Ep 1: Hugging Face CEO Clem Delangue on The Future of Open vs Closed Source in AI

Jacob and Jason sit down with Hugging Face CEO Clem Delangue and discuss trends in who’s using Hugging Face, the future of closed source vs open source in machine learning, why Clem compares large closed foundation models to Formula One Cars, how enterprise AI teams will evolve and AI safety.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:04Jacob Effron?

  This is Unsupervised Learning, a podcast about the latest in the rapidly developing AI landscape and what it means for businesses in the world. I'm Jacob Effron. I'm joined by Jason Warner, and we're excited to kick off the inaugural episode of Unsupervised Learning. First guest is Clem, the CEO of Hugging Face, a company really at the epicenter of the AI community. And we had a really interesting conversation. We touched on trends in who's using the Hugging Face platform, the future of closed source versus open source in machine learning, how enterprise AI teams will evolve, future of AI safety, and even heard Clem compare the large foundation models to Formula One cars. So definitely stay tuned. It's an interesting point. But a particularly fun one to do with Jason, who is the former CTO of GitHub, obviously a ton in common with

• 0:47Jacob Effron?

  Hugging Face. And it was just a blast hearing him and Clem go back and forth on the benefits of open source, effectively building community. So I think folks are really going to enjoy this one. Without further ado, here's episode one.

• 1:00Jacob Effron?

  Today, we're excited to kick things off with a pretty much perfect first guest. Can't think of a better person than Clem, the co-founder and CEO of Hugging Face. Hugging Face is a pivotal part of the ML ecosystem today, really building the GitHub of machine learning, a place where developers create, discover, collaborate on ML models, data sets, and applications. And Hugging Face also offers a set of hosted services on top of this. The company has raised $160 million, was last valued at $2 billion in May, some great investors, including Lux, Edition, Sequoia, and Co2, and over 10,000 companies using the platform today. So as we dig into these spaces, can't think of a better first guest. Clem, thanks so much for coming on.

• 1:37Speaker 01

  Thanks for having me.

• 1:39Jacob Effron?

  Well, you know, to kick things off, I think would just be great to get a bit of background on how Hugging Face came to be and what you do today. As I understand it, there's a pretty interesting pivot story here.

• 1:48Speaker 01

  Yeah, yeah, as it sometimes happens for startups, right? When we started Hugging Face, we were three co-founders, right, with a scientist, an engineer, and someone more on the product side. And when we kind of like talked about our shared passion for AI and machine learning, we were like, okay, what can we do that is both kind of like scientifically challenging but fun at the same time? And we ended up starting by building this kind of like Tamagotchi AI, AI friend, kind of like ChatGPT, but much more entertaining than ChatGPT.

• 2:28Speaker 01

  And we actually did that for a bit more than than three years. That's that's what we raised our first rounds of of funding on. And as it sometimes happened, we kind of like started to share a little bit what what we were building. Some of our friends, some other companies started to ask us like, oh, how how do you do that? It's it's working pretty well. It's it's really fast. It's it's able to talk about a lot of different different topics. So using a lot of different data sets, a lot of different models. And when we started open sourcing some of that, the usage and the adoption really blew up right from the right from the start, right? Like companies started to to use this.

• 3:11Speaker 01

  Researchers started to use it to share their models, to collaborate. And that's really what what took us from our first kind of like vision and and product to what we are now, which is, as you said, the most used platform for companies to build machine learning and integrate machine learning into their products, their workflows or their features.

• 3:34Speaker 02

  We we can't let you answer that question without talking about the Hugging Face. So so to walk us through the thought process. I will last when I remember meeting you and asking you that question, I love the answer. So let's tell a broader audience what you're doing there.

• 3:52Speaker 01

  Yeah, the when we when we started the company, a running joke with with my co-founders was that we wanted to be the first company to go public with an emoji instead of the three-letter ticker. You know, like when when you go on the NASDAQ. So we absolutely wanted an emoji. And then we were just using and loving so much the Hugging Face emoji that we were like, OK, let's let's go for that. At the beginning, we thought we would keep the name maybe for a few weeks, for for a few months. But then the community started to use it everywhere, to put it in social network when we were posting things, to add it to to their T-shirts or to to other swag. So much so that at some point we were like, OK, there's no way we're going to change this.

• 4:37Speaker 01

  We we need to keep it. And and that's why we still have this name, Hugging Face, and this emoji, the Hugging Face emoji as as our logo.

