2026-01-30
- ARENA W3D5
- Experiments
- When does the refusal direction emerge?
- exists in base models
- makes sense? because conversations with AIs are in training data
- later layers are more important
- exists in base models
- Mech interp on "end thinking" tokens
- Does it matter if you do helpful training first or harmless training first?
- When does the refusal direction emerge?
- Experiments