Descriptions:
AI safety and security researcher Roman Yampolskiy joins Lex Fridman for a deep and often unsettling conversation about his thesis that the development of superintelligent AI will almost certainly end catastrophically for humanity. Yampolskiy, a computer scientist at the University of Louisville and author of AI: Unexplainable, Unpredictable, Uncontrollable, argues that controlling a superintelligent system is mathematically analogous to building a perpetual motion machine—not just difficult in practice, but impossible in principle.
The conversation carefully unpacks the key distinctions between narrow AI safety (where systems can be tested exhaustively and edge cases identified) and general AI safety (where a single failure can be irreversible and civilization-scale). Yampolskiy points to a database of AI incidents he has contributed to as evidence that not a single large language model has ever been fully controlled by its developers, arguing this track record offers little reason for optimism as capabilities continue to scale. He places the probability of catastrophic AGI outcomes at 99.99%, while acknowledging a spectrum of failure modes ranging from extinction to a world where humans are kept alive but stripped of autonomy—”like animals in a zoo.”
Lex Fridman provides substantive counterpoint throughout, noting that most working AI engineers place P(doom) at 1–20%, and the two carefully examine why that gap exists and what evidence could resolve it. The result is one of the more rigorous public treatments of the AI control problem available.
📺 Source: Lex Fridman
🏷️ Format: Interview
