The workshop frame
The most useful split from the session is not local versus cloud. It is vague versus bounded. When the question is still broad, exploratory, or poorly framed, a frontier model is the better instrument. It has the width to tolerate ambiguity. Local AI becomes useful after the work has been reduced into a narrower operating lane.
That lane is created by the SOP. A model should not be asked to own the whole workflow. It should own the fuzzy part: the natural-language request, the messy note, the uncertain input, the image or audio extraction, the summarization step, or the private coding context.
The core note is simple: the model is not magic glue. It is one component inside a workflow that has to be explicit enough to inspect.
What changes when it runs locally
Local AI makes the privacy argument stronger, but it also makes the operating constraints more honest. A model that barely fits into memory is not practical just because it technically loads. If the operating system, browser, and tools have no headroom, the system will fall into swap and the experience will be slow enough to fail in daily use.
This is why smaller biased models matter. The target is not the biggest model a machine can technically run. The target is the smallest model that can do the bounded job with acceptable behavior. Adapters, post-training, and quantization are practical tools for that job because they move the model toward the task without forcing every workflow through an oversized base model.
A chat box is not an operational system. A harness gives the model controlled access to files, search, commands, calendars, APIs, or MCP tools. That power only works if the model has clear permissions and narrow action boundaries.