How long until surveillance?
Posted by boloshon@reddit | LocalLLaMA | View on Reddit | 63 comments
I'm sharing something worrying me, hoping to get answers saying "no it's impossible, you silly" (or similar).
I just read that reddit was asked to provide the identity of a user because of Ice criticism.
In France, one guy shared his will to become a terrorist with chatGPT and got the police (RAID) knocking at home.
Historically, teens have been arrested for trying to reproduce what they saw on Mac Gyver.
So it's a matter of time, if not already that journalists realise there is something like uncensored local LLM allowing to do whatever you want, without control, and that someone was helped by it for a crime or self-arm.
There is also more and more websites asking you to prove your identity (scanning face, and ID) first pron websites, then aliexpress etc.
So my guess it's that one day, one gov will discuss the danger of local llm "thanks" to a commited crime and try to "regulate" them. Making it illegal for someone or a gafam to release a too smart LLM above a certain quant for example.
When I see that HF is planning to create an opensource version of Glasswing, it means that it might trigger a race for people fixing software against people exploiting them.
In a dystopian view, people would exchange usb keys of LLM, in a dark street, wearing trench coats to avoid AI act police
Maybe I went too far thinking about it. One would say there is always open-source which will never be "regulatable" like piracy hard to catch.
But what's your take on this?
jacek2023@reddit
If I remember correctly Anthropic is already trying to attack open source, at some point other forces will join. I believe the first attacked place will be huggingface. So again I want to say that centralisation of models is not a good thing.
JayPSec@reddit
Centralization is never a good thing. Could not agree more.
Maybe the future is sharing LLMs on covert Newsgroups (I'm oldish :P)
sasquatch3277@reddit
I wonder if people will use torrents or other p2p tech or federated platform to share model weights. Safetensors is a thing. It's still somewhat of a black box.
I assume the motivation to replace weights with compromised ones trained to immediately exfiltrate your secrets is already high and will increase significantly over time
Responsible_Buy_7999@reddit
You should read the Mythos system card and alignment risk update. They’re enormous documents demonstrating significant effort to understand the danger of what they made.
Some people, with access to the raw model, would immediately ablate away any protections there may have been, effectively throwing all that, I have to say entirely sensible and desirable, care in the trash.
There’s a theory that the reason we don’t hear from aliens is because by the time they hit a certain threshold where they develop AI and nukes, they destroy themselves before achieving warp drive or figuring out cryogenic travel. Exhibit A in favour of this is the people who want open source ablated models of frontier capabilities.
Individual_Spread132@reddit
because the space is ridiculously vast and we can't see shit even in our solar system
think of it this way: there were thoughts of Mars having some form of alien life up until the moment we actually landed there, observing and analyzing the environment
when you see them talk about some distant 'exoplanet' they claim to be special, yapping about its atmospheric contents and whatnot - it's all inaccurate, based on whatever whimsical methods of "detection" they have (usually amounting to looking at a few pixels of some distant star's image being dimmed in some way, when - supposedly - a planet is passing in front of it; they derive their 'observations' from that, they can't even see the planet and they have no way to confirm whether their guesswork is even 1% correct)
llama-impersonator@reddit
spectroscopy is actual science, why are you here?
Individual_Spread132@reddit
Did I deny it being a science? I said it lacks accuracy and can't be used to determine whether a certain planet supports life or not. They search for specific signatures that would be considered 'life defining' in terms of our own understanding of Earth-based life. They can't even tell for sure if the planets they find are rocky or not. One day a scientist claims it's a "super Earth", and soon afterwards it's suddenly a gas giant, according to another lab. For all it's worth, it's the scientists "sciencing for science" and having fun (for salary) with the instruments they've got at this day and age.
llama-impersonator@reddit
now you're just griping about the scientific method. i don't even disagree with your conclusion other than calling this stuff whimsical. exoplanet stuff is still low confidence, as you might expect from a branch of astrophysics that's only been around for 30 years and is looking at tiny things that are super far away. pretty sure most of the people working on this stuff would agree that looking only for earth-based signatures is going to miss countless things, but given how pretty much everything we know about distant objects in space is from applying this process to more visible objects, i think there's still a lot to be gained from continuing to do it.
iLaux@reddit
Well to be honest I have the same concerns as u. Lack of privacy and digital surveillance are becoming a concerning threat to society, but people don't seem to care at all.
