J-space.

This research into J-space blew my mind. Anthropic found that a small internal workspace can emerge on its own during training. When a concept enters it, the model can report it, hold it deliberately, and reason with it. Much of the rest runs automatically underneath.

The model’s grip is loose, though. Told not to think about something, it half fails, the way people told not to think of a white bear do. The researchers do not yet know what decides what enters the workspace. And when they swapped one concept for another, the model reported the implant as the thought it had chosen.

This is a functional distinction, not a claim that models have a subconscious. Humans build external structures that do something related. Seymour Papert called such artefacts objects-to-think-with. Language, maps, diagrams, notebooks, and interfaces give parts of complex systems a form we can inspect, manipulate, and share.

The researchers found J-space by looking for representations the model could put into words. The J-space they can currently see is built almost entirely out of words. Maps and diagrams are workspaces made of something else, and they sit where their maker can check them. The model already borrows the trick, thinking out loud on a scratchpad when the silent workspace runs out of room. I suspect that is what external objects-to-think-with are for. They extend the workspace into forms words cannot hold, on surfaces we can check.