Local LLM, Humanity's Last Library

[CuriousBread]

The apocalypse is upon us. The internet is no more. There are no more libraries. No more schools. There are only local networks and people with the means to power them. 

How'd you build humanity's last library as an LLM that contains the entirety of human knowledge with what you have? It needs to be easy to power and rugged. 

Potentially it'd be decades or even centuries before we have the infrastructure to make electronics again. 

For those who knows Warhammer. I'm basically asking how'd you build a STC. 

 

What model would you use? What hardware would you use? How'd you set it up? 

[Hinjima]

3 minutes ago, CuriousBread said:

The apocalypse is upon us. The internet is no more. There are no more libraries. No more schools. There are only local networks and people with the means to power them. 

How'd you build humanity's last library that contains the entirety of human knowledge with what you have? It needs to be easy to power and rugged. 

Potentially it'd be decades or even centuries before we have the infrastructure to make electronics again. 

For those who knows Warhammer. I'm basically asking how'd you build a STC. 

 

What model would you use? What hardware would you use? How'd you set it up? 

The Book of Eli with Denzel Washington comes to mind.

If you haven't seen it, great movie.

[danalog]

I'd probably burn a bunch of those optical disks that can hold 125TB and carry them around in a comically large spindle. All of the human knowledge worth saving wouldn't be more than 100lbs of those 

[CuriousBread]

5 minutes ago, danalog said:

I'd probably burn a bunch of those optical disks that can hold 125TB and carry them around in a comically large spindle. All of the human knowledge worth saving wouldn't be more than 100lbs of those 

The issue is CD-R degrades after 5-10 years. Commercially pressed one last longer but that's not quite a STC. More so a big digital encyclopedia 

[Kisai]

1 hour ago, CuriousBread said:

The apocalypse is upon us. The internet is no more. There are no more libraries. No more schools. There are only local networks and people with the means to power them. 

How'd you build humanity's last library as an LLM that contains the entirety of human knowledge with what you have? It needs to be easy to power and rugged. 

Potentially it'd be decades or even centuries before we have the infrastructure to make electronics again. 

For those who knows Warhammer. I'm basically asking how'd you build a STC. 

 

What model would you use? What hardware would you use? How'd you set it up? 

 

You would need something theoretical like the isolinear chips in star trek.

 

Basically you need a kind of EPROM (not EEPROM/NOR/NAND flash) that has been purposely designed for storage endurance over everything. All of humanity's knowledge in an array of EPROM's arranged in order of necessity to survival. So in the event it is physically damaged, all the knowledge about pokemon and harry potter gets lost first, while the information on material sciences, chemicals, pathogens, and so forth are stored in the most protected/armored parts.

 

Like at a practical level, there is no physical material out there that would survive a nuclear, biological, or "zombie" apocalypse scenario. You would literately need a 100% entirely self contained (as in powered, processing, ram, display, audio output) computer that can be used even after being stored for 10000 years. Right now because commercial interests are not aligned on this, the most durable material we have is in fact acid-free paper. Basically printed books, that would likely have to be kept sealed in airless vacuum sealed vaults. UV light will erase any printed material exposed to light over time.

 

Optical discs and magnetic media have zero lifespan, because they are not stable at outside room temperture, and the drives themselves require moving parts that wear out and require knowledge of how to repair. Good luck fixing a cd-rom drive any more than a 5.25" floppy drive. Remember the zip disk click of death? Same reason why mechanical hard drives are not viable long term storage.

 

So where are we at:

- Magnetic media can be erased by accident, stray magnetic waves, improper storage. Drives are impossible to replace

- Optical media can be scratched, or rot in storage, drives are impossible to replace

- NV/EEPROM can be erased from energy pulses, or decay over time (about 10 years)

- Paper burns, gets water damaged, physically gets eaten by bugs, and rots. But paper is infinitely easier to copy by hand, which leads to transcription errors in making copies. Remember like most of the people employed by churches before the advent of the printing press were just people making copies of bibles or other religious scripture by hand. That is the only reason why it still exists while the only other written information we've been able to recover is written in literal stone and clay.

 

Imagine trying to have a stone wikipedia. each stone tablet weighs a kilogram and stores about half the amount of text as it would had it been on paper. Wikipedia is 63,337,468 pages big. (63.3 giga-grams... had to actually find a measurement) like I can not even find a physical object on earth to compare that to, but that makes it heavier than the great giza pyramid.

