Textbook of AI

Abstract. LeCun's manifesto on next-generation AI architecture. Argues that current frontier autoregressive language models are dead-ends because of four properties, they cannot plan, cannot reason, lack persistent memory and have no world model. Sketches a hierarchical architecture composed of a configurator, a perception module, a world model, an actor, a short-term memory and a critic, all coordinated through energy-based learning. The paper is the most prominent statement of the position that scaling autoregressive LLMs will not produce artificial general intelligence and informs Meta's V-JEPA and JEPA-2 line of research.