Most of our understanding of the Universe stems one way or another from star-light. However, only by building realistic models of convection and radiative transfer in the dynamic stellar atmospheres, can we predict the detailed strengths and shapes of lines in the emitted light and accurately infer the properties of stars. Recently, we reached the milestone of making 3D non-LTE line formation of highly complex species like Fe numerically feasible for late-type stars and can for the first time predict their metallicities without invoking free parameters. I will demonstrate how improved models affect our understanding of the origin of elements and chemical evolution in stars, stellar populations, and the Galaxy. The next challenge lies in incorporating physically more realistic models into automated spectroscopic analysis pipelines for million-star spectroscopic surveys, e.g. the ongoing Gaia-ESO and GALAH. Looking ahead, their successors WEAVE and 4MOST will expand the stellar inventory by another order of magnitude: what questions about the history of our Galaxy can then be asked and answered?