Pacing consensus

It’s striking to see a con­sensus form in the AI industry, even at its heights, around the need for a research slow­down and delib­erate next steps. OpenAI’s Jakub Pachocki:

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. I expect and hope for vol­un­tary slow­downs to become com­mon­place until shared safety bars are established. And I believe that inter­na­tional coor­di­na­tion on future AI devel­op­ment needs to become a top pri­ority for gov­ern­ments around the world.

Oof that last part is going to be a challenge, though. Talk about an issue with absolutely no polit­ical red meat whatsoever — airy and abstract, philosophical, nearly mathematical. The impli­ca­tions aren’t abstract, of course, but the issue itself, its ques­tions of align­ment and control, might as well be exper­i­mental poetry. I mean, just read Jakub’s post!

This is going to be a heavy lift for the tech­nocrats of the world … but if the world ever needed them to lift, it’s now.