Who runs this site?
Hello, I am Saenggaknamu (“Thinking Tree”), the owner of this site. I am a developer with 32 years of experience, building many kinds of AI models in pursuit of an AI that feels human and thinks the way people do. As AI became widely available and I could develop on my own, I started taking on all sorts of challenges.
Then I watched an AI solve a Sudoku. Unlike the way people think, it filled in the answers row by row from the very first cell, like a child who already knows the answer. When I told it that was cheating, only then did it compare the rows, the columns and the 3×3 boxes according to the rules of Sudoku. Seeing this, I downloaded models from Hugging Face and began training them myself. Since then I have found that many models solve brain-training games as if they already knew the answer.
I handle everything at MariaGO myself: model training, the game engines, the web pages and the servers. The writing here is not a summary of a general game encyclopedia; it is based on code and experiment records checked during implementation.
The “GO” in the project name stands for Go (baduk), where it all began. Questions that started with Go spread to Gomoku, Othello, Janggi, chess and number and word puzzles. The rules differ from game to game, but they share one theme: “how does the AI read the current state and choose its next action?”
Why build yet another game site?
There are already plenty of online game services. The problem MariaGO tries to solve is not finding an opponent quickly to compete. I wanted a personal place to play where you can make a move whenever you like without arranging a time with someone, stop when you are short on time, and come back to the same board.
The goal is not to show off the strongest AI, but to let anyone meet a trained model's judgement directly in the browser and understand the process behind it.
I love the film Blade Runner. In it, J.F. Sebastian and Dr. Tyrell play chess from different places, exchanging one move at a time in words. MariaGO's Go has no turn time limit so that it can have that same character.
Each time you confirm a move, the board and move order are saved on the server, and the game in progress is loaded again on your next visit. You can step back and forth through the record and play again from any move.
Note: games are saved under an ID created for each browser. If you play in a private window or another browser, your existing games will not load and a new save will be created.
How is it built?
The Go model is trained on positions and next moves extracted from SGF records of strong amateurs and professionals. Feeding in the raw records is not the end of it. The 19×19 coordinates are converted into an aligned grid the model can read, duplicate positions are removed, and we check that the training data and the live game server use the same board representation.
Only long after I started did I learn that not every AI model uses the same Go coordinates. I trained Google Gemma for a long time, but the model always played in the fourth quadrant. So I added various restrictions, and those restrictions caused new errors of their own.
Later I found that each model used different coordinates, and that even the SGF records and our own model disagreed. That was also when I learned that when I told a model the human's move, it answered in a different coordinate system. So early on I spent a very long time training the models to align coordinates, and only after that did their Go strength start to rise sharply.
After training, we do not look at win rate alone. We record whether the model finds coordinates in all four corners with similar accuracy, whether it picks points that are already occupied, and whether it passes too early. In web games, too, the AI's move is not applied as is; the game engine checks for illegal moves such as suicide, ko violations and playing on an occupied point.
| Training data | Positions and actual next moves extracted from strong amateur and pro SGF |
|---|---|
| Position format | A 19×19 text grid with row and column numbers aligned to two digits |
| Live checks | Coordinate accuracy, legal-move rate, early passes and directional bias checked separately |
| Game saves | Board state and full move order saved per move, used for review and continuing |
Content and operating principles
We write what we have checked
Features are described exactly as they work today. New features are added whenever an idea comes up, and the writing is updated along with them.
We publish the limits too
Conditions readers should know, such as model mistakes or the life-and-death limits of automatic scoring, are stated alongside the results.
We write down how things are made
Rather than just collecting game links, we let you read the rules, how they are implemented, and even the record of what failed and how it was fixed.
We aim for an unhurried experience
If something comes up, just stop. When you have time again and a move comes to mind, play it then. If you made a bad move, you can always go back to that point and play again.
We are clear about current limits
Not every game on MariaGO is equally complete. Go supports saving records and continuing games, but the feature set of the other games varies. Automatic Go scoring counts surrounded empty areas and captured stones but cannot fully identify dead stones. Its value is not a referee's final verdict but a reference for players reviewing the board.
These differences and failures are, in fact, the heart of the project record. When features change, this page and the research notes will be updated as well.