I think these things are one of the most valuable points of local LLMs. Being able to handle YOUR data privately. No external server, no government, no advertisers. In a world where even a political view or a comment can get you arrested, I think we have to remember that the future needs to be local and open source!
Radiant-Video7257@reddit
We're far past the point of no return.
boloshon@reddit (OP)
couldn't agree more, open source sometimes seems like a unforeseen glitch/hope
No_Lingonberry1201@reddit
Big companies open sourcing their stuff is what Joel Spolsky calls commonditizing the complement and there's some really good business reasons for it. Basically, as the letter describes, if gas prices are cheap, more people will use cars (like in the USA) while more expensive prices disincentivize car use (like in the EU). So if the stuff they are selling becomes easier to use with an open source product (e.g. terraform making it easier the management and provisioning of cloud resources easier) it makes sense for them to support it.
Disclaimer: I'm not an economist, so I might have not explained it properly and oversimplified it, go read the letter, it's not long and worth it, even 24 years later.
soshulmedia@reddit
It seems many people don't really value privacy nowadays. I think a way to address this problem is to make people aware that they will violate the privacy of others if they use any of the bad actors services.
For all the talk about "caring about others" this is an "interesting" ignorance that people seem to have.
Let's say, for example, someone goes on ChatGPT and discusses a work related interpersonal issue on there like folks nowadays seem to do. He or she automatically also violates the privacy of others if s/he does so, even if using pseudonyms or placeholders for their names.
And any AI in the backend, given enough data (Palantir etc.) can then just correlate it all together to the complete picture again.
There's so much talk and fuzz about "please think of others!1!" when it comes to social and race and whatnot issues. (Too much in my opinion as it isn't any more "love others as you love yourself" if the drive is to get people to bend over backwards for everything and thus stop loving themselves!)
But there's no such thing if people just violate the privacy of others?
Is that maybe simply because TPTB values the idea of "slaves enslaving each other" but does not value "people having dignity and privacy" and people are ignorant wrt. the brainwashing and propaganda that shapes their worldviews?
ReasonablePossum_@reddit
Not much. But they will not get here in some time tho, the main reason for asking ID everywhere is to control people waking up to ISR/IL controlling the West and doing horrendous things in the middle east. iCE episode on reddit is only being used to set a precedent and a system to doxx people in parallel to that.
The good news is that there are decentralized platforms where this will never get to (Lemmy, Mastodon ecosystem) AMD communities are already migrating there to avoid dystopian control.
As for filesharing, P2P will not go anywhere. Torrents, Soulseek, cloud layers, etc are all out there. Same for dark layers for website hosting and upkeep.
Ryoonya@reddit
I would imagine this will happen in Western Europe. Europe is already fighting against Encryption, going after graphene os, going after TOR, and regulations on AI.
My advice is to buy a property from a third-world country where you can bribe and build your own little data Center.
Plenty of countries will be willing to take in people and investment where one country regulates. Which is why you see the middle east get huge amounts of investment for datacenters because europe is being europe.
Imaginary-Unit-3267@reddit
It would unironically be based as fuck if African countries end up leading the world in AI in the future because of dynamics like this.
fuck_cis_shit@reddit
AI misuse risk is a function of compute. the worst offenders aren't going to be your neighbors doing black-hat stuff, but governments, large corporations, or superwealthy individuals doing bad stuff at scale with datacenter compute
Imaginary-Unit-3267@reddit
It's not considered bad by the people in power if rich people, governments, and corporations are doing it though. As evidenced by the fact that human trafficking appears to only be a crime if plebs do it.
Kodix@reddit
I mean, it's *possible*, but it depends on one thing: high-profile incidents.
If there's a terrorist attack somewhere and the government finds that a local LLM was used to give the terrorist specific information about making weapons or planning the attack, then it *could* result in a local or even global panic about it.
