Tech Entrepreneurs Shift From Chatbots to Physical AI and World Models
Computer scientist Louis Castricato, after eight years of large‑language‑model research at Brown University, announced his departure to launch Overworld, a startup built around AI "world models." He explained that the field has "passed the point of doing real fundamental LLM research" and is now focused on teaching machines how to understand space, time, light, and the laws of physics. Overworld aims to create AI that can navigate and manipulate physical environments, a goal echoed by Fei‑Fei Li of World Labs, who described world models as learning "the statistical structure of space and time" – a step beyond predicting the next word.
Yann LeCun, former chief AI scientist at Meta, has also championed the concept. In a recent podcast he called the world model a "buzzword" that enables AI agents to "predict the consequences of its own actions." These ideas are crystallizing into three practical categories: renderers (visual fidelity), simulators (physically accurate training grounds), and planners (decision‑making in unstructured worlds). The ultimate ambition is to build robots that can plan and act safely in real settings, a prospect that is drawing substantial venture funding.
Investors such as Steve Jang of Kindred Ventures have backed Overworld and related ventures like Causal Labs (weather prediction) and Extropic (chips optimized for world models). Jang believes the future will host many model types rather than a single dominant architecture. This sentiment aligns with recent corporate moves. BlackBerry, traditionally a security and communications firm, announced a strategic pivot toward "physical AI" through its General Embedded Management (GEM) platform, which underpins AI‑enabled industrial and robotic systems. CEO John Giamatteo highlighted the need for "deterministic and safety‑certified" AI, especially as autonomous vehicles become "robots on wheels."
The shift is already influencing product development. Researchers like Martin Hebert at Carnegie Mellon note that robots must handle geometry, force, and balance—challenges that go beyond language prediction. Overworld’s prototype video‑game worlds allow characters to interact with dynamic environments, a testbed for the same physics‑aware AI that could someday run factory robots or autonomous drones.
Industry analysts see this evolution as a natural response to the limits of pure language models, which excel at text but falter when tasked with real‑world interaction. By integrating perception, simulation, and planning, the next generation of AI promises safer, more adaptable systems that can operate alongside humans across transportation, manufacturing, and entertainment.
