Making predictions on new tasks without any task-specific training or fine-tuning on labeled examples.