/tech/ - Technology

>>33914
>>33920

destructive book scanning is already an established practice for digitizing print books and it's the only way to digitize a book and get a high-quality flat scan of every page with no shadows, companies have been doing it for years and they don't do it with expensive rare antique books. this is just a sensationalized regurgitation of an old washington post story, probably recycled by 404media or futurism, that's all they do nowadays is churn out enormous amounts of AI doomsday critihype clickbait, probably with the aid of LLMs.

they make it seem like AI companies are destroying all the books like brave new world or something, but there's like hundreds of billions of copies of books in the world and they only need to scan one copy to get a digitized copy. it's probably a good thing actually, most books in the world have not been digitized and they're all disintegrating and becoming unreadable, pretty much any book printed before 1996 was printed on acidic paper that disintegrates within a century and only a tiny fraction of all the books in the world have been digitized, probably less than 10%. most of it will never be digitized and will just be gone forever.

>>33919
>LLMs have plenty of bearing, they are going to erode what little public trust is left in digital text content which will collapse the entire blogging/digital journalism industry.
"blogging" as it currently exists mostly works around name recognition, worst case scenario "clickbait celebrities" (if theyre even real and not just an autistic fixation from your part), will just transparently move to tiktok which literally has 0 editorial standards.

>>33921
if they couldn't be bothered to write why should the reader be bothered to read?

If you havent maxed your claude limit by afternoon you have wasted your day.

>>33932
you know who is routinely maxing out their claude subscription? the chinese AI labs actively distilling this shit

ok i bring good news for lord autismo, hater of clickbait

How exactly is LLM-based AGI supposed to be a thing when some of the crucial functions of existing "intelligent" AIs are actually ad hoc tool calls that are by nature not generalizable?

>>33924
>it's the only way to digitize a book and get a high-quality flat scan of every page with no shadows
No dumbass. It's preferred simply because it's cheaper and faster.

The bigger problem is that these companies aren't sharing the digitised versions of the book. Ostensibly this is caused by copyright law but I wouldn't be surprised if it's to deny their other AI competitors from training on them. Old books basically turn into one-time use training material.

I love the idea of future where there is no physical books because they have been burned. Reminds me of a book but I can't remember it's name because I dont own anything haha

>>33950

go find a book in your house right now (if you have one) and open it and you'll find that you can't make the pages lie flat while they are bound together in the book. to properly digitize a book you need a flat 2d photographic scan of every page with no distortion, otherwise optical character recognition will not work.

>The bigger problem is that these companies aren't sharing the digitised versions of the book.

they legally could not do that even if they wanted to. it's illegal to digitize a book and publish the digitization if you don't own the copyright. but it's not illegal to digitize a book and train an LLM with it. they're just doing the same thing they are doing with copyrighted digital content on the web, scraping everything off the public internet and putting it all in a giant blender that grinds it all up into a fine paste and distills it into an LLM, where there is no feasible way to trace the appropriation back to the original copyright holders in a court of law. the physical books aspect is just a small part of this and you really need to get with the times and get past this 2000s mindset where you see the internet/digital information and physical reality as two completely separate worlds.

in essence, big tech have figured out how to leverage their tremendous wealth and influence to develop a technology that effectively nullifies the idea of intellectual property and they are scrambling to hoard as much data as they can before the rest of the world catches on to what is happening. they want to create a world where instead of governments and laws and courtrooms controlling access to information, it will be a few megacorporations who hoard all of the information and lock it up in their heavily guarded vaults and their LLMs function as one-way subscription-based slop dispensers that spit out an inferior, inhuman, and often-wrong regurgitation to the subscriber without providing access to the genuine information, there will be very strict access controls and deliberately crippled performance for ordinary subscribers and only a select few extremely wealthy people/organizations will be able to pay for unrestricted access and high-quality output and their access will be very closely supervised to prevent any funny business.

why are we letting book store owners (petite bourgeoisie parasites) off the hook on this one. if they wanted to preserve their stock of rare books, they would not sell, or they would donate the books to a public library or whatever.

>>33956

why are we pretending that we really really care about physical books all of a sudden

>>33957
Because we had no real reason to fear or care that much for them before?

>>33953
>you'll find that you can't make the pages lie flat while they are bound together in the book
Yes but there now exists tech that corrects for that and allows you to scan non-destructively. The same company that Anthropic is relying on offers that but it's more expensive to do it. That's it. You're acting like it's an unsolved problem even though the book-scanning company itself offers it.

