"After LLMs"... Fierce Competition in World Models That Understand the Real World

The competition in generative artificial intelligence (AI) is expanding from large language models (LLMs) to world models that can understand physical environments and predict the outcomes of actions.

On the 3rd (local time), at the Fira Gran Via exhibition center in Barcelona, Spain, Honor's robot demonstrated walking at the 'Mobile World Congress (MWC) 2026'. 2026.3.3 Photo by Jin-hyung Kang

On June 2, major global big tech and AI companies judged that world models are a stepping stone toward artificial general intelligence (AGI) and are thus accelerating related research and development (R&D). A world model is a technology that understands the physical laws of the real world and predicts dynamic changes in situations. Unlike conversation-oriented LLMs that understand language and images, world models autonomously interpret new environments and unstructured data to make independent decisions. As a result, this is considered a core technology that enables the realization of physical AI, which moves autonomously without relying on predefined rules or limited data inputs.

Recently, Google DeepMind connected real images from Google Street View to its general-purpose world model, Genie 3, which can create 3D virtual environments through text. This enables the simulation of environments where AI agents and robots can interact based on real-world locations.

NVIDIA is also generating virtual environments for robot and autonomous driving AI training using its Cosmos model. With components like Cosmos Reason, which understands physical environments, and Cosmos Policy, which controls behaviors, it is possible to create synthetic data to train physical AI.

In South Korea, NC AI is leading the development of world models as a project operator in the national physical AI research and development initiative by the Agency for Defense Development. Their strategy is to combine expertise in building large-scale, high-precision 3D virtual worlds with proprietary 3D generative AI technology. Naver has also developed the "Seoul World Model," a city-scale generative model based on Seoul, ensuring accuracy in representing real-world space and time.

Koray Kavukcuoglu, Chief Technology Officer (CTO) of Google DeepMind, explained, "World models are a key part of advancing toward AGI. It's not just about understanding video; we need models that comprehend the rules of movement and physical laws to simulate the real world. In high-level reasoning, both the actual physical world and textual information must be considered when making decisions."

The Software Policy & Research Institute's report, "World Models: The Evolution of AI That Understands Reality," recommends that South Korea leverage its manufacturing strengths to turn data into assets and strategically foster industry-specific world models. The report states, "By systematically collecting physical behavior data generated at high-density production sites to build 'manufacturing datasets for world models,' and developing world models specialized in Korea's areas of strength, the country should secure a global competitive edge."

This content was produced with the assistance of AI translation services.