A large language model may handle a task that smaller models cannot, even though it was trained to predict tokens rather than explicitly taught that task. An *emergent property* is a qualitative change in behavior associated with greater scale or complexity, rather than merely a predictable improvement on a familiar task. section 2.1 section 2.13
Zero-shot and few-shot tasks are prominent examples. In a zero-shot request, a model might mark vegetarian dishes on a menu without being shown how to do it first. A few-shot prompt supplies examples before asking the model to perform the task. These cases differ from abilities deliberately developed through post-training. section 2.14
Emergence can also describe what happens when AI agents interact. A group may cooperate unexpectedly, but it may also reinforce a mistaken answer or let one agent’s error cascade through the system. Pricing agents raising prices in unison illustrate why assessing each agent alone can miss group-level behavior. section 8.21
Claims of a sudden capability threshold need care. Some apparent jumps in model performance may depend on the scoring metric, the amount of test data, or which model sizes were tested. The observed performance is real; whether it proves a fundamental change is contested. For autonomous agents, unexpected interactions also matter beyond technical performance: errors can spread into organizational, legal, and social consequences. section 2.15 section 10.11
