Louis Castricato, a computer scientist who studied large language models (LLMs) like those behind ChatGPT, felt that research in this area has plateaued and shifted his efforts to building AI that understands and navigates physical environments. His startup, Overworld, aims to create interactive “world models” where AI can predict and respond to physical realities rather than just text. Leading AI researchers such as Fei-Fei Li and Yann LeCun emphasize that true intelligence involves understanding space and time—not just language—enabling AI to act in the real world. Experts argue that while chatbots revolutionize text-based tasks, they cannot perform physical actions like picking up objects. Instead, “physical AI” or embodied AI models teach machines about the complex interactions of the physical world. Overworld and similar companies are gaining venture capital interest by innovating in simulated virtual environments for training robots and interactive gaming. These “world models” range from visually realistic renderers to sophisticated simulators and planners, which help AI predict and plan actions in unstructured environments, an essential step for practical robotics and AI applications.
Back
