BREAKING
Anthropic Finds 'Global Workspace' in Claude
0
%
of activations (under)
0
+
concepts at a time
How Claude Uses This Space
1
Silent thought forms
↓
2
Steps surface in space
↓
3
Edit steps
↓
4
Output changes
Delete the Workspace: What Breaks
Survives
●
Fluent text generation
Degrades
sharp
●
Multi-step reasoning
●
Summarization
●
Planning
A Safety Lead, With Caveats
Research Stage, Paper Unconfirmed
AI NEWS BLITZ
Anthropic says it can observe a global workspace inside its Claude models.