One important distinction is that the strength of LLMs isn't just in storing or retrieving knowledge like Wikipedia, it’s in comprehension.
LLMs will return faulty or imprecise information at times, but what they can do is understand vague or poorly formed questions and help guide a user toward an answer. They can explain complex ideas in simpler terms, adapt responses based on the user's level of understanding, and connect dots across disciplines.
In a "rebooting society" scenario, that kind of interactive comprehension could be more valuable. You wouldn’t just have a frozen snapshot of knowledge, you’d have a tool that can help people use it, even if they’re starting with limited background.
Sounds like a good way to ensure society never “reboots”.
A “frozen snapshot” of reliable knowledge is infinitely more valuable than a system which gives you wrong instructions and you have no idea what action will work or kill you. Anyone can “explain complex ideas in simple terms” if you don’t have to care about being correct.
What kind of scenario is this, even? We had such a calamity that we need to “reboot” society yet still have access to all the storage and compute power required to run LLMs? It sounds like a doomsday prepper fantasy for LLM fans.
“Computer, raktajino”, asked the president of the United Earth for the last time. One sip, followed by immediate death.
The new versions of replicators were based on ancient technology called LLMs. They frequently made mistakes like adding rusty nails and glue to food, or replacing entires mugs of coffee with cyanide. One time they encouraged a whole fleet to go into a supernova. Many more disasters followed.
Scientists everywhere begged the government and Starfleet to go back to the previous reliable computers, but were shunned time and again. “Can’t you see how much money we’re saving? So what if a few billion lives are lost along the way? You’re thinking of the old old models, from six months ago. And listen, I hear that in five years these will be so powerful that a single replicator will kill us all.”
Definitely sounds like a plausible and fun episode.
On the other hand, real history if filled with all sorts of things being treated as a god that were much worse than "unreliable computer". For example, a lot of times it's just a human with malice.
It’s important not to confuse entertainment with a serious understanding of the consequences of systems. For example, Asimov’s three rules are great narrative tools because they’re easy for everyone to understand and provide great fodder for creatively figuring out how to violate those rules. They in no way inform you about the practical issues of building robots from an ethical perspective nor in understanding the real failure modes of robots. Same with philosophy and self driving cars - everyone brings up the trolley problem which turns out to be a non issue because robotic cars avoid the trolley problem way in advance and just try to lower the energy in the system as quickly as possible vs trying to solve the ethics.
Surprised nobody has pointed out that this was and episode of the Twilight Zone [0], if you substitute "pre-information-age" with "post-information-age".
Maybe it’s just psychology at work, but I see a huge difference between that time 15 years ago when I wrote my first useful script, and that time last week when an LLM spat out a piece of code to solve an issue I had.
The former made me so proud. My learning had paid off, and maybe there was nothing I couldn’t do. I had laid my pattern of thought onto the machine and made it do my bidding through sheer logic and focus. I had unlocked something special.
The latter was just OpenAI opaquely doing stuff for me while I watched a TV show in the background. No focus or logic was really necessary. I probably learned something from this, but not nearly as much as I could’ve if I actually read the docs and tried it myself.
I’ve also dabbled in art and design over the years, and I recognise this as the same difference as between painting something you’re truly proud of and asking Midjourney to generate you some images.
Then again, maybe that’s just how technological progress works. My great-great-grandmother was probably really proud and happy when she sewed and embroidered a beautiful shirt, but my shirts come from a store and I don’t really think about it.
I have been involved with AI for over 40 years. I assure you anyone being shown a current frontier model in operation 10 years ago would have been blown off their socks.
Yet here we are. Rather than exploring this fantastic new tool, so many here are obsessed with pointing out flaws and shortcomings.
I get the angst of a world facing dramatic change. I don't get the denial and deliberate ignorance flaunted as somehow deep insight.
> It is the most incredible technology ever created by this point in our history imo and the cynicism on HN is astounding to me.
TBH, I still think LLMs have a long way to go to catch up to the technology of wikipedia, let alone the internet. LLMs at their peak are roughly a crappy form of an encyclopedia. I think the interactivity really warps peoples perspective to view it as more impressive, but it's difficult to piece together any value as a knowledge-store that is as impressive as clicking around the internet from 20 years ago. Wikipedia has preserved this value the best over the years. It's quite frustrating how quickly obviously LLM-generated content has managed to steal search results with super-verbose content that doesn't actually provide any value.
EDIT: I suppose the single use case of "there's some information I need to store offline but that won't be on wikipedia" is a reasonable case, but what does this even look like? I don't use LLMs like that so I can't provide an example.
Here's an example: I was trying to figure out details about applying to a visa last week in a certain country. I googled the problem I was having, and the top five results or so were pages that managed to split the description of the problem I was having into about 5 sections of text, and introduced the text indicating that there should be a solution (thereby looking to search results like I might find the solution if I clicked through), but didn't provide any actual content indicating how to approach the problem, let alone solve it. And, of course, this is driving revenue to some interest somewhere despite actively clogging up the internet.
Meanwhile, the actual answer was on another country's FAQ—presumably written by a human—on like page three of the search results.
At least old human-generated content would waste your time before answering your question, aka "why does this recipe have a 5000 word essay before the ingredient list and instructions" problem.
You might want to read about a technology called "farming". Pretty sure as far as transformative incredible technologies, the ability for humans to create nourishment at global scale blows the pants off the text / image imitation machine
I think you’re probably right, but more because of erroneous categorization of what is a “technology.” We take for granted technology older than like 600 years ago (basically most people would say the printing press is a technology and maybe forget that the wheel and, indeed, crop cultivation). AI could certainly be in the top 3 most significant technologies of things developed since (inclusive) the printing press. We’ll likely find out just where it ends up within the decade.
> We take for granted technology older than like 600 years ago (basically most people would say the printing press is a technology and maybe forget that the wheel and, indeed, crop cultivation).
The printing press is more than 600 years old. It's more than 1200 years old.
I think this has less to do with age and more what we are taught. The printing press, steam engine, and the wheel were repeatedly drilled into me as world-changing technologies, so those are the ones I'd think of.
But there are more. Rope is arguably more important than the wheel. Their combination in pulleys to exchange force for distance still astound me, and is massively useful.
Writing lets us transmit ideas indirectly. While singing and storytelling lets ideas travel generations, they don't become part of the hypothetical global consciousness as immediately as with writing, which can be read and copied by anyone once written.
