DeepMind's new AI program, AlphaZero, taught itself to play and beat the world's best AI in three different board games: chess, shogi, and Go
Jackson Ryan / CNET :
Context & Ripple Effects
AlphaZero is the second step in DeepMind's move away from human game knowledge. In 2017 its self-taught Go agent beat the AlphaGo version that defeated the world champion, winning 90% of games using only reinforcement learning with no human input; AlphaZero generalizes that recipe across chess, shogi, and Go in a single program.
The significance is breadth plus method: one self-taught system now tops the best dedicated AI in three games, and the related coverage already shows where the line goes next — StarCraft II with AlphaStar's 10-1 run against professionals, then MuZero mastering games without even being told the rules.
First-order effects
- DeepMind replaces three separate game-specific systems with one self-taught program, validating that its reinforcement-learning approach transfers across games rather than being a Go-only result.
- The world's best existing chess, shogi, and Go engines are beaten by a system that learned purely from self-play, removing human opening theory and game records from the training pipeline.
Second-order effects
- DeepMind applies the same self-play template beyond turn-based boards, taking it into real-time strategy with AlphaStar against professional StarCraft II players.
- The method is pushed further into domains where the rules themselves aren't given: MuZero learns games without knowing the rules and is applied to YouTube video compression, the first sign of the technique leaving games entirely.
Third-order effects
- The lineage from IBM's Deep Blue through AlphaGo to AlphaZero marks a structural shift in how superhuman play is achieved — from hand-coded search plus human knowledge to systems that bootstrap their own expertise — making human game data optional rather than essential.
- If the pattern holds, games serve as the proving ground while the underlying learning machinery migrates into commercial workloads like video compression, turning research demos into applied infrastructure.
The trend: DeepMind's game-playing AI is evolving from human-trained specialists to fully self-taught generalists, with each generation shedding more human-supplied knowledge on its way out of games and into practical applications.