How DeepMind worked with former world chess champion Vladimir Kramnik to make its AI, AlphaZero, learn new variants, which could make AI chess more creative
Tom Simonite / Wired :
Context & Ripple Effects
AlphaZero arrived in late 2018 as a pure self-learner: after teaching itself to beat the world's best AI in chess, shogi, and Go, it seemed to have exhausted the point of board games as benchmarks. That followed DeepMind's earlier proof with AlphaGo Zero winning 90% of games against its champion-beating predecessor with no human input at all.
The Wired story closes that loop differently: instead of pushing further past humans, DeepMind brings former world champion Vladimir Kramnik in to help AlphaZero learn new chess variants — repositioning the system from opponent to creative collaborator.
First-order effects
- Kramnik gains a research instrument for exploring rule variations that human theory has never mapped, while DeepMind gains a world champion's intuition for judging whether AlphaZero's novel play is genuinely creative.
Second-order effects
- The framing pressures competing machine-chess approaches to justify themselves — including the AI trained by parsing expert commentators' reactions rather than playing, which bet on human judgment as the signal where AlphaZero uses self-play.
Third-order effects
- If expert-in-the-loop variant exploration works on chess, it strengthens DeepMind's stated trajectory — laid out by principal scientist David Silver — of carrying reinforcement learning from conquered games into open-ended real-world problems, with humans defining the rules worth exploring.
The trend: Reinforcement-learning research is pivoting from defeating humans in fixed games to collaborating with domain experts on new problem definitions, with board games as the testbed.