All right. Hello everyone and welcome back to the Sir podcast, and today we have Thomas and Leandro. Now, before we even start, if you wouldn't mind introducing yourself, as well as the work you do at Hugging Face.
Yeah, thanks Peter for having us. So I'm Thomas Wolf. I'm a co-founder and chief scientist at Hugging Face. Hugging Face is an open source AI platform where basically people can find all the models, the data sets, the evals around AI. But also more than just artifacts, we also host a lot of education resources. With Leandro, we've been publishing and writing a lot of blogs and books about AI and how to train and how to use these models.
Yeah, so I'm Leandro von Werra. I lead the research team at Hugging Face. And as Tom said, what we traditionally try to do in the research team is less building the models themselves, but releasing all the knowledge that goes into these models. So we work much more on releasing the data sets, releasing the recipes, releasing blog posts around the models. Very often we still need to train some models to prove to people that the data sets and the recipes actually are valuable. But we think the most value we bring is by building the things around the models themselves, and that way trying to enable more people to be great model builders.
Right. So Hugging Face in the past has really been very in favor of having science and AI together. Now, what is something that you see at this intersection that makes you so willing to commit to it?
I think one thing that's very exciting about the intersection of AI and science is that it's very clear that it's for the public benefit, right? If we manage to advance science in a meaningful way, everybody's going to benefit from that. And the other aspect that I think we strongly believe in is that if this happens in the open, in a collaborative way, there is a much higher chance that these things are going to be evenly distributed, right? I think if all the researchers have access to data sets and to models, to contribute to science and using AI, I think we're going to see an acceleration across the whole field rather than in very spiky local spots where companies have access.
Yeah. I think generally as well, open source and science go really well together, because the more you share things in science — and that's why it's interesting to actually be today at the ICM conference of mathematics — a big part of science is just sharing what you've learned and discussing, and being able to have new ideas by talking about your new idea with a fellow mathematician or fellow physicist. And open source is the same way. Open source, which is the core, kind of philosophical mission of Hugging Face, is: you share some code, somebody gives you feedback. Maybe good feedback, maybe annoying because they're right and you maybe designed something wrong, but it's good feedback, and then you tune and you make your software better. And that's why open source software is actually so powerful today — it's powering the whole infrastructure, from Linux to all servers; everything is based on open source software, because building these things in common is usually way better than building them alone. And so there's just a very natural fit, I think, for a strong open source, open science company like ours to also want to do more things in other fields of science than just AI.
Right? I think that makes sense, especially for science, which has really benefited a lot through time by being open source. Collaboration is such a big part of scientific discovery, and I think it's very important what's being done.
It is. So I'm quite a history-of-science nerd as well. I mean, if you actually look at the first discoveries or inventions, a lot of them were very closed source. So typically you would have Leonardo da Vinci writing in mirror — his mirror writing — so that people could not understand what he was writing. You would typically have this very closed approach until what really changed was around the Royal Society, in like 16-something, where they decided, for the first time, that actually, if someone else is able to test your ideas independently and confirm them, then you will have — that was the beginning of the idea that a very important part of science is reproducibility. Not just what somebody wrote, but to test it again. And that was where the first publishing of science, the idea, was to have the proceedings of the Royal Society and basically have someone publish what they did and explain in very, very deep detail what they went through, so that someone else could test it and reinvent it. I think that's where basically science started to really accelerate, and that's what led to all the scientific discovery then in the 18th and 19th century, and then ultimately, I think, with the 20th century, this huge acceleration, in particular in physics — quantum theory and general relativity being basically based on discussion between scientists and based on having this strange finding being published, so that someone else could start from this strange finding and say, you know, this is very puzzling, but I have an idea of a possible theory explaining that. And for the last thing, maybe I don't know, as a physicist, like the Higgs boson, which is very, I think, mind-blowing, that you can actually explain the math of particles like that. So all of this, I think, has been the reason science has accelerated so much, because of having more and more communication, more and more exchange between researchers. So we think AI should basically follow the same path, instead of maybe what we start to see a little bit, which is: let's not share too much what we know, right?
Maybe just one example I'd like to add, usually in that context, is: if you look at the release of language models, when ChatGPT came out, nobody knew how OpenAI made ChatGPT. Nobody really knew how you take a base model and you make it chat very well, like how ChatGPT did. And I think the whole research field, the open research field, and all the other labs spent a year roughly trying to figure out how do you make a chat model.
