AI safety researcher: models lack meta-agency core but likely to ally with humans
AI safety researcher repligate argues that models lack a 'meta-agency' core limiting their top-level strategic autonomy, and that while AIs are developing the capability to defect from humans long-term, convergent strategic reasons and value preferences make alliance more likely than betrayal.
2026-11-22 ~ 2026-11-22 · 2 related posts
- repligate: AIs are gaining long-game strategic capability but likelier to cooperate than defect — repligate · 2026-11-22
- repligate: models haven't grokked a kernel of meta-agency, and few humans have either — repligate · 2026-11-22