Brian H. Liou
Backend and distributed systems engineer building AI infrastructure, game solvers, and measurement tools.
Active work
mistboard.comActive
A platform for original chess, xiangqi, shogi, and strategy-board variants, built around fair rules, serious online play, and strong engines for every supported game.
Recent writing
-
Building a Flip Jungle EngineA Flip Jungle engine build report: the search, the exact endgame tablebases, the skill ladder, and what a near-perfect engine says about the game's skill ceiling.Jul 2, 2026 -
Building a Dou Shou Qi EngineA Dou Shou Qi engine build report: the ruleset, the open-source benchmark, the measured search gains, the tablebase frontier, and why the obvious neural and AlphaZero-style upgrades are still open work.Jun 27, 2026 -
One Word Can Point at Six Cards, Not NineA single clue word pins down all nine of your cards on only 0.286% of boards. The embedding geometry permits the separation; no common word realizes it. And the clues a similarity decoder does accept still fail on a real reader.Jun 22, 2026 -
Building a Banqi EngineA Banqi engine build report: the rules, the benchmark, the evaluation fixes, two pathologies with interactive replays, and the AlphaZero climb I have not paid for yet.Jun 19, 2026 -
Solving Six Men's MorrisSix men's morris is strongly solved: a draw over all 42,372,745 reachable positions, the first published solution.Jun 19, 2026 -
Solving Three-Player Chinese CheckersAn exact tablebase for the smallest nontrivial three-player Chinese Checkers board, where the opening is a pure kingmaker position and draws appear only deeper in the game.Jun 15, 2026 -
Toward Solving Micro ShogiA 64 GB cloud solve of the K+P+G rung, a dense-rank rerun that cut memory 8.8x, and the current cost model for a complete Micro Shogi tablebase.Jun 12, 2026 -
Toward Solving Shogi4 (4×4 Shogi)Recovered rules, a validated solver, a 2.1-billion-position closed subgame, and a costed full solve left unrun.Jun 8, 2026 -
Solving Dōbutsu ShōgiDōbutsu shōgi, a children's shogi variant, is strongly solved: the second player wins in 78 plies. The results of the solve, and why one rule makes a tiny board so deep.Jun 4, 2026 -
We Are the LLM's MemoryIndependent engineers are assembling the OS layer for LLM coding agents in public. Here's the mechanism, who's winning, and what happens in the next 12 months.May 7, 2026