A new technique enables huge machine-learning models to efficiently generate more accurate quantifications of their uncertainty about certain predictions. This could help practitioners determine whether to trust the model when it is deployed in real-world settings.