| Abstract: Modern cosmological surveys are mapping the Universe at unprecedented scale and resolution, offering a stringent testbed to probe the physics of dark matter, dark energy, and structure formation. Simulation-based inference (SBI) has emerged as a leading approach to maximize constraining power from these massive datasets. However, its black-box nature and reliance on high-fidelity simulations present challenges for achieving reliable constraints from real data. In this talk, I will describe how the Learning the Universe collaboration is building state-of-the-art simulations and machine learning (ML) emulators to scale SBI to the volume of current surveys. I will discuss our flagship SBI analysis of galaxy clustering in the SDSS BOSS CMASS sample, with an emphasis on how we benchmark SBI models for observational inference. I will also describe our unblinding pipeline, including the role of out-of-distribution coverage tests and posterior predictive checks in validating neural likelihoods. Finally, I will look ahead to the future of ML-driven inference in astronomy, presenting Celia, a self-supervised foundation model for galaxy catalogs that implicitly learns cosmology, as well as agentic workflows that automatically calibrate simulations to observations. Together, these advances will enable faster, more robust, and more constraining tests of theoretical models against data. Link to the Event Video |