Playable demos & bots
Honing puts small open decision models into Red Alert 2 and a bouncing-shell tank game, then feeds what happened back into their training and the code around them.
Only you can see your notes.
Recording unavailable. Open the experiment ↗
In the Red Alert 2 clip, NanoJev sells its own construction yard and loses; after supervised training on a teacher's matches, the same model builds a base and wins. The Tank War clips work differently: the weights never change, only the controller does, and version 4's dodging comes from local geometry code. Haoze He says no human has beaten the trained models yet, while the README calls the training loop unfinished.
You can test how an artificial intelligence model plays by running the tank game locally. Your developer connects your chosen model to the game through simple health and decision web addresses. You can also play against a preconfigured model directly in your browser on the project's demonstration website.
You can also borrow the Tank War lesson. The same model played much better once the code around it changed what it was told, which moves it could pick and how its choices were checked, while simple geometry code dodged incoming shells. The training loop is still early work, and the public arena does not accept outside models yet.