I don't understand why this is set up the way it is. Maintaining a BitTorrent tracker for models seems like a reasonably good idea, but you have to "apply to join" to be able to contribute?
Its standard for private trackers, you don't want a malware infested hell like piratebay, plus generally you would want people to seed back what they downloaded at least.
> generally you would want people to seed back what they downloaded at least
But the goal is preservation rather than sharing.
I'm confused about how the economics of that would work. If I wanted to contribute to this, I'd have to download the models, so I'd immediately be in the red, and then I could never get out of it because who's going to download these obsolete models? There's no download demand for them; if they were, the whole project would not be necessary.
>then I could never get out of it because who's going to download these obsolete models?
Private trackers usually overcome this issue by also rewarding the length of time you share, even if no one downloaded. Either you get a ratio of 1 or share for a full week/whatever length they decide is enough.
The ai description of this specific site seems to imply they do it as well.
You shouldn't be confused about just the economics of the bittorrent tracker.
You should be worried about the entire economy because it is founded on the principle of paying running costs with one time payments. Basically, neoclassical economists are pathological liars. You cannot model a realistic economy with just a scalar price vector without farming out the planning process to an omnipotent entity.
You see, everything in the economy has two costs: the cost of existence which is a recurring cost (which also happens to be a fixed cost) and the cost of performing the action, usually known as variable costs. The options market has figured that stuff out a long time ago.
Yep, we rely on an economy and are unwilling to pay for its existence. We truly are incorrigible idiots.
Paying running costs with one time expenses is a recipe for growth dependence and planned obsolescence. The only way to bring in money is via a sale, which is a one time event, but your fixed costs are recurring. So you have to sell more products. If you build a product that lasts too long you run into a problem because you didn't charge the customer the running costs.
Oh, I forgot to mention the actual point. Availability/readiness costs money and therefore needs to be compensated even when nothing is actually done.
Does anybody use ONNX for LLMs? I've only encountered it for small models packaged to run on consumer devices, and never seen it show up in any workstation or datacenter scenario.
If you have Google Chrome or Microsoft edge installed on your computer then you have a few gigabytes of onnx llm (Gemma for chrome, Phi for edge iirc) installed. On my side I use onnx for all local ai on https://ossia.io ; works great.
You can use ONNX directly with WebAssembly (Wasm) through the ONNX Runtime Web library (onnxruntime-web). If you use transformer.js [1] then it can load the entire LLM into memory and you can run it from the browser.
Not an LLM, but I had it load the whisper transcriber in a page when I built a podcast timecode logger. [2] I plan to try to get the AI generated podcasts into my MTH120 study notes [3] from Gemini transcribed.
Edge does run Phi-4-mini using Chromium's ONNX Runtime, but Chrome uses LiteRT-LM to run Gemini Nano, and doesn't ship it to the browser in ONNX format.
I see the merit in keeping a decentralized archive for open models. But there has to be a better way to do it vs keeping it behind auth and an application form.
it's light, not like usual trackers, it's only to make sure people seed 1:1 and to reduce load on a couple of people seeding large models to everyone. the only way to do that is to enforce it, and for that you need a place where you can possibly enforce it without making it platform run, so, that's why you have this.
Surely nobody is downloading these torrents (today) because they want the data. The only motivation to use this website is to be part of an archive. So who would leech from it? Why put up barriers to entry?
24 comments
[ 3.9 ms ] story [ 25.7 ms ] threadTo what end?
But the goal is preservation rather than sharing.
I'm confused about how the economics of that would work. If I wanted to contribute to this, I'd have to download the models, so I'd immediately be in the red, and then I could never get out of it because who's going to download these obsolete models? There's no download demand for them; if they were, the whole project would not be necessary.
Private trackers usually overcome this issue by also rewarding the length of time you share, even if no one downloaded. Either you get a ratio of 1 or share for a full week/whatever length they decide is enough.
The ai description of this specific site seems to imply they do it as well.
You should be worried about the entire economy because it is founded on the principle of paying running costs with one time payments. Basically, neoclassical economists are pathological liars. You cannot model a realistic economy with just a scalar price vector without farming out the planning process to an omnipotent entity.
You see, everything in the economy has two costs: the cost of existence which is a recurring cost (which also happens to be a fixed cost) and the cost of performing the action, usually known as variable costs. The options market has figured that stuff out a long time ago. Yep, we rely on an economy and are unwilling to pay for its existence. We truly are incorrigible idiots.
Paying running costs with one time expenses is a recipe for growth dependence and planned obsolescence. The only way to bring in money is via a sale, which is a one time event, but your fixed costs are recurring. So you have to sell more products. If you build a product that lasts too long you run into a problem because you didn't charge the customer the running costs.
Oh, I forgot to mention the actual point. Availability/readiness costs money and therefore needs to be compensated even when nothing is actually done.
Not an LLM, but I had it load the whisper transcriber in a page when I built a podcast timecode logger. [2] I plan to try to get the AI generated podcasts into my MTH120 study notes [3] from Gemini transcribed.
1. https://github.com/huggingface/transformers.js/
2. https://chrissherlock.github.io/mth120/podcast_timecode_logg...
3. https://chrissherlock.github.io/mth120/
Surely nobody is downloading these torrents (today) because they want the data. The only motivation to use this website is to be part of an archive. So who would leech from it? Why put up barriers to entry?
https://pirateface.co/audnai/penclaw-GLM-5.3-abliterated