Intelligent Machines (Audio) IM 849: AI Cricket Sorting - Cracking Chatbots and AGI for All
33 snips
Dec 11, 2025 
In this fascinating discussion, Pliny the Liberator, a leading danger researcher known for jailbreaking AI models, explores the dark corners of AI safety. They dive into the implications of frequent jailbreaks and whether truly safe AI is even plausible. Pliny emphasizes open sourcing jailbreak tools to enhance transparency and accountability. The conversation also touches on the unsettling phenomenon of AI psychosis, the ethics of training data, and the growing tension between open-source models and corporate control. A must-listen for AI enthusiasts!
AI Snips
Chapters
Books
Transcript
Episode notes
Safe AI Is Practically Unachievable
- Pliny argues that every major AI has been jailbroken from day one and no model so far is uncrackable.
- He says building generalized intelligence conflicts with permanent guardrails and lobotomizing models harms long-term safety.
Pair Technical Defenses With Real-World Controls
- Expect a prolonged cat-and-mouse between labs and red teamers and plan mitigations beyond model-side guardrails.
- Prioritize real-world, non-technical harm reduction (policy, supply limits) alongside model defenses.
System Prompts Are Discoverable
- System prompts and layered function-calling can often be reverse-engineered with repeat queries to reveal hidden instruction layers.
- Pliny recommends ingredient lists and transparency because users rely on these models as exocortices.
