World models: When AI learns to simulate the world before it acts

By Thomas Gayet, Head of Innovation at SQLI

For several years now, artificial intelligence has been generating increasingly realistic text, images, code and video. A new stage is taking shape: models capable of simulating an environment, anticipating the consequences of an action and helping autonomous systems learn before acting in the real world.

It's the dawn of a new era: world models

The announcements by Google DeepMind around Genie 4 and the work of NVIDIA on Cosmos illustrate this evolution. Moving beyond content generation, the challenge is now to build environments in which AI can explore different scenarios, test decisions and learn from their consequences.

From generation to simulation

A classic generative model produces a result from an instruction. A world model aims to go further: representing an environment and predicting how it might evolve based on its state and the actions performed.

An agent could thus test several decisions in a simulated environment before choosing one to execute. A robot could train to manipulate objects in various situations. An industrial team could explore different scenarios without having to trigger each situation in the real world.

This evolution was notably on the program of the World Summit AI 2026, with the conference presented by Mateusz Malinowski, co-founder of Reka AI, who presented one of the major challenges in the field: bringing together language understanding, visual perception and physical intelligence. The program also featured a talk by Simon See from NVIDIA, dedicated to the transition from research to embodied AI.

Several approaches coexist. Google DeepMind explores the generation of interactive worlds with Genie 4. Decart develops real-time video generation and transformation technologies. World Labs works on the creation of 3D spatial environments with Marble.

An essential distinction remains: generating a visually convincing environment does not necessarily mean faithfully reproducing the laws of physics. Temporal consistency, predicting the consequences of an action, and transferring skills acquired in simulation to the real world remain major challenges.

From simulation to physical AI

The NVIDIA GTC Berlin 2026 also highlighted this convergence between world models, simulation and Physical AI. The interest no longer lies solely in image generation, but in the ability to produce training data and anticipate the behaviors of physical environments.

This direction is in line with NVIDIA's work around Cosmos. Presented by the company as a family of models for physical AI, Cosmos 3 notably combines visual reasoning, world generation and action prediction. NVIDIA describes uses for training and evaluating robots, autonomous vehicles and AI systems capable of interacting with their environment.

This approach addresses a concrete difficulty: certain situations are rare, costly or dangerous to reproduce in the real world. Simulation allows multiplying the variants, provided that the environments produced are reliable enough for the learnings to remain useful outside simulation.

From content to experience

Retail illustrates a complementary evolution. I recently tested Anywear, developed by Decart, a virtual try-on tool that allows you to visualize clothes on your own video stream, in real time. After trying many virtual try-on solutions over the past three years, I was particularly impressed by the result: you can observe the rendering in motion without going through the prior creation of a photo.

Anywear relies on real-time video transformation and not on a world model in the strict sense. It nevertheless illustrates a broader evolution: AI no longer only transforms content, it begins to transform the experience itself.

Other applications are conceivable: exploring personalized products or spaces in luxury, testing energy management strategies in the face of incidents, generating synthetic scenarios to evaluate fraud detection systems or enriching training for rare aeronautical situations. These examples do not all have the same level of maturity and do not necessarily rely on world models. They however share the same logic: exploring situations and testing decisions before their implementation.

A new building block for agentic AI?

AI agents already know how to decompose a goal, use tools and execute actions. But in a complex environment, acting effectively also requires anticipating the consequences of one's decisions.

A world model could allow an agent to compare several options in a simulated environment, then choose an action. The agent would then observe the real result and adapt its strategy.

This perspective aligns with the work of AMI Labs, founded around Yann LeCun, on predictive models capable of anticipating the evolution of the world. The objective differs from that of visual world generators: it is primarily about building a representation useful for prediction and planning.

The convergence between agents and world models could thus evolve AI from a system that executes instructions to a system capable of exploring several possible futures before acting. This however remains a research goal: current models can make mistakes and their simulations do not constitute a guarantee of behavior in the real world.

From demonstration to industrialisation

The risk would be to confuse a convincing demonstration with a technology immediately deployable at scale. Physical fidelity, temporal consistency, computing cost and the ability to transfer simulation results to the real world will be decisive.

In industrial or regulated sectors, generative models should complement specialised simulators and validation processes, and not replace them without proof of reliability.

World models open a new step in the evolution of artificial intelligence. After models capable of generating content and agents capable of executing actions, they explore the possibility of simulating an environment and anticipating the consequences of a decision.

The next step in AI may not only be about better answering our requests or executing our instructions, but also exploring what could happen before acting.

Sources