Doom Debates! The Facade of AI Safety Will Crumble (Video)
Feb 17, 2026
A provocative take on why current AI safety efforts are shallow and may miss extinction-level risks. It questions psychoanalysis-style testing for advanced systems and explores the gap between abstract goals and real implementations. The discussion highlights how maturing AI could outmaneuver human-centered safety checks and why that makes future outcomes especially worrying.
AI Snips
Chapters
Transcript
Episode notes
Facade Of Current AI Safety
- AI companies treat safety as mechanistic psychoanalysis of current models rather than confronting core risks.
- Liron Shapira warns this approach will fail once truly mature outcome-optimizers arrive.
Know The Desired Superintelligence First
- The real challenge is knowing what a superintelligent system that does what we want would look like before we build a dangerous one.
- Shapira stresses we have little traction on this foundational design problem.
Level Separation And Why Brains Mislead
- Level separation explains why implementation quirks of brains let psychoanalysis work for humans but won't for mature AIs.
- Shapira contrasts weak level separation in brains with stronger separation in serious computing implementations.