💭 given adequate spatial representation and an engine providing legal moves and simulations, llms could play any board game badly. Combine this with techniques like mcts (llm provides intuitive strategy, mcts prefers winning states) and you might have a decent baseline ai for any game where those things can be provided.

Can an engine be made such that llms can code in needed rules on the fly? I imagine if you already have an mtg engine adding an arbitrary new card is something a modern llm could do with sufficient tooling. If the game can be iteratively built up that way, all that's left is the gui