We raised $100M for open and collaborative machine learning(huggingface.co)
huggingface.co
We raised $100M for open and collaborative machine learning
https://huggingface.co/blog/series-c
47 comments
Wish them luck but having gone through something similar with another deep learning company that raised a ton of money, things will probably get a lot harder for them after the C round when investors start looking at revenue and not growth, and with a large valuation that limits the pool of potential acquirers.
I recently searched through the ML companies listed on https://topstartups.io/.
One repeating pattern that surprised me is that there are few successful ML startups, the ones that are there don't seem to be remotely close to self-sustaining - or even product market fit in many cases.
Why is ML such a struggle? are we overstating the impact relative to "old fashioned" data collection and analytics? Is the tech to expensive for companies to adopt in terms of man hours?
One repeating pattern that surprised me is that there are few successful ML startups, the ones that are there don't seem to be remotely close to self-sustaining - or even product market fit in many cases.
Why is ML such a struggle? are we overstating the impact relative to "old fashioned" data collection and analytics? Is the tech to expensive for companies to adopt in terms of man hours?
A lot of people come up with an idea and then have no product. And investors are just happy to throw money at them. It's that whole idea that if you want to start an AI company, hire a bunch of people in India to do your work, then once the funding round is done, actually build the system.
Hugging Face actually has a product and I think they'll be fine. I actually think $100 million is undervalued, because when I think "open models" or "training datasets" I think "Hugging Face."
If you really want free money as an AI startup just say you're going to solve safe / friendly AI. People throw money at that without even showing anything. Hugging Face actually recently put out a job for someone who can work in "bias mitigation": https://nitter.net/mmitchell_ai/status/1520483233132990464
Hugging Face actually has a product and I think they'll be fine. I actually think $100 million is undervalued, because when I think "open models" or "training datasets" I think "Hugging Face."
If you really want free money as an AI startup just say you're going to solve safe / friendly AI. People throw money at that without even showing anything. Hugging Face actually recently put out a job for someone who can work in "bias mitigation": https://nitter.net/mmitchell_ai/status/1520483233132990464
> If you really want free money as an AI startup just say you're going to solve safe / friendly AI. People throw money at that without even showing anything. Hugging Face actually recently put out a job for someone who can work in "bias mitigation": https://nitter.net/mmitchell_ai/status/1520483233132990464
This is the hardest part about going from Fintech to AI/ML for me, it all seems like it's vaporware and you really don't know where to apply your skillset from previous work in several Industries. My focus is on AI/ML based solutions for Supply Chain and Logistics, because it's already broken and needed a re-work for the last 2 decades.
But I'm realizing that there is almost no way to raise these kind of funds and still have a viable business model beyond just setting arbitrary benchmarks and re-tooling your term sheet to reflect the 'new way' things are done in this space now.
I'm still wondering if this twitter stream [0] is just a calloused AI guy ranting, or is it as prophetic as it was when I projected practically the same thing for enterprise based 'blockchains.'
My dillema is wondering how Tesla and Mercedes are all trying to solve FSD, but at the same time we have things like Kiwibot which rely on people piloting them via Wifi from developing countries and pay them sub-standard wages ($2-3/Hr) to deliver food to affluent college kids on campuses. We already have doordash and all of it's incarnations that are bloated with VC money, how can this be a thing?
0: https://twitter.com/amanrsanger/status/1521654821211631616
This is the hardest part about going from Fintech to AI/ML for me, it all seems like it's vaporware and you really don't know where to apply your skillset from previous work in several Industries. My focus is on AI/ML based solutions for Supply Chain and Logistics, because it's already broken and needed a re-work for the last 2 decades.
But I'm realizing that there is almost no way to raise these kind of funds and still have a viable business model beyond just setting arbitrary benchmarks and re-tooling your term sheet to reflect the 'new way' things are done in this space now.
I'm still wondering if this twitter stream [0] is just a calloused AI guy ranting, or is it as prophetic as it was when I projected practically the same thing for enterprise based 'blockchains.'
