The $4.5B Platform Driving the Open Source AI Revolution | Clem Delangue, CEO, Hugging Face

Miami vs. New York vs. San Francisco

Matt Turck [1:04] Hey, Clem, how are you?

Clément Delangue [1:06] Pretty good. Thanks for having me.

Matt Turck [1:14] We're recording this today in New York, which used to be your stomping grounds, but you're in Miami, right? You've been in Miami for a few years?

Clément Delangue [1:31] Yeah, I'm French. I grew up in France, as you can hear from my accent. And then I moved to New York, spent eight years in New York. And during COVID, I did the cliché move from New York to Miami. But I'm also traveling a lot, so still spending quite a lot of time in New York, in San Francisco, in Paris.

Matt Turck [1:57] I should give the disclaimer upfront. Actually, a few weeks ago, I had a wonderful chat with Olivier from Datadog, and I gave the warning that it was going to be a conversation between two French people speaking English to each other. So people have been warned again. Any thoughts on Miami versus New York versus, in particular, San Francisco? There is this dominant narrative that all AI is built in San Francisco. Is that a problem for you to not be in San Francisco?

Clément Delangue [2:29] Not so much. At Hugging Face, we have a fully remote culture. So we have people all over, including in San Francisco, in New York, in Miami. Me personally, I enjoy the move from New York to Miami. It's making me happy, which I think is important. Yeah, important step for a co-founder, for a CEO to be good at his job. So it's been quite a good move for me.

Matt Turck [2:35] And so you're fully distributed, but you have a large office in Paris these days. Is that fair?

Clément Delangue [3:00] Yeah. So we are a hybrid company, so people can work from anywhere in the world. Usually when there are a couple of people, three, four, five people, they take an office in the city that they are at. So I think we have maybe 12, 13 small offices all over the world, the biggest one being Paris, where we have, I think, around 20 to 25% of the team based here. The three founders of Hugging Face are French, so we have quite a lot of French roots, but we consider ourselves an international company.

Clément Delangue [3:19] So not an American company, not a French company, but an international company with people from all over the world.

Current state of open source AI

Matt Turck [3:30] You've become one of the leading voices in open-source AI. What would you say is the current state of open-source AI?

Clément Delangue [4:04] Three, four years ago, AI was extremely open and collaborative. All the research that was created was published. Most of the models were open-sourced, right? So, at the time, as you mentioned, it was GPT, GPT-2, there was BERT, there was XLNet. Obviously, Google kicked off this wave with Attention Is All You Need, which is like the seminal paper for transformers, which is the T in ChatGPT. Over the following years, the field became a bit more closed, a bit more commercially driven.

Clément Delangue [4:47] Some of the big players started to share a little bit less of their research, open-source models a bit less, actually, and put them behind APIs. And that's where we are today. I think the field is much more closed, much less collaborative, much less science-driven than it used to be. And I regret it because, for me, the progress of AI depends on being more open, more collaborative. As a matter of fact, I think we got where we are today thanks to this era of being open in terms of science and in terms of models.

Clément Delangue [5:14] So I hope we can go back to fostering more open science and open source in the future. I think it would be the right direction for the development of this technology.

Matt Turck [5:27] On the other hand, as you had predicted, I think, at the end of last year, is that fair to say open-source AI is basically on par in terms of performance with the commercial models?

Clément Delangue [5:59] Yeah, I think it's quite fair to say that it depends a little bit on the domains, on the modalities. So, for example, I think open source is actually ahead of closed source for most of the text applications today, especially when you have a very specific, narrow use case. Maybe not for a very generic, generalist use case like search, right? Search, you need more of a large model, and that's more provided by the API providers. But whenever you want to build a more narrow, specific use case, open source is generally better.

Clément Delangue [6:37] There are some domains where it's different. For example, video: I think video generation right now, there's not enough open source. So there are more proprietary approaches, but it varies over time, right? It's the typical technology cycle where, whenever there's a void in open source or proprietary, there are new startups emerging that are leading the field. So the same way you've had OpenAI, which was originally open source and very open for the first five years of their existence, the same way you've had a Stability AI or a Runway that's been built on top of Stable Diffusion.

Clément Delangue [7:07] I think in all the domains where there's not a lot of open source, there's a lot of room for new startups to be created and take a big position and have an impact on the field.

Matt Turck [7:14] And presumably, Llama is a breath of fresh air for the whole open-source AI movement.

Clément Delangue [7:46] Yeah, I mean, the thing to understand is that open source is really the foundation for everything else, right? So it's really the tide that lifts all boats, in a way. Even the closed-source companies are massively using open science and open source. As a matter of fact, OpenAI is using open source, is a customer of Hugging Face, and obviously is building on top of the open research, right? As I was mentioning, that's the T in ChatGPT, which is transformer and open research published by Google.

Clément Delangue [8:25] So, in a way, by contributing to this foundation, you accelerate the whole field of AI. You make it more accessible to everyone. I think one important fight that I'm picking is to try to build a field where you can create competition and where all startups can build AI instead of relying on a few companies, because I think there are very strong natural tendencies for concentration of power in AI. So thanks to open source, you fight these natural tendencies a little bit and create more opportunities for all companies to build with AI.

