StarCraft, a complex real-time strategy game, has long been a proving ground for AI. While traditional reinforcement learning methods have made strides, they often require vast amounts of training data and produce “black box” solutions that are difficult to interpret. A new research paper explores a fascinating alternative: using Large Language Models (LLMs) to generate decision trees that guide in-game actions. This approach, called LLM-SMAC, tackles the StarCraft Multi-Agent Challenge (SMAC) where the AI controls multiple units against an opponent. Instead of learning through trial and error like traditional reinforcement learning, LLM-SMAC uses LLMs to analyze the game scenario and generate code that represents a decision tree. This code dictates the units’ actions based on different in-game conditions. The model is designed in three parts: a planner that devises a high-level strategy, a coder that translates the strategy into executable Python code, and a critic that evaluates the code's performance and suggests improvements. The results are promising. LLM-SMAC achieves high win rates in various SMAC scenarios, often with minimal training. This method also addresses some of the shortcomings of traditional AI. The decision tree code is highly interpretable, showing clearly how the AI makes decisions. This “white box” approach is crucial for understanding and improving AI strategies. Furthermore, the policies learned by LLM-SMAC show good transferability. Strategies learned in one scenario can often be applied to similar scenarios without retraining, something traditional reinforcement learning struggles with. By leveraging an LLM’s existing knowledge base, LLM-SMAC significantly reduces the need for lengthy training periods, increasing efficiency and making the process more practical. The research also highlights the importance of incorporating game-specific information, such as unit attributes and map layouts, to improve the AI's decision-making. While this is early-stage research, it opens exciting possibilities for using LLMs to solve complex decision-making problems in gaming and beyond. Future work will focus on further refining the LLM’s ability to generate code and exploring the potential of training LLMs directly on decision tree code itself, pushing the boundaries of what’s possible in game AI.