AI Researchers Walk Away From Chatbots to Build World Models

Louis Castricato was in his eighth year studying large language models when he left to build AI world models for video games. The Brown University doctoral candidate, who founded Overworld in Providence, Rhode Island, told The Associated Press the field had moved past its foundational moment. “We basically have passed the point of doing real fundamental LLM research,” Castricato said. “Now it’s just applications.”

Investors are still counting on chatbots. They have committed trillions of dollars to leading developers like Anthropic and OpenAI, and OpenAI is preparing a public listing. A growing circle of AI researchers and venture firms is placing a different wager. They are funding world models, systems that learn the physics of physical environments, the spatial counterpart to a language model.

Walking Away From Chatbots

Castricato quit his doctoral program at Brown and founded Overworld, a small startup in Providence, Rhode Island. The startup’s name states its thesis. Overworld is building AI that knows the shape of a world, with language as one input among many.

The first product is a video game engine. Overworld is building scenes, such as a spooky forest, that adapt as a virtual character moves through them. “There’s no other world model where you can just walk through doors or where you can interact with a detailed environment like this,” Castricato said. “We optimize for interaction above anything else.” Castricato says interaction is the test a chatbot cannot pass.

The pivot is small in headcount. The signal it sends is large: Castricato, LeCun, and Li are each putting their time behind AI built for three-dimensional space, not for the next token in a sentence. The world that text-trained models cannot read is the next round of work.

The Three Researchers Placing the Bet

Three of the field’s most prominent scientists are staking their next chapters on the world model idea. Fei-Fei Li, the Stanford computer scientist often called the “Godmother of AI,” is the founder of the San Francisco startup World Labs. Yann LeCun quit his post as Meta’s chief AI scientist last year to start Advanced Machine Intelligence Labs, a Paris-based research operation. Castricato rounds out the trio with Overworld. Each is putting their time behind systems that perceive, reason about, and act within three-dimensional space.

The scale of the move varies. World Labs has published a taxonomy essay on world models that has become a reference point for the field. LeCun’s AMI Labs research operation spans Paris, New York, Montreal, and Singapore. Overworld is a small Rhode Island team. All three are betting on AI built for three-dimensional space, with language as a layer on top.

What a World Model Does

Li’s essay is the cleanest attempt to pin the term down. “Where language models learn the statistical structure of text, world models learn the statistical structure of space and time,” she wrote. The point is that text-trained AI has no grip on how a room looks, how a cup tips, or how light falls on a surface.

A world model, in Li’s framing, learns how light falls on a surface, how a garden looks from an angle no camera has captured, how objects respond to force, and how they follow the laws of physics. It is the same statistical machinery that produces text from text, redirected at the geometry of the real world. World Labs calls this spatial intelligence, the skill the company was founded to build.

LeCun offers a complementary definition. On a recent “Unsupervised Learning” podcast, he called a world model the system that enables an AI agent “to predict the consequences of its own actions.” Both definitions share one move: replace the next word with the next state of the world.

That shift is what the wager turns on. It is also what makes the term slippery. Researchers, founders, and investors all use “world model” to mean slightly different things, and the field has not agreed on a single definition.

A Three-Part Taxonomy of World Models

Li divided world models into three categories, ordered by how close each is to a system that can actually do work in the world. Li’s three-part taxonomy sorts the competing bets by the kind of value they promise. The three categories are renderers, simulators, and planners.

The taxonomy separates visual polish from physical accuracy from agentic planning. Renderers can produce convincing imagery but cannot teach a robot. Simulators can build training grounds. Planners are the unfinished category, the one the field is racing to claim. “A robot that can plan is a robot that can work, and the entire industry is racing to be the one that gets there first,” Li wrote.

Why the Robot Application Matters

The most concrete application is the one chatbots cannot do. Chatbots cannot pick up a coffee mug, notes Martial Hebert, dean of computer science at Carnegie Mellon University.

