BREAKING
Anthropic Finds Hidden 'Workspace' in Claude
Meet J-space via the J-lens
What J-space actually holds
Hidden reasoning
●
Math intermediate steps
●
Bug detection in code
●
Prompt injection recognition
Key properties
●
Reportable and modulatable
●
Broadcast downstream
●
Bypassed by simple recall
Inhibiting J-space breaks reasoning
1
Fluent output kept
↓
2
Inhibit J-space
↓
3
Reasoning degrades
Access vs phenomenal consciousness
Access consciousness
claimed
●
Reportable
●
Controllable
●
Usable in reasoning
Phenomenal
avoided
●
Subjective experience
●
No claims made
●
Caution urged
Code and demo now public
AI NEWS BLITZ
Anthropic says Claude has a hidden internal region that reasons without writing it down.