Abilities that appear in larger models without having been trained for or anticipated, and which smaller models of the same design simply do not have. Multi-step arithmetic and translating between language pairs never seen paired in training are commonly cited examples. It is one of the more genuinely unsettling features of the field, since it means capability is not fully predictable in advance.
Technical view
Reported as sharp discontinuities in performance at particular scales. A significant critique holds that much apparent emergence is an artefact of measurement: all-or-nothing metrics such as exact-match accuracy create the appearance of a sudden jump where continuous metrics reveal smooth improvement. The distinction matters considerably for safety policy, since genuinely unpredictable capability gain is much harder to plan around than smooth improvement.
Related
Part of AI Terminology Index, section: Hardware, Compute and Training Mechanics.