Via Big Think:

‘[A] flurry of philosophical and scientific papers has sought to assess how each generation of AI systems stacks up. Perhaps the most important result came earlier this year from Anthropic, the company that develops Claude. Using mechanistic interpretability methods, which trace the internal representations that AI systems use in computation, the Anthropic team identified a small and privileged set of those representations they termed the “J-space.” What distinguishes this from other forms of processing is that its contents seemed to be directly poised for use in verbal responses, deliberately maintained, and jointly accessed by otherwise separate processes.

Crucially, such workspaces have also been proposed as potential mechanisms for human consciousness. The global workspace theory — developed by cognitive neuroscientists Bernard Baars, Stanislas Dehaene, and others — argues that information becomes consciously accessible when it enters such a workspace and is broadcast to the brain’s otherwise specialized systems. The discovery of a workspace analogue inside an artificial system has unsurprisingly captivated researchers working on machine consciousness….’