 

Maybe it's not unreasonable if some kind of stone-archive was built into a "pyramid" like such, but I think we'd have to cull an awful lot of material AND also write it in something like unicode pictograms for 1 glyph per word so that it isn't dubiously interpreted, but also far more concise. So essentially re-inventing the ancient Egypt hieroglyphics.

 

[Bitter]

Laser etched into gold plates stored in a former mine in a geologically stable area of the earth at high enough above sea level that even all of the ice melting would not flood it. Written in the 5 most spoken languages on earth along with a 'rosetta stone' showing how to translate from one language to another in case parts get damaged from one language but not the others. Gold is stable long term. Aluminum might also work. Stainless steel could work too probably. Audio could be written like a record. Data could be stored on metal discs like masters for CDROM's so that copies could be made in the future. So on and so forth. 

[Kisai]

5 minutes ago, Bitter said:

Laser etched into gold plates stored in a former mine in a geologically stable area of the earth at high enough above sea level that even all of the ice melting would not flood it. Written in the 5 most spoken languages on earth along with a 'rosetta stone' showing how to translate from one language to another in case parts get damaged from one language but not the others. Gold is stable long term. Aluminum might also work. Stainless steel could work too probably. Audio could be written like a record. Data could be stored on metal discs like masters for CDROM's so that copies could be made in the future. So on and so forth. 

Gold is stable, but is too soft for long term storage. There's also the risk that some future generation doesn't know what the gold plates are and melt them down into jewelry much like what has happened with the last several centuries.

 

Like it seems to me that some kind of etching is the right direction, but it might have to be into materials that would not tarnish/oxidize, but also not be seen as worthless. Like as much as we talk about forever chemicals being bad, perhaps that might be what we should use as the "ink" on something to fill the etching in on another material that is difficult to degrade or break, like a graphene material or something.

 

[Bitter]

54 minutes ago, Kisai said:

Gold is stable, but is too soft for long term storage. There's also the risk that some future generation doesn't know what the gold plates are and melt them down into jewelry much like what has happened with the last several centuries.

 

Like it seems to me that some kind of etching is the right direction, but it might have to be into materials that would not tarnish/oxidize, but also not be seen as worthless. Like as much as we talk about forever chemicals being bad, perhaps that might be what we should use as the "ink" on something to fill the etching in on another material that is difficult to degrade or break, like a graphene material or something.

 

https://en.wikipedia.org/wiki/Voyager_Golden_Record

Gold is too soft for storage? How? It's just sitting there, it's not like wax that melts. 

[Kisai]

1 hour ago, Bitter said:

https://en.wikipedia.org/wiki/Voyager_Golden_Record

Gold is too soft for storage? How? It's just sitting there, it's not like wax that melts. 

That is a gold PLATED copper record.

 

Also, link the source. https://science.nasa.gov/mission/voyager/making-of-the-golden-record/

Quote

• The record is constructed of gold-plated copper and is 12 inches (30 cm) in diameter. The record's cover is aluminum and electroplated upon it is an ultra-pure sample of the isotope uranium-238. Uranium-238 has a half-life of 4.468 billion years.

 

You ain't doing this with pure gold, because it's too soft. That's why jewelry isn't 24K gold because it will bend and scratch easily. It's the most scratch-prone metal material in existence.

 

Just taking a cue from NASA, you'd probably do micro-dot etchings on a disc with a stable forever-chemical dye on a non-reactive graphene surface. So you would need only a microscope to read the data, and a microscope could be assembled at any point in the future from glass, or a digital camera, or something along those lines.

 

Either way, the main obstacle to preserving knowledge is the inability to make a material that would survive more than a century. The secondary problem is recalling the data after the device used to read it no longer functions. Hence my earlier comment that if you wanted a digital data (Eg an AI or LLM) to recall the data, you'd need an entire self-contained computer to play it. That's why it doesn't work past a few years. Whatever you do to store data indefinitely needs to be analog, and needs to be readable by the human eye.

[05032-Mendicant-Bias]

7 hours ago, CuriousBread said:

How'd you build humanity's last library as an LLM that contains the entirety of human knowledge with what you have? It needs to be easy to power and rugged. 

