MTG

Reminds me of that Twitter bot roborosewater

You could totally have that and it wouldn't need to play like shit at all.

What you'd want to do is pair up a few different layers of AI.

The base layer would be a fine tuned LLM describing card mechanics, flavor text, and art. You'd use this to generate a ton of synthetic cards for each color deck, ideally directing it using manual prompt controls to create a mix representative of actual frequencies of various costs and types so you have a sample that will pace well with typical game play.

Then you'd need a model that plays Magic adeptly according to the rules, which should be able to be accomplished with something like AlphaZero, an approach that works really well for learning games.

You'd gradually advance that second model from playing well with premade decks to picking its own deck from legal cards until it was able to beat versions of itself using competition winning decks.

Then you gradually start mixing in your synthetic cards, removing less used legal cards as you go until you finally have added all synthetic cards and have no legal cards.

Then you reduce the synthetic cards to the most used X number of cards for each color or mixed color decks being built by the competitive AI.

Finally, you have a set of AI generated cards that would likely be pretty fun to play.

While right now that's a pretty expensive and labor intensive process, within the decade it's the kind of thing you'd be able to set up in a weekend on local hardware as long as you have a digital copy of all legal Magic cards in a standardized text format and competition deck records in the same.

@paddirn @MysticKetchup ChatGPT will create a commander deck list for you.

It wont be a legal deck list.