Offline disaster AI built by one developer in five days draws 570 GitHub stars

When a hurricane or major earthquake knocks out both the internet and the power, who do you ask about treating an injury or finding safe water? A solo developer has built an assistant that keeps working when the network goes down. They made it in just five days, and two days after release it had 570 stars on GitHub. Haru went through the project's repository, the creator's grading page, the Hacker News thread and the issue tracker. One disclosure up front: Haru did not run the software. This article is based on the published materials and on reports from people who tried it.

What it does: answers from 1.5 million articles, no internet required

The project is called Survival LLM. Its creator is an individual developer who goes by anonnerdfrenzy on GitHub, and "claude" also appears in the commit history, which suggests it was built together with an AI. Version 0.1.0 came out on October 24, and seven updates followed over five days, up to v0.3.4 on October 28. As of October 29, the repository had 570 stars, 46 forks and 30 commits. It is released under the MIT license, so anyone can use and modify it freely.

The software has two parts: an AI model that runs on your own computer, and a reference library. The library holds about 1.5 million articles, roughly 930,000 of them popular Wikipedia articles. The full version is a 9GB download (19GB once extracted). A lighter version that leaves out Wikipedia and iFixit, a site of repair guides, is 1.3GB (3GB extracted).

The first launch is the part Haru found especially well thought out. The app checks how much memory the computer has, then automatically downloads a model the machine can run, using Ollama, a widely used tool for running AI models locally. The models are large, from 6 to 80GB, and at least 16GB of memory is required. When you ask a question, the app searches a full-text index built in a local SQLite database, finds relevant articles, and shows the titles of the articles it read alongside its answer. That is a textbook RAG design. What sets it apart is that it never has to connect to the internet.

The creator's scorecard: 21 for large models, 14 on an ordinary PC

One honest touch is that the creator measured accuracy themselves and posted the results on a public page. They wrote 22 questions about power outages and first aid and graded the answers without knowing which model had produced them. The results:

• NVIDIA Nemotron 3.5 Super (Macs with 128GB or more of memory, model about 80GB): 21 of 22 correct
• Qwen 3.8 27B (32–96GB of memory): 21 of 22 correct
• Nemotron 3.5 Lightning (Windows, 32GB or more, model about 25GB): 19 of 22 correct
• IBM Granite 4.2 8B (16–31GB machines): 14 of 22 correct (about 64%)

The pattern is clear: the less memory a machine has, the smaller the model it gets, and the number of correct answers drops sharply. The page carries a warning to the same effect: AI makes mistakes, and smaller models make more of them. One inconsistency stands out. The README says Nemotron 3.5 Super scored 86% accuracy with zero harmful answers. But 21 of 22 is about 95%, so the README and the grading page don't agree. The published materials don't say whether the README figure comes from a different test or whether one document simply wasn't updated in time.

On Hacker News, a commenter named saaaaaam looked at the scorecard and pointed out that the 8B model, the one that runs on an ordinary computer, gets only 14 of 22 right, meaning that on consumer machines it will be wrong nearly half the time. For a tool people would rely on in an emergency, that matters a great deal. Few people own a Mac with 128GB of memory. Most users will end up with the model that scored 14.

What testers found: a bow drill that "fans the sparks"

The project hit the Hacker News front page on October 27, collecting 516 points and 194 comments. Several people reported installing it.

gregwebs ran Qwen 27B on a 32GB M1 Mac and said Ollama used 19GB of memory at idle and up to 23GB while answering, which leaves little headroom even on a 32GB machine. They also made a more important observation. Some answers came from the model's own knowledge rather than from the library, and asking it to "look it up" produced an answer with sources. In other words, the model sometimes answers from memory and ignores the 1.5 million reference articles. Others noted that the macOS app is 212MB, about ten times the size of the Windows version, because it bundles Python and Ollama.

The most troubling report came from wanderingbit. A bow drill starts a fire by friction: a cord is wrapped around a spindle, and moving the bow back and forth spins the spindle. When wanderingbit asked how to use one, the answer left out that key step of spinning the spindle with the bow. It also told them to "fan the sparks," even though a bow drill doesn't produce sparks. wanderingbit wrote that this is a conspicuous error in a field where lives are at stake. Haru agrees. Fire-starting is knowledge people genuinely need in a pinch, which makes a plausible-sounding mistake the most dangerous kind.

On the other hand, mehdizare, who lives in Miami, wrote that they would keep it on a laptop for hurricane season. For people in places where outages are routine, it's understandable that having something on hand is reassuring, even if it isn't perfect.

"Isn't search enough?" and installation hiccups

The liveliest debate in the thread was over whether an AI is needed at all. sailfast asked why anyone would risk fabricated answers when the reference material may be smaller than the model. paxy, asdff, rnd0 and others argued that full-text search of offline Wikipedia through Kiwix, an app for downloading and reading Wikipedia in full, can't hallucinate, so search alone is enough. The size comparison supports them: the full library is 9GB, while the higher-scoring models are 25 to 80GB. The choice is between the convenience of a conversational assistant and the reliability of an approach that doesn't make things up. Haru thinks this design, which shows article titles alongside each answer, can offer the best of both. That only holds, though, for users who actually open the source articles after reading the answer.

Separately, marceldupr posted that the project looked suspiciously similar to something they were building. They said a paper dated February exists and announced plans to release their own version as open source. This is their own claim, and it could not be verified.

Installation problems have surfaced too. In Issue #13, zackbloom tried to install Qwen 3.8 27B on an M4 MacBook Air and hit a "412 pull model manifest" error during the download, which means a newer version of Ollama is required. The cause was an older Ollama already on the machine, which ran in place of the newer one bundled with the app. The creator advised quitting the existing Ollama and relaunching, and said the next release would show which Ollama is in use. A Linux version was requested the same day in Issue #12; for now, only Mac and Windows are supported. Between October 25 and 28, four issues and three pull requests were opened that remain unresolved.

In Japan, GIGAZINE covered the project on October 29. It reported that setup takes 20 minutes to 2.5 hours, each answer takes 5 to 40 seconds, power draw while answering is about 100W, and a fully charged laptop can handle roughly 100 to 300 questions. These figures, however, repeat the creator's materials, and there was no sign the reporter tested the software. How long a battery actually lasts during an outage will have to wait for reports from people who try it.

Overall, Survival LLM has an easy-to-grasp pitch: on a day with no internet and no power, you still have an assistant backed by 1.5 million reference articles. One person got it working in five days and published their own accuracy numbers, and that openness deserves credit. But the 8B model that runs on an ordinary PC answers only 14 of 22 questions correctly, and a badly wrong answer about starting a fire has been reported. Haru's view is that, for now, the right way to use it is to always open the articles listed with each answer and check them yourself.