Remix.run Logo
▲ mochidusk 6 hours ago

I built Brewhaven[1], a card game harness with agents that can play MTG's 4-player commander.

I was curious to see how frontier models fare when it comes to playing a complex game AND if it can scheme, barter, bluff, banter, etc...

All the frontier models are very capable at gameplay but only Claude impressed me with table-talk/diplomacy. They're not superhuman at MTG yet, but the more interesting takeaway is this:

Current day AI is the ultimate translation machine; if you can translate something X -> Y, you can also go in reverse and do Y -> X; think text-to-image and image-to-text.

So if we have something that was trained on a corpus of all human data -> plays MTG and scheme/bluff/make deals then it's possible something like Brewhaven can generate trajectories that not only improve MTG gameplay but also general LLM performance on tasks that requires long-term planning (and sparse rewards). Basically, my belief is that if it can play MTG competently, it can generalize to do just about anything competently in other domains.

[1]https://x.com/duskqho/status/2119890372663153140