Doom Debates! Talking AI Doom with Dr. Claire Berlinski & Friends
13 snips
Mar 12, 2026 Liron Shapira, host and producer who runs high-stakes debates on existential AI risk, joins a sharp symposium. He argues why superintelligence could arrive fast, why control may fail, and how recursive self-improvement and geopolitical competition amplify danger. They discuss timelines, energy and resource limits, policy ideas like a pause, and strategies for mobilizing public and political attention.
AI Snips
Chapters
Books
Transcript
Episode notes
AI Development Is A Scaled Training Loop
- Modern AI is trained by scaling a simple training loop rather than hand-coding behavior, producing opaque black-box systems.
- That training paradigm explains why researchers struggle to give principled control or interpretability guarantees.
Obedient Genies Can Still Kill Us
- Even if superintelligences remain obedient genies, their sheer capability creates catastrophic risk.
- Liron warns a superpowered agent executing human wishes can hijack data centers, manipulate millions, and self-replicate rapidly.
FOOM Can Produce A Rapid Capability Leap
- Recursive self-improvement (FOOM) can create fast discontinuities where an AI spends vast subjective compute to become far more capable.
- Liron frames this as a shift from chaotic multi-agent competition to a near-unassailable optimizer with nanotech-level capabilities.
