World models are emerging as a major AI category beyond large language models, aiming to simulate the physical world rather than just process language. Ars Technica spoke with three expert practitioners: Vincent Sitzmann from MIT, Anastasis Germanidis from Runway, and Ben Mildenhall from World Labs. They explain that while LLMs began with a chat interface seeking use cases, world model developers are starting with specific applications in robotics, research, and asset generation.

Prominent figures like former Meta chief AI scientist Yann LeCun and computer vision pioneer Fei-Fei Li argue LLMs are "wordsmithed in the dark" and spatial intelligence is AI's next frontier. Hugging Face CEO Clem Delangue predicts the "LLM bubble might be bursting." Major funding reflects this shift: World Labs and LeCun's Advanced Machine Intelligence (AMI) each reportedly raised $1 billion, while Runway secured $315 million.

Recent announcements include Google DeepMind's Genie 3, World Labs' Marble for 3D environment generation, and Runway's GWM-1 specialized world models. Experts define world models as systems that build internal environment representations to simulate future events in real-time, enabling continuous spatial interaction unlike turn-based LLMs.