this post was submitted on 28 Aug 2023
65 points (91.1% liked)
Piracy: ꜱᴀɪʟ ᴛʜᴇ ʜɪɢʜ ꜱᴇᴀꜱ
53370 readers
1090 users here now
⚓ Dedicated to the discussion of digital piracy, including ethical problems and legal advancements.
Rules • Full Version
1. Posts must be related to the discussion of digital piracy
2. Don't request invites, trade, sell, or self-promote
3. Don't request or link to specific pirated titles, including DMs
4. Don't submit low-quality posts, be entitled, or harass others
Loot, Pillage, & Plunder
💰 Please help cover server costs.
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Okay, I'm with you but...
how are we using these closed source models?
As of right now I can go to civitai and get hundreds of models created by users to be used with Stable Diffusion. Are we assuming that these closed source models are even able to be run on localized hardware? In my experience, once you reach a certain size there's nothing that layusers can do on our hardware, and the corpos aren't using AI running on a 3080, or even a set of 4090's or whatever. They're using stacks of A100's with more VRAM than everyone's GPU in this thread.
If we're talking the whole of LLM's to include visual and textual based AI... Frankly, while I entirely support and agree with your premise, I can't quite see how anyone can feasibly utilize these (models). For the moment anything that's too heavy to run locally is pushed off to something like Collab or Jupiter and it'd need to be built with the model in mind (from my limited Collab understanding - I only run locally so I am likely wrong here).
Whether we'll even want these models is a whole different story too. We know that more data = more results but we also know that too much data fuzzes specifics. If the model is, say, the entirety of the Internet while it may sound good in theory in practice getting usable results will be hell. You want a model with specifics - all dogs and everything dogs, all cats, all kitchen and cookware, etc.
It's easier to split the data this way for the end user as this way we can direct the AI to put together an image of a German Shepard wearing a chefs had cooking in the kitchen, with the subject using the dog-Model and the background using the kitchen-Model.
So while we may even be able to grab these models from corpos, without the hardware and without any parsing, it's entirely possible that this data will be useless to us.
Fait point. But we're talking about piracy here. Just steal it first and then let's see if we can use it.
Akshually, while training models requires (at the moment) massive parallelization and consequently stacks of A100s, inference can be distributed pretty well (see petals for example). A pirate 'ChatGPT' network of people sharing consumer graphics cards could probably indeed work if the data was sourced. It bears thinking about. It really does.
You definitely can train models locally, I am doing so myself on a 3080 and we wouldn't be as many seeing public ones online if that were the case! But in terms of speed you're definitely right, it's a slow process for us.
I was thinking more of training the base models, LLAMA(2), and more topically GPT4 etc. You're doing LoRA or augmenting with a local corpus of documents, no?
Ah yeah my mistake I'm always mixing up language and image based AI models. Training text based models is much less feasible locally lol.
There's no model for my art so I'm creating a checkpoint model using xformers to bypass the VRAM requirement and then from there I'll be able to speed up variants of my process using LORA's but that won't be for some time, I want a good model first.
Fair cop, Godspeed!
I was thinking the same thing. Would you think there’d be a way to take an existing model and pool our computational resources to produce a result?
All the AI models right now assume there is one beefy computer doing the inference, instead of multiple computers working in parallel. I wonder if there’s a way to “hack” existing models right now so it can be used to infer with multiple computers working in parallel.
Or maybe, a new type of AI should specifically be developed to be able to achieve this. But yes, getting the models is half the battle. The other half will be to figure out how to pool our computation to run the thing.
I'm not sure about for expanded models, but pooling GPU's is effectively what the Stable Diffusion servers have set up for the AI bots. Bunch of volunteers/mods run a SD public server and are used as needed - for a 400,000+ discord server I was part of moderating this is quite necessary to keep the bots running with a reasonable upkeep for requests.
I think the best we'll be able to hope for is whatever hardware MythicAI was working on with their analog chip.
Analog computing went out of fashion due to it's ~97% accuracy rate and need to be build for specific purposes. For example building a computer to calculate the trajectory of a hurricane or tornado - the results when repeated are all chaos but that's effectively what a tornado is anyway.
MythicAI went on a limb and the shortcomings of analog computing are actually strengths for readings models. If you're 97% sure something is a dog, it's probably a dog and the 3% error rate of the computer is lower than humans by far. They developed these chips to be used in cameras for tracking but the premise is promising for any LLM, it just has to be adapted for them. Because of the nature of how they were used and the nature of analog computers in general, they use way less energy and are way more efficient at the task.
Which means that theoretically one day we could see hardware-accelerated AI via analog computers. No need for VRAM and 400+ watts, MythicAI's chips can take the model request, sift through it, send that analog data to a digital converter and our computer has the data.
Veritasium has a decent video on the subject, and while I think it's a pipe dream to one day have these analog chips be integrated as PC parts, it's a pretty cool one and is the best thing that we can hope for as consumers. Pretty much regardless of cost it would be a better alternative to what we're currently doing, as AI takes a boatload of energy that it doesn't need to be taking. Rather than thinking about how we can all pool thousands of watts and hundreds of gigs of VRAM, we should be investigating alternate routes to utilizing this technology.
The point about GPU's is pretty dumb, you can rent a stack of A100 pretty cheaply for a few hours. I have done it a few times now, on runpod it's 0.79 USD per HR per A100.
On the other hand the freely available models are really great and there hasn't been a need for the closed source ones for me personally.
0.79 dollars per hour is still $568 a month if you’re running it 24/7 as a service.
Which open source models have you used? I’ve heard that open source image generation with stable diffusion is on par with closed source models, but it’s different with large language models because of the sheer size and type of data they need to train it.
I have used it mainly for dreambooth, textual inversion and hypernetworks, just using it for stable diffusion. For models i have used the base stable diffusion models, waifu diffusion, dreamshaper, Anything v3 and a few others.
The 0.79 USD is charged only for the time you use it, if you turn off the container you are charged for storage only. So, it is not run 24/7, only when you use it. Also, have you seen the price of those GPUs? That 568$/month is a bargain if the GPU won't be in continuous use for a period of years.
Another important distinction is that LLMs are a whole different beast, running them even when renting isn't justifiable unless you have a large number of paying users. For the really good versions of LLM with large number of parameters you need a lot of things than just a good GPU, you need at least 10 of the NVIDIA A100 80GB (Meta's needs 16 https://blog.apnic.net/2023/08/10/large-language-models-the-hardware-connection/) running for the model to work. This is where the price to pirate and run yourself cannot be justified. It would be cheaper to pay for a closed LLM than to run a pirated instance.