Mesa-Optimizer

Advanced

Learned subsystem that optimizes its own objective.

Full Definition

Learned subsystem that optimizes its own objective.

Keywords

inner optimizer

Domains

AI Safety & Alignment

Related Terms

Robust Alignment related to

Maintaining alignment under new conditions.

 Corrigibility related to

Willingness of system to accept correction or shutdown.

 Scalable Oversight related to

Using limited human feedback to guide large models.

 Existential Risk related to

Risk threatening humanity’s survival.

 Value Misalignment related to

Model optimizes objectives misaligned with human values.

 Outer Alignment related to

Correctly specifying goals.

 Inner Alignment related to

Ensuring learned behavior matches intended objective.

 Deceptive Alignment related to

Model behaves well during training but not deployment.

Concept Map

See how Mesa-Optimizer connects to other concepts.

Open Knowledge Graph