World Models: AI and the Challenge of Structured Data

The Importance of Structured Data for the Future of AI

The artificial intelligence (AI) of tomorrow relies on the ability to manipulate structured data. Knowledge graphs play a crucial role by providing the necessary context to generate reliable and explainable outcomes. These tools enable AI systems to better understand and interpret the world around them.

Companies are investing colossal sums in the development of AI, hoping that it will one day understand the real world. "World models" embody this ambition, supported by influential figures like Yann LeCun and laboratories such as AMI Labs. Their goal is to surpass the current limitations of AI systems to simulate, anticipate, and reason more effectively. However, a crucial question often remains overlooked: what representation of the world can these systems actually rely on within companies?

The Challenges of Integrating AI in Business

In the corporate context, the challenge is no longer about deploying high-performing models, but about achieving reliable results. Over the past two years, the adoption of AI has been rapid, but the outcomes have not lived up to the investments. This issue is more related to architecture than to the technology itself. AI systems must now integrate into complex environments.

The main obstacle has evolved. It is no longer the creation of models that poses a problem, but rather the quality of the data that feeds them. An AI cannot perform better than the informational framework within which it operates.

Context: A Blind Spot for AI in Business

Whether discussing generative models or autonomous agents, the quality of results directly depends on the available information and its structuring. Information systems tend to describe the "what" and the "how," but rarely the "why." Understanding a situation also requires understanding the relationships between pieces of information. This implicit knowledge, made up of rules, trade-offs, and business logic, is often absent from data architectures. Consequently, models accumulate information but struggle to derive meaning from it.

Without explicit structuring, context becomes a problem in itself. Too much unorganized information can degrade the quality of responses, introduce contradictions, or amplify errors. In other words, adding data without organizing it makes AI less reliable.

"World models" promise to overcome these limitations by integrating a comprehensive understanding. However, they will not resolve the central issue of knowledge organization.

Structuring Data Before Simulating the World

This is why new approaches focus on context engineering. The challenge is no longer to accumulate data, but to select, organize, and connect relevant information at each step of reasoning.

Context graphs are crucial here: by explicitly representing the relationships between data, rules, and processes, they allow for a transition from mere accumulation of information to a structured understanding of situations. For example, an AI observing a train stopped at a station knows it is at a standstill through perception. But it does not realize that a delay of a few minutes can lead to missed connections, require a different crew, or disrupt an entire line. This understanding relies on knowledge of how the network operates, which knowledge graphs can provide by linking trains, stations, schedules, crews, and their dependencies.

This shift in perspective is often underestimated. It is not the AI that needs to become smarter, but the data that needs to become intelligible.

An Organizational Frontier to Cross

For leaders, the conclusion is clear: even the most sophisticated AI will not compensate for poorly managed data assets. The challenge is not to choose the right model, but to build the right framework within which this model can operate, with reliable, traceable, and explicitly linked data.

"World models" raise a fundamental question: can we industrialize an AI capable of understanding the world without organizing its representations? At this stage, all indications suggest no. The next frontier of AI will not only be algorithmic; it will be organizational and structural.

Organizations that grasp this before others will not only gain a technological advantage but also a sustainable competitive edge. They will have transformed their collective knowledge into structured assets, while their competitors will still be trying to understand why a decision was made months ago.