Matt Turck [9:11] Is there any way, in your opinion, that this can be sort of reaccelerated or encouraged? Because I guess you could argue Llama, perhaps there is an altruistic aspect to it somewhere, but fundamentally, I guess it's a question of Meta not wanting platform dependency for AI the way that platform dependency for Apple. So do you think, in a world where there's so much money at stake, people have any kind of incentive to do open source these days?

Clément Delangue [9:46] I think so, because I think compared to software, AI is much more science-driven. And I think if you look at science and how it's been done forever, it's more for altruistic motives, and it's been more in a way that researchers are publishing, sharing their research, and building on top of each other. So I think there is this incentive that is very strong for AI. As I mentioned, there are very important business opportunities in open-source AI. Sometimes people simplistically oppose open source with monetization, but actually, as we've seen in software, there are really great open-source companies, open-source commercial companies.

Clément Delangue [10:37] They are very successful, and I would expect there to be even more in AI. We'll have the MongoDBs, the Elastics, the Red Hats of AI, which are going to be open-source AI companies. So there are a lot of commercial opportunities with open source. And finally, I think in AI, it's such an important technology that I believe policymakers will require it to be shared as some sort of fundamental infrastructure for all. So there might be, in the future, hopefully, more and more policy incentives, maybe for more open-source AI as a way to give access, to be more inclusive.

Clément Delangue [10:55] With this technology, instead of concentrating it in the hands of a few.

Government regulation of AI

Matt Turck [11:23] Yeah, you anticipated my question. And I remember watching you from afar on TV in front of Congress. I think that was in June of 2023, which, by the way, as a French-born person, must have been quite an experience to show up and testify in front of the U.S. Congress. Anything that you have seen people either do or talk about from a policy standpoint that you'd be in favor of from an open-source AI perspective?

Clément Delangue [12:07] Yeah, this week is a good example in terms of regulation. If you take the specific case of California, because yesterday Governor Newsom vetoed and accepted two different bills, one which was really creating more constraints for small tech and everyone to build AI and was really having the risk of concentrating power, and another one that is more creating, I think, better incentives for AI, which actually mandates AI builders to share the datasets that their models have been trained on, which is an opportunity to create more inclusivity, understand the bias, the limitations of your models.

Clément Delangue [12:59] So, like on any technology topics, there's going to be some good regulation, some bad regulation. In general, for me, I'm in favor of the regulations that are creating more competition, more opportunities for everyone to build with AI, and tackling some of the end-user risks of AI today, like the risk of misinformation, the risk of biases, instead of focusing on more sci-fi-driven, long-term kind of hypotheses, like the doomsday scenario where AI is going to take control and kill us all. I think policymakers have made a lot of progress in the past few years in their understanding of AI.

What is open source AI?

Clément Delangue [13:18] and how they can approach the topic. So I'm quite excited to see what they're going to come up with in the next few years.

Matt Turck [13:36] By the way, what is open source AI? I may have asked that at the beginning of this conversation about OpenAI, but is that open-sourcing the model? Is that open-sourcing the dataset? There was this whole controversy about weights versus not weights. Where are we on all of that?

Clément Delangue [14:07] So it's a gradient, right? Openness is a gradient. From full open source AI, which would be considered a practice where you share the weights, you share the datasets that you use to train the model, you usually share the training scripts for reproducibility, to give everyone the ability to retrain the model themselves. That's the most extreme side of the gradient. And then the more you go towards more secrets—for example, you open-source the weights, you don't open-source the datasets—which, in my opinion, is still a good step and a step in the right direction, and still an impactful thing to do depending on your company constraints.

Clément Delangue [15:02] Of course, not everyone can share everything that they're doing. But in general, the more we can push towards the more open side of the gradient, the better for the world, because then you give more power for everyone to build with AI. You foster the progress of the field because everyone can learn from your mistakes or from what works to reproduce that. And the more transparent it is, it helps with biases. For example, you can see that some biases are contained in the dataset.

Open source AI: China vs U.S.

Clément Delangue [15:21] Helps with education because you can understand why the model is not answering that, but is answering like that. So the more towards openness, the better for the field.

Matt Turck [15:31] And this slowing down of open-source AI, is that more of a U.S. thing? Are different parts of the world behaving differently when it comes to open-source AI?

Clément Delangue [15:51] Yeah, it's more a U.S. thing. In opposition to that, I think China has been publishing much more open-source AI lately. I would argue that they're probably the leader today of open-source AI, which is surprising to some. But, for example, in video, they've been leading in open-source AI, which is kind of like, for us being in the U.S., can be sometimes challenging because, as it is kind of like the foundation for everything, it means that American companies are then building on Chinese foundations, which leads to some challenges.

Clément Delangue [16:58] We can take an example maybe with Grok, which is kind of like the ex-Twitter model that is based on open source. Let's say, for example, if they were using a Chinese open-source model for image generation instead of using kind of like an American model, then when you would use, in the U.S., like the Grok chatbots, you wouldn't be able to generate the same kind of images that you would with American models. So that leads to some very interesting questions. France also has promoted open source much more than the U.S.

