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

Summary

Clem Delangue, co-founder and CEO of Hugging Face, discusses his vision for democratizing AI through open source. He argues that a world where only a few companies can do AI would be “a very scary dystopian world” and positions Hugging Face as a platform to enable tens of thousands of startups, nonprofits, and governments to build with AI. The company recently launched Richie Mini, an open-source desktop robot that can be 3D printed and fully programmed by AI builders - already with over 5,000 pre-orders.

The conversation traces Hugging Face’s origin story from a Tamagotchi-style conversational AI in 2016-2017 to becoming the “GitHub for AI.” The pivotal moment came when co-founder Thomas Wolf spent a weekend porting Google’s BERT model from TensorFlow to PyTorch, getting 1,000 likes on Twitter - which felt like “breaking the internet” for the small French team without a network. Other researchers then started adding their models, and the platform emerged organically from community need.

Delangue emphasizes Hugging Face’s unusual organizational structure: extremely flat with distributed responsibilities where anyone can tweet from the company account, everyone is expected to hire rather than relying on HR/talent teams, and there’s a belief that everyone can do technical work, communication, and recruiting. He discusses the company’s aligned incentive model where making open source more popular directly benefits them, contrasting with API-based companies whose monetization incentives push them toward closed source.

Highlights

”A world where only a few companies can do AI would be a very scary dystopian world”

yt-dlp --download-sections "*0:00-1:00" "https://www.youtube.com/watch?v=b0iJZS9HgJA" --force-keyframes-at-cuts --merge-output-format mp4 -o "b0iJZS9HgJA-0m00s.mp4"

“Something that we should be scared about is to end up in a world where only a few companies are able to do AI. I think that would be a very scary dystopian world. Open source is a way to fight these natural tendencies. We believe that tens of thousands, hundreds of thousands of startups, of nonprofits, of governments should be able to build with AI.” — Clem Delangue, 0:00

”We were three random French guys without much network”

yt-dlp --download-sections "*10:35-11:40" "https://www.youtube.com/watch?v=b0iJZS9HgJA" --force-keyframes-at-cuts --merge-output-format mp4 -o "b0iJZS9HgJA-10m35s.mp4"

“If you think of it, if you look at us at the time, the founders, we were three random French guys without much network, without much special access to any kind of network or any kind of organization. It’s really the community that made us who we are.” — Clem Delangue, 10:35

”We got a thousand likes on Twitter and thought we broke the internet”

yt-dlp --download-sections "*9:20-10:20" "https://www.youtube.com/watch?v=b0iJZS9HgJA" --force-keyframes-at-cuts --merge-output-format mp4 -o "b0iJZS9HgJA-9m20s.mp4"

“Thomas worked the whole weekend and on Monday he tweeted about the release of PyTorch pre-trained BERT. And we got like a thousand likes on Twitter at the time which to us must have been a huge thing - for us it was like we broke the internet, what happened? Why are so many people liking this?” — Clem Delangue, 9:20

”I try not to trust people, but instead trust systems and incentives”

yt-dlp --download-sections "*12:00-13:05" "https://www.youtube.com/watch?v=b0iJZS9HgJA" --force-keyframes-at-cuts --merge-output-format mp4 -o "b0iJZS9HgJA-12m00s.mp4"

“When you think about doing good in the world, I tend to try not to trust too much people, but instead trust the systems and the incentives. We built Hugging Face in a way that if we’re making open source more popular, we’re going to do well as a company. It aligns the incentives.” — Clem Delangue, 12:00

”Anyone from the team can tweet from the Hugging Face Twitter account”

yt-dlp --download-sections "*15:05-16:10" "https://www.youtube.com/watch?v=b0iJZS9HgJA" --force-keyframes-at-cuts --merge-output-format mp4 -o "b0iJZS9HgJA-15m05s.mp4"

“Our social accounts - we don’t have someone who’s responsible for them as a community manager. Anyone from the team can tweet from the Hugging Face Twitter account. And that creates kind of like a distribution of responsibilities that shows everyone that it’s their responsibilities to interact with the community.” — Clem Delangue, 15:05

”We believe everyone can do technical work, communicate, and hire”

yt-dlp --download-sections "*16:42-17:45" "https://www.youtube.com/watch?v=b0iJZS9HgJA" --force-keyframes-at-cuts --merge-output-format mp4 -o "b0iJZS9HgJA-16m42s.mp4"

“For us we’re taking a little bit of a different approach where we believe everyone is able to do technical work, everyone is able to communicate, everyone is able to hire. We rather try to hire generalists and help them do all of that.” — Clem Delangue, 16:42

