New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Now The Bad Guys Have AI As Powerful As Mythos (and they will let you use it free NO KILL SWITCH)
stable_genius
Member
Unlike the good guys models the Zhipu's GLM-5.2 is open-weight, that is, the bad guys made their model free to use by anyone who has the hardware and the know-how.
NO KILL SWITCH!!!
NO SUPERVISION!!!
UNFETTTTERED ACCESS!!!
What do you think of this, will it change anything in the low end?
As it can be taken freely I'm sure that the smaller bad guys will want to use it too in their nefarious actions so I wonder, will this translate into more business for LET providers?
Comments
I don't really follow this but who are those supposed good guys that heroically safeguard mankind? I don't think they exist, so i guess it'll just even the playing field. As much as i don't like AI it is what it is and i can't bring myself to care that much.
GLM-5.2 seems more around Opus level than Mythos/Fable level. Still very impressive, and it shows that the open models are always just a matter of months behind closed ones.
I don't like AI either, most of it is just hype anyway. However, this is different because these models excel at finding bugs in code and that will have an impact on you whether you like it or not. You can ignore the hype but you cannot ignore this.
Well, those bugs would be found sooner or later anyways and the good guys keeping a thumb on it basically just means giving the CIA a head start. Finding + patching is the only solution.
Why?
One potential very bad side effect of this is hardware getting even more expensive because now the only way the good guys have to stop you from using the most powerful features of these models is by starving you of hardware.
They once believed they could simply restrict their models to stop you from accessing the most powerful features but now they know their only option is to make hardware so expensive that scarcity of hardware will take care of stopping you.
Now the bugs will be found sooner, the CIA good guys will have to work harder and on a more level playing field.
Because you'll have to patch your software more often, you will have to use unpatched vulnerable software more than you should. You will have to work harder to fix the vulnerabilities faster and risk suffering serious loss.
This shit works at scale and subtle software vulnerabilities that otherwise would never emerge will from now on become very apparent.
This is not just hype that you can safely ignore.
What is all this good / bad guys nonsense here?
I get the feeling I'm reading some book for children 🫠
Some call that SARCASM.
I call it that too and it certainly is no child's business.
Declaring any kind of secret service to be the proverbial good guys is a bit weird to me but in the end it probably simply comes down to many people not liking to be left vulnerable just because whoever wants to stack up its 0day collection.
I mean just look back at the whole Flash debacle. How many of those exploits were also known to third parties (other services, generic cybercrime, ...) will forever stay an unknown variable. Sure, it's nice to be able to exploit your adversaries but when the cost includes leaving yourself open to the very same thing the whole concept starts looking a bit sketchy. It's basically security by insecurity, which is as much of an oxymoron as it sounds like.
What I would like to know is if there are any LET providers capable of providing the hardware needed to run these models competitively.
I'm doubtful.
Sure, just get the VPS with 1TB VRAM for $7/yr
No i am not.
Same feeling.
What is good, what is bad?
Who is going to be the judge for that?
YOLO. Less worries more silly.
For a while that's likely going to be the case but exploitable bugs are kind of a finite resource or at least the rate at which new one's are going to be added will sooner or later become the limiting factor. Sure, settling down will take longer on larger more complex codebases but even there it will happen.
Which is pretty much already the case. Exploits have been a commodity for quite a while now. Expecting even every second critical vulnerability that is found to end up with the actual vendor and being patched would be kind of optimistic. These things tend to get bought up and stashed away. Chances are a lot of what we (as mere mortals) get presented as new and scary has been known in interested circles for years. It's just that some random rediscovery has ended the exclusivity.
Well, maybe but i'd guess the overall effect to be rather small. Most serious bugs aren't very universal and will just affect a small subset of systems. Even if they do chances are circumstances would hinder or stop successful exploitation (even more or less free-for-all exploits like blaster could be pretty much ignored if your box wasn't directly exposed to the internet for example). Assuming that there will be some (temporary) uptick is probably a safe bet but effects on individual users will just happen rather sporadically.
Likely some will but in the end it's all still just cooking with water.
Ignoring security would be kinda bad with or without AI
today i learned that glm 5.2 is on par with mythos
China is already distilling mythos, don't worry guys.
Either huggingface or specific hardware might get banned next.
HuggingBay is up soonTM
Mythos distillation is a lot harder due to mandatory kyc.
I think they don't need to do any distillation to surpass Mythos, all they have to do is keep their focus on doing real work as efficiently as possible instead of investing in hype generation and creative financial operations first.
Those I've been amusingly calling "The Good Guys" have been focusing far too much on Financial Creativity and hype generation, that may prove to be their undoing.
"The Good Guys" (that was SARCASM, in case you failed to notice) work in tremendously wasteful ways, anyone who works efficiently will be able to do double the work using half the talent and 10% of the material resources.
Elon.
Haha, good one
That's awesome man!
I am so glad for you.
Sam Altman,
Alex Karp,
Jeff Bezos,
and more...
can it do $7 per year?
lots of talk about GLM 5.2 but there is no way to run that shit. It needs at least 300GB of vram.
Exact that's the problem, you can't do that shit on old second hand or modded hardware, the hardware that is needed is out of LET providers reach.
People have been talking of some cheaper modded GPUs with double the VRAM but they are only 20GB and fairly slow, reliability might not be great either. You'd need to use 15 of those to get to 300GB VRAM.
But the AI giants cant keep making a loss forever and the price of AI services will rise very sharply (it already is) so there are some serious opportunities here, one day someone will be able to offer something here on the low end.
But that will only be after the burst when hardware gets cheaper.
The CIA aren't the good guys.
yea.. and I'm not trying to spend money on hardware or rent gpu server with the current prices just so I can tell local models to do my work. I own the base M4 Pro Macbook Pro with 24GB of unified memory so it can run some basic local models. But I rather just purchase a claude/gpt subscription so I can use the latest models that can do the job a lot better and faster.