How much intelligence can run locally, at the edge, before the cloud is needed at all?

Small models keep getting better faster than large models keep getting cheaper, and the gap between “runs on my machine” and “good enough for the task” narrows every month. The question is where the line actually sits — per task, not in general.

Worth mapping: which tasks degrade gracefully at small scale (routing, extraction, summarisation, triage) and which collapse (deep reasoning, long-context synthesis)? And how far does a hybrid get you — local first, cloud only for escalation?

There’s a systems angle too: latency, privacy and cost all change character when inference moves from a datacenter to the device. The experiment is to find the smallest model that still clears the bar for a real workload, and to measure what that buys.