Instructions:

1. Windmill near a river, get around 3KW of power
2. Build a NAS with the highest endurance drives you can find. As little as 20 TB worth of drives will do it. Get 100TB worth of spares.
3. Get an offline copy of wikipedia
4. Get an offline copy of wikipedia
5. Get an offline copy of wikipedia
6. Get a copy of every patent ever (not sure how much that takes)
7. Get python. ALL of it. You'll curse your ancestry when pip install no longer works.
8. Now you can start getting LLM. NAY get ALL the LLMs 
9. Find out you backed up knowledge but not food
10. Congratulation, you have now become a doomsday prepper!

 

P.S. Github did a cold store in 2020. 

 

Quote

The AWA (Arctic World Archive) is a joint initiative between Norwegian state-owned mining company Store Norske Spitsbergen Kulkompani (SNSK) and very-long-term digital preservation provider Piql AS. AWA is devoted to archival storage in perpetuity. The film reels are stored in a steel-walled container inside a sealed chamber within a decommissioned coal mine on the remote archipelago of Svalbard. The AWA already preserves historical and cultural data from Italy, Brazil, Norway, the Vatican, and many others.

 

[Spindel]

Why the F build a LLM for this?

 

Why not just an indexed database? 

 

The ML brain rot in people is real...

[Dedayog]

14 hours ago, CuriousBread said:

The issue is CD-R degrades after 5-10 years. Commercially pressed one last longer but that's not quite a STC. More so a big digital encyclopedia 

No, it doesn't.  I have burned discs I still use 20 years later. 

[Kisai]

4 hours ago, Spindel said:

Why the F build a LLM for this?

 

Why not just an indexed database? 

 

The ML brain rot in people is real...

Cause people think a LLM will evolve into an AGI, and it will not.

 

Like if you even think about the idea for more than 2 seconds, you realize that prior to the internet, encyclopedia Britannica was the go-to. It's been published since 1768. It stopped printing the physical version in 2012. Why would you keep mass printing a version when the internet version is readily available, and the internet exists.

 

If the internet ceased to exist, there would probably be a quick scramble to make print versions of Britannica again. Wikipedia would not be useful as a print edition since many of it's use cases requires the internet. There is no index book for wikipedia. There is for Britannica (or at least there was.)

[CuriousBread]

5 hours ago, Spindel said:

Why the F build a LLM for this?

 

Why not just an indexed database? 

 

The ML brain rot in people is real...

The entirety of human knowledge in a library without a know how to where to access and how to get into it is useless. 

 

It needs to be structured and LLM can process that monolithic behemoth data into biteable chunks. 

 

A LLM's benefit over raw data is it can lay out the data from simple to in depth. 

[Bitter]

https://en.m.wikipedia.org/wiki/Wikipedia:Terminal_Event_Management_Policy

[Biohazard777]

36 minutes ago, CuriousBread said:

The entirety of human knowledge in a library without a know how to where to access and how to get into it is useless. 

 

It needs to be structured and LLM can process that monolithic behemoth data into biteable chunks. 

 

A LLM's benefit over raw data is it can lay out the data from simple to in depth. 

How exactly do you think we got where we are today 

[Kisai]

3 hours ago, Bitter said:

https://en.m.wikipedia.org/wiki/Wikipedia:Terminal_Event_Management_Policy

It's interesting to note that they basically came to the same idea I mentioned above.

 

 

The thing is, realistically, there are four scenarios you have to consider:

a) imminent disaster. Let's say... oh an EMP pulse from somewhere that just renders all electronics made after 1970 non-functional.

 

To quickly describe an EMP, is that it's a high energy burst that turns everything with metal in it into an antenna and has nothing to dissipate the energy, so that in turn burns everything with wire in it. There is no escaping a EMP, everything that is exposed to the atmosphere just becomes useless in seconds.

 

If you had not a offline, printed, copy of wikipedia on paper at that point, it's gone. This is also the most likely scenario where the continuity of civilization is over. At best we lose the last three centuries of progress. At worst, we end up back in the bronze age as all of industrialization has to be re-discovered.