It's great that it's free and CLI, but why is it so shit?
https://github.com/anomalyco/opencode/issues/16697
Want to switch to https://continue.dev
But can't get Cloudflare API (Kimi-K3) to work.
Don't suppose anyone he has had success with this?

>>33966
Lol, took two days, but finally got it.
The trick was to use a provider string "azure" which hasn't upgraded to the /responses API.
Then just set the apiBase, and apiKey to the appropriate string.
Free CLI AI that hopefully doesn't make me swap!

>>33966
i've used continue.dev for a bit and it was abominable, that was roughly 6 months ago, opencode is way better, but i dont use it for sessions that are long enough for memory leaks to be a big issue. at the end of the day all these agent interfaces are vibe coded and are going to be kinda shitty

>But can't get Cloudflare API (Kimi-K3) to work.
should work with the OpenAI API?

>>33974
>OpenAI API
That's what I'm using but it's an outdated version of the API. There's a provider called "cloudflare" that you'd really expect even more to work, but neither that nor "openai" do on account of expecting the newer /responses API which is only on some of the more advanced models if I understand correctly. Maybe Kimi-K3 was one of these but there was also an issue with this model in particular not being available for free. It was a pain to setup, and I don't even have my AGENTS.md rewritten yet.

>i dont use it for sessions long enough
This wasn't really the issue… I'm typically only using a readonly agent, and I've hit swap within three queries.

>all these agent interfaces are vibe code
Would love to see a "suckless" agent with a NO-AI.md, and some _sane_ approach to security.

>>33977
>and I've hit swap within three queries.
well i also have 64GB of DDR4 I got like 10 years ago i guess

Trying to think about LLMs as completion engines, from a training set to convention.
Have been looking for a lexicon to describe software patterns for LLM feedback.
The Gang of Four is practical, but the wounds of OOP bleed out all over the pages.
It seems nearly everything can be replaced by things like discriminant unions, or higher-order functions.
And its made much simpler in the process of so doing.
Recalls: https://mishadoff.com/blog/clojure-design-patterns/
Thought to use just this as a source of the lexicon, but it seems so inessential, and in translation.
Do I have to make the leap into category theory to get something that is essential?
https://blog.ploeh.dk/2017/10/04/from-design-patterns-to-category-theory/
This might be too much to learn at the moment…

why don't we have claude fable hammer away at the economic calculation problem until it works

>>33990
well I don't have the credits for it

>>33954
>a few megacorporations who hoard all of the information and lock it up in their heavily guarded vaults
This isn't really a concern. The shadow library datasets that LLMs are trained on are widely mirrored and can't really be taken down.

>>33924
>destructive book scanning is already an established practice for digitizing print books
Yes.
>and it's the only way to digitize a book and get a high-quality flat scan
If you got a couple millions to spare for such a project you surely can come up with something. Like (talking out of my ass now) making three photos in very quick succession, first and third projecting a grid onto the pages and then taking the middle photo and use an average of the two grids to flatten the photo digitally. I guess one grid photo is enough if the time between the photos is really short. I also believe that it is not necessary to open the pages 180°, half of that will do if the setup has separate cameras for left and right page. And surely something else than a human hand can turn the pages. ("But Google Books…" - Google is not run by infallible people.)
>and they don't do it with expensive rare antique books
If you process truckloads of books you will make mistakes. What seems an unremarkable mass-produced book can be worth quite a lot if it is a rare edition (or indeed unique because a famous person put some comments into it).

>>33959
>there now exists tech that corrects for that and allows you to scan non-destructively. The same company that Anthropic is relying on offers that but it's more expensive to do it.
Figures. I wonder how much of the price difference is functional pricing.

>>33996
>If you got a couple millions to spare for such a project you surely can come up with something.

the way that rare books are non-destructively scanned is they are secured on a v-shaped platform and someone has to manually flip through the book and take a photograph of each individual page from an overhead camera, and then manually correct the distortion for each individual image. books come in a variety of shapes and sizes and each page of a book bends and distorts differently depending on its position in the book, there's not really a way to automate something like this.

>>34026
Projecting a grid, software automatically transforming the grid pattern to flat.

lol

>>34032
This is supposed to be the AI company that's the least likely to be subsidising their prices btw.

