ai
AI Researchers Shift Focus to World Models for Physical Environment Understanding
Computer scientists are moving from language models that process text to systems designed to interpret space, time, and physical interactions. Several startups and established researchers are developing these models for applications in robotics and interactive simulations.
news.mit.eduAudio version
Tap play to generate a narrated version.
Developing·Limited corroboration so far. This page will refresh as more sources emerge.
Computer scientist Louis Castricato left doctoral studies at Brown University to found Overworld, a company developing AI systems that can navigate physical environments rather than process text alone. Castricato said large language model research has shifted from fundamental work to applications.
Overworld is building interactive video game environments where virtual characters respond to objects and spaces in real time.
Li, founder of World Labs, wrote in an essay this month that world models learn the statistical structure of space and time, including how light falls on surfaces and how objects follow physical laws. Yann LeCun, who left Meta last year to start Advanced Machine Intelligence Labs, described world models as systems that enable AI agents to predict the consequences of their actions.
Martial Hebert, dean of computer science at Carnegie Mellon University, said language models cannot capture physical interactions such as grasping objects or adapting movement. Hebert noted that world models could accelerate development of physical AI systems.
Steve Jang, co-founder of Kindred Ventures, said the firm is investing in Overworld, Causal Labs, and Extropic. Jang stated that multiple model architectures will likely coexist rather than a single dominant system. Li categorized world models into renderers focused on visual output, simulators that replicate physical structure, and planners that guide agent actions in unstructured settings.
