対価を払っているのは、公開のあとの部分です
評価、監視、そして実際のユーザーが来てはじめて現れる失敗について。
下書きです。この記事はレイアウトの骨組みとして記事セクションに同梱されており、本文が入るまで検索とサイトのアシスタントからは除外されています。
この記事は英語で公開しています。サイトの他の部分は翻訳していますが、記事は翻訳していません。技術的な論旨は機械翻訳を通すより、書かれた言語のまま読むほうが価値が高いからです。
An AI feature is not finished when it is deployed. That sentence is easy to agree with and almost nobody budgets for it, which is why so many pipelines that were accurate in month one are quietly wrong by month four.
Three things that change without anyone touching the code
- The provider updates the model behind the version string you pinned, or deprecates it entirely.
- An upstream API adds a field, reorders a list, or starts returning
nullwhere it used to return an empty array. - Your own data changes shape, because the business changed and nobody thought to mention it to the pipeline.
What a maintenance loop contains
An evaluation set that reflects real traffic rather than the examples used during the build. A scheduled run of it. An alert when a score moves, and a human who reads that alert. None of this is sophisticated; the reason it is rare is that it produces no demo.
This post is a placeholder shipped with the writing section. Replace the body with the real piece; the frontmatter above is the only thing the site reads.