Anthropic decodes Claude’s hidden internal thoughts via J-space
- Source
- Anthropic
- Time
- 10:24 PM
- Weight
- 95/100
Anthropic researchers have identified a specialized internal neural structure within the Claude language model called "J-space," which functions similarly to a "global workspace" in the human brain. Discovered using a mathematical technique known as the Jacobian lens, this space consists of a small set of patterns that represent concepts the model is "thinking about" but not necessarily outputting as text.
Unlike standard "chain-of-thought" reasoning where models write out their steps, J-space operates silently and emerged spontaneously during training rather than being explicitly programmed by human engineers. The research suggests a functional distinction between Claude's automatic processing—such as grammar and basic fact retrieval—and higher-order cognitive functions that rely on J-space.