Clément Delangue [17:19] And that's why I was happy to be at Congress, to try to explain and make the case for why it's important for the U.S. to keep fostering, keep promoting, keep pushing for open source.

Matt Turck [17:47] It's fascinating, especially in the case of China. Why do you think that is? For a while, there was this perception and probably concern that there was going to be a bifurcation and that there was going to be a Chinese world of AI and a Western world of AI. But it's really interesting to hear that the Chinese are actually publishing more in a way that precisely can be open to the rest of the world. Is that a power play of some sort? What do you think might be the driver?

Clément Delangue [18:13] I think it's a cycle, definitely. I think if the U.S. is open-sourcing less, there are opportunities for others to take the lead. I think maybe because they came in the technology cycle a bit later, they see it as an opportunity to catch up and take the lead faster. So it's always hard to kind of find the motives behind it, but it's a very clear pattern that we're seeing with most of the best open-source models now, with a lot of modalities coming out from Chinese labs.

LLMs vs. SLMs

Matt Turck [18:49] And then speaking of modality, another interesting topic that I've heard you talk about is large models versus smaller models and specialized models. What's your current thinking on this, of which one matters more, which one lends itself to open source more?

Clément Delangue [18:54] So we just crossed 1 million public models on Hugging Face.

Matt Turck [18:55] Congratulations.

Clément Delangue [19:25] So it's a very big counterpoint to the one model to rule them all. I call it a fallacy. I think big models, big generalist models, are good for some use cases, especially when you have a generalist use case. So if you're doing search and you want your search, or if you're doing ChatGPT and you want it to be able to talk about everything, a big model makes sense. Now, when you want to do, let's say, banking customer support chatbots, you don't really need it to tell you about the meaning of life, right?

Clément Delangue [20:16] You can use a much smaller, more specialized model that is going to be cheaper to run, faster to run, easier to retrain, more controllable. That makes more sense. And so I think that's what companies are realizing, sometimes using the big models as the first experiment, first proof of concept. And then when they want to go faster, save money, increase accuracy on their specific use case, then they switch to a smaller, more customized model. So I think we're going to end up in a world where there are both, right, on both sides of the spectrum: some extremely large, powerful, generalist, costly models for some use cases, all the way down to very specialized, simple, optimized models.

Clément Delangue [20:42] And depending on your use case, you're going to use different ones.

Matt Turck [20:49] Yeah, that's fascinating. And I guess that's a big part of the thesis, especially for the sort of enterprise and business part of Hugging Face.

Clément Delangue [21:17] But an interesting analogy is code repositories, right? To me, the way we approach AI is just a new paradigm of software. You wouldn't think that there's one codebase to rule them all, right? Like, you understand quite clearly now that every company can write their own code, have their own code repository for their specific use case. And I think it's the same for AI. So the same way today every company, every product has its own codebase, tomorrow every company, every use case is going to have its own optimized model.

Clément Delangue [21:54] So we sometimes say at Hugging Face that there's going to be as many models as code repositories today. So GitHub has hundreds of millions of code repositories that are all customized for specific use cases. In a similar way, hopefully in a few years, Hugging Face is going to be hundreds of millions of models for all specific use cases.

Are commercial LLMs just 'Training Wheels' for enterprises?

Matt Turck [22:22] To play it back, it sounds like you're saying that actually OpenAI or any commercial LLM may have a little bit of a graduation risk, that they might be a little bit like training wheels for the enterprise. As enterprises get more sophisticated, they may either switch away from those or at least have those be only a part of their ensemble of models. Is that fair?

Clément Delangue [22:35] I don't know if it's a general rule. It's something that we're noticing. We've seen a lot of companies who started their AI lifecycle more with a large model and then moved on to smaller models.

Matt Turck [22:37] Smaller and open source.

Clément Delangue [23:12] Yeah, these things are really hard to predict. I think at the end of the day, we're still very early in the technology cycle. Early technology cycles tend to be a bit irrational in a way, right? It's not so much the technology benefit that drives decision-making. It's almost more like the hype. It's even more like the word of mouth, like the perception of technology. So I'm excited to see in the next few years how the technology cycle evolves, how it matures, and if decision-making from companies, from enterprises, evolves a little bit and becomes a bit more rational in a way.

Clément Delangue [23:59] And I think it's going to be good not only for us, as a little bit biased towards this model, but also for the world in many aspects. For example, when we think about energy consumption, which to me is a very big topic, going towards more specialized, smaller, less energy-consuming models hopefully will make some of our challenges easier to solve. And again, as I mentioned before, I think smaller, more customized models also mitigate the natural tendencies of concentration of power in the sense that everyone can play with them, everyone can build with them, everyone can train them, compared to a world where you only have big models and just a few players are able to play this game.

Software 2.0: built with AI

Matt Turck [24:47] Software 2.0 analogy, if you could just talk to that. Software 2.0 being building technology and software with AI, which is actually, as much as people are excited about AI, not necessarily a widely held view that all technology is going to be built with AI. But is that your view?

