AI is entering a new phase, moving beyond language-based systems like large language models (LLMs) to what industry experts call world models. Unlike LLMs, which focus on text, these new models interpret and simulate the physical world — using video and spatial data to understand objects, scenes, and dynamics.
These world models represent a paradigm shift: they can form internal representations of real environments, not just process information about them. That capability opens up applications in robotics, video games, and any domain that needs an AI to “think” in three dimensions, rather than just stringing words together.
Crucially, these models also tie into the concept of digital twins — virtual replicas of real physical systems or spaces. By accurately simulating a system’s behavior, world models can support predictive maintenance, scenario testing, or real-time monitoring.
However, building these systems isn’t easy. One of the biggest hurdles is data bottlenecks: collecting rich spatial and video data for real-world environments is significantly harder than scraping text from the web.