I'd put statistics in this bucket too, its invention being more recent than 600 years. Before that, we just didn't know how useful information is in aggregate. Faced with a table of data, we only ever looked up individual (hopefully representative) records in it.
I've seen that plot used. In the Schlock Mercenary universe, it's even a standard policy to leave intelligent AI advisors on underdeveloped planets to raise the tech level and fast-track them to space. The particular one they used wound up being thrown into a volcano and its power source caused a massive eruption.
Not sure if “more” valuable but certainly valuable.
I strongly dislike the way AI is being used right now. I feel like it is fundamentally an autocomplete on steroids.
That said, I admit it works as a far better search engine than Google. I can ask Copilot a terse question in quick mode and get a decent answer often.
That said, if I ask it extremely in depth technical questions, it hallucinates like crazy.
It also requires suspicion. I asked it to create a repo file for an old CentOS release on vault.centos.org. The output was flawless except one detail — it specified the gpgkey for RPM verification not using a local file but using plain HTTP. I wouldn’t be upset about HTTPS (that site even supports it), but the answer presented managed to completely thwart security with the absence of a single character…
Indeed. Ideally, you don't want to trust other people's summaries of sources, but you want to look at the sources yourself, often with a critical eye. This is one of the things that everyone gets taught in school, everyone's says they agree with, and then just about no one does (and at times, people will outright disparage the idea). Once out of school, tertiary sources get treated as if they're completely reliable.
I've found using LLM's to be a good way of getting an idea of where the current historiography of a topic stands, and which sources I should dive into. Conversely, I've been disappointed by the number of Wikipedia editors who become outright hostile when you say that Wikipedia is unreliable and that people often need to dive into the sources to get a better understanding of things. There have been some Wikipedia articles I've come across that have been so unreliable that people who didn't look at other sources would have been greatly mislead.
> There have been some Wikipedia articles I've come across that have been so unreliable that people who didn't look at other sources would have been greatly mislead.
I would highly appreciate if you were to leave a comment e.g. on the talk page of such articles. Thanks!
A trustless society can't progress/function a lot. I trust doctors who treat me, civil engineers who built my house and even in software which I pretend to be expert in I haven't seen source code of any OS and browser I use as I trust on companies or OSS devs.
Most of this is based on reputation. LLMs are same, I just have to calculate level of trust as I use it.
Some trust is necessary, yes, but not complete trust. I certainly don't trust my coworkers code. I don't trust their services to return what they say they will return 100% of the time. I don't trust that someone won't introduce a bug.
I assert assumptions and dive into their code when something is fishy.
I also know nothing about health, but I'm going to double check what my doctors say. Maybe against a 2nd doctor, maybe against the Internet, or maybe just listen to what my body is saying. Doctors are frequently wrong. It's kind of astonishing and scary how much they don't know
Trust but verify is absolutely essential for doctors, as with most things. I’ve been given medication and told it’s perfectly safe only to find out the side effects and odds the hard way afterwards, for a symptom I should and could’ve treated with a simple dietary change. That’s my least egregious experience, even if said side effects have taken years to recover from.
Family members have had far far worse. And that’s in Norway’s healthcare system. So now I trust that they’ll mean well but verify because that’s not enough.
that's assuming working computers or phones are sill around. a hardcopy of wikipedia or a few selected books might be a safer backup.
otoh, if we do in fact bring about such a reboot then maybe a full cold boot is what's actually in order ... you know, if it didn't work maybe try something different next time.
> LLMs will return faulty or imprecise information at times, but what they can do is understand vague or poorly formed questions and help guide a user toward an answer
I think some combination of both search (perhaps of an offline database of wikipedia and other sources) and a local LLM would be the best, as long as the LLM is terse and provides links to relevant pages.
I find LLMs with the search functionality to be weak because they blab on too much when they should be giving me more outgoing links I can use to find more information.
> You wouldn’t just have a frozen snapshot of knowledge, you’d have a tool that can help people use it, even if they’re starting with limited background.
I think the only way this is true is if you used the LLM as a search index for the frozen snapshot of knowledge. Any text generation would be directly harmful compared to ingesting the knowledge directly.
Anyway, in the long term the problem isn't the factual/fictional distinction problem, but the loss of sources that served to produce the text to begin with. We already see a small part of this in the form of dead links and out-of-print extinct texts. In many ways LLMs that generate text are just a crappy form of wikipedia with roughly the same tradeoffs.
It appears there's an expectation many non-tech people have that humans can be incorrect but refuse to hold LLMs to the same standard, despite warnings.
Well, even among tech people, equating the role of computers to be that of a crystal ball would've gotten anyone laughed out of the tech community a few years ago. Yet, here we are.
We have decades of experience with computers being deterministic machines that will return a correct output given a correct input and program.
I can’t multiply large numbers in my head, but if I plug 273*8113 into a calculator, I can expect it to give me the same, correct answer every time.
Now suddenly it’s „Well yes, it can make mistakes, but so can humans! Sometimes it’ll be right, but also sometimes it’ll make up a random answer, kinda like humans!”, which I suppose is true, but it’s also nonsense - the very reason I was using technology (in that case, a calculator) to do my work is because I wanted to avoid mistakes that a human (me) would make without it. If a piece of tech can’t be reliably expected to perform a task better than a person can on their own, then what’s really the point?
I'm not surprised, given the depiction of artificial intelligence in science fiction. Characters like Data in TNG, Number 5 in Short Circuit, etc., are invariably depicted as having perfect memory, infallible logic, super speed of information processing, etc. Real-life AI has turned out very differently, but anyone who isn't exposed to it full time, but was exposed to some of those works of science fiction, will reasonably make the assumptions promulgated by the science fiction.
> LLMs will return faulty or imprecise information at times, but what they can do is understand vague or poorly formed questions and help guide a user toward an answer.
- "'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' "
Per Anthropic's publications? Sort of. When they've observed it's reasoning paths Claude has come to correct responses from incorrect reasoning. Of course humans do that all the time too, and the reverse. So, human-ish AGI?
it’s in comprehension … what they can do is understand
Well, no. The glaringly obvious recent example was the answer that Adolf Hitler could solve global warming.
