Many Are Building Cathedrals on Quicksand

Short summary

AI vendor landscapes shift every 11-37 days with releases and deprecations, making tightly-coupled model-specific code obsolete within months. Teams that abstracted prompt formats, API calls, and model behaviors behind controllable interfaces survived seamlessly; those that hard-coded GPT-4-specific logic faced expensive rewrites. Key lesson: hard-code stable concepts (tokens, embeddings, RAG), externalize unstable ones (APIs, pricing, behaviors).

• •Frontier model releases occur every 11 days (2026) vs 37 days (2023), with 60-day deprecation windows
• •Teams that abstracted model-specific behaviors into configuration survived transitions; tight coupling required full rewrites
• •Distinguish stable (tokens, embeddings, RAG) from unstable (API params, pricing, behaviors) when architecting

Generated with AI, which can make mistakes.