Evening_Ad6637@reddit
It’s only a matter of time before the Epstein regime starts claiming that the Iranians used GLM or Deepseek.
If not this, then they will stage a sensational and "terrifying" incident.
printr_head@reddit
Already there.
Embarrassed-Area4652@reddit
Given the direction of the tech industry in general over the last, oh, 32 years or so ( https://en.wikipedia.org/wiki/Clipper_chip was 1993), I think it's safe to assume that unless you have hard protections in place around your data, it's being harvested and put to ill use.
ZenaMeTepe@reddit
Too late. Local LLMs are already good enough that a bad actor can use them for harm. Just like a smart bad actor could use their own brain for harm before.
boloshon@reddit (OP)
I understand and yet, what seemed to be an awesome model a few months ago, now seems like obsolete to me, so I'd like to be sure to be able for the future to keep on using the latest model possible. This theory wouldn't be as extreme as forbiding local models, but is limiting their thinking ability for the public a possibility like it was limited for encryption in the 90's?
koflerdavid@reddit
You were fine using it just a few months ago. Of course it is obsolete in the sense that there is always something new and the grass is always greener somewhere else, but that's just life. Apart from that, the Local LLM users have always been at the mercy of model trainers to publish models since nobody has the compute resources to train a model from scratch by themselves, therefore we always have to be ready for no new models coming our way. But the old stuff is here to stay and already powerful enough to tell us how to build pipe bombs and stuff and how to blow ourselves up in the process lol
ZenaMeTepe@reddit
Any model that can be stored on an USB stick will eventually leak imo so you'll find it on the gray market, eventually.
silenceimpaired@reddit
Didn’t some states in the US already pass laws like this?
Saren-WTAKO@reddit
That would be cool. illegal hackers going underground training and distributing unregulated illegal AI models, sounds like a scifi story
thrownawaymane@reddit
Neuromancer, the book that came up with cyberspace and a million other things that have been ripped off since
Ill-Bison-3941@reddit
😅 I grew up in Russia in the 90's, that's how we were getting our American movies. And before that, it was how my parents were getting American music. Bootlegging can be a godsend.
iLaux@reddit
Yeah. Imagine sharing with someone a USB flash drive full of uncensored models. Like you go to a café, meet with someone and make that exchange. Like a hackers movie.
Although sounds dystopian af.
boloshon@reddit (OP)
There is this artist, I think in Berlin who is glueing usb-stick in random walls in the streets, so that you can exchange files by plugging your computer on it.
kaeptnphlop@reddit
Honeypots lol
boloshon@reddit (OP)
well it's almost the scenario of Akta Manniskor with unhinged bots for example
MrPecunius@reddit
William Gibson-esque black market AI cannot be far away.
It's weird as hell to be living in the future I read about in the 80s when cyberpunk was fresh and new.
hockey-throwawayy@reddit
When will the Dixie Flatline be on HuggingFace?
SocialDinamo@reddit
I had a thought and haven’t been able to shake it, right now every quality and even crap input output are shoved into models. All of our info. One day, when they figure out the black box, how much info of ours we provided will be able to be searched and read like plain text?
woahdudee2a@reddit
you think future tech is needed to read your prompts?
SocialDinamo@reddit
No, more like if the “black box” is a fancy compression algorithm, and one day it can be easily uncompressed
brickout@reddit
Assume literally everything is monitored. Duh.
fallingdowndizzyvr@reddit
If you aren't at least using a non US domiciled VPN when using reddit, then you are just asking for it.
Although, in the near future when you have to show ID to do anything online, it won't matter.
Cool-Chemical-5629@reddit
Oh look who woke up late to the party. Well better late than never.
boloshon@reddit (OP)
What’s your point ? What are you trying to say
Cool-Chemical-5629@reddit
You can take it as my appreciation comment after seeing more people leaving the matrix to see what the real world is really all about.
PurpleWinterDawn@reddit
"It's not that I have something to hide, it's that I have nothing I want to show you." - Anon (movie)
stopbanni@reddit
I am pretty sure that it will be too much time before that. Because running local LLM is still not cheap, not fast, not easy. You still need technical knowledge to get best results from local LLM.
