AFAIK that's right, it would be very unlikely to generalize on random labels, which is why I read the comment as suggesting the network shouldn't have low training loss in that situation.