i watched one of those videos about running small local ai models and could not figure out what i would actually do with one. so i asked for the conceptual layer instead of a product list, and the framing finally clicked. a cloud model is a consultant. you call it for hard problems, it bills you, it's smart. a small local model is the machine in the back room that does one repetitive function ten thousand times with no supervision and no bill. the mistake everyone makes is asking the small one to think. the win is asking it to work.
a task belongs in the back room when it checks four boxes. narrow: one function, known input, bounded output. repetitive: it runs often enough that privacy and cost become real. low stakes: a wrong answer costs nothing. self contained: the answer sits fully inside the input, no web lookups, no tool chains. the archetypes fall out of that filter. the classifier that sorts and tags whatever flows through your pipes. the extractor that pulls fields out of messy text. the summarizer that compresses logs and threads. the writer for low stakes copy like descriptions and captions. the router that decides which expensive pipeline something deserves. the embedded brain that ships inside a product where you cannot put an api key. and the fallback, a dumb but present brain that still works when the network dies. an intelligent battery backup.
then i ran the filter over my own stack and the best targets are embarrassingly obvious. session mining and topic extraction over old chats. wiki capture summaries. the anonymize and rewrite pass before anything private gets published. weekly review digests. all extract and classify and summarize work, all reading my most private text, all currently shipped to a cloud model. the pattern to remember: expensive model for judgment, local model for volume. every flow has a volume half and a judgment half. cut the volume half first. the one exception, stated plainly: anything where a wrong output becomes a wrong action, a ticket or a money decision, stays on cloud until small models prove otherwise.