The majority of existing AI products are powered by large language models. They’re the technological foundation for chatbots like ChatGPT and Gemini, creating the natural-sounding text you see on your screen. But it may not be the most consequential AI technology.
The AI evolution has gone through many phases, and this next one is going to be less concerned with creating words and more focused on understanding our natural world. World models are built to translate our physical world — such as the laws of physics, object detection and movement — into a digital blueprint that AI can understand.
You likely won’t interact with world models in the same way you do with LLM-powered tech, like through chatbots. However, the world models will demonstrate how AI can create realistic videos, guide surgical robots and enhance autonomous vehicles’ driving capabilities. They’re important building blocks in developing what’s called physical AI — tech that not only understands our world but can take actions in it.
Don’t miss any of our unbiased tech content and lab-based reviews. Add as a preferred Google source.
A variety of AI pioneers have signaled a shift toward building a world model. Yann LeCun, a leading AI pioneer, has recently left his role leading Meta’s AI efforts to join a startup focused on building world models. Fei-Fei Li, colloquially known as the godmother of AI, has said spatial intelligence — the ability to understand your physical environment — is the next frontier for tech innovation.
“Spatial intelligence will transform how we create and interact with real and virtual worlds — revolutionizing storytelling, creativity, robotics, scientific discovery and beyond,” she wrote in a November blog post.
Nvidia CEO Jensen Huang also dedicated a portion of his CES 2026 keynote to the company’s efforts in world models. Building an AI model that’s grounded in the laws of physics and ground truth starts with the data used for training, Huang said.
Watch this: Every Announcement from the Nvidia Live CES 2026 Stream
AI models of every flavor require immense quantities of data to build and refine their outputs. Typically, AI companies rely on content created by real humans — with and without their permission — which has led to major legal showdowns. World models can be built with human data, including simulations. That data is essential to building world models that can reason and make cause-and-effect judgments.
Nvidia’s world model Cosmos uses text, image and video to understand the physical world.
One area where Nvidia is using world models is in self-driving cars. In a live demo, Nvidia demonstrated how its world model, Cosmos, uses a car’s sensors to understand its own position and that of every other nearby car on the road to create a live video of its surroundings. Developers can use that information to run scenarios, like car accidents, to see how the vehicle would respond and make necessary safety improvements. Synthetic data, or nonhuman-generated data, can also be helpful in tandem with world models to help predict rare “edge cases.”
As AI continues to be integrated into every part of our online lives, it’s essential that it can understand our physical world, rather than continue to hallucinate and make mistakes. Renewed research and investment from industry leaders in spatial intelligence, world models and physical AI show that the industry isn’t just going to build more chatbots — it’s working on building AI that’s more rooted in our reality, rather than the other way around.
