Can intelligence exist without a body? This question, which seemed to have a negative answer a few decades ago, is being reconsidered in light of advances in artificial intelligence. If you are a user of ChatGPT, you probably think the answer is affirmative: a machine can be intelligent if we understand intelligence as the ability to comprehend and solve problems. However, the issue is more complex. The intelligence of current AIs is based on language. The models dominating the landscape, like ChatGPT, are essentially language models: their "intelligence" lies in predicting the next word in a text, thanks to massive data processing. Language has been key in human evolution, which is why these AIs are so effective in tasks requiring a certain type of intelligence. But their scope is limited by their nature: they do not interact with the physical world. Machines currently lack a body and, therefore, direct experience of the three-dimensional environment in which we live. This deficiency limits their learning and explains, for example, the difficulties of autonomous cars. Thus, general artificial intelligence (AGI), that AI capable of surpassing humans, remains a distant promise, despite predictions from some technologists. In response to this challenge, the scientific community has begun to explore a new avenue: "world models." Unlike language models, these aim to provide machines with a deeper understanding of the physical context, bringing their intelligence closer to human intelligence. Although AIs will still lack a body, they will no longer rely solely on language to learn. Richard Sutton, a pioneer of reinforcement learning and Turing Award winner 2024, recently warned that natural language models have hit a ceiling. According to Sutton, these systems only imitate, lack their own goals, and do not learn from real experience. For him, true intelligence involves interacting with the world and achieving goals, not just predicting words. This view is shared by key figures in AI. Yann LeCun, one of the fathers of deep learning, left Meta after the Llama 4 fiasco and the internal reorganization that sidelined him. Now, LeCun is driving a new company dedicated to developing AI capable of understanding the physical world, with persistent memory and reasoning and planning skills. According to LeCun, current models are useful but limited: they can describe how to rotate a cube, but do not truly understand the concept of spatial rotation. Fei-Fei Li, creator of ImageNet and a leading figure in the field, also supports this new direction. From her start-up World Labs, she has launched Marble, a tool capable of generating 3D worlds from a single image. Li argues that real intelligence requires the ability to act and understand space, as babies do when learning physics before they can speak. Danijar Hafner, creator of the Dreamer series, has demonstrated the potential of world models in robotics. Dreamer 4 managed to obtain diamonds in Minecraft just by watching videos, without playing beforehand. It built an internal model of the environment and practiced in its "imagination" before facing the real challenge. This approach could revolutionize robotics, allowing robots to learn from observation and simulation, rather than costly trials in the physical world. The rise of world models poses challenges for the industry. As Gillian Tett notes in the Financial Times, if these models require different architectures and hardware, the current infrastructure could become obsolete in a few years. The future of AI may depend less on the scale of language models and more on their ability to understand and act in the real world.