The phenomenon: a skill that switches on
Plot a particular skill — say, three-digit addition, or following a tricky instruction — against model size. For small models the score sits near zero, flat and useless. Then, past some size, it shoots up. Researchers named this an emergent ability: a capability essentially absent in smaller models that appears, seemingly abruptly, in larger ones. The word borrows from physics and biology, where a system as a whole does things its parts cannot.
The most famous examples are abilities the models were never explicitly trained for. In-context learning — picking up a new task from a few examples in the prompt, with no weight updates, also called few-shot prompting — strengthens sharply with scale. So does chain-of-thought reasoning, where asking a big enough model to "think step by step" suddenly helps, while on a small model the same trick does nothing.
Why it feels almost magical
Emergence is unsettling because scaling laws had promised smoothness. The loss curve is a tidy power law with no kinks — yet here is a downstream skill that looks like a switch flipping. If the underlying training loss glides down continuously, where does the sudden cliff in the skill come from? These emergent reasoning skills are exactly the ones that make large models commercially exciting and scientifically puzzling at the same time.
The 'mirage' counter-argument
In 2023 a Stanford group lit the emergence debate by arguing that much of the apparent emergence is an artifact of how we measure, not of the model. Their key observation: the sharp jumps tend to show up under harsh, all-or-nothing metrics like exact-match accuracy. Get a 30-digit answer right except for one digit and you score zero — until, past some size, you finally nail every digit and jump to one.
A harsh exact-match metric multiplies per-token probabilities, so a smoothly rising p turns into a sudden cliff — the heart of the mirage argument.
Loss is smooth; the scoreboard need not be
The cleanest way to hold both sides at once is the distinction between downstream-versus-loss scaling. The pretraining loss improves smoothly — that is not in dispute. What can look abrupt is a downstream benchmark score, because the function turning loss into a benchmark number can be steep or thresholded. Emergence, on this view, is real as an experience for the user but not a discontinuity in the model's underlying competence.
Pretraining loss follows a power law in model size N — it keeps improving smoothly, which no one disputes.
Two practical lessons fall out. First, beware benchmark saturation and brittle metrics — a flat-then-jumping curve may say more about your ruler than your model. Second, smooth metrics are friends of planning: they let you forecast capability instead of being ambushed by it. The next guide builds directly on this, turning curves into forecasts you can bet a budget on.