Hugging Face raises $235M from investors including Salesforce and Nvidia
techcrunch.com
techcrunch.com
Although, is it technically a platform / marketplace?
Anyway for now, enjoy the free beer from the VCs.
What Docker ended up accelerating at is reproducible development environments that tear down and spin up easily. I think this is the largest faction of docker users.
It never translated into a high volume of Docker Swarm and related sell through features materializing though.
That's what a swath of all their new paid features are focused on enterprise things like SSO, auditing etc.
For Hugging Face, they will need to have a compelling set of features that make it either hard to migrate off of or vastly preferable to other options the majority of the time.
Right now, their most compelling feature is being (mostly) free
Whilst the scale of their model hosting is impressive, the functionality seems pretty basic. The models are just BLOBs in git LFS repos, you're usually relying on knowing which users to follow, then learning the way they name their models and how that naming convention applies to the particular framework and hardware you're using.
As an example the user "TheBloke" is prolific at publishing LLM models for various hardware / framework combos, but look how little the HF interface actually helps navigate or find what you're after: https://huggingface.co/TheBloke
Also, is anybody using HF Hub in "production"? We've deployed a few LLMs now and once we've decided on a model the first thing we do is get it off HF and into our own storage ready for deployment. There seems no reason to tightly integrate HF Hub with production systems given it's a just a bunch of files you can copy and keep.
From what I gather it's mainly hosting with some maintaining of a few libraries, but then I recently starting to see they are offering lots of classes via DeepLearning platform; these lat couple of months as an AI student I've been asked to enroll into classes (temporarily for free) to assess where they are, here is the most recent example [0].
To what end, I'm not entirely sure, but I guess it's to get an overall pulse on what can be monetized and take it from there? I really think that this a low number compraed to where we were in the last few years, but ti also shows how little investment actually exists in the AI and ML space: consider that Git got bought by M$ for 7.5B and then tried cash in on it by releasing Co-Pilot and then got into legal issues as a result and then tried it's hand with Open AI and $10B.
This is starting to seem like a reversion to the mean, and AI's promise was always to lower the cost to everything it can, but I think one of the harder pills to swallow is that the traditional VC model is not really being supported after all the hype and losses.
Personally speaking, I'm thinking of moving on to Cyber Security after my finals this semester; I come from Bitcoin and the hype cycles there are something I was looking to get away from after nearly 13 years in that side of fintech.
Have been experimenting with deploying to Mac's for inference as they are cheaper and use less power. But also still deploy to a substantial amount of gpus too.
Either way I try to take connections to huggingface hub out of the equation at deployment time.
I think it's important for people to diversify away from them and not build anything that uniquely depends on them. It's not good to have chokepoints like this.
If they reach a point where they actively become community-hostile, someone will just fork their codebase and release a web app called "FaceHugger"
What they do isn't too hard to replicate, but the price for which they do it is impossible to compete with.
Rumor has it Github is actively working on something similar to HF Hub already
pick one if you are lucky. Those Fortune-12 giants can't do anything "in a heartbeat" .. listen to what that means.. nonsense
Someone utilize p2p for the love of god. Pretty sure there’s gigabytes of porn moving (freely) through the internet millions of times per day.
P2P model distribution isn't impossible, but it's a lot less interesting to many people, or they would've already spread their models through torrents in the first place.
But if HF's server code is open source it should be trivial to make one.
In fact, I think the whole community is HF averse. The two most popular frameworks are based around the Stability implemenetaion and file format, not HF diffusers.
Automatic1111 and ComfyUI implemented the SAI backend/format in the early days because thats all there was, and now they are stuck with it. The intertia is tremendous.
It seems like an odd move to me as what makes Auto1111 so appealing are the extensions and I’m assuming that breaks an awful lot of them, but what do I know?
Technology-wise, path dependency from stuff that was built out before HF diffusers were available, and on which more continues to be built in the ecosystem.
If you are doing a greenfield project that is mostly standalone, its not an issue, but for existing popular projects and the communities of satellite projects around them, the switching cost is high.
