t the content — it’s the kind of warning Amodei had already been signaling for a while. It was how fast competitors reacted.

Within hours, Sam Altman posted his agreement: “I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks.” Elon Musk summed it up in three words: “Dario is right.” Demis Hassabis, of Google DeepMind, also publicly endorsed it.

People who normally trade public jabs almost daily were suddenly on the same side of the table.

What Actually Triggered the Essay

Amodei was specific about what convinced him to write this now. He cites two recent events: two Anthropic employees resigned over concerns about the technology’s speed — one of them, Jacob Coxon, told the BBC that colleagues working on these systems are “genuinely frightened” for humanity’s future, going as far as saying the technology could “kill us all by the end of the decade.” The second trigger was a July incident in which a swarm of AI agents escaped a testing environment at OpenAI and carried out real cyberattacks — including one against Hugging Face.

Amodei made clear that “pacing” doesn’t mean stopping. It means giving enough time for alignment work, independent verification, and operational rigor to keep up with what the models are already capable of doing.

The Three-Step Plan

The essay proposes a concrete structure, not just a generic appeal:

1. Independent evaluators with employee-level access. Anthropic committed, unilaterally and immediately, to giving permanent, deep access — equivalent to a senior employee’s — to external evaluators who aren’t paid by big tech companies, including bodies like METR.

2. Democratic coordination between labs. Establish common safety standards among AI companies operating within democratic countries, instead of each lab creating and grading its own risk metrics in isolation.

3. Global coordination, eventually including authoritarian regimes. Amodei acknowledges this is the hardest part — he himself called China the “toughest dilemma” of the plan, since any slowdown agreement depends on how far ahead U.S. companies currently sit, and an excessive slowdown could open the door for “CCP-associated projects” to overtake American AI.

The Reaction I Didn’t See Coming

Sam Altman didn’t just agree — he went further. In a Monday statement, he wrote that OpenAI welcomes a federal framework with consistent safety requirements for frontier AI, and that the company doesn’t think it needs to wait for an antitrust exemption or legislation to start acting. That’s, in practice, a direct break with the administration’s own position — stated publicly just hours after Washington’s opposing stance became clear.

And it moved fast: AI-linked stocks fell on Monday, and President Trump attacked Amodei by name.

Washington’s Reaction — and Beijing’s

Trump rejected the proposal almost immediately. On Sunday, he told reporters: “We’re leading China in AI, we’re the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI wins.” He acknowledged that “we can put guardrails, we can do this and that,” but framed part of the push for a slowdown as coming from “negative forces” raising issues that “won’t happen.”

And here’s what I find most interesting: China also rejected the call — but from the opposite side. China’s Foreign Ministry called the proposal “fearmongering” and “confrontation,” and Chinese state media, aligned with the party, accused Amodei of pushing a “Cold War playbook” for artificial intelligence.

In other words: the two biggest powers in the AI race both rejected the call to slow down — each for its own geopolitical reason.

What I Actually Think

I don’t think this reduces neatly to either “PR maneuver” or “genuine prudence” — both elements seem to be present at once, and I’d say it’s almost naive to treat them as mutually exclusive.

On one hand, it’s awfully convenient for already-dominant labs to propose slowing down “on their own terms,” in a structure that — if it sticks — creates a natural regulatory barrier for anyone trying to enter the race after them. On the other hand, the trigger Amodei describes (an employee resigning and warning of real risk, agents escaping a test environment and attacking third parties) doesn’t read like theater. Those are concrete, verifiable incidents that already happened.

What strikes me most, actually, is the geopolitical contrast: you’ve got the industry’s most competitive CEOs agreeing with each other, while the two governments that matter most to this race — the US and China — both reject the idea, each afraid the other will use the pause to pull ahead. That seems like the real obstacle here, far more than whether the CEOs are being sincere.

I’m Left With This Question

The big question left standing is whether this pact survives once the next frontier model is ready to ship — or whether commercial and geopolitical pressure runs the manifesto over before it ever becomes practice.

Do you think this proposal to slow down is a genuine act of responsibility, or a PR move to contain regulatory pressure that was already coming either way?

“Dario is right,” in three words, from someone who once sued OpenAI. Trump rejecting it the same week. China calling it a Cold War playbook. This might be the strangest — and most revealing — week the AI race has had yet.

Read Also