We develop machine-learning interatomic potentials that bring near-quantum chemical accuracy to molecular simulations of reactive processes in condensed phases, from bulk water and interfaces to complex chemical environments, while reaching time and length scales inaccessible to direct ab initio dynamics. A key focus of our work is data-efficient training (active learning) and the tight coupling of machine learning potentials with enhanced and path-sampling methods to reveal mechanisms, rate-limiting steps, and rare reactive events. Recent applications include uncovering the molecular mechanism of proton transport in water, achieving microsecond reactive sampling to dissect prebiotic reaction pathways relevant to origins-of-life chemistry, and the prediction of reactive properties in the low-data regime. These advances open a route to predictive, mechanistic modeling that can guide the design of new catalysts by connecting atomistic dynamics directly to reactivity and selectivity.