Hacker News Reader

Comment by braza | Hacker News Reader

Maybe it’s a naive question, but here we go: Following HF for a while in the industry, despite all the UI/UX, what’s the business use case to use them instead a “s3” to distribute the models or some kind of torrent for a non-centralised model distribution?

After the whole debacle around the GPT4chan[1] and the whole gating mechanism for models, it’s hard for me to think how some entity can trust that they are not going to shutdown or do gating due to some ToS shenanigans. In other words: if you’re this man in the middle between models and clients, it’s not better to treat yourself as a “dumb pipe”?

N.B.: I think the company has a great culture seeing from the outside and I assume that I can be misinformed about their business model.

[1] - https://en.m.wikipedia.org/wiki/GPT4-Chan

View on HN

It's not naive and it's hard to answer, because the answer is equally naive:

It's sort of like asking what the business case is to have repos on GitHub instead of having a private git server / GitLab.

The value is because "that's where everything is happening": ex. I just did a 4 day hackathon wrapping llama.cpp on _all_ platforms for my as-yet unreleased app. If you need a local AI / llama.cpp model, you go to HuggingFace, full stop.

Then, I want to host these models on my own - I don't want to rely on the HF repos of 3rd parties being stable. Few clicks later, started my own, and uploaded the models. Then, I translate a Python function to Dart, and I can download these models, ranging from 2 GB to 28 GB, using the app, for free, without an API key.

That's much easier than S3, both in cost and integration time.

But still, the answer sounds naive and marginal I'm sure.

But in this case, Github had 3 major tailwinds: (1) the popularization of VCS, (2) the network effect due to the number of users and (3) it's more or less the mandatory part of modern software engineering.

For the use case that you mentioned, I agree that it is important and I understand. Still, for the folks that aren't relying upon LLMs or are doing some vanilla/traditional ML in some laggard industry, I have a hard time believing that those folks are going to HF.

(1) popularization of pytorch

(2) network effect due to the number of users

(3) uploading weights, distributing weights, running demo on GPUs is mandatory part of ML engineering

=== interlude ===

I hope it doesn't sound like I'm being argumentative; discussion is especially interesting to me because it often weighs on me how hard it is to explain HuggingFace. So I enjoy trying and improving at it.

=== longer analogy ===

Imagine if all mobile developers in 2008* needed to host demos on iPhones captive in a server farm somewhere.* Some company offered that for free. On top of it apps were 30 GB, but the company hosted downloads for free. So everyone is putting their stuff on there. Then that feedback loop continues while the field takes a historic spike in interest and it's 4 years later.

* AI developers in 2020.

** GPUs captive in a server farm somewhere.

== Musings ==

This sort of highlights a thread of discussion for startups, the unreasonable effectiveness of specialization. Data scientists in 2020 use Python because they can, they're not really familiar with GitHub as in VCS so their mental model of it is more a dropbox. All of a sudden there's an $X billion (so far) opportunity to clone GitHub, but make it marginally easier to use via hiding stuff that's necessary for all other software, and then light money on fire hosting GPUs and S3.

my experience with HF is that it's incredibly useful for exploring models, and making prototypes, demos and MVPs.

Once you want to scale to production, you're right, it doesn't make sense to use the HF repository, it makes more sense to clone it into S3 or something else that you have more ownership over.

It sounds like SketchUp for LLMs.

To be honest, I do not care about this gpt-chan comtroversy, I doubt most people who use them do, and they have built up a community and ecosystem around their offerings.

I am also surpriswd this has its own wiki page, although it looks like it has been rather quickly put together woth not the most fluid writing.

> have built up a community and ecosystem around their offerings

The community it’s great and I am user for a while. My doubt is that if there’s a lot of use cases where companies and/or MLEs/DS will do some “git pull model_v0.1” from any of the HF model store.

I work at a major tech company and we do this.

I do care in the sense that I don’t want to live in some sterilized clean room Disney world.

See also the neutering of all the big commercial models. No one is running / giving access to a high quality virtually (or completely) uncensored model.

Horses for courses...

Big Corp wanting to replace their call center with the robots will certainly not want them to be offensive to someone who just wants to pay their bill.

Reply on news.ycombinator.com