AI board game laboratory
MariaGO is not a portal full of game links. It is a personal development project where people play board games against AIs trained on real game records, and where we write down what we discover while implementing the models' judgement and the game rules ourselves.
What makes it different
Rather than winning quickly, we value watching why the AI chose a move and coming back to continue a game you paused.
The Go model learns the relationship between positions and next moves from strong amateur and professional game records. Coordinate bias in the training data and illegal moves are checked and fixed through real games.
Go games have no byo-yomi or turn timer. Tap an empty point once to preview the move, then tap again to confirm, so there is no need to hurry even on a small screen.
The board and game record are saved after every move. Come back in the same browser and your game in progress is restored, and you can branch a new variation from any point in the record.
Field notes
Instead of repeating general game rules, we record the problems we actually ran into while building MariaGO and the limits of the current implementation.
Five games against R5, a match where the scorer chose the moves, and three games Gemma 4 lost after studying the heuristic. The on-screen AI choice applies only to that browser.
We reduced premature passes and double-atari risks, and told ko apart from capturing after filling, based on the current board shape. Per-client ranks and records and improvements to the evaluation heuristic are also covered.
The AI now reads ladders where escaping only leads back into atari and avoids those moves, and in ladder positions it looks for another legal move instead of passing.
The AI no longer steps into an opponent's tiger mouth where it would be captured at once, nor fills its own without reason. Moves that capture stones or save nearby groups remain allowed as exceptions.
When the grid was written differently, the model kept playing only in the lower right corner. Here is why we aligned the coordinate format between training and play and checked accuracy per quadrant.
Counting empty points is not the same as deciding which stones are dead. We explain what the current scoring calculates and what to keep in mind when reading its result.
Playground
Here is what you can do before you follow a link. Each service opens in its own game room, and features and polish differ from game to game.
Play against the trained model, review saved game records and check the score of the current board.
Take on the AI's reading in a game of trapping and flipping discs.
Read threats across rows, columns and diagonals as you and the AI race to connect five.
Korean chess, where you weigh the moves of the chariot, cannon and horse against the limits of the palace.
From developing pieces to king safety, experiment with choices that change with every position.
Fill rows, columns and 3×3 boxes without repeats and check your reasoning step by step.
Use the single empty space to put the tiles in order and try to shorten your path.
A strategy game where you clone and jump to spread out and turn your opponent's pieces into yours.
Complete the words using the letters where across and down clues meet (Korean).
Start with 30 billion chips, choose your cards and bets, and compete with the Cheolsu, Younghee and Minsu AIs for the last stack standing.
From the maker
Behind the single sentence “the AI plays” lies a chain of separate problems: selecting training records, converting coordinates, checking illegal moves, saving and reviewing games, and scoring.
MariaGO does not hide this process. For example, the current Go scoring counts surrounded empty areas, captured stones and komi, but cannot fully decide life and death on its own. The models make mistakes too, and we record those failures and feed them into the next round of training and implementation.
Why we started and how we run the project →