OpenAI chief scientist Jakub Pachocki has said, publicly, that no lab has solved alignment to a degree that would justify scaling at maximum speed, and that he hopes voluntary slowdowns become a normal part of how the industry operates.
Coming from the person who runs research at the company that set the pace, that is worth pausing on. It is also, read carefully, a statement with a lot of room in it. “No lab has solved it” is a description of the field. “Voluntary” is doing heavy lifting in the second half.
The coordination problem nobody has fixed
A voluntary slowdown works only if the labs that would gain from not slowing down also slow down. There is no mechanism for that. There is no verification, no penalty, and no way for a company that pauses to know whether a competitor paused with it or simply said so. The first mover into restraint hands ground to whoever declines.
Which is why statements like this tend to arrive from the technical leadership rather than the commercial side, and why they are usually framed as hopes. Pachocki is describing the state of the science honestly. He is not describing a policy his company has adopted.
What would make it real
Three things would change the reading. A published threshold, meaning a stated capability level at which OpenAI commits to stopping. An external body with access to check the claim. And evidence that a slowdown has actually happened at least once, with the cost absorbed.
Absent those, the remark still has value as a marker. When the chief scientist of the leading lab says the safety work has not caught up with the capability work, that is a data point regulators can quote back, and one that competitors will find awkward to contradict. Whether it changes a single training run is a separate question.