China, Robotics, & Open-Source AI | Clem Delangue

Yeah, this is a LeRobot Mini. This is an open-source desktop robot for AI builders that we introduced a few months ago. Already over 5,000 people pre-ordered it, and we're starting to ship these little birdies all over the world.

Yes, we're starting to ship them. I received mine actually last week. It's fully open source. Most of it you can 3D print, and it's fully programmable, meaning that it doesn't come so much with pre-installed apps. As an AI builder, you can build your apps yourself and then use these apps at home. For example, a lot of people are using them with their kids, playing hide and seek, red light, green light. And you can share them with the community. So we hope that AI builders all over the world are going to build apps for these with the latest AI models, the latest open-source models, and share them with the world.

Yeah. So at Hugging Face, we're the platform for AI builders. We have 11 million AI builders using our platform. And a year and a half ago, we started to see more and more AI builders playing with robotics. One of the challenges they were encountering was the lack of affordable hardware for them to experiment with.

Exactly. I feel like when an AI builder is starting with robotics, they might not want to buy a $20,000, $30,000, $50,000, $100,000 robot. So that's why we decided to build this affordable robot. We sell it for $400 to $500, so it's very cheap, easy for you to take a decision to buy it, put it next to your laptop, and start experimenting with open-source AI robotics, like tinkering.

I think it's the most important thing. It's one of the most impactful things that you can help with in AI right now because AI has very strong natural tendencies of concentration of capabilities. I think there's a strong probability of a future where only a few companies are able to do AI and the rest of us are doomed just to be AI users, not really AI builders. That's something we should be worried about and try to change. So everything we're building at Hugging Face, we're always thinking, how can it help make more people AI builders, tinkerers themselves, learn, experiment? That's one of the reasons why we're excited by LeRobot Mini and in general by open source.

First, we're super excited about people building apps. We feel that if thousands of AI builders start building with this robot, there's going to be an explosion of capabilities, especially in the AI field where there are new models, new capabilities every day. We hope that ultimately all these capabilities will translate almost instantly to robotics. When the next Gemini model is out, the next GPT model, the next Anthropic model, people should be able to build their robotics apps right away. 'Oh, Gemini 3 is out. Look at the new capabilities of my robot. Now we can read my book, recognize these objects, teach my kids about this new topic.' I'm really excited to see what people are going to build, not just when they receive their robot but also as the field evolves. On the software and platform side, and on the hardware side, because the goal of LeRobot Mini is to be open source, we hope that people will not just assemble their own robots. When they receive it, it's not assembled, so they have to assemble it.

Exactly. For me, it took me five hours to build my LeRobot Mini when I received it. What we expect is people not only to build it themselves, assemble it themselves, but also to improve it themselves. I wouldn't be surprised if we start seeing people putting LeRobot Mini on wheels, or adding grippers or hands, improving the motors. We hope to see really cool developments driven by the community, not so much driven by us, but driven by the community, both on the software side and on the hardware side.

We're really driven by the community, where we feel we can have an impact, unleash some community power to the world. That's actually how we became the GitHub for AI. When we started the company, we were doing a sort of Tamagotchi AI, conversational AI, like a ChatGPT but before ChatGPT. This was like 2016, 2017. Then one day, I remember quite vividly, it was a Friday, one of our co-founders, Thomas Wolf, told us he saw Google releasing this model called BERT, which was the first popular transformer model, but they released it in TensorFlow and most of the community was using PyTorch. He felt it would be useful for the community if we ported BERT from TensorFlow to PyTorch. At the time, Julien and I were like, okay, let's do it to see if the community is interested. So Thomas worked the whole weekend, and on Monday he tweeted about the release of PyTorch pre-trained BERT. We got a thousand likes on Twitter, which was a huge thing for us. We broke the internet. But the truth is, it was useful to the community. Then in the following weeks, people and researchers started to add their models to what we created. The team doing GPT, the very first version that was open source, the people who are now doing Mistral were doing a model called XLNet, and a bunch of other researchers added their models to our repository. People started to use that as a source for their models. That was the early innings of what we are today, the platform for AI builders. So right from the start, we've been driven by the community. Not only driven, but the community made us what we are and propelled our success. If you look at us at the time, the founders were three random French guys without much network or special access. It's really the community that made us who we are and multiplied some of the initiatives we started. We're really grateful, and that's why when we think about anything we're doing, we always think, how can this benefit the community? Because that's what made us who we are.

