Author argues search and self-play let AI surpass humans

The author argues that human intelligence serves as a floor rather than a ceiling for artificial intelligence, because once models reach human-level competence through imitation, verification methods like search, self-play, and proof checkers allow them to surpass human performance. Citing examples from AlphaGo to automated mathematical theorem proving, the piece contends that verifiable environments enable AI systems to scale beyond human limitations. Consequently, the author suggests people should focus on directing AI and defining verification rather than worrying about matching task execution. Open article →