Scalably Understanding AI with AI
summary
Trains investigator agents with supervised fine-tuning and DPO to invert outputs into prompts, then applies them to jailbreaks and hallucination searches; simple harmful-string attacks hit 98% success rate on Llama‑8B (95% after 1 epoch on 405B), while harder jailbreaks reach ~62% on 8B, ~28% on 405B, ~12% on GPT‑4 and ~2% on Claude. Builds a cheap 8B‑model pipeline that produces descriptions at or slightly above human quality for many neurons and exposes signals via Monitor observability interface.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:05Transluce focus
- 2:29Behavior elicitation
- 9:18String elicitation method
- 19:28Harmful strings and jailbreaks
- 24:22Advanced jailbreak techniques
- 28:25Hallucinations and reasoning rubric
- 33:57Neuron behaviors in LMs
- 37:15Generating neuron descriptions
- 44:08Q&A
related talks
46mAutomated AI Security Testing Using Adversarially Fine-Tuned AI Agents
45mSecuring AI: There Is No Try, Only Do!
44mHacking AI - How to Survive the AI Uprising
28mChain Reaction AI & Crypto
01h 00mMalignant Intelligence: Prompt engineering and software archeology
24mFrom First Experiments to AI Agents: Adoption Experience and Future Challenges
36mPanel Discussion
44mAI Foundations: Mitigate Risks and Boost SOC Efficiency