My dillema is wondering how Tesla and Mercedes are all trying to solve FSD, but at the same time we have things like Kiwibot which rely on people piloting them via Wifi from developing countries and pay them sub-standard wages ($2-3/Hr) to deliver food to affluent college kids on campuses. We already have doordash and all of it's incarnations that are bloated with VC money, how can this be a thing?
0: https://twitter.com/amanrsanger/status/1521654821211631616
I'm becoming more suspicious that there is a chasm between what investors think ML can do, and what it can do. These trends never end particularly well, but by the same token - it's hard to tease out the real wins from the noise.
Part of this is probably the software lifecycle. Everyone wants AGI, but AGI is always further than we think.
Part of this is probably the software lifecycle. Everyone wants AGI, but AGI is always further than we think.
Interesting you mentioned revenue and not net income / profits. How are revenue and growth detached?
A lot of companies use other metrics to show growth, like github stars and package downloads
The brand name only makes me think of Half Life head crabs (they hug the face). Maybe that's the joke on us.
It comes from the emoji[1], but I always think of Boaty McBoatface[2]
[1] https://emojipedia.org/hugging-face/ [2] https://en.wikipedia.org/wiki/RRS_Sir_David_Attenborough#Nam...
[1] https://emojipedia.org/hugging-face/ [2] https://en.wikipedia.org/wiki/RRS_Sir_David_Attenborough#Nam...
Congrats to the Huggingface team! They closed this round at a great time. I except there will be much stronger pressure over the next 18-24 months for them to show revenue growth.
We tried huggingface.co with great hope recently. Unfortunately, though their system was well orchestrated, we could not make progress on baby steps.
We uploaded our model to github and then downloaded to hugging face. Why the package installed correctly, it failed because underlying Glibc headers were compiled with a version that is different from Hugging face's. So, while the platform works for some situations, it still has a long way to go.
We uploaded our model to github and then downloaded to hugging face. Why the package installed correctly, it failed because underlying Glibc headers were compiled with a version that is different from Hugging face's. So, while the platform works for some situations, it still has a long way to go.
I don't get how they can earn those $100M back without becoming significantly less open.
The gist of huggingface is that they host pre-trained open source models for free. Anyone can download them and then use them offline without HF. And their paid offerings aren't even close to being cost competitive with buying a few 3090 workstations.
Also, HF doesn't really have any technical moat. Other people research and train those models, they just provide the hosting. In my opinion, their biggest value is the community. But how do you monetize a group of motivated volunteers?
What stops GitHub from offering free LFS hosting for AI models, thereby copying the foundation of HFs community?
And lastly, is there really much of a market in making SOTA AI beginner-friendly? You still need to buy/rent that A100 GPU server. Who's going to be greedy on salary for people operating a $20k/month machinery?
The gist of huggingface is that they host pre-trained open source models for free. Anyone can download them and then use them offline without HF. And their paid offerings aren't even close to being cost competitive with buying a few 3090 workstations.
Also, HF doesn't really have any technical moat. Other people research and train those models, they just provide the hosting. In my opinion, their biggest value is the community. But how do you monetize a group of motivated volunteers?
What stops GitHub from offering free LFS hosting for AI models, thereby copying the foundation of HFs community?
And lastly, is there really much of a market in making SOTA AI beginner-friendly? You still need to buy/rent that A100 GPU server. Who's going to be greedy on salary for people operating a $20k/month machinery?
The moat is network effects, not the hosting service. GitHub is indeed a potential competitor but, again, you are just thinking about the hosting, while Huggingface provides an API compatible with all the uploaded models and datasets. GitHub would have to create their own API and then hope that users would switch to their service.
