Large language models (LLMs) have shown remarkable abilities in generating text, translating languages, and even writing different kinds of creative content. But can they actually understand and model the complexities of the real world? Researchers are exploring this question by using LLMs as “world models” for decision-making. Imagine an AI agent trying to solve a problem, like figuring out how to make a campfire or correctly giving a blood transfusion to a patient. Instead of relying on traditional programming, the agent uses an LLM to simulate the consequences of its actions. This research dives into how effectively LLMs can predict outcomes and guide an agent toward successful decisions. The study uses 31 diverse tasks, ranging from everyday chores like washing clothes to complex procedures like forging a key. By testing two powerful LLMs, GPT-4o and GPT-4o-mini, across these various scenarios, the researchers investigated three key aspects of world modeling: Can the LLM accurately verify whether a sequence of actions will achieve a goal? Can it suggest helpful actions the agent could take? And, finally, can it plan an entire sequence of actions to solve the task? The results reveal a nuanced picture. While GPT-4o generally outperforms GPT-4o-mini, especially in tasks requiring domain-specific knowledge, both models struggle with long-term planning. This suggests that while LLMs demonstrate promise as world models, there's still a significant gap between their abilities and true human-like understanding. Furthermore, combining multiple functions within the world model—like predicting outcomes and suggesting actions—introduced instability, highlighting the complexity of building reliable AI decision-making systems. This research is a stepping stone toward more robust and capable AI agents. Future work could focus on developing more stable systems and improving the LLMs' ability to reason about long-term consequences, bringing us closer to AI that can truly understand and interact with the world like we do.