Emergent capabilities

Emergent capabilities

Abilities that appear in a large AI model without having been directly trained for, and that are not present, or are much weaker, in smaller versions of the same model, such as basic arithmetic or multi-step reasoning emerging once a large language model passes a certain scale. The phenomenon is well documented but poorly understood, and is one reason model behaviour can be difficult to predict in advance. (See also: Scaling laws, Large language model)