The Age of a Billion Answers: Promise and Peril in AI's Rare Outputs

Join this technological conversation on how we are transitioning into a new era: Language models now answer billions of questions each day, from everyday how-to requests to supporting doctors and scientists in their work. Yet each response is only one sample from a much larger distribution of possible answers. Zachary Horvitz, an AI researcher at Columbia University, argues that both AI's greatest promise and its greatest danger may lie in the rare outputs hidden within this distribution. He will explore how unusual outputs, and the methods researchers use to elicit and verify them, have shaped the development of artificial intelligence and what they could mean for scientific discovery. He will also examine the risks that rare model behaviors pose for a society increasingly reliant on AI systems.