Maybe these AI models have kind of the consciousness of the end of like the workspace in their mind where there are some concepts.
That's the goal, right? Is to stop having to work.
We have more than 4 million models on the hub. I think it's a constant revolution.
Hello, guys. So here now at Machine Summit in Paris, here with you on a digital form from the AI Crypto Minds, and our guest Thomas Wolf, who is co-founder and chief scientist of Hugging Face. And actually it's like GitHub for AI models and a platform for open-source AI models, and now as well on robotics too with the acquisition of Pollen Robotics. So maybe you can introduce yourself to us.
Thank you. Hello, everyone.
So tell me about your story. It's so fascinating. You studied physics as well, then too you used to be a patent lawyer, and then you started Hugging Face. What's the idea behind it? So the super early days, how it came to this machine.
I think I always worked with people I really wanted to work with. First, I felt that we needed a connection with my PhD director. I found him a very interesting researcher. He had crazy ideas, so I thought, 'Actually, I want to work with you.' And the same for other things. And then my co-founders also, when I met them, I found that these were actually guys who I could work with.
And then the topic, I would say I've always loved to understand how the world works. So that's why I studied physics first. I did work in law for a time, and law was my way to understand how human society works, basically. And AI was this way to maybe understand how intelligence itself works. And we saw now, maybe today—I don't know when this broadcast will be—but today, for instance, Anthropic released this very interesting paper around maybe these AI models have kind of a workspace in their mind where there are some concepts. And so I think AI was this fascination about how is intelligence made.
And you are the big proponents of open-source models, so what is the most fascinating thing about these open-source models?
I think it's a constant revolution. So last year DeepSeek was the big wave, right? Everyone realized that this was also pretty good, and actually a lot of people just discovered open-source AI at this time. This year we had that again just a few weeks ago with GLM-4, which was surprisingly close to frontier models. And every year, every few months having this open-source model that suddenly makes a jump, extremely good, I think it's fascinating.
So right now, first of all, the models I really like: GLM-4 is really impressive, really amazing. And the current models, a series of pretty small models that are really good because they're quite easy. I really like the series of models from Google. Google has this series called Gemma, and the recent model Gemma 2 is extremely good. And it's also pretty small—you have a Gemma 2 27B, so you can run them. We just released a few days ago a demo of talking live with this model, and it's actually really, really fast, so you can discuss as with a person basically. The model answers so fast and it also reasons well. So Gemma is my personal favorite right now. But yeah, there are constantly models. We have more than 4 million models on the Hugging Face Hub.
I guess you couldn't try them all, but you tried quite a lot of them.
Yeah, tried it. So we have a ranking of the trending ones, which is usually models people are excited about. And usually you should go on the homepage of Hugging Face; that's a good place to start. You look at what is trending, what are people excited about right now, and usually it's really interesting models.
And let's talk about robotics and your acquisition of Pollen Robotics. So what was the decision behind it? Why do you support open source for this project?
So we started to work on software for robotics first. So we had this open-source library called LeRobot, which a lot of people are using. But then what we discovered is hardware is very expensive. Even the cheapest humanoid is still like several thousand dollars—seven, ten thousand dollars, twenty, thirty thousand dollars typically. And this is very expensive. So we thought there is a need for people who want to start trying physical AI or robotics. There are software developers and pretty deep in AI, and they would love to do robots.
There is a need for an entry-level, accessible robot that you could buy. So we started with a robotic arm, the SO-100, that costs $100. And then we saw a lot of excitement about that, like a lot of people trying to build it or buying kits. And so that brought us to deciding maybe we should do more. And so at the same time I was talking with the Pollen team, and then I asked them why not join us, so why must we just gather this basically entry-level, accessible, yet still highly reliable and high-quality robots.
Is it commercially viable to sell robotics at this price?
We already sold more than 10,000 of these. That was the first one we launched last year. Probably on track to sell 20,000 this year. We're going to have a second one released this summer that we also think is going to be a huge success. Exactly, I'm very, very excited about this new one. It's also the same order of magnitude, between $200 and $500. And yeah, being so many years in the business.
That's great. Are there any other companies you're looking to acquire, Speechmatics or—