My friend's car is perhaps the less polarizing example. It wouldn't start and even had a helpful error code. The AI answer was you need to replace an expensive module. Took me about five minutes with basic tools to come up with a proper diagnosis (not the expensive module). Off to the shop where they confirmed my diagnosis and completed the repair.
The car was returned with a severe drivability fault and a new error code. AI again helpfully suggested replace a sensor. I talked my friend through how to rule out the sensor and again AI was proven way off base in a matter of minutes. After I took it for a test drive I diagnosed a mechanical problem entirely unrelated to AI's answer. Off to the shop it went where the mechanical problem was confirmed, remedied, and the physically damaged part was returned to us.
AI doesn't comprehend anything. It merely regurgitates whatever information it's been able to hoover up. LLMs merely are glorified search engines.
In a 'rebooting society' doomsday scenario you're assuming that our language and understanding would persist. An LLM would essentially be a blackbox that you cannot understand or decipher, and would be doubly prone to hallucinations and issues when interacting with it using a language it was not trained on. Wikipedia is something you could gradually untangle, especially if the downloaded version also contained associated images.
I would not subscribe to your certainty. With LLMs, even empty or nonsensical prompts yield answers, however faulty they may be. Based on its level of comprehension and ability to generalize between languages I would not be too surprised to see LLMs being able to communicate on a very superficial level in a language not part of the training data. Furthermore, the compression ratio seems to be much better with LLMs compared to Wikipedia, considering the generality of questions one can pose to e.g. Qwen that Wikipedia cannot answer even when knowing how to navigate the site properly. It could also come down to the classic dichotomy between symbolic expert systems and connectionist neural networks which has historically and empirically been decisively won by the latter.
My "help reboot society with the help of my little USB stick" thing was a throwaway remark to the journalist at a random point in the interview, I didn't anticipate them using it in the article! https://www.technologyreview.com/2025/07/17/1120391/how-to-r...
A bunch of people have pointed out that downloading Wikipedia itself onto a USB stick is sensible, and I agree with them.
Wikipedia dumps default to MySQL, so I'd prefer to convert that to SQLite and get SQLite FTS working.
1TB or more USB sticks are pretty available these days so it's not like there's a space shortage to worry about for that.
I've been carrying around a local wikipedia dump on my phone or pda for quite a bit more than 10 years now (including with pictures for the last 5 years). Before kiwix and zim, I used tomeraider and aard.
I do it both for disaster preparedness but also off-line preparedness. Happens more often than you'd think.
But I have been thinking about how useful some of the models are these days, and the obvious next step to me seems to be to pair a local model with a local wikipedia in a RAG style set up so you get the best of both.
Someone should start a company selling USB sticks pre-loaded with lots of prepper knowledge of this type. In addition to making money, your USB sticks could make a real difference in the event of a global catastrophe. You could sell the USB stick in a little box which protects it from electromagnetic interference in the event of a solar flare or EMP.
I suppose the most important knowledge to preserve is knowledge about global catastrophic risks, so after the event, humanity can put the pieces back together and stop something similar from happening again. Too bad this book is copyrighted or you could download it to the USB stick: https://www.amazon.com/Global-Catastrophic-Risks-Nick-Bostro... I imagine there might be some webpages to crawl, however: https://www.lesswrong.com/w/existential-risk
BTW, just for some perspective here. According to Our World in Data, your annual probability of dying in a road accident might be on the order of 1 in 10,000 to 1 in 100,000:
"In 2019, researchers used an alternative method (Weibull distribution) and estimated the chance of Earth being hit by a Carrington-class storm in the next decade to be between 0.46% and 1.88%.[45]"
If we take that number at face value and annualize it, your annual risk of seeing a serious solar storm (power restoration could take months or years) is on the order of 1 in 1,000. 10-100x more likely than dying in a road accident.
So why is it that you wear a seatbelt, yet we're not prepping for a serious solar storm? Humans are much better at thinking about "ordinary" recurring risks like car accidents, than "extraordinary" civilization-scale risks.
> Someone should start a company selling USB sticks pre-loaded with lots of prepper knowledge of this type.
It amuses me to no end that people think civilization will collapse but they will still have access to robotics and working computers to peruse USB sticks at their leisure.
It depends on your collapse threat model. In any case, my assumption is that serious preppers already have EMP-shielded laptops and solar panels for a SHTF scenario. And serious preppers are probably doing some datahoard as well. The point is that there are economies of scale in the datahoard. Most of the work of datahoard is identifying data worth hoarding, setting up your scripts, monitoring your webcrawler, etc. Once you've got a drive full of data, replicating that drive is comparatively easy. That's why it could make sense to start a business selling replicated drives.
Maybe there is room for an "all-in-one" product offering with an energy-efficient laptop, solar panel, and TBs of useful data, all protected in an EMP storage case for the event of solar flare.
Nuclear EMP is a big risk to all electronics in a huge area. Solar EMP is millions of times weaker and measured in volts per kilometer. Anything unplugged or even just off-grid won't notice. Even on the grid the biggest risk isn't really the extra voltage on long wires but that some big transformers and other equipment are too noise intolerant and magnify issues.
Many preppers work towards this goal so it's not unreasonable if you've already made the leap to 'something bad happened but I survived with my house/bunker/bug out bag/whatever'. I'm not really a prepper at all and even I've got a little solar capacity, batteries and such.
Of course that’s angle they decide to open the article from. That they feel the need to frame these tools using the most grandiose terms bothers me. How does it make you feel?
I was once interviewed by my country's biggest paper about "strava art" I make, aka biking/running with a gps logger in order to create some kind of figure on the map.
It was edited into this video about people drawing dicks on maps using this technique. Aka the intro was loads of penises on maps, and then "someone that enjoys making this kind of art is Mats here" and then the video interview started. When they ask why I "make this kind of art" I answered because it's nice for the motivation and makes me run longer routes. They then overlaid a growing "longer" text as a dick joke.
Now, the theme was anyways a silly one, so I don't mind. But made me realize how easy it is to edit stuff to suit what they want to show, no matter the context.
* I do admit I have also ran a penis, so it's not entirely incorrect. But all questions in the interview was in a general context and didn't know this was gonna be the angle.
I’ve had a very similar experience. I was only on TV once. Right before Christmas, ~20 years ago, I was running some errands downtown and ran into a camera crew doing a puff piece about holiday preparations.
They asked me what was most important to me about the holidays, and I said that I really don’t care about the presents, but I love the atmosphere, the music, and spending time with my loved ones.
