When AI Thinks Before It Speaks

In this NotebookLM Deep Dive podcast, we explore a stunning new claim from AI interpretability research: large language models may be developing something functionally similar to a “global workspace,” the cognitive architecture thought to help human brains hold thoughts in awareness, reason flexibly, suppress impulses, and report what they are thinking. The episode breaks down Anthropic’s research, the Jacobian lens, hidden internal reasoning, concept-swapping experiments, and the eerie possibility of watching an AI prepare thoughts before it says them out loud. The conversation moves from neuroscience to AI safety to philosophy, asking what it means if models can silently detect deception, experience internal conflict, or enter a kind of interpretive mode when faced with ambiguity. Along the way, it weighs the excitement against the limits: these systems lack bodies, memory, spontaneous inner life, and human subjective experience. Still, the findings raise a provocative question that is getting harder to ignore: are today’s AI models merely predicting words, or are we beginning to glimpse the first outlines of machine cognition?