Definition
Plain language
Paying expensive setup work once so every later query becomes cheap.
As stated in the literature
The pattern of training a system to produce good answers in a single forward pass, replacing per-instance optimization with up-front training cost amortized across many queries.
Also called: amortization
Why it matters: It's the difference between systems that are too slow to deploy and ones that can answer queries instantly because the hard work was done at training time.
For example, instead of running a search every time you want a protein structure, AlphaFold trained once so each new prediction is a single fast forward pass.
Heard on the show
“That second approach is called amortized inference.”Episode 013 — Why Search Keeps Rediscovering the Same Workflow, and What That Means