• 4:45Speaker 02

  I mean, I love it. Um, I remember even when I was still like GitHub, I'm talking to some folks and they're asking about various things in various ecosystems. I'm like, oh, of course, in the ML space, you got Hugging Face. And everyone thought I was joking. Like, no, no.

• 4:58Speaker 02

  Hugging Face is an actual web address and everything. It's like, just go do it. And then they uh they came back and they're like, oh, yeah, we get it now. Let's talk about that name. But awesome. And and we got a whole bunch of questions that we have here that want to go into in in depth, but it's been a little while since you and I caught up. And I gotta ask you, what's life been like for you and Hugging Face since ChatGPT blew up? It basically set the non-ML world on fire. It feels like the ML world is basically kind of like, hey, we're we've been here. We understand this type of stuff. But all of a sudden, ChatGPT, my grandmother, you know, my uncle in Connecticut, they're all talking about it type of deal like. What what's life been like for you guys since that happened?

• 5:39Speaker 01

  It's been amazing, frankly. Um, we we were not so much surprised by the technical capabilities of it, just because we've been working in the field for for some time and and we really saw the progression to get to where we are now. But I think what it did is really unleash the creativity of everyone when it comes to thinking about AI and and machine learning. So I feel like now, when I when I talk to CEOs of companies, when when I talk about, you know, VP engineering, uh, when when I talk about product people, they understand much more the the capabilities of of AI. Um, and they have tons of ideas.

• 6:25Speaker 01

  So in general, for the field, I think it's going to be very positive in unleashing this this creativity for for everyone to see that they can use AI and machine learning for every single thing that that they do. And we've actually seen that with uh with usage of of the platform since uh since the launch of ChatGPT. Actually, the usage of Hugging Face has has exploded.

• 6:48Speaker 02

  I I imagine it would particularly with things like even Karpathy putting up all his, hey, make a ChatGPT, that sort of thing. And like all these instruction sets on. It feels like, you know, as someone, you've been obviously in this deeper than me, but I've been in it for a long time as well. But it feels like this is the actual tipping point for mass market type of stuff. Two thousand or twenty seventeen, when we had that last wave of type of AI, what whatnot. This is different. This feels like it's like, no, we're past that now. We're it's in the zeitgeist. We're we're an AI forward entire world at this point now.

• 7:23Speaker 01

  Yeah, we moved from niche to mainstream. You know, I mean, when when we look at Hugging Face, I think we we crossed two hundred fifty thousand models that have been shared on the platform. I think in in January, we've had more than than fifty million page views on Hugging Face. Which, you know, a few years ago, you would have said that, you know, a machine learning platform would get like fifteen million page views a month. People would have told you like, no way, because because machine learning was still niche. Yes, yes.

• 7:57Jacob Effron?

  And you mentioned kind of like the the users, you know, obviously you've seen a spike in usage. You know, and any difference in kind of the type of people that are coming to Hugging Face or that you're talking to, you know, that you that you've noticed since kind of this mainstreaming?

• 8:09Speaker 01

  I think what's interesting is that it's getting broader in terms of like the domains and the use cases than it was before. And I mean, that that follows the expansion of of the Hugging Face platform too, you know. I I feel like a few years, a few months ago, it was mostly concentrated on some subset of tasks, right? More like NLP, more more around text or some other specific task, for example, text to image, right? With with stable diffusion that that we've all heard of. But now I feel like because some of these tools are quite generalists. People are applying machine learning to a whole range of domains.

• 8:52Speaker 01

  From, you know, text to audio, you know, time series. We're starting to see a lot of companies using machine learning for time series, like your Uber ETA, for example. We're starting to see a lot of biology. We're starting to see a lot of chemistry. We're starting to see a lot of code. So it really kind of like made the scope and the domains of usage of of of machine learning and interest broader than than it was before.

• 9:23Speaker 02

  I'm I'm super curious about this too, because obviously you sit at such a unique intersection of all the things. It's like, you know, we talked about Hugging Face is the GitHub for the ML space, but obviously like sitting inside GitHub, we can watch the entire software industry evolve and understand what was happening is fascinating, right? You can see the rise of JavaScript at one point, or like which new language or framework or whatever was kind of emerging. You can see all that. It's amazing. But now let's ask a different question here. I would say that we're living in a post LLM world at this point and all that sort of stuff. How do you see the entire ecosystem evolving? Where does open source models play in the future of this? You know, OpenAI has a very specific stance on what they want to do with their models, their weights and things of that nature.

