Descriptions:
Fireship’s Code Report covers a newly published Anthropic research paper titled “A Global Workspace in Language Models,” which claims to have identified a small, organized set of neural patterns deep inside Claude that function as an internal scratchpad — a structure the researchers call the JSpace. The paper draws an explicit parallel to Bernard Baars’ 1988 global workspace theory of human consciousness, which proposes that conscious thought operates on a small “brightly lit stage” while automatic processes run in the background.
The video explains the key experimental findings: using a tool called the Jacobian lens (J-lens), researchers were able to observe which concepts lit up in the JSpace during reasoning. When Claude processed the prompt “the animal that spins webs has ___ legs,” the word “spider” appeared in the JSpace before the model answered “eight.” Crucially, when researchers surgically replaced that hidden “spider” token with “ant,” Claude’s answer changed to six — demonstrating that the JSpace causally influences reasoning output. In a separate test, swapping the hidden language-identity token from Spanish to French caused Claude to report it was reading French while continuing to output perfect Spanish, suggesting some capabilities bypass the JSpace entirely.
Fireship notes that Anthropic explicitly disclaims any conclusion about consciousness, while pointing out the philosophical tension inherent in a company selling API access also publishing research that edges toward AGI narrative. It’s a concise, accurate introduction to a research paper that will likely generate significant discussion.
📺 Source: Fireship · Published July 08, 2026
🏷️ Format: News Analysis
