Survival LLM: AI assistant and emergency library that runs on your own computer

This topic is bringing out the hallucination-panic crowd.

It's perfectly safe to use LLMs to surface underlying data and knowledge without having to trust that it's giving you correct advice.

Use the LLM as search technology: string match -> semantic search -> search intelligence.

This is especially helpful when you don't know the right question to ask. Semantics only help you find text that is semantically near the input query. Semantics won't help you when what you are looking for is connected to what you have by something other than semantics, like causation, mechanism, or interpretation.

Trying to figure out why your crops aren't growing well will likely require you to read several articles across a range of sciences, probably scoped to the specific conditions you are dealing with. A semantic search is not going to surface those first.

Allowing an LLM to chain together several sources for you to read first does not require you to ask it for a conclusion.

Shoutout to project nomad.

https://www.projectnomad.us/

its got a ton of useful options, preconfigured, integrates pretty easily, and can use any local LLM or data.

I set this up on my machine and tested it with the following question:

How can I build a bow drill so I can start a fire?

It got it partially right: spindle, bow with string tied the correct way, and fire board. But then it flubbed by saying you put the spindle between both hands and move your hands back and forth to rotate the spindle and create the necessary friction… this totally ignores using the bow to rotate the spindle! It also mentioned “fanning the sparks” but there is absolutely no way your bow drill is going to create sparks.

I asked a frontier model to check the survival LLM’s conversation trace and it pointed out the errors.

I didn’t see obvious errors in the other 3 prompts I used, but this one seemed pretty egregious for a survival LLM.

I thank the creator for making this, but take it with a grain of salt.

Considering the tendency even now of LLMs to hallucinate, and the hardware requirements for any kind of LLM -why not just gather those resources into some sort of local archive that can be searched without an LLM?

This seems redundant and requiring more overhead while providing less reliability than something like Kiwix to me.

From what I can tell: - zero firearms or bowhunting content

- no books on hunting or fishing

- no instructions on field dressing animals

- no FM 7-8 or ATP 3-21.8

- no info on reloading ammunition or flint knapping

This feels more like a tool for city dwellers than a serious survival tool.

> Mac or PC with 16–31 GB | IBM Granite 4.2 8B | 14 of 22 right | 6 GB

So on a regular consumer machine this is going to be wrong nearly half the time. Great! Let’s hope it’s only on the simple stuff and not the “how do I perform an emergency appendectomy”.

Honestly, just buy some books.

this reminds me of something I built pre-LLMs that i called a "doomsday computer"

it was a raspi with a couple 18650s and bms taken from a drill battery combined with a pico-projector and a laser keyboard stuffed in a waterproof acrylic cube with a stack of sd cards filled with stuff like the survivor library and an offline copy of wikipedia.

complete with fold out solar panels and circuitry taken from one of those amazon solar phone chargers.

worked well for a few hours at a time at night while camping. eventually i was planning to add a retropi sd and a best of music sd (like classical and the rolling stone magazine top 500)

This downloads a library of mostly kiwix wikipedia information and produces a full text SQLite index that is live searched by the model and then the model reads and summarizes the information.

Interesting thing is that when I asked a follow up question it answered but showed: This answer came from what the AI already knows, not the library, so there's no source to check. Ask “look that up” to get one.

When I asked "look that up" it was able to. It highlights relevant sections.

The information it was working off in my case was WikiHow and there aren't any citations in that information.

Running on a 32GB Mac M1 using Qwen 27B. Ollama is showing as taking 19GB steady state and 23GB while answering on Activity Monitor (sometimes a little more and sometimes a little less).

Its amazing this can be done, but it seems like someone should be able to develop a much smaller model that can be used reliably for this use case. There are smaller models available to use and download but it is stated that they have some reliability issues.

I give it about 50/50 the internet goes down or becomes effectively unusable for weeks, maybe months, at some point in the next 5 years given abysmal commitment to cybersecurity from almost every company that exists.

Some kid with the right prompt will supply chain attack a dependency of a dependency of firmware for border routers just to watch the world burn, or runaway AI swarms will saturate all bandwidth. Or a Linux kernel zero-click vuln. So many ways this can happen.

Better to prepare a bit and never need it.

If every single example posted has hallucinations and the response is "it's your fault for not checking the sources" then it's obvious that an LLM isn't the right tool for the job. A search engine works perfectly fine for these kinds of datasets.

Cool project.