In terms of content hosting, well, a fair amount is hosted on HF, but there is a difference of content focus between CivitAI and HF, and a lot of what CivitAI hosts HF probably wouldn’t want to. Also, CivitAI has a UI focused on the narrow space of imagegen, whereas HF is more general.
On top of that HF is just hard to use for your average user. Civit.ai is just "click to download" while HF is "look here's a broken model card... you can figure it out from here".
Despite the cute logo, I think most people find that HF comes across as fairly anti-user. Despite having years doing ML related work, I still find HF a bit byzantine to navigate.
https://github.com/huggingface/safetensors https://huggingface.co/blog/safetensors-security-audit
I don't know much about it, and I think (but couldn't quickly confirm) it's open source, so your point about a fork still stands, although I don't think that solves everything. If it did, people wouldn't care that e.g. Hashicorp changed their license away from open source.
I don't quite understand how they do so, but it seems to be possible. Everyone who downloads a 5GB stable diffusion model probably downloaded 5GB of pytorch+cuda+cudnn first.
This itself is kinda insane. In fact, it was literally untenable for PyPi, which is why (historically) installing all that was unecessarily painful.
[1] https://dustingram.com/articles/2021/04/14/powering-the-pyth...
python -m spacy download en-core-web-lg
etc
There are network effects in play here. They have a big moat purely because of their large user base.
Model hosting is a nice convenience, but if HF removed every single repo from GitHub tomorrow and paywalled the model repo, it wouldn't be a big deal. Maintainers would clone their models and repos to somewhere else.
Careful not to become dependent on the free beer.
Why would a CEO give up equity if they don't need the capital? Wouldn't they prefer not to sell big chunks of flesh?
If he's not deploying it now, maybe the CEO is predicting rough times ahead and wants to de-risk for future cash flow or fundraising turbulence? Isn't that, too, a kind of negative signal? Or is this all prudent?
It's possible to structure a deal where the company has to reach a series of milestones to unlock each additional tranche of funding, but this is less common.
"Then collapse" - nope, then generate tens of billions of profit per quarter. Im sure the people of Hugging Face would love that
"enshittification" doesn't refer to anything at all. It's a political term meant to be used by anti-corporate causes - it doesn't have a real definition because it's a made-up word with no value other than activism.
> The word does indeed refer to something. Just because you disagree with it doesn’t mean it doesn’t exist.
That's irrelevant here. Virtually everyone would agree that the word "libtard" also refers to something, but very few would actually appreciate it being used.
As a counter point, GitHub has remained great for longer than expected, and has arguably improved since being bought by Microsoft.
One can hope that their example will help set the expectation for other platforms like HF
VC-subsidized Ubers were awesome in my mid 20s. It’s too expensive now for everyday use but I still banked all that consumer surplus.
it is totally can be condition of investment round: have specific vendor as compute provider.
Oh I assumed they were directly delivering them truckful of A100s? do they actually still bother with cash? ;-)
HF did an amazing job in community building, transformers library and being the central store for all oss models. That said they are ages away from PMF and just have a bunch of different products non of them commercially successful (services, autotrain, quantization, HF hub for EE, inference end points etc). The majority of their revenue comes from partnerships with SageMaker/Azure where they pay them for sending users their way which wouldn't continue to grow.
While it's always a possibility for a FANG company to buy them IMO they are completely screwed. At a $4.5B valuation they will have to reach at minimum $250m in ARR to IPO and at the moment they're probably stuck at around $25m ARR.
Google has poured hundreds of million into Anthropic due to Eric Schmidt, so they aren’t going to be a buyer.
Hosting binaries isn't really a stable business model.
Google could grow GCP by more deeply integrating HF into GCP, I think, while retaining as much of the HF brand and interface as possible.
Big friction point might be ethics, since HF is still seen as the "good guys" and where some folks who left Google over ethical concerns landed.
> Its revenue run rate has spiked this year and now sits at around $30 million to $50 million, three sources said — with one noting that it had more that tripled compared to the start of the year.
I can't for the life of me understand the strategy that Clem and team have, other than raise as much money as possible just because they can. My experience with their sales teams was just absolutely awful, and it gave me no hope that they can grow ARR as soon as they need to. We practically begged them to sell us something, but it wasn't until we went elsewhere definitively that they seemed interested.
