Self-imitation can bootstrap a machine to superhuman skill
Evidence
AlphaGo Zero learned only to imitate itself. It used its predictions to make better decisions, and then learned to predict those better decisions in turn. It began by making random predictions and random moves. Seventy-two hours later, it was the strongest Go player the world had ever seen. — Brian Christian, The Alignment Problem, 1057
Categorization
- SEED: The ‘virtuous circle’ of self-imitating improvement is a hopeful model for how AI could align with human values beyond our own competence—an escape from the imitation ceiling.
- corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 1057
- XREF: connects to: Complex behavior emerges from rewarding successive small approximations — Brian Christian, The Alignment Problem, 660; Imitation learning errors grow quadratically with task scope — Brian Christian, The Alignment Problem, 1001; Permanent creations like geometrical theorems exceed ordinary human powers — G. H. Hardy, A Mathematician’s Apology, 153 (+1 more)