Imagine an AI that not only plays games but also designs them, complete with rules and logic. That's the promise of PIANIST, a new framework that leverages Large Language Models (LLMs) to generate game world models. Traditionally, creating AI agents for games requires painstakingly programming the rules and environment dynamics. PIANIST flips the script. By breaking down a game's world model into seven digestible components—information sets, hidden states, actors, action functions, transition-reward functions, information partition functions, and information realization functions—PIANIST allows LLMs like GPT-4 to generate the code for each piece. Give it a natural language description of a game like Goofspiel (a card game) or Taboo (a word-guessing game), and PIANIST guides the LLM to produce a working model. This model can then be used with algorithms like Monte Carlo Tree Search (MCTS) to let the AI plan and make decisions. Researchers tested PIANIST by pitting LLM-generated agents against ground-truth models (using actual game rules) and even human players. While the AI held its own against the perfect-information models, showing it could accurately learn the game dynamics, it struggled against human cunning. This highlights a key challenge: while LLMs excel at understanding and representing game logic, they still lack the strategic nuance of human players, especially in games with partial information, like Taboo or poker. The exciting part? This is just the beginning. PIANIST demonstrates that LLMs can grasp the core mechanics of complex games. As LLMs continue to evolve, we can expect even more sophisticated game-playing AIs and potentially entirely new games designed by AI, blurring the line between player and creator.