AlphaGo vs AlphaZero

Training Philosophy: Human Knowledge vs. Tabula Rasa

The most fundamental difference between AlphaGo and AlphaZero is their relationship to human expertise. AlphaGo was bootstrapped on millions of moves from expert human Go games using supervised learning. This gave the system a strong foundation—it could predict expert moves with reasonable accuracy before any self-play training began. The subsequent reinforcement learning phase refined this foundation, but human knowledge remained the bedrock.

AlphaZero rejected this approach entirely. Starting from random play and knowing only the rules of each game, it discovered everything from scratch through self-play. The result was not just equivalent performance—it was superior. AlphaZero surpassed every version of AlphaGo, demonstrating that human game knowledge was not merely unnecessary but actually a limiting factor. This "tabula rasa" insight has proven to be AlphaZero's most enduring contribution: it showed that AI systems unconstrained by human assumptions can explore strategy spaces more effectively than those guided by centuries of accumulated expertise.

Recent research published in PNAS (2025) has gone further, extracting novel chess concepts from AlphaZero's internal representations and successfully teaching them to human grandmasters—demonstrating that the knowledge AlphaZero discovered from scratch genuinely extends beyond human understanding.

Architecture and Efficiency

AlphaGo's architecture was comparatively complex. It used separate policy and value networks, a fast rollout policy for Monte Carlo Tree Search simulations, and required enormous distributed compute—1,920 CPUs and 280 GPUs for the Lee Sedol match. Each component served a specific function, and the system's performance depended on their careful orchestration.

AlphaZero dramatically simplified this. A single neural network with two output heads (policy and value) replaced AlphaGo's multiple networks. The MCTS was streamlined by eliminating rollouts entirely, relying solely on the value network for position evaluation. Training required just 4 TPUs. This architectural elegance—achieving superior performance with a simpler, more general design—exemplified a trend that would define the next decade of AI research: scale and simplicity often beat complexity and specialization.

The efficiency gains were staggering. Where AlphaGo trained for weeks, AlphaZero reached superhuman chess in 9 hours and superhuman Go in about 34 hours. This wasn't just faster—it made the approach practically applicable to new domains without months of engineering effort per game.

Generality: Specialist vs. Polymath

AlphaGo was a Go specialist. Every architectural decision, every training choice, every evaluation function was designed specifically for Go. Moving it to another game would have required fundamental re-engineering. This is not a criticism—specialization was necessary to achieve the first breakthrough, and AlphaGo's victory over Lee Sedol was the proof of concept that made everything after it possible.

AlphaZero, by contrast, was deliberately designed as a general algorithm. The same code, the same hyperparameters, and the same architecture mastered Go, chess, and shogi. This generality was the point: it demonstrated that reinforcement learning from self-play was a domain-general capability, not a game-specific trick. The progression from AlphaZero to MuZero (which added learned environment models) to AlphaProof (which applies the same principles to mathematical theorem proving) confirms that this generality was real, not a one-off achievement.

Legacy in Modern AI (2025–2026)

Both systems' legacies extend far beyond board games. AlphaGo's primary legacy is cultural and conceptual: it proved that deep learning combined with search could conquer problems previously thought to require human intuition. Move 37—the unconventional fifth-line stone placement in Game 2 against Lee Sedol—remains the iconic symbol of AI creativity.

AlphaZero's legacy is more architectural and methodological. Its self-play reinforcement learning paradigm directly inspired: AlphaProof, which achieved silver-medal performance at the 2024 International Mathematical Olympiad; AlphaEvolve, DeepMind's coding agent that discovered a novel matrix multiplication algorithm (what Demis Hassabis called its own "Move 37 moment"); and the reasoning capabilities in modern LLMs. OpenAI's O-series models and DeepSeek-R1-Zero both draw on AlphaZero-style principles of learning to reason through search and self-improvement rather than pure imitation of human outputs.

A January 2026 analysis argued that current LLMs are at the "AlphaGo stage"—trained heavily on human data—and that an "AlphaZero-style upgrade" through self-play and reinforcement learning represents the next frontier for language model development. If this analogy holds, AlphaZero's most transformative impact may still be ahead.

Impact on Human Players and Knowledge

AlphaGo's impact on professional Go was immediate and profound. After the Lee Sedol match, top professionals began incorporating AI analysis into their training. Move 37 and other unconventional strategies prompted a reassessment of fundamental Go theory. The game's professional community went through a period of disruption as centuries-old positional principles were overturned.

AlphaZero had a similar but broader impact across multiple game communities. In chess, grandmasters described AlphaZero's style as "alien" yet beautiful—it favored dynamic piece activity and long-term compensation over material, playing in ways that recalled romantic-era chess but with modern precision. A 2025 PNAS study demonstrated that specific concepts from AlphaZero's play could be formally extracted and taught to grandmasters, who were then able to apply these machine-discovered ideas in their own games. This represents a genuine case of AI-to-human knowledge transfer at the highest level of expertise.