Error bars on DFT predictions: sense or nonsense?