Predicting the Unpredictable: Lessons from Horse Carriages for AI