Thanks for explaining :)
Congrats. I'm so jealous. Personally, I found Huggingface's models clunky to work with and for me it was easier to build and train my own. I really didn't like how they tried to make a "unified api" and treat PyTorch, TensorFlow 2, etc as backends. Though I realize that for those who are not adept in ml, or don't have the need for much customization, using Huggingface's models makes a lot of sense. No doubt they'll make themselves rich and their investors even richer.
Getting that kind of raise in the current environment is a huge accomplishment. Congratulations.
Huggingface + WandB are some of my favorite tools doing ML.
What is this?
Question: _________ is just an attention mechanism
Answer: Huggingface
https://www.microsoft.com/en-us/research/uploads/prod/2020/0...
Question: _________ is just an attention mechanism
Answer: Huggingface
https://www.microsoft.com/en-us/research/uploads/prod/2020/0...
how will you earn that money back to the investors?
So this question can only come from a place where you have no idea what they do in their field. For every news article or arxiv post that you see talking about how this amazing new GPT-N model has broken all sorts of language benchmark scores, you'll notice that basically nobody can reproduce those results. That's mostly due to the barrier of entry with respect to hardware for training the models.
Huggingface is releasing APIs and model checkpoints that allow any random internet user to execute (almost) SOTA language models in production. FYI - that's an amazing leap forward and a strong piece of kit for MLEs to have access to.
So let me rephrase your question: Is general access to SOTA language models worth 100mm to the software market?
I suspect the answer is a resounding yes.
Huggingface is releasing APIs and model checkpoints that allow any random internet user to execute (almost) SOTA language models in production. FYI - that's an amazing leap forward and a strong piece of kit for MLEs to have access to.
So let me rephrase your question: Is general access to SOTA language models worth 100mm to the software market?
I suspect the answer is a resounding yes.
> So let me rephrase your question: Is general access to SOTA language models worth 100mm to the software market?
That doesn't answer the original question. The question was:
"How will you extract $100M+ from the software market?"
Your answer was:
"I think that Huggingface will produce $100M worth of value."
Which may or may not be true, but just because something produces X amount of value doesn't mean the project will be able to extract that value. See any open source project.
That doesn't answer the original question. The question was:
"How will you extract $100M+ from the software market?"
Your answer was:
"I think that Huggingface will produce $100M worth of value."
Which may or may not be true, but just because something produces X amount of value doesn't mean the project will be able to extract that value. See any open source project.
So if you want someone to answer precisely how they'll extract hundreds of millions of dollars from an emerging market, I have to imagine this isn't the correct forum to expect such answers.
Enabling the general software community access to SOTA language models will absolutely unlock an order of magnitude more money (than 100mm) over time. At least for now their obvious strategy for capturing this value is providing these APIs to enable it, and I suspect they'll gladly host such versions for the orgs that don't have the capacity to fine tune / host their own LLMs.
Enabling the general software community access to SOTA language models will absolutely unlock an order of magnitude more money (than 100mm) over time. At least for now their obvious strategy for capturing this value is providing these APIs to enable it, and I suspect they'll gladly host such versions for the orgs that don't have the capacity to fine tune / host their own LLMs.
Let me try to make the point clearer:
1. Investors expect Huggingface to extract more than $100M from the market. Otherwise they'd be called 'donors'.
2. If they openly publish models, then their APIs will be undercut by other providers who can take the published model and host it for cheaper. It would be cheaper for other companies because: they don't need to pay the cost of training the model, and they can specialize in simply hosting models.
3. Because of 2), Huggingface would need to avoid allowing other companies to host models, including internal APIs (because then providers would simply spin up to making hosting those internal APIs easy).
4) Because of 3), their policy of publishing trained models openly has to change.
So the question that the original poster was asking is: what Huggingface policies will change, given the need to make returns on this investment?
The original poster is likely thinking of OpenAI, which went down a similar route (starting training open models, took in a bunch of money, realized that openly publishing them wasn't sustainable, kept the models secret and created locked down APIs for accessing them).
> So if you want someone to answer precisely how they'll extract hundreds of millions of dollars from an emerging market, I have to imagine this isn't the correct forum to expect such answers.
