Product1 publisher2 min readPublished
Amodei's pacing plan leaves two of its three steps to parties Anthropic does not control
Altman and Hassabis backed the essay within a day, the president answered by citing his criminal and regulatory power over AI companies, and nothing on the record changes what ships to developers this quarter.
The Product Desk · Product desk
What happened
- Amodei's Saturday essay argues for slowing AI development and sets out three steps: embedded third-party evaluators, standards coordination among frontier companies in democratic countries, and government-to-government pacing talks.
- Anthropic is unilaterally committing to the first of those steps, evaluators who verify whether a company sticks to its safety practices and commitments and who can report incidents.
- Altman said shortly afterwards that he agrees with Amodei on pacing the frontier, and committed to having independent evaluators at OpenAI, which he called a great idea.
- Hassabis replied that the essay points towards the right path forward, and referred back to the framework for an AI standards body he published in July.
- Trump posted that his administration has stopped AI people from doing bad things, naming Amodei and Anthropic and saying Amodei is now pretending to be a perfect little angel.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Only one of the three steps sits inside a company that can act alone. A customer expecting the pacing agenda to change release cadence is waiting on rival labs and on governments, and no government has signed anything.
- exposure Teams whose production path runs through Claude now have a vendor in a public dispute with the administration that regulates it, and that goes to legal review.
- decision Two vendors have now said in public that outside evaluators should see their safety work. Renewal is the moment to test whether the commitment includes telling customers what an evaluator finds.
- contradiction OpenAI asked for federal safety requirements while the vice president described that kind of request as a trojan horse, so a buyer cannot forecast US rules from what the labs say they want.
A team that ships on Claude or GPT reads an essay like this looking for two things: whether the model version they call in three months arrives later, and whether they hear anything useful when a model misbehaves.
On the first, the authors closed the question themselves. Amodei wrote that pacing "does not mean halting model training or technical progress, ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this" [4]. Altman's version was that progress should be "slower than it otherwise could be" [9]. House Speaker Mike Johnson ruled out the harder option from the government side, telling CNN's Jake Tapper on Sunday that "We cannot put a moratorium on this because China will overlap us" [18].
The second and third steps are where the plan stops being one company's to keep. Amodei asks for coordination between frontier AI companies in democratic countries on standards and limits, and for global coordination between democratic and authoritarian governments on pacing, a step he qualified with the phrase "to the extent this is possible" [2]. One of the three steps is inside a single company. Two require agreement from rivals or from states [20]. Demis Hassabis, who backed the direction, put the gap plainly: "The details need working through, but the direction is correct for meeting this critical moment," he said, pointing to a framework for an AI standards body he published in July [12][13].
The White House replied by naming the president as the guardrail. Trump posted that the only control or guardrails AI needs "is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!" [15].
For a buyer, the weekend raises two questions. Does the promise bind one vendor, or does it need many parties? Does it change what ships to you, or what you get told? The evaluator commitment is single-vendor and informational, and it is the only item with a named party attached. The standards body and the intergovernmental talks are multi-party, and neither exists yet.
That leaves a contract question. Whether the evaluator commitment and its incident reporting appear anywhere in your agreement, and whether an incident reported to an evaluator also reaches the customer running the model in production, is the part a vendor can answer in writing this week. The Verge's roundup does not say who the evaluators would be or when they would start [22].
What to watch
- Whether Anthropic names its embedded evaluators and publishes their access terms and incident-reporting path.
- Whether OpenAI's request for a federal framework produces bill text, given that Vance called that kind of request a trojan horse.
- Whether the standards body Hassabis floated in July gets a second frontier lab to sign on.