Your Fine-Tuned Model Is Tech Debt: A 50x ROI House of Cards — Dan Bjornn, Lease End
Source
youtube.com
Author
AI Engineer
Date
Context
A short training run can conceal a slow repair process. In Dan Bjornn's case study, fixing a customer-intent classifier took about a week because examples had to be gathered, expanded, checked and categorized; training itself took roughly an hour. He reports replacing that system with prompts, skills and context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → that could be updated as files, bringing fixes below an hour. Although each message became more expensive, he says total cost fell. His account makes maintenance effort part of the cost comparison: cheaper model calls can coexist with expensive corrections.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.