Anthropic Finds Claude Contemplating Concepts in Hidden Space
New interpretability technique reveals previously unobserved internal reasoning processes in large language models.
ARWF dimensions
- Existential criticality
- Does the threat involve irreversible systemic failure?
- Probability vectoring
- Theoretical, or active proof of concept?
- Timeline imminence
- How close to current deployment?
- Mitigation gap
- Identified solution, or currently unaligned?