Thomas Wolf - Cofounder at Hugging Face⁩ on LLMs, robotics and the future of technology

0:00Thomas Wolf:Awesome. We are back here at Vips and Pretzels with Thomas from Hugging Face. Thomas, for those of us out there who don't know you, can you give us a quick intro, who you are, what you do? Of course, I'm a co-founder and chief science officer at Hugging Face. So Hugging Face is a platform for open source models, data sets, and spaces, which are kind of AI apps. So it's a place where you fill the community, you will find a community who is building AI. Awesome. And maybe again for those who are not so deep in these topics, what exactly does open source mean versus closed source and how does that apply to AI models?

0:34Yeah. So you have two ways to use AI models nowadays. The first way is to go to chat GPT. So you use what is called a closed source model, which means you cannot really download the model on your computer. For instance, you have to run it through the API. And the other way is called open source model or open weights more specifically. And the main difference is like a software you can download it, you can modify it, you can tweak it, you can control it, you have much more control on the AI model itself.

1:01Thomas Wolf:And then you can like put it into your own apps or do more things with it. Yeah, for instance if you want to fine-tune it because you think it needs to know a lot about legal and such if it doesn't know enough you can take an operative model, a legal data set like your contract and you can fine-tune the model which means you will train the model a bit more so it knows more about your own domain. And what is Can you maybe also explain how do open source and closed source models differ? I understand the flexibility, but in terms of speed, monetization, what are other things that are pros and cons?

1:32It's quite different. So when you use open source, you leave on the table a little bit of the easy use, I would say, with closed source. So closed source, you just pay basically a subscription, the thing works out of the box. Open source or open weights, you need to download them, you need to find GPU to run them, which can be more complex. But on the other hand, you have much more control. So let's say you want to run them super fast because you have a type of application with like instant answer and not seeing the, you know, like the thing writing. You can run them on some of these novel chips that exist like Grok, Cerebrass, and then you get like extremely fast, 5 ,000 tokens per second output.

2:09So for many novel application or exploration, you're actually better with open source model. ANGELA GALPINIRAJANI - Another example, there was this summer a very interesting release called Genie 3 from Google, which is kind of a picture where this model generates a picture, and you can walk in the picture and interact with the picture. For this, you will use open source model, because you will take a video model, you will tune it, and you will deploy it in your application. So all of the things that are a bit exploratory, but you have the knowledge to implement it yourself, usually people choose to go with open source.

2:41If you want the out-of-the-box simple to use solution, usually go with process.

2:45Thomas Wolf:And open source is just usually more specific. You can go into, you know, give more context to the LLM. And I would say the other thing is you can also save a lot on the price because you can control on which GPU you run it, so you can run it on much cheaper GPU. But yeah, as I said, this comes with the fact that you need more knowledge. You need to learn. So that's also why we have a big part of our website. It's about education on how to use that learning. And you recently acquired a robotics company. you've got the reaching mini coming out i think the order is rolling out now i think yeah can you touch a bit more on why that acquisition and and how you're leaning into robotics a bit yeah of course so i mean one of my job at a gift is to is to keep an eye on the new trends in robotics we're eight years old now there have been like already many waves in ai i would say in this eight years old the lla and the agents the all of this and and i think the next wave that's going to come the next place where ai will kind of solve something is robotics it's not i mean it's happening maybe the end of this year mostly next year I think and so that's why we decided first 18 months ago to start a software library called Le Robot and what we discovered when we open source this software was like there was a lot of interest for that like the community around this the open source robotics community grew exponential and so this year we decided to double down and not only do software but also do some hardware and I would say the specificity of the hardware we do is two ways it's extremely cheap very accessible.

4:11It's the same idea around learning. We want people to be able to learn about robots and we see actually a lot of non-technical people buying our robots to get a first food. They're curious about robots. They don't want to buy you know 30 ,000 humanoids but it's$300 that's fine. So this is accessibility and I would say the second one is building an open source platform so that people can take this robot and just like the model they can tweak it, they can control it, maybe make it as a you know like a receptionist like an agent to welcome people to a conference you can do all of that with this hardware

4:41Thomas Wolf:It sounds a bit like you care a lot about the education piece and sort of making things accessible to the wider public is that right or can you talk a bit more about that? If you think super long term I think KIA is gonna be everywhere and my my goal is that we feel that's a technology we own and we control and that's it's not a technology that controls us. For me the way to make that is actually to not only let people use the technology like Charg-A-P-T which Which is already great, but also let people build the technology themselves so that they know how it works, you know, internally, what type of data they use to train a model, how you train a model, how you build a robot.