This market isn't new; Google, AWS, OpenAI, etc. all have APIs they charge for. They also have services to host trained models for you. How will Huggingface make money without resorting to hiding its models?
1. Investors expect Huggingface to extract more than $100M from the market. Otherwise they'd be called 'donors'.
2. If they openly publish models, then their APIs will be undercut by other providers who can take the published model and host it for cheaper. It would be cheaper for other companies because: they don't need to pay the cost of training the model, and they can specialize in simply hosting models.
3. Because of 2), Huggingface would need to avoid allowing other companies to host models, including internal APIs (because then providers would simply spin up to making hosting those internal APIs easy).
4) Because of 3), their policy of publishing trained models openly has to change.
So the question that the original poster was asking is: what Huggingface policies will change, given the need to make returns on this investment?
The original poster is likely thinking of OpenAI, which went down a similar route (starting training open models, took in a bunch of money, realized that openly publishing them wasn't sustainable, kept the models secret and created locked down APIs for accessing them).
> So if you want someone to answer precisely how they'll extract hundreds of millions of dollars from an emerging market, I have to imagine this isn't the correct forum to expect such answers.
This market isn't new; Google, AWS, OpenAI, etc. all have APIs they charge for. They also have services to host trained models for you. How will Huggingface make money without resorting to hiding its models?
Hugging Face is selling CPU cycles. They're also letting you upload your own datasets that aren't "limited" like others. I'm not quite sure where you think their approach of "open models" is wrong, they still sell the CPU cycles.
The idea that restricting access to the data is the only way to profit is such an archaic way of thinking. Hugging Face, if they keep making a good user interface and a good front end, will very much be able to fill the niche it is designed for: people who can't afford a $10-20k rig to run a model but who need to run it for their backend project.
Also, it may be due to using HN, but when I think of "where can I run a model" or "get a dataset" I think Hugging Face. They are leveraging the democratization of the data.
The idea that restricting access to the data is the only way to profit is such an archaic way of thinking. Hugging Face, if they keep making a good user interface and a good front end, will very much be able to fill the niche it is designed for: people who can't afford a $10-20k rig to run a model but who need to run it for their backend project.
Also, it may be due to using HN, but when I think of "where can I run a model" or "get a dataset" I think Hugging Face. They are leveraging the democratization of the data.
Thanks for clarifying, I misunderstood what Huggingface's product was.
I see the niche. The risks are:
- the mid market is constantly churning; either players become too big and you can't meet their requirements or they go bankrupt. Customer acquisition becomes a pretty big expense.
- selling CPU cycles is a cutthroat business which competes pretty directly with AWS, Azure, and Google Cloud. Their edge will likely be ease of use, but at some scale, the larger providers will be able to undercut them hard.
- selling a solution for managing datasets and training models using cloud CPUs is a crowded market.
- not sure how trustworthy the company is with private datasets. Easier to trust an established vendor.
But it wouldn't be a startup if there weren't risks.
I see the niche. The risks are:
- the mid market is constantly churning; either players become too big and you can't meet their requirements or they go bankrupt. Customer acquisition becomes a pretty big expense.
- selling CPU cycles is a cutthroat business which competes pretty directly with AWS, Azure, and Google Cloud. Their edge will likely be ease of use, but at some scale, the larger providers will be able to undercut them hard.
- selling a solution for managing datasets and training models using cloud CPUs is a crowded market.
- not sure how trustworthy the company is with private datasets. Easier to trust an established vendor.
But it wouldn't be a startup if there weren't risks.
I don't think OpenAI is a valid comparison. Huggingface's mission, unlike in the case of OpenAI, is not training models, but being the standard service for sharing them. The vast majority of models and datasets available at the Huggingface Hub have been provided by third-party companies or researchers. They aim to be the Github of ML models and data, not an AI-building startup.
Yep - "Huggingface Enterprise" just like there's "Github Enterprise" seems like a straightforward way to make money, at least to me? Does Microsoft make good money from Github Enterprise?
