Developers increasingly reach for OpenAI or Anthropic APIs, but the practice ties apps to fragile cloud services. When a provider’s server stalls or a credit card lapses, the feature disappears. The author argues that this dependence erodes privacy and inflates costs, turning a simple UX tweak into a distributed system that can fail at any moment, and forces users to reinstall or lose core functionality.

A concrete counterexample comes from the Brutalist Report, a news‑aggregator that now ships an iOS client capable of summarizing articles entirely on‑device. Using Apple’s SystemLanguageModel APIs, the app processes up to 10 KB text chunks, generates concise bullet points, and returns a typed Swift struct—no network round‑trip, no user logs, and no vendor account required, and demonstrates how privacy can be baked in.

The piece stresses that local AI excels when the task is data transformation—summarizing, classifying, extracting—rather than answering open‑ended questions. By keeping inference on the handset, developers avoid latency, rate‑limit headaches, and compliance burdens tied to cloud providers. For most app features, the trade‑off favors on‑device models, making privacy‑preserving intelligence the default rather than an afterthought, especially for regulated industries handling sensitive data.