Emergent ability
A skill that shows up suddenly only once a model gets big enough, absent in smaller ones.
Think of it like
Like water staying liquid, liquid, liquid — then abruptly boiling at 100°C. Nothing, nothing, then a phase change.
Example
Small models flunk multi-step arithmetic entirely, then past a certain scale a model starts getting it right, seemingly out of nowhere.
How it actually works
Emergent abilities are capabilities that don’t improve smoothly with size but appear to switch on past a threshold. They’re real but contested — some “jumps” are artifacts of harsh all-or-nothing metrics, and smoother scoring reveals gradual gains. Either way, they’re why bigger models sometimes surprise their own builders.
For product teams
Why a scale-up can unlock features you didn’t explicitly train for — and why they’re hard to predict.
For engineers
Capabilities that appear near a scale threshold rather than improving continuously; partly metric-dependent.
Related
- Capacity — The parameter growth that triggers it.
- In-context learning — A key one: learning from examples in the prompt.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome