What Do You Mean by "Open World"?

Abstract

This paper explores the different understandings of the term "open world" in AI research, categorizing them into partially open world and fully open world. By distinguishing between intelligence and skills, I argue that a fully open world is essential for evaluating intelligence, while a partially open world risks shifting focus towards problem-specific skills. However, both partially open and closed worlds can still be valuable research tools if the underlying assumptions of AI systems are carefully examined.