This wasn’t just another software update. It was the first time an AI beat a professional human player (Yoshio Ishida, 9p) at even odds using a neural network.
He kept the MCTS engine, but he added a as a "co-pilot."
[Your Name] Category: Go & AI History
Before the deep learning explosion of 2016, there was . And in 2014, the world saw its true turning point: Crazy Stone Deep Learning, The First Edition .
If you only know Lee Sedol vs. AlphaGo, you are missing the prequel—the scrappy, brilliant, and often-overlooked origin story of modern Go AI. Crazy Stone Deep Learning The First Edition
Liked this? Check out my post on "The 3 Ancient Joseki that Still Break Modern Neural Nets."
Crazy Stone Deep Learning, The First Edition: The Moment the Machine Learned to "Feel" the Board This wasn’t just another software update
Let’s rewind and look at why this "first edition" was so crazy. Classic Go bots (Gnugo, early Crazy Stone) relied on Monte Carlo Tree Search (MCTS) . They played millions of random games in their head and guessed the best move based on statistics.
This worked well for amateurs but hit a wall at the professional level. Why? MCTS is terrible at intuition . It doesn't know a good shape from a bad one; it just knows brute-force probability. The "First Edition" of Crazy Stone with deep learning was a hybrid beast. The developer, Rémi Coulom (a French programmer), did something radical. And in 2014, the world saw its true
Crazy Stone Deep Learning, The First Edition wasn't perfect. But it was the first time a machine stopped looking like a calculator and started looking like a Go player.