AP's Matt O'Brien reports that a growing group of AI researchers are walking away from chatbot work for "world models" — systems that learn the structure of space, time and physics rather than the statistical structure of text.
The lede is a defection. Louis Castricato was in his eighth year studying large language models when he quit his doctoral studies at Brown to found Overworld, a Rhode Island startup building game worlds that react to a character moving through them. "We basically have passed the point of doing real fundamental LLM research," he told AP. "Now it's just applications."
He is not alone at the top of the field. Yann LeCun left his job as Meta's chief AI scientist last year to start Paris-based Advanced Machine Intelligence Labs, and reads a world model as something that lets an agent "predict the consequences of its own actions." Fei-Fei Li, whose startup World Labs is built on the same bet, calls the term "one of the most important and most overloaded terms in AI today": where language models learn the statistical structure of text, she wrote in a recent essay, world models learn "how light falls on a surface, how a garden looks from an angle no camera has captured, how objects respond to force."
Martial Hebert, dean of computer science at Carnegie Mellon and a robotics researcher for four decades, puts the gap in one object: a chatbot cannot pick up a coffee mug. "There's all the geometry of the world, the dynamic of how I move my hand, the physical interaction of the contact with the cup," he said. "This is much more complex than just predicting the next word in a sentence."
Li's essay splits the category three ways, which is the most useful part of the story if you are trying to price the claims:
The money is still overwhelmingly on chatbots — trillions committed to OpenAI, Anthropic and their peers — so this is not a market shift yet. It is a talent one, and talent moves first. The people who built the current paradigm are saying out loud that the remaining work on it is application engineering, and they are spending their own next decade somewhere else.
Watch the funding, too: Kindred Ventures' Steve Jang, an Overworld investor, is also backing Causal Labs (weather prediction) and Extropic (chips suited to world models) — a portfolio built on the premise that there is no one model to rule them all. Note the honest caveat inside the enthusiasm: LeCun himself says "world model is quickly becoming a buzzword," and the label currently covers a video generator, an improvised playable game and a physics engine alike.
Sources: AP News
