I’d like to share a project I’ve been working on related to KataGo AI. I used an existing TensorFlow.js-compatible KataGo bare neural net from this repo. This is a small, early KataGo network adapted for browser use.
To keep the AI strength manageable, the app doesn’t use all the network’s input features, limiting the strength to about 2 dan. A hack was implemented to simulate a weaker level close to GnuGo, around 6 kyu.
On top of this, I built a simple, clean GUI so you can play against the AI or with a friend locally or AI vs AI, directly in your browser—no installation needed.
For “learning from the AI” this is a great interface.
The only questionable thing so far was that there were a few times when it said I should rethink the choice, but the move that I’d chosen was marked as “A” - the best alternative move to choose
In my humble opinion, this is a fantastic resource - thank you so much for putting it together.
The one tiny User Interface quibble I would bring up is - I strongly suggest considering having Background Music OFF as the default rather than having the user’s browser immediately launch music when the site loads. It might be a Western Hemisphere kind of preference, but I know it’s a pet peeve for a lot of internet users. Just a heads up.
That would require to send the board state(or a board screen shot) to LLM (chatGPT) or a very complicated heuristics. (but it is not impossible, probably would be slow)
I’ve been wanting to do something like this. Make bot accounts on OGS for weak katago networks of various strength for people to play with.
I played your 6k. I won one and lost one as a 10k. I’d like a surrender option for early analysis.
