Forgive me for newb question, but why Hugging Face is a thing at all? Couldn't models be simply distributed via torrents? And the whole aggregator thing would come down to indexing magent links published by model providers?
There's a herding effect with torrents where popular things are quick to get, unpopular things may be very difficult to get, because the majority of users delete files to save space.
I might host 3 or 4 models that I've downloaded recently, but I won't be hosting the 50 or so that I've tried in the last two years.
And who seeds it? If you're going to pay a host to seed it, they might as well just provide HTTP and let you add them as a web seed to your torrent, which we could be doing already.
But if you don't pay a host to seed it, either the host's business model won't play well with torrents, or there's no host and a torrent will rot.
Or even “why would I pay for a flight when I can simply walk across the country”
It’s all about the convenience, minimising the time and effort from “looks interesting” to “running the model”.
Nvidia don’t care if you do it on their cloud, someone else’s cloud, or on your own machine - they win either way, as it further propagates the technology on which they are building their future.
The important thing about HuggingFace isn't the safetensors files, it's the READMEs and the Google search placement. It's easy to reproduce their distribution with torrents, it's very hard to then establish yourself as the default option.
It would be far more efficient to decentralize the data. But HF provides convenience, they subsidize the cost. They gained millions of users, and many enterprises. They can now sell the popularity of their platform. Nobody can buy a p2p network of people.
hugging face is basically a real easy way to run llms. They also provide a bunch of libraries to do things like split compute across all your resources.
