BREAKING
Anthropic finds 'J-space' in Claude
A quiet internal workspace
0
%
of activations
0
Sonnet version
How the Jacobian Lens works
1
Scan vocabulary
↓
2
Find activation patterns
↓
3
Apply layer by layer
↓
4
Watch J-space shift
What it means, and what it doesn't
Safety potential
upside
●
Detect prompt injection
●
Visualize reasoning steps
Big caveat
caution
●
Access, not phenomenal
●
Still research stage
Open-sourced for verification
AI NEWS BLITZ
Anthropic says it found a hidden thinking space inside its Claude models.