Key Points

• Dystopian world of few AI companies (0:00) - A world where only a few companies can do AI would be scary; open source fights these natural monopolistic tendencies
• Richie Mini robot launch (1:05) - Open-source desktop robot for AI builders; 5,000+ pre-orders; can be 3D printed; took Clem 5 hours to assemble his
• Robot apps explosion (5:10) - Expects community to build apps like hide and seek, red light green light with kids; new AI models translate instantly to robot capabilities
• Hardware modifications expected (6:40) - Anticipates people putting Richie Mini on wheels, adding grippers, improving motors
• Origin as Tamagotchi AI (7:55) - Started as conversational AI like ChatGPT before ChatGPT, back in 2016-2017
• BERT to PyTorch moment (8:25) - Thomas Wolf ported Google’s BERT model over a weekend; 1,000 Twitter likes felt like breaking the internet
• Community-driven success (9:45) - Researchers started adding models; early GPT team, Mistral founders (then at XLNet) contributed; community made Hugging Face
• Three random French guys (10:55) - Founders had no network or special access; community multiplied their initiatives
• Platform 99% free (11:40) - Built business model that fosters doing good; trust systems and incentives rather than people
• Aligned incentives (12:25) - If open source becomes more popular, Hugging Face does well as a company; avoids API monetization trap
• Flat distributed structure (14:35) - Very little HR/talent teams; everyone can tweet from company account; everyone responsible for hiring
• Generalists over specialists (16:45) - Believes everyone can do technical work, communicate, and hire; specialization is boring and limiting
• Engineers hiring themselves (17:30) - Engineers directly reach out to people they want to work with rather than relying on talent teams

Mentions

Companies

• Hugging Face (0:45) - AI platform company; “GitHub for AI”
• Google (8:35) - Released BERT model in TensorFlow
• OpenAI (5:35) - Referenced for GPT models; early GPT team added models to Hugging Face
• Anthropic (5:45) - Referenced for Entropic models
• Mistral (9:55) - Current company of people who contributed XLNet to early Hugging Face
• Toyota (17:05) - Referenced for big tech companies putting people in boxes

Products & Technologies

• Richie Mini (1:05) - Hugging Face’s open-source desktop robot; can be 3D printed
• BERT (8:35) - Google’s first popular transformer model; ported to PyTorch by Thomas Wolf
• PyTorch (8:45) - ML framework that community was using; BERT was ported from TensorFlow to PyTorch
• TensorFlow (8:40) - Google’s framework that BERT was originally released in
• GPT (9:50) - First open-source version was added to Hugging Face early
• XLNet (10:00) - Model created by people now at Mistral; added to Hugging Face
• Gemini 3 (5:55) - Referenced as example of new model that should translate to robot capabilities
• Transformers library (8:55) - What emerged from BERT PyTorch port; researchers added their models to it

People

• Clem Delangue (0:45) - Co-founder and CEO of Hugging Face
• Thomas Wolf (8:25) - Hugging Face co-founder who ported BERT to PyTorch over a weekend
• Julien (8:30) - Hugging Face co-founder
• Ty Morse (0:42) - Host of Relentless podcast

Surprising Quotes

“So I think something that we should be scared about is to end up in a world where only a few companies are able to do AI. I think that would be a very scary dystopian world.” — 0:00

“We got like a thousand likes on Twitter at the time which to us must have been a huge thing - for us it was like we broke the internet, what happened? Why are so many people liking this?” — 9:20

“If you think of it, if you look at us at the time, the founders, we were three random French guys without much network, without much special access to any kind of network or organization.” — 10:55

“When you think about doing good in the world, I tend to try not to trust too much people, but instead trust the systems and the incentives.” — 12:05

“I think a lot of founders or companies sometimes do this mistake where they probably want to do good, but they end up building systems where it’s not really rewarded for them to do good.” — 12:50

Transcript

0:00 So I think something that we should be scared about is to end up in a world where only a few companies are able to do AI. I think that would be a very scary dystopian world. Open source which is basically the ability to share models to share data sets openly is a way to fight these natural tendencies. We believe that you can create a world where not just a few organizations are able to build and dominate but really any organization. Tens of thousands, hundreds of thousands of little tech, of startups, of nonprofits, of governments should be able to build with AI.

0:42 Today I have the pleasure of sitting down with Clem Delangue. And he is the co-founder and CEO of Hugging Face, which is basically like AI GitHub pretty much. Let’s start off with what is that little robot in the center there?

