Hugging Face's Clem Delangue on Open Source AI and the LLM Bubble | MTS Live

The idea of restricting a technology like AI based on risks is just like, for example, you would say, okay, some people can punch other people, so let's tie down everybody's hands, right? Because it's too dangerous, some people can punch. But in reality, you don't want to do that because your hands are so useful. The way you want to control it is untie everyone and then regulate or fight the bad actors. So for example, if hacking that creates cybersecurity risks, it's illegal, right? So you have to fight it, but not by preventing everyone from getting these capabilities. Otherwise, you slow down progress, you create massive gaps in terms of controls, in terms of capabilities, and you create actually additional risks.

Yeah, of course. Thanks for having me.

Yeah. So I mean, historically the US was super, super strong with open source, right? That's kind of what led to the current AI revolution, right? Like the T in ChatGPT is actually coming from Transformer, which was open source from Google. Unfortunately, for the past few years, this trend has changed and things tended to kind of close down in the US, with frontier labs more kind of sharing their models behind closed-source APIs. In China, it's the complete opposite movement. They're the strongest open-source contributors today. If you ask most startups, most academia in the US that are using open source, they're usually using Chinese open-source models, right? You've probably heard of DeepSeek, of Qwen, of Kimi. There are a bunch of companies and organizations in China contributing massively to the field of open source.

Well, I was asked if we were in an AI bubble, and I said we're probably not in an AI as a general field bubble. But I feel like if there's one specific domain of AI where there's so much investment that there's maybe a risk of overinvesting, it's large language models distributed behind APIs, right? You see the building of crazy data centers for it, and obviously you see a lot of revenue growth, but with uncertain margins, uncertain long-term sustainability and moat for it. So if there's a bubble, it's probably an LLM. But we'll see what happens in the next few months.

So the interesting thing is that we've had these conversations and this talking point for a while in AI. When we were at Hugging Face, I think six, seven years ago, at the time it was GPT-2, and there was already a lot of people saying that it was too dangerous to release in open source. At the time, right, it was six, seven years ago, when basically it was nothing more than just an autocomplete. I think we've seen progressively that these were quite overblown, and I think they're also overblown today. And proof point is that, you know, Mythos, I think when it was announced, was it like three weeks ago, a month ago, it was crazy dangerous, and now it's starting to be deployed kind of everywhere. I think they just got access to the first international organization, it's in South Korea, I think yesterday or something like that, and probably in a few weeks or in a few months everyone is going to be using Mythos and not kind of destroy the world as a result. So I think with the current models, it's safe to release behind APIs, it's safe to release in open source, and it's actually the safest way because it gives everyone kind of the capabilities to not only build the systems but also build the protection systems. So if we talk, for example, for cybersecurity, the biggest risk is that a few players have capabilities that other people don't have, right? And so the attackers could have capabilities that the defenders don't have, whereas if you make it more open, actually it's usually easier for the defenders to react and kind of make the whole system safer. So that's kind of what we see with each release, where there are always overblown concerns before, and then progressively we all adapt and the benefits kind of outweigh the risks.

Got it. It's kind of like the story of technology, you know. The idea of restricting a technology like AI based on risks is just like, for example, you would say, okay, some people can punch other people, so let's tie down everybody's hands, right? Because it's too dangerous, some people can punch. But in reality, you don't want to do that because your hands are so useful. They're creating so many good things in the world. You need your hands. The way you want to control it is untie everyone, give the freedom to everyone, and then regulate or fight the bad actors, right? So, for example, if hacking that creates cybersecurity risks, I mean it's illegal, right? So you have to make it illegal, you have to fight it, but not by preventing everyone from getting these capabilities, because otherwise you slow down progress, you create massive gaps in terms of controls, in terms of capabilities, and you create actually additional risks.

Yeah, I mean, I'm excited to see conversations about open source AI. Probably there's going to be some conversations about distillation, about collaborations between two countries. I hope both countries will be able to agree on fostering more transparency, more openness to kind of help more people access this technology. I'm glad that Jensen hopped into the plane and joined these conversations because I think he has a lot of the right perspectives on this topic to kind of create more collaboration between countries and kind of shared progress.

Yes, I have two little robots behind me, the Reachy Mini. We've shipped almost 10,000 of them all over the world. So it's probably one of the most widely distributed robots of the year at this point. I think what's really cool with robotics is that it enables very new use cases and better use cases for AI. So for example, for the Reachy Mini, you have an app store. Anyone can build apps. So there's been over 300 apps that have been created for it already. And when you see it in action, for example with kids, empowering kids to interact with AI in a different way than looking at a laptop or looking at a phone, you realize that it's very empowering when you see the Reachy Mini on a kitchen table looking around and helping you cook. You realize that it's enabling, empowering, creating new use cases that are just not possible just with a laptop and a phone, right? That's why OpenAI and Sam Altman, for example, have talked a lot about their excitement about bringing new devices to market. There's an important China-US component there because it's very likely that Chinese are going to dominate robotics, or at least they're already dominating. And so on this topic too, it's really important that we build more in the US on this topic. And we obviously have a lot of strength for it with the strength of the startup ecosystem in the US, the strength of the frontier models. I hope to see a lot more in the coming months on the topic.

Yeah. I mean, I don't blame them. They have a lot on their plate, right? I think with the coding assistant, they've kind of been dealing with their own set of issues. The reality is that hosting and sharing AI artifacts is quite different than hosting code. So even if people have been calling us the GitHub of AI, I think it's two very different things. For example, for us, the volume of files of data that we're dealing with is much, much larger than what GitHub is doing. For example, just last week, we added two petabytes of data to the platform just last week. It's kind of a matter of comparison. It's the equivalent of 500,000 two-hour movies that have been uploaded to Hugging Face just last week. So you have a lot of structural differences. And we managed to build our infrastructure capabilities in a way that makes it just better for people that are building in AI to use Hugging Face to host their models, their datasets, both publicly but also privately. We have a lot of private usage now. So that's kind of some of the reasons why we managed to do it, whereas GitHub focused on other things.