I have some property that doesn't really get cell reception in the summer, and even in the winter the reception is at-best spotty. I use granite and qwen together with a manually curated set of books, articles, etc. to get work done while disconnected. I tend to update the set of resources before leaving town because I'm generally intentionally disconnecting to focus on deep work. Building a workflow that makes that possible has helped me build out a setup that replaces a ton of my day-to-day anthropic/openai use even when I'm near fiber.

It is interesting how, even pre-llms, offline computers change the nature of a lot of activities. E.g., about once a year I'll walk do a random walk in the woods for a few days without paying much attention to maintaining a bearing. With maps and compasses it's a whole skill set to find your way back. With a phone and downloaded maps you just kinda press the power button and check your heading every hour or two.

The "survival" use-case is interesting. I fear a lot of the practical applications of that knowledge are either time-sensitive (e.g., RAG on a "Stop The Bleed" course won't help because by the time you learn the patient is dead) or require tools+skills that aren't easily improvised completely from scratch in high-stress situations (e.g., building shelter or even really simple stuff like building a fire when it's very cold out). And for other stuff common sense prevails (e.g., power out -> keep fridge closed).

The required specs seemed a bit high to me, and I was just looking for alternatives.

I found Project Nomad (https://projectnomad.us) which looks more polished. Any other suggestions?

Emergencies = No Internet

What about no power? You gonna run your generator for your local model? What are we doing here? Have we lost our way completely?

Even in their table it shows the LLM never getting all the questions right.

It's pretty cool. But if it's such a small model and small corpus, isn't it almost just as good just to have a static wikipedia-style copy of those key sources (WikiMed, MedlinePlus, Ready.gov, CDC etc.)? Maybe if they (and that could be static too), gave you a "what's going on?" question and then told you what you might not know - i.e. "power is out" -> "make sure you open your fridge as little as possible", "if you have backup battery ensure you turn power hungry devices off" etc.?

I'm not a doomsday prepper but man if you need to run an LLM locally during the apocalypse to figure out how long your food is going to last, I feel like you are in for a tough time.

This kind of stuff really feels like knowledge that probably shouldn't be offload to AI but rather learned and practiced. Again, I don't have a bunker in my backyard but seems like a weird use case.

This is much better waste of space than Apple AI. It was something I was imagining to build, got pleasantly surprised to see it having been done.

Living in Miami, this is the kind of thing I'd actually keep on a laptop for hurricane season, when power and cell service can be out for days.

Smallest model download is 6 GB. I'd rather store a collection of relevant ebooks.

All this is interesting, but I always had a question: in a state of disaster or some mishap even if you have an light LLM running - most important is a power source/electricity. Without it nothing will work.

Why should I download an 80 GB model that I need proper hardware for, when a couple MB of PDFs is all that's needed?

Reminds me of the time when Docker was thrown at everything.

Back in 2018 I was working on a project for educating schools in certain parts of Africa and Middle East, on AI. The issue was it had to be air-gapped (no internet guarantees) and small/portable. I ended up downloading Wikipedia (not all of it), and then created a lightweight inverted index (for search) and TextRank (for summaries). The whole thing was on a Raspberry Pi, and installed in a small plastic enclosure. We had many of these and could easily ship them. There was an "AI Learning Module" as well we built, that started with Scratch, and then moved into Python. But if I was to do it all over today, I would probably add a tiny LLM and maybe a small ASR and a TTS model too, and you basically have a completely self-contained AI knowledge assistant in a cigarette box. The point being, you basically also have a "survival AI" you can take with you. Of course you still need electricity...

212MB for MacOS and 25MB for Windows? Why are MacOS binaries so large? Is it typical for them to be 10x Windows binary size?

That said, I'm really looking forward to small and quick (latency wise) models are a reality. I can easily see a world where every keystroke, terminal command, etc. goes through a "autocorrect" type model. You could even blend natural language plus terminal program segments and the AI just infers what you intend. Either built-in or a second classifier model, determines if the autocorrect change needs human review first.

Great idea, I do a lot of work testing small LLM's in this space and agree if you add some guardrails and train a small embedder, you won't need these large models, as your info is static, you can ship it all as one.

Or try these with your trained embedder: https://ollama.com/library/minicpm-v4.6 https://ollama.com/library/bonsai-2

I went through a category 5 hurricane that took down everything. Prepared by downloading as much as I could ahead of time but it wasn’t enough. This, or something similar, would have been a blessing.

Love this. And as a trend, I'd love to see more local/small task-specific models/agents. Experts at helping you with that one thing, leaving the everything-and-the-kitchen-sink to the frontier models/agents

A detail from the benchmarks in that page is that Nemotron Lightning has good results for a MoE model close to Qwen 27B which is dense. Very interesting considering that Qwen 35B MoE stopped being in Alibaba's plans, and curious if Nemotron is equally capable in other bemchmarks.

I like to use llamafile for cases like this: https://github.com/mozilla-ai/llamafile

How are we powering this computer? Maybe be need a bootstrap llm that can answer only that question and runs on a solar device? Or some kinda clockwork bad boy, now you're talking :)

We need this, but running on smartphones.

Also, I am surprised that the difference between 32GB and 128GB in right answers is so small:

* Mac with 128 GB or more memory NVIDIA Nemotron 3.5 Super 21 of 22 right 80 GB

* Windows PC with 32 GB or more NVIDIA Nemotron 3.5 Lightning 19 of 22 right 25 GB

I made it a while ago now, but similar: https://github.com/x3haloed/eowkit

Helps you download a local LLM and offline Wikipedia. Probably easier to just ask your favorite agent to glue something together yourself these days.

I saw emergency and I immediately thought in case I couldn't access online docs or get my coding/platform questions answered day-to-day.

Apocalypse emergency I don't need help on. My crisis brain has collected more than enough info and capabilities over the decades.

Really cool idea! Will give it a spin.

I’m currently using https://kiwix.org for the offline access to Wikipedia, but your Survival LLM looks even better.

At what point would it be better to download a few real sources and do plain text search? Why risk a hallucination for content that would probably be smaller than the size of the model?!

this is more like .. survival harness, no?

>The app is the same everywhere. When it opens, it checks your computer's memory and downloads the AI model that fits. More memory means a smarter model.

first thing an emergency tool should require, internet access.

at first glance I figured it'd be an ultra small specialty trained survival model that'd be small and runnable anywhere, sorta like an emergency flavored version of those tiny tts models that have been all over recently.

Maybe that's an idea for someone.

This could be made a lot more practical (and safer) if made searchable via a simple vector search / embedding that can run on any ordinary computer.

The source (data collection) seems the most important feature here worth reusing.

What this needs is a recipe or script that you can point your local agent at and it builds a bespoke implementation that’s, by definition, highly resilient.

Edit: see project nomad.

Alternate idea: Download Unsloth Studio https://unsloth.ai/docs/new/studio#quickstart

From the command line: unsloth studio

Browse to the URL it gives you (eg http://127.0.0.1:8888/)

Download a good fast local model e.g. Gemma-4-12B (small machine) or Qwen3.6-35B-A3B (larger memory). Or a Nemotron model, if you prefer that.

Then you'll have both your backup/emergency assistant that runs on your own computer and be gaining skills in a common, frequently-updated, well-supported program.

Sorry if this is a dump question but I was wondering why not fine tuning each model to get 100% accuracy on those topics?

Wonder why they didn’t run Qwen 3.5 9b on 16th Macs. That’s a super impressive model for size. Is granite better?

Odd that it doesn’t run on phones. It isn’t that large, 212MB on a Mac. Phones are basically everyone’s EDC and most are quite powerful.

With all due respect: why does this need an LLM, and this much data? Surely a "classic" knowledge base with a more traditional search bar, running on top of the same data this LLM was trained on would be way smaller, and actually as trustworthy as the source material it surfaces for you?

The only way Big AI labs will sell us appliances is when the internet was somehow destroyed.

Not sure where this would fit in.

pretty cool I like it. could you support more (small) models in the future, e.g. from the qwen or gemma family?

Nice idea! I wonder how the installed IBM Granite 4.2 8B LLM does on machines with "modest" RAM

I am not going to accuse anyone of anything, but this looks suspiciously like the project I have been working on. The dates, the time, the folder structure, the concept. It is a bit disheartening, because it missed the whole point, and it butchered the implementation. No LLM makes sense. Having an LLM in the loop is having the worst of both worlds? Example: My initial use case was to provide schools and traders in rural Africa with at least some sort of free LLM-like capability. They have phones that cost $5. Power cuts, drop of the cellular network. Not just 'No-LLM to prove a point'.

It is also only 5% of the wider scope and I was about to release it under OSS 2 days ago, but wanted add some final finess. Now that this is out, I might as well release it so the concept actually ties to its original goal, and release the scientific papers that date back to Feb. Please ping me on github. Same username as here. Always could do with help.

