AI scientists who feel they have hit limits in large language model research are jumping into developing "world models" that understand and respond to physical environments, the Associated Press reported on June 25.

World models focus on getting AI to learn not only text but also the statistical structure of space and time. This includes how light falls on surfaces, how objects respond to force and the laws of physics.

Fei-Fei Li (페이-페이 리), founder of World Labs, described world models as "one of the most important and most discussed concepts in AI today," the report said. Yann LeCun (얀 르쿤), who resigned last year as Meta's chief AI scientist and set up AMI Labs in Paris, defined world models as "something that lets an AI agent predict the outcomes of its actions."

Marshall Eber (마샬 에베르), dean of the School of Computer Science at Carnegie Mellon University, pointed to the fact that chatbots cannot pick up a coffee cup as a limit of existing language models. "How to move a hand and physical contact with a cup are far more complex than predicting the next word," he said.

Luis Castricato (루이스 카스트리카토), who left a PhD programme at Brown University and founded Overworld, is focusing on having AI create interactive game worlds. "There was no world model where you can walk through a door or interact with detailed environments," he said.