Safety
53d agoAnalyzing Anthropic's Global Workspace Theory
Section 2.1 of Anthropic's latest paper introduces J-space to identify verbalizable representations in language models. This technical breakdown attempts to reverse-engineer the methodology used to detect a global workspace. It focuses on the conceptual calculation of these spaces rather than empirical data.
The Signal
The analysis helps researchers better understand how internal model states map to human-readable concepts.