Major AI labs are converging on the concept of "World Models," which focus on predicting future states or observations in various environments, such as video games, codebases, or markets. Yann LeCun is leaving Meta to establish a new lab dedicated to this area, while Ilya Sutskever describes emotions as value functions in a way that emphasizes their role in planning and simulation. Google introduced Genie 3 for world simulation, and Demis Hassabis hinted at dedicating significant research time to this topic. OpenAI's Sora and Meta's code world model (CWM) demonstrate the importance of understanding not just patterns but also the causal laws governing environments, with the 32B CWM outperforming larger models on benchmarks related to code execution. World models have existed in various forms for decades, even if the term is new. Recommendation engines, algorithmic trading systems, and weather models all function as world models by predicting outcomes based on interventions. These models are essential in adversarial domains like finance and geopolitics, where opponents adapt and respond. Traditional pattern matching fails in these dynamic environments, necessitating a shift to models that incorporate the actions of adversaries. Current language models (LLMs) struggle with this because they primarily imitate past data rather than simulate competitive dynamics. Training on actual competitive outcomes can enable models to learn the underlying dynamics of these environments. Language plays a pivotal role in making world models more accessible. With LLMs, it’s possible to input complex market signals and translate them into actionable predictions without human intervention. This capability suggests that world models could be integrated into transformer architectures, allowing for the simultaneous processing of language and state transitions. The addition of value functions enhances these models by providing a way to evaluate the desirability of predicted outcomes. This framework not only aids in multi-step tasks but also addresses the long-standing "credit assignment" issue in reinforcement learning, enabling models to learn which actions lead to favorable states. Ilya's comparison of emotions to value functions offers a human perspective on how we evaluate potential outcomes, further linking human intuition to AI development.