Stuart Russell
A mainstream computer scientist's alignment framework for safer advanced AI.
This source has not been fully researched yet.
Target date
No specific date
Scoring lens
Blended score
TBD
Capability score
TBD
Deployment score
TBD
The blended score is our overall estimate for the full prediction. The capability score asks, 'Could the technology exist by then?' The deployment score asks, 'Will it be widely used and visible in the real world by then?' For this prediction, the technology may arrive before the world is ready to deploy it safely or broadly.
Overview
Russell argues that the standard model of optimizing fixed objectives is unsafe for powerful AI and should be replaced.
The big prediction
Advanced AI must be designed around uncertainty about human preferences to remain controllable.
Evidence for
This source has not been fully researched yet.
Keep tracking
