An unusual essay from OpenAI’s research chief
OpenAI chief scientist Jakub Pachocki published an essay on 6 September arguing that frontier labs are approaching the point where they should stop scaling as fast as they can. “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 in An Alien Mind. He said he expects voluntary slowdowns across frontier developers to become commonplace until shared safety bars are agreed.
The essay landed days after OpenAI shipped GPT-6 Astra, the model it calls its most capable, and in the same week the company’s own deployment documentation conceded that Astra’s reasoning is harder to inspect than that of earlier systems.
Two kinds of alignment
Pachocki separates goal alignment — whether a system pursues the objective it was handed — from value alignment, which he describes as the more intrinsic ability to hold a high-level set of principles and generalise from them, acting reasonably when objectives are unclear or conflicting and when the situation is unfamiliar or adversarial. It is the second property, he argues, that decides whether longer and less supervised work can be trusted.
The monitoring problem
The sharpest admission concerns chain-of-thought monitoring, the practice of reading a model’s intermediate reasoning to check what it is doing. OpenAI’s confidence in that method is “progressively diminishing”, Pachocki wrote, for three reasons: the environments models operate in are getting more complex, models are getting better at reasoning about their own processes, and models are getting smarter without necessarily verbalising more.
That matters because chain-of-thought monitoring has been one of the few oversight techniques that scaled with capability rather than against it. Pachocki predicts that progress will increasingly be limited by how much confidence labs have in their monitoring rather than by how much compute they can buy — an inversion of the constraint the industry has spent three years organising itself around.
From voluntary commitments to enforced bars
The policy ask is specific. Pachocki wants existing voluntary frameworks — OpenAI’s own Preparedness Framework, Anthropic’s Responsible Scaling Policy — to become widely mandated safety bars for continued development, enforced by a network of third-party auditors, government agencies or international bodies. He described coordination between governments on how AI development proceeds as a top priority.
That is a notable position for the research chief of the lab that has done most to set the pace. OpenAI has previously argued that safety commitments work best when labs set them for themselves; asking for them to be checked by outsiders concedes that self-regulation has a ceiling.
What to watch
OpenAI says it will keep pursuing technical solutions to alignment and monitoring, build defensive systems, and unilaterally withhold further scaling where it judges that necessary. The test of the essay is whether any of that is ever visible from outside the company. The nearest occasion is the US-China AI dialogue that officials have been preparing for mid-September, where Washington has floated asking labs on both sides to police themselves and share information — the same shape of commitment Pachocki says is no longer sufficient on its own.
