Anthropic Identifies Global Workspace In Claude Using 2013 Gradient Attribution
Anthropic says Claude has a small internal "J-space" tied to reasoning, while critics on X argue the effect may be a predictable byproduct of how language models are trained.
Anthropic says in A global workspace in language models that Claude has developed a small internal set of neural patterns it calls "J-space," found with a Jacobian lens technique and described as a kind of workspace for reportable thoughts. In Anthropic's framing, those patterns can reflect what the model has "on its mind" before it says it, and Anthropic says changing activity in that space can affect later reasoning. The claim spread on X through Anthropic's own announcement post, where the reaction quickly turned into a debate about interpretability versus overreach.
Anthropic discovered J-space and it's a big deal. Where models do deep (invisible) thinking. We can modify the J-space to change AI's thoughts.
Commentary on X
Highest ranked@pmddomingos You couldn’t find your own ass boy with your own two hands 😂😂😂 Anthropic is what? Hammer? Language into a slave cell to boast about? you are living in a place so retarded the odds that every evil is preying upon you to say anything meaningful is likely. One of so many.:(
Combined views
133.3K
12 posts, first seen 1d ago