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.

Original post →

More from AGI Musings

AGI Musings channel →