The first wave of generative AI changed the way we work almost overnight. Large Language Models brought remarkable fluency, speed, and reach. They helped us write, code, summarize, analyze, challenge and automate parts of our digital routine. For many organizations, this was the moment they realized AI was no longer just a fancy research topic but a potential productivity engine.
But with some distance, it becomes clear that this first surge is only the ver very beginning. LLMs are increasingly powerful (and power hungry), yet they operate in a relatively narrow slice of reality. They read, predict, and generate language. They don’t build a lasting internal sense of the physical world. They don’t form an adequate memory of how systems behave. They don’t model precise causes, consequences, or temporal changes with the consistency required for real-world execution. They talk. Some say they reason. They don’t act.
The next chapter will look very different. We are entering a stage where agents, large action models, and large multimodal models expand the perimeter way beyond text. Agents introduce more precise planning and independent, autonomous execution. LAMs bring an understanding (and projection) of how actions change an environment. LMMs fuse text, image, sound, video, sensor input and spatial information into a unified, cohesive perception. This combination opens the door to operational AI that interacts with the world instead of describing it. For a CIO, this shift is profound. Our systems, processes, and risks no longer sit only in documents and databases and data lakes. They sit in physical assets, buildings, supply chains, vehicles, human routines, weather patterns and live contexts where language alone is not enough. Agentic AI will analyze, decide, and act across these layers. It will coordinate (and execute) workflows end-to-end, monitor and adapt processes in real time, and handle complexity that today demands entire teams. Forget yet another and bigger chatbot this is a different paradigm altogether.
It’s in this transition that a name reappears on the radar: Yann LeCun. For decades, he has been one of the foundational and authoritative voices in AI. Turing Award co-laureate. Pioneer of convolutional neural networks (CNNs), a breakthrough in artificial intelligence. LeCun developed the foundational CNN architecture, notably the LeNet network in the late 1980s and 1990s, originally applied to handwritten digit recognition for tasks like reading checks. CNNs are deep learning models specifically designed to process and recognize patterns in data with a grid-like topology, such as images. They utilize convolutional layers to automatically and efficiently extract spatial hierarchies of features from input data, making them highly effective for tasks such as image and video recognition, as well as speech analysis. LeCun (65) is the architect of the ideas that eventually shaped computer vision and many modern deep learning techniques. He was the Chief AI Scientist at Meta. One of the few people who can connect the dots from research to industrial-scale systems.
Yann LeCun is, in his own way, one of France’s strongest ambassadors in the global AI landscape. Born in Paris and shaped by our country’s academic ecosystem before becoming a central figure in deep learning, he embodies the French tradition of scientific excellence with international impact. Alongside Mistral -today’s flagship of French AI entrepreneurship- LeCun represents the other essential pillar: fundamental research, long-term vision, and the ability to push the field into new territories that others eventually follow. For a country determined to stay relevant in the global AI race, this blend of science, ambition, and intellectual independence is a strategic asset.
Now he has left Meta to found a company focused on what he calls large world models. These models do way more than predict text; they form deep internal representations of the physical world. They learn how objects move. How environments change. How actions propagate. In LeCun’s view, this is the missing ingredient to build AI that approaches human-level intelligence. Not because it becomes sentient or intelligent (there are way too many definitions of these human centered concepts to even open a decent debate 😊), but because it understands the world it operates in.
If he succeeds, the implications are considerable. World models combined with agents could give machines a form of situational awareness that transforms robotics, automation, planning, simulation and digital operations. They could help us manage complexity at a scale no LLM can approach. They could power new forms of enterprise systems: adaptive, predictive, context-aware. For organizations, it means a shift from task automation to operational intelligence. They also promise a more energy-efficient path forward, because a model that understands its environment needs far less brute-force computation than today’s massive LLMs. Instead of scaling endlessly, world models focus on learning structure, dynamics and causality, making AI both more capable and fundamentally more aligned with the constraints of the real world.
The first surge of GenAI taught us the value of language. The next will teach us the value of understanding, reasoning, and acting. CIOs will need to prepare for an architecture where AI is no longer a peripheral assistant but an operational participant. Governance, integration, security, model oversight, and change management will need to evolve accordingly.
We are still early. But if the last two years have shown anything, it is that early becomes mainstream very quickly. The second wave will not arrive with fanfare; it will arrive with capability. The organizations that start preparing for it now -technically, operationally, culturally- will be the ones that shape the curve rather than chase it.
This is how the real transformation begins: with an adaptive mindset…