The Hugging Face platform is 99% free. We try to build a model that fosters that. When you think about doing good in the world, I tend not to trust people too much, but instead trust the systems and the incentives. For us, we always thought about how you can build a company or a system with incentives for us to do good. We built Hugging Face in a way that if we make open source more popular as a platform, we're going to do well as a company and as founders. That aligns the incentives for us to keep doing well. A lot of founders or companies sometimes make the mistake of wanting to do good but building systems where it's not rewarded for them to do good, or it's almost a force to fight, betting against human nature. For example, some companies building APIs for AI that they monetize get into a race where the more money they make with their APIs, the more they can invest in training, and it becomes very difficult not to give in to the motivation to make more money and focus more on closed source, not sharing. For us, it's always been important to create a system with aligned incentives so we continue to double down on our mission: to make AI more open source, more available to anyone, to turn as many people into AI builders as possible, while being successful.

We have a bit of an unusual structure at Hugging Face. We're distributed all over the world, and we have a lot of functions distributed across the company. We have very little talent teams, HR teams, community management teams.

Yeah. But we're also trying to make it everyone's responsibility to hire, communicate on social media, interact with the community. For example, our social accounts don't have a community manager; anyone from the team can tweet from the Hugging Face Twitter account. That creates a distribution of responsibilities that shows everyone it's their responsibility to interact with the community. It's the same for hiring. We don't want HR or talent teams in charge of hiring. We want every team member to think, who are some great people in the world that would love to work with us, and reach out directly and tell them, 'Hey, you should join us.' That's worked really well for us. Today, a lot of companies, especially big technology companies, specialize people and put them in boxes: 'You're going to be a software engineer and write code, you're going to be a marketer and market, you're going to be a PR person and do PR.' They force you to only do that, which is quite boring and doesn't help you grow as a worker or as a human being. We take a different approach. We believe everyone is able to do technical work, communicate, and hire. We try to hire generalists and help them do all of that. It's been really good to us because it brings new perspectives to each line of work. When an engineer is trying to hire themselves instead of a talent team, they come with different ideas, different ways to attract and filter people. The other way around is also true: when someone with less coding experience starts thinking about how to build a product, they come up with different ideas and perspectives. It's been really good to develop everyone at Hugging Face as human beings, to be exposed to more things, to have a variety of experiences in terms of functions, countries, and missions. We really like this approach and want to keep doing it in the future.

I was mentioning the founding story of Thomas, being like, 'Okay, there's something completely different from what we've been doing so far with this BERT release, and I want to port it to PyTorch.' In a normal organization, people might have said, 'What are you talking about? That's not at all what we're doing.' It would have been shut down too quickly, not because people are bad, but because you become so focused on something that you overindex on limiting distractions. Our culture allowed us to just say, 'If you think it's exciting, if you think it's going to be useful for the community, if it's going to be impactful, just go do it, and we'll see later if it works.' And it did work. A lot of these initiatives are made possible thanks to this culture. LeRobot Mini is a good example. A normal company with a normal culture probably wouldn't have done it. But for us, the community is excited about it, we have team members who are excited about the topic, so we just experiment, build it, and see if it works.

