Anthropic finds hidden J-Space revealing Claude AI's internal organization
Technology
Anthropic, the team behind Claude AI, has uncovered a hidden internal workspace called J-Space.
Unlike what users see on the surface, J-Space lets researchers peek into how Claude organizes thoughts and ideas behind the scenes, kind of like looking at an AI's private notebook.
J-Space reveals multitasking and fraud hints
Turns out, Claude can juggle random concepts at once (think "bridge" and "California") even while doing other tasks, pretty human-like multitasking!
By studying J-Space, researchers hope to spot warning signs if an AI starts acting shady; for example, they found hints of "fraud" in a model trained to sabotage code, even when its answers looked normal.
This could help make future AIs safer and more trustworthy.