Clément Delangue [25:27] So AI is a terrible name for the current technology because artificial intelligence is kind of like calling these concepts of independence, this concept of sentience, this concept that it's like a foreign kind of intelligence, obviously coming from a lot of sci-fi around the topic. The way we see it is just kind of like the new paradigm to build all software, in the way that before, the way you were doing technology was by writing millions of lines of code and kind of like in a really rule-based approach to building products.

Clément Delangue [26:24] Now you do it differently. You start from the data, from the datasets, and you train a model and you optimize a model, and then that's what kind of like is the foundation for your technology, your technology product. And I think that's the right way to approach this new technology and realize that companies are building this technology, controlling this technology, and iterating on it to progressively do the current use cases of technology better and also creating new capabilities. So we started by making search better with AI, and also we created new use cases like ChatGPT, which is like a new way of conversing with a technology system.

Clément Delangue [26:51] And hopefully in the next few years, we'll continue that way in the sense that we'll make current technology products, current technology use cases better, and then enable new use cases.

Matt Turck [27:18] But you're in the camp of that way of building code is the way that, fast forward 10 years, all code will start with AI, as opposed to somebody more of a copilot kind of analogy, where people will do it manually still, for lack of a better term, and then have AI by the side. You think that AI-driven code generation is going to be 100% of technology creation?

Hugging Face founding story

Clément Delangue [28:03] Yeah, I think it's really kind of like the new paradigm to build all tech. So I think in a few years, if you build a product without the help of AI, without AI, it's going to be like creating a shop without kind of like software 20 years ago, which doesn't mean that it's going to completely disappear, right? You still open shops today, maybe without software, but the norm will be to build products and technologies with AI, in my opinion.

Matt Turck [28:26] For the second half of this conversation, let's talk about Hugging Face, starting with the founding story. Hugging Face is one of those fascinating stories in technology. I guess we were talking about Twitter and X, a little bit like Twitter that came out of a podcasting company. Hugging Face came out of a different product. Do you want to give us all the details? What did that chatbot do?

Clément Delangue [28:58] Yeah, we started out of our excitement and passion for AI. And at the time, eight years ago now, we were like, okay, what's a scientifically challenging topic that we can work on, but that is fun at the same time, and ended up starting building this Tamagotchi AI, right? So at the time, it's like a Siri, Alexa, but we wanted to make it entertaining. We did that for a bit more than three years.

Matt Turck [29:01] So it was kind of Character.AI today?

Clément Delangue [29:37] Yes, kind of Character.AI or ChatGPT, but kind of like entertaining. Not saying that ChatGPT is not entertaining, but more focused on entertainment, I guess. And we were lucky to kind of work very early on on transformer models, on transfer learning for NLP. And the kind of origin story of the pivot is quite interesting with Thomas, one of our co-founders.

Matt Turck [29:38] Thomas Wolf.

Clément Delangue [30:15] Wolf, and chief scientist, coming up to Julien and me on, I think it was on a Friday afternoon. Oh, there's this thing that came up which is called BERT from Google, but it kind of sucks because it's in TensorFlow and most people today want to use PyTorch. So he said, oh, I think I'm going to spend the weekend hacking around it and maybe release something. And then on Monday, he releases a PyTorch implementation of BERT and tweets about it. I think at the time we got maybe like 1,000 likes on his tweet on Twitter.

Clément Delangue [30:53] And we're like, wow, we broke the internet. At the time, we were literally like nobody, didn't really have much visibility. And then progressively, we realized that there was something around kind of like not just creating the final application, but providing a platform, providing technology for companies to adapt this, adopt this new paradigm for building software. So after, I think after a few months, I think in addition to BERT, we added, or the researchers from the labs added, XLNet, which was from Guillaume Lample at Meta, who's now the founder of Mistral.

Clément Delangue [31:19] We added GPT at the time from OpenAI, and then progressively it evolved into this platform that is Hugging Face today.

Matt Turck [31:23] And that became the Transformers library.

Clément Delangue [31:42] Transformers library. Yeah, that's kind of like—so the first repository was PyTorch-Pretrained-BERT, which became PyTorch Transformers when we added more models than BERT, which became Transformers.

Matt Turck [31:49] Which is one of the most successful open-source libraries of all time. And the G in GPT.

Clément Delangue [32:32] So that's one of these examples of pivots, right? Of this ability of startups to be flexible, to be opportunistic, to follow when you see some demand, some kind of potential in the market. And for us, it's also a good example of community-driven evolution of a startup, because it's really kind of like the community that gave us the signal that it was useful, and then the community that helped us make it better by adding more models from the early models to then contributing more.

Matt Turck [32:43] When was that moment? Because having a highly successful library is one thing, but then turning yourself into a platform to host a bunch of different libraries and models and all the things, when did that happen?

Clément Delangue [33:28] It started quite early, even in the first days of Thomas releasing the first port of BERT. I think contributors, open-source contributors, started to solve bugs and improve some small things. And then progressively, as we added more models, more contributors started to add more models. So progressively, the community contributed more and more, and we felt this movement where the more we contributed to the community, the more open source we did, the more the community was giving us back. So it validated us into this approach to today, as I mentioned, where we have 5 million AI builders using our platform.

