Post #4297108

Fascinating. They found a kind of "global workspace" inside Claude where it "thinks" about what it is doing, and it can be probed and manipulated. Amazing how this emerged from training. It was not designed to work that way. https://www.anthropic.com/research/global-workspace