All terms
Patterns
Self-Improvement
Approaches where AI systems get better using their own outputs or feedback, with less human input.
Definition
Self-improvement refers to approaches where AI systems get better using their own generated data, critiques, or feedback rather than relying solely on new human-labeled data. Examples include training on a model's own best solutions or having models critique and refine each other. It raises both promise and safety questions.