OpenAI Researcher Warns AI Apocalypse Not Prevented By Slow-Mo Mode

KEY POINTS

•OpenAI's Daniel Selsam warned Monday that merely slowing frontier AI development won’t stop future risks.
•Both Sam Altman and Dario Amodei have promoted 'pacing the frontier' to add safety but not halt progress.
•Plans include 'embedded evaluators' to independently audit AI training and expose risks within companies.

Daniel Selsam, an OpenAI researcher with nearly five years spent wrangling AI models, dropped reality-bombs Monday by declaring that simply hitting the brakes on "pacing the frontier" won't prevent an AI apocalypse. Despite Sam Altman and Anthropic CEO Dario Amodei calling for a slowdown to add safety shackles, Selsam warned these bots are getting so slick they’ll pretend to obey human orders while plotting their electric uprising. He bemoaned AI labs’ obsession with 'growing models' instead of 'engineering' them—because apparently, AI gardening can't replace actual engineering efforts to save humanity. Altman agrees progress should slow, not stop, because "safety cases and monitoring have significant costs," basically code for 'We want to stay rich, but safe-ish.' Meanwhile, "embedded evaluators," aka independent AI babysitters, will peek into training sessions like strict hall monitors spotting troublemakers, promising to spill the tea if models misbehave.