Don't Let LLMs Make Decisions: They're Terrible At It

The author, an NPC developer for an online game, argues against using Large Language Models (LLMs) for business logic and decision-making. Instead, LLMs should serve as a user interface, translating user requests into API calls and results back into natural language. LLMs are shown to be inferior in performance, debugging, and adjustment compared to specialized systems, using a chess-playing bot example. The author advocates for using LLMs for tasks like text transformation, categorization, and intent understanding, while relying on purpose-built systems for core logic. While LLMs will continue to improve, this architectural principle remains crucial.