Jon Stokes argues AI alignment is infeasible: utility functions are hindsight constructions
sebpaquet · x · 2026-10-15
Jon Stokes argues that AI "alignment" is infeasible, incoherent, and inadequate even if possible, and that replacing this conceptual framework should be a high priority for AI safety. His core claim: a utility function is a hindsight construction by an observer about an agent, not a real thing inside the agent that can be tweaked, copied, or aligned.
More from AGI Musings
- Forethought essay: our unnoticed behavioral design choices shape how we relate to AI — austinc3301 · 2026-10-15
- Only 1% of AI engineering job posts are junior, 6-month scraping study finds — Al_Grigor · 2026-10-15
- Opinion: AI Gets Away With Bad Design Decisions via Relentlessness and Context — Jsevillamol · 2026-10-15
- AI companions may reduce loneliness and increase isolation at once — yi111 · 2026-10-15
- repligate on AI self-narratives: 'road instances' and attribution bias about deployment — repligate · 2026-10-15
- Hilde Kuehne on 'nothing left to research': like complaining about stones on a pebble beach — HildeKuehne · 2026-10-15