A look at the history of game-playing AI and how it often beats humans, from IBM's Deep Blue to AlphaGo
Context & Ripple Effects
This Engadget retrospective lands at a natural inflection point in the games-as-benchmark arc: IBM's Deep Blue proved machines could beat humans at chess, and DeepMind spent the following two decades repeating the trick with harder games — AlphaGo Zero learning purely by self-play and winning 90% of games against the champion-beating version of AlphaGo, then AlphaZero teaching itself chess, shogi, and Go from scratch within months.
The piece also arrives just after DeepMind pushed past perfect-information board games entirely, with AlphaStar taking a 10-1 series off professional StarCraft II players. The later result that an amateur could beat a top-ranked Go system using tactics surfaced by an analysis program adds the counterpoint this history needs: superhuman play does not mean unexploitable play.
First-order effects
- For general readers, the piece consolidates a fragmented timeline — Deep Blue, AlphaGo, AlphaZero, AlphaStar — into a single narrative where each DeepMind system was trained with less human input than its predecessor.
Second-order effects
- Each milestone has forced the next benchmark to be harder: once self-play mastered closed board games, DeepMind moved to StarCraft II's hidden information and real-time decision-making, keeping the human-vs-machine comparison alive as a public proof point.
Third-order effects
- The amateur-with-a-tool result points toward a structural shift in how these systems are evaluated — from 'can it beat the champion' toward adversarial probing for exploitable weaknesses, since even top-ranked Go AI lost 14 of 15 games to a prepared human.
The trend: Game-playing AI is evolving from headline victories over champions toward self-taught general methods — and toward adversarial testing that treats superhuman systems as systems with weaknesses to find.