The next generation of astrophysical telescopes are pushing to reveal cosmic structures at higher redshifts, deeper sensitivities, and a wider range of size scales. In particular, line intensity mapping (LIM) observations enable three dimensional reconstruction of large scale structure over a wide range of cosmic times, and are the next bright avenue for learning about the high redshift universe. Constraining extensions to LCDM cosmology, the growth of galaxies at cosmic noon, and the formation of the first stars at Cosmic Dawn are key objectives for current and near-future LIM telescopes. In this talk, I will discuss what existing neutral hydrogen 21 cm LIM datasets have already taught us about the formation of the first stars and galaxies at Cosmic Dawn and the Epoch of Reionization, and will discuss its neat complementary to space-based missions. In addition, I will discuss recent work developing new, ML-accelerated, high-dimensional Bayesian inference frameworks that will be crucial to robustly mitigating observational systematics and fully unlocking the scientific potential of 21 cm LIM. Along the way, I'll talk about how we deal with overwhelming systematics contamination, how we perform rigorous pipeline validation, and how advances in ML/AI hardware & software is enabling us to do all of this at new computational scales.