I agree in principle. But I disagree that you cannot test GNFS on small numbers. You can do this, but you are right that you might have trouble to easily verify that it has a better asymtotic complexity than other algorithms.