Name: The Next Frontier of Open-Source
Uploaded: 2025-05-06
Duration: 17 min 8 s

An early pioneer of large language models, Hugging Face is best known for its vast repository of open-source and “open-weight” AI models. With its recent acquisition of robotics company Pollen, Hugging Face is now exploring the development of AI “world models” and their transformative role in robotics. We’ll discuss what this means for the future of AI and its growing impact on the physical world. 

The Next Frontier of Open-Source

Conferences

OK. Thomas, thanks so much for joining us. Thank you. Um, so, hugging face, as many of you probably know, is, is a, is a repository of open source and open weight AI models. Uh, and for a long time, it really seemed like the open uh weight models or the open source models weren't quite as capable as those that were being developed. as proprietary models by some of the leading tech companies like Anthropic or OpenAI. But then along came DeepSeek and it's R1 model back in January. And can you tell us just, you know, from your view what the significance of that moment was and what's really kind of changed with open source since since Deepsea? Yeah, I, I often call this moment the, the chat GPT moment for open source AI which is the, just like chat GPT was the moment the, the whole world kind of discovered AI I would say outside of the the AI sphere. I feel like deep sig was the moment the whole world discovered there was kind of this open source AI stuff, you know, outside of the open source field. So for us we've seen, we've seen the Deep Sync team uh shipping extremely impressive models since more than 18 months, I would say. So we were not very surprised to see this model, but I think that's really the moment people discovered that existed. This being said, I still, I still think a lot of people didn't really understood what it was, right? So a lot of people thought you had to go through Chinese servers to use it or through the Chinese government app or through Deepsy, but the reality is an open source model, an openweights model is basically a model you can download and use on your own laptop. Nobody's going to check what you send and that's the main interest of Openweights model is the privacy is very strong because you don't send your data to an external provider or server. And so this is the moment for open source, but the main reason this became famous is because for the first time we had a model that was really challenging the performance of OpenAI or entropic model and on some benchmark actually beating them. So now open up uh catch up and the 03 or 4 models I think again above this this threshold, but Deep R2 is coming, so I think the, the jury is still out. So basically there is now a blurry, I would say line and for I think everyone probably has this, this, this feeling which is it's getting hard to really tell what is the best model. They all seem like. The the top model 3, the latest Enroy, Gemini, Deepig, they all seem to be actually very close in performances. I know there's this concern actually that benchmarks are becoming kind of meaningless. Uh, what are you guys doing at Hugging Face in that regard in terms of trying to to make sure that there are benchmarks that people can actually rely on to see which of these models is better at a particular task than another one? Yeah, it's a good point. I think the world of benchmark has evolved a lot. So we used to have this very academic benchmark. They were mostly measuring the knowledge of the model. I think the most famous being MMLU, which was basically a set of like graduate level or PhD level questions. The model had to answer. Um, these benchmarks are mostly all saturated right now, and the most interesting is, I think, are two types of benchmarks going into 20225. The first one are more agenttic benchmarks where the models are not only asked to answer on the fly, you know, knowledge question but acting to do tasks. So I think there is still a lot of questions around how to measure this. And the second part is that I think benchmarks that are more, you know, adapted to each use case or each application. We had a, we have released just 3 weeks ago an interesting project on this which is called Your Bench. I can talk a little bit about it. Basically in your bench you'll give a couple of documents of the type of work that you do or like if you're in a company and working on legal domain, you put a couple of legal documents and it automatically generates for you a specific benchmark for your documents that you can then, you know, apply on different models to see which one is actually good for your application. Because if these models are all roughly the same on this academic benchmarks, it doesn't really mean they're all exactly the same on each vertical use case, of course, right, of course. Um, I just want to remind you, I will come to the audience for, for questions in a moment, so please think of your questions. Um, but first, about a month ago, Hugging Face made a big announcement that it was acquiring a company called Pollen Robotics, which is a French company that makes an open source humanoid robot, and this is a big shift for you guys. I mean, you guys have always been doing software. Now suddenly you, you, you bought a robotics company. Why? Why did you buy a robot company? Right? Um, I think it makes a lot of sense, um, for two reasons. The first one is, um, We're always trying to understand where AI is going, right? So when we started, we were a game company. We discovered open source and we kind of embedded on text AI and then we expanded to video and to speech, basically trying to to follow, you know, where is is going and in each of these areas to try to build open source communities of people who are building AI. Um, what I've seen in the past 18 months is that robotics is on the path to become, you know, yet another field that AI will tackle successfully, which means I'm pretty confident that in 2 years if we are here sitting again