Skip to the content.

Context

A feature needs to work in production now, but the efficient, dedicated version of it (a fine-tuned or classical model) needs training data or engineering that does not exist yet. A general-purpose LLM can do the task immediately, if expensively.

Example

Zeeguu estimates the difficulty level of every article with an LLM, an expensive call it makes constantly. This is also the bootstrapping case: those LLM-generated difficulty labels are accumulating as the training set for a cheaper classical classifier that will take the task over once enough have been gathered. The LLM ships the feature and earns its keep today, while quietly producing the data that will one day retire it.

A plainer instance without the bootstrapping twist: article topic classification runs on an LLM now, to be replaced by a dedicated topic-detection model once the taxonomy of available topics settles.

Problem

How can a feature ship and start earning its keep before the cheaper, dedicated version that will eventually run it has been built?

Forces

Solution

Use the LLM to perform a task in production while building a more efficient replacement. The arc is rent, then build: the LLM’s general capability is rented (paid per call, costly but available immediately) to ship the feature now, while a cheaper, dedicated implementation is built to eventually own the task. The rented LLM is convincing to users and good for early beta-testing and feedback, but intended to be temporary.

Bootstrapping variant: In the strongest form, the LLM generates the training data for its own replacement (as with the difficulty labels above): the rented capability directly produces the labeled data the dedicated successor is trained on, so running the stand-in funds and enables the build.

Consequences

Known Uses

These are distillation and self-instruct methods from the literature; we did not find a first-hand account of a production feature completing the full rent-then-build arc, and Zeeguu’s own replacement is still in progress.

Notes


💬 Open an issue about this pattern