Scalable Oversight

Advanced

Using limited human feedback to guide large models.

Full Definition

Using limited human feedback to guide large models.

Keywords

weak supervision

Domains

AI Safety & Alignment

Related Terms

Existential Risk related to

Risk threatening humanity’s survival.

 x-Risk related to

Existential risk from AI systems.

 Takeoff Speed related to

Rate at which AI capabilities improve.

 Fast Takeoff related to

Sudden jump to superintelligence.

 Deceptive Alignment related to

Model behaves well during training but not deployment.

 Mesa-Optimizer related to

Learned subsystem that optimizes its own objective.

 Robust Alignment related to

Maintaining alignment under new conditions.

 Corrigibility related to

Willingness of system to accept correction or shutdown.

Concept Map

See how Scalable Oversight connects to other concepts.

Open Knowledge Graph