The story of a baby dropping a spoon in a highchair isn’t just a parenting rite of passage—it’s a profound lesson in how intelligence begins. Each fall is a hypothesis tested, a law of physics confirmed through direct experience. The child isn’t reading about gravity; they’re living it. This simple act of observation and trial mirrors the core of what many believe is the next evolution of artificial intelligence. Today’s leading AI systems, like those powering ChatGPT, are linguistic marvels. They can write poetry, draft emails, and explain complex ideas with astonishing fluency. But they do so without ever having felt the weight of an object, seen a fall, or experienced cause and effect in the real world. They are fluent in language but blind to reality. Yann LeCun, Meta’s Chief AI Scientist and a pioneer of deep learning, argues this gap is not just a minor flaw—it’s a fundamental limitation. He believes the future of AI lies not in text, but in video. Not in predicting the next word, but in predicting what happens next in a visual world. The current dominant architecture—autoregressive models—builds language one word at a time, relying on statistical patterns from massive text datasets. But this approach is inherently fragile. Each prediction carries a small error margin. Over long sequences, these errors compound exponentially, leading to hallucinations—confident, coherent fabrications. No amount of data or compute can fix this if the underlying model is built on a flawed premise: guessing the next token instead of understanding the world. LeCun points out a stark contrast in data intake. Human language, all of it, amounts to about 10^14 bytes. A four-year-old, however, absorbs between 10^14 and 10^15 bytes of visual information every year—watching objects move, fall, collide, and change. This sensory flood is how children build a deep, intuitive model of reality long before they speak. LeCun’s vision is to build AI that learns the same way: first through observation, then through language. He calls for “world models”—systems trained on video to understand physics, motion, and object permanence. These models wouldn’t just describe the world; they’d anticipate it. They’d know that if a cup is on the edge of a table, it will fall—just like a baby who’s dropped a spoon a thousand times. This shift is already underway. Meta’s V-JEPA 2 is trained on video to predict future frames, building an internal model of physical laws. Apple’s SlowFast-LLaVA-1.5 learns to separate objects from motion in long videos, enabling reasoning about complex sequences. These systems don’t just parse images—they learn the narrative of action. The implications are vast. A video-trained AI could reason about real-world problems in ways text-based models cannot. It might predict the spread of disease based on human movement patterns, simulate climate effects with physical fidelity, or design safer robots that understand how objects behave. LeCun’s message is clear: we’ve spent years building AI that mimics language. Now, we must build AI that understands the world. The future isn’t in bigger text models—it’s in systems that watch, learn, and predict like a child who’s dropped a spoon and learned something real. For researchers, the call is to move beyond autoregressive LLMs and invest in multi-sensory, video-driven models. For companies, the path forward is open-source collaboration and infrastructure for visual learning. If history is any guide, LeCun is rarely wrong. And if he’s right, we’re not just building smarter machines—we’re building ones that finally get it.