Something unusual happened at OpenAI this weekend.
Jakub Pachocki, the company's chief scientist, said AI labs may need to slow down when safety work cannot keep up with model progress.
His argument is simple. AI could become much more capable over the next few years, but our ways of checking what it is doing are still imperfect. A model can look helpful in one test and still find ways around the people watching it.
Pachocki wants voluntary safety promises to become real standards, checked by outside auditors, governments, or international groups. That would be a huge change for an industry that mostly polices itself.
The hard part is the word slow. Who decides when development has gone too far? What exact warning triggers a pause? And who makes every major lab follow the same rule?
The race is getting faster, but the brakes are still being designed.
This is especially striking because OpenAI just launched its newest model with confident claims about safety and capability. Now one of its top scientists is saying the industry cannot keep scaling at maximum speed forever.
My take? A slowdown sounds sensible, but it only matters if someone can define the speed limit.
Top comments (0)