BREAKING
Anthropic finds 'J-space' in Claude
0
+
pages
0
x
GitHub release
How the Jacobian Lens works
1
Scan vocabulary
↓
2
Read Jacobian
↓
3
Extract J-space
Four key properties of J-space
Reportable
verbalize
●
Model can state its thoughts
●
Modulable by instruction
Selective
limited
●
Mediates multi-step reasoning
●
Only part of processing
Access vs subjective experience
Tool aimed at alignment audits
AI NEWS BLITZ
Anthropic says it discovered a hidden inner-thought space inside its model Claude.