Rich Sutton presents the OAK architecture, a vision for achieving superintelligence through experience-based learning. He emphasizes the importance of domain-general AI designs that learn from runtime experience rather than pre-programmed knowledge. Sutton introduces the concept of "options and knowledge" (OAK), where agents learn high-level transition models to enable planning with larger, more meaningful steps. He argues for the necessity of runtime learning due to the vastness and complexity of the world, advocating for agents that can discover and adapt like humans. The OAK architecture involves creating sub-problems from highly ranked features, learning solutions to these sub-problems, and using models of these options to plan and improve behavior, all while maintaining metadata on utility and curating new possibilities.
Sign in to continue reading, translating and more.
Continue
