I think its implying that ML often works because the manifold assumption is partially correct. so samples will lurk in the same neighborhood.