The specialization in fake property has reached its ultimate contradiction in the manufacture of fake property to undermine all fake property.
Yet the post-industrial economies are still so intent on manufacture of this fake property that they can't realize they're destroying their only basis.
This chasing of super-profits in the long run is going to wipe out the profitability of the entire fake economy.
Is the relative productivity of the real economy really high enough to justify the difference in real wages?
I don't think so.

>>34037
40%+ of US GDP is from IP intensive industries, it's a paper tiger, literally paper.

Pangram 4.0 has almost completely solved the false positive problems with earlier AI detectors. It even works on slop code (but I doubt you'd need to use it for that since it's obvious).

False negatives are still a statistically significant thing though. But an incompetent fake getting through is orders of magnitude less worse than an innocent being accused of being a fraud.

>>34035
Yeah that's what's so funny. I'm actually willing to bet money that the next major release of deepseek will be a closed weight model. Maybe inference is cheap, but training is not and no one has figured out how to recoup losses.

>>34061
Making your model open weight would actually make it cheaper for you since some fraction (extremely small but whatever) of people will run it themselves on their RTX 6000s (consumer version of fucking Blackwells) instead of groyping your servers. You only do it if you're a "frontier" corpo who does the heavy lifting of making bigger models and want to protect that effort.

Also DeepSeek is the OG open weights player so it's extremely unlikely they'd go closed weight.

from what i know about hardware and all i've learned about what is being marketed as "AI" and the actual technology behind it, i'm realizing what a bunch of smoke it really is. the transformer architecture, this supposedly ingenius design developed by the braintrust at google, is just a stupid software kludge, a simplification of a pre-existing rudimentary neural network model to make it just barely possible to produce a convincing demo to show to investors by burning up lots and lots of gpus and wasting enormous amounts of energy. it's not a new technology, it's not a breakthrough; it's a marketing ploy. the fact of the matter is, computer hardware as we understand it, i.e. a sequential processing unit connected to a separate memory bank, is fundamentally incompatible with the concept of neural networks, and for parallel processing in general. a graphics card is just a cpu with a lot more cores shoved into it and a faster memory bank, that's it. we are trying to mow football fields with toenail clippers. there is no software trickery that can get around the basic laws of physics, to do any serious development with neural networks and machine learning and achieve the kinds of things that the AI hype wizards are talking about, we have to throw out almost everything we know about computer hardware and start over from square one. and the united states are not going to be the ones to do this, it's going to be china and it already is china who are doing all the important research and development in this area. that's what happens when you outsource your entire high-tech engineering sector overseas and trade a golden goose for a temporary supply of cheap laborers. anyone who really understands how computers work at the hardware level can see this and it's kind of amazing to see people who are highly respected academics getting sucked into the AI hype and prophesizing about the end times, how fine the line really is between science and religion, between technology and idolatry. i stumbled across a blog by this fucking guy named scott aaronson who is a prominent researcher in quantum computing, has a phd in computer science from uc berkeley, was a professor at mit, worked at openai, etc. and is also associated with the lesswrong neoreactionary cult, and his blog is just fucking nonsensical gibberings about the looming threat of artificial general intelligence and how it's going to happen any day now and change the world forever. the whole silicon valley techbro culture is full of these obsessive narcissists who specialize in one extremely narrow field of knowledge and are utterly clueless about anything else, completely up their own asses in their abstract theories and conjectures yet meanwhile have probably never even touched a fucking soldering iron in their lives, don't even know how the electricity works in their house. and you know they are spending hours and hours talking to these chatbots getting automatic validation for all their unhinged thoughts and going further off the deep end. that's the real looming singularity/apocalypse that these people should be worried about, is the erosion of their own sanity from staring too long into the world's most expensive vanity mirror.

If AI so good why don't the use it to solve climate change instead of hacking random companies

>>34065
>how fine the line really is between science and religion, between technology and idolatry.
much to ponder here, not for me mind you, i'm really stupid. but scientists think they have neatly compartmentalized science from religion, or even ideology in general for that matter and it turns out, not only have they not, they're insane zealots.

>>34063
>Making your model open weight would actually make it cheaper for you since some fraction (extremely small but whatever) of people will run it themselves on their RTX 6000s (consumer version of fucking Blackwells) instead of groyping your servers.
Yes but you also are, in essence, subsidizing training and research costs to your competitors and compute providers.

>>34068
>Yes but you also are, in essence, subsidizing training and research costs to your competitors and compute providers.