at the 14 conference, we'll have robots around that will be maybe serving drinks or whatever, you know, maybe interviewing me or I'll be a robot. We don't know yet. But what, what we see is that preemption, walking, all these kind of complex tasks. are getting sold now and what I see as well is that there is a lot of companies starting there, but they're all closed source company and the same we had when OpenAI started getting close. We don't really want a future where you actually, you know, always rely on a couple of incommandants to provide, you know, a technology that's that's very important. So we started one year ago with a software project called E Robots which was led by Remi who joined us from Tesla where where he worked on on the Tesla robots. And, and we started with this, which was a pure software library basically giving access to the latest AI breakthrough in robotics on software level, and this got a huge, huge traction. So this robotics, this robotics community has been growing amazingly. And so earlier this year when I was talking with our basically longtime friends at Poland. I kind of understood that we also had to focus on the hardware stack because in robotics, the hardware models and how to build robots is as important as the software. So see if we can just like we did in software we try to to to give the basic building blocks so that everyone can build a company on top, can build a successful startup on top. We want in robotics hardware to be to give the basic building blocks so that everyone can actually build increasing incredibly big startups. So the idea is that the hardware itself would be open source. The accentuators, all the things that make. A humanoid robot move, you would be able to either obtain those the plans for those so you could build your own versions of them. Yeah, so there's two main ideas. One, the long term is what you're describing making open source actuators, firmware, all of that. So you could either get the plans or anyone can build them which will drive the price cost quite lower and which will also bring some standardization. And the other, the other part is to build a community of roboticists. Uh, which are people who actually know how to make them and and uh feel empowered to actually create their own. I want to ask, what is, what is having sort of, you know, an open source robotics project unlock in terms of some of the remaining challenges in AI and and getting to this kind of goal of of humanlike AGI? I know you've talked a little bit about this, but there's certain things you feel like having an embodied system really will help unlock. Yeah, I think um Well, there are things you can learn live, I think there are things you can do. I mean, one thing I'm aligned quite well with with Sam Altmann, we had a discussion around this open sourcing OpenAI, which is if we have a future where actually AI is extremely intelligent. We don't want the remaining thing to be just human basically working to bring stuff to the AI. That would be a very sad future, right? We'll just bring the contract to get signed because that's the only thing the AI cannot do. So, and for the same reason we had this meme which I think is very true, which is I wanted AI to do my laundry and clean my dishes so that I can be creative the rest of the time. And right now we kind of have the opposite, right? AI is very creative and we are still like cleaning dishes by hand, right? So we need to, to, to at least give the, the power of human to to automate the second part as well, right? I know we've got at least one question in the audience. Let's go to the, go back there. Um, please say your name. Samsa is a business school. So thank you for this very interesting, very interesting. I have a, I have a question about the sustainability of the business models of the, the companies that are providing the open weights models, whether in robotics or LLMs. Should we expect them to continue developing and providing models? Yeah, it's a good question because we've we've seen a lot of companies starting by open sourcing some models and then they move to closer, closer system, right? I mean OpenAI being maybe the first example, but more recently we've seen that quite a lot, um. So it depends. I mean, I would say it's quite counterintuitively hugging face being, for instance, a very open source company is cash flow positive, right? We have, we have actually a sustainable business model. So you can definitely build good business model around open source. It's not, it's not as paradoxical as it may seem. Open source typically brings you a lot of visibility, a big community that you can leverage in other ways. Um, another thing we've seen repeatedly and and I would be ready to bet that it will continue is each time we every company kind of stop open sourcing we see. This has a lot of people start to see this as an opportunity to establish themselves as the new company who now open source. So Deep Sick gaming maybe because M was open sourcing less. If Deep Six starts to stop open sourcing, maybe another, another team will, you know, come and say, Hey, we're going to open source the best model to be the providers of the best model. So there is this kind of constant stream of people and teams joining the open source movement that I'm quite confident we stay. And for a hugging face, your cash flow positive, you say, what, what are you making money on? What's the revenue from? So we, we, we're quite lucky to have a business model that's really aligned with open source. So for us, our goal is really to have the maximum number of people participating in this kind of open community and sharing models, right? So we try to encourage that. So we have two ways we partners. One is very large partnership with a couple of companies to kind of help them incorporate our catalog of models in their own offerings. So basically the more models we have, the better it is for them as well. And the other is the enterprise version. The public hub. So if you're