For two centuries the slow part of materials science was imagination: guessing which arrangement of atoms might hold together long enough to be worth a trip to the furnace. In late 2023 DeepMind's GNoME model predicted 2.2 million new crystal structures, of which some 380,000 were computed to be stable, roughly ten times the number humanity had catalogued in all the years before. Almost none of them exists outside a database. The bottleneck of the field has moved, in one paper, from thinking of materials to making them.