A global workspace in language models

Investigate interpretability research on Claude’s internal thinking and how J-space serves as a privileged workspace in language models. The page summarizes findings about Claude’s internal activations, introduces how to influence which patterns light up in J-space, and discusses implications for deliberate reasoning. Aimed at researchers and engineers, it highlights experiments, theories like the global workspace, and links to the full paper.

🔗 Full discussion: https://news.ycombinator.com/item?id=48808002