Emergent abilities

Wei and coauthors published "Emergent Abilities of Large Language Models" in 2022; Schaeffer, Miranda and Koyejo answered in 2023, in "Are Emergent Abilities of Large Language Models a Mirage?", that the jumps owe part of their sharpness to the choice of metric and soften when it changes. Emergent abilities are new behaviors or skills that appear in a large language model when it crosses a scale threshold, in number of parameters or volume of training. Nobody programmed them explicitly, and smaller models gave no way to predict them. Wei and coauthors, who popularized the term, show that certain skills absent from small models arise fairly suddenly beyond a critical size. For risk management the consequence is serious: a model's behavior cannot be fully deduced from its design, and scaling up can bring out abilities, and also failures, that nobody anticipated. The AlgoPolis paper relies on this unpredictability to place algorithmic risk under Knightian uncertainty rather than priceable risk: the execution environment changes faster than actuaries can collect stable data on it. The notion is still debated among researchers, and part of the observed suddenness comes from the metrics chosen to measure it. The practical unpredictability of large models remains.