5:16We want this thing to be like common knowledge.

5:18Thomas Wolf:Yes, and so it doesn't really matter whether it's hardware, software, just whatever it is, people need to understand it. This is great for startups, right? Because here, if you want to build the startups, you don't want to just depend on someone providing you the tech, you want to build the stuff yourself. And I think that's one reason a lot of them build an opportunity. Yeah, and I mean, the closed source LLMs are always like a shortcut because they know a lot already and it kind of works. But it's sort of like, it's like the Ozempic of AI because it just, you know, it's a quick fix to a solution, but you don't really understand.

5:50Exactly.

5:51Thomas Wolf:What is, I mean, you also do a lot of angel investing and you just said, obviously, AI is the main thing in startups right now. What are signs of greatness that you look for in founders building with AI? So I would say the founder themselves is usually the bigger thing, so they need to have this ambition. I think AI, the one cool thing about AI is it's a whole green field. You can build a lot of things in AI. And I would say the counterpart of this is you have to be ambitious. There's a lot of small things you can build, like small tweaks. But there is also what is much more interesting to me is you can completely change the world in some parts.

6:29So I'm looking for these founders who have like huge idea about things they really want to change in the world. And the second thing is the world of technical people. You have to be very technical in there. You have to understand, like I was saying, you have to understand a bit how the models work, you know, to understand also what is coming. Because it's nice to build with what is today. It's even better to build with the tech you think there's going to be in, you know, six months, one year. So you have to be very technical to understand that. And then I think you're just good mentoring. and that's why I do a lot of angel investment, which is basically I try to help these people to grow.

7:01Thomas Wolf:How many angel investments have you done? Quite a lot. A lot is good. I mean, you just kind of touched on it. What is next for AI? Your job is to look at what's next. What should we all be getting prepared for? I mean, three things I would say. The first one is robotics, like I was saying. Another one I'm very excited about is AI and science. So basically using AI to not only do a chatbot, which is nice, but kind of solve huge problems or like change the world in a world like let's invent new materials that are just stronger, you know, better for batteries and like more energy efficient, like let's invent new proteins or new like, or here, you know, I talk a lot with Proxima here in Munich, they're making fusion reactor.

7:45That's cool, right? If you get like, so I think being able to, what we discovered is, it's not directly just asking TGPD can you solve fusion that doesn't work but you can do is you learn how this model works and then you train specific models for science you know in each field and this model can accelerate the rhythm of discovery so this is something I'm very excited about I think when you're technical when you're out of a great university like here it's the perfect place right you know something very well like the field of physics biology or something you combine that with knowledge of AI, boom, great things happen.

8:21And I would say the last thing is maybe more for entrepreneur is I think there's a lot of areas that become a little bit static in a way and like AI can totally change them. So I was talking yesterday with the founders working on like compliance, that's very boring area. Here you can come with new eyes, you can come with, you know, knowledge of what you can do with Asia, with AI, you can reinvent this field really. So that's maybe less scientific, I would say, but the impact on the world can be just tremendous.

8:52Thomas Wolf:Maybe something I've heard you have your takes on is sort of, does the current paradigm of LLMs actually get us to AGI? Could you talk a little bit about your take there? Yeah. So I'm one of these persons who don't talk too much about AGI. I'm a bit aligned with, in many ways I disagree with Dario, but on this I'm aligned with Dario from Anthropic, which is, I like to call them more like very powerful AI, which is, I think we'll just keep improving them and there's like no end. AGI, this is the idea that there is a ratio where you're done, right? We reached the max and now the machine continues on itself.

9:28I think it will be more like we'll keep improving and then maybe GPT-6 will still, you know, won't be able to count the number of R in strawberry or like this weird thing. And then we keep solving this thing and pushing the frontier and they will be more and more powerful. I don't think there is really a threshold where we have like self. Now we have it. Yeah. And the same is I think we'll have always this kind of frontier model. They are the frontier. But we have a lot of models that are slightly less intelligent. But they're also very interesting maybe because they're super fast to run. So I think I'm very bullish on models that can run on the edge, on smartphone, maybe new earbuds.

10:04And maybe they cannot solve complex math problems, but probably I don't need that in my earbuds. right? And so I think all this like long tail of intelligence and capabilities is super interesting to explore.

10:17Thomas Wolf:And they have a you know intelligence for specific things rather than having AGI solving all the problems. Exactly. Just like GPT-5 now tend to reroute you on different type of model depending how complex your queries. I think we'll have this kind of multi-step models a lot. Interesting. And you touched on, we touched on robotics a little bit. What else are you kind of building now that is you taking your bets for the future? And how are you kind of preparing, hugging face for that future? So I think small model is something we pushed a lot. One reason is also that we try to push where basically no one is really doing something that's necessary.