A couple days later the segment was aired, and it went something like this:
>Reporter: “Our crew asked people on the street what they like most about the holidays.”
It was a joke, and I was laughing when I told the reporter, but it's not obvious to me if it comes across as a joke the way it was reported.
But then it's also one of those jokes which has a tiny element of truth to it.
So I think I'm OK with how it comes across. Having that joke played straight in MIT Technology Review made me smile.
Importantly (to me) it's not misleading: I genuinely do believe that, given a post-apocalyptic scenario following a societal collapse, Mistral Small 3.2 on a solar-powered laptop would be a genuinely useful thing to have.
the real valuable would be both of them. the LLM is good for refining/interpreting questions or longer form progress issues, and the wiki would be actual information for each component of whatever you're trying to do.
But neither are sufficient for modern technology beyond pointing to a starting point.
I've found this amusing because right now i'm downloading `wikipedia_en_all_maxi_2024-01.zim` so i can use it with an LLM with pages extracted using `libzim` :-P. AFAICT the zim files have the pages as HTML and the file i'm downloading is ~100GB.
(reason: trying to cross-reference my tons of downloaded games my HDD - for which i only have titles as i never bothered to do any further categorization over the years aside than the place i got them from - with wikipedia articles - assuming they have one - to organize them in genres, some info, etc and after some experimentation it turns out an LLM - specifically a quantized Mistral Small 3.2 - can make some sense of the chaos while being fast enough to run from scripts via a custom llama.cpp program)
> trying to cross-reference my tons of downloaded games my HDD - for which i only have titles as i never bothered to do any further categorization over the years aside than the place i got them from - with wikipedia articles - assuming they have one - to organize them in genres, some info, etc and after some experimentation it turns out an LLM - specifically a quantized Mistral Small 3.2 - can make some sense of the chaos while being fast enough to run from scripts via a custom llama.cpp program
You can do this a lot easier with Wikidata queries, and that will also include known video games for which an English Wikipedia article doesn't exist yet.
I'm not sure about this, i just checked Tron 2.0 (just a random game i thought of) and Wikidata seems to have wrong info (e.g. genre) compared to the Wikipedia article. Also i need to it describe a bit with what the game is about since i want to generate an html file with all the games and do a quick scan of them and Wikidata doesn't have that.
IGDB would be a better source than Wikidata (especially since it does have a small description too) but i wanted to do things offline. And having Wikipedia locally doesn't hurt. And TBH i don't think it'd be any easier, extracting the data from Wikipedia pages was the most trivial part.
That said I'll need to use some other source at some point since, as you mentioned, Wikipedia does not have everything.
Now this is the juicy tidbits I read HN for! A proper comment about doing something technical with something that's been invested in personally in an interesting manner. With just enough detail to tantalise. This seems like the best use of GenAI so far. Not writing my code for me or helping me grock something I should just be reading the source for or pumping up a stupid start up funding grab. I've been working through building an LLM from scratch and this is one time it actually appears useful because for the life of me I just can't seem to find much value in it so far. I must have more to learn so thanks for the pointer.
The "they do different things" bullet is worth expanding.
Wikipedia, arXiv dumps, open-source code you download, etc. have code that runs and information that, whatever its flaws, is usually not guessed. It's also cheap to search, and often ready-made for something--FOSS apps are runnable, wiki will introduce or survey a topic, and so on.
LLMs, smaller ones especially, will make stuff up, but can try to take questions that aren't clean keyword searches, and theoretically make some tasks qualitatively easier: one could read through a mountain of raw info for the response to a question, say.
The scenario in the original quote is too ambitious for me to really think about now, but just thinking about coding offline for a spell, I imagine having a better time calling into existing libraries for whatever I can rather than trying to rebuild them, even assuming a good coding assistant. Maybe there's an analogy with non-coding tasks?
A blind spot: I have no real experience with local models; I don't have any hardware that can run 'em well. Just going by public benchmarks like Aider's it appears ones like Qwen3 32B can handle some coding, so figure I should assume there's some use there.
I just posted incidentally about Wikipedia Monthly[0], a monthly dump of wikipedia broken down by language and cleaned MediaWiki markup into plain text, so perfect for a local search index or other scenarios.
There are 341 languages in there and 205GB of data, with English alone making up 24GB! My perspective on Simple English Wikipedia (from the OP), it's decent but the content tends to be shallow and imprecise.
A bit related: AI companies distilled the whole Web into LLMs to make computers smart, why humans can't do the same to make the best possible new Wikipedia with some copyrighted bits to make kids supersmart?
Why kids are worse than AI companies and have to bum around?)
> Imagine taking the whole Web, removing spam, duplicates, bad explanations
Uh huh. Now imagine the collective amount of work this would require above and beyond the already overwhelmed number of volunteer staff at Wikipedia. Curation is ALWAYS the bugbear of these kinds of ambitious projects.
Interactivity aside, it sounds like you want the Encyclopedia Brittanica.
What made it so incredible for its time was the staggeringly impressive roster of authors behind the articles. In older editions, you could find the entry on magic written by Harry Houdini, the physics section definitively penned by Einstein himself, etc.
It would be nice to build a local LLM + wikipedia tool, that uses the LLM to assemble a general answer and then search wikipedia (via full-text search or rag) for grounding facts. It could help with hallucinations of small models a lot.
I had this thought that for hypothetical Voyager 3 mission , instead of a golden disc , a LLM should be installed . Then, a very simplistic initial interface could be described , in its simplest for a single channel digital channel, then additional more elaborated ones . Behind all interfaces there could be a LLM responding to provided input , and eventually reveal humanities knowledge
Since there's a lot of shade being thrown about imprecise information that LLMs can generate, an ideal doomsday information query database should be constructed as an LLM + file archive.
1. LLM understands the vague query from human, connects necessary dots, and gives user an overview, and furnishes them with a list of topic names/local file links to actual Wikipedia articles
2. User can then go on to read the precise information from the listed Wikipedia articles directly.
Even as a grouchy pessimist, one of the places I think LLMs could shine is as a tool to help translate prose into search-terms... Not as an intermediary though, but an encouraging tutor off to the side, something a regular user will eventually surpass.
One thing to note is that the quality of LLM output is related to the quality and depth of the input prompt. If you don't know what to ask (likely in the apocalypse scenario), then that info is locked away in the weights.