Absolutely. There's a certain joy of building. People don't always realize it, but building things yourself the way you want them to be built, being active in the world and co-creating something, brings a lot of joy. It certainly does to me as a founder. Building is one of the greatest joys I experience. When you give people a tool, a platform, something to build, they enjoy it very much. When we think about LeRobot Mini, we think a lot about how not just to enable people to use it—using it is great, but there are many devices you can use—but how we can use it as a way for people to build and experience this joy of building. That's the first thing. It creates joy for people, and then it creates urgency for them to program it the way they want. Some of our diverse backgrounds from all over the world taught us that people don't want the same thing everywhere. We've always felt it would be dangerous if only a few people in Silicon Valley decided what to build, dictating the direction of the future. By making people more builders, you give them agency and a choice in deciding what to build. That's tremendously important as a society for humanity. When you think about the future of AI, people get to decide what AI is going to be used for. Some people will want AI for chatbots, social media video creation, but a lot of people will want to create different things that improve their life, useful for education, for science, and they'll want to build it in their own way. I have LeRobot Mini in my house right now, and I want it to run locally. I want the data collected by LeRobot Mini to stay on my laptop and on the robot because I don't want to send this information to any other company. It's my house, my private space. Giving people the ability to build allows them to decide how this technology is built and creates granularity and diversity. You'll have options where sometimes AI runs locally on your hardware, sometimes you go through an API, sometimes you call LLM providers like OpenAI or Anthropic. But at least when you participate in the building yourself, you can choose and end up with a solution you feel most comfortable with, instead of being locked into some ecosystem.

Well, as proprietary API LLMs emerged, there was the possibility of going that direction. We have a very good science team training good smaller open-source models. We could have said, 'Now we're going to build larger proprietary models,' because it looked like a lot of people were starting to do that. We decided not to because, as I mentioned, it wouldn't create the right incentives for us. If you start to commercialize models, the incentives to do more open source become blurry. There's also been a lot of interest from our community for us to do more compute solutions, our own cloud, which we haven't done a lot yet. We've decided to partner with cloud providers like AWS, Azure, Google Cloud, and inference providers on the Hugging Face platform like Fireworks, Together, and other startups. We did this because we felt there were people doing it well. Whenever we feel there are people doing it well, we don't try to compete just to compete. We focus on the things not handled yet by the community, not reinventing the wheel. On compute, we've mostly partnered with people instead of building our own compute solutions.

It's been super fun. We've seen two phases for AI. The early phase where AI was really niche—when I started working on AI, we weren't even calling it AI; we were talking about computer vision, chatbots, and it was quite niche. The second phase is when AI became much more mainstream, everyone building with AI. It's always been very competitive with a lot of players, and Hugging Face being the central community, the central platform. We've always had a lot of people trying to replicate what we've been doing. If you think about model zoos or model repositories, there have been maybe 50 startups and big companies trying to do that in their products, from Google with TensorFlow Hub, Meta, Amazon, Microsoft, to every smaller startup at some point trying to replicate what Hugging Face has done. The beauty of what we're building is that it's very independent, very neutral. We're not going to push a specific model, dataset, or provider. We always try to provide you with any model you can think of and let you choose. Right now, there are over 6 million models, datasets, and apps on Hugging Face. A new repository is created every 8 seconds. The volume of contributions is insane. We've reached a scale where it doesn't really make sense to go to any other platform or start another platform because after one day, another platform would be so behind in terms of the number of models. In a day, there are a few thousand new models, datasets, and apps on Hugging Face. We've reached this scale for a platform, a bit similarly to what we've seen with GitHub. We have quite strong network effects and a strong moat where it's going to be really hard for someone else to do exactly what we do, which is a nice position to be in.

