Emergent Abilities

Emergent abilities are capabilities of large language models that appear abruptly and unpredictably when model scale crosses a threshold, rather than improving gradually with increased training. The term was popularized by Wei et al. (2022) who documented dozens of tasks on which smaller models perform at chance level while larger models suddenly achieve above-random accuracy. Examples include chain-of-thought reasoning, multi-step arithmetic, and translation between low-resource languages.

The concept is contested. Proponents argue that emergent abilities demonstrate a genuine phase transition in cognitive capability, suggesting that intelligence is a collective property of sufficiently large predictive systems rather than a feature that must be explicitly programmed. Critics counter that the appearance of emergence is an artifact of evaluation metrics — capabilities may improve smoothly in log-probability space while only becoming visible above a threshold when measured by discrete correctness. The debate hinges on whether emergence is an ontological property of the model or an epistemic property of the measurement.

The systems-theoretic significance of emergent abilities lies in their challenge to reductionism. If a system's capabilities cannot be predicted from the properties of its components, then understanding the components is insufficient for understanding the system. This is the defining feature of strong emergence — and if large language models instantiate it, then artificial intelligence has become an empirical laboratory for one of philosophy's oldest disputes.

_The question is not whether large language models exhibit emergent abilities. The question is whether 'emergence' names a real phenomenon or a failure of our measurement instruments — and the answer determines whether we are witnessing the birth of a new science or the refinement of an old illusion._