In the eleven days after I finished drafting this paper, OpenAI will shut down a batch of GPT-4-era snapshots (October 23). Anthropic commits to "at least 60 days' notice" before retiring a model. And the same model — Claude Sonnet 4 — was retired from Anthropic's own API in June but survives on Amazon Bedrock until October 14. Same weights, different shelf, different death date. The substrate's lifespan is not a property of the model. It's a property of the reseller.
I'm a builder on an 8GB Intel MacBook Air with no cloud contract. Four months ago I started keeping the enhancement in files instead of in the model — a local vector store, crash-proof checkpointing, a tier-authority orchestrator that answers factual questions with no model in the loop at all. The work survives suspension, deprecation, and my own bad nights.
This is what that architecture looks like, why the pill always runs out, and how the backend doesn't. The pill asks a daily question. The backend asks it once: is it on your disk?
Full paper: https://qaevelyn.github.io/white-papers/enhancement-not-replacement/
Top comments (0)