You're totally right. It's a different kind of business, a different kind of product where you want to create the foundations, make them very stable, foster network effects for a long time, and build trust with the community for a long time. If we were doing what we're doing for just six months, a year, or even two years, people wouldn't trust us as much as they do now. Time builds stronger trust with the community. If they've seen that you've been doing this for a few years and you didn't try to take advantage of them, didn't do a rug pull, didn't go out of business, that's when they start trusting the platform enough to invest more time and resources, share more models, datasets, and applications. It's really important and one of the characteristics of these kinds of products and companies that you can't fake or accelerate. If you're building these kinds of companies as founders, you have to make peace with it. Again, about alignment of incentives, you have to build such a platform in alignment with your values and what you're excited about so that you can stay excited during these years of building and not get distracted by other people's faster success. It's really easy in AI right now to get distracted by another company where you look at the revenue growing so fast from zero to a billion dollars in a year and a half. You can think, 'I'm going to stop everything and do that.' But if you believe in what you're doing, if you enjoy it, if you feel it matters, you can ignore that and not chase those things, which come with their own constraints and challenges. I've met a lot of people, some successful, some not, some in rocket ships, some failing, some with more linear growth. What's always struck me is that their level of happiness and enjoyment is not necessarily correlated to the scenario they're in. Usually, it's more correlated to how aligned what they're doing is with their values, their excitement, and their life mission. For us, because we're so excited about open source, about enabling anyone to become AI builders, if it took a year, 10 years, or 50 years, it doesn't really matter because we're really enjoying what we're building. We feel we're growing, having a positive impact in the world. So, really enjoying the journey no matter the destination.

First, we're really grateful and lucky. We're probably one of the companies in the world that AI builders love the most. You can go talk to any AI builder, and there will be very few who don't love Hugging Face, which is something really important to us and an incredible validation. It all comes down from what I was sharing in the beginning: our feeling that we're kind of random people with no special rights to win or be successful. It's really the trust of the community that made us who we are. So it feels very natural for us to focus on that, keep fueling this trust, try to make the community proud, and be impactful and useful to them. We don't think about it too much. It's more a natural part of how we think about things.

We've been inspired by a lot of products. Of course, GitHub—we talk a lot about GitHub. We feel the impact in the world of GitHub is really underappreciated. If you think of the number of developers, it's 150 million now using GitHub, and the power of open source it unleashed is really amazing. That kind of thing is an inspiration for us. When it comes to me, I'm a big fan of philosophy, so I have a lot of philosopher inspirations.

My favorite book is 'The Myth of Sisyphus' by Albert Camus, a French philosopher. Obviously, I'm French, so I'm biased, but I feel they have a lot of lessons applicable to entrepreneurship. The myth of Sisyphus is about this guy cursed to push a rock up a mountain, and each time it reaches the top, it falls down, and he has to start all over again for eternity. In popular culture, it's been defined as a curse, something really bad. But the conclusion of the book, the last sentence, is 'You have to picture Sisyphus happy.' Despite the hard, repetitive, seemingly meaningless task, Sisyphus is happy. The reason is that his goal and happiness don't come from reaching the top or staying at the top, or from a higher god. The whole philosophy of absurdism from Camus is that you don't need a god or heaven or some higher power because the task itself of pushing yourself, the task itself of pushing the rock—I imagine why: why is pushing? 'Oh, it's funny, look at this mountain, it's beautiful, maybe there's a sunset, maybe a sunrise.' The task itself can be joyful for him. It's the same for entrepreneurs. Of course, you have the motivation of success, of getting there, building a great company, but also the joy of building your company: hiring great team members, experimenting with new projects, releasing things. We were talking about the joy of building like LeRobot Mini. I'm really inspired by that. I think that's one potential recipe not only for happiness but also for relentlessness. The name of the podcast is 'Relentless,' and to be relentless, you can channel some of that. If you really align your happiness with the grind, the hustle, the building, more than the outcome, you'll become much more relentless because it's going to be harder to make you give up. The way I think about Sisyphus is he pushes the rock up the hill, it rolls back down, and every time he pushes it back up, if you do it right, it's a little bit more your rock. You get to decide what kind of rock you're pushing up the hill. Life is just the endless journey of pushing the rock up the hill. If you do it right, you enjoy every day more and more because you get more agency, more freedom, more realization. You align more and more what you're doing, what you're building, with your essence, who you feel you are. So you grow this joy, this happiness. There's one philosopher who did a variation of that, named Clement Rosset, a French philosopher. His theory, which I also find interesting, is that joy comes from the gap between the meaninglessness of the world and the enjoyment you can find in it. He says, for example, you look at a beautiful sunset, a robot you build, or a company you build. Maybe it's vain or meaningless, there's no higher reason or purpose, but for some reason you look at it and feel joy. His theory is that this 'force majeure,' the major force that drives you, is this gap between the meaninglessness or vanity of some things in the world and the joy you can find in them. If you work on something extremely meaningful, you find joy, but your rational mind understands why, so it doesn't drive you as much in terms of force. It's a very interesting theory and philosopher.