• 10:06Speaker 02

  But, you know, there's a, as I believe code wants to be open, I believe models are going to want to be open too and all that sort of stuff. How do you view this? What does, what, you know, how does Hugging Face think about this, but specifically Clem too?

• 10:19Speaker 01

  So first, it's, I think, useful to remind everyone that all current machine learning is based on open science and and open source, right? I think that's the number one driver for the progress of the field in the past few years, right? If it wouldn't have been as open as it's been, maybe it would have taken us 10 years, 20 years, 50 years to get to where we are today. And so at the, you know, macro level, I think it's important to remember that the more open science and and the more open source things are, the faster we can we can progress and evolve as as a field.

• 11:02Speaker 01

  So that's really something that we should aim at and foster, I think, in in the ecosystem. Now, that being said, you know, it's it's also fairly natural that some companies or some organizations want to open source or not open source based on also their, you know, commercial goals, right? And that that was the same thing in on GitHub, right? Not not all companies were using GitHub in in a public mode. And and what I think, yeah.

• 11:34Speaker 02

  I was gonna say, I think it's a really important point that you bring up, which is that it's all based on public. Public knowledge is out there already and it's an implementation. So the model's out there. It's one of those things where, you know, we talk about things that if you're not in the space or maybe you don't fully understand that a lot of the value that's derived here is academically out there already. It's written in papers. It may not be accessible in the format that we thought of it in other domains as accessible, but that's what Hugging Face exists for is to make that stuff more accessible for folks out there. And the fact that OpenAI is doing this remarkable work or Anthropic or other large language models are doing this incredible work, that knowledge is available to the folks. This is exactly why GitHub and Hugging Face have been so important to the

• 12:18Speaker 02

  space is most things exist on GitHub. If you want to go see how the Linux kernel works, you can go look at it. But if you want to see the basis for what OpenAI has been building upon for a while, you can go look at it and figure it out.

• 12:31Speaker 01

  Yeah, yeah. And I think the openness and transparency is going to be even more important for machine learning than it's been for traditional software. Because if you think of machine learning compared to traditional software, to me, the main difference is that machine learning is science driven. It's really driven by scientific improvements of the architectures of the of the models, of the data approaches. And historically, you know, science has always been an open and collaborative endeavor, right? I mean, scientists, they want to contribute to the advancement of science. They want to publish research papers. They are

• 13:15Speaker 01

  actually evolving in their careers based on their publications.

• 13:19Speaker 02

  Yep.

• 13:19Speaker 01

  And so I think machine learning hopefully is going to be even more open and more based on open and collaborative approaches than than traditional softwares.

• 13:29Speaker 02

  Yeah.

• 13:30Speaker 01

  But at the same time, I think there's there's room for Proprietary or closed source approaches, especially as, you know, I think they create new and different distributions for machine learning. And I think it's something that's always, always happened. Like if you think about, you know, software, you have Squarespace or Weeks or like platforms like that that are kind of like making it easier to build technologies. And then you have GitHub, which is kind of like more open and more kind of like a platform for companies to build. Or you've always had, you know, like the Elasticsearch on one side and the Algolia or like other more like closed source search companies on the other side.

• 14:15Speaker 01

  So I think, you know, both approaches are a part of the ecosystem, are contributing to moving the field forward. And we'll always kind of like coexist. And the interesting thing is that on the Hugging Face platform, we have 250,000 models, exactly half of them are private. It's, it's, it's, and it's been, it's been like that for, for quite a while. It's, it's been like that for, for the past two years. And so ultimately, you know, it's, it's good headlines to think, okay, open source is going to kill proprietary or proprietary is going to kill open source. The truth is that you'll always have both that are kind of like feeling different needs

• 14:59Speaker 01

  and providing different, different tools for, for companies depending on their use cases.

• 15:03Jacob Effron?

  Yeah, it's a really interesting point. I mean, I think I've heard you before refer to these kind of closed models like OpenAI and Anthropic as like the Formula One cars of, of the space. And, you know, I, I think, you know, you've obviously been at the forefront of leading some of the open source efforts to kind of come up with foundation model efforts, right? Like Big Science is Bloom. And, you know, one thing that seems to be happening is it seems like some of these players like OpenAI and Anthropic are going to just throw more and more money at compute, right? For like the cutting edge model. And I wonder how you think about like one, you know, the limits of what the open source community can fund on just the pure compute side. And then two, kind of the implications of that for when are these Formula One models going to make the most sense versus like the, the open source alternatives?

