SafetyPage

Alignment Forum and LessWrong · Forum post · LessWrong

Personas in a Pre-Internet Language Model

David Africa · 13 Oct 2026

Read at lesswrong.com ↗Read at sourceNo licence statement found

Outline · 152 paragraphs

  1. TLDR¶2
  2. Why run experiments on a model from 1930?¶12
  3. Setup¶17
  4. Generalising Misalignment¶28
  5. Writing period-native datasets¶30
  6. Emergent misalignment¶37
  7. Qualitative differences¶65
  8. Subliminal Learning¶88
  9. In-Context Persona Generalisation¶95
  10. Utility Engineering¶103
  11. What Talkie likes¶105
  12. Coherent preferences come from post-training¶108
  13. How Much of This Comes From Modern Post-Training?¶114
  14. Personality Traits in Talkie¶122
  15. Conclusion¶130