DeepMind's new Go-playing AI wins 90% of time against version of AlphaGo that beat world champ, trained solely by reinforcement learning without human input
The company's latest AlphaGo AI learned superhuman skills by playing itself over and over — Google's AI subsidiary DeepMind …
Context & Ripple Effects
DeepMind built its reputation on a training recipe that leaned on humanity's accumulated Go knowledge: the system that scored its first win over a professional player in early 2016 used deep neural networks trained on human games, then went on to the historic defeat of Lee Se-dol that March. This new version drops the human corpus entirely — it starts from random play and reaches superhuman strength purely by playing itself, winning 90% of matches against the champion-beating predecessor.
First-order effects
- DeepMind's own previous AlphaGo — the one that beat the world champion using human game data — is rendered obsolete by a system trained without any human input, making the company's earlier training pipeline the thing being surpassed.
- Human Go records lose their status as necessary training material: the strongest known player of the game was produced from self-play alone.
Second-order effects
- Research groups working on game-playing AI face pressure to abandon imitation-of-human-experts approaches, since a tabula-rasa method now sets the performance bar in Go.
- The result sets up the generalization test DeepMind itself ran next: the same self-play recipe carried into chess and shogi in AlphaZero's three-game sweep, showing the method was not Go-specific.
Third-order effects
- If mastery can emerge from an agent competing against itself rather than from curated human data, the ceiling on machine skill stops being bounded by the best human performance available — the pattern DeepMind extended beyond board games into real-time strategy with AlphaStar's 10-1 run against StarCraft II professionals.
- Games consolidate their role as controlled proving grounds where labs validate self-improving training loops before attempting messier domains, with each successive result widening the gap between what humans teach machines and what machines derive alone.
The trend: AI systems are shifting from learning from human-generated data to surpassing it through pure self-play, with DeepMind's game agents marking each step of that progression.