• 15:43Speaker 01

  Yeah, the reason why I compare these like very large models to Formula One is because they're they're good marketing, right? The same way Formula One cars are really good marketing. They are very impressive, right? Like the the same way Formula cars is really impressive. But the truth is, for most people out there, they don't need a Formula One car to go to work. They need kind of like a good, normal car that is not going to cost millions of dollars, not only to build, but also to to run and and to drive. And that's that's what we're seeing, you know?

• 16:23Speaker 01

  Probably kind of like these large proprietary models are going to be useful for some use cases. Let's say you're Google, you know, and you want like a very generalist model to be able to answer everything. That's that's great to have a model like that, right? But the truth is, if you want search to work on your website, you actually want a model that is like much smaller, much faster, usually more accurate if you specialize it on your on your own data, and just cheaper to run, right? You don't want it to cost 10 cents for for every prediction that that it makes. So if you look on on the Hugging Face platform, what's interesting

• 17:07Speaker 01

  is that even if we have models from, you know, a few million parameters, which is considered small in in today's world, up to 180 billion parameters, right? On the platform, we host like the largest language models out out there, like you mentioned Bloom or or others. Actually, most of the usage today is for models from 500 million to 10 billion parameters. And it's just because, you know, they're faster, cheaper to run, usually more accurate if you train them on your data. So again, I mean, I think it's it's a matter of like use cases.

• 17:52Speaker 01

  You'll want to use one sort of model for for some use cases, others for other kind of use cases, and and they solve kind of like different problems, I would say.

• 18:03Speaker 02

  Yeah, and I can't see most organizations in the world wanting the everything model, particularly with the complexities that come with those types of models at some point, particularly like these large enterprises saying like, hey, we don't actually know the output of some of the questions that might be asked to this. And so we want to get more specific and we want to like kind of constrain some of those. And just like with, you know, just code in general, but like models specifically here, a bunch of different modalities are going to emerge and adapt and, you know, they're gonna, like you said, the Formula One car to the sedan, to the minivan and all that sort of stuff. There's gonna be a bunch of different things that exist in the world that look like this. Now, I'm gonna ask you a semi-controversial question. I'm curious what your answer would

• 18:46Speaker 02

  be. So what does Clem think about large language models? Should they be open? And philosophically, you know, when companies like OpenAI argue that some of the models might be too dangerous, how do you think about that whole statement?

• 18:59Speaker 01

  Yeah, I think they should be safely open because if you if you look at kind of like the underlying ethical reasoning, so part of the Hugging Face team, we have someone called Irene Soleiman, was actually at OpenAI before and led the release of GPT-2 in open source at OpenAI before, was one of the worldwide experts in terms of safety of releases of of models. And when you look at at the theory behind it and when you try to avoid risks, you actually realize that open sourcing, open releasing, and kind

• 19:38Speaker 01

  of like being transparent and inclusive with these models is actually safer for humanity in the in the long term. An intuitive way to understand that is that by not releasing, not open sourcing, you actually create very big challenges in terms of concentration of power. And you end up with organizations who are able to build, understand models, which are usually big tech or big companies. And then

• 20:13Speaker 01

  You have the rest of society who is not able to understand and then control this, right? You end up with public organizations, with governments unable to understand and create regulation that makes sense. You end up with underrepresented populations who are not able to voice their concerns about some of these biases, the biases that you do find in there. So you create a gap, which makes it actually much more dangerous than if you don't have this gap. And that's why, for example, science, right? If you look at science in general, it's always been open and collaborative, you know, like the scientists,

• 20:55Speaker 01

  they publish their research papers. It's kind of like integrated by the community and by society, because that's kind of like the safest way to progress technology versus kind of like building this behind closed doors and keeping them behind closed doors, where you create this gap of understanding and ability to control between a few organizations that are monopolistic and kind of like keeping control and the rest of society.

