Soliciting surplus last gen graphics cards for USIPS

gagabobo1997 said:
If you are looking at the v100 route, the /lmg/ general on /g/ has this V100MAXXING guide for their erp models or whatever.

Prices probably have changed (most things have only gotten more expensive this year nevermind im just a doomer, they seem to have gotten way cheaper, particularly the 32gb models) but it's still good advice.
Click to expand...
That's a really good idea. Do we have any AI bros super deep into this shit and know the hw really well?
 
Null said:

anyone wanna spot me one lmao
Click to expand...
I’ll give you a GTX970 in exchange for True & Honest.

Can’t really help you with anything but cash since you won’t say specifics about what you plan to do, so I’ll see about getting some crypto for you to buy whatever. I do think you need to carefully document everything you do, you have a lot of enemies and if there’s even a hint you used illicit images to train this thing you are going to end up smeared and in legal trouble. Yes you only got it because a furry paedo kept putting it on your site to sabotage you, but courts really don’t like free speech advocates on the best of days, less so when the opposing counsel accuses you of being as degenerate as his own client. Training an image recognition AI does mean distilling sample images of whatever it’s meant to recognise into the model, and judges won’t understand that the sample material can’t be reconstructed from the baked model.

You say you’re doing something different, but it’s worth mentioning again that training a model on illicit material is illegal in most locales. Yes Facebook and Twitter do exactly that for their own moderation models, but you can’t afford the lawyers and bribes they can. And the troons will jump at any chance to get you in legal trouble, so make sure to document whatever it is you’re doing.
 
I'm not training anything. I have a strategy to use existing legal models to look for certain identifiers to secure the site and possibly eventually other sites. It is not a complex idea. I don't want to say sepcifics because I know he tries to beat these things and enjoys trying to fight countermeasures.
 
Null said:
possibly eventually other sites.
Click to expand...
Those v100 rigs would probably be great for this since they'd be scalable easily, just buy another one and throw it in the rack (in theory).

Would you do a fundraiser? You were able to do those USIPS annual memberships, I'd be more than willing to throw a couple hundred bucks towards the charity if you ran a drive.
 
I will complete the NCMEC reporting system first before asking for more money. I will also wait and see if maybe someone has one of these v100 things their company wants to get rid of already.
 
Are the servers only needed for training the model you want to train? And then once it is trained you no longer need them?
If so, why do you not just rent servers for this project from somewhere like this: https://www.runpod.io/pricing There are hundreds of companies just like that one around and you can rent this hardware for much less than actually buying it.

Keep in mind even if you get the servers you wish to get, you still require motherboards, extensive amounts of expensive RAM, SSDs, PSUs and it will cost an arm and a leg in terms of electricity.

Think of it like Salad thinks about Chantal - she is now in Canada and he is away from her, so he just manipulates her to get her YouTube money without having to actually deal with her IRL. TLDR: Why have the cow when you can have the milk for free? The hardware you're trying to get is extremely expensive to own and operate.

Perhaps do not abandon the approach you're doing already, but perhaps contact a few cloud providers and ask for a deal for the "social good" project you are trying to action and maybe one of them would be willing to support with a few free hours as credits?

(I did search the thread for "rent" so don't get mad at me for suggesting this)

edit: just to add that with the rent option you'll get access to the cream of the crop hardware instead of handmedowns that are old and clunky
 
If you want to do some testing before deciding on a specific GPU model https://vast.ai also allows you to rent all kinds of models from other people and they take crypto. They give you SSH access to a container with the GPU and you can install basically anything you want.
 
cohle said:
If you want to do some testing before deciding on a specific GPU model https://vast.ai also allows you to rent all kinds of models from other people and they take crypto. They give you SSH access to a container with the GPU and you can install basically anything you want.
Click to expand...
It may end up cheaper to do the entire thing like that - having the hardware would be nice but that will take time @Null
You can start immediately with minimal cost, so if it doesn't pan out you can cut your losses almost immediately instead of waiting weeks or months for actual hardware
 
As an aside Josh, have you messed around with Torch or TensorFlow before? Custom models are kind of fun to throw together. IIRC sklearn has some of the sample datasets like MNIST or cat/dog available for download if you want a starting point. As much as modern AI tech has aggrieved the world, the underlying concepts are very neat to play around with.
One of my personal favorite projects was training an AI to be able to guess where ships are on a battleship board
 
Null said:
I am interested in using AI to build a tool to auto flag illegal content. I believe this can be trivially achieved using last gen cards and open source models.

I've received offers of free hardware in the past. I know AI is this really hot thing right now so I'm hoping there are companies people work for phasing out 4000/5000 Adas for 6000s and I can get ahold of that.

If you have access to this let me know please.
Click to expand...
I know you're not retarded, so I don't mean this as a gotcha, but how are you planning on training this thing? Obviously you can't and shouldn't use the data you want to detect, so what's the plan?
 
gagabobo1997 said:
I'd be more than willing to throw a couple hundred bucks towards the charity if you ran a drive.
Click to expand...
Same here, this project has me interested. I have a negative appraisal of the usefulness of LLM's and their image generating counterparts but Machine Learning models that can enable moderation against these disgusting attacks for small website operators seems feasible to me. It is something I'd be happy to support, will stand back and stand by.
 
Null said:
That's a really good idea. Do we have any AI bros super deep into this shit and know the hw really well?
Click to expand...
I know the V100s pretty well; we had several servers with them, I deployed applications on them, and we still move refurbs at my current gig. It's IMO a very solid machine, but I don't know what else to tell you. $3300 gets you a 4x V100 server on Ebay. It's a pretty good price. What kinds of models do you want to run?
 
Honestly what is the point of all this trainload of antiCSAM dedicated organizations and LEO task groups and multibillion dollar computer clusters and all this spending if certain bad actors are apparently freely and openly wandering around and firehosing CSAM material at anyone they want to shutdown and nobody in the Grand antiCSAM Alliance lifts a finger to stop them? If I didn't know any better I'd say the millions/billions being spent on 'antiCSAM' are seemingly not so much to stop it but to facilitate it being built into a weapon to be used against undesirables.
 
Last edited:
Silver Coin Anon said:
I have a negative appraisal of the usefulness of LLM's and their image generating counterparts but Machine Learning models that can enable moderation against these disgusting attacks for small website operators seems feasible to me
Click to expand...
we'll see, he can definitely get something running but the effectiveness is up in the air. He has a couple big problems: it's easy to overload the scanner unless he can rate limit attachments (residential proxies remain a major risk here), and the relevant models aren't very hard to fool. He has options for dealing with these issues but his initial hesitancy was well-founded imo, if he's not careful he could wind up spending thousands just to give attackers an easy way to DoS attachments.
 
You must log in or register to reply here.
Back
Top Bottom