Definition
Plain language
A pattern where one expensive model proposes a few high-quality examples and a cheaper model uses them as templates to generate many more.
As stated in the literature
A data-generation technique in which a strong agentic model produces a small set of high-quality seed tasks or trajectories, then a cheaper model uses them as in-context exemplars to generate many more examples at scale.
Also called: seed-and-amplify
Why it matters: It lets teams get the quality of an expensive frontier model at the scale of a cheap one, which is often the difference between an affordable training dataset and an unaffordable one.
For example, a top-tier model might write fifty carefully-crafted research questions, and a cheaper model uses those as templates to generate ten thousand more in the same style.
Heard on the show
“They call it propose-and-amplify.”Episode 017 — When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers