Johannes Heidecke, the Head of Safety Systems at OpenAI, is leaving. I know what you’re thinking: Hey, didn’t the head of safety at OpenAI just leave?
In fact, it feels like a head of safety at OpenAI is pretty much always leaving. Working in safety leadership—loosely defined—at OpenAI is a little like working as a drummer in the band Spinal Tap; lots of turnover. I’m not the world’s premier OpenAI Kremlinologist, so I might be missing some details and nuance, but here’s my basic timeline:
According to Wired, those previously reporting to Heidecke’s safety teams will be led by Mia Glaese, who is a VP, and also the head of alignment. However, there does seem to be an other replacement for Heidecke, according to Wired. Saachi Jain, former leader of safety teams, will now be an “interim head of safety systems” under Glaese.
What exactly keeps happening inside OpenAI’s offices is anyone’s guess, but OpenAI research chief Mark Chen did at least give Wired a hint, saying, “The demands on safety continue to increase—we are training models at a much faster cadence, and release cycles have come down greatly in turn,” and added, “As a result, we have bigger coordination challenges around safety today than ever before.”
The generous reading is that this is still an immature industry. The points along the chain where safety considerations are needed genuinely may keep jumping around as OpenAI figures out how best to build its products. Perhaps today’s sensible safety test procedure is tomorrow’s unnecessary bottleneck.
And there’s no actual direct evidence for a less generous reading of Heidecke’s departure—for instance, one in which any such consideration is a post-hoc rationalization for a pruning of safety procedures in service of faster product rollouts.
Read the full article here
