I think he already covered both, the second with "or I have to write my own separate implementation for every and each type of graph I'll use, without any significant way to reuse code".