IMO, the real question isn't how much momentary disruption it can cause to the infrastructure, but how hard it would be for emergency services to quickly restore it in a basic "offline" capacity. Like, power and logistics functioned back in 1970s, and Ukraine shows that essential infrastructure can be kept on life support even if someone is physically destroying it with warheads. But a lot of advice here seems to assume, indeed, literal Mad Max levels of civilizational failure, where there's a real threat of dying of starvation and exposure all alone. This doesn't really seem threat-appropriate...?
Or am I failing to appreciate the range of possibilities here, the fragility of all these systems? But again, it doesn't seem that cities in states that are literally under heavy physical bombardment are reduced to that, as long as the city in question isn't literally on the frontline.
Water-sewer systems are extremely vulnerable and fragmented: https://www.thefai.org/posts/america-s-sewer-systems-are-surprisingly-vulnerable-to-cyber-attacks
I think we can assume a "hackening" can do things at least as bad as the recent Outlook outage and FAA shutdown. Note that I'm not saying these were (or weren't) related to AI cyberattacks, merely that this is a baseline level of severity we should expect regularly.
Brangus has built Survival LLM, a local LLM and dataset that can answer-survival related questions like "how do I purify water in an emergency?". If you have money to burn you can buy it preinstalled on an airgapped, faraday-caged laptop with a solar power kit.
I am exactly the kind of nerd this thing appeals to, but I'm a little confused about when I would use it. Seems like i'll have pretty limited tools during an actual emergency and the more important thing is to figure out what I'll need ahead of time.
One could probably get partway there for free/cheap by downloading, say, this library of homesteading/survival PDFs - especially onto an e-reader - or spending half an hour on ebay ordering cheap used first aid, survival, etc books.
Broadly, sure - like, I don't know the quality of every book in it, but at a glance: "Where There Is No Doctor" and "Where There Is No Dentist" are commonly-used and referred-to books in the field of low-resource medicine, the Foxfire series is well-regarded scholarship about practical survival and craft traditions, there are US Military and Red Cross texts in here. It looks like whoever assembled it put thought in.
I'm a bit skeptical of the value-add of the small local LLM, here. It seems like stripping the LLM out and just rigging up something along the lines of RAG[1] on a bunch of popular survival manuals would get you a faster, more reliable, and leaner system without losing any of the usefulness. This is especially true given that open LLMs are optimized for parameter efficiency, meaning data that doesn't relate to math, programming, or other common benchmarks/usecases isn't really emphasized as much in the training sets.
Without the 'G'. Basically a local search engine that uses whatever the most popular RAG setup uses to map your questions to sections in the text likely to answer them.
FWIW, I think RAG - G = IR
darkbook.io doesn't say what the laptop's hardware is...
FWIW this is also extremely appealing to me conceptually, but on quick sober assessment doesn't seem very threat-appropriate. Like, you're not going to be trapped in the wilderness cut off from civilization, and there isn't going to be an immediate mass die-off such that a lot of essential knowledge is going to be lost. (I think?) It's not unhelpful to have, but doesn't seem particularly pivotal.
Couldn't you just put any laptop with local model weights downloaded in a trash can?
https://thehackening.vercel.app/
I took a stab to first draft to consolidate all the threads/posts on The Hackening into a public-facing site with some step-by-steps on security hardening (focused on high impact/low cost actions first), as well as info on threat models. I also added an incidents page as a newsfeed for relevant and developing events, though there is no automation to keep this current yet.
This was inspired by Raemon's call to action in the comments here. Since I trusted Fable 5.2 with a lot of the information gathering and guides, I will still need to go through and check some specifics. If anyone wants to help with the site, please let me know. Here's what I think could add the most value:
1. Review of guides, especially from anyone with a background in security.
2. A more detailed write-up of the threat models.
3. Anything I missed for threat models or recommended actions (I know there's a lot of prepping/speculative info that was left out).
4. Review of guides as you work through them. Any roadblocks that we need to preempt in the guides? Corrections or better sources?
5. General site updates/bug fixes/design/copy improvements (Started a task list of these on github).
6. Any other good ideas to add to the site, of course.
I think it's good to have offline maps and addresses, so you can still find your way to both: a) people you care about and would want to find and help in an emergency, e.g. family members; and b) people you might be able to get help from.
Folks who don't have much local trusted community may want to take down the address of the nearest LDS church, so as to be able to attempt to volunteer with them in an extended shutdown.
(It doesn't seem like we're close to really being able to make great use of it but I just snagged www.thehackening.ai if someone wants to run a project making a good general resource site about it)
Put this together when Ukraine/Russia kicked off, think it holds up pretty well: https://www.lesswrong.com/posts/dZrqheHkQqRMoxbYk/an-effective-grab-bag
I can recommend Kiwix, an offline reader. On Arch, its installation is simply:
pacman -S kiwix-desktop
and then inside the app, with a few clicks you can download various useful things like maps, Project Gutenberg books (eg. science, technology, medicine among others), the entire English Wikipedia with or without pictures, etc...
Boring, standard advice that is good regardless applies doubly now (approaching this mostly from an electronics perspective). Depending on your threat model, some of it will be more or less useful, but none of it is a bad idea in general:
Backups: Keep 2 backups of all important data, of which one is airgapped and one is off-site (you can also have one that is both and one that is neither). Check that you can restore these (without the help of LLMs, ideally). Include copies of legal documents, such as insurance information, ownership records, birth and marriage certificates.
Cryptography: Use a password manager, ideally with hardware two-factor authentication (e.g. a YubiKey). Then you just need to memorize one strong master password. Use the password manager to generate and store strong passwords for all other services. Rotate the weak passwords that might be left. Encrypt your devices and your backups, and you may choose to memorize those passwords instead of or in addition to storing them in your password manager. Keeping physical break-glass copies of important passwords and recovery codes (stored e.g. in a safe) is also a good idea.
Updates: Run updates when offered to you by your phone and computer. If you are responsible for servers, keep those up to date too. Out-of-date software is generally the higher risk, not supply chain attacks.
Self-sufficiency: Plan for surviving up to two weeks without grid electricity. Get a large powerbank with an inverter and a 100 W solar panel, a hand-crank radio, and a flashlight.
As with everything, LLMs can help you with these tasks. Just don‘t paste or store passwords in clear text.
Medical supplies: it's good to have a stockpile of your daily medications, and a first aid kit (I use this one).
If you think we'll be more on the Mad Max side of things, Amble sells emergency kits with multiple antibiotics and antivirals.
I've also been eyeing https://pawprintoxygen.com/ for home oxygen without a prescription, although I'm leaning against.
In February 2020, LessWrong hosted Coronavirus: Justified Practical Advice Thread . It was a gathering place for people to share and debate small pieces of the overall survival policy and let the wisdom of crowds lead to better outcomes than anyone could get alone.
In that spirit, I notice The Great AI Hackening is at least as imminent as covid was in January 2020, maybe closer. What specific things should people do to personally prepare? What are your unresolved cruxes that would let you answer that question?
All suggestions should be accompanied by reasoning but you don't need to be 100% convinced it's good advice, the whole point is to reason it out together.