Whether you are among believers in machine learning or among nonbelievers, these days, you have heard a lot about a somehow seductive and revolutionary theme saying: let us replace classic designs with pure machine learning-based solutions. For instance, in the network community, there are different efforts to replace current congestion control heuristics with clean-slate learning-based ones. In this talk, I will not introduce yet another clean-slate learning-based design and I will not promote the revolutionary philosophy of replacing classic designs with learning-based ones! Instead, I will try to introduce a more pragmatic and evolutionary design philosophy for using learning-based techniques. To that end, I will show how learning-based techniques can be put wisely in harmony with classic heuristics in the context of congestion control. And how this harmony can lead to way higher performance and way lower overhead for the system compared to the clean-slate learning-based designs.