At the opening keynote of the MIT Generative AI Impact Consortium workshop, Yann LeCun, Meta’s Chief AI Scientist and a professor at New York University, once again made bold claims that have become his hallmark. The 2018 Turing Award winner reiterated his long-standing view that large language models (LLMs) are a dead end, arguing they will never achieve human-level intelligence. He also delivered a scathing critique of the humanoid robotics industry, calling it a "big secret" that no company knows how to make robots smart enough to be truly useful. LeCun’s skepticism is rooted in decades of research. In 1987, while completing his PhD in Paris, he worked on connectionist learning models—laying the foundation for the backpropagation algorithm in neural networks. At the time, the AI field was dominated by expert systems, which encoded human knowledge into rules. LeCun saw a different path: intelligence should emerge through self-organization, not pre-programmed rules. "Maybe I was naive, or maybe humanity just isn’t smart enough to design intelligence directly," he said. "Intelligent systems must build themselves." He recalled struggling to find a PhD advisor, until one "very kind gentleman"—Maurice Milgram—agreed to sign the paperwork, even if he couldn’t help technically. That small act set LeCun on a path that would later redefine AI. In 2016, LeCun introduced the "cake theory" of AI: if AI is a cake, self-supervised learning is the main body, supervised learning the frosting, and reinforcement learning the cherry. "At the time, DeepMind and others were obsessed with reinforcement learning, but I never believed in it," he said. "It’s too inefficient. You should use it only as a last resort." LeCun’s vision was to train systems to learn the underlying structure of data—without task-specific labels—so they could form a model of the world. The idea worked for text: predict the next word in a sequence, and the model learns language. But it failed for video. "The future of a video is too vast to predict," he explained. "You can’t predict every person’s face, room size, or floor texture. Forcing a model to do so kills its ability to learn." In 2022, LeCun and his team at Meta’s FAIR lab began developing a new approach: JEPA (Joint Embedding Predictive Architecture). Instead of reconstructing pixels, JEPA learns a compressed, meaningful representation of data and predicts that representation. The key challenge? Preventing the model from simply outputting a constant, uninformative state. The system must retain enough input information while discarding the unpredictable. This method has shown strong results. In a direct comparison, non-generative models like V-JEPA outperformed traditional generative autoencoders in image and video tasks—even surpassing supervised models in some cases. LeCun called this a "clear empirical signal" that for natural sensory data, generative models are not the way forward. One of the most promising applications is in robotics. By training a world model—where the system learns to predict how the world changes in response to actions—robots can plan and act without task-specific training. "Given a state and a possible action, can the model predict the next state?" If so, the robot can simulate sequences, optimize for goals, and act with foresight. This is not just imitation; it’s true planning. LeCun’s team has already demonstrated zero-shot task execution in robots, using self-supervised world models trained over 62 hours of real or simulated movement. No reinforcement learning. No task-specific data. Just a model of the world. He believes this will become the dominant AI architecture within three to five years. "No one will use the LLMs we have today," he said. "This will be the decade of robots." On safety, LeCun is more optimistic than many. He advocates for "goal-driven" systems: models that act to achieve a defined objective, with safety rules hardwired into the goal function. "You can’t make a robot that knocks someone down to get coffee if the goal function includes a rule: don’t move if someone is in the way," he said, referencing a common AI safety thought experiment. "The system is structurally incapable of breaking the rules." He even compared this to human law: we don’t rely on perfect morality, but on rules that shape behavior. "Designing these safety layers is hard, but not harder than building a safe jetliner." To young researchers, LeCun offers a clear path: study deep, fundamental concepts. "If you're a student, pick quantum mechanics over app development," he said. "It teaches you path integrals, which are the same idea behind decoding speech or optimizing sequences. These are universal tools." He envisions a future where AI assistants handle the low-level work, allowing humans to focus on higher-level thinking. "The real secret is that students teach their professors," he said. "The future is not about doing more, but about thinking more." In the end, LeCun’s message is clear: the path to true intelligence isn’t in scaling LLMs or building human-like robots. It’s in building machines that understand the world—through self-supervised learning, world models, and goal-driven planning. The real revolution, he says, is just beginning.