an AI builder, if your team is actually training AI models, it's quite likely that you would like to have the same hub that we have internally to version your models, to share them internally, and this is our latest product since we started to commercialize this last October and it's been growing extremely fast. We have several 1000 companies already using this product. OK. Other questions. He's also the question. I've got more if I don't see any. We got one down here. get them wait for Mike to get to you. I can also say again the question if you want. There we go. Hi, my name is Matthias Schroeder from the Mindshift. Uh, I just wanna, um, ask you about art and AI. Um, hugging Faith is like the number one platform for people who are active in AI and music and other fields. Uh, are you aware of that? Is that important for you? Um, what's your thinking about the future of AI and art? Yeah, um, it's a good question. I would say something I would like to push more because when you look at, so I see that mostly as uh. My way is to see that mostly as artists using AI tools to actually expand what they can do or to have new ideas. I think that's the most exciting use of AI. I personally am involved also in the AI art communities in Amsterdam with a small side project, and what I would love to do is to see them use more open source tools because most of the tools they use right now are closed source, which is a bit sad because if they use open source they could actually Given some knowledge and a bit they could actually tweak them a lot more and play with these tools a lot more. So if you know actually someone I've been trying to find the right person to carry to push this project hugging face and to see how we can, you know, maybe educate artists or maybe make some specific tools that make it easier for them to use open source AI tools would be amazing, I think. Uh, you just mentioned you're based in Amsterdam yourself. Uh, here we are sitting in the UK, um, the CEO of Hugging Face Clem lives in the US. How do you view, you know, the evolving geopolitics of AI? You know, there's getting increasing the sense that we're in this bipolar world and you have to choose to either use US models or Chinese models. So what do you see as the role for sort of Europe and is there a, can there be a Third pole, uh, in that, in that equation. You want me to give an answer in 3 minutes the geopolitics of AI, um, and what I see is I'm, I'm extremely, so we, we're kind of a world company. We, we're based both in the US, indeed in Europe with a very small team in Asia as well, and we try to kind of stay with this feeling. That we would like. This being said, I'm still European living in Europe. I would love to see Europeans strive. I would love to see a lot of startups grow there. I'm pretty optimistic at the moment. I think I see amazing young teams starting. I think that the move from a big focus that we had last year on only foundation models and only talking about the cost of training them. The move from that to what we see now, which is basically people assuming this foundation model exists. You have a lot of them you can choose, um, take the best of them and actually focusing on building application is very exciting because this can be built with small teams in Europe. I see a lot of teams. I mean, maybe the biggest thing we saw recently was Lovable emerging out of Stockholm, but there's many other teams that are basically growing making amazing apps. So yeah, view on this. I'm very optimistic about coming. Great. We're running low on time, but I wanted to ask you, you know, you talked a little bit about about human robotics, humanoid robotics, and what's why that's important. What else do you see is sort of next for AI, and do you think that our current kind of architectures, these transformer-based architectures, will really get us to the level that we want to get to next? A good question on many aspects actually. I, I, I was talking today with some people. There is an amazing team in London building uh novel chips, novel superconductor, semiconducting, uh, things for yeah, and one big question for them is can we focus on transformers and these architecture? Are they going to be, um, my gut feeling is we can still do a lot of things with. Forers and what we have now with better data, we saw this other shift to reasoning model where basically we keep the same type of model. We just let them think for longer and we keep getting improved performance. So that's quite exciting. So on this, I think we can still extract a lot of intelligence from the same basic building blocks that we have. Um, and then the other part I'm very excited about, and I think maybe we talked about it a little bit last time, which is AI and science, which is basically all this knowledge we have on training, very good generative and predictive models applying that to other things like predicting good materials. Also here in the UK there is a very great company called Second Low predicting weather forecasting and trying to predicting another one is predicting proteins. All of this basically using all the things we learn in AI to actually have. Maybe you could say real impact on the world in a way, but that's the physical impact on on science. Yeah, so AI is a tool for science. That's a great optimistic note to end it on. Thank you so much. Please give Thomas a round of applause and thank you all for listening to you. But what a fabulous day we've had. A big thank you to our founding partner Accenture and all of our speakers, moderators, and of course to all of you for an amazing first day. Tomorrow we kick off with roundtable discussions over breakfast at 8:45 a.m., then we'll meet back here for sessions starting at 10:00 a.m. Join us now for a drink and light nibble before you head off to dinners. Uh, coaches will be leaving for those dinners from the front drive promptly at 6:45. So get your drink quickly and then then go. Um, have a wonderful night and we'll see you all tomorrow.