# OpenAI's Chief Scientist Calls for a Coordinated Slowdown in Frontier Model Development

Jakub Pachocki argues that no lab's safeguards can support full speed scaling for much longer, and he wants voluntary commitments turned into mandatory thresholds enforced by outside auditors or governments.

- Published: 2026-09-08T06:18:45.692Z
- Canonical: https://polylog.news/ai/2026-09-08/openai-s-chief-scientist-calls-for-a-coordinated-slowdown-in
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Anthropic News](https://www.anthropic.com/claude-fable-and-mythos-5-1)

Jakub Pachocki, the chief scientist of OpenAI, published an essay titled "An Alien Mind" on 6 September, arguing that the industry should be prepared to decelerate. He wrote that competing to build the most capable systems regardless of the risk is indefensible once the scale of that risk is understood, and that labs should slow down together when their safety work cannot match the pace of their capability gains. [Bloomberg reported](https://www.bloomberg.com/news/articles/2026-09-07/openai-chief-scientist-urges-extreme-caution-with-pace-of-ai) that he urged "extreme caution" on the pace of development, and [Decrypt summarized](https://decrypt.co/377635/openai-chief-scientist-warns-ai-slow-down) his position that no lab's current safeguards are adequate to keep building more powerful systems at full speed for much longer.

The concrete request is institutional rather than technical. Pachocki proposes that today's voluntary company commitments become mandatory safety thresholds, [enforced by independent auditors, governments or international bodies](https://www.datastudios.org/post/openai-chief-scientist-ai-slowdown-recursive-self-improvement-safety-bars). His stated worry is recursive self-improvement: models that run research projects and eventually improve themselves without a human involved in the process.

Two things are worth separating. OpenAI has announced no pause, no delayed launch and no change to its compute plans, so this is an argument made public by an executive, not a corporate decision. And a leading lab that proposes the rule usually finds it easier to meet than a smaller challenger does, which is the standard objection to safety-threshold proposals like this one. What is verified is that this kind of restriction is already happening in practice: OpenAI limited the strongest cyber capabilities of its Astra model to vetted testers, and Anthropic ships [Mythos 5.1 only to organizations cleared through its Cyber Verification and Life Sciences Verification programs](https://www.anthropic.com/claude-fable-and-mythos-5-1), currently inside the United States only.

## What this means

The mechanism to watch is access restriction, not computing power. When frontier capability is released through vetted-organization programs and regional limits rather than an open application programming interface, the developers who lose access are small companies and firms outside the United States, while the winners are large enterprises with compliance staff who can pass a verification review. That reshapes who can build on frontier tools well before any law arrives, and it pushes everyone left out toward open-weight alternatives.

## What to watch

- Whether any lab actually delays a launch or reduces a training run, rather than only saying deceleration may be needed. An actual delay would be the first evidence that this argument actually changes company decisions.
- Whether other frontier labs endorse a shared safety threshold or publicly decline it. Endorsement would point toward an industry standard body, and refusal would point toward regulation written without input from the labs.
- How many customers Anthropic's and OpenAI's vetted-access programs admit, and whether organizations outside the United States get in. That determines whether this restriction functions as a safety measure or as an export control under another name.