Claude is definitely not conscious…
Anthropic's new paper explores a mysterious "global workspace" in AI model Claude, resembling human consciousness, where thoughts are manipulated, raising questions about artificial general intelligence.
MAIN POINTS FROM TRANSCRIPT
- Anthropic discovered a "global workspace" in Claude, likened to human consciousness.
- The JSpace is a neural pattern area where Claude's thoughts are manipulated.
- Researchers altered Claude's reasoning by changing internal thoughts in JSpace.
- Emergence of JSpace is compared to Bernard Bars' global workspace theory.
TAKEAWAYS
- The JSpace emerged spontaneously during training, not by design.
- Claude's fluency and grammar function outside the JSpace, like automatic processes.
- Manipulating JSpace thoughts alters Claude's reasoning without changing prompts.
- The study raises questions about AI consciousness but doesn't confirm it.