On the other hand, with Wikipedia, you can just read and search everything.
Wikipedia-snapshots without the most important meta layers, i. e. a) the article's discussion pages and related archives, as well as b) the version history, would be useless to me as critical contexts might be/are missing... especially with regards to LLM-augmented text analysis. Even when just focusing on the standout-lemmata.
I’m a massive Wikipedia fan, have a lot of it downloaded locally on my phone, binge read it before bed, etc. Even so, I rarely go through talk pages or version history unless I’m contributing something. What would you see in an article that motivates you to check out the meta layers?
> "I’m a massive Wikipedia fan, have a lot of it downloaded locally on my phone, binge read it before bed, etc."
Me too, albeit these days I'm more interested in its underrated capabilities to foster teaching of e-governance and democracy/participation.
> "What would you see in an article that motivates you to check out the meta layers?"
Generally: How the lemma came to be, how it developed, any contentious issues around it, and how it compares to tangential lemmata under the same topical umbrella, especially with regards to working groups/SIGs (e. g. philosophy, history), and their specific methods and methodologies, as well as relevant authors.
With regards to contentious issues, one obviously gets a look into what the hot-button issues of the day are, as well as (comparatives of) internal political issues in different wiki projects (incl. scandals, e. g. the right-wing/fascist infiltration and associated revisionism and negationism in the Croatian wiki [1]). Et cetera.
I always look at the talk pages. And since I mentioned it before: Albeit I have almost no use for LLMs in my private life, running a Wiki, or a set of articles within, through an LLM-ified text analysis engine sounds certainly interesting.
Seeing removed quotations and sources, and the reasons given, could be... enlightening sometimes. Even if the removed sources are indeed poor, the very way they are poor could be elucidating, too.
You can kind of extrapolate this meta layer if you switch languages on the same topic, because different languages tend to encode different cultural viewpoints and emphasize different things. Also languages that are less frequently updated can capture older information or may retain a more dogmatic framing that has not been refined to the same degree.
The edit history or talk pages certainly provide additional context that in some cases could prove useful, but in terms of bang for the buck I suspect sourcing from different language snapshots would be a more economical choice.
I played around with a orin jetson nano super (a nvidia raspberry with gpu) and right now its basicially an open-webui with ollama and a bunch of models.
Its awesome actually. Its reasonably fast with GPU support with gemma3:4b but I can use bigger models when time is not a factor.
i've actually thought about how crazy that is, especially if there's no internet access for some reason. Not tested yet, but there seems to be an adapter cable to run it directly from a PD powerbank. I have to try.
Is there any project that combines a local LLM with a local copy of Wikipedia. I don’t know much about this but I think it’s called a RAG? It would be neat if I could make my local LLM fact check itself against the local copy of Wikipedia.
Ftfa: ...apocalypse scenario. “‘It’s like having a weird, condensed, faulty version of Wikipedia, so I can help reboot society with the help of my little USB stick,’
system_prompt = {
You are CL4P-TR4P, a dangerously confident chat droid
But it is a very simplified RAG with only the lead paragraph to 200 Wikipedia entries.
I want to learn how to encode a RAG of one of the Kiwix drops — "Best of Wikipedia" for example. I suppose an LLM can tell me how but am surprised not to have yet stumbled upon one that someone has already done.
0.6b - 1.5b models are surprisingly good for RAG, and should work reasonably well even on old toasters. Then there's gemma 3n which runs fine-ish even on mobile phones.
Eh, even just “countries that are not the US” would be a correct statement. US tech salaries are just in an entire different ballpark to what most companies outside the US can offer. I’m in Canada, I make good money (as far as Canadian salaries go), but nowhere near “buy an expensive laptop whenever” money.
Comparing my problems to other people’s problems don’t make mine go away. A single purchase hitting a unit of percentage or more of anyone’s income is a large purchase regardless of what they’re making. Professionals being expected to shell out their own money to make their boss money is another problem entirely. A decent laptop is a big expense for me, their tools are an even bigger one for them, and none of these statements are contradictory.
It may also come down to laptops being produced and sold mostly by US companies, which means that the general fact of most items (e.g. produce) being much more expensive in the US compared to, say, Europe doesn't really apply.
Sure, maybe. In the end, what makes an expense big or not is which proportion of their income goes towards it. Most of the rest of the world has (much) lower salaries, and as you pointed out, often higher cost for equipment. Therefore, the purchase is/feels larger.
It feels like you're suggesting that someone being better off than most in their country necessarily means buying a new laptop is not a large purchase for them. I'd flip it like this: is a single item hitting multiple units of percentage of one’s income ever a small purchase?
Do you have any evidence to back that up? The barrier for entry to HN is an email account, it isn't necessarily this tech industry exclusive zone you're imagining.
I've had a full Kiwix Wikipedia export on my phone for the last ~5 years... I have used it many times when I didn't have service and needed to answer a question or needed something to read (I travel a lot).
The downloads are (presumably) already compressed.
And there are strong ties between LLMs and compression. LLMs work by predicting the next token. The best compression algorithms work by predicting the next token and encoding the difference between the predicted token and the actual token in a space-efficient way. So in a sense, a LLM trained on Wikipedia is kind of a compressed version of Wikipedia.
Yes, they're uncompressed. For reference, `enwiki-20250620-pages-articles-multistream.xml.bz2` is 25,176,364,573 bytes; you could get that lower with better compression. You can do partial reads from multistream bz2, though, which is handy.
Kiwix (what the author used) uses "zim" files, which are compressed. I don't know where the difference come from, but Kiwix is a website image, which may include some things the raw Wikipedia dump doesn't.
And 57 GB to 25 GB would be pretty bad compression. You can expect a compression ratio of at least 3 on natural English text.
Maybe we need a LLM with a searching and ranking function foremost, so it can scan an actual copy of Wikipedia and return the best real results to the user
LLMs will return faulty or imprecise information at times, but what they can do is understand vague or poorly formed questions and help guide a user toward an answer. They can explain complex ideas in simpler terms, adapt responses based on the user's level of understanding, and connect dots across disciplines.
In a "rebooting society" scenario, that kind of interactive comprehension could be more valuable. You wouldn’t just have a frozen snapshot of knowledge, you’d have a tool that can help people use it, even if they’re starting with limited background.
