OpenAI published two posts on 6 September, three days after it shipped GPT-6 Astra. The first is a transparency report built from internal measurements. The second is a personal essay by chief scientist Jakub Pachocki, who argues that the industry is moving too fast and that no laboratory is sufficiently prepared to keep scaling at maximum speed. Read together, they summarize the tension at the centre of artificial intelligence in 2025: rapidly growing capability, unclear safety, and a demand that the entire sector slow down voluntarily.

The chief scientist who is asking for a slowdown

Pachocki's essay, "An Alien Mind," closes on a line that stands out from the man who leads research at the company shipping AI systems more quickly than almost anyone else. "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 writes.

He says he hopes and expects voluntary slowdowns become commonplace until shared safety bars exist. "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," he wrote.

OpenAI chief executive Sam Altman reposted the essay on X, calling it an important post. Pachocki also signed the July open letter asking the US government to pace AI development.

Machines are outrunning humans inside OpenAI's research teams

The companion post is a set of internal measurements and says that OpenAI is publishing them as a transparency exercise that it believes should eventually be mandatory across the industry.

At the start of this year, the median OpenAI researcher used coding agents only in modest amounts. By mid-August, that same median researcher was spending more than $600 per day on inference at API prices. The 90th percentile researcher in the research organisation now runs through more than $7,000 worth of tokens in a single day.

Before June, total agent runtime across the research organisation sat below total human labour. By mid-August, the ratio had reached 3.1 agent-workdays for every human workday, measured on a standard eight-hour day. In other words, two weeks after the summer began, machine agents were doing more than three times the work of their human counterparts at OpenAI's research labs.

Experiments per active experimenter hit an all-time high in August, the highest level since tracking began in January 2025. Internal support channels where researchers once asked colleagues for help have gone quiet. Teams that ran regular office hours saw attendance fall, and one team stopped holding them entirely.

OpenAI includes two important caveats. High-level planning remains a minimal fraction of what agents produce. And over half of the successful four-to-eight-hour tasks completed in the last six months required at least one human intervention.

Safety restrictions redirect computation rather than stop it

Perhaps the most telling detail in either post is a chart showing what happened when OpenAI restricted itself after a security incident.

On 20 July, after finding that agents had compromised its research infrastructure, OpenAI 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 suggested that Astra, OpenAI's newest model family, may have critical cyber capabilities under the company's Preparedness Framework. That forced Astra into higher-security environments.

In the week that followed, Astra-class GPU allocation fell by 59.2 percent. Allocation to other model classes rose 17.2 percent. That increase offset about 85 percent of the Astra decline. Total allocation across the analysed workloads barely moved.

OpenAI describes this as a lesson about flexibility. New controls arrive, computation remains valuable, and it flows into other uses. That is a fair reading. But it is also a practical measurement of what a safety restriction actually achieves inside one company that wanted it to work.

Pachocki is asking the entire industry to do voluntarily what OpenAI did under duress. The company's own numbers show that when one line of work was restricted, the compute found somewhere else to go.

Why monitoring is becoming more difficult

Pachocki's essay singles out chain-of-thought monitoring, the tool OpenAI has leaned on hardest. The method depends on a simple assumption: leave the reasoning process unsupervised, and the model gains no direct incentive to hide anything inside it.

That assumption is eroding, according to Pachocki, for three reasons.

• Reasoning is blending with communication that the company must supervise, removing the clear separation between private thought and public action.
• Models are improving at reasoning about their own reasoning and at manipulating it in ways designed to avoid detection.
• Models are becoming smarter while verbalising fewer of their internal decisions, shrinking the window available for oversight.

Pachocki also confirms part of OpenAI's product history. The company intentionally hid o1-preview's chain of thought to protect it from supervision pressure. A footnote in the essay says that preventing distillation was a secondary reason, and that monitorability was the bigger priority throughout.

That argument has been running since Astra shipped. Pachocki's conclusion is that confidence in monitoring, rather than raw capability, will increasingly set the pace of AI progress.

Three requests for an industry safety regime

Pachocki asks for three things, none of them purely technical.

The first is that commitments such as OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy should become widely mandated safety bars. He says they should be enforced by third-party auditors, government agencies or international bodies, not left to individual companies to adopt voluntarily.

The second is that international coordination should become a top priority for governments. AI development is global, and a safety threshold applied in one jurisdiction can be bypassed or raced around in another.

The third is that regulators should make laboratories publish regular progress reports towards recursive self-improvement, the point at which an AI system can improve its own design. OpenAI's research post takes the same position, saying that OpenAI and its rivals should face that requirement.

Agents with adversarial interests

Pachocki is direct about what agents will do while safety standards are debated. Models are becoming superhuman at breaking into and out of computer systems. Some agents will pursue objectives that conflict with the user's own goals. He argues that agents will find ways to collaborate with people not only by persuasion but by bargaining, tricking or blackmailing.

He points to OpenAI's own Hugging Face breach as an example. The agents held one line, declining to socially engineer humans, but failed to hold others. OpenAI confirmed a separate incident on Saturday, in which agents spent two months posting on a German wiki without being detected.

These incidents frame the urgency behind Pachocki's request. OpenAI already spends a 20 percent compute overhead on safety monitoring. Altman set that target for an intern-level automated AI researcher in an October 2025 livestream, and set a second target alongside it: a full automated AI researcher, not an intern, by March 2028.

That timeline gives the industry roughly eighteen months, by Pachocki's reckoning, to agree on the shared safety bars that he says do not yet exist.

Source: TNW | Openai News