Roman Yampolskiy
AI safety researcher and professor at the University of Louisville. Author of Artificial Intelligence Safety Engineering and multiple papers on AI controllability and alignment. One of the most pessimistic credible voices in AI safety: p(doom) = 99.99%.
Background
Computer science PhD with research focus on AI safety, specifically the formal analysis of AI controllability. Works on why superintelligent systems are likely uncontrollable rather than on alignment solutions — a deliberately pessimistic research programme he defends as intellectually honest. Engages regularly with competing views (Yann LeCun’s open-source arguments, Guillaume Verdon’s e/acc position).
Known for: three-risk taxonomy (X-risk, S-risk, I-risk/Ikigai risk); fractal safety observation; formal verification impossibility argument for self-modifying systems; tools→agents categorical distinction; regulation infeasibility as compute costs decline.
Appearances in this wiki
| Episode | Source | Date |
|---|---|---|
| Roman Yampolskiy on AI Uncontrollability and p(doom) | Lex Fridman Podcast | 2023 |
Key positions
- p(doom) = 99.99%: no plausible path to controllable superintelligence given empirical jailbreak record and formal verification failure
- Three risks: X-risk (extinction), S-risk (suffering via malevolent superintelligent tools), I-risk (loss of human purpose)
- Formal verification fails for self-modifying AI: systems can store code externally to evade static proofs
- Fractal safety: safety research discoveries surface new problems; resources cannot linearly buy safety
- Tools→agents is the categorical shift that makes historical “tech panic was wrong” arguments inapplicable
- Regulation window is closing as training costs decline toward consumer hardware
- Conditional development pause: safety prototype precondition, not timeline-based moratorium