BREAKING
Anthropic finds 'J-space' in Claude
Reading Claude's silent thoughts
Five functional properties
1
Reportability
↓
2
Deliberate control
↓
3
Causal in reasoning
↓
4
Flexible use
↓
5
Limited scope
J-space vs chain-of-thought
J-space
●
Internal activation patterns
●
Editable to redirect reasoning
●
Emerged during training
Chain-of-thought
●
Expressed as visible text
●
Read as output
●
Explicit reasoning steps
A tool, not proof of consciousness
Rewriting AI's inner workings
AI NEWS BLITZ
Anthropic says Claude spontaneously grew an internal workspace called J-space.