 

b) impending extinction event, something like yellowstone erupting that makes significant parts of the planet uninhabitable in a matter of years. The internet wouldn't be gone immediately, but it's likely half the planet would just be gone in minutes. So saving wikipedia at that point is futile if there aren't digital copies on every continent, including Antarctica. An impending ELE, would require data to be stored, quite literately, off planet. So for this reason, as silly as it sounds, we need to send a printed version of wikipedia, to the moon and mars. Just we haven't even mastered materials that would survive years on another planet yet, never mind one that needs to survive a planet routinely pelted by meteorites.

 

c) Impending, man-made global disaster, so thermonuclear war or climate change-triggered event. So "Fallout". Some parts of the world will be spared, but the rest of the world will just be unlivable for centuries, or maybe even 10000's of years. It's not unreasonable to assume that let's say Antarctica, maybe Australia and New Zealand might end up being the only places not rendered unlivable. All manufacturing that is global, would just be over, so if you haven't printed wikipedia at that point, good luck printing it when the supply of ink and paper are gone.

 

d) slow-crawl disaster. So the most plausible version of this scenario is that oil and coal are banned from being used as energy sources, and as a consequence there are no backup energy sources when something like a nuclear plant fails (see fukushima) which amplifies the disaster to a point where it has a amplifying feedback loop where power failures start happening on continental scale.  It may eventually recover, but it could be months until power is restored to regions that had no hydroelectric facilities to "warm start" other power plants (eg nuclear.)

 

Like no matter which way you consider it. Having a computer, doesn't factor into the preservation of our knowledge. There is no scenario where a computer made today would still work 20 years from now and still be able to read the data stored today, because the mechanical parts (eg HDD's) would be worn out, and materials that makes up EEPROM/NOR/NAND flash would certainly be dead. Likewise the helium in current HDD's would escape, so the drives would not be usable in 20 years either. We do not presently have any digital storage that can be read randomly, indefinitely. Optical and Magnetic storage have short shelf life, and the drives needed to read them have even shorter lives.

 

Like what we need is permanent "flash" like storage that can be written to, even if super slowly, and then blow the fuses on the write-channel so that the result is permanent, and can not be erased after that point. But it also needs to not be damaged from EMP. So that means hardening and sealing the media from being electrically exposed unless physically in the "drive." In a sense, it would be re-inventing the floppy disk disk cover. Good luck when the trend for flash is lower durability and smaller physical sizes when it should really be going in the other direction. Wikipedia, without images is 24GB, compressed. That can be stored on a 32GB USB flash drive at present (probably around 64GB uncompressed.) 645TB when you include all media.

https://en.wikipedia.org/wiki/Wikipedia:Size_of_Wikipedia

 

https://commons.m.wikimedia.org/wiki/Special:MediaStatistics

 

So, we need a 1PB drive, currently, if we wanted to save everything but the edit history.

 

Then there is this:

So "how much" is worth saving?

 

So, off-the-cuff, what is actually valuable to preserve:

1. Science

2. Biology Health and Medicine

3. History

4. Business

5. Geography

6. Culture and Arts

7. Society

8. Biography

 

Of which saving science and biology is the most important, because without that, there is no starting point from which to rebuild from.

 

Everything else is useful but not necessarily important to save if you don't have time to save it. 1000 years from now, the Biography, Society and Culture stuff might not even even be interesting to all but a handful of people. Do you know anything about ancient Egyptian culture? Probably not. But the sciences are still in use.

[Spindel]

16 hours ago, CuriousBread said:

The entirety of human knowledge in a library without a know how to where to access and how to get into it is useless. 

 

It needs to be structured and LLM can process that monolithic behemoth data into biteable chunks. 

 

A LLM's benefit over raw data is it can lay out the data from simple to in depth. 

And a library with the help of a LLM is still useless if you don't know how to access the LLM it's a null point. 

 

You act like all the knowledge isn't already structured and you are wrong. Every piece of knowledge recorded up until like 2 years ago was recorded and structured entirely without an LLM and hasn't simply disappeared suddenly. All human knowledge isn't simply just a database dump it is already structured in bitable chunks one example is Principia Mathematica that was written in 1687 long before computers even was a thing. 

 

I stand by my point the ML brain rot in people is real!

[Spindel]

18 hours ago, Kisai said:

Cause people think a LLM will evolve into an AGI, and it will not.