1:05 Yeah, this is a Richie 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. Are you already shipping them? Yes, we’re starting to ship them. I received mine actually last week.

1:32 It’s fully open source. Most of it is - you can 3D print and 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 that you build at home for you. For example, a lot of people are using them with their kids playing hide and seek with their kids, red light green light with their kids. And also you can share them with the community.

2:00 So we hope that the AI builders all over the world are going to build apps for these with the latest AI models with the latest open-source models and share them with the world.

2:15 How did you kind of make the decision to enter something that seems so unrelated from the hugging face platform to a robot that seems like a relatively big jump. I mean, a big strategic jump as a company.

2:32 Yeah. So it’s uh it’s something that we’ve been kind of like thinking for quite some time. You know, because at the end, our mission is to democratize AI, make AI more accessible to people, especially machine learning specifically. And one thing that’s always kind of like concerned us is that it’s too difficult to tinker with AI.

2:55 You know, I I grew up uh in the 90s, early 2000s, and at the time there was this kind of like movement of hobby makers, tinkerers. They would build their own computers and experiment with how computers work and all of that. And that’s kind of like what we’re missing with uh with AI today.

3:18 It’s very difficult for anyone today to tinker with AI and really learn and develop hands-on skills. It’s very easy to consume AI, right? You go to Claude, you go to ChatGPT, you go to Google Gemini, and you have a nice chat. But that doesn’t mean you actually really know how it works or have developed skills for the age of AI.

3:45 And so uh we feel like today only a very small fraction of people are actually AI builders, people who actually build, tinker, create with AI. And that’s something we should be worried about and that we should try to change. So everything that we’re building at Hugging Face, we’re always thinking, okay, how can it help make more people AI builders, tinkerers themselves, learn, experiment?

4:15 And so that’s one of the reasons why we’re excited by Richie Mini and in general by open source of course.

4:25 How do you think something like that is going to evolve? Are you gonna basically build the first version, get it into a bunch of people’s hands? You said 5,000 or so people have already bought it. What are you looking to get back even as far as people just using it and working out the kinks?

4:48 So, first we’re super excited about people building apps, right? We feel like if thousands of AI builders are starting to build with this robot, there’s going to be kind of like an explosion of capabilities, right?

5:10 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. Right? I was mentioning before next time the next Gemini model is out, the next GPT model is out, the next Anthropic model is out, people should be able to build their robotics apps right away, right?

5:45 Oh, Gemini 3 is out. Look at the new capabilities of my robot. Now we can read my book. Now we can recognize these objects. Now we can teach my kids about this new topic that it wasn’t able to teach before. So 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.

6:12 So that’s on the software side, on the platform side. And then on the hardware side, because the goal of Richie Mini is to be open source, we hope that people are not just going to assemble their own robots. So when they’re going to receive it, it’s not going to be assembled. So they have to assemble it.

6:35 A little bit like the IKEA Lego set sort of situation. Exactly. For me, it took me five hours to actually build my Richie Mini when I received it. And what we are expecting is people not only to build it themselves, assemble it themselves, but also to improve it themselves, right?

7:00 So, I wouldn’t be surprised if we start seeing when we’re shipping it people who are putting Richie Mini on wheels, right? Or people who are adding grippers or hands to Richie Mini. People who are improving the motors, right?

7:20 So yeah, 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.

7:35 Hugging Face did not start off as kind of like this AI GitHub platform. And it’s - I certainly would not have expected a year and a half ago that you guys were building robots or trying to ship a bunch of robots. What kind of guides you when you’re making decisions on where to take the company?

7:58 We’re really kind of like driven by the community, right? Where we feel like we can have an impact, unleash some community power to the world. And that’s actually how we became kind of like the GitHub for AI. When we started the company, we were doing some sort of a Tamagotchi AI kind of like a conversational AI kind of like a ChatGPT but before ChatGPT.

8:25 This was like 2016, 2017-ish? Exactly yeah it was quite a long time ago now. And then one day I remember quite vividly it was a Friday, one of our co-founders Thomas Wolf told Julien and us, the other co-founders - “I’ve seen kind of like Google releasing this model called BERT which was kind of like the first popular transformer model but they released it in TensorFlow.”

8:55 And most of the community right now is using PyTorch, right? So I feel like it would be useful for the community if we ported BERT from TensorFlow to PyTorch, right? And at the time Julien and I were like okay yeah let’s do it to see if the community is interested, if it’s useful to the community.

9:20 And so Thomas worked the whole weekend and on Monday he tweeted about the release of PyTorch pre-trained BERT, right, which was like the BERT port in PyTorch. And we got like a thousand likes on Twitter at the time which to us must have been a huge thing - for us it was like we broke the internet, what happened? Why are so many people liking this?