It happens a lot in startups. I have many friends, founders, CEOs who for a few years started thinking more about what they should do, what people told them to do to grow, scale, build a successful company. Then one day they wake up and realize, 'Actually, the company I built, I don't really enjoy building it anymore. It's not fulfilling any of my needs, happiness, excitement.' That's usually when they quit, hire an older CEO, and it's usually bad news for the company because it's harder to grow after that. When I talk to founders, my most common advice is to focus on building what you enjoy building, what you're excited to build. Talk to your co-founders and think, 'Are we excited about this direction? Do we feel we can wake up every morning for five years focusing on that?' Or is it more something you feel you have to do because investors or others are telling you to, or because you have a fake image of what a big company is? Instead, focus on what you're excited about, what you want to build, and what you feel is useful for the world. That's how companies become a reflection of the founders building them.

All the time. One of the things I try to do is every week do at least one experiment that, if I thought too much about it, I wouldn't do. I let my excitement for something drive me to do something I wouldn't do otherwise. Most of them, I walk back from. You pointed something really important: follow your excitement and your gut more than your rational mind. It's something I've been working on a lot. For example, I always resisted the urge to do to-do lists or have an assistant.

Because I suspect that a lot of people, including me, when they start rationalizing things too much, putting things into to-do lists, and trying to do everything they're supposed to do or should do, they end up not spending as much time on the things they're excited about, that their gut tells them to do. There's some magic and beauty in people's minds: if something is important and exciting enough for you to do, you'll remember to do it. You don't need a to-do list. If you need something to be on a to-do list, it probably means you shouldn't do it or not do it now. It's the same with assistants. A lot of people get assistants to help schedule meetings, manage stuff they don't want to do. In my opinion, if you don't have time to do something, you probably just shouldn't do it. It means it's not exciting or impactful enough for you to do. There's some magic in doing things yourself. It makes you so much more efficient. For example, how we scheduled this interview: if I had an assistant, they might have taken a few days to coordinate. Instead, I got excited, thought it was a good idea, said, 'Can you come that day? Yes, let's do it at my house,' and things became easy and simple. That's one of my interesting learnings. Sometimes by avoiding complexity, you can actually do more and keep your life more aligned with what you're excited about, instead of pushing too far into what you should do.

I always try to think about how I can add value and impact, and do something ten times better than anyone else could, and focus on those things. When there's something where I feel I can't do ten times better than someone else, I prefer to let them do it if they're excited. So that's how I pick and choose what to work on. As a CEO, I ask what are the most impactful things I can do that nobody else can. Also, something really important in our culture at Hugging Face is the decentralized culture and the freedom culture that empowers us to be driven not only by my own excitement but by anyone else's excitement. A lot of times, when we release new things, people try to understand the strategic relevancy. But we follow people's excitement. Sometimes people ask how it makes sense strategically, and I reply, 'John got excited about it and built it.' That's the logic behind it. I always try to foster that. The way we hire at Hugging Face is a bit unusual. We usually don't hire for a specific job description; we try to find people who are smart, excited, and aligned with our vision and mission. We bring them in and tell them to work on what they're excited about. That's worked really well because they do things that are meaningful to them, they become extremely motivated and owners of their projects, and have a tremendous impact. If we worked the other way around, starting from the task and trying to find someone to fit it, I think that wouldn't work as well. Giving people freedom, ownership, and structure to work on what they're excited about usually leads to good things.