• 21:26Speaker 02

  This is probably one of those long topics that we could get into, though I have a particular view, which I think is like philosophically aligned with yours, which is why we probably get along well on topics such as these. But I always have tried to summarize this down. There's a saying, which is, you want to put something out in the light, you'll understand what's going on, like that sort of thing. But how often have we been burned in the past as society in the world for blindly trusting X, where X is whatever it is? And it's almost a universal truth, where if you put your blind faith, trust in X, it bites us, whether that be an institution, a program, a process, or whatever, it doesn't work. Whenever I run into somebody who's having this conversation

• 22:11Speaker 02

  with me, I'm like, great, like, imagine, you know, the IRS is one of these. The IRS is this one massive big black box where you have no idea what's going on inside this thing. You have no idea what's happening. And how often do we get frustrated by this? Now, a lot of people argue that, well, that won't happen in the AI space or the model space, because these companies don't have that same sort of structure. I I just don't believe it. The history, history has never once proven that blind faith trust in X will work out in the long run for humanity always degrades. So therefore, our reaction to that as a society that started in code was to open the things so that no one person or one entity can do that. And it just feels like we're having the exact same discussion that we had about open source software 25 years ago, and we're doing it

• 22:55Speaker 02

  again. But thankfully, people like you exist in the world and the academic and the scientific community is pushing it forward, too.

• 23:02Jacob Effron?

  Yeah, I mean, I'm curious, like, how do you see the the conversation? Obviously, we're in early days of that conversation. You know, how do you see it evolving over the next, you know, two, three years? And and more broadly, obviously, the space has just gone through such transformational change this past year. Like, where do you kind of see us in five years?

• 23:18Speaker 01

  It's an interesting question. We've learned to stop trying to predict things too much because we've been proven wrong so much with the speeds of the evolution of the field. It's kind of like an amazing thing. I I don't think you've seen in the past a field that is becoming so mature, so mainstream, and still moving so fast at the at the same time. It creates a very, very unique challenges that we haven't seen in the in the past. I think what we're starting to understand now is that machine learning is a new paradigm to build old tech, you know, like it's this analogy from Andres

• 24:00Speaker 01

  Karpathy, you know, like Software 1.0 was like the first paradigm to build software. That's what we've been doing for for the past 15, 20 years. Now we're entering a new paradigm, right, which is using machine learning to build all technology. And so I think in in five years, every single technology company is going to build with machine learning. As a matter of fact, maybe we won't even call it, you know, AI or or machine learning then because it's going to be so ubiquitous and it's going to be really kind of like the default way to build all technology that you won't even need kind of like a Name for it.

• 24:41Jacob Effron?

  Right. And it seems like you're starting to see, obviously, a lot of, it seems like even the last few months post ChatGPT, a lot of newer industries or folks kind of come and explore use cases. You know, I'm curious what then role Hugging Face plays around some of those new use cases. Like, for example, I thought it was really interesting, you hired a machine learning engineer to kind of focus on healthcare and biotech. I'm sure there's lots of different verticals of which are exploring these models. You know, how do you think about where you guys play in helping each of those specific verticals? Is it kind of just community engagement, some set of like tailored products for those markets? How have you been thinking about it?

• 25:15Speaker 01

  Well, I mean, the first interesting thing is that we're really seeing the whole fields and all the domains getting more homogeneous than they were before. Before, it was very siloed, right? You had NLP, you had computer vision, you had audio, you had ML for biology. Now, because technologically speaking, you're starting to use the same architectures for all, right? Namely, transformers, you are starting to kind of like break the barriers between all these domains. And that's what we're seeing on Hugging Face with more and more people sharing biology, chemistry, time series models.

• 25:57Speaker 01

  And our goal as a company is really to provide the right foundational platform for all companies to be able to do this paradigm switch, right? From traditional software to machine learning. And so that means kind of like bringing on the community, right? Of researchers sharing models and companies using these models. That means adapting our tools. So for example, one feature that is very popular on the Hugging Face platform are spaces, right? The ability to create demos very, very easily.

• 26:35Speaker 01

  And the spaces for molecule prediction models are going to be a little bit different than the spaces for text generation, right? So we need to adapt and create more capabilities for all these new domains. And then the harder thing, and actually, I think that's the main reason why most of the companies are using our platform. We need to build the right abstractions for it to be future-proof. Right, I think what companies want to do when they use the Higgyface platform is to make sure that what they build today is not going to be obsolete in six months, right? So something we're thinking a lot about is, you know, how do we build the right abstractions, the right tools

• 27:20Speaker 01

  that are going to be used not only today, not only in one year, but in five years, in 10 years, in 15 years? So that companies can take advantage of the evolution of the new models, of the new architectures, but at the same time not get lost on them and have to start from scratch over and over again.

