Research Notes

The Emergent Abilities Debate

By Satwik · March 11, 2026

"Emergent abilities" named a striking observation: on certain tasks, model performance stayed near chance as scale increased, then jumped sharply past some threshold, rather than improving smoothly. Reported across models like GPT-3 and PaLM, this framing suggested capabilities could appear suddenly and unpredictably as models grew, which is unsettling if you are trying to forecast and contain what a bigger model will be able to do.

The idea had real safety weight. If dangerous capabilities can switch on abruptly at scale, then evaluating a smaller proxy tells you little about the next model, and pre-deployment red-teaming faces a discontinuity problem. Emergence became a load-bearing concept in arguments for caution about frontier training runs.

The debate is genuinely open, and this is where researcher discipline matters. A prominent counterargument holds that much apparent emergence is an artifact of the metric: sharp, discontinuous, all-or-nothing scoring (like exact-match accuracy) can manufacture apparent jumps even when the underlying model capability improves smoothly, and that switching to continuous or partial-credit metrics makes many "emergent" curves look gradual. Under that view, emergence is at least partly in the measurement, not purely in the model.

For our purposes both readings carry a lesson. If emergence is real, forecasting frontier risk from smaller models is unreliable and we should assume unpleasant surprises. If it is largely a metric artifact, then our evaluation choices can hide or fabricate capability transitions, and benchmark design becomes a safety-critical activity in its own right. Either way, the practical mandate is the same: build evaluations with graded, sensitive metrics, and do not assume that a capability absent at one scale will stay absent at the next.