What it does This repo rebuilds the AlphaZero pipeline—self-play, Monte Carlo Tree Search, and a dual-head neural network—inside Jupyter notebooks. It targets Connect4 rather than Go, which keeps training times sane and the code readable. The linked blog post walks through algorithmic steps and run instructions.

The interesting bit Most AlphaZero ports chase performance; this one chases clarity. By staying in notebook form and shrinking the problem to Connect4, it turns a famously opaque system into something you can step through cell-by-cell. The trade-off is explicit: you sacrifice speed for comprehension.

Key highlights

• Pure Python/Keras implementation with no exotic dependencies
• Self-contained notebooks covering MCTS, network training, and agent play
• Connect4 as the working example—small enough to train on modest hardware
• Accompanying blog post with algorithm summary and setup guide
• ~2K stars suggests it has found an audience among learners

Caveats

• README is extremely terse; all documentation lives in the external blog post
• No candidate images provided, so visual learners are out of luck here
• Jupyter format means this is pedagogical scaffolding, not production infrastructure

Verdict Grab this if you’re a developer or student who has read the AlphaZero papers and wants to watch the loop actually run. Skip it if you need a battle-tested engine or GPU-cluster-scale training.