Wei, Tay, et al. (2022) defined the term this way: an ability is emergent if it is not present in smaller models but is present in larger ones, so it cannot be predicted by extrapolating the performance of smaller models (2022, sec. 2). The typical plot shows near-chance performance over several orders of magnitude of compute and then a jump: on three-digit arithmetic, GPT-3 takes off only around 2 × 10²² FLOP, about 13 billion parameters (2022, sec. 3). The GPT-3 paper had already shown something similar: the 13-billion model adds two-digit numbers correctly only half the time; the 175-billion model, always (Brown et al., 2020, sec. 3.9.1). The authors frame the idea with Philip Anderson's (1972) motto: "more is different."
Schaeffer et al. (2023) argued that it is a mirage. Over 92% of BIG-Bench's emergent abilities appear under all-or-nothing metrics such as exact match; under continuous metrics, the same curves become smooth and predictable (2023, sec. 1). Arithmetic shows why: if a model gets each token right with probability p, it gets a 10-token answer right with probability p¹⁰.
p = 0.5 → 0.001 · p = 0.8 → 0.11 · p = 0.9 → 0.35 · p = 0.99 → 0.90
A gradual per-token improvement looks like a jump when measured on whole answers. Critics of the mirage view reply that nobody has any use for a sum with 80% of its digits right: the discontinuous metric may be precisely the one that matters.
The debate touches a loaded philosophical concept (emergence). Krakauer et al. (2026) distinguish emergent capabilities, which a complex system manifests ("more is different"), from emergent intelligence, which would consist in using those capabilities ever more economically ("less is more"). Is a jump in a plot a new property of the system or of our way of looking at it? The question also matters for safety: if abilities appear without warning, scaling laws are not enough to anticipate them.