Imagine a robot learning to navigate a new environment, not through explicit programming, but by observing the consequences of its actions and figuring out the "rules of the world." This is the fascinating challenge of domain induction in symbolic planning, where AI agents need to deduce the underlying logic governing a domain without human intervention. Traditional symbolic planners rely on meticulously hand-crafted action semantics (essentially, if-then rules about how actions affect the world). But what if we could automate this process? Researchers are exploring the exciting potential of Large Language Models (LLMs) to infer these action semantics from environmental feedback. In a new approach called PSALM (Predicting Semantics of Actions with Language Models), LLMs work in tandem with symbolic planners, proposing plans, observing the outcomes, and iteratively refining their understanding of the environment's logic. PSALM takes advantage of the LLM’s general knowledge and language processing capabilities to generate potential action sequences and then infers the preconditions and effects of each action based on what happens when those sequences are executed. This approach has shown remarkable success in several simulated environments, achieving 100% planning success rates after learning from just a single goal-directed task. The key insight is that LLMs, combined with a feedback loop and a structured way to represent knowledge, can learn complex rules by simply trying things out and seeing what works. This approach has exciting implications for robotics and broader AI applications. Imagine robots that can learn new tasks in unfamiliar environments without extensive programming, or AI agents that can deduce the rules of a game simply by playing it. While still in its early stages, PSALM opens up a promising pathway towards more autonomous, adaptable, and intelligent systems.