Concept
Emergent capabilities
- Programmers
- Writers and journalists
Abilities that appear in bigger AI models without anyone programming them, sometimes abruptly and hard to predict in advance.
Nobody teaches a language model arithmetic, translation or programming one skill at a time. Its makers train it to predict text, make it bigger, and find out afterwards what it can do. A 2022 paper by researchers at Google and elsewhere called an ability “emergent” if it is “not present in smaller models but is present in larger models”, meaning it couldn’t be predicted by extrapolating from smaller models.
How sudden these jumps really are is debated. A 2023 paper argued that many of them come from how the tests are scored: with all-or-nothing scoring, improvement looks abrupt, while with measures that give partial credit it is often smooth and predictable. Both sides agree on the part that matters for safety: new abilities come from scale rather than design, so a model’s makers learn what it can do by testing it after it is built.
That’s why testing before release matters, and why models keep surprising even their own developers. The AI Safety Atlas notes that AI capabilities “have consistently surprised experts for over a decade”.
