AI Researchers Pivot from LLMs to Physical World Models

AI researchers and entrepreneurs are pivoting away from large language models and chatbots toward the development of world models. This shift aims to create physical and embodied AI capable of understanding the statistical structure of space, time, and physics, allowing systems to navigate physical environments and predict the consequences of their actions rather than simply predicting the next word in a sentence.

Fei-Fei Li, founder of World Labs, is leading this movement by developing a taxonomy that categorizes world models into renderers, simulators, and planners. Other central figures include Yann LeCun, who left Meta to establish Advanced Machine Intelligence Labs, and Martial Hebert of Carnegie Mellon University, who researches the application of these models to robotics.

The technology is being applied to a variety of sectors, from robotic planners to interactive video games. Louis Castricato's startup, Overworld, is specifically developing virtual environments where AI can interact with detailed surroundings. This transition has triggered increased venture capital interest, with Kindred Ventures investing in world-model companies such as Overworld, Causal Labs for weather prediction, and Extropic for specialized hardware chips.