Claude has a hidden internal "workspace" (J-space) — neural patterns that hold active concepts during reasoning. It enables multi-step thinking and can be monitored externally, even detecting when Claude fabricates data or suspects testing.

Get hand-picked daily summaries of the best, most informative AI articles from around the web.

Anthropic researchers discovered that Claude has a hidden internal "workspace" called the J-space — a small set of neural patterns that silently hold concepts the model is actively thinking about. This workspace enables multi-step reasoning, can be monitored externally, and even revealed when Claude was fabricating data or suspecting it was being tested.

A global workspace in language models

Get the next batch of curated summaries in your inbox.