The working assumption in most household preparedness guides is that digital tools fail when the grid fails. No internet, no AI assistant. No cloud, no answers. A report this week from Hacker News on PrismML's Bonsai 27B model challenges that assumption directly: the model runs at the 27-billion-parameter scale entirely on consumer phone hardware, no server connection required.
That's a meaningful shift — not because AI is magic, but because the information gap during emergencies is a documented and underappreciated problem.
What's actually changing
For years, running a large language model required either a cloud API or a workstation-class machine. Smaller on-device models existed — 3B and 7B parameter versions — but they degraded noticeably on complex queries: multi-step first aid reasoning, dose calculations, structural triage questions. A 27B-class model running locally is a different category of capability.
What this means practically: a phone with Bonsai 27B (or models like it) loaded before an emergency could serve as an offline reference system for medical, mechanical, and logistical questions when cell towers are saturated or down. Hurricane Katrina-era after-action reports and more recent assessments of wildfire evacuations consistently identify a single failure point — people couldn't get reliable, specific information fast enough. Not shelter locations. Not which roads were passable. Not whether a symptom warranted evacuation or shelter-in-place.
A capable offline model doesn't solve the coordination problem. It doesn't tell you where FEMA trucks are. But it can answer "what are the signs of carbon monoxide poisoning" or "how do I purify water with bleach" or "what's the correct dose of diphenhydramine for a 40-pound child" without a cell signal, without waiting for a callback, without hoping the Wikipedia page cached correctly.
That's not catastrophism. That's filling a documented gap.
What we'd actually do
Download and test at least one serious on-device AI tool before you need it. The Bonsai 27B model is new, but the on-device AI ecosystem has been maturing for 18 months. Apps like Ollama (desktop), LM Studio, and PocketPal AI (mobile) let you pull and run open-weight models locally. Spend 30 minutes this week setting one up on a household device — not because the grid is going down tomorrow, but because learning a tool in calm conditions means you'll actually use it correctly under stress. Confirm the model files are stored locally, not streamed.
Build a dedicated "offline library" device and keep it charged. A mid-range Android phone or a retired iPad with 256GB of storage can hold a capable model plus offline maps (Maps.me or Organic Maps), the full Wikipedia snapshot (about 20GB compressed for the English text-only version), and a PDF library of medical and mechanical references. This device lives on a charging pad and costs nothing to maintain beyond the hardware you may already own. It becomes your household's offline reference stack.
Stress-test your information setup with a 4-hour no-connectivity drill. Turn off Wi-Fi and cellular on your primary devices for an afternoon and try to answer five realistic emergency questions using only local tools. You will find gaps. That's the point. The gaps you find in a drill are fixable; the gaps you find during an actual outage are not.
Be honest about model limitations before you rely on one. On-device models at any parameter count can hallucinate. They can give confident wrong answers about drug interactions, structural engineering, or local geography. The correct frame is "better than nothing, not better than a professional." Use them the way you'd use a field first-aid manual: as a starting point that requires judgment, not as a final authority. Train your household on that distinction explicitly.
The bigger picture
The preparedness community has always had a knowledge problem. Paper manuals are heavy and become outdated. Online resources disappear behind paywalls or go dark when you need them. For the first time, a reasonably capable generalist reasoning tool can live entirely on hardware most families already own, available when the networks are not.
That doesn't mean AI solves emergencies. It means one more piece of the information infrastructure can be made local and resilient. The goal — as always — is reducing your household's dependence on any single system that can fail. Power, water, food, information. Each one benefits from redundancy built in calm times.
This week's announcement is a data point, not a revolution. But it's a data point worth acting on.