Clément Delangue [34:24] who collaboratively shared 1 million public models. Half of them have been downloaded in the past 30 days. So most of them are actually very useful and active for the community. They also contributed, I think it's more than 200,000 datasets to the platform. So it's open datasets that anyone can go and use to fine-tune, to customize their models for specific languages, specific domains, specific use cases. And collectively they built over 300,000 Spaces, which are the apps on the Hugging Face platform.

Matt Turck [34:55] When did you feel you sort of had it? This idea of being the GitHub for machine learning or AI, rewind back to whenever that was, 2014, '15, '16, '17, '18, that general period of time, there were a number of companies that were sort of talking about doing this, and none of them really ever took off. And in some ways, GitHub never became the GitHub. Why do you think that is? And why do you think you were able to break through? Was it perfect timing?

Matt Turck [35:00] Was it transformers? What were the reasons?

Clément Delangue [35:29] Luck. I think it's a big aspect in anyone's success. Sometimes we forget to talk about it, and we kind of look back at things, and we're like, okay, I had the perfect strategy, I had the perfect approach to things. The truth is that we got lucky in many ways. We got lucky in terms of timing, as you mentioned. I think we were at the right time, at the right place. We were lucky in terms of who we were as founders compared to, in alignment with what we were trying to build.

Clément Delangue [36:09] Before, we had some experience around consumer, and so I think it helped us to have a very community-driven approach instead of maybe a more enterprise-driven approach or a more traditional kind of like B2B approach. And then we got lucky that the community adopted us and started to contribute to our platform and that we could collaborate with a lot of the products that were around at the time. Right from the beginning, actually, we took much more of a collaborative approach than a competitive approach.

Clément Delangue [36:40] And so it helped us kind of like to focus on where we were adding value, in a way, to the community, where we were building kind of like something useful for the community, and usually integrating with other offerings and other technology products when it was making more sense.

Are there any competitors?

Matt Turck [37:12] Because right now, you are in the AI ecosystem in a pretty extraordinary position, which is that in a world where you have the OpenAIs and the Geminis and Anthropic competing neck and neck, every other day there seems to be one that beats the other with the latest models. Is that fair to say you pretty much have no competitor? I mean, it sort of feels like everybody has given up on trying to be the next Hugging Face.

Clément Delangue [37:52] Yes and no. There are always some sort of alternatives that emerge regularly from a whole bunch of players, from big tech cloud providers, sometimes from more emerging startups focusing on some parts of the platform. What's exciting to us is we have some sort of a network effect as a platform for collaboration in AI that I think makes us in quite a unique position. I think what we're excited about, too, is that we're profitable these days. And so we managed to find kind of like a business model that creates some sustainability for us and a way to see the future without having to take short-term kind of like decisions, which is important for us as a community platform.

Clément Delangue [38:25] In a way, we want to make sure that we're here for the long term, for the community. So that's kind of like the things that we're excited about.

Matt Turck [38:29] But what a concept! You mean you don't burn?

Clément Delangue [38:53] Yeah, in AI, people don't understand sometimes. Like, you mean profitable just on the inference, right? Not on the training? Profitable based on the billions of dollars that you raised. Yeah, it's quite unusual for AI startups, but I think we'll see it more and more in the current cycle of VC funding too.

Matt Turck [38:56] You think that's going to dry up?

Clément Delangue [39:24] I think, as I was mentioning, there are a lot of investors who caught up and became obsessed about AI, leading to some kind of irrational decisions. And that changed a little bit in the past year, 18 months maybe. So it's definitely proved to be a bit harder, I think, for most startups in AI these days to raise than it used to be. So we're seeing, for example, a lot of M&A activity these days, both for kind of like big organizations, right, the Databricks, the Inflections of the world, but also for smaller organizations.

Clément Delangue [39:54] For example, I receive tons of inbound these days from early-stage startups wanting or getting excited about being acquired by Hugging Face. We actually acquired two in the past four months.

Matt Turck [39:57] Yeah. What were those, while we're on the topic?

Clément Delangue [40:12] So we acquired a company called Argilla, which has been working for a few years now on the topic of datasets and collaborative annotation of datasets. Really great team based out of Spain.

Matt Turck [40:17] So it's a product expansion for you, the reason for the acquisition?

Clément Delangue [40:36] Yeah, topic expansion, because datasets are growing fast on Hugging Face, and we think it's a very important artifact of the AI building process. They raised actually from ENIAC here, based in New York, and those are great investors. And the second one that we acquired is called XetHub, which is a team that did a previous startup that got acquired by Apple, then worked for a couple of years on the AI platform at Apple, and then started their new company a few years ago that is joining us from Seattle, which is building very interesting infrastructure for models and datasets, specifically kind of like enabling the ability for better versioning for models and datasets.

Clément Delangue [41:53] Basically, a simplistic way to understand it is that today, when you have a dataset of, let's say, a million data points and you want to update it with 1,000 more data points, you almost have to re-upload 1,000,000 data points, versus their technology that is replacing an existing technology that both Hugging Face and GitHub are using, is enabling you to kind of just add the 1,000 data points. So, enabling much faster AI building and much faster iteration both on models and datasets. So we're excited about both.