"This market isn't new; Google, AWS, OpenAI, etc. all have APIs they charge for."
And if they were standalone businesses they'd be losing money, it's neither a big nor profitable market.
When the business model for a project is not 'really obvious' it's usually a bad sign.
AirBnB, Uber, Stripe etc. - 'how' they make money is obvious, it's intrinsic to the product.
And if they were standalone businesses they'd be losing money, it's neither a big nor profitable market.
When the business model for a project is not 'really obvious' it's usually a bad sign.
AirBnB, Uber, Stripe etc. - 'how' they make money is obvious, it's intrinsic to the product.
> So if you want someone to answer precisely how they'll extract hundreds of millions of dollars from an emerging market, I have to imagine this isn't the correct forum to expect such answers.
I don’t think the idea is to get a precise answer. I think the idea is to answer what their business model is. Like will they sell subscriptions? Or ads? Or patronage? Or enterprise support? Or conferences? Or what.
I don’t use huggingface, but am a little familiar with them and think they are a great group with great software. But since it’s OSS, I’m not sure how they would make such a huge amount as $100M. Not to mention that they probably need to make much more than that to have happy investors. So there’s probably a $1-2B plan for making money somewhere and knowing the general idea for their business model would be cool.
I’m a bit bitter over what happened with OpenAI, and many other great opensource projects that turned into crappy companies boxed into making way more than they naturally could make (eg, elastic).
I don’t think the idea is to get a precise answer. I think the idea is to answer what their business model is. Like will they sell subscriptions? Or ads? Or patronage? Or enterprise support? Or conferences? Or what.
I don’t use huggingface, but am a little familiar with them and think they are a great group with great software. But since it’s OSS, I’m not sure how they would make such a huge amount as $100M. Not to mention that they probably need to make much more than that to have happy investors. So there’s probably a $1-2B plan for making money somewhere and knowing the general idea for their business model would be cool.
I’m a bit bitter over what happened with OpenAI, and many other great opensource projects that turned into crappy companies boxed into making way more than they naturally could make (eg, elastic).
It's expensive to hire NLP labor right now, and has been for awhile. Seems like one strategy could be: HF provides a cheaper & more scalable alternative to having to hire an in-house NLP team. Basically NLP becomes synonymous with HF.
And they amortize the cost of hiring their own NLP engineers by developing a few models/model-based services that lots of businesses would be willing to pay for. E.g. 'foundation models' for different verticals like healthcare etc. Then it'll also be a lot easier to either fully automate or at least scale up work that's specific to each paying customer (because fine-tuning should go much more quickly, just essentially be a hyperparameter tuning cycle in as many cases as they can get away with).
And they amortize the cost of hiring their own NLP engineers by developing a few models/model-based services that lots of businesses would be willing to pay for. E.g. 'foundation models' for different verticals like healthcare etc. Then it'll also be a lot easier to either fully automate or at least scale up work that's specific to each paying customer (because fine-tuning should go much more quickly, just essentially be a hyperparameter tuning cycle in as many cases as they can get away with).
NLP engineers spend 10% of their time training models, and 90% preparing the dataset and learning about the specifics of the task. I don't think this scales like selling software.
It's a completely different thing to produce something of value and to get paid for it. There are plenty of examples. Immediately come to mind operating systems or compilers. Or a lot of completely underfunded but fundamental open source tools that everybody uses but nobody pays.
So this answer can only come from a place where you have no idea how business works.
There's a gaping difference between 'value creation' and 'value capture'.
Some products create incredible value for many parties, but don't have an easy way to capture value.
Some products create negative value for the system, but are oriented towards capturing a lot of money.
Wikipedia, Web Browsers a lot of Open Source libs. - examples of the kinds of things that can be invaluable, but whereupon it's difficult to capture value.
There's a gaping difference between 'value creation' and 'value capture'.
