Large language models excel at processing text but fundamentally cannot predict what happens when you act on the physical world. This critical limitation is driving a major shift in AI research toward world models, a new category of AI that learns to simulate how the physical world actually works by training on video and action data. Companies like DeepMind, World Labs, and AMI Labs are racing to build these systems, with AMI raising $1.03 billion and World Labs securing $1 billion in recent funding. Why Can't Language Models Understand Physics? Today's most advanced AI systems, including large language models (LLMs) and vision-language models (VLMs), suffer from what researchers call "jagged intelligence." They can solve complex math olympiads but fail at basic physics because they lack grounding in physical causality. Turing Award recipient Richard Sutton explained the core problem in a recent interview: LLMs "just mimic what people say instead of modeling the world, which limits their capacity to learn from experience and adjust themselves to changes in the world". This means these models show brittle behavior and break with even tiny changes to their inputs. Google DeepMind CEO Demis Hassabis echoed this concern, pointing out that today's AI models are missing critical capabilities regarding real-world dynamics. As AI moves out of web browsers and into physical spaces like robotics, autonomous driving, and manufacturing, this limitation becomes a hard ceiling. You cannot safely deploy a robot or self-driving car that cannot reliably predict the physical consequences of its actions. What Are the Three Different Approaches to Building World Models? Researchers have developed three distinct architectural approaches to solve this problem, each with different tradeoffs and applications. - Latent Representation Learning (JEPA): Endorsed by AMI Labs and pioneered by Yann LeCun, this approach learns abstract features instead of predicting every pixel. It mimics how humans understand the world by tracking a car's trajectory and speed without calculating light reflections on every leaf. This method is highly compute and memory efficient, requires fewer training examples, and runs with low latency, making it ideal for robotics and self-driving cars. - Generative 3D Environment Creation: Adopted by World Labs, this method takes an image or text description and generates a complete 3D environment using Gaussian splats, a technique representing 3D scenes with millions of mathematical particles. These environments can be imported directly into standard physics engines like Unreal Engine, drastically reducing the time and cost to create complex interactive 3D environments for industrial design and robotics training. - End-to-End Generative Simulation: Used by DeepMind's Genie 3 and Nvidia's Cosmos, this approach continuously generates scenes, physics, and reactions in real-time without relying on external engines. The model itself acts as the physics engine, maintaining strict object permanence and consistent physics at 24 frames per second, making it ideal for synthetic data generation for autonomous vehicles and robotics. Yann LeCun, co-founder of AMI Labs, explained that world models based on the latent representation approach are designed to be "controllable in the sense that you can give them goals, and by construction, the only thing they can do is accomplish those goals". How Are Companies Using World Models Today? The applications emerging from these systems are already tangible. Waymo, an Alphabet subsidiary, built its world model on top of DeepMind's Genie 3, adapting it specifically for training its self-driving cars. Nvidia Cosmos uses its end-to-end generative approach to scale synthetic data and physical AI reasoning, allowing autonomous vehicle and robotics developers to synthesize rare, dangerous edge-case conditions without the cost or risk of physical testing. In healthcare, AMI is partnering with healthcare company Nabla to use its latent representation architecture to simulate operational complexity and reduce cognitive load in fast-paced healthcare settings. The enterprise value is clear: Autodesk, a major industrial design software company, is heavily backing World Labs to integrate these models into their applications. Beyond enterprise applications, world models are still in what researchers call the "GPT-2 era," a phase of mass exploration rather than mass commercialization. Jeff Hawke, CTO at Odyssey, explained that "we are in the GPT-2 era of world models. This is a phase of mass exploration, not mass commercialization." Current models can predict 1 to 2 minutes of coherent video, and early use cases include gaming, retail simulations, and robotics applications. What Are the Key Limitations and Challenges? Despite the excitement, world models face significant hurdles. The end-to-end generative approach requires enormous computational resources to continuously render physics and pixels simultaneously. Training these models demands internet-scale public video data, and the selection of that data matters significantly. Current models struggle with prompt sensitivity, hallucinations, looping artifacts, and maintaining coherence over extended periods. The field also faces a terminology challenge. Researchers debate whether to call these systems "world models," "spatial intelligence models," or "video generators," reflecting disagreement about what these systems fundamentally are and what they can do. Additionally, questions remain about how to handle sensors beyond video, such as lidar data for autonomous driving, and how to balance physics realism against computational efficiency. Why Should You Care About World Models Right Now? World models represent a fundamental shift in how AI systems will interact with the physical world. Unlike LLMs, which process abstract knowledge through text prediction, world models learn to predict dynamics from visual observations and actions. This allows them to reduce situations that are computationally difficult to simulate at scale into a single fixed-cost operation in a neural network. The implications are profound. In traditional computing, simulating N agents or objects is at least an O(N) or O(N2) problem, meaning complexity grows exponentially with the number of elements. A world model simulates an entire stadium of fans, each with independent behavior, as a fixed-cost forward pass through the neural network. This efficiency breakthrough could unlock progress in embodied AI, robotics, and autonomous systems in ways that current model architectures cannot. The convergence of massive funding, competing architectural approaches, and real-world applications suggests world models are moving from research curiosity to foundational infrastructure. As these models mature, we are seeing the emergence of hybrid architectures that combine strengths from each approach, such as DeepTempo's LogLM, which integrates elements from LLMs and latent representation models to detect cybersecurity anomalies.
