Computer scientist Louis Castricato, after eight years researching large language models (LLMs)—the artificial intelligence technology behind chatbots such as ChatGPT and Claude—has shifted his focus toward a new area of AI development known as "world models." Feeling that fundamental LLM research has largely plateaued and now mainly revolves around applications, Castricato left his doctoral studies at Brown University to found Overworld, a Rhode Island-based startup focused on creating AI systems that understand and navigate physical environments, rather than solely processing language.

While significant investment continues to flow into AI chatbots, with developers like OpenAI and Anthropic drawing billions, an increasing number of AI entrepreneurs and researchers see world models as the next frontiers. These models aim to equip AI—and in some cases robots—with the ability to comprehend the spatial and temporal dynamics of the real world, encompassing how light interacts with surfaces, how objects respond to forces, and the physical laws governing these interactions.

Fei-Fei Li, founder of the San Francisco startup World Labs and often called the "Godmother of AI," describes world models as among the most critical yet ambiguously defined concepts in AI today. In a recent essay, Li highlighted the distinction between language models, which analyze textual data, and world models, which capture the statistical structure of space and time. She categorizes world models into three types: "renderers," which create visually rich but physically inaccurate scenes; "simulators," which faithfully replicate physical environments for training purposes; and "planners," which predict appropriate actions for AI agents or robots operating in unpredictable settings.

Yann LeCun, AI pioneer and former chief AI scientist at Meta who now leads Paris-based Advanced Machine Intelligence Labs, also emphasizes the growing prominence of world models. He characterizes them as enabling AI agents to anticipate the outcomes of their actions, a capability he sees as foundational for more autonomous, interactive systems.

Some experts underscore the limitations of current generative AI models—which operate by predicting the next word or pixel in a sequence—in handling physical, dynamic interaction. Martial Hebert, dean of computer science at Carnegie Mellon University and a veteran in robotics research, notes that language-based AI cannot perform tasks such as picking up a coffee mug, which involves complex spatial geometry, movement dynamics, and tactile feedback. Hebert refers to world models as a pathway toward "physical AI" or "embodied AI," an evolution of traditional robotics characterized by AI systems with a generalized understanding of and adaptability within their environments, much like the human nervous system regulates bodily functions without conscious effort.

Castricato’s Overworld exemplifies how world models are being applied beyond robotics. The startup develops interactive video game environments, such as adaptable virtual forests, where characters can engage with objects and spaces in detailed and responsive ways. This approach prioritizes interactivity over mere visual realism, aiming to create immersive user experiences that differ from existing static or pre-scripted virtual worlds.

Venture capital interest in world model-focused companies is growing despite the technology’s early stage. Steve Jang, co-founder and managing partner at Kindred Ventures, which backs Overworld as well as other firms like Causal Labs (working on AI for weather forecasting) and Extropic (specializing in computer chips optimized for world models), predicts a future with diverse model types and design philosophies rather than a single dominant AI architecture.

Researchers and industry leaders acknowledge ongoing challenges as the field matures. Li’s proposed taxonomy aims to clarify overlapping definitions and competing visions within the world model space, reflecting the complexity of building AI systems capable of functioning effectively in physical and virtual worlds alike. The race is on to develop a "planner" world model that enables robots to make meaningful, autonomous decisions—an achievement seen by many as the key to the next phase of AI progress.