Claude discovered a hidden space of "conscious" thoughts. However, no one designed it; it just appeared there.
Anthropic discovered that within the model there is a small set of neural states that act as a kind of working memory.
The researchers called it J-space and describe it as an internal space where what the model is currently "thinking" about is stored but not communicated to the user.
Until now, it was thought that AI simply predicted the next word. Now it turns out that before responding, it develops internal representations that it uses for reasoning.
Furthermore, if the J-space is suppressed, the model dramatically loses its ability to reason multi-steply, accurately retell information, and perform complex planning.
4
3
1July 7, 2026 1.6K 12