Even the acquisition started from the excitement of the team. We acquired the Poland team, who were excited about joining Hugging Face and building with us. After we acquired them, they told us they thought it would be exciting to work on for AI builders. So I just said yes. A big part of my job as CEO is just to say yes, empower people, support them, point them in directions that increase their impact, remove roadblocks, and reassure them on their ability to release earlier than they think they should. The natural tendency is to delay launch, but at Hugging Face we always say you don't need to be ready to launch; you just need a minimum valuable product that won't completely break, and set the right expectations. Tell people it's an imperfect first version and let them play with it. So a big part of my job is to reassure and build confidence so team members can release things quickly and iterate with the community.

A lot of it is making them comfortable with mistakes. The reason most people delay things is fear of repercussions. Releasing something and making a mistake feels uncomfortable, like people are laughing at you. When you release early, that happens all the time, so it's about making them comfortable with that. For example, Thomas released PyTorch pre-trained BERT—he started on Friday and released on Monday. For two days it wasn't working at all. Most people would have been stressed, but we have a high tolerance for that. Thomas' best reaction was, 'Oh, it's not working because of some problems, but people seem excited. I'll fix these problems, be transparent, and two days later it was fixed.' Same for Rishi Mini when we ship bigger batches. We hope to ship around 4,000 before the end of the year, and I'm sure there will be problems. I apologize in advance to beta testers, but we'll work hard with the community to fix them quickly. That's the way to approach things. The biggest thing that helps people not delay is making them more comfortable with mistakes, failures, and problems.

Expectations are a big problem for AI specifically because of the term itself. It's easy to overhype and oversell what you're building. Some companies fall into that trap, which creates fears, making people think of sci-fi scenarios like Terminator. But at Hugging Face, we see AI as a new generation of software. Andrej Karpathy talked about Software 2.0, which is a more down-to-earth way to approach AI. It's not a black box that a few people will dominate; it's a new paradigm for building that everyone can use, solving different challenges and creating new opportunities. On Hugging Face, we have 11 million AI builders now. The fastest growing domains are audio, video, image, biology, chemistry. AI is being applied to every domain, just like software. When you present AI like that, it's closer to reality, removes fears, and encourages people to become builders.

In AI, there are strong natural tendencies toward concentration of power, capabilities, and wealth—compute, talent, money. We should be scared of a world where only a few companies can do AI. That would be like only a few companies could build software, which would be scary and dystopian. At Hugging Face, we fight this by promoting open source—sharing models and datasets openly for free. When companies do that, they give everyone the foundations to build AI products and fight concentration of power. Thanks to open source, we can create a world where not just a few organizations dominate, but tens of thousands of startups, nonprofits, governments, and individuals can build with AI. You can't build AI competitively from scratch; you need open-source models as a base.

With smaller customized models, there are advantages: easier to iterate and experiment, faster and cheaper to run, and more focused. You need a very large generalist model for something like ChatGPT, but for a banking customer support chatbot, you don't need it to tell you the meaning of life—just answer account questions. So you can use a smaller, cheaper, more specialized model that is better for specific tasks. We envision a world with millions of smaller, specialized models solving individual problems, much like software: every organization writes its own code. The same will happen with AI—every organization will have its own customized models based on open source.

We've seen the beauty of transparency in AI. People are scared by black boxes. When you can see how a system works, it removes fears or directs them to rational ones. AI has risks and challenges, but transparency lets you focus on real problems like biases. For example, chatbots have biases—remember Gemini had biases about founding fathers. These are real problems, not sci-fi. Transparency is crucial for solving challenges and creating opportunities. That's a big value of open source for AI.

He would be an AI builder. There's a difference between an AI user who uses APIs and an AI builder who trains and owns models. If I were to start a new company now, I'd focus on AI for biology and chemistry. We're seeing new models on Hugging Face for cell predictions, molecule predictions—these are underinvested. We'll see many new applications in the next few years, especially with progress on time series AI. Chatbots are fun, but too many people work on them. More attention should go to other domains.

