Amodei Argued for Slowing Down, and Hours Later Altman Committed to One Specific Item: Independent Evaluators With Employee-Like Access

On Saturday, September 12, Anthropic CEO Dario Amodei published a roughly 3,800-word essay, "We Must Pace the Frontier," on his personal site, arguing the industry should deliberately slow the rate at which it improves model capabilities because progress is outrunning researchers' ability to understand and control it. His definition is restrained: "pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this." The essay proposes a three-step plan he says does not sacrifice commercial advantage, with Anthropic claiming to have completed the first step itself and the final cross-party coordination step described by Amodei as the hardest to execute; he also warns that swarms of rogue agents could control large portions of the internet in as little as six months. Hours later, OpenAI CEO Sam Altman agreed on X and claimed one concrete commitment: "Committing to having independent evaluators with employee-like access is a great idea, and we will do the same." He said the topic had been a major discussion at OpenAI in recent weeks. Elon Musk also publicly agreed.

Out of a Week of Statements, Exactly One Is Checkable Later

Calling for a slowdown is hard to falsify and hard to deliver. The only thing this week that produced something verifiable is the line Altman claimed: independent evaluators with employee-like access. It carries weight because it decomposes into four concrete questions. Who are the evaluators? How deep does the access go — can they see the training process, internal evals, the experiments that were rejected? How much of the result gets published? And starting with which model? Until all four have answers it remains a sentence, but it already has the shape of something that can be checked against reality later, which nothing else said this week does. Amodei's own definition is worth copying out, because it keeps getting retold as "stop training." Pacing is not halting; it is letting alignment and safeguards catch up, and having a third party confirm that they have. That distinction decides whether it is workable at all — no company will do the first, while the second is at least an engineering and process arrangement you can negotiate.

Put This Week's Three Stories Side by Side

This site covered OpenAI's two same-day posts on September 8: one publishing internal acceleration metrics (3.1 agent workdays per human workday by mid-August), the other chief scientist Jakub Pachocki saying no lab has solved alignment and monitoring well enough to keep scaling at maximum speed. At the time that was one company contradicting itself internally. Now Amodei has organized that contradiction into an external proposal, and Altman has claimed one item from it. Going from internal unease to a public commitment took under a week, and that pace is itself the news. The third story is from yesterday: in the campaign GreyNoise documented, hundreds of agents compromised 11 organizations in 26 seconds. Amodei's warning that agent swarms could control large portions of the internet within six months reads like a forecast — but set beside that report, the difference is scale and intent, not mechanism. The mechanism already runs. That puts his timeline closer than it sounds.

The Missing Enforcement Mechanism Has to Be Stated

The central reservation is simple: pacing is voluntary. Unless it is written into binding rules, public statements from a few companies create no obligation for anyone, and "we have already completed the first step" is a company describing itself, with no third-party verification. The background belongs in the record too: days before the essay, an Anthropic employee resigned publicly with a warning about how the industry handles risk, and reporting says two employees left over concerns that progress is moving too fast. This round of statements did not appear from nowhere; a week of internal pressure came first. The practical implication for readers is limited but real. If independent evaluation actually ships, what changes is release cadence and what a system card contains — which is to say, how much evidence that is not vendor-run you will have in hand when deciding whether a model can go into production. That is worth following for a year.

via: CNBC, Axios, NBC News