AGI as Unpredictable Multi-Agent Systems: The Need for New Constraints

sebkrier · x · 2026-08-22

Nate Witkin highlights Séb Krier's argument against the 'omniscient genie' model of AGI, suggesting instead a future dominated by millions of semi-reliable, sub-AGI agents. The discussion emphasizes that human institutions rely on incentives (like firing or promotion) which agents currently lack, making behavior constraint difficult. As we enter a multi-agent world, better harnesses, protocols, and mechanism design are deemed critical for safety.

Original post →

More from AGI Musings

AGI Musings channel →