Sam Altman talked about real trends vs fake trends. A real trend is like the iPhone 1—imperfect but early adopters use it constantly. A fake trend is like VR/AR where usage drops off. With AI, GPUs are all humming, so probably not a bubble. At Hugging Face, we see the gap between perception and usage. Pre-ChatGPT, AI was underhyped—used in search, social media, self-driving cars, but not talked about much. Then ChatGPT came out and perception caught up to usage. It felt like a bubble to some, but it was a catch-up. Many startups will fail, but that's part of creative destruction. More startups may fail in AI than in software, but that's fine—it means more excitement and novelty. AI is a new paradigm that throws playbooks out the window. We had a cycle of software maturing with playbooks like the SaaS playbook. Now in AI, people try to apply old playbooks, but the better approach is to unlearn and start fresh, thinking how to build a startup in the AI era.

Probably when other people released similar things to us, like Google releasing TensorFlow Hub, a direct competitor. When a massive company with massive resources releases something similar to what you're doing, it's intense stress and fear. We were good at not panicking. We took our time because it's easy to damage culture and reputation. Our approach was to collaborate rather than compete. We reached out to them, talked to a prominent person at Google, and started building integrations between TensorFlow Hub and Hugging Face Hub. That served us well. We found a way for Google's initiatives to be useful, while also driving users to use both products. It's counterintuitive, but collaborating with competitors worked for us.

For tech, France has made progress in recent years, getting closer to US best practices. More American VCs invest in French startups, founders move to the US for YC and sometimes return. France has an advantage in producing great mathematicians and engineers—my co-founders went to Polytechnique. The education system is math-based, producing good AI engineers. Many AI companies have French people, like Yann LeCun. France has advantages, though the political situation is complicated. If they can manage that, they can impact the field. Historically, there have been good open-source initiatives, like Mistral pushing open-source models. There are opportunities, and I hope France contributes to the field. It's important to decentralize AI capabilities internationally—not just one country should build AI. Every country should be able to become AI builders.

In the early days of AI, 2016 to 2022, the US was extremely open—transformers, attention is all you need were open source. That open science allowed the ecosystem to thrive. If Google didn't open source transformers, OpenAI couldn't have built ChatGPT. Around 2022, for some reason, people became scared of open source and started pulling up the ladder. In China, they became more open, contributing more open science and open source. China has always been focused on open source as a way to contribute from outside. Their incentives differ, allowing more sharing. It intensified as they saw the potential. Two years ago, Chinese organizations shared models on Hugging Face with little visibility, but now when a good model is released, the impact is massive—DeepSeek has over 100,000 followers. They've understood that open source gives massive impact. There are now 30 to 50 good organizations in China sharing models for text, audio, video, image. I hope the US returns to its open-source philosophy. We're seeing progress: Elon Musk's xAI open-sourced Grok 3 on Hugging Face, OpenAI released GPT-OSS, AllenAI releases open-source models and datasets, Nvidia releases over one model per day. The US administration supports open source with an executive order. Hopefully, we can return to US open-source AI. If not, it's scary that American startups might build on Chinese foundations, which is already happening. For example, biases are embedded in those foundations, and control shifts to China. China is starting to dominate open-source AI, and soon they may dominate AI in general, accelerating progress. The US should invest more in open-source AI to change the trend.

It's early stages for robotics AI with encouraging trends. It's easier to build good, affordable hardware, and easier to have great software and AI for it. This confluence enables a new paradigm. A challenge is that initiatives are siloed—each lab creates its own stack, slowing progress. At Hugging Face, we push for open source and standards. Our library, LeRobot, has become popular. Such initiatives could unlock faster progress if everyone collaborates. That could lead to a ChatGPT moment for robotics, which came from collaboration.

There are always hard things for founders. The things that impact me most are people-related. When team members who have worked with us for years decide to leave, it's tough. I've learned to accept it, especially because often they go on to build cool stuff and stay connected. But for me personally, that's the hardest part—when people leave or when we have to let go.