Two Minute Papers · 2026-07-15 · notable
Two Minute Papers: 'Claude's Brain Has A Secret... And Scientists Found It'
Károly Zsolnai-Fehér walks viewers through Anthropic's global-workspace paper: an internal 'J-space' inside Claude that mirrors a leading neuroscience theory of conscious access, discovered with a new Jacobian-lens interpretability tool.
Two Minute Papers explains Anthropic's J-space — a workspace inside Claude that reads like a language-model version of the global workspace theory.
What is it?
'Claude's Brain Has A Secret... And Scientists Found It' is Károly Zsolnai-Fehér's July 15 video walkthrough of Anthropic's 'A global workspace in language models' paper. The paper describes a small collection of internal patterns inside Claude — J-space — that each correspond to a word the model might later produce.
How does it work?
The video explains the Jacobian lens (J-lens), the tool Anthropic built to find J-space. For every word in Claude's vocabulary, J-lens computes the internal activity pattern that most increases the chance the model says that word later. Zsolnai-Fehér uses the paper's diagrams to show how these patterns cluster, emerge naturally during training, and let researchers read which concepts Claude is 'holding in mind' before it speaks.
Why does it matter?
Two Minute Papers is a common on-ramp for engineers to fresh interpretability research. J-space bridges to global workspace theory from neuroscience — one of the leading accounts of conscious access — while the paper itself is careful to talk only about 'access consciousness,' not phenomenal consciousness. For anyone building on Claude, the takeaway is that model-internal state is becoming steerable and inspectable, not just its outputs.
Who is it for?
AI engineers, interpretability researchers, and viewers who follow Two Minute Papers for accessible model-mechanics videos.