World models in AI have evolved significantly, distinguishing between different approaches. Traditional 3D models, like those from Fei Fei Li and Google, focus on visual representation. The Meta school explores more abstract concepts through frameworks like JEPA and Code World Models. A third frontier involves adversarial reasoning, where AI can understand and anticipate interactions among multiple agents. This is critical for applications in fields like law and finance, where the ability to predict how others will respond can mean the difference between success and failure. The article emphasizes the disconnect between how experts and AI evaluate situations. While AI, like ChatGPT, generates seemingly appropriate responses, it often fails to account for the nuances of human interactions. An example illustrates this: a poorly phrased Slack message could be misinterpreted by an overloaded colleague, while a more precise request would be prioritized. This highlights the importance of understanding the environment and the perspective of the recipients, which is something experienced professionals excel at. In the realm of adversarial models, the piece contrasts perfect information games, like chess, with imperfect ones, such as poker. Chess requires only knowledge of the board state, while poker demands an understanding of hidden information and opponent psychology. The AI Pluribus exemplifies advanced adversarial reasoning by calculating potential moves across all possible hands, making it unpredictable. This strategic depth is what sets effective human experts apart from AI that relies on surface-level pattern recognition. The true measure of success in these scenarios isn't whether a response sounds good, but how well it performs in a dynamic, adaptive environment.