What is it?
Emergent ability
An emergent ability is a skill that a model does not have when small but suddenly displays once it crosses a certain size. Below the threshold the ability is essentially absent; above it, the ability appears. Examples historically included certain kinds of arithmetic and multi-step reasoning.
Emergence is part of what made scaling so exciting: simply making models bigger sometimes unlocked qualitatively new capabilities, not just smoother improvements.
From the Foundational Modelling glossary
As the Planning and Reasoning glossary puts it
An emergent ability is a skill a model lacks when small but suddenly shows once it crosses a certain size. Chain of thought is the classic example (Chapter 1): prompting a small model to reason step by step barely helps or even hurts, but past a scale threshold the same prompting produces a large jump. In Chapter 4, careful reasoning behaviors emerge in a different sense, appearing during reinforcement learning because they earn rewards, without anyone programming them in.