AI researchers pivot from chatbots to world models as physical AI emerges as next frontier

AI Research Shift Signals New Direction Beyond Language Models

Computer scientist Louis Castricato spent eight years studying large language models before concluding that fundamental research had reached a plateau. After leaving his doctoral studies at Brown University, he founded Overworld, a startup building AI systems that understand and navigate worlds, not just words 2

2

. This pivot reflects a broader AI research shift gaining momentum across the industry, as entrepreneurs and scientists confront the limitations of large language models that power ChatGPT, Claude, and similar chatbots.

The shift from chatbots to physical AI represents more than incremental progress. AI world models aim to teach systems the statistical structure of space and time—how light falls on surfaces, how objects respond to force, and how environments change from different angles 3

3

. "Where language models learn the statistical structure of text, world models learn the statistical structure of space and time," wrote Fei-Fei Li, founder of San Francisco startup World Labs, describing world models as "one of the most important and most overloaded terms in AI today" 2

2

.

Fundamental Limitations Drive Pivot to Physical AI

Despite impressive capabilities—passing bar exams, matching top students in mathematical competitions, and generating poetry—large language models face practical and conceptual constraints. The AI systems that understand physical environments address a core weakness: LLMs do not experience the world they describe and cannot test hypotheses or probe environments 1

1

. They learn patterns of cause and effect indirectly, which proves insufficient when success depends on understanding action consequences in real-world scenarios.

"Chatbots can't 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. This is much more complex than just predicting the next word in a sentence" 4

4

. Scaling laws described in a heavily cited 2020 paper suggested performance improves with model size, training data, and computing power, but companies are exhausting high-quality public text data and facing power constraints as data centers demand gigawatts of electricity 1

1

.

AI Agents in Simulated Worlds Enable New Capabilities

The next frontier in AI involves systems where AI agents in simulated worlds learn by acting within generated environments. At General Intuition, co-founder Adam Jelley demonstrated an AI program that creates video game worlds on the fly, with autonomous agents navigating rooms and confronting characters in real time. "An AI playing in the mind of another AI," Jelley described the system, betting it will eventually outsmart large language models 1

1

.

Yann LeCun, who left his role as Meta's chief AI scientist to start Paris-based Advanced Machine Intelligence Labs, views world models as enabling AI agents to "predict the consequences of its own actions" 2

2

. "I joke that the smartest systems we have today are not as smart as a house cat," LeCun says, noting that while cats can't code like LLMs, they survive by understanding their environment 1

1

.

Spatial and Temporal Awareness Opens Path to Robotics

For scientists with decades in robotics research, physical AI and embodied AI represent the evolution of their field. Hebert explains that humans possess general models in their nervous systems allowing bodies to adapt quickly—balancing, walking, adjusting when injured. "You don't need to think about that. You have a general model somewhere in your nervous system and brain that allows your body to adapt very quickly," he said 3

3

. This spatial and temporal awareness enables causal reasoning about physical interactions that text-trained models cannot replicate.

Overworld optimizes for interaction, building video game environments where spooky forests adapt as virtual characters move through and interact with objects. "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 3

3

. Applications extend beyond gaming: Causal Labs builds AI world models for weather prediction, while Extropic develops specialized computer chips suited to world model computation 2

2

.

Venture Capitalists Back Emerging World Model Companies

Despite less obvious near-term applications compared to AI coding tools, simulated worlds and physical AI are attracting venture capitalists. Steve Jang, co-founder and managing partner at Kindred Ventures, is investing in Overworld, Causal Labs, and Extropic, signaling confidence in world models as a viable path forward 2

2

. While trillions of dollars continue flowing to chatbot developers like Anthropic and OpenAI, this emerging investment pattern suggests the industry recognizes that achieving artificial general intelligence requires AI systems capable of reasoning about space, causality, and action consequences—especially for controlling humanoid robots, operating factories, and exploring other planets 1

1

.