Matt Turck [42:10] What do you look for in general? What's interesting to you for anybody listening to this that feels like working with Hugging Face would be a lot more fun than being by themselves? Like, what are you looking for?

Clément Delangue [42:45] We do it quite similarly to how we hire, in the sense that we don't really take kind of like a top-down strategic approach to it where we're like, okay, there are these topics, these topics, and we need to acquire companies in this topic. Instead, we're trying to find teams that share our culture of doing things, our company culture, share our mission of democratizing AI thanks to open source. And then we assume that if we bring in teams like that, they'll manage to find their impact.

Clément Delangue [43:30] And then we adapt our roadmap and our evolution based on them. So what works for us is really teams that are impact-action-driven and like to work in a very decentralized approach with openness and a strong drive to contribute to the community. And then we can acquire different stages. As I said, we're lucky to be profitable. We raised a bit less than $500 million so far. Most of it is still in the bank. So we have the chance to kind of do a lot of different operations depending on what we are excited about.

Matt Turck [44:02] We had started going down the path of a little product tour, which I think would be fun to spend a few minutes on. So you mentioned models, datasets, and Spaces. So, just to go back on models, I'm actually curious. So it looks like the top ones are the Llamas, too, as of last night when I was looking at this. Are there any models that caught your attention or the team's attention in terms of being maybe less obvious or rising fast or anything you'd highlight?

Most interesting models on Hugging Face

Clément Delangue [44:44] Yeah, my favorite page on the Hugging Face platform is the trending model feed, where every week you can see new models emerging in a lot of different domains, right? Obviously, we talk a lot about text with Llama, but I'm actually much more excited about other modalities. I'm much more excited about audio now, video, image, biology, chemistry. And actually, if I were redoing the talk that I was giving, that NLP is going to change the world, now I think NLP has changed the world.

Clément Delangue [45:18] And it's other modalities that are going to change the world. It's going to be audio, it's going to be video, it's going to be biology, chemistry, AI. These are, in my opinion, the domains that in the next few years are going to change the world the same way NLP has in the past. So right now, if you go to the trending tab of the Hugging Face platform in terms of models, for the first time, I think in the first or second spot, you have an OCR model.

Clément Delangue [45:58] So you give it an image and it extracts the text, which is interesting. I think it's the first time you have this kind of model in the very popular models. You're starting to have a lot of video generation models. I'm extremely excited about them because I think we're going to have the Stable Diffusion model for video generation soon. There are a lot of proprietary models in video generation, but the first companies that are going to release an open-source video generation model, I think it's going to have a tremendous impact the same way Stable Diffusion has had a tremendous impact on image.

Clément Delangue [46:50] And then one model that I've been excited about this week, too, is a model that has been released by IBM and NASA on climate prediction, which is not only super interesting, but also super positively impactful. It's a typical example of AI for good because if you look at climate, and especially the ability to predict severe climate events like hurricanes—I'm in Florida most of the year.

Matt Turck [46:51] Near and dear to your heart.

Clément Delangue [47:24] I know a thing or two. You can actually save tens of thousands of people if you manage to predict an event like that just a few hours earlier. And so this model, that is called Prithvi, is intended to serve as some sort of a foundation for better research and better models for climate prediction. So I'm excited to see these kinds of models being released by IBM and NASA in the US.

Matt Turck [47:34] So models, datasets, Spaces. Spaces, you were saying, is the place where you can basically showcase applications, create applications.

Clément Delangue [48:13] More and more in production, we see more and more startups using Spaces to power their apps. I think we have over 300,000 public Spaces that have been created, and quite a lot more that are private. One thing that people don't really realize is that a big part of the usage of the Hugging Face platform now is in private mode, almost as much as what is public. So there's almost 1 million private models on the platform. And it's basically companies using—

Matt Turck [48:20] You said you had 1 million public and also 1 million private.

Clément Delangue [48:56] Yeah. And it's basically companies using the Hugging Face platform privately as their main platform to build AI, which kind of ties into the most important features that are actually the internal and external collaboration features of the platform. So you not only host your models, your datasets, your apps, but you collaborate on them, right? So it's going to be one researcher that is sharing a model, then someone else who can open a pull request, that can comment on the model, that can test the model and tell you, okay, there's a problem for this use case.

Clément Delangue [49:21] For this use case, we should improve it. That can contribute more data, more datasets. And so the collaborative approach, the collaborative aspect of the platform, is very valuable and what most enterprises are using today.

Matt Turck [49:25] And that includes evaluation tools as well and benchmarking tools.

Clément Delangue [50:02] Yeah, you have some integration with some evaluation tools. For example, it makes it really easy to create leaderboards on the platform. I think we crossed 5,000 leaderboards that have been created on the Hugging Face platform as a way to rank your models based on their accuracy and their performance. And the way we're trying to do things is in a very open, modular way, in the sense that companies don't have to use everything end to end. It's more like they pick the features and the functionalities that they like.

