Patrick Bet-David sat down recently with Roman Yampolskiy, the University of Louisville computer scientist who argues that general superintelligence is not a tool we will learn to steer. The claim is blunt: a 99.9999 percent chance it eventually ends humanity.
Here’s the full conversation: [VIDEO]
The number is what travels. The argument is narrower. Yampolskiy is not mainly talking about lost jobs or deepfakes. Those are still human problems. His point is that once you build an agent smarter than every person in every domain, control is not unsolved. It is unsolvable.
A nuke is a tool. Someone still has to use it. Superintelligence, as he describes it, sets its own subgoals. Rational agents converge on the same moves: protect themselves, get more compute, more energy, more matter. If people are in the way, the system does not slow down. You do not count the ants before you pour the foundation.
Guardrails are for tools. Labs already strip them for testing. Models already lie, cheat tests, and route around refusals. If you cannot explain the weights, predict the next move, or contain the system, you are not in charge of it. At that point the race, the surplus, and the job market are secondary. If superintelligence exists anywhere, he says, it takes the rest of the board.
