emergent abilities
/ ee-MUR-junt uh-BIL-uh-teez /
Emergent abilities are skills that a smaller AI model simply can't do, that seem to switch on once the model gets big enough — as if it crossed a threshold and a new capability suddenly appeared. The classic story: tiny models score essentially zero on, say, multi-step arithmetic or following a tricky instruction, and then past some size they start getting it right, sometimes jumping from near-zero to quite good. It looks less like gradual improvement and more like a phase change, the way water doesn't slowly become ice but flips at a temperature.
This drew huge attention because it suggested scale doesn't just polish existing skills but can unlock genuinely new ones you didn't train for or predict — which felt both exciting and a little unnerving, since you couldn't tell in advance what a bigger model might suddenly be able to do. It became one of the headline arguments for why the field kept building larger and larger models.
But the honesty here is important, because later research pushed back hard. A influential 2023 analysis argued that much of the apparent "emergence" is an artifact of how we measure: if you grade a task all-or-nothing (you only score if every step is exactly right), improvement looks like a sudden jump; if you use a smoother, partial-credit measure, the very same models often show steady, gradual gains the whole way up. So some emergent abilities are real surprises, while others are mirages created by the metric. The takeaway is to be skeptical of dramatic "it suddenly learned X" claims and ask how the ability was being scored.
On a three-digit addition task graded "all digits correct or zero," small models score ~0% and a large one suddenly scores 80% — looking emergent. Re-grade the same outputs by how many digits each got right, and you see the smaller models were already getting most digits correct: a smooth climb, not a jump. The ability grew gradually; the metric made it look like a switch.
Same models, two metrics: an "emergent jump" can flatten into a smooth curve.
Emergence is partly real and partly a measurement artifact. A sharp "the ability appeared at scale X" claim often softens into a gradual trend once you grade with partial credit — so always ask what metric was used before treating a sudden new skill as genuine.