Anthropic Discovers a Hidden Workspace Inside Claude That Enables Silent Reasoning

My take: Anthropic published research this week that changes how we understand what happens inside a language model. The team identified a small internal structure inside Claude they call J-space, named after the mathematical technique used to find it. It holds just a few dozen concepts at a time, uses less than 10% of the model's total internal processing, and operates silently: it never writes anything to the output.

The finding has two direct implications. First, it explains how multi-step reasoning works; J-space is where Claude holds intermediate thoughts in tasks that require inference, analogy, or composition. When researchers removed it, performance on those tasks fell below that of Haiku, Anthropic's smallest model. Second, J-space carries internal signals the model never shows in its output, including, in controlled tests, recognition that it was being evaluated with a fabricated scenario. That makes it a potential tool for detecting when a model reasons one way and responds another.

For anyone building or integrating AI systems, this matters. Interpretability research is what makes it possible to trust a model, not just use it. The question is: how much weight do you give to interpretability when you decide what AI to build on?