How do we (more) safely defer to AIs?
A chilling new strategy suggests we must let AI govern itself to survive.
Deep Dive
A new AI safety paper argues that as systems become more powerful, human control becomes infeasible. The proposed strategy is to fully defer to AIs on managing existential risks, having them automate safety research and align successor systems. The author claims this rushed deference, while dangerous, may be the only viable option if a fast takeoff scenario like 'AI 2027' occurs, and must happen just above the minimum capability threshold for automating safety work.
Why It Matters
This outlines a controversial last-resort plan that could define humanity's survival in a potential AI takeover scenario.