A mechanism-first guide to optimizers: what SGD, momentum, RMSProp, Adam, and AdamW each solve, why AdamW became a strong default for modern deep learning, and when other optimizers still matter.