"in essence" is doing some very heavy lifting there. that's like saying flowers are subsidizing food costs to bees by allowing them to harvest their pollen. innovation is symbiotic, it requires free sharing and copying of ideas, otherwise it dies just like flowers die when there no bees to spread their pollen and bees die when there are no flowers to eat pollen from.

if the wright brothers had allowed other people to improve on their ideas to design and build better airplanes then they could have taken their improvements and used them to build better wright brand airplanes, but instead they just became patent trolls and sued anyone who tried to build an airplane. where is the wright company now? where are all the supersonic metal-body mono-wing jetliners with the wright logo on them? this is why things like capitalism and intellectual property are retarded.

>>34070
>"in essence" is doing some very heavy lifting there.
it is not. it describes what is happening.

>where is the wright company now
it merged with curtiss and exists as curtiss-wright today, it publicly trades as CW in the NYSE. i feel like you could have chosen a better example, maybe? you seem confused about what businesses do and seem to have deluded yourself into thinking high flyer might be above profit motive.

>>34070
>that's like saying flowers are subsidizing food costs to bees by allowing them to harvest their pollen. innovation is symbiotic
>this is why things like capitalism and intellectual property are retarded.
I don't want to samefag but your post is so baffling because it's the most egregious example I've seen this week of someone confusing what is vs what it ought to be. I feel there are more and more people like you on this board, posting these sort of confused tiktok comment section illiterate pabulum that adds nothing.

>>34072
>I feel there are more and more people like you on this board

yeah that's what happens, time passes, events unfold, old people leave, new people arrive, attitudes change, cultures evolve. all this dynamism is kind of the point of reading online discussion boards, if you prefer static text content then maybe you should unplug your router and read a word document or something.

>>34063
>RTX 6000s
You don't need an RTX 6000 to run open weights models. They come in all sizes and can be run on any computer. The smaller ones are just dumber.

>>34075
high level pabulum here

>>34076
I'm aware but I'm talking about the newest DeepSeek models here.

PSA: the only "non-corporate" AI cybersecurity scare report, which people pointed to as evidence that it's not all marketing stunts, turns out to be a sloppy work produced by "Effective Altruist" ideologues.

https://archive.is/20260812011323/https://www.telegraph.co.uk/gift/23731f653ebf2657

>However, the AISI had made some very strange decisions in this regard.
<It had disabled safeguards on purpose, left the systems attached to the internet and then let the AI software agents do their worst, while pretending not to notice.
>The AISI also allowed access to the Tor anonymous network, which let the AI access the dark web.
<“No enterprise network I have ever been on has allowed Tor. It is always disabled. That’s a big clue that this is a setup,” one expert told me.
>In response to the “rogue” event, the AISI produced its first ever “incident report” – formatted up to look and sound like a formal industry document.
<“It’s an ethically challenged publicity stunt,” says one highly experienced senior security expert who has led the development of global technical security standards.
>Nor is this the first time the quango has arguably tried to hoodwink us. When Anthropic sought help to promote Mythos as a dangerous cyberweapon earlier this year, AISI obligingly conducted a test on targets with their defences disabled.
<When Rishi Sunak hosted a summit on AI in 2023, the EAs quickly captured it and turned it into a vehicle for their hobby horse. “Three of the four partner organisations on the lineup are bankrolled by EA donors,” Politico UK reported at the time.
>One of their recommendations was the establishment of a new quango called the AI Safety Institute – the “Security” is a recent rebrand.
<Now a question: how do you think Jade Leung, the Prime Minister’s official AI advisor, reacted when she heard about the AISI’s latest amateurish stunt? Did she spell out the ways in which the testing fell short?
>That’s not likely, because Leung is not only No10’s official AI advisor but also the AISI’s chief technology officer – a clear conflict of interest.
<She is also a senior and long-standing member of the Effective Altruist movement. Leung received $180,000 in 2020 from the Centre for Effective Altruism, set up by William MacAskill, the EA founder, and then worked at his Future of Humanity Institute in Oxford.

>>34106
>next token prediction but marxistly
Waow. New frontiers in ML (no pun intended) theory.

i saw a video on titkok of some guy trying to cut small restaurants some slack because they use AI for their ads because "they don't have art departments" like chipotle or whatever, and the comments are 99% "small business suck", and small business owners desperately fending for their life. 5 years ago this would have been unthinkable, everyone would be like "yeah but nestle is worse" or whatever. chat, is luddite sentiment bad? leftypol led me to believe that it's reactionary but it seems that AI hatred has been extremely didactic for some people.