The internal mechanisms by which a model achieves a low negative log likelihood become irrelevant as it approaches perfect simulation of the true data distribution.