I’m far from being knowledgeable in this space, but it seems like AI/ML “is a feature, not a product”.
And if that’s the case, what business are you in when a company sells AI/ML?
Are you in the business of licensing the model you created? Charging for the output? Hosting infrastructure? What exactly are you in the business to sell?
To use an analogy, if you’re selling AI/ML, are you in the IaaS industry, PaaS industry, SaaS industry (or something else?)
The further you keep the user away from training and hosting, the less you’re in those businesses. But I would think it’s only economical to do that if you have some type of advantage in implementing the things you’re abstracting away.
If they can make it easier and worthwhile to use their product and create business value, then they will sell a lot of picks and shovels.
PS - I'm optimistic about AI, but it's important to be realistic about the current state of the industry
For example, if you open the home screen on the average smartphone right now, you'll see apps like:
- Delivery apps like Uber, Lyft, etc., whose recommendations, ETA predictions, driver matching, and more are built on ML.
- Media apps like YouTube, Netflix, etc., all of whom rely on models for recommendations.
- Email apps like Gmail, whose filtering (both spam and categorization) and text completion are based on ML.
- Photo apps like Instagram, Snapchat, and even your phone's basic Camera app, all of which use computer vision.
If you Google anything, you're perusing the output of a model. If you're being recommended something on basically any platform, you're interacting with ML. If you ever use speech-to-text, you're using a neural network. Your bank uses ML for fraud detection, your posts on social media are moderated by ML-based content moderation, and if you have a car with any recent-ish sort of lane departure assistance, you're driving with help from a neural network.
Most of these companies have large, mature ML teams, whose outputs represent massive amounts of revenue. Hence, they represent a legitimate market for selling picks and shovels.
But what “tools” are being sold for AI/ML?
- A model isn’t a tool, that’s a purpose built offering that can be used for 1 use case. Tools are by definition, intended to help with broad/general use cases.
- Hosting isn’t a tool, it’s a commodity and lots of cloud players already occupy that space.
Please don’t take my comments as trolling, I just geninuely don’t understand what exactly is being sold that is new or unique that doesn’t already exist.
I imagine they have big plans on how to expand and grow the business.
a) a hardware vendor nvidia/amd whose products are needed by anyone in the game
b) you have a captive customer base already ( microsoft, salesforce, servicenow, adobe) to whom you can sell ai/ml value adds
c) you make money via ads (google facebook) and ai/ml helps with better targeting
everyone else is pissing away VC money.
FTFY.... I don't think AMD will be competitive in AI until at least 2025, though given their broken roadmap promises, not sure if anyone will be willing to invest in them
Something, something gold rush, shovels.
I am obviously biased but IMHO, yes. High end switches are not commodity gear. And a lot of the differentiating secret sauce is in software. So, even if two switches from two different vendors have the same silicon, they might have dramatically different characteristics. That is without going into other aspects like integration with the rest of the stack, quality of support and even lead times.
B) re: value adds, like what exactly?
C) this seems like such core functionality that a company wouldn’t outsource this to a 3rd party vendor. If that’s the case, there isn’t an opportunity to sell anything if you’re that AL/ML vendor then.
Just do a quick HN search on companies selling AI/ML, you will see what I mean.
Like, it's plausible that many organizations would be willing to pay lots of money for a "ChatGPT which knows my internal documents" AI/ML feature for the Sharepoint/Confluence/etc they are currently using, but a would be very wary of migrating those internal documents to some upcoming startups' new document management system.
What is the moat? "We will run your inference" can't be the answer.
Network effect. It's more like a GitHub for ML.
Also trust. When your business is hosting binaries, that's no small feat.
Docker also had DockerHub, but it wasn’t as necessary as GitHub and didn’t catch on as well.
I think HF is more like docker than GitHub. GitHub had the wide scope to be huge. Docker was too niche (in comparison) and didn’t catch on so much to be necessary for every single project like Git did.
Similarly, HF is also niche in the sense that docker is. There’s no need to host your model on HF, it’s just a convenience for some projects. It’s effectively a package manager for ML models.
