How AI Works: The Four Core Mechanics

Demystifying AlphaGo's Move 37, Diffusion Models, Autoregressive Hallucinations, & The Cumulative Stack

The Emergence of Non-Human Creativity

In March 2016 (Game 2 vs Lee Sedol), AlphaGo placed a black stone on the 5th line (coordinate row 10, column 5). Human masters considered 5th-line shoulder hits fundamentally mistaken because early game logic demands securing 3rd or 4th line territory.

Human Expert Policy Likelihood: 1 in 10,000 (0.01%)
AlphaGo Training Baseline: 160,000 Expert Games
Self-Play Honing: 30,000,000 Self-Play Games
Current Board State Win Probability (Black): 52.4%
Black: 52.4% | White (Lee Sedol): 47.6%
Key Insight: AlphaGo initially calculated that human masters wouldn't play Move 37 (0.0001 probability). But its value network—honed over 30 million self-play games—evaluated that the move exerted long-range cosmic central influence, jumping Black's projected win rate to 68.2%.