9:48 But the truth is it was useful to the community, right? And then after that the following few weeks, people and researchers started to add their models to what we created. Right? At the time it was the team doing GPT, the very first version of GPT that was open source. It was the people who are now doing Mistral that were doing at the time a model called XLNet.

10:15 And a bunch of other researchers started to add their models to our repository and then people started to kind of like use that as a source for their models. So that was kind of like the early innings of what we are today like the platform for AI builders.

10:35 So right from the start, we’ve really been driven by the community, and not only driven, but really that’s the community that made us what we are. And really propelled our success because if you think of it, if you look at us at the time, the founders, we were three random French guys without much network, without much special access to any kind of network or any kind of organization.

11:05 And it’s really the community that made us who we are and multiplied a little bit some of the initiatives that we started. So we’re really really grateful and that’s why when we’re thinking about anything we’re doing we’re always thinking okay how can this benefit the community because that’s really what - who made us who we are.

11:30 What have been the biggest points where you forgo kind of making money or like early monetization opportunities in order to kind of build a culture of just helping builders?

11:45 Yeah. Well, I mean the platform - the Hugging Face platform is 99% free, right? And we try to kind of like build a model that fosters that, right?

12:00 When you think about kind of like doing good in the world, I tend to try not to trust too much people, but instead trust the systems and the incentives. So for us, one thing we always thought about is how you can build a company or a system with incentives for us to do good.

12:28 Right? We built Hugging Face in a way that if we’re making open source more popular as a platform for open source, we’re going to do well as a company, as founders. And so I think it creates and it aligns the incentives for us to keep doing well, right?

12:55 I think a lot of founders or companies sometimes do this mistake where they probably want to do good, right? But they end up building systems where it’s not really rewarded for them to do good or sometimes it’s almost kind of like a force to fight.

13:12 It’s basically betting against human nature almost. Yeah. Yeah. Like for example, some of these companies when they’re building APIs for AI that they monetize, right? And they get kind of like into this race where the more money they make with their APIs, right, the more money they can invest into training and you get into this race that I think at the end kind of like makes it very difficult for you to not kind of give in into this motivation to make more and more money and to focus more and more on API, on closed source, not sharing and all of that.

14:00 And so I think for us it’s always been important to try to create a system with aligned incentives so that we continue to really double down on our mission which is to make AI more open source, more available to anyone, to turn as many people into AI builders as possible while kind of like being successful.

14:30 Did you do anything in the early days that was kind of unusual to find employees that were very culturally aligned and like mission aligned?

14:42 Yeah, we have a bit of an unusual structure at Hugging Face because we obviously kind of like distributed all over the world. But also we have a lot of functions that are distributed all over the company. So we have very little kind of like talent teams, HR teams, community management teams.

15:05 Are you trying to maintain like a pretty flat structure? Yeah. But also we’re trying to make it so that it’s everyone’s responsibilities to hire, everyone’s responsibilities to communicate on social media, to interact with the community.

15:25 So for example, our social accounts - we don’t have someone who’s responsible for them as a community manager. Anyone from the team can tweet from the Hugging Face Twitter account. And that creates kind of like a distribution of responsibilities that shows everyone that it’s their responsibilities to interact with the community.

15:50 And it’s the same for hiring, right? We don’t want to have kind of like HR or talent teams in charge of hiring. We want every single team member to think okay who are some great people in the world that I would love to work with and then reach out to them directly and tell them hey dude you should join us, you should come to Hugging Face.

16:15 And that’s something that’s worked really really well for us. I think today a lot of companies, especially big technology companies, they really specialize people in a way, right? And they put people in boxes. They’re like, “Okay, you’re going to be a software engineer and you’re going to write code. You’re going to be a marketer and you’re going to market. You’re going to be like a PR person and you’re going to do PR.”

16:42 It’s a little bit like they’re looking through a specific lens on just their thing and not like a broader view of the company. Yeah. And they kind of like forcing you to only do that, which is in my opinion quite boring and doesn’t really help you to grow not only as a worker but as a human being.

17:05 So for us we’re taking a little bit of a different approach where we believe everyone is able to do technical work, everyone is able to communicate, everyone is able to hire, right? And we rather try to kind of like hire generalists and kind of like help them do all of that.

17:30 And I think it’s been really good to us because it also brings kind of like new perspectives to each kind of like line of work. When you have kind of like an engineer who is trying to hire themselves instead of kind of like a talent team, they come with different ideas, right?