Paul Christiano: no alignment approach has >50% chance of working — 'that is the plan'
trevposts · x · 2026-11-19
Dwarkesh Patel cites Paul Christiano (OpenAI board member, foundational AI safety figure): no single current alignment approach has more than a 50% chance of getting superintelligences to actually do what a particular human user wants. Dwarkesh presses the implied plan — months before superintelligence, unleash AIs on many approaches each with a few percent chance, hope they solve alignment while still on our side, then build 'crazy amounts' of superintelligent minds aligned with us. Christiano: 'That is the plan. I agree that that is the plan.' A striking admission of how thin current alignment confidence is.
More from AGI Musings
- Dean Ball challenges AI's "silent consensus" on human primacy — deanwball · 2026-11-19
- David Deutsch: Musk's Childhood Trance Shows Why No Behavioural Test Can Detect AGI — Paimaamu · 2026-11-19
- Tokens spent explained 80% of performance on Anthropic's browsing benchmark — Nir777 · 2026-11-19
- If AI hits targets, move to a four-day workweek? — MikkoH · 2026-11-19
- Moral Machine Creator Iyad Rahwan Interviewed on How AIs Should Make Moral Decisions — iyadrahwan · 2026-11-19
- Nina Schick: the risk is not having SuperIntelligence — first felt in cyber — NinaDSchick · 2026-11-19