#What constitutes a world model in artificial intelligence?
Fei-Fei Li, a prominent figure in AI, aims to clarify the ongoing debate regarding what defines a world model. In her recent publication, she delineates a framework that categorizes world models into three crucial functionalities: rendering, simulating, and planning. This classification highlights the components that form a cohesive understanding of spatial intelligence in AI, allowing machines to effectively engage with physical environments.
#How do world models function?
The first component, rendering, serves a fundamental role in generating visual representations based on data inputs. This functionality is primarily seen in many existing AI systems, and Li asserts that those limited to rendering do not qualify as true world models.
Simulation takes this a step further by incorporating the principles of physics, causality, and temporal object interactions. While a renderer may illustrate a ball approaching a cliff, a simulator comprehends the underlying mechanics, predicting that the ball will fall.
Lastly, the planning function leverages insights gained from simulation to develop actionable strategies. This distinguishes between AI that merely observes a scenario and one that can autonomously navigate a kitchen to prepare a sandwich without causing a mess.
#How do these functions interlink?
Li emphasizes that these three functions are not isolated; they exist within an interconnected loop. Each function informs and enhances the others, creating a dynamic system. For instance, the renderer contributes visual context for the simulator, the simulator informs the planner with physical predictions, and the planner establishes priorities that shape the output of both the renderer and simulator.
#Why is this framework essential for robotics?
Li argues for the necessity of world models to bridge the divide between simulation and real-world applications. Achieving a highly accurate digital representation of the physical environment enables robots to undergo training in virtual spaces before operating in reality.
To put this theory into practice, World Labs has introduced its commercial product, Marble, demonstrating a tangible application of Li's framework. Marble creates immersive, high-fidelity 3D environments from multimodal input—allowing users to dictate environments using text or images. This innovative approach has practical implications for robotic simulations. Unlike traditional videos, which consist of fixed sequences, Marble maintains cohesive geometry and adheres to physical laws. Robots training in these environments can explore shelves from various angles and consistently identify objects in the same positions.
#What is the financial backing for this technology?
World Labs has garnered significant financial support, raising $1 billion in February 2026, building on a prior round of $230 million. Among its investors are notable names including AMD, Autodesk, NVIDIA, and Fidelity.
In total, the company has amassed $1.23 billion, placing it in an elite category among AI startups that emphasize spatial intelligence rather than participating in the current arms race of large language models.