World models are becoming one of the most watched directions in the field of artificial intelligence. Google DeepMind CEO Demis Hassabis recently stated explicitly that the current large language model path, represented by ChatGPT, needs to be supplemented by world models. Although OpenAI, Google, xAI, Anthropic and other companies are all betting on LLMs, Hassabis believes that to achieve true artificial general intelligence (AGI), AI must understand the physical world and causality, and possess the ability for long-term planning and hypothesis verification.
What is a world model?
A world model refers to an AI's internal simulation of the dynamics of the external world, enabling it to predict physical scenes, object motion, trajectory changes, and the consequences of actions. Unlike LLMs, which only process text or images, world models emphasize the understanding of causality. For example, how will an object fall after being pushed over? How will a driver decide when seeing an obstacle ahead? Such models are not only used for content generation; more importantly, they allow AI to perform mental simulation and planning.
Major progress in the industry
Google DeepMind released the Genie 3 system in August 2025, which can generate interactive 3D environments from text. This is seen as a major breakthrough for world models in the generation domain. In addition, video models such as Veo also implicitly embody an understanding of world dynamics. Hassabis stated that future AGI will be an integrated system that combines foundation models (such as Gemini) with world model capabilities.
Tesla is another active practitioner of world models. Over the past two years, its FSD (Full Self-Driving) system has used end-to-end neural networks to predict frames from eight driving cameras, for use in simulation and real-time driving decisions. Tesla has also applied similar principles to the Optimus humanoid robot, using vision-based end-to-end models to achieve manipulation, navigation, and task planning. Although Tesla does not publicly publish papers, its patent portfolio extensively covers neural networks, data processing, and simulation, closely related to world models.
Nvidia plays the role of infrastructure provider. It supplies simulators and development tools to most humanoid robot companies, and its world simulation technology is also in a leading position.
In addition, Fei-Fei Li and her World Labs have contributed significantly at the theoretical level, providing solid academic support for world model research.
Hassabis's view: The bottleneck and timeline of AGI
In recent interviews, Hassabis has repeatedly emphasized that scaling laws are still valid, but the returns have diminished. He believes that merely expanding model scale is not enough to achieve AGI; one or two more major innovations at the level of AlphaGo are also needed.Current AI systems exhibit “jagged intelligence”—they perform well on certain tasks but can unexpectedly fail on simple ones. Missing capabilities include: continual learning, genuinely novel creativity, consistent performance, long-term planning, and deeper reasoning.
Hassabis predicts that AGI will still take 5 to 10 years. Computing power and energy shortages are the main bottlenecks, but he also points out that AI itself can help solve these problems, for example by optimizing materials design, solar efficiency, and controlled nuclear fusion. Models such as DeepMind's Gemini Flash have already achieved roughly a 10-fold annual improvement in energy efficiency through distillation techniques.
LeCun's Stance and Departure from Meta
Yann LeCun has long criticized LLMs as a “dead end” on the road to human-level AI. He believes world models are the key. At the end of 2025, LeCun left Meta to found Advanced Machine Intelligence Labs (AMI Labs), continuing his world model research program. The company's goal is to develop systems that can understand the physical world, have persistent memory, and are capable of reasoning and planning complex sequences of actions.
Future Outlook
World models have enormous potential in robotics, autonomous driving, and scientific discovery. Studies have shown that AI agents trained in simulated worlds perform 20% to 30% better on reasoning tasks than conventional agents. Hassabis envisions a “golden age of science” for AI in the future, where breakthroughs like AlphaFold in protein folding will repeatedly emerge in fields such as materials, physics, mathematics, and weather.
Although world model technology is still in its early stages, there is a growing consensus that it may be the essential path to true artificial general intelligence.