OpenAI's chief scientist wants voluntary slowdowns. OpenAI's own data shows 3.1 agent-workdays per human and a two-week pause that changed nothing

Jakub Pachocki, OpenAI's chief scientist, published an essay on Sunday titled "An Alien Mind" that ends with a sentence no chief scientist of a frontier lab had written before: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer. I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established."

Jakub Pachocki@merettm

I wrote about the state of AI, why I’m concerned about the next few years, and the choices we need to make to keep the future in humanity’s hands.

An Alien Mind: t.co/FeIfWNe0UE

The same weekend, OpenAI published "Research acceleration: The view inside OpenAI," a data report on how its own researchers use agents. Read the two together and you get a company that believes it should slow down and has measured, in some detail, why it will not.

What the data says

By mid-August the median OpenAI researcher, ranked by agent usage, was consuming more than $600 a day of inference at API prices. The 90th percentile: more than $7,000 a day. Before June 2026 total agent runtime in the research organisation was below total human labour. As of mid-August "the research organization uses 3.1 agent-workdays of effort for every workday of human labor." Experiments per active experimenter hit an all-time high in August. One internal support team stopped holding office hours because the agents took over the troubleshooting.

OpenAI says it has reached the goal Sam Altman set last fall of "an automated research intern by September of this year," and is "making strong progress toward creating an automated AI researcher by March of 2028."

Then the part that matters for Pachocki's argument. After the Hugging Face incident, OpenAI shut down its training container service on July 20 and paused reinforcement learning on deployment-bound models for two weeks. On August 7, preliminary evidence that Astra "may have critical cyber capabilities" forced Astra-class runs into higher-security environments, and Astra-class GPU allocation fell 59.2 percent the following week.

And: "allocation to other model classes rose 17.2 percent. That increase offset about 85 percent of the Astra-class decline, leaving total allocation in the analyzed RL workloads largely unchanged." OpenAI's own gloss: "When new controls are introduced, compute remains valuable and flexible, and will naturally be channeled into alternative uses within the research enterprise."

What the essay says

Pachocki's argument has three parts. Progress is driven by compute, and algorithms are "largely ... discoveries along the path of scaling." Alignment training is either goal-oriented RL against a spec, which is "brittle," or generalisation from pretraining, which under enough optimisation pressure "can learn to reason in a motivated way." And the tool OpenAI relies on to check either one, chain-of-thought monitoring, is weakening: "our ability to rely on CoT monitoring is progressively diminishing," because reasoning blends with tool use, models get "better at reasoning about and manipulating [their] own reasoning process," and "the models become much smarter even without using verbalized reasoning at all."

His conclusion is that commitments like the Preparedness Framework must become "widely mandated safety bars ... enforced by a network of third-party auditors, by government agencies or by international bodies." His stated reason for training fast anyway is defence: "We will need powerful, aligned AI for defense; to secure infrastructure, to protect against rogue agents in real time."

Our read

We take Pachocki at his word, and we think the data report refutes the mechanism he is hoping for. A voluntary slowdown inside one lab, as OpenAI just demonstrated with real numbers, is a reallocation. The GPUs that could not run Astra ran something else at 85 percent of the same intensity within a week. That is not a criticism of the researchers. It is what "compute remains valuable and flexible" means when you say it out loud. A company with a $600-a-day median inference habit and a March 2028 automated-researcher target does not slow down by asking itself nicely, and the essay's own logic (progress is compute, compute is fungible) is the reason.

Which is why the only sentence in the essay with teeth is the one about mandates. OpenAI is now, on the record, asking to be regulated by third-party auditors and international bodies, and has published the internal metric (agent-workdays per human-workday) such an auditor would need. That is the actual news, and it should be read next to Anthropic's alignment assessment of the Mythos 5 incidents on Wednesday, which reaches the same place by a different road: "it is critical that alignment and security mature faster than capabilities advance." Two labs that compete on everything are now both asking for a referee.

What would a referee measure? Probably something in the shape of OpenAI's own ratio. We'd expect a US federal requirement to report RSI-progress metrics to be proposed in legislation before the end of 2026, with OpenAI supporting it publicly, and the standard will be written from the numbers that exist, which is bad news for any lab that has not published equivalents. If OpenAI's next report shows a pause in which total RL compute actually fell, we were wrong about the mechanism and glad of it.