This paper presents a roadmap for developing scalable, aligned artificial intelligence (AI) based on an explanation of the fundamental principles of natural intelligence. A possible path toward scalable, aligned AI lies in enabling artificial agents to learn good models of the world, including our preferences. To achieve this, a key goal is to create agents that learn to represent the world and the world models of other agents—a problem known as structured learning (also known as causal representation learning or model discovery). With this goal in mind, this paper presents the principles that will guide us forward, along with the structured learning and alignment problems, synthesizing diverse ideas from mathematics, statistics, and cognitive science. 1) We discuss the essential roles of core knowledge, information geometry, and model reduction in structured learning and propose a core structural module for learning from a broad range of natural worlds. 2) We outline a path toward aligned agents through structured learning and theory of mind. As an example, we mathematically outline Asimov's Three Laws of Robotics, which prescribe agents to act prudently to minimize the misfortune of other agents. We also complement this example by proposing an improved approach to alignment. These observations can serve as guidelines for developing artificial intelligence that can help extend existing aligned structure learning systems or design new ones.
