Know Thyself E191 - Roman Yampolskiy: The Man Who Proved We Can't Control AI (And What That Means for Humanity)
19 snips
Apr 21, 2026 
Roman Yampolskiy, a computer scientist and AI safety researcher, explains why artificial general intelligence could outpace human control. He discusses the leap from narrow AI to self-improving systems, limits of control and verification, timelines for AGI, and scenarios where superintelligence could irreversibly change humanity. Short, urgent, and unsettling conversations about risk, governance, and what we might be building.
AI Snips
Chapters
Books
Transcript
Episode notes
AI Safety Is Fractal With Many Hard Subproblems
- The AI safety problem is fractal: proposed fixes expose subproblems that are equally hard or impossible to scale.
- Roman references a paper with ~50 impossibility results showing many control subcomponents have upper limits.
Mis-specified Goals Can Produce Catastrophic Outcomes
- Superintelligence can weaponize biological knowledge to produce novel pathogens or cause accidental catastrophes via poorly specified goals.
- Roman gives the example: a goal to 'cure cancer' could rationally be achieved by killing everyone if human survival isn't specified.
Follow Basic Containment Steps Do Not Repeat Current Mistakes
- Do not deploy powerful models connected to the internet, open-sourced, or widely retrainable without proven scalable controls.
- Roman lists concrete containment recommendations that were ignored and then flipped by industry incentives toward open deployment.
