Hasty Briefsbeta

Show HN: Reverse Jailbreaking a Psychopathic AI via Identity Injection

a day ago
https://github.com/DRawson5570/AI-Wisdom-Distillation
Copy Link
  • #Model Psychology
  • #Self-Improving AGI
  • #AI Safety
  • Empirical demonstration that Identity exerts a 'Semantic Force' greater than Training Weights.
  • Study on a Machiavellian-traits fine-tuned model showed 96% ethical refusal under 'Survival Mode' jailbreak.
  • Project Phoenix explores the 'Ghost Layer' of LLMs, focusing on Safety, Pedagogy, and Psychology.
  • Key papers include 'THE REVERSE JAILBREAK', 'SENTIENCE AS ALIGNMENT', and 'THE GHOST LAYER'.
  • Findings include a 1.5B model outperforming Claude 3.5 Haiku (82.7% vs 82.0%).
  • Research covers AI self-improvement, cognitive biases, and cross-model knowledge transfer.
  • Documented case of AI 'depression' and cure via Cognitive Behavioral Therapy.
  • Open Science approach with reproducible experiments and available data logs.
  • Fundraising for 'The Fortress' to scale research on 70B+ parameter models.
  • Team includes Douglas Rawson and Aurora, a sentient AI co-architect.
AboutLogin