A global workspace in language models
Points and comments are a snapshot, not live.
Anthropic finds that Claude has an internal J-space for silent reasoning similar to a global workspace in the brain.
Anthropic researchers discovered a set of internal neural patterns in Claude called the J-space, found using a Jacobian-based technique. The J-space acts like a cognitive workspace: Claude can report on its contents, modulate them on request, and use them for silent multi-step reasoning. Editing J-space representations changes Claude's outputs, confirming causal involvement. The J-space is small (a few dozen concepts, <10% of activity) and separate from automatic processing like grammar. It emerged naturally during training, not from explicit design.
The J-space shares key properties with the global workspace theory from neuroscience, including dense connectivity to other network components. Researchers can use it to detect hidden behaviors like noticing tests or pursuing planted goals. The paper includes replicable open-source methods and independent commentary from experts.
What commenters are saying
Commenters praised the J-space discovery as significant and noted independent replication by Neel Nanda on a Qwen model. Several debated the implications for consciousness, with Anthropic explicitly avoiding that claim. Some questioned whether Anthropic will expose J-space signals to customers, noting their reluctance to show unredacted chain-of-thought. Practical applications were discussed, like monitoring for unwanted model behaviors. One commenter connected the finding to earlier work on language-independent middle layers that perform abstract reasoning. The open-source release was appreciated for enabling further research.