Yann LeCun, the AI researcher and Meta's chief AI scientist, used his stage time at VivaTech in Paris to argue that the next big leap in artificial intelligence won't come from bigger language models. Instead, he pushed for something he calls 'world models' — systems that learn by interacting with the real world, not just by digesting text.
His message was direct: the industry's current obsession with scaling up language models is a dead end. Real progress, LeCun said, demands that AI build internal representations of how the physical world works. That shift, he argued, could redirect the billions flowing into AI research.
What LeCun proposed
LeCun described world models as AI systems that can predict the consequences of actions, much like humans do when they reach for a cup or step off a curb. Rather than training on ever-larger datasets of text and images, these models would learn from direct experience — through sensors, robots, or simulated environments.
He didn't offer a ready-made product or a timeline. But he made clear that the current approach — dump more data and compute into a transformer — has diminishing returns. 'We need a different paradigm,' he said, according to reports from the event. The goal is an AI that understands causality, not just correlation.
Why scaling language models isn't enough
For years, the dominant bet in AI has been that bigger models trained on more data will keep getting smarter. That bet produced GPT-4, Gemini, and Llama. But LeCun has been publicly skeptical. At VivaTech, he laid out the limits: language models don't grasp physics, they don't plan, and they can't learn from trial and error the way animals do.
He pointed to a simple test: show a large language model a video of a ball rolling off a table. It might describe the scene, but it can't predict where the ball will land. A world model, by contrast, would simulate the trajectory. That gap, he argued, is the reason autonomous driving, robotics, and even reliable chatbots still stumble.
What this means for investment strategies
LeCun's pitch isn't just academic. If the industry takes world models seriously, it could shift where venture capital and corporate R&D spend their money. Right now, most of the AI investment goes into training larger models and buying GPUs. A pivot to world models would require funding new kinds of hardware, simulation platforms, and robotics companies.
He didn't name specific companies or amounts. But the implication is clear: the next trillion-dollar AI company might not be the one that builds the biggest language model. It might be the one that builds a system that can actually navigate the world.
Whether investors will follow LeCun's lead remains an open question. The industry has a long habit of doubling down on what works. But with a figure as respected as LeCun publicly calling for a new direction, the conversation is at least shifting. The next major AI conference will show whether anyone is listening.
