Hacker News Reader

Comment by lysecret | Hacker News Reader

I’m highly curious of hugging faces role in the new world of huuuge models and few shot learning. Seems to me to be quite an existential thread to them.

View on HN

All (?) the "open" (to various degrees of openness) LLMs are on HuggingFace and have code to support their unique options. For example here is Llama[1] and here is Opt[2].

All the code I've seen for training them uses Hugging Face. When EleutherAI releases their open foundation LLM then everyone will want to fine tune it and do RLHF on it and they'll all use HuggingFace code to do it.

I think they are in the best position of any company including OpenAI.

[1] https://huggingface.co/docs/transformers/main/model_doc/llam...

[2] https://huggingface.co/docs/transformers/main/model_doc/opt

They are at high risk of becoming another Docker.

Everyone uses it but nobody pays for it.

They might succeed if they can convince people to subscribe for model updates like we do for security patches and game updates. For example release any new model for paying members in the first 30 days, and for everyone else later. Or some kind of paid API for LLMs, like "Open"AI, because it's too much of a hassle to run your own LLM.

How so? They host a lot of the FOSS and are a huge infrastructure provider in the space. No matter if Stable Diffusion, GPT or some other thing is the hot new thing in AI/ML space, you can be sure that huggingface is hosting some part of it either today already, or tomorrow.

> They host a lot of the FOSS and are a huge infrastructure provider in the space

You could say the same about Docker, I think? It seems to have been a mixed blessing for them.

Absolutely, you can say the same for a bunch of different companies. Some did really well (GitHub) while others did less well (Docker). I don't think looking at Docker is representative of the entire "FOSS Hosting" angle, and won't predict the future of huggingface.

I have used them extensively and am a big fan, but none of their models are even close to the type of thing ChatGPT can do (for now). They are built for a world where you take a pre-trained model (say Bert) and fine tune it to your specific task. I feel like ML is moving towards: We have these fantastic model with 50 page context, just give it a few examples and it can do whatever you want out of the box. If this is the future, then most of HuggingFaces code and infra is obsolete.

But that's models from people uploading models to huggingface, huggingface themselves don't write/maintain those models, it's a community of AI/ML practitioners.

And LLMs are just a small part of the ecosystem, huggingface is hosting much more than just those models. It hosts models for text-to-image, classification, text-to-speech, and everything in-between those. And beyond that, they also host datasets that are being used for training a bunch of models.

As long as there is a FOSS AI/ML ecosystem, huggingface will remain being relevant (granted no other similar platform appear and takes over). For example, if you want to do anything Stable Diffusion today, it's more likely than not that something is being pulled down from huggingface, one way or another.

Even in your own example, where we have one-shot models that don't need fine-tuning, those models still need to be hosted somewhere. Today, that hosting happens most commonly on huggingface.

He talks about this topic, it's in the transcript.

I ctrl+f'd for "large" of large language model

Reply on news.ycombinator.com