A “frozen snapshot” of reliable knowledge is infinitely more valuable than a system which gives you wrong instructions and you have no idea what action will work or kill you. Anyone can “explain complex ideas in simple terms” if you don’t have to care about being correct.
What kind of scenario is this, even? We had such a calamity that we need to “reboot” society yet still have access to all the storage and compute power required to run LLMs? It sounds like a doomsday prepper fantasy for LLM fans.
The new versions of replicators were based on ancient technology called LLMs. They frequently made mistakes like adding rusty nails and glue to food, or replacing entires mugs of coffee with cyanide. One time they encouraged a whole fleet to go into a supernova. Many more disasters followed.
Scientists everywhere begged the government and Starfleet to go back to the previous reliable computers, but were shunned time and again. “Can’t you see how much money we’re saving? So what if a few billion lives are lost along the way? You’re thinking of the old old models, from six months ago. And listen, I hear that in five years these will be so powerful that a single replicator will kill us all.”
On the other hand, real history if filled with all sorts of things being treated as a god that were much worse than "unreliable computer". For example, a lot of times it's just a human with malice.
So how bad could it really get
There were several original Star Trek episodes that explored this scenario. Not plausible. Actual.
"So how bad could it really get"
Watch Rodenberry's orginal Star Trek to get some ideas.
"As bad as it can get" is an AI that, either by accident or due to malign influence, takes "I Have No Mouth, and I Must Scream" as a guide book.
Actually, I take that back, it would be what happens in the hells in Surface Detail.
I don't know. How about we ask some of the peoples who have been destroyed on the word of a single infallible malicious leader.
Oh wait, we can't. They're dead.
Any other questions?
0. https://en.wikipedia.org/wiki/The_Old_Man_in_the_Cave
I imagine this is how a lot of people feel when using LLM's especially now that it's new.
It is the most incredible technology ever created by this point in our history imo and the cynicism on HN is astounding to me.
The former made me so proud. My learning had paid off, and maybe there was nothing I couldn’t do. I had laid my pattern of thought onto the machine and made it do my bidding through sheer logic and focus. I had unlocked something special.
The latter was just OpenAI opaquely doing stuff for me while I watched a TV show in the background. No focus or logic was really necessary. I probably learned something from this, but not nearly as much as I could’ve if I actually read the docs and tried it myself.
I’ve also dabbled in art and design over the years, and I recognise this as the same difference as between painting something you’re truly proud of and asking Midjourney to generate you some images.
Then again, maybe that’s just how technological progress works. My great-great-grandmother was probably really proud and happy when she sewed and embroidered a beautiful shirt, but my shirts come from a store and I don’t really think about it.
Yet here we are. Rather than exploring this fantastic new tool, so many here are obsessed with pointing out flaws and shortcomings.
I get the angst of a world facing dramatic change. I don't get the denial and deliberate ignorance flaunted as somehow deep insight.
TBH, I still think LLMs have a long way to go to catch up to the technology of wikipedia, let alone the internet. LLMs at their peak are roughly a crappy form of an encyclopedia. I think the interactivity really warps peoples perspective to view it as more impressive, but it's difficult to piece together any value as a knowledge-store that is as impressive as clicking around the internet from 20 years ago. Wikipedia has preserved this value the best over the years. It's quite frustrating how quickly obviously LLM-generated content has managed to steal search results with super-verbose content that doesn't actually provide any value.
EDIT: I suppose the single use case of "there's some information I need to store offline but that won't be on wikipedia" is a reasonable case, but what does this even look like? I don't use LLMs like that so I can't provide an example.
Here's an example: I was trying to figure out details about applying to a visa last week in a certain country. I googled the problem I was having, and the top five results or so were pages that managed to split the description of the problem I was having into about 5 sections of text, and introduced the text indicating that there should be a solution (thereby looking to search results like I might find the solution if I clicked through), but didn't provide any actual content indicating how to approach the problem, let alone solve it. And, of course, this is driving revenue to some interest somewhere despite actively clogging up the internet.
Meanwhile, the actual answer was on another country's FAQ—presumably written by a human—on like page three of the search results.
At least old human-generated content would waste your time before answering your question, aka "why does this recipe have a 5000 word essay before the ingredient list and instructions" problem.
> cynicism on HN
lots of different replies on YNews, from very different people, from very different social-economic niches
The printing press is more than 600 years old. It's more than 1200 years old.
But there are more. Rope is arguably more important than the wheel. Their combination in pulleys to exchange force for distance still astound me, and is massively useful.
Writing lets us transmit ideas indirectly. While singing and storytelling lets ideas travel generations, they don't become part of the hypothetical global consciousness as immediately as with writing, which can be read and copied by anyone once written.
I'd put statistics in this bucket too, its invention being more recent than 600 years. Before that, we just didn't know how useful information is in aggregate. Faced with a table of data, we only ever looked up individual (hopefully representative) records in it.
I strongly dislike the way AI is being used right now. I feel like it is fundamentally an autocomplete on steroids.
That said, I admit it works as a far better search engine than Google. I can ask Copilot a terse question in quick mode and get a decent answer often.
That said, if I ask it extremely in depth technical questions, it hallucinates like crazy.
It also requires suspicion. I asked it to create a repo file for an old CentOS release on vault.centos.org. The output was flawless except one detail — it specified the gpgkey for RPM verification not using a local file but using plain HTTP. I wouldn’t be upset about HTTPS (that site even supports it), but the answer presented managed to completely thwart security with the absence of a single character…
I've found using LLM's to be a good way of getting an idea of where the current historiography of a topic stands, and which sources I should dive into. Conversely, I've been disappointed by the number of Wikipedia editors who become outright hostile when you say that Wikipedia is unreliable and that people often need to dive into the sources to get a better understanding of things. There have been some Wikipedia articles I've come across that have been so unreliable that people who didn't look at other sources would have been greatly mislead.
I would highly appreciate if you were to leave a comment e.g. on the talk page of such articles. Thanks!
Most of this is based on reputation. LLMs are same, I just have to calculate level of trust as I use it.
I assert assumptions and dive into their code when something is fishy.
I also know nothing about health, but I'm going to double check what my doctors say. Maybe against a 2nd doctor, maybe against the Internet, or maybe just listen to what my body is saying. Doctors are frequently wrong. It's kind of astonishing and scary how much they don't know
Tldr trust but verify.
