Everyone spent two years mocking LeCun for calling LLMs “glorified autocomplete.” He wasn’t complaining. He was diagnosing. While the industry poured billions into making next-token prediction marginally better, LeCun mapped an entirely different architecture. World models.