.svg)
.jpg)
At the 2026 World Artificial Intelligence Cannes Festival (WAICF), Yann LeCun reframed the current conversation around AI.
While much of the industry continues to double down on larger models, more data, and faster generation, LeCun made a different case:
“The current path that a lot of AI companies are following is limited.”
If that’s true, then the next leap in AI won’t come from scaling what we already have. It will come from building something fundamentally different. That “something” is world models.
Today’s AI systems are undeniably powerful. They can generate code, summarize complex information, and perform at a level that would have seemed impossible just a few years ago. But when it comes to understanding the real world, they still fall short in surprisingly basic ways.
LeCun pointed to a simple but telling example: a human can learn to drive in about 20 hours, yet even with millions of hours of driving data, AI systems still struggle to achieve the same level of reliability. At the same time, models can pass professional exams but remain far from enabling something like a truly useful household robot.
This disconnect highlights a deeper issue. Modern AI is exceptionally good at recognizing patterns, but real intelligence depends on something more—context, memory, reasoning, and the ability to plan.
A central theme in the keynote was that language alone won’t get us there.
“Most of human experience and most of human intelligence has nothing to do with language.”
Large language models operate primarily in what LeCun describes as “System 1”—fast, reactive, and pattern-driven. What’s missing is “System 2”: deliberate reasoning, where a system can simulate outcomes and plan before acting.
World models are designed to close that gap.
Rather than simply predicting the next output, a world model builds an internal representation of how the world works. It allows a system to imagine what might happen next—not just statistically, but structurally.
At a high level, this means taking the current state of the world, combining it with memory and context, and using that to predict how things will evolve over time. It’s the difference between reacting to a prompt and anticipating the consequences of an action.
This is also where today’s agentic systems begin to break down. As LeCun put it:
“Building an agentic system that does not have the capability of predicting what the effect of its actions are going to be, is an extremely bad way of building an agentic system.”
Most agents today rely on learned patterns rather than true understanding. They can execute, but they don’t reason. Without the ability to simulate outcomes, they operate without foresight—performing well in controlled environments but struggling in dynamic, real-world conditions.
One of the more provocative ideas in the keynote was the need to move beyond generative models as the foundation for intelligence.
The challenge is that predicting every detail of the world—such as every pixel in a video—is fundamentally intractable. There are too many possible outcomes, and systems end up averaging them into something that lacks clarity or meaning.
This is where approaches like Joint Embedding Predictive Architecture (JEPA) come into play. Rather than reconstructing the world, these models aim to understand it—capturing the underlying structure that allows for reasoning and prediction.
This approach also mirrors how humans learn. Infants don’t learn through labeled data or predefined tasks. They observe, form expectations, and build an understanding of how the world behaves. When something violates those expectations, they notice.
World model-based systems are beginning to show similar behavior, developing early signals of common sense by identifying when outcomes don’t align with learned patterns.
What emerges from this is not just a new model, but a new direction for AI.
Instead of systems that simply generate outputs, we move toward systems that can reason, plan, and adapt. This has significant implications for how organizations think about AI today.
Much of the current focus is on copilots and automation—valuable capabilities, but still rooted in pattern recognition. The next wave will be defined by systems that can simulate outcomes, navigate complexity, and make decisions with context.
That shift requires more than new models. It requires new foundations—better data, integrated architectures, and operating models designed for AI that acts with intent.
World models don’t replace what we have today—they extend it.
They represent a move from prediction to understanding, from reaction to reasoning, and from tools to systems that can operate with intent.
LeCun’s perspective is a reminder that while the current wave of AI is powerful, it is not the end state. The real opportunity lies in building systems that don’t just generate outputs, but actually understand the world they operate
👉 Contact us to start building what’s next.
At the 2026 World Artificial Intelligence Cannes Festival (WAICF), Yann LeCun reframed the current conversation around AI.
While much of the industry continues to double down on larger models, more data, and faster generation, LeCun made a different case:
“The current path that a lot of AI companies are following is limited.”
If that’s true, then the next leap in AI won’t come from scaling what we already have. It will come from building something fundamentally different. That “something” is world models.
Today’s AI systems are undeniably powerful. They can generate code, summarize complex information, and perform at a level that would have seemed impossible just a few years ago. But when it comes to understanding the real world, they still fall short in surprisingly basic ways.
LeCun pointed to a simple but telling example: a human can learn to drive in about 20 hours, yet even with millions of hours of driving data, AI systems still struggle to achieve the same level of reliability. At the same time, models can pass professional exams but remain far from enabling something like a truly useful household robot.
This disconnect highlights a deeper issue. Modern AI is exceptionally good at recognizing patterns, but real intelligence depends on something more—context, memory, reasoning, and the ability to plan.
A central theme in the keynote was that language alone won’t get us there.
“Most of human experience and most of human intelligence has nothing to do with language.”
Large language models operate primarily in what LeCun describes as “System 1”—fast, reactive, and pattern-driven. What’s missing is “System 2”: deliberate reasoning, where a system can simulate outcomes and plan before acting.
World models are designed to close that gap.
Rather than simply predicting the next output, a world model builds an internal representation of how the world works. It allows a system to imagine what might happen next—not just statistically, but structurally.
At a high level, this means taking the current state of the world, combining it with memory and context, and using that to predict how things will evolve over time. It’s the difference between reacting to a prompt and anticipating the consequences of an action.
This is also where today’s agentic systems begin to break down. As LeCun put it:
“Building an agentic system that does not have the capability of predicting what the effect of its actions are going to be, is an extremely bad way of building an agentic system.”
Most agents today rely on learned patterns rather than true understanding. They can execute, but they don’t reason. Without the ability to simulate outcomes, they operate without foresight—performing well in controlled environments but struggling in dynamic, real-world conditions.
One of the more provocative ideas in the keynote was the need to move beyond generative models as the foundation for intelligence.
The challenge is that predicting every detail of the world—such as every pixel in a video—is fundamentally intractable. There are too many possible outcomes, and systems end up averaging them into something that lacks clarity or meaning.
This is where approaches like Joint Embedding Predictive Architecture (JEPA) come into play. Rather than reconstructing the world, these models aim to understand it—capturing the underlying structure that allows for reasoning and prediction.
This approach also mirrors how humans learn. Infants don’t learn through labeled data or predefined tasks. They observe, form expectations, and build an understanding of how the world behaves. When something violates those expectations, they notice.
World model-based systems are beginning to show similar behavior, developing early signals of common sense by identifying when outcomes don’t align with learned patterns.
What emerges from this is not just a new model, but a new direction for AI.
Instead of systems that simply generate outputs, we move toward systems that can reason, plan, and adapt. This has significant implications for how organizations think about AI today.
Much of the current focus is on copilots and automation—valuable capabilities, but still rooted in pattern recognition. The next wave will be defined by systems that can simulate outcomes, navigate complexity, and make decisions with context.
That shift requires more than new models. It requires new foundations—better data, integrated architectures, and operating models designed for AI that acts with intent.
World models don’t replace what we have today—they extend it.
They represent a move from prediction to understanding, from reaction to reasoning, and from tools to systems that can operate with intent.
LeCun’s perspective is a reminder that while the current wave of AI is powerful, it is not the end state. The real opportunity lies in building systems that don’t just generate outputs, but actually understand the world they operate
👉 Contact us to start building what’s next.