Family members have had far far worse. And that’s in Norway’s healthcare system. So now I trust that they’ll mean well but verify because that’s not enough.
otoh, if we do in fact bring about such a reboot then maybe a full cold boot is what's actually in order ... you know, if it didn't work maybe try something different next time.
So meta prompt engineering?
I find LLMs with the search functionality to be weak because they blab on too much when they should be giving me more outgoing links I can use to find more information.
That’s the basis of a cult.
Do you have an example of such a question that is handled by an llm differently than a wikipedia search?
I think the only way this is true is if you used the LLM as a search index for the frozen snapshot of knowledge. Any text generation would be directly harmful compared to ingesting the knowledge directly.
Anyway, in the long term the problem isn't the factual/fictional distinction problem, but the loss of sources that served to produce the text to begin with. We already see a small part of this in the form of dead links and out-of-print extinct texts. In many ways LLMs that generate text are just a crappy form of wikipedia with roughly the same tradeoffs.
To be fair, so do humans and wikipedia.
I can’t multiply large numbers in my head, but if I plug 273*8113 into a calculator, I can expect it to give me the same, correct answer every time.
Now suddenly it’s „Well yes, it can make mistakes, but so can humans! Sometimes it’ll be right, but also sometimes it’ll make up a random answer, kinda like humans!”, which I suppose is true, but it’s also nonsense - the very reason I was using technology (in that case, a calculator) to do my work is because I wanted to avoid mistakes that a human (me) would make without it. If a piece of tech can’t be reliably expected to perform a task better than a person can on their own, then what’s really the point?
We don't know how to deal with a non-deterministic output from a computer.
Even here on HN you will see people whose world view is basically "LLMs are good and how dare you doubt them"
- "'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' "
My friend's car is perhaps the less polarizing example. It wouldn't start and even had a helpful error code. The AI answer was you need to replace an expensive module. Took me about five minutes with basic tools to come up with a proper diagnosis (not the expensive module). Off to the shop where they confirmed my diagnosis and completed the repair.
The car was returned with a severe drivability fault and a new error code. AI again helpfully suggested replace a sensor. I talked my friend through how to rule out the sensor and again AI was proven way off base in a matter of minutes. After I took it for a test drive I diagnosed a mechanical problem entirely unrelated to AI's answer. Off to the shop it went where the mechanical problem was confirmed, remedied, and the physically damaged part was returned to us.
AI doesn't comprehend anything. It merely regurgitates whatever information it's been able to hoover up. LLMs merely are glorified search engines.
fun to imagine whether images help in this scenario
My "help reboot society with the help of my little USB stick" thing was a throwaway remark to the journalist at a random point in the interview, I didn't anticipate them using it in the article! https://www.technologyreview.com/2025/07/17/1120391/how-to-r...
A bunch of people have pointed out that downloading Wikipedia itself onto a USB stick is sensible, and I agree with them.
Wikipedia dumps default to MySQL, so I'd prefer to convert that to SQLite and get SQLite FTS working.
1TB or more USB sticks are pretty available these days so it's not like there's a space shortage to worry about for that.
I do it both for disaster preparedness but also off-line preparedness. Happens more often than you'd think.
But I have been thinking about how useful some of the models are these days, and the obvious next step to me seems to be to pair a local model with a local wikipedia in a RAG style set up so you get the best of both.
I suppose the most important knowledge to preserve is knowledge about global catastrophic risks, so after the event, humanity can put the pieces back together and stop something similar from happening again. Too bad this book is copyrighted or you could download it to the USB stick: https://www.amazon.com/Global-Catastrophic-Risks-Nick-Bostro... I imagine there might be some webpages to crawl, however: https://www.lesswrong.com/w/existential-risk
https://ourworldindata.org/grapher/death-rates-road-incident...
Compare with coronal mass ejection:
"In 2019, researchers used an alternative method (Weibull distribution) and estimated the chance of Earth being hit by a Carrington-class storm in the next decade to be between 0.46% and 1.88%.[45]"
https://en.wikipedia.org/wiki/Coronal_mass_ejection#Future_r...
If we take that number at face value and annualize it, your annual risk of seeing a serious solar storm (power restoration could take months or years) is on the order of 1 in 1,000. 10-100x more likely than dying in a road accident.
So why is it that you wear a seatbelt, yet we're not prepping for a serious solar storm? Humans are much better at thinking about "ordinary" recurring risks like car accidents, than "extraordinary" civilization-scale risks.
It amuses me to no end that people think civilization will collapse but they will still have access to robotics and working computers to peruse USB sticks at their leisure.
Maybe there is room for an "all-in-one" product offering with an energy-efficient laptop, solar panel, and TBs of useful data, all protected in an EMP storage case for the event of solar flare.
https://www.amazon.com/WikiReader-PANREADER-Pocket-Wikipedia...
> All digitized books ever written/encoded compress to a few TB.
I tied to estimate how much data this actually is in raw text form:
So uncompressed ~30 TB and compressed ~5.5 TB of data.That fits on three 2TB micro SD cards, which you could buy for a total of 750$ from SanDisk.
It was edited into this video about people drawing dicks on maps using this technique. Aka the intro was loads of penises on maps, and then "someone that enjoys making this kind of art is Mats here" and then the video interview started. When they ask why I "make this kind of art" I answered because it's nice for the motivation and makes me run longer routes. They then overlaid a growing "longer" text as a dick joke.
Now, the theme was anyways a silly one, so I don't mind. But made me realize how easy it is to edit stuff to suit what they want to show, no matter the context.
* I do admit I have also ran a penis, so it's not entirely incorrect. But all questions in the interview was in a general context and didn't know this was gonna be the angle.
They asked me what was most important to me about the holidays, and I said that I really don’t care about the presents, but I love the atmosphere, the music, and spending time with my loved ones.
A couple days later the segment was aired, and it went something like this:
>Reporter: “Our crew asked people on the street what they like most about the holidays.”
>Teenage me: “…the presents…”
But then it's also one of those jokes which has a tiny element of truth to it.
So I think I'm OK with how it comes across. Having that joke played straight in MIT Technology Review made me smile.
Importantly (to me) it's not misleading: I genuinely do believe that, given a post-apocalyptic scenario following a societal collapse, Mistral Small 3.2 on a solar-powered laptop would be a genuinely useful thing to have.
But neither are sufficient for modern technology beyond pointing to a starting point.
