December 20, 2025
Heavy leaves, heavier drama
Shallow trees with heavy leaves (2020)
Brains over brawn? Chess AI tactics spark a spaceship-hunting flame war
TLDR: A new Game of Life search tool favors fewer, smarter checks using chess-style AI and logic solvers, finding exotic patterns. Comments explode over brains vs brawn, credit to a 2001 pioneer, and whether picking solvers with 'reinforcement learning' is genius or just gambling.
Chess meets pixel art and the comment section is on fire. The post compares Stockfish’s brawn (searching zillions of positions) with Leela Chess Zero’s brainy vibe (fewer positions, deeper thinking), then ports that debate to hunting “spaceships” in Conway’s Game of Life. Old-school tool ntzfind digs forever; new kid ikpx2 prunes hard, asks SAT solvers (logic puzzle engines) for help, and even uses a tiny dash of reinforcement learning to pick between kissat and CaDiCaL. Strong opinions explode: one camp yells “speed rules,” the other chants “think smarter,” while a third insists credit belongs to Paul Tooke’s 2001 trick that ikpx2 revives.
Drama escalates when folks call solver-picking a “casino bandit” and meme “heavy leaves fall faster.” Others swoon over parallel threads and priority queues because it “finally fits in memory.” Jokes fly: “choose your fighter: 100,000,000 moves vs 40,000 big brains,” and “trees so shallow they’re influencers.” Some worry kissat’s lack of incremental support; others ask for CryptoMiniSat and neural pickers next. The feel-good camp cheers that simulating extended partials finds weird tails and debris the main search misses. The skeptics clap back: “that’s just lucky accidents.” The vibe: chaotic, funny, and deeply split—classic internet brawl, now with gliders.
Key Points
- •Stockfish and Leela Chess Zero exemplify contrasting search strategies: wide alpha-beta vs. deep neural MCTS.
- •Ikpx2 applies a “shallow tree with heavy leaves” approach to Game of Life spaceship search, unlike ntzfind’s depth-first enumeration.
- •Ikpx2 uses SAT solvers for deep lookahead (typically ~30 rows) to prune near-dead ends without full subtree traversal.
- •Kissat (optimized but non-incremental) and CaDiCaL are used, with RL (multi-armed bandit) selecting the solver; CryptoMiniSat is a candidate addition.
- •Extended partials are simulated directly in the automaton, sometimes yielding new objects—a concept first explored by Paul Tooke in 2001.