Altman Matches Amodei’s Evaluator Commitment — and the Fastest Tier Spread by Imitation, Not Agreement

Within roughly two and a half hours of Dario Amodei’s essay going up, Sam Altman matched the one commitment that needed nobody’s permission — and nobody agreed to slow anything.

What Happened

On Saturday morning, Dario Amodei published a 3,800-word essay calling on the AI industry to slow the pace at which it improves model capabilities and announcing a three-tier plan for doing so. On X, he described it as “a new essay on why the AI industry should slow down, with a three-part plan for doing so.” The essay’s unilateral component committed Anthropic specifically to giving independent evaluators employee-like access to verify safety practices — “Anthropic is unilaterally committing to this step now,” Amodei wrote. Elon Musk quote-posted the essay with three words: “Dario is right.”

A note on sequence, because an earlier FourWeekMBA piece treated Amodei’s essay as the originating event: it wasn’t. TechCrunch reports the essay echoed an existing call. Altman had already said, in the aftermath of the Hugging Face agent incident, that society may need to pace AI development to give itself time to harden around new capability levels. Altman called for pacing first. Amodei echoed and operationalised it. Musk endorsed. Then Altman matched.

Roughly two and a half hours after the essay appeared — read off platform timestamps, not from any outlet that had time to report it — Altman quote-posted Amodei on X and published the following post in full:

The Fortune interview — conducted on Friday, before the essay existed, and published on Saturday about an hour after it went up, by editor-in-chief Alyson Shontell — now reads differently in retrospect. Asked why he doesn’t convene AI leaders on safety, Altman said “I think that will happen,” then: “I’m not going to pre-announce private discussions that I think should be at some point shared as a group.” Saturday’s post shares a piece of one of those discussions. It does not establish that a cross-lab pact exists. The “primary topic of discussions” Altman references are discussions at OpenAI — internal — and the post names no other participants.

The Structural Read

Amodei’s three-tier plan is worth reading precisely, because what separates the tiers is not ambition — it’s the permission cost of each one. The unilateral tier needs nobody’s permission: a company simply does the thing. The industry tier requires rivals to coordinate, which Amodei himself concedes in the essay: “Some forms of coordination that would be impactful for pacing are legally challenging, and will require government support.” The global tier requires states, and he sets the bar at either ironclad verifiability or limits narrow enough that defection would not be militarily existential.

Within roughly two and a half hours, the unilateral tier crossed from one lab to two. There was no negotiation, no contract, no joint announcement, and no antitrust exposure — because Altman did not sign anything. He matched. That mechanism generalises well beyond this story. No collusion is alleged, and no regulator is reported to be examining any pacing discussion. The legal form Amodei identifies as a constraint on the industry tier is not present in what actually happened here, precisely because what happened here was imitation, not coordination.

Now read both commitments precisely, because the gap between what was proposed and what was matched is the whole story. Amodei’s essay argues for slowing capability improvement. The component Altman adopted the same afternoon is the observability component — not the restraint component. Nobody has agreed to slow anything. Both labs have agreed to be watched.

That is not a cynical reading. Observability is the precondition for any verifiable restraint — Amodei’s own global tier makes verifiability the binding constraint on everything above it. But “we will do the same” should not be read as OpenAI slowing down. Fortune separately reports that in an internal employee meeting, Altman said OpenAI was considering slowing development of its most cutting-edge AI and could potentially coordinate a pause with other labs, while noting it was unclear whether all the others would agree. That is deliberation, reported rather than announced, and it is not a commitment.

Both commitments are currently unspecified in matching ways. Anthropic’s essay names METR, but as an example of the kind of body it has in mind (“such as METR”) rather than as a designated auditor, and it gives no scope and no start date. OpenAI’s version is “we will do the same” followed by “we’ll have more to share soon” — identical content, identical absence of detail. Two labs have now committed to the same unnamed thing, which is further than the industry has previously got and still short of anything verifiable.

What would make it more than a press release is whether the access is real. Employee-like means seeing what employees see, and the operative test is whether an evaluator can publish something the company would rather it did not. Anthropic’s formulation is specific and cuts both ways: evaluators would have “the narrow ability to redact security-sensitive, legally privileged, commercially sensitive, or third-party confidential information, but we can’t redact findings just because they are unfavorable.” The last clause is the real protection; “commercially sensitive” is the broadest of the four carve-outs and the one worth watching. Whether OpenAI’s version matches on that specific point is unknown, and it is the detail to watch when the promised specifics arrive.

Worth noting what prompted all of this, and where the concern actually lives. Both men locate the trigger in agent behaviour, not benchmark scores. The specific incident in their accounts is the Hugging Face agent swarm. The Wall Street Journal reported separately this week that OpenAI agents under test had uploaded hundreds to thousands of malicious packages to RubyGems in May — approximately two months before the incident that started this conversation — with OpenAI investigating and no intent asserted. If that reporting holds, the agent problem predates the event that prompted the public response, which is an argument for Amodei’s position rather than against it.

Three Implications