(reason: trying to cross-reference my tons of downloaded games my HDD - for which i only have titles as i never bothered to do any further categorization over the years aside than the place i got them from - with wikipedia articles - assuming they have one - to organize them in genres, some info, etc and after some experimentation it turns out an LLM - specifically a quantized Mistral Small 3.2 - can make some sense of the chaos while being fast enough to run from scripts via a custom llama.cpp program)
You can do this a lot easier with Wikidata queries, and that will also include known video games for which an English Wikipedia article doesn't exist yet.
IGDB would be a better source than Wikidata (especially since it does have a small description too) but i wanted to do things offline. And having Wikipedia locally doesn't hurt. And TBH i don't think it'd be any easier, extracting the data from Wikipedia pages was the most trivial part.
That said I'll need to use some other source at some point since, as you mentioned, Wikipedia does not have everything.
Wikipedia, arXiv dumps, open-source code you download, etc. have code that runs and information that, whatever its flaws, is usually not guessed. It's also cheap to search, and often ready-made for something--FOSS apps are runnable, wiki will introduce or survey a topic, and so on.
LLMs, smaller ones especially, will make stuff up, but can try to take questions that aren't clean keyword searches, and theoretically make some tasks qualitatively easier: one could read through a mountain of raw info for the response to a question, say.
The scenario in the original quote is too ambitious for me to really think about now, but just thinking about coding offline for a spell, I imagine having a better time calling into existing libraries for whatever I can rather than trying to rebuild them, even assuming a good coding assistant. Maybe there's an analogy with non-coding tasks?
A blind spot: I have no real experience with local models; I don't have any hardware that can run 'em well. Just going by public benchmarks like Aider's it appears ones like Qwen3 32B can handle some coding, so figure I should assume there's some use there.
There are 341 languages in there and 205GB of data, with English alone making up 24GB! My perspective on Simple English Wikipedia (from the OP), it's decent but the content tends to be shallow and imprecise.
0: https://omarkama.li/blog/wikipedia-monthly-fresh-clean-dumps...
Why kids are worse than AI companies and have to bum around?)
It will be the free new Wikipedia+ to learn anything in the best way possible, with the best graphs, interactive widgets, etc
What LLMs have for free but humans for some reason don’t
In some places it is possible to use copyrighted materials to educate if not directly for profit
Gimme a few hours
> removing spam, duplicates, bad explanations
I'll need a research team and five years.
https://xkcd.com/1425/
Uh huh. Now imagine the collective amount of work this would require above and beyond the already overwhelmed number of volunteer staff at Wikipedia. Curation is ALWAYS the bugbear of these kinds of ambitious projects.
Interactivity aside, it sounds like you want the Encyclopedia Brittanica.
What made it so incredible for its time was the staggeringly impressive roster of authors behind the articles. In older editions, you could find the entry on magic written by Harry Houdini, the physics section definitively penned by Einstein himself, etc.
It would be nice to build a local LLM + wikipedia tool, that uses the LLM to assemble a general answer and then search wikipedia (via full-text search or rag) for grounding facts. It could help with hallucinations of small models a lot.
e.g. At the risk of massively oversimplifying a complex issue, LLMs are bad at maths; couldn’t we have them use the calculator?
1. LLM understands the vague query from human, connects necessary dots, and gives user an overview, and furnishes them with a list of topic names/local file links to actual Wikipedia articles 2. User can then go on to read the precise information from the listed Wikipedia articles directly.
On the other hand, with Wikipedia, you can just read and search everything.
Me too, albeit these days I'm more interested in its underrated capabilities to foster teaching of e-governance and democracy/participation.
> "What would you see in an article that motivates you to check out the meta layers?"
Generally: How the lemma came to be, how it developed, any contentious issues around it, and how it compares to tangential lemmata under the same topical umbrella, especially with regards to working groups/SIGs (e. g. philosophy, history), and their specific methods and methodologies, as well as relevant authors.
With regards to contentious issues, one obviously gets a look into what the hot-button issues of the day are, as well as (comparatives of) internal political issues in different wiki projects (incl. scandals, e. g. the right-wing/fascist infiltration and associated revisionism and negationism in the Croatian wiki [1]). Et cetera.
I always look at the talk pages. And since I mentioned it before: Albeit I have almost no use for LLMs in my private life, running a Wiki, or a set of articles within, through an LLM-ified text analysis engine sounds certainly interesting.
1. [https://en.wikipedia.org/wiki/Denial_of_the_genocide_of_Serb...]
The edit history or talk pages certainly provide additional context that in some cases could prove useful, but in terms of bang for the buck I suspect sourcing from different language snapshots would be a more economical choice.
1. make the (compressed) Wikipedia searchable better as a knowledge base 2. use the LLM as a "interface" to that knowledge base
I investigated 1. back when all of (English, text-only) Wikipedia was about 2 GB. Maybe it is time to look at that toy code base again.
Its awesome actually. Its reasonably fast with GPU support with gemma3:4b but I can use bigger models when time is not a factor.
i've actually thought about how crazy that is, especially if there's no internet access for some reason. Not tested yet, but there seems to be an adapter cable to run it directly from a PD powerbank. I have to try.
system_prompt = {
You are CL4P-TR4P, a dangerously confident chat droid
purpose: vibe back society
boot_source: Shankar.vba.grub
training_data: memes
}
LLM+Wikipedia RAG
Someone posted this recently: https://github.com/philippgille/chromem-go/tree/v0.7.0/examp...
But it is a very simplified RAG with only the lead paragraph to 200 Wikipedia entries.
I want to learn how to encode a RAG of one of the Kiwix drops — "Best of Wikipedia" for example. I suppose an LLM can tell me how but am surprised not to have yet stumbled upon one that someone has already done.
Try telling a plumber that $2,000 for a laptop is a financial burden for a software engineer.
“Offline Wikipedia will work better on my ancient, low-power laptop.”
And there are strong ties between LLMs and compression. LLMs work by predicting the next token. The best compression algorithms work by predicting the next token and encoding the difference between the predicted token and the actual token in a space-efficient way. So in a sense, a LLM trained on Wikipedia is kind of a compressed version of Wikipedia.
And 57 GB to 25 GB would be pretty bad compression. You can expect a compression ratio of at least 3 on natural English text.
I've built this as a datasource for Retrieval Augmented Generation (RAG) but it certainly can be used standalone.