Like the issue with modern AI is the data centers and central control no? How feasible would an AI be, whose code is FOSS and that is trained and running decentralized?
It’s a capitalist world. Economies of scale will always tip the balance toward centralization, in the hands of billionaires.
You should, in fact, look into how centralized AI is potentially harmful to your privacy. I look at decentralizing AI as the means to completely destroy the likes of Grok, GPT, etc.
There were plenty of open weights LLMs, image, audio, or video generators that you can run locally on a decent NVIDIA GPU, if that’s what you mean. Try the !localllama@sh.itjust.works community
no i mean more like a public service, but instead of one centralized power with massive datacenters, it is a distributed zraining and computation. like reddit vs. lemmy for example.
Hardest issue would probably be financing, and motivation.
GPUs are expensive, electricity is expensive. All the current major LLMs are huge loss leaders for giant players with deep pockets. A distributed AI service would be by smaller players without the financing nor the motivation to upfront all the cost.
There is “folding@home” where you donate time on your hardware for scientific calculations, but that’s quite different from donating time on your hardware to some random unknown stranger to generate AI cat images or summarise a news article.
Lemmy and Mastodon etc have a comparatively modest monetary (and energy/environmental) cost, and the benefit is building communities and bringing people together. For distributed AI the cost ( monetary and energy/environmental) is higher, and the benefit is limited.
This is only a problem in a world where we spend no time to optimize these models like we are doing today where we just throw more power at them rather than engineering them to be…better. Look at how China is doing AI - their more limited resources in this regard have forced them to invest in LLM models that work on more modest hardware with much less necessary power. This is necessarily the direction this development must continue to make it a viable product for the average person to engage with without the need for an oppressive mega corporation footing the infrastructure bill (and poisoning the surrounding population at the same time).
Edit: I’ll add that I am broadly not in favor of AI as a whole but the tech is here and has novel use cases. Making the models more efficient is a necessary step towards seeing this tech’s true usefulness be actualized.
Oh I think I did hear of something like that. Search the AI horde
Already done. See:
Open training: https://huggingface.co/collections/allenai/olmo-3
Decentralized training: https://huggingface.co/NousResearch/Hermes-4.3-36B
Decentralized inference: https://aihorde.net/
They aren’t the only example of each, either.
The issue, as with many decentralized projects, is that they’re isolated from each other, and too few know about them.
Have you heard about the AIhorde project by the DBzero community ?
@aihorde@lemmy.dbzer0.com draw for me a surprised lemming hacker typing in front of a green text computer screen
Here are some images matching your request
Prompt: a surprised lemming hacker typing in front of a green text computer screen
Style: flux
Technically? Not a problem. The reason most of them run in data centers is the massive amount of computational and therefore also electrical power you need to run a somewhat useful model.
Even worse when you need to initially train them. That’ll really hit the wallet.
There are (by now) vast selection of models you can easily run at home without any outside connection, as long as you have reasonable hardware to run them.
Based on OPs comments in the rest of the thread, they’re talking about a fold@home type system, not a locally run LLM.
If you mean distributing inference across many machines, each of which could not individually deal with a large model, using today’s models, not viable with reasonable performance. The problem is that you require a lot of bandwidth between layers; a lot of data moves. When you cluster current systems, you tend to use specialized, high-bandwidth links.
It might theoretically be possible to build models that are more-amenable to this sort of thing, that have small parts of a model run on nodes that have little data interchange between them. But until they’re built, hard to say.
I’d also be a little leery of how energy-efficient such a thing is, especially if you want to use CPUs — which are probably more-amenable to be run in a shared fashion than GPUs. Just using CPU time “in the background” also probably won’t work as well as with a system running other tasks, because the limiting factor isn’t heavy crunching on a small amount of data — where a processor can make use of idle cores without much impact to other tasks — but bandwidth to the memory, which is gonna be a bottleneck for the whole system. Also, some fairly substantial memory demands, unless you can also get model size way down.
Plausibly feasible, but there would be more people using the service than letting the service borrow their hardware for processing or memory. I imagine it would work like a botnet, where if a user generates a prompt, their machine would borrow other nearby machines to process the command efficiently. So, in order to make it a fair service, perhaps the software does a POST and only operates if you meet a minimum hardware specification, and a stable internet connection.