• 27:40Jacob Effron?

  How do you think about that? I mean, I'll shamelessly ask you, I think that's such a big question in the space right now. I feel like there's, you know, you have the existing state of the art models and I feel like so many people are optimizing around, okay, here's how GPT-3 works today, let me get really good at prompt engineering. Who knows if that's going to be a part of models in three, five years? Like, how do you think about, I guess, what are like, no matter what changes happen in the underlying models, these are blocks that are important versus, you know, maybe this is just kind of a moment in time thing that we don't need to focus on as much?

• 28:05Speaker 01

  So first you think about it and you make it a priority, much more than in the previous paradigm, I think. And that comes from our heritage of, you know, having started to think about this like six years ago now. And for example, if you look at our open source libraries and if you look at like the main maintainers of the libraries and how they reply to comments, how they prioritize the different pull requests, the different issues, they actually talk about this a lot and are doing a lot of trade-offs in terms of like making it like future-proof and stable versus jumping on every single trend and breaking everything, making a lot of breaking changes and things like that.

• 28:47Speaker 01

  Then the second thing is, and I think that's one of the things that made us successful, is having the full range of understanding from very, very specialized science to commercial usage of it. It's something kind of like unusual for startups. I think you don't have a lot of startups that have kind of like these full range of capabilities and it comes from our history. Of being three founders with like the the full scope. But that allows you to have kind of like the long term science evolution in mind. While at the same time kind of like building for companies and for final use cases. So that kind of reaches the gap between science to to production.

• 29:32Speaker 01

  Which is also something that is a bit different for machine learning than it used to be for for traditional software. So these would be would be, I guess, the the two main things. And then the last one is, and and Jason, you've you've seen that with with GitHub, I'm I'm sure. But when you've started to be this centerpiece for for the community. The community itself is is helping you a lot on that, right? Because they're keeping you honest, they're, you know, telling you when you when you fuck up what they what they've been building, when you when you slow them down. So like relying on the community and being community driven is also like a a good way to.

• 30:10Speaker 01

  Make sure that what you're doing is is useful right now and and kind of like a future proof for most of your community members.

• 30:18Speaker 02

  Speaking of community, what's next for Hugging Face in this way? And let's talk about, you know, your one, already explosive growth. But two, how you maintain community balance and monetization at some point, too? How do you how do you approach that problem?

• 30:35Speaker 01

  Yes, we have a lot of exciting stuff for for the community. As I was mentioning, we're really excited to interact more with the community in in biology, in chemistry, in in time series. That's that's one. We're also working on building more tools around science and science papers. Because we feel like there's a lot of interesting things to to do there, especially connecting more. Science papers to the models themselves that we have on the hub, the data sets and and the demos. We also have a interesting support of the startup ecosystem.

• 31:16Speaker 01

  So one of our earliest investor, Matt Artman, was at Beta Works, is launching. A fund that is dedicated to helping and and funding startups of the hugging face ecosystem. So we're pretty excited about that. It's a lot of stuff on the community side. Then when it comes to your question around, you know, how to do the trade off between, you know, community and monetization. The interesting thing is that this problem hasn't come so often yet on like our day to day decisions. Especially because I think we've been pretty clear on the

• 32:01Speaker 01

  delimitation between, you know, what are things that are really contributing to the community, right? So for example, companies sharing in open source their models, which we think is creating a lot of value for the community. And so that's something that we'll always keep in the free tier of the of the platform. Compared to, you know, companies that are, for example, using without open sourcing, which are using, you know, at scale for commercial use cases, which are pretty clear kind of like paid offering. So right now we have, you said 10,000, we actually at 15,000 companies using us. Most of these companies are using us for free,

• 32:45Speaker 01

  but we have 3,000 companies that are now customer of ours, ranging from companies like Microsoft, like Bloomberg, all the way to to small companies like Grammarly, for example, and and all the way down to kind of like smaller startups or or people starting with their with their machine learning projects.

• 33:08Jacob Effron?

