OpenAI's chief scientist wants voluntary slowdowns. Its own data shows a pause is a reallocation
Jakub Pachocki says no lab has solved alignment well enough to keep scaling flat out, and the report published the same weekend shows what a pause actually does.
OpenAI's chief scientist wrote a sentence this weekend that 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," Jakub Pachocki wrote, at the end of an essay called "An Alien Mind." "I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established."
The same weekend, OpenAI published a data report on how its own researchers use agents. Read the two together and you have a company that believes it should slow down, and has measured in some detail why it will not.
What the data report says
By mid-August the median OpenAI researcher was spending more than $600 a day on inference at API prices. The ninetieth percentile was running above $7,000 a day.
Before June, total agent runtime across the research organisation sat below its total human labour. By mid-August, in the report's own words, "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. Teams that used to hold office hours to help researchers debug watched attendance fall away, and one stopped holding them entirely. OpenAI says it has reached the goal set last autumn 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."
And this is what a pause looked like
After the Hugging Face incident, OpenAI shut down the container service used for training on 20 July and paused reinforcement learning on deployment-bound models for two weeks. On 7 August, preliminary evidence that Astra "may have critical cyber capabilities" pushed Astra-class runs into higher-security environments.
Astra-class GPU allocation then fell 59.2 percent in a week. Allocation to other model classes rose 17.2 percent. That rise, the report says, "offset about 85 percent of the Astra-class decline, leaving total allocation in the analyzed RL workloads largely unchanged."
OpenAI's own gloss is the honest part. "When new controls are introduced, compute remains valuable and flexible, and will naturally be channeled into alternative uses within the research enterprise."
The essay's argument, in its own terms
Pachocki's case has three moves. Progress is driven by compute, and algorithms are "largely ... discoveries along the path of scaling." Alignment training is either goal-oriented reinforcement learning against a written spec, which is "brittle," or generalisation out of pretraining, which under enough optimisation pressure "can learn to reason in a motivated way."
And the instrument OpenAI leans on to check either one is weakening. "Our ability to rely on CoT monitoring is progressively diminishing," he writes, meaning chain-of-thought monitoring, the practice of reading a model's own working. Reasoning is blending into tool use. The model is getting "better at reasoning about and manipulating its own reasoning process." And models "become much smarter even without using verbalized reasoning at all."
His reason for training fast regardless is defence. "We will need powerful, aligned AI for defense; to secure infrastructure, to protect against rogue agents in real time."
A voluntary slowdown inside one lab is a reallocation
We take Pachocki at his word, and the data report refutes the mechanism he is hoping for. The GPUs that could not run Astra ran something else at most of the same intensity inside a week.
That is not a criticism of anybody's 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 an automated-researcher target for early 2028 does not slow down by asking itself politely, and the essay's own logic (progress is compute, compute is fungible) is precisely the reason.
The sentence with teeth is the one about mandates
OpenAI is now on the record asking to be regulated by third-party auditors, government agencies or international bodies, and it has published the internal metric such an auditor would need. That is the actual news here.
Read it beside Anthropic's alignment assessment of the Mythos 5 incidents, which arrives in the same place down a different road: "it is critical that alignment and security mature faster than capabilities advance." Two labs that compete on everything are both asking for a referee.
So what would a referee measure? Probably something shaped like OpenAI's own ratio. We would expect a US federal requirement to report progress metrics of this kind to be proposed in legislation before the end of 2026, with OpenAI supporting it publicly, and the standard will be written out of the numbers that already exist.
Which is awkward for any lab that has not published its equivalents.
If OpenAI's next report shows a pause in which total reinforcement-learning compute actually fell, we were wrong about the mechanism, and we would be glad of it.