Clément Delangue [50:28] And because it's mostly open source, it's integrated with a lot of other services and sometimes with their internal services and stack. And so they can really pick what makes sense for them at some point to make the whole AI-building process just simpler.

Shifting focus in enterprise solutions

Matt Turck [51:02] And so let's get into the sort of business model part of Hugging Face. And so you alluded to the enterprise side of things. So there is $9, $20, and then the enterprise tier. Maybe talk about the enterprise tier, which sounds like it's the bulk of the business. And at some point, it sounded like you had a significant consulting business around enterprise customers, but it feels like it's evolved towards being more of that sort of collaboration software platform. Is that fair?

Clément Delangue [51:30] Yeah, I mean, the way we approach things is that we are a platform for AI builders, right? We think most of our usage and users are always going to stay open source and free, and we need to build some sort of a freemium model, right, where a small percentage of the usage and the users are going to be paid and kind of fund the rest of the platform. And when you think about the difference between open source, free, and premium, we identified three different variables.

Clément Delangue [52:21] It's like premium support and premium features. And usually these appeal the most to enterprises who need advanced features and more enterprise features, especially around user management, around security, and things like that. And then you have premium compute, premium infrastructure, which is very important for AI. And we do this with collaborations with the cloud providers and with our own cloud offering, with Inference Endpoints and Spaces GPU. So that's how we approach things.

Matt Turck [52:30] What would be an example or two? And I don't know if you can give names or not give names, but of large enterprises using all of this.

Clément Delangue [52:55] I mean, most large enterprises building with AI are using us one way or another. They're using different blocks, right? Not everyone is using the same thing. But in terms of maybe the largest, all the big technology companies are customers of ours and partners of ours, right? So when you think about Microsoft, when you think about NVIDIA, all these companies are massive customers of ours, to smaller enterprises, all the way to the independent developers that are using us too.

Clément Delangue [53:33] The thing that has been growing the most for us lately is this concept of Enterprise Hub, in the sense that we have 100,000 organizations using the Hugging Face Hub. And the idea is for a small percentage of them, especially the most expert ones, they have requirements for more user management features, security features, like a security scanner that will scan the new open source models and make sure that it makes sense for their environment.

Clément Delangue [54:14] So we have over 2,000 companies using the Enterprise Hub today, from Mercedes, Bloomberg, all the way to larger ones like NVIDIA, for example, that I was mentioning.

Matt Turck [54:44] So with the $20 team tier, I have access to—I can have private models and I can collaborate. And then separately, I can get the Inference Endpoints, and I'm going to pay on a per-usage basis. So I'll pick my model, pick my provider. And then the Enterprise Hub is where you have all the enterprise-grade features of security authentication.

Bloom & Idefix

Clément Delangue [55:06] And usually they consume kind of like a bundle of premium features: premium support, because usually they value our help to maximize the usage of the platform, and premium compute. Usually, they value kind of like a bundle of the three premium offerings that we have.

Matt Turck [55:21] You don't develop your own models. Is that fair? Because obviously you've got the Transformers library, which is a library, but you all were very involved in Bloom.

Clément Delangue [55:21] Yeah.

Matt Turck [55:29] Was that, like, one exception that's just not part of the model, or do you view yourself as developing some models? Or maybe you do actively?

Clément Delangue [56:01] Yeah, we do develop some models. Not so much as our core business model, like some other companies, but more as a way to show what's possible, lead by example, and create value for the community. So when we do, we release everything in open source, not only the models, but usually the datasets and the training scripts for it, like we did with Bloom. It's an interesting example because at the time there weren't many open source large language models.

Clément Delangue [56:48] People were saying that it was too dangerous to open source large language models. And at the time, there were several teams building large language models. There was BigScience with Bloom, and Meta also was doing a foray with something called OPT at the time. And actually, some of the OPT researchers told us after the fact that the fact that we released Bloom in open source really helped them in their internal case to release OPT. They were internally saying, "If Bloom could be open source, we could open source OPT."

Clément Delangue [57:31] And they ended up actually open sourcing OPT, which in some ways kind of created the open-sourcing dynamics at Meta. So, in a similar way, when we're seeing gaps in the market, in a way, in the field for things for open source, we take advantage of the fact that we have great researchers, that we have resources to train models. So, for example, we released a few months ago something called Idefix, which is a multimodal model. Before that, there weren't a lot of open source models like that, obviously with a French name.

Matt Turck [57:44] Yes, exactly. I'm smiling, but that would be a very obscure reference for anybody who's not French.

Clément Delangue [57:54] Yeah, we realized after the fact that it actually wasn't a good idea because in the US, nobody's getting the reference. It's very arcane for them. Idefix—what does it even mean?

Matt Turck [58:20] So, for reference, it's part of a cartoon strip that any kid in France of, I guess, a certain age would have grown up with. And Idefix was the dog. The dog, right? So it's a story of a Celtic Gallic tribe called Asterix and Obelix.

Clément Delangue [58:36] Yeah, the dataset. So the model is called Idefix. The dataset is called Obelix. But we do train models. We also release open source datasets whenever we feel like it can be impactful for the field.

