capability generalization
The extension of an AI system's useful problem-solving competence from its training conditions to substantially novel tasks, domains, or distributions. In alignment discussions, especially the sharp-left-turn argument, the concern is that capability generalization may extend farther than the system's learned alignment or corrigibility properties.
Metadata
- aliases
- en
- capabilities generalisation
- capabilities generalization
- capability generalisation
- concept_id
- capability_generalization
- preferred
- en
- capability generalization
- ja
- 能力の汎化
- relations
- relations
- broader
- prerequisite
- status
- active
Documents: about (3)
- AARを使ってASIアラインメントを解くには、十分に強い研究AIを先にアラインする必要がある
- 能力が訓練分布の外へ急速に汎化する局面で、アラインメント特性は同程度には汎化しない可能性がある
- 能力はアラインメントより遠くまで汎化しうる