AGI Needs World Models: Current Status and Prospects
Recently, Google DeepMind CEO Demis Hassabis publicly stated that the development path of large language models like ChatGPT needs to incorporate world models to move toward true Artificial General Intelligence (AGI). This statement has sparked widespread discussion in the AI field.
Why World Models Are Key
Hassabis believes that existing large language models excel at processing multimodal data such as text, images, and video, but they still lack understanding of the physical world, causality, long-term planning, and hypothesis testing. World models aim to enable AI to build an internal simulation of how the world operates, allowing it to predict outcomes and propose new hypotheses through mental simulation, much like humans do.
He pointed out that scaling laws remain valid, but simply increasing computational power and data may not be enough to achieve AGI; one or two major innovations on the level of AlphaGo are also needed. World models are the most likely breakthrough in his view.
Research Progress at Various Companies
In August 2025, Google DeepMind released Genie 3, which can generate interactive 3D environments from text, marking significant progress for world models in generative simulation. Hassabis revealed that future AGI will merge large models with such world simulation capabilities to form more powerful systems.
Tesla's FSD system already uses technology similar to world models, predicting driving scenarios frame by frame through eight cameras to achieve autonomous driving planning and decision-making. This technology has also been extended to the Optimus humanoid robot, enabling it to understand and operate in the physical world.
NVIDIA, meanwhile, provides world simulation tools for many robotics companies through its simulation platforms. Yann LeCun, former chief AI scientist at Meta, founded AMI Labs after leaving Meta, focusing on world models and persistent memory to support more advanced intelligent systems.
Challenges and the Future
Hassabis expects AGI to take another 5 to 10 years. Computing chip shortages and energy constraints are the main bottlenecks, but AI itself is expected to accelerate breakthroughs in energy technology. He emphasized that world models combined with continuous learning, long-term memory, and planning capabilities are essential to truly achieve human-like broad cognitive abilities.
Currently, when AI learns by itself in simulated environments, its reasoning capabilities can improve by 20% to 30%, demonstrating the potential of world models. As research teams continue to advance, world models are moving from theory to engineering practice and are expected to become an important cornerstone of AGI.