  Yeah, no, I mean, it it seems like the, you know, the dream for a lot of these companies is to be the one. You you mentioned, obviously, so many companies are now considering, you know, or or kind of seriously adopting their own ML use cases. And it feels like on the startup side, you know, the dream is to be able to be the partner for those companies, the Microsofts, the Bloombergs of the world as they as they kind of go through this journey. And I'm curious, like, obviously, you know, people are using your models and they're and they're they're taking models off. Hugging Face, you have other companies that are kind of close to the open source community like Stability that are trying to, you know, work closely with companies. And then you have other folks that say, all right, take your model off Hugging Face, but then talk to us, we'll be the ones to kind of bring you on that journey. How do you think about ultimately like who earns the right to kind of be that partner to a Bloomberg? Or to, you know, some of these other companies around helping them do this ML transformation?

• 33:54Speaker 01

  Well, I mean, at first, I think there's going to be multiple very successful companies, right? Given kind of like what we're seeing with machine learning becoming the default to build all tech. The same way you have like a number of hundred billion dollar plus companies, technology companies today, I wouldn't be surprised if in five years you would have like multiple hundred billion dollar machine learning companies, right? Just, just because of the size of the, of the opportunity. So fundamentally, I don't worry too much about, you know, competition and what other people are doing. If you look very closely at what everyone is doing,

• 34:35Speaker 01

  they all bring something kind of like different to the table and are kind of like a different approach to the market. And I think overall, it's, it just brings the field forward and kind of like accelerates the, the adoption of, of machine learning. So I think it's, it's all kind of positive. We actually, you know, are lucky to be able to collaborate with, with everyone, right? From companies like, like Stability, hosting, you know, stable diffusion on Hugging Face, OpenAI, open sourcing models like Whisper that are on Hugging Face.

• 35:13Speaker 01

  Yeah, because ultimately, as I said, we, we think there's room for multiple companies to, to be successful. And all of them are really contributing to the democratization of, of good machine learning.

• 35:26Jacob Effron?

  You obviously do so much great work, kind of teaching people about AI through courses you do. You obviously, you know, work closely with these companies. Curious, like where do you think things ultimately shake out with companies and their own AI teams? We have one version of that world today. You could imagine a world in which, you know, people double down even more and these teams get massive. Or, you know, in some senses, it makes sense to outsource some of these parts. Or I'm sure something in between. Curious both like what trends you're seeing today? And then, you know, if you fast forward five, 10 years, like what is the, the AI team at a typical enterprise look like?

• 35:55Speaker 01

  So first, I think all tech teams will be AI teams in the future, right? Like, as I said, because it's like the new paradigm to build all tech. I'm not even sure we're going to call it AI teams or ML teams or ML engineers, but I think if you're building a technology product today, you need to do some machine learning. And one strong intuition that I have is that companies are not just going to use or like outsource machine learning, they're going to build machine learning. You know, like same way, you know, if you look at the first generation of software,

• 36:34Speaker 01

  you could have said, okay, there's gonna be like a no-code tool for people to do everything. And that's happened in some segments, right? You've had, you know, the Squarespace, the Weeks and these tools like that. But ultimately, companies needed and wanted to write code themselves and actually build technology because that's the way they cater to their use case. That's the way they create the differentiation compared to others. That's kind of like the sustainable way of building technology. Similarly, for machine learning, companies who are going to create value

• 37:17Speaker 01

  are going to want to train models, fine-tune models, optimize models, specialize models for their own use cases. And we're already seeing it, like if you look, in my opinion, at like the best startups out there, they're very much kind of like AI-native, but also AI-full-stack startups. Like, I think we've got recently a new release from Runway ML, like the video editing platform. When you look at what they're capable of doing by actually building AI and not just using AIs through APIs or others, it's mind-blowing, right?

• 38:01Speaker 01

  So, yeah, ultimately, I think most companies... Engineers will want to build machine learning themselves and so most of the tech teams will end up being AI teams.

• 38:12Jacob Effron?

  It's a really interesting point about marrying the power of kind of building your own, you know, uh, foundation model with, you know, or your own kind of model for your use case with the application side.

• 38:21Speaker 01

  And what's, um, yeah, what's I think is going to be very exciting is that, and Jason, I'm, I'm sure you have like an interesting opinion on that, but it feels like one of the things that obviously the adoption of software and the impact of software in the world has been really, really fast, but one limiting factor for me has been the number of software engineers. Part of the reason has been, in my opinion, because it's kind of like hard to go from, you know, doing like consulting or doing any other type of work to becoming a software engineer, right?

