Richard Sutton, Turing Award recipient and reinforcement learning pioneer, delivered a pivotal address at the closing session of the 2026 World Artificial Intelligence Conference (WAIC), outlining a definitive roadmap for artificial intelligence beyond current large language models. Making his first public speech since founding Oak Lab and departing Keen Technologies, Sutton reiterated his long-standing thesis that contemporary foundation models will not evolve into superintelligent systems, but will instead solidify as highly reliable knowledge repositories. Grounding his argument in his seminal 2019 paper, The Bitter Lesson, Sutton noted that AI research history consistently demonstrates that methods relying on human expertise and handcrafted rules are ultimately superseded by approaches that leverage massive computation, search, and machine learning. While modern LLMs exemplify this trajectory by scaling with data and compute, Sutton argued they remain fundamentally constrained by their dependence on static, human-generated corpora. They excel at organizing and retrieving existing knowledge but lack the capacity to explore, experiment, and generate novel insights through real-world interaction. This limitation, he warned, represents the inevitable ceiling for foundation models. To transcend this bottleneck, Sutton advanced the Big World Hypothesis, asserting that the complexity of the physical and social environment indefinitely outstrips any single computational system. No matter the parameter count or training volume, a model cannot encapsulate the entirety of reality or replicate the nuanced, contextual understanding humans develop through lived experience. Consequently, Sutton proposed a bifurcated future for AI. The first category comprises specialized, boundary-defined systems optimized for specific tasks. The second, which he views as the true trajectory toward advanced cognition, consists of embodied agents capable of continuous, lifelong learning through direct environmental engagement. He characterized these as systems that do not claim omniscience but instead accumulate personal expertise through action and feedback. Within this framework, Sutton redefined the role of current foundation models. Rather than pursuing unattainable general intelligence or generating speculative answers, they should be engineered as super encyclopedias. Their primary function is to accurately catalog, organize, and transparently deliver humanity’s established knowledge while clearly demarcating the boundaries of what is known. He emphasized that addressing hallucination and improving factual reliability must precede further scale expansion. This positioning aligns with his broader advocacy for an Era of Experience, a paradigm shift where artificial intelligence evolves from passive data consumption to active, interactive exploration. The conference also featured reflections from social anthropologist Xiang Biao, who cautioned against allowing algorithmic text generation to erode human capacities for patience, critical analysis, and independent judgment. Sutton’s technical vision and Biao’s sociological perspective converge on a singular conclusion: while AI can efficiently curate established information, profound understanding and knowledge creation require direct human or agent engagement with the physical world. As the industry navigates post-LLM development, Sutton’s WAIC address establishes a clear directive. The next generation of AI will not emerge from larger static datasets, but from architectures designed to learn continuously through experience, interaction, and real-world trial and error.