iterated amplification
An alignment and scalable-oversight approach that repeatedly alternates amplification and distillation: a human or current overseer uses many copies of the current AI to solve a task better than either alone, then a new AI is trained to imitate or distill that amplified overseer, allowing capability to bootstrap while aiming to preserve alignment.
Metadata
- aliases
- en
- IDA
- iterated distillation and amplification
- concept_id
- iterated_amplification
- preferred
- en
- iterated amplification
- relations
- relations
- broader
- prerequisite
- status
- active