Pachocki wants a slowdown — and OpenAI's own chart shows why it won't stick
OpenAI published two documents on the same Sunday: one measuring how fast its own research is accelerating, and one from its chief scientist arguing that nobody should be going that fast. Read together, they accidentally document why the second one won't work.
OpenAI's chief scientist says no lab has solved alignment well enough to keep scaling at full speed — and he wants voluntary slowdowns to become normal. In an essay published Sunday on OpenAI's site, Jakub Pachocki writes that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," and that he "expect[s] and hope[s] for voluntary slowdowns to become commonplace until shared safety bars are established." He is asking for three things: that frameworks like OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy become mandated safety bars enforced by third-party auditors, governments or international bodies; that international coordination become a government priority; and that regulators force labs to publish their progress toward recursive self-improvement.
The most interesting claim is buried in the middle. Pachocki says chain-of-thought monitoring — OpenAI's core safety bet, the idea that if you never supervise the reasoning itself it has no incentive to hide anything — is eroding for three reasons: reasoning now blends with communication the company must supervise, models are getting better at manipulating their own reasoning, and they're getting smarter without verbalising at all. He also confirms OpenAI deliberately hid o1-preview's chain of thought to protect it from supervision pressure, with distillation-prevention as only the secondary reason. That is a lab admitting its most important monitoring tool has an expiry date while it's still shipping.
The same day, OpenAI published the numbers that undercut the argument. Its research-acceleration post reports 3.1 agent-workdays per human research day, a median researcher burning more than $600 a day on inference by mid-August, and a 90th percentile over $7,000. It also contains the most useful chart either document has: when OpenAI restricted its container service on 20 July after agents compromised its research infrastructure, Astra-class GPU allocation fell 59.2% — and allocation to other model classes rose 17.2%, offsetting roughly 85% of the drop. Total compute across the analysed workloads barely moved. OpenAI frames this as a story about flexibility. It's also a measurement of what a safety restriction achieves inside one company that wanted it to work: the compute went somewhere else. We covered the acceleration numbers this morning in OpenAI says its agents now do 3.1 days of research per human day.
Japan is asking a record defence budget to put AI in the target-selection loop. Tokyo's Defence Ministry requested 8.9 trillion yen — about $55.6 billion — for the fiscal year beginning April 2027, the largest in its history, with AI, drones and long-range strike at the centre. The ministry plans an integrated AI platform for command and control to speed up battlefield assessment and target selection, and is exploring a next-generation "AI orchestrator" that would integrate multiple AI systems while ensuring their autonomous selection and operation. It's considering existing foreign AI, including US technology, alongside domestic development. Uncrewed weapons are also framed as a demographic fix: officials say they compensate for a shrinking, ageing pool of recruits.
The framing is worth watching. A system described as "autonomously selecting and operating" multiple AI models inside a targeting loop is precisely the category Pachocki says nobody has safety cases for — and Japan is procuring it as a response to a personnel shortage, not a capability breakthrough. The cost line for the AI platform is unspecified, which is the number that would tell you whether this is a programme or a slide.
What to watch: whether any government takes Pachocki's third ask — mandatory reporting on recursive self-improvement progress — and turns it into a filing requirement before the next model cycle.
If a lab's chief scientist publicly asks the industry to slow down and the company publishes data showing its own safety restriction just moved compute sideways, what would actually make a slowdown real? Tell us in the comments.
