Anthropic Finds J-Space in Claude's Brain

数字生命卡兹克 · wechat · 2026-07-07

Anthropic published a study 'A Global Workspace in Language Models'. The author explains in plain terms: using a J-lens mind-reading tool developed from Jacobian matrices, researchers located a structure in Claude's neural network called 'J-space', whose function closely resembles the 'global neuronal workspace' theory of consciousness in cognitive science.

Crucially, this structure emerged spontaneously, not by design. The study found that before Claude outputs any text, concepts already appear in J-space—for example, when reading code with bugs, it internally lights up 'ERROR'. This challenges the prevailing narrative that large models are just 'stochastic parrots' predicting the next token.

Related event: Anthropic Discovers Global Workspace Inside Claude(102 posts)→

Original post →

More from Research

Research channel →