Some products create incredible value for many parties, but don't have an easy way to capture value.
Some products create negative value for the system, but are oriented towards capturing a lot of money.
Wikipedia, Web Browsers a lot of Open Source libs. - examples of the kinds of things that can be invaluable, but whereupon it's difficult to capture value.
Since everyone seems to be avoiding direct answers, I have a few ideas:
1. Open source the free, lower quality models (I.e. fewer epochs trained) but sell perpetual licenses to higher quality ones.
2. API access to high quality pre-trained models. Similar to OpenAI, but code would be open source.
3. Licensed access to a Databricks/Collab-esque style development environment. Jupyter is great, but once they establish a big enough community and enough killer features, they could adopt paying users.
4. ML Ops infrastructure as a service for enterprise
5. Consulting
1. Open source the free, lower quality models (I.e. fewer epochs trained) but sell perpetual licenses to higher quality ones.
2. API access to high quality pre-trained models. Similar to OpenAI, but code would be open source.
3. Licensed access to a Databricks/Collab-esque style development environment. Jupyter is great, but once they establish a big enough community and enough killer features, they could adopt paying users.
4. ML Ops infrastructure as a service for enterprise
5. Consulting
One of the draws to use Hugging Face is because the open source community is giving those trained models away. What they are doing is permitting people to use those models affordably (something like a million words would be $10 if I recall correctly, you need to spend $10k minimum to run GPT-Neo-x in your bedroom just to print out one prompt).
I'm not saying these are bad ideas but they need to more focus on their friendly front end and clean API access. There will be a DALL-E 2-Neo-x in a few months. People are going to want to run it without being limited by OpenAI's terrible interface.
I'm not saying these are bad ideas but they need to more focus on their friendly front end and clean API access. There will be a DALL-E 2-Neo-x in a few months. People are going to want to run it without being limited by OpenAI's terrible interface.
Looks like they are selling a chat bot to Monzo and allegedly Bing uses them for search?!
Wonder how this news will impact cohere
great - somewhere in your game plan, please recognize that not everyone is going to use opaque, trained-elsewhere binary blobs as base models; nor does everyone require CNN/DeepLearning to do useful, real research or invention. Things like python skLearn, an internally built supervised model or a few hundred of them; analysis runs that are done without logging into a cloud with AUTH linked to some records behind a closed glass door somewhere..
If you are FAIR, then you let people try things without being attached to you. agree?
If you are FAIR, then you let people try things without being attached to you. agree?
>great - somewhere in your game plan, please recognize that not everyone is going to use opaque, trained-elsewhere binary blobs as base models; nor does everyone require CNN/DeepLearning to do useful, real research or invention.
Shocking that a company focused on Deep Learning, NLP and CV might focus their efforts on people who do Deep Learning, NLP and CV??
Shocking that a company focused on Deep Learning, NLP and CV might focus their efforts on people who do Deep Learning, NLP and CV??
You can turn off the part where it logs into the cloud and insert your own logging backend fairly easily.
Also if your research/biz needs are satisfied by sklearn, then why not just use sklearn? But for a lot of NLP systems, BERT is actually really really useful. And if you don't want to use a pretrained BERT, you can easily initialize their BERT implementation randomly and train it on your own data.
Also if your research/biz needs are satisfied by sklearn, then why not just use sklearn? But for a lot of NLP systems, BERT is actually really really useful. And if you don't want to use a pretrained BERT, you can easily initialize their BERT implementation randomly and train it on your own data.
Can't use sklearn so much for NLP. Sure, it's great for small ML tasks, but you need big honking transformer models to compete in the field.
On the other hand prompting GPT-3 or fine-tuning text transformers is quite accessible now. About as easy as using sklearn.
On the other hand prompting GPT-3 or fine-tuning text transformers is quite accessible now. About as easy as using sklearn.
An example is HF's push toward ONNX export support for their major AI models in the transformers package, which allows faster model inference and could theoretically compete with their managed inference service.