Product1 publisher3 min readPublished
OpenAI's chief scientist asks the whole industry to slow down days after GPT-6 Astra shipped
Jakub Pachocki's post on OpenAI's website asks every AI company to slow down and governments to coordinate on what comes next. The post is undated, and nothing a customer has already signed changes.
The Product Desk · Product desk
What happened
- OpenAI chief scientist Jakub Pachocki published a post on the company's website arguing that AI development should be slowed across the whole industry, not only inside OpenAI.
- He also said international coordination from governments on future AI development is needed, and concluded that using both levers together is the best way to proceed.
- It went up days after OpenAI released GPT-6 Astra, the business-focused model that customers are now migrating to.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision A roadmap that assumes each release is cheaper and more capable than the last now sits against the public position of the lab's own chief scientist, so teams have to decide whether to keep pricing that curve into commitments.
- constraint The lever OpenAI describes first lives in customer workflows. If monitoring and human sign-off become requirements, the rollout owner builds and staffs them, and the lab does not.
- contradiction TechRadar called the timing surprising because Astra shipped days earlier. Buyers are left weighing an argument for restraint against a shipping cadence that looks the same as before.
- precedent A frontier lab putting government coordination in writing hands regulators an industry participant to cite when they draft capability rules.
Somebody has a migration ticket open for GPT-6 Astra this week, with the prompts rewritten and the evals half done. Jakub Pachocki's post does not change that ticket. TechRadar's account of it records no date, no capability threshold, and no change to a rate limit or a contract term [15]. OpenAI shipped Astra, its business-focused model, days before the post appeared [10].
What the post does give a product team is a precise description of the two things OpenAI thinks it can pull. "The main levers we have are either steering the process to strengthen alignment and monitoring alongside the AI and find ways to keep people in the loop; or coordinating to slow down future development as needed to build confidence in these measures," Pachocki wrote [5]. His conclusion is that a combination of both is the best way to proceed [6].
The two levers land in different parts of a roadmap. Monitoring and keeping people in the loop means review queues and approval screens, built and staffed on the customer side. Coordinated slowdown is about models still to come. A feature whose value depends on the next release being cheaper and better than Astra is the one the second lever reaches, and OpenAI's chief scientist has now argued in public for slowing that down [1][2].
The reasoning he gives rests on how little the field knows. Pachocki wrote that "study of deep learning-based AI is largely an experimental science" and that "we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways" [9][8]. He said AIs "present clear new dangers" for computer security [4]. Pointing at OpenAI's own plan for an automated research assistant and its progress on recursive self-improvement, where an AI develops its next iteration [11], he wrote: "I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence" [3].
TechRadar called the timing surprising given how close it sits to the Astra release [12]. Christian Cawley, who wrote the TechRadar piece, asked whether the rest of the industry would work to a reduced pace and answered, "It doesn't seem incredibly likely" [13]. TechRadar also notes that the post does not discuss energy, cooling, or data center opposition [14].
Running a roadmap against this is cheap. Two questions do it: does a feature's value depend on a model more capable than Astra, and does it route a consequential action without a human check? The features that answer yes twice are where both levers point. The post leaves alone the ones that answer no twice. The other two cells say which risk is being carried: schedule risk if future capability arrives later than planned, or rework if monitoring and sign-off stop being optional. The post gives no date for either lever [15].
What to watch
- Whether OpenAI turns either lever into published policy: a gated capability tier, a deprecation schedule, or a required human review step in the API terms.
- Whether any named government or standards body takes up Pachocki's call for international coordination with a concrete process.
- Whether other frontier labs answer in kind or keep shipping on their existing cadence.