Insight5 min read
Artificial intelligence keeps surprising us and reinventing itself. After the wave of innovation brought by Transformer models, a new generation of AI is emerging.
A new wave of innovation in artificial intelligence
Artificial intelligence keeps surprising us and reinventing itself. After the wave of innovation brought by Transformer models, able to understand and generate language with unprecedented accuracy, a new generation of AI is emerging. It promises to redefine the way we interact with digital technology and to open up new possibilities in the physical world.
Gemini 3 and SAM 3D: AI enters the real world. Last week, Google unveiled Gemini 3, a model able to generate high-quality images and use multimodal information for complex tasks. Meta, for its part, launched SAM 3D, a technology that can reconstruct any object or person in three dimensions from a single image.
These advances show that AI is no longer limited to text: it is learning to perceive and interpret the real world, taking a decisive step towards spatial intelligence.
AI pioneers are leading the way
Meanwhile, Yann LeCun and Fei-Fei Li are already sketching out the next decade. They are calling for the development of "world models": systems able to understand space, perceive their environment and reason about what they see — and what they don't.
- Yann LeCun, a French researcher and co-inventor of convolutional neural networks, is one of the pioneers of deep learning. He was head of AI research at Meta and is a professor at New York University. He is now leaving Meta to found his own company dedicated to world models, aiming to build systems that can perceive and reason about space.
- Fei-Fei Li, a Stanford professor and world-renowned computer vision expert, created ImageNet, the database that revolutionised image recognition. She now leads World Labs, her own company, which aims to build AI that understands its environment and acts intelligently and safely.
So what exactly are world models?
World models are a major step forward for artificial intelligence. While today's models, such as Transformers, excel at analysing text or generating content, they remain limited when it comes to understanding space, movement or depth. They cannot perceive the world the way we naturally do.
Our brain works mainly with images: more than 90% of the information we process is visual and three-dimensional. This perception allows us to navigate complex situations with precision. For example:
- A pedestrian crosses a street by instinctively judging the speed of cars and the likely path of other passers-by.
- A footballer anticipates the movements of teammates and opponents while controlling the ball.
- A surgeon looks at an organ and plans their actions in three-dimensional space, taking into account depth, texture and the position of instruments.
- An autonomous drone analyses a complex environment, avoids obstacles and chooses the best route to its destination.
These abilities rely on spatial heuristics: unwritten rules our brain applies automatically to anticipate and act in the real world.
World models aim to reproduce this kind of spatial intelligence in machines. They don't just absorb data; they learn to interpret space, understand the relationships between objects and anticipate events. For example:
- predicting the movement of a ball in a video game or in reality;
- understanding how vehicles interact at a complex junction to avoid accidents;
- determining the location and condition of out-of-sight items in a warehouse to optimise logistics;
- inferring the likely path of a pedestrian in an urban environment, for safety systems or self-driving cars.
In short, these models no longer just "react" to information: they are starting to understand the world dynamically and in context, as a human would. This ability to anticipate, plan and perceive spatial relationships paves the way for far more complex and powerful applications than those we know today.
To do this, they rely on latent representations: internal structures that let them infer information that is not directly visible from partial data. In practice, this allows them to fill the "gaps" in their understanding of a space or situation.
A few examples:
- Autonomous navigation: a vehicle can predict the path of a pedestrian hidden behind an obstacle.
- Robotics and logistics: a warehouse robot can anticipate the movements of other machines or workers.
- Games and simulation: in a video game or augmented reality simulation, the model can anticipate characters' actions and react in real time.
- Medicine and assisted surgery: an AI system can track a surgical instrument, predict how tissues or organs will evolve and guide the surgeon precisely.
- Drones and exploration: an autonomous drone can map an unknown environment, detect obstacles and plan the best route.
This ability to perceive and plan at the same time turns AI systems into genuine agents that understand context and act autonomously, while respecting the laws of physics and the logic of space.
Major impacts for business and society
The rise of world models is not just a technical breakthrough: it has profound consequences for companies, organisations and society. These models allow AI to understand, anticipate and act in complex environments, opening up new opportunities.
- Industry and logistics: optimised flows and production lines, anticipation of needs.
- Transport and mobility: self-driving vehicles that predict the movements of other road users and reduce accidents.
- Health and medicine: patient monitoring, early warning of complications and surgical assistance.
- Safety and security: anticipating risky behaviour or situations, preventing accidents, coordinating emergency responses.
- Creativity and innovation: generating realistic scenarios, simulating experiences and designing innovative solutions.
These capabilities will change the way organisations operate and make decisions. Companies that adopt these technologies will gain a strategic edge: optimised processes, lower costs, greater safety and new opportunities.
World models mark a new era of artificial intelligence: systems closer to reality, adaptive and able to reason in dynamic environments. They herald a decade in which AI will no longer simply respond to requests, but anticipate, plan and act autonomously and intelligently.
This revolution goes beyond technological innovation: it will change the way we work, innovate and interact with both the digital and the physical world. Understanding this new generation of AI is therefore essential to stay at the forefront of digital transformation.
In short, AI is entering a phase of maturity: closer to reality, more intuitive and able to reason in complex environments. The revolution does not stop with Transformer models: it heralds a decade of reinvention, in which AI will learn not only to respond, but also to understand and anticipate the world around it.
