This page traces the canonical components of a language model and its training pipeline, from the sequence-to-sequence era of 2014 to the consolidated decoder-only design of the present. The striking feature of that history is convergence: since 2017 the architecture itself has changed little, and each era's progress has come less from new components than from scale, data, and new ways of training the same ones. Where the next improvements come from is an open question. One avenue might be generalization: models that abstract more from less data, recombine what they know to handle problems unlike any they were trained on, and remain reliable outside their training distribution rather than merely interpolating within it. Another might be analytic distillation: compressing what large models have absorbed into smaller, cheaper networks, and increasingly into explicit, inspectable structures that can be checked and reused. Both avenues treat understanding, rather than parameter count, as the scarce resource. Drag the slider to move between years; select any linked component to open the corresponding article.