Yann LeCun’s Quest for Real AI: Moving Beyond the Limits of Large Language Models
For the past couple of years, the technology sector has operated under a singular, massive assumption: that if we just make Large Language Models bigger, feed them more data, and throw more computing power at them, we will eventually achieve true artificial general intelligence. But one of the field’s most respected pioneers is openly challenging this consensus, embarking on a new venture to prove that the current crop of AI is fundamentally hitting a wall.
According to a new report from BBC News, Turing Award winner Yann LeCun is actively working on a new startup project aimed at developing a far more flexible, adaptable AI system. LeCun, who also serves as the Chief AI Scientist at Meta, has long been a vocal skeptic of the industry’s over-reliance on generative text models. His latest move translates those academic critiques into practical development, aiming to build systems that can actually understand the physical world.
The core of LeCun’s argument lies in how current generative models learn—or rather, how they fail to learn. He argues that modern LLMs are fundamentally incapable of reaching human-level, or even animal-level, intelligence. The issue is that they are trained almost exclusively on text, which is a highly compressed, human-designed abstraction of reality. They lack the ability to deal with messy, unstructured, real-world sensory data. A house cat, LeCun often points out, understands gravity, object permanence, and basic physics within the first few weeks of its life without ever reading a word. Current AI systems possess none of this common-sense understanding.
By building a new startup focused on “flexible” AI, LeCun is attempting to construct “world models”—AI systems that learn by observing video, physics, and sensory inputs, allowing them to predict consequences and reason through actions before taking them. If successful, this approach could bypass the massive energy demands and hallucination errors that plague current generative models, paving the way for autonomous systems that can safely navigate the physical world, from advanced robotics to truly smart digital assistants.
Ultimately, LeCun’s endeavor serves as a crucial reminder that the future of AI is far from settled. While the tech industry’s giants remain locked in an expensive arms race to scale up existing architectures, the real breakthrough might not come from doing the same thing bigger, but from starting over with a completely different philosophy. If we want AI to truly assist us in the physical world, we may first have to teach it how to look at the world, rather than just read about it.