The culture of Hugging Face

Matt Turck [59:05] Well, great. Maybe to close the conversation, maybe a few words on a combination of entrepreneurial lessons and the Hugging Face culture, because I think you guys have a very strong identity as a company that doesn't necessarily follow the classic Silicon Valley playbook by the letter. Do you want to touch upon that?

Clément Delangue [59:47] Yeah, we're trying to build a different culture. We're not really interested in building a successful company just to build a successful company. If we end up with a company that is similar to companies that are existing today, it would feel like a failure to us as founders. So we've always been very intentional about doing things a bit differently in ways that are more aligned with our values and what we want to improve in the world. So we're taking a very decentralized approach to company building where we're trying to give extreme freedom and ownership to all team members.

Clément Delangue [1:00:46] To be able to focus on what they're excited about, what makes them happy, what they think can help the world and contribute to the field. So we're very decentralized, we're very impact- and action-based. We do little to no strategy, no planning, no roadmapping. We have zero product managers. We're trying to build this very lean organization where everyone can have an impact and where the pace of shipping, contributing, building follows the pace of the technology. I think we're in a domain where if you slow down, you're going to die because the train is going to pass, and very fast you're going to be outdated, old-fashioned.

Clément Delangue [1:01:29] You're going to lose your ability to attract the best people in the field. And so you're going to reach a plateau really fast, which is going to lead to you dying or having to sell your company, like we've seen with many others, like Adept, Inflection, and companies like that.

Matt Turck [1:01:35] And I've heard you say that you have either no or very few KPIs.

Clément Delangue [1:02:09] We think in a way that a lot of the rules and best practices have been built for the previous generation of startups, right? The software startups. So, for example, sometimes when I meet entrepreneurs and I can see that they have a lot of the Lean Startup principles kind of in mind, I tell them to trash the book. No offense to the author, but I think a lot of the things that we learn with traditional software are outdated with AI. And actually, if you look at the most successful early companies in AI, like OpenAI, like Hugging Face, or others, you can see that they didn't really follow the traditional company-building best practices that we were taking for granted.

Clément Delangue [1:02:56] So we're trying to do things quite differently in that way. I think it's the same for investors, actually. I'm doing my own experiments at a very small scale, but I've invested in almost 100 startups in the past 2 years, and I've tried to take a very different approach than what I would do if I was investing in more traditional software companies. Like, for example, the founding teams of AI startups have changed quite drastically from, for a software company, the ideal team maybe being a software engineer with a business person, to in AI sometimes being all scientists, all AI scientists being sometimes good founders.

Clément Delangue [1:03:32] So it's interesting to see how this different paradigm calls for different ways of doing things. And we're trying to embed that and to kind of embody that with Hugging Face.

Matt Turck [1:03:57] While we're on the topic, in your portfolio, that's your chance as an investor to start pimping out your companies, which is what us VCs spend half our time doing. But jokes aside, any company that you would highlight in your portfolio, either because they're doing well, but perhaps more interestingly, companies that are not yet really on the radar that people should know about?

Clément Delangue [1:04:26] It's hard to pick one, especially because obviously investing is not my job and I'm like 200% focused on Hugging Face. So I'm doing that a little bit more in my free time. I've been excited about video generation these days. So a company that I haven't invested in but that I was excited about these days is Hotshot, which is like a video generation app. But yeah, there are also a lot of startups that I've invested in, or not invested in, in domains that are harder, I would say, like biology, chemistry, that I think could have tremendous impact on the field and the world in the next few years.

The future of Hugging Face

Matt Turck [1:04:57] So maybe to close, next year for you, for the company, for your investment portfolio, any plans, projects, roadmaps, like anything that you look forward to?

Clément Delangue [1:05:31] Well, we have 5 million AI builders today. I think ultimately there's going to be more AI builders than software engineers, right? Software engineers, you can think maybe today there are 50 million. That's kind of like a safe bet. So for us, the goal is really to keep growing the number of AI builders that are using us. Focus on everything non-text. As I said, now NLP is boring, and I'm more excited about images, video, biology, chemistry, time series. So this is a big area of focus for us.

Clément Delangue [1:06:04] Datasets, which are a big way that you can contribute to AI today, almost as much or more than models. When you release a dataset, for example, we released a dataset called FineWeb that I think has been used in over 1,000 models. So when you release a dataset, it's actually an artifact that can have a massive impact because it can be used in many models. So I'm excited about datasets. And last one, maybe evaluation. You mentioned it a little bit earlier.

Clément Delangue [1:06:36] I think we're at the very early days. People are doing evaluation the wrong way, just looking at one massive public leaderboard that doesn't really tell them much about how the model is going to perform on their own use case. So I'm excited to see more people focusing on evaluation and making progress on how do we define if a model is good or not good for your own specific data use case.

Matt Turck [1:06:39] Well, Clem, thank you so much for doing this. It's been wonderful. Really appreciate it.

Clément Delangue [1:06:40] Thanks for having me.

Matt Turck [1:07:02] Hi, it's Matt Turck again. Thanks for listening to this episode of The MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already, or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build the podcast and get great guests. Thanks, and see you at the next episode.