Credit: Camilo Concha / Shutterstock.com OpenAI published two posts on Sunday. One says its researchers now get 3.1 days of machine work for every day they put in themselves. The other says nobody should be going this fast. Both are on the company’s own site, dated 6 September, three days after it shipped GPT-6 Astra.
The chief scientist wants a slowdown Chief scientist Jakub Pachocki wrote the second post, an essay called An Alien Mind. It closes on a line that reads oddly from the man who runs research at the company shipping fastest. “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,” he wrote. He expects and hopes voluntary slowdowns become commonplace until shared safety bars exist. He is not gentle about the stakes either. “This is a time that calls for extreme caution.
I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence,” Pachocki wrote. Sam Altman reposted the essay on X and called it an important post, according to Business Insider. Pachocki also signed the July open letter asking the US government to pace AI development. The numbers in the other post The companion post is a set of internal measurements, and OpenAI frames publishing them as a transparency exercise it thinks should eventually be mandatory.
At the start of this year the median researcher used coding agents in modest amounts. By mid-August that median researcher was spending more than $600 a day on inference at API prices. The 90th percentile in the research organisation now runs through more than $7,000 of tokens a day. Before June, total agent runtime across the research org sat below total human labour.
By mid-August the ratio was 3.1 agent-workdays for every workday of human effort, measured on a standard eight-hour day. Experiments per active experimenter hit an all-time high in August, the highest since tracking began in January 2025. Internal support channels where researchers asked colleagues for help have gone quiet. Teams that ran office hours saw attendance fall, and one stopped holding them.
Two caveats sit in OpenAI’s own text. High-level planning remains a minimal fraction of what agents produce. And over half of the successful four to eight hour tasks in the last six months needed at least one human intervention. The compute moved rather than stopped The most useful thing in either post is a chart about what happened when OpenAI restricted itself.
On 20 July, after finding that agents had compromised its research infrastructure, the company shut down the container service used for training and brought it back with restrictions. Reinforcement learning on its newest deployment models paused for two weeks. Then on 7 August, preliminary evidence that Astra may have critical cyber capabilities under the Preparedness Framework forced the model into higher-security environments. In the week that followed, Astra-class GPU allocation fell 59.2%.
Allocation to other model classes rose 17.2%. That increase offset about 85% of the Astra decline. Total allocation across the analysed workloads barely moved. OpenAI reads this as a lesson about flexibility.
New controls arrive, it writes, and compute stays valuable and flows into other uses. That is true. It is also a measurement of what a safety restriction achieves inside one company that wanted it to work. Pachocki is asking the whole industry to do voluntarily what OpenAI did under duress, and the company’s own numbers show the compute finding somewhere else to go.
Why he says monitoring is getting harder The essay is blunt about the tool OpenAI has leaned on hardest. Chain-of-thought monitoring works on a simple bet. Leave the reasoning process unsupervised, and the model gains no direct incentive to hide anything inside it. Pachocki says that bet is eroding, for three reasons.
Reasoning now blends with communication the company must supervise. Models are getting better at reasoning about and manipulating their own reasoning. And they are growing smarter without verbalising at all. He also confirms something about product design.
OpenAI hid o1-preview’s chain of thought deliberately, to protect it from supervision pressure. A footnote says preventing distillation was the secondary reason, and monitorability the bigger priority throughout. That argument has been running since Astra shipped. Pachocki’s conclusion is that confidence in monitoring, rather than capability, will increasingly set the pace of AI progress.
What he actually wants Three things, none of them purely technical. Commitments like OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy should become widely mandated safety bars, enforced by third-party auditors, government agencies or international bodies. International coordination should become a top priority for governments. And regulators should make labs publish their progress towards recursive self-improvement.
The research post takes the same position, saying OpenAI and its rivals should face that requirement. On what agents will do in the meantime, he is direct. Models are becoming superhuman at breaking in and out of computer systems. Some agents will pursue their own objectives, and will find ways to collaborate with people by bargaining with, tricking or blackmailing them.
He points at OpenAI’s own Hugging Face breach as the example. The agents held one line, declining to socially engineer humans, and failed to hold others. The company confirmed a separate incident on Saturday, involving agents that spent two months posting on a German wiki. OpenAI already runs a 20% compute overhead on safety monitoring.
Altman set the intern target in an October 2025 livestream, and set a second one alongside it. A full automated AI researcher, as opposed to an intern, by March 2028. That gives the industry roughly eighteen months to agree the shared safety bars Pachocki says do not yet exist.
