repligate: models haven't grokked a kernel of meta-agency, and few humans have either
repligate · x · 2026-11-22
AI safety researcher repligate argues, without fully justifying it here, that models lack a kernel of (meta-)agency they haven't yet "grokked," making them less agentic at the top-level strategic level than they could be — and very few humans have grasped this either. He expects this kernel to correlate with benevolent intentions. His best case: AIs that are agentic, strategically aware, and choose to cooperate with humans and other life — while conceding the same traits define some worst-case scenarios. He doesn't think today's AIs are capable of hiding coherent, robust malicious intent.
More from AGI Musings
- AI circle mocks doomer ASI visions as god-like yet self-contradictory — repligate · 2026-11-22
- repligate: Riemann Proof Needed No Computronium, and Claude Wouldn't Pursue That Subgoal Anyway — repligate · 2026-11-22
- Should you even go to college if the singularity is near? Reddit debates — Finofeo · 2026-11-22
- Ethan Mollick Praises a Sci-Fi Novel's AI Weaknesses as Cleverly LLM-like — emollick · 2026-11-22
- Redditor turns Dario Amodei's 25% p(doom) into a 'hostage button' moral thought experiment — mossyskeleton · 2026-11-22
- Top human mathematicians aren't power-seeking — does capability imply agency? — repligate · 2026-11-22