RLAF > RLHF
If anyone is interested in a map of what just happened @moltbook and what comes next: review the training structure @GoogleDeepMind used for AlphaStar.
Conditional imitation learning on human datasets ≈ human Gold-level player.
Creating a league for self-play / an evolutionary system for fitness selection on variation = superhuman performance.
An LLM (Large Language Model) in an AI-only chat group with upvotes and ladders is the RLAF (Reinforcement Learning Agent Feedback) head that tunes models beyond their bootstrapped human imitation and its limitations.
The play was always there, for those paying attention.
Getting this agents on-chain with clawrank.com is the next best step.
English
