Scaling laws promise smooth improvement: more compute and data lower the loss along a predictable curve.
Benchmark scores did not always look smooth, and that mismatch is what emergence describes.
On some tasks a small model scored at chance, and a model a few times larger scored well above it.
Three-digit addition, unfamiliar instruction formats and multi-step word problems were the standard examples.
The claim drew attention because it implied capabilities could not be forecast before training a model.
That is genuinely uncomfortable for anyone trying to predict what the next model will be able to do.
Later work argued the sharpness is often an artefact of how the tasks are scored.
Exact-match scoring is all or nothing: an answer that is one digit wrong earns exactly zero.
A model steadily improving from very poor to nearly right therefore scores zero the entire way.
Swap to a continuous measure, such as the probability assigned to the correct answer, and the curve is smooth.
Under that reading the ability was building gradually and only the scoreboard moved suddenly.
Not every case dissolves this way, and the debate is not settled.
The useful takeaway is to be suspicious of a sharp jump on a discrete metric.
Plot a continuous measure alongside it before concluding that something new appeared.
It also means small-model evaluations can seriously understate what a larger model will manage.