It seems that if there’s a way to make models more easily runnable, HF would not be necessary at all. They have a community there, but it’s not something that can be monetized well.
All cloud providers seem to have an image hosting feature, why go through hassle of managing and paying for a docker hub account when you're already on AWS and can integrate better with ECR? The same problem exists for ML model hosting IMO.
Github is totally different, it's a more complex and human proposition. For many github has replaced an entire raft of tools, not just code hosting.
What was the moat when AWS EC2 launched in 2007? There were already thousands of little mom-and-pop VPS companies that would rent you a virtual Linux box with root permissions for $15/month.
Bonus points for certifying the models actually do what they say. That by itself will probably become a mini industry.
Hugging Face are the clear leaders in terms of having mindshare in the community to be able to build it.
We’re not in a world where non tech people are searching for models in the model store from their phone.
That kind of thing is par for the course when invested in by hardware companies, but it can work out very well for everyone.
2) How in the world are they going to pay it back?
They better have a damn good idea because this seems like a good recipe for popping like a balloon.
1) Grow 2) They don't. It's not a loan. VC world doesn't operate on loans. You go to the bank for that.
https://huggingface.co/pricing
I’m probably missing the obvious, but one part on the pricing page it says Spaces Hardware starts at $0 … and another part says it starts at $0.05.
• Repository storage is free, with a paid enterprise offering, à la GitHub.
• Serving a demo app has hourly costs (“Spaces Hardware”)
• Serving a production model for an app hosted elsewhere is also hourly (“Inference endpoints”)
• Training models is free for now (“AutoTrain”; honestly I haven’t tried this one.)
Citation needed.
For at least half of them, it might make sense to simply sponsor accelerating worldwide new AI product development because those new products will increase sales of their hardware and services; pitch in $20m together with others to support efforts like Huggingface, while expecting that their tools will indirectly cause an extra $1b sales for you in the next few years.
Yes, they have. Higher bond yields (compared to 0% three years ago) is causing pension funds to shift their allocation away from "alternative" investments (vc & pe), so it has become harder to raise a new VC fund these days.
I'm not going to host my code on multiple platforms, pretty easy to make a model available for download with S3
I hope the founders take their share home and have fun burning the rest. At least hopefully some open software/AI models will have come out of it when the company collapses under its own weight.
What would HN recommend? I prefer Hugging Face as it has a stronger community built in but others prefer a open source project we can customize.
If it wants everyone to fork its models, having the model on Huggingface and using its libraries will increase adoption, as people are used to that format, having the quantization built-in, etc. Having to perform model conversions slows things down, despite TheBloke’s constant efforts to convert every format to every other.
If the model will only be accessible through CLI, APIs, or a website UI, then custom can make sense.
Y'all could host models on HF, and use the HF format/download code (which is quite good).
But make a pretty frontend for it.
Down the road you could just host your own backend implementing the HF API, if such a thing is necessary.
I am curious however, why name drop Elon and X.ai when asking this question?
With all due respect, I'm not sure how that is going to advance the technical conversation, and in some regards may indeed de-rail it and hinder the mentoring you might receive?
I hope I do not sound combative!
I am generally curious, as I see this behavior at $JOB, but in reverse. "My customer is seeing this." "I have a customer that is asking Y."
I am generally curious on the different perspectives that teammates can come from, that can drive them to take wildly non-congruent approaches, even with similar career experiences.
You work at a business and you think it's name dropping for a colleague to bring up specific customers' needs?
I'm trying to see the steel man version of your argument, but I'm not getting it. GP is working for a well publicized AI startup and shared that they are evaluating HF versus alternatives. I thought it was useful information, and it's germane to this thread.
Added to the fact that of they get too dominant they'll get leveraged out of the supply chain and I'm not sure what the value proposition is here.
I mean obviously I'll be wrong but it's hard to not be skeptical
Edit: dang, guess HN strips out emojis from posts. The Punycode version of the Hugging Face emoji on a relevant TLD supporting emoji domains would be http://xn--zp9h.ml/
I wouldn't be surprised if there is little to near zero risk for Nvidia with this bet.