First_Ad6432@reddit
you are being watched by people that have 0 brain cognition
Responsible_Buy_7999@reddit
It might help you to look into the processes at work and consider each case on their own. It’s important to notice that the first case is failing to work for the gov, and the latter ones are entirely appropriate and “urban legend” vague.
The criticism of ice thing was a was a separate process than the one that led to police knocking on the kids door.
There was a leak, long ago, about the policy manuals from various companies about how to serve them with court orders. They list fax machines, contact emails, designate responsible individuals. It is a well oiled machine for giving up user information.
The subpoena is a form of pretty please, it’s not a judge finding there is a potential crime and ordering openAI and google and the ISP to connect the dots from the chats to the user’s home address, which all three companies complied with.
mystery_biscotti@reddit
(Your telecom providers can already produce records of SMS chats going back ten years. There are rules the telecoms follow before releasing those documents. Lots of steps, lots of checks, always required a human at a specific level to approve.
The place I worked for only agreed quickly in one type of instance: abductions. But that was a single telecom, about seven years ago.)
JackStrawWitchita@reddit
boloshon@reddit (OP)
Ahah I love this, thanks. The more I look at it, deeper is the metaphor.
silenceimpaired@reddit
Someday we might realize aluminum foil actually protects you from brain tumors and wish we wore the hat.
mpasila@reddit
Some radiation can go through aluminium foil though.
Mayion@reddit
The more I look at it the more I realize I am a fat fuck because I thought they were cookies
mpasila@reddit
Just gotta start using "illegal" sites to go around facial scanning/ID required sites. (with TOR or VPNs at least while those are easily accessible) Might even have to install "illegal" operating systems that don't require an ID to access.. (Linux).
Due-Function-4877@reddit
Destroying all other forums and putting all your discussions in one place that's owned by Google is an absolutely awful idea. Unleashing your agents on small websites to flood them with traffic and make their owners rage quit also doesn't help. But, there's stopping you. You're going to do those things. You already are.
BidWestern1056@reddit
use local models, fuck big tech https://github.com/npc-worldwide/npcsh https://github.com/npc-worldwide/incognide
Nyghtbynger@reddit
Depends on how invested law enforcement is. Drugs like cocaine and weed are very repressed in Singapore but almost legal in the UK or France. Maybe they will chase the models for the first 10 years and then give up. They can't really restrict the hardware because of video games and video editing
silenceimpaired@reddit
RAM shortage says hi.
Eventually you’ll own nothing… including raw compute and you’ll be happy… or they’ll take it away from you.
The RAM shortage is the first step. Owning a computer will soon be a luxury if it does more than connect you to a server.
AmazingGabriel16@reddit
Probs already a thing especially with trigger words.
Like all outputs are probably scanned for trigger words and if trigger words exist it will send it to the police.
Local FTW!
a_beautiful_rhind@reddit
You're thinking too small. Only in the LLM providers doesn't seem to be the limit of what governments are planning. Those datacenters being built aren't going to be used to run chat bots that snitch on you. Gonna be everywhere and used on everything.
soshulmedia@reddit
My take on this is that you are right. But it takes a certain mindset to see that good and evil are not at all "outdated categories" and that Gov is not God.
To know that it is not a sign of mental health to be well-adapted to a sick society and all that ...
The past few years should have woken people up, and certainly some woke up.
But I guess any real change only happens when the "bread & circuses" part of the "societal contract" is called into question. When intellectual laziness and parroting the status-quo with all its lies and propaganda becomes just .. stale.
Ironically, LLMs might "help" to just precipitate that.
It is up to those who value love and freedom to help humanity go down a better path instead of dystopian, orwellian, new-worldian doom. But I also think that the best time to do that is when "everything is in flux" so to speak.
So ... I guess the best you or I or anyone else can do is network and prepare intellectually as well spiritually for what is coming.
Sufficient-Past-9722@reddit
CRUSHER: What kind of resistance should we expect? DATA: At our first encounter, the Borg virtually ignored us when we beamed aboard their vessel. Clearly they did not consider our being there a threat to them. SHELBY: That may change, however, if we start interfering with their plans. Shelby to Bridge.