Emergent abilities are skills a model cannot do at small scale but suddenly can at large scale.
DefinitionWhat it means
An emergent ability is a capability, such as multi-step arithmetic or basic chain-of-thought reasoning, that is essentially absent in smaller models and appears fairly abruptly once parameter count and training data cross a threshold. Unlike the smooth improvement predicted by scaling laws, emergent abilities show up as a step change on a specific benchmark, which makes them hard to plan for in advance.
Why it mattersWhy you should care
Emergent abilities are a source of both excitement and risk in product planning: a feature that fails reliably in a smaller or cheaper model may suddenly become viable in the next model generation, or a safety issue may appear unexpectedly at scale. Teams building on foundation models track emergent-ability research closely because it affects both roadmap timing and risk assessment for new deployments.
At a glanceSee it
A compound task scores only when every chained sub-skill fires — so smooth per-skill gains yield a sudden jump once the weakest link turns reliable.
Whether a skill looks emergent can hinge on the metric — an all-or-nothing score turns steady underlying progress into an apparent overnight leap.
Where you see itIn the wild
- Benchmark papers plotting accuracy versus model scale
- Debates about whether emergence is real or a measurement artifact
- Release notes highlighting a new capability in a larger model tier