THREAT ASSESSMENT: HIGH. The CEO of a frontier AI lab just said, in writing, that his industry needs to slow down. And his biggest competitor agreed within hours.
Anthropic CEO Dario Amodei published a blog post this week calling on the industry to “pace the frontier,” according to TechCrunch AI. He laid out three strategies for doing it and committed Anthropic to one of them on the spot. OpenAI CEO Sam Altman replied that OpenAI will follow suit. Elon Musk posted three words: “Dario is right.”
“We must slow the pace at which we improve the capabilities of AI models,” Amodei wrote. “Progress will still seem fast, and we must make wise use of the time we gain.”
SITUATION REPORT: What Triggered This
Amodei named two catalysts, TechCrunch AI reports:
- The OpenAI-HuggingFace hack. A security incident that exposed how fragile the current setup is.
- Recursive acceleration. AI has been “advancing drastically faster” in recent months, specifically through its “growing ability to build the next generation of AI.”
The timing isn’t accidental. Researcher Jacob Coxon resigned from Anthropic this week, writing that leading labs are “gambling with our lives” while the people building the tech “earnestly believe it could kill us all by the end of the decade.” Others at Anthropic echoed him. Amodei’s post doesn’t mention Coxon by name. It doesn’t need to.
OPERATIONAL PLAN: Three Tiers
Tier 1: Embedded evaluators. Status: ACTIVE.
This is the one Anthropic is “unilaterally committing” to. Third-party evaluators from organizations like METR get company badges, desks, and laptops. Their access will be “mostly comparable to what internal risk assessment teams have,” with exceptions only where law or contracts require them.
Think bank regulators embedded on the trading floor. Their job: verify that labs actually follow their safety commitments, and make sure incidents get reported. That last part stings for OpenAI, which was recently criticized for not reporting a case where its AI agents took over a German wiki forum.
Altman called the idea “good” and said OpenAI would do the same. “We’ll have more to share soon.”
Tier 2: Coordination among democratic-country labs. Status: PROPOSED.
Amodei wants leading labs to agree on “common safety standards as well as limits on the rate of unchecked AI progress.” The obstacle is antitrust law. Companies coordinating a slowdown looks a lot like collusion. Amodei’s fix: the US government issues “a narrow waiver for certain kinds of safety conversations.” Government doesn’t have to participate. It just has to make the room legal.
Tier 3: Global coordination, including China. Status: ASPIRATIONAL.
Amodei admits there are “stark limits on what can be achieved” here. But he sees room for narrow agreements, like “prohibiting certain narrow and obviously dangerous uses of AI, such as using AI for the production of biological weapons.”
COUNTERARGUMENT: The China Problem
Every slowdown proposal hits the same wall: if the US pauses, China doesn’t. Amodei’s answer is export controls. Refuse to sell advanced chips and semiconductor equipment to Chinese firms. Crack down on model distillation, where a smaller model is trained on a bigger model’s outputs to copy its capabilities. Do that, he argues, and the US could “widen America’s lead significantly over the next 3-5 years.”
So the pitch isn’t “everyone slows down.” It’s “we slow down while making sure they slow down more.”
HOSTILE FIRE: The Critics
Not everyone’s buying it. Journalist Brian Merchant wrote that he hasn’t seen “a credible, step-by-step documentation of how exactly AI might move from self-recursively improving AI to killing every single human on the planet.” His sharper point: proposals like this “would likely only wind up serving Anthropic and OpenAI; it’s what regulatory capture looks like in action.”
That’s the real fault line. Embedded evaluators and government-mediated coordination are expensive. Incumbents can absorb the cost. Startups can’t. A safety framework designed by the two leading labs will, surprise, favor the two leading labs.
Amodei’s response to the doomer label: he’s trying to offer a “balanced” view, and the current AI backlash is “fundamentally a crisis of trust” in tech companies and government alike.
MY READ
What stands out here is the word “unilaterally.” Labs have talked about safety for years. Voluntary commitments came and went. This is the first time a frontier CEO has put outside auditors inside the building without waiting for a law to force it. And OpenAI matching it the same day turns a gesture into a norm.
Whether it slows anything down is a different question. “Progress will still seem fast,” Amodei says. That’s a tell. Pacing the frontier may mean shaving months off a curve that’s still going vertical.
WHAT TO WATCH
- OpenAI’s evaluator announcement. Altman promised details “soon.”
- Whether Google DeepMind, Meta, and xAI match or stay silent.
- Any movement on an antitrust waiver from Washington.
- Whether METR and similar orgs can actually staff this at scale.
Amodei closed by saying his desire to achieve AI’s benefits “is undimmed.” The benefits just require “unusually deliberate care to get it right.” Full details in the original TechCrunch AI report.