Functional Workspace Structures in Language Models: Open Research Inquiry with Research Overview

Current research on advanced AI models has described an emergent structure that was notexplicitly programmed and that exhibits functional properties of a Global Workspace. GlobalWorkspace Theory is an influential neuroscientific theory of how information becomes availableto different cognitive processes. Anthropic calls the structure found in its models J-Space. Thestudy shows that it is involved in the models’ internal reasoning. To situate this finding and thesubsequent research, this inquiry invites frontier developers and AI researchers to provideexpert feedback on the current state of research.