When I quit the app, why is it still running in the background?

On the setup process: "You only click for about 5 minutes; the rest is downloading." On the energy consumption: "A small generator or power station runs it easily"

Five minutes of clicking and a (small) power station is a small price to pay when survival could be at stake.

Can this work with ollama?

Seems like a remarkably stupid idea as opposed to just having an extensive encyclopedia downloaded with a very thin client you can use to read from it.

In an actual survival scenario power isn't guaranteed and errors of any kind can be life or death. So you need to minimize the resources used to get the thing working.

From the site:

>Survival LLM >It Can be wrong

thank you!

Have you considered just having a library?

LLM-authored website convinces you to waste time and energy on chatting with an LLM when struggling to survive.

I'd want an Android version...

Bold of you to assume there will be running computers past the nuclear apocalypse. It is much, much more likely that a lot of electronics will be permanently fried by the inevitable EMPs. Distributing practical skills in how to survive is much more useful.

I recently evaluated LLMs for use for offline navigation, with this very kind of survival scenario in mind: https://fig.blue/posts/directions-from-memory/

tl;dr: Local LLMs are really really bad at offline navigation.

That said, pairing the survival guide with a library is totally the right move.

LLMs have become the software equivalent of using a hammer to solve all your problems. Like others have said, there are much better solutions to provide an enhanced search engine for your survival knowledge.

it would be useful if it would shrink a little bit

It's disappointing that Gemma E4B seems to perform poorly on this (42% accuracy according to the GitHub). It would be nice if a $25 phone could be loaded up with this and stuffed in a EMP proof metal box with a small solar charger. Maybe a finetune or Gemma 5 will get us there.

"Survival LLM, the power is out and there's no food, how do I turn your computer on to ask your questions?"

Just few days ago I thought digital preppers are soon gonna make something like that. It's a generally good idea.

I'd put in this also a ton of software documentation, minimally for Rust.

And a bunch of scientific literature.

I can almost guarantee you that by the time you need a survival assistant, you won't be willing to spend gasoline on generating electricity for your GPUs.

(It's helpful for shorter outages. But "survival assistant", I think not, unless you run a compound)

Talos Principle vibes.

This will come in handy after AI takes over the world.

But kidding aside, this is an interesting reveal on the ease with which a custom LLM can be spun up to fit a particular niche application, especially for systems not meant to host an LLM.

If you're actually offline and need information in a 'survival' situation, I'm not so sure the hallucinatory nature of an LLM is the based way to distribute a repository of information. Additionally, there's no need to fit an LLM entirely in memory.

In a theoretical scenario where the internet is gone and isn't coming back: If you can deal with a very low token/second rate there's no reason not to store 'offline' a much larger model like qwen 3.8-flash-next or deepseek v4 flash 1017 and use an inference engine which can stream-read it off disk to extract the information you want, albeit slowly. Of course also download and use things that'll fit in RAM, sure, like the qwen 3.8 27B in the linked page. But no need to restrict yourself to only smaller models.

if your computer doesnt work during an emergency because power went out, how will your LLM work?

No GNU/Linux? Meh.

How on earth are you going to operate this if you have no electricity?

This is just crap. Why would you want something that can give you wrong answers (like 4 out of 20 like in the tables, that is much!) - in an emergency?

Please, please do not trust a chatbot in emergencies. There’s a reason why actual survival and emergency preparedness guides include manuals rather than recommend you just use a chatbot: practicality.

Just look at those specs needed for accurate-ish results. You really think your 128GB Mac Studio or Framework desktop is going to be able to work in a real emergency situation? Even if it is, how long will it last? What about those precious seconds waiting for inference?

The danger here is less hallucinations and more, well, everything else about it. Instead of setting this up, take the time to make sure you have a restocked first aid kit and enough water for a week of use by your household. Stock up on meds and non-perishable food. Plan exit strategies and contingencies. Acquire and store little first aid booklets or manuals of common survival tasks in physical form (or hell, throw onto an eReader - infinitely more battery life than an LLM host in an emergency).

This is one of those “sounds good in my head” projects that folks with emergency experience would bristle at. This is a hindrance, not a helper.

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[flagged]

[dead]

[flagged]

[flagged]

[dead]

[dead]

[dead]

No thanks, I'll use a generic model along with trusted sources.

Also: "Apple could not verify “Survival LLM” is free of malware that may harm your Mac or compromise your privacy." This is obviously some vibe coded trash from an inexperienced dev.