The development in the last decades in the sciences of complex systems and statistics has allowed us to define the limits of predictability better and -often- to extend those limits. The rapid progress of computing techniques and capabilities has even further bolstered those two branches of science to the point that it is now feasible both to compute the trajectory of plasma and magnetic field structures under the MHD equations in domains as large as the Heliosphere and to try to predict -just by analysing full disk images or the magnetograms- whether a given solar Active Region will release part of its stored energy as high energy photons and particles, or shoot out a coronal mass ejection. Yet, a robust forecast of flare eruption still escapes us. Even the apparently more straightforward problem of the propagation of a coronal mass ejection in the interplanetary medium has not been solved to the limit we would like. At the same time, we fight with the uncertainties associated with the boundary conditions. Consequently, some of us turned to the dark side and applied this hybrid approach of numerical methods, complex system science and statistics, which usually goes under Machine Learning (ML). Considering how large the ML field is, I can only share my personal experience of how ML contributes to the forecast of space weather and why it is probably here to stay. I will focus on some aspects, reporting on recent approaches that show how ML methods can go further than the common “black-box” approach. In particular, I will zoom in on two recent ML branches: employing attention-based tools to help the interpretability of Deep Learning models and the possibility of enforcing some of the system physics in training, thus limiting the results to be adherent to physical laws.