Trillion-dollar LLM era challenged: World models learn space, not just text

For machine learning engineers, LLMs have been a triumph, but they're blind to the physical world. World models tackle the hard problem of spatial and temporal reasoning, as explained by Fei-Fei Li: learning how light falls and objects obey physics. This technical pivot is drawing top talent away from chatbots and into a new research era.

A significant shift is underway in the artificial intelligence community as leading researchers and entrepreneurs pivot from the language models that powered chatbots like ChatGPT and Claude to 'world models' designed to understand and interact with the physical world. This transition was crystallized recently when Louis Castricato, an eighth-year doctoral student specializing in large language models (LLMs) at Brown University, quit his PhD to launch Overworld, a startup dedicated to AI that comprehends spatial and temporal environments rather than just text. 'We basically have passed the point of doing real fundamental LLM research,' Castricato told the Associated Press. 'Now it's just applications.' His move is part of a broader trend that includes some of AI's most respected names: Fei-Fei Li, the 'Godmother of AI' and founder of World Labs, and Yann LeCun, a pioneer who left his post as Meta's chief AI scientist in 2025 to found Advanced Machine Intelligence Labs in Paris. These departures underscore a growing belief that the next frontier for artificial intelligence lies not in generating ever more fluent text but in mastering the physical realm—a domain that could unlock robotics, autonomous navigation, and hands-on human-AI collaboration.

“

Overworld, World Labs, and Advanced Machine Intelligence Labs are all early-stage efforts, but they carry the cachet of names that have defined modern AI.

At the heart of world models is the ambition to replicate how humans intuitively understand space and physics. In an essay published this month, Li explained: 'Where language models learn the statistical structure of text, world models learn the statistical structure of space and time: how light falls on a surface, how a garden looks from an angle no camera has captured, how objects respond to force and follow the laws of physics.' This approach moves AI from a flat, symbolic representation of knowledge into a dynamic, embodied intelligence. LeCun, appearing on the 'Unsupervised Learning' podcast, acknowledged that 'world model is quickly becoming a buzzword,' but the underlying technical challenge is profound: building systems that can predict, plan, and act in three-dimensional environments. For Castricato, this meant walking away from a nearly completed doctorate—a personal gamble that reflects both disillusionment with the diminishing returns of LLM scaling and the tantalizing promise of what comes next.

The trend is unfolding against a backdrop of unprecedented financial commitment to traditional LLM development. Investors have poured trillions of dollars into companies like Anthropic and OpenAI, betting that language models will become the backbone of a new digital economy. Yet even as these firms raise ever larger rounds, some top talent is looking elsewhere. The exodus suggests a maturation point: fundamental research into transformer architectures and pretraining paradigms may be giving way to applied engineering, while the scientific vanguard turns its attention to a more complex problem. Overworld, World Labs, and Advanced Machine Intelligence Labs are all early-stage efforts, but they carry the cachet of names that have defined modern AI. Their emergence signals to venture capitalists and corporate strategists that the next generation of value creation may come from AI that can physically manipulate its environment.