By Anne Canal

What does it mean to build a machine that learns and reasons as elegantly as a living mind? In his lecture How Could Machines Learn as Efficiently as Animals and Humans?, AI pioneer Yann LeCun—Chief AI Scientist at Meta and professor at NYU—presents not just a roadmap for artificial intelligence, but a philosophy of intelligence itself.
LeCun, one of the three “godfathers” of deep learning, begins by confronting the limits of today’s systems. Chatbots and image models, he notes, may astonish us with fluency, yet they lack what even a small child—or a cat—possesses in abundance: a world model, an intuitive understanding of physics, agency, and consequence. To transcend these limitations, LeCun proposes a new architecture: Joint Embedding Predictive Architectures (JEPA), a framework in which machines learn by predicting the structure of the world, not merely by mimicking human text or labeling data.At the core of LeCun’s argument is self-supervised learning, a process by which an intelligent system teaches itself from the data it already observes, discovering order without explicit instruction. This, he suggests, is how animals—and perhaps human infants—grasp reality: by filling in the missing pieces of their sensory streams, predicting what will happen next.