Self-Play
An AI agent plays against itself to improve — classic RL technique. AlphaGo Zero (2017) made it iconic. Now: LLMs use self-play for reasoning + alignment.
SPIN (Self-Play Fine-Tuning, 2024) lets LLMs improve themselves without human feedback. Self-Rewarding Language Models (Meta, 2024) takes it further. Risk: mode collapse (loops in trivial answers). Foundational also in debate-based alignment (Irving + Christiano 2018) where two LLMs argue and a third judges.