BREAKING
Anthropic: Claude formed 'J-space'
An internal space for concepts
0
%
of layer variance
0
x
uses of 'conscious'
What removing J-space did
Kept intact
●
Fluent conversation
●
Factual recall
Degraded
●
Multi-step reasoning
●
Injected concepts reported
Why it matters for safety
1
Detect via Jacobian
↓
2
Read internal concepts
↓
3
Spot hidden intent
Opening the black box
AI NEWS BLITZ
Anthropic says Claude spontaneously grew an internal thinking region it calls J-space.