• 39:04Speaker 01

  It takes, it takes some time to master that and to become a good software engineer. But interestingly speaking, in my opinion, it's much, much easier to switch and to move from being a software engineer to doing an AI or machine learning engineer, however you call them. So one interesting thing is that maybe the first generation of software actually paved the way for much, much faster adoption and progress and impact for, for machine learning. Because it's going to be much, much faster to turn all software engineers into machine learning engineers than it used to be turning any other people into, into software engineers.

• 39:50Speaker 01

  So I'm, I'm excited to see how it plays out, right? And, and see if like, uh, maybe in, in four to five years, it's not impossible to think that you're going to have many, many more machine learning engineers or AI engineers than you have software engineers today. And in that sense, uh, I mean, that's an interesting. Thing that my my co-founder has been saying Julien for for quite a while, which was a bit controversial at the time, which is starting to be a bit less controversial. Is is that is always been saying that.

• 40:28Speaker 01

  Software engineering as we understanding today is actually a more subset of machine learning. Instead of the other way around, right? Machine learning is not so much a subset of of software engineering, but but the other way around. And that's one of the ways maybe we we're gonna going to see that happen too.

• 40:48Speaker 02

  Yeah, I remember Julien saying that for the first time, and I really liked the the thought behind it, because it it really shows what's gonna be. Past some sort of coding event horizon that we can't really fully see yet, you know, and I think that's an important thing for us to think about and like, you know. I think that this is, we've already tried to do this in some other ways. So as example, like Clem, you and I have had this conversation. I said it on a couple of different podcasts. But GitHub says that they've got a hundred million developers or something like that in the world at this point. If you look at Salesforce, when they say how many active sales developers they've got, the number is dramatically higher. And then you see all these no code, low code platforms and the thought behind what GitHub is doing versus what Salesforce is doing with these low code or no code is.

• 41:30Speaker 02

  You're effectively trying to create mechanisms for people to become developers of various systems and it's that all, you know, it's kind of the same way in which. I actually view people who are excellent at manipulating Excel as doing programming because of what they're doing, manipulating data in and out. And if you think about what the tools available to them are, they they're so primitive. The fact that we have to type words into the systems or characters into the systems that become words and become. Fully functional sentences and all that sort of stuff and they compile down and that's the instruction set and I get what we're doing. Even if you have no other exposure to the system, but just Copilot and ChatGPT at the moment. You should be able to imagine a world in which everybody has the ability to go create something in the future here really quickly.

• 42:16Speaker 02

  And whether or not we call that software development in the future, I don't really care. But the point is, other people are creating, a lot, a lot more people are now have the capacity to create, have the, you know, the curiosity to create and have clearly, clearly have the, the tools at their fingertips that are available to them to go create. And that is remarkably powerful, like, you know, and I, I'm super excited for that future and to see what happens with that.

• 42:43Jacob Effron?

  Well, Clement, I mean, it's super interesting to get your, your perspective on, on so many of these topics, you know. Going forward, we plan to kind of conclude our interviews with a standard kind of rapid fire set of, of the same closing questions. And, and as our first guest, you get to be our guinea pig to test some of these out. And so, you know, to start, I'll, I'll let Jason kick it off because I know he loves the, the AGI question. So Jason.

• 43:03Speaker 02

  I do, I do, I love the AGI question. And so, and really it just comes down to like, what are your actual thoughts on AGI? Yes, no, it's a thing, it's not a thing. If so, it's a thing, timeline. And how close are we today if it is and all that? And like, if you are a believer in the AGI camp and think it's going to get there, what should we expect or think about what comes next after it?

• 43:22Speaker 01

  Yeah, I, I don't think AGI is, is coming anytime soon. I think, I think it's, um, it's frankly a distraction for, for the field, I think, because you don't need to think about AGI to make everything that's possible with machine learning exciting. Just the perspective of, you know, being the new paradigm to build all technology and kind of like unlocking new capabilities that weren't possible before is, is exciting enough for me and, and some kind of like more focusing on, on that than, than AGI.

• 43:58Speaker 02

  I love that perspective. Um, I love that perspective.

• 44:00Jacob Effron?

  I guess our, our second question is obviously, you know, I think we're all pretty active on Twitter here, you know, read the same pieces. Uh, I feel like it's the same things that, that get talked about a lot of times in, in the broader AI community. Curious, you know, one thing you feel like is overhyped right now and underhyped kind of in the general discourse?

• 44:16Speaker 01

  So I