10:56So small model is a perfect example. The big research lab don't really want to give you small models because their business model is based on you paying for tokens. So if you get the model running on your phone, it doesn't really work, right? I mean, there might be some licensing agreements in the future, of course, but short term, it's not really their own interest. So they have not even pushed that. So two years ago, we started to explore small models. Can we train them better? And the idea is we take the recipe that we know works for large models, we adapt them and we put them in small models.

11:25And what we discover is you can actually do quite cool stuff with small models. You can actually do really interesting things. And so this is an area we've pushed. Now it's getting pushed more and more. So now there is also a good Chinese team, there is also good like NVIDIA is trying nice model models. So this area is starting on its own. So I would say it needs us a bit less. So now we're looking for the next area where people would need to see us. We have a couple of things. I don't know which one we will explore. I was talking with our head of research who is in Switzerland, Leandro. We were discussing one thing for instance, people, nobody's really exploring.

11:59I mean people are exploring but not as much as diffusion model. These are models that generate one shot like text, long paragraph of text.

12:05Thomas Wolf:Yes. So it's very fascinating. Very few teams are working on that. Oh, that's interesting. Diffusion for text. Yeah. Interesting. All of these things that can go really fast, I think are very interesting because they fundamentally change the paradigm in a way. Like if instead, you know, instead of waiting, if you go to Lovable, you generate website. Right now you wait for the thing to generate, right? If it was instance, maybe you could generate 20 versions of this website and choose maybe more based on the graphical view. So there's a lot of user interface, interaction with the machine thing that might totally change if we go one order, two order of magnitude faster.

12:41That's really interesting.

12:43Thomas Wolf:Maybe one more thing around career advice, because you've kind of had a couple of careers doing a PhD in physics, which maybe informs your interest now in real world physical applications of AI, then you were a patent lawyer, which I found very, very interesting to learn. And then obviously now at Hugging Face, what's your advice for young people coming out of university today or maybe going into university? I think it's a very different learning environment. It is. How would you go about it? I would say in retrospect, maybe you don't have to follow what people say you should follow. Usually people tell you there is one track and it happens to me several times in my life, Like when I was doing master in physics, you had to do a PhD.

13:27And that was like, which is super. When I was a patent attorney, you had to become a partner. And every time at the end, I was like, I'm not sure I really want to do that. And I took this shortcut. But some people, for instance, in the law, some people are still like, I don't really, I mean, I cannot get it why you leave, why you, you know, you left this thing. It was such a great thing, yeah. Why did you decide not to become a partner? So I think in life, you can do this thing where people say you have to do that. And you decide not to. And you go in the other way. I think that's something I have to do a lot.

13:56And the other thing I would say is work with people on topics you really like. So I choose to always be on the people I worked with. First, I mean, my PhD advisor is still a good friend. We were playing music together. We had an amazing time during a PhD. The law firm was also I decided because I had a good friend there. And then my co-founder also was someone I really liked. I think this is, each time you do this, a big career change, you and if you start a startup, keep in mind you're going to be like six to ten years with these people. You want to be with good people. I think the number one reason startup fails is often the co-founders just don't go along very well.

14:35So I think this idea of like developing quickly your sense for you man, who you want to work with, it will also help you for hiring because hiring is a lot of, I mean there is technical stuff, do these people code a lot, but there's a lot of vibe based actually, do I feel this person is going to be you know, great or not? Do I feel I can work with this person? I think building this type of thing is very important. It's maybe a bit hard when you're very technical because you tend to be introverts and not want to talk with people. So I would say when you're at uni, for me it was like participating in clubs, activities.

15:06I was doing a lot of music, science, computer science as well. So I think these things are surprisingly maybe more important than just the raw knowledge that you learn. Especially nowadays. Yeah, exactly.

15:17Thomas Wolf:Yeah, I have a friend who's an AI researcher and he said, usually I would tonight be at home coding, but I feel like the way the world is developing, I should probably come to your party. I think it's a better use of my time now. Come to bits on press. Exactly. I think I need to meet more people because this is going to be tough. Yeah, and I think I love what you said about, you know, maybe don't look for the local maximum, but look for your global maximum and that might mean like going out of a track. especially in Europe. I think in Europe people tend to stick to local maximum and have a little bit of difficult thinking way bigger.

15:51I can do you know a global company from the start I don't have to start in Germany for the German market I can think global.

16:02Amazing! Thank you so much for joining us. Thank you so much.