“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,” Hebert said. “This is much more complex than just predicting the next word in a sentence.” For Hebert, the most useful application for world models is as a faster and cheaper path to physical AI, another tech industry buzzword. “Some people may have different definitions, but physical and embodied AI are kind of the evolution of what we used to call robotics,” he said.

The robot case makes the bet legible. A machine that can pick up a coffee mug and walk differently on a sore knee is a machine that has internalized a world model. Hebert’s analogy is the body itself: a general model in the nervous system that lets the body adapt quickly without conscious thought. Hebert’s CMU faculty profile lists more than four decades of robotics research behind that comparison.

Where the Capital Is Going

Capital is following the wager, even if the dollar volume is small next to the trillion-dollar LLM round. Steve Jang, co-founder and managing partner at Kindred Ventures, is investing in Overworld and other world model-focused companies. The Kindred Ventures post on Overworld is the clearest map of what the wager looks like in practice. Kindred’s portfolio spans three of the startups in the space.

• Overworld (Providence, R.I.): real-time world models for interactive video game environments
• Causal Labs: AI models for weather prediction through what it calls a “Large Physics foundation Model”
• Extropic: specialized computer chips designed to run world model workloads

“I think that the future is many different types of models with many different philosophies and architectures,” Jang said. “I don’t think that it’ll be one large, dense model to rule them all.” The bets are plural, not concentrated. The round is also early.

A Buzzword That Means Different Things

The wager has a definition problem. LeCun calls it a buzzword, and Li describes the term as one of the most important and most overloaded in AI today. Their agreement is rare in a field that has spent two years arguing about what a world model even is.

World model is quickly becoming a buzzword.

Yann LeCun, AI pioneer and founder of Advanced Machine Intelligence Labs, on the Unsupervised Learning podcast.

A video model that produces gorgeous but physically impossible flames, a language model improvising a playable game, and a physics engine that faithfully simulates combustion all go by the same name, Li wrote. The bets being placed sound similar in pitch. The work they fund is not similar at all.

The trillion-dollar LLM round is paying for one bet. The world model round, much smaller in dollars, is paying for the other. The two are not in direct competition. Investors are still counting on chatbots to keep printing revenue, and Castricato, LeCun, and Li are putting their time into world models.

Frequently Asked Questions

What is a world model in AI?

A world model is an AI system trained to learn the statistical structure of physical space and time, including how light falls, how objects move, and how they obey physical laws. In a June 2026 essay, Fei-Fei Li defined it as the spatial counterpart to a language model, which learns the statistical structure of text.

How is a world model different from a large language model?

A large language model trains on text and predicts the next token. A world model trains on representations of physical environments and predicts the next physical state, including motion, contact, and the effect of an agent’s own actions. Predicting the next state of a room is a more general skill than predicting the next word in a sentence, in the framing researchers like Li and LeCun are using.

Who is building world models today?

The most prominent efforts include World Labs, founded by Fei-Fei Li in San Francisco; Advanced Machine Intelligence Labs (AMI Labs), founded by Yann LeCun in Paris; Overworld, founded by Louis Castricato in Providence, Rhode Island; Causal Labs, which is building AI for weather prediction; and Extropic, which is building specialized computer chips for world model workloads. Kindred Ventures has invested in several of these companies.

Why are some AI researchers leaving chatbot work?

Some researchers believe the next phase of AI requires systems that understand physical space. Louis Castricato, a former Brown University doctoral candidate, told The Associated Press he had passed the point of doing real fundamental LLM research and that the field had moved into applications. Yann LeCun, formerly Meta’s chief AI scientist, left to start AMI Labs and pursue world models.

How much money is being invested in world models?

The dollar volume is much smaller than the trillion-dollar LLM round. Most of the world model wager is coming from venture firms rather than the tech giants, with Kindred Ventures visible across several portfolio companies. The bets are plural, with weather prediction and video game environments both represented in Kindred’s portfolio. Capital is also still flowing into the LLM round at the trillion-dollar scale.