 

Like if you even think about the idea for more than 2 seconds, you realize that prior to the internet, encyclopedia Britannica was the go-to. It's been published since 1768. It stopped printing the physical version in 2012. Why would you keep mass printing a version when the internet version is readily available, and the internet exists.

 

If the internet ceased to exist, there would probably be a quick scramble to make print versions of Britannica again. Wikipedia would not be useful as a print edition since many of it's use cases requires the internet. There is no index book for wikipedia. There is for Britannica (or at least there was.)

It's scary how (mostly young) people use LLMs to look up things. If you are the least knowledgeable in any topic you see that the answers most LLMs give you are straight up wrong. They are well worded (which can make you think the answer is correct) but they are still wrong. 

[05032-Mendicant-Bias]

5 minutes ago, Spindel said:

And a library with the help of a LLM is still useless if you don't know how to access the LLM it's a null point. 

Say what you will, but LLM are great to just chat with. It doesn't get more untuitive than that. Even better you can install a STT and a TTS and literally speak with it.

 

My go to approach to learn a new library is to use an LLM after all. Then I delve on my own on the docs and doing it properly.

 

Storing an LLM alongside wikipedia would unironically be great if you plan to start something. "My plants arent growing, why?" "There can be many reasons plants arent growing, such as...." Then you follow the breadcrumb and look into the specific knowledge and technique you are looking for.

  

1 minute ago, Spindel said:

answers most LLMs give you are straight up wrong.

What are you asking LLMs? If you ask for news, they get it wrong because that's not how they work. But if you ask general knowledge that is not really news, they work reliably well.

[X-System]

On 6/17/2025 at 1:23 AM, CuriousBread said:

The issue is CD-R degrades after 5-10 years. Commercially pressed one last longer but that's not quite a STC. More so a big digital encyclopedia 

No all CD-R, I have old CD-R 650MB Maxell Gold from 1996-2000, they work perfect.

[Kisai]

5 hours ago, X-System said:

No all CD-R, I have old CD-R 650MB Maxell Gold from 1996-2000, they work perfect.

Those are still stored on organic dyes, and the label side getting even a tiny scratch from accidently dropping it, would damage the entire disc.

 

Like the reason why CD-R and DVD-R are not considered archive quality is because the vast majority of the media made is made cheaply to be disposable.

 

https://en.wikipedia.org/wiki/M-DISC 

Quote

In 2009, testing was done by the US Department of Defense (DoD) producing the China Lake Report^[22] testing Millenniata's M-Disk DVD to current market offerings from Delkin, MAM-A, Mitsubishi, Taiyo Yuden and Verbatim with all brands using organic dyes failing to pass the series of accelerated aging tests.

https://archive.org/details/millenniata-archival-dvd-m-disk-china-lake-full-report-november-10-2009

 

 

Again it has to be said that ODD's will not be functional even if the discs survive.

 

[SpaceGhostC2C]

On 6/16/2025 at 7:23 PM, CuriousBread said:

The issue is CD-R degrades after 5-10 years. Commercially pressed one last longer but that's not quite a STC. More so a big digital encyclopedia 

CD-Rs don't have 125TB capacity, though. @danalog may have something like M-Disc in mind - those will not degrade in 10 years.

[danalog]

4 minutes ago, SpaceGhostC2C said:

CD-Rs don't have 125TB capacity, though. @danalog may have something like M-Disc in mind - those will not degrade in 10 years.

neither has any CD I've owned... sure they might last 10 years if you leave em outside but the newest decayed discs I've found have been early 00s

[CuriousBread]

On 6/18/2025 at 1:32 AM, Spindel said:

And a library with the help of a LLM is still useless if you don't know how to access the LLM it's a null point. 

 

You act like all the knowledge isn't already structured and you are wrong. Every piece of knowledge recorded up until like 2 years ago was recorded and structured entirely without an LLM and hasn't simply disappeared suddenly. All human knowledge isn't simply just a database dump it is already structured in bitable chunks one example is Principia Mathematica that was written in 1687 long before computers even was a thing. 

 

I stand by my point the ML brain rot in people is real!

I mean as in useable form. 

 

Yes what Pi is is explained in Wikipedia, however how will you structure the mathematical knowledge in a way that is teachable to the next generation? 

 

Textbooks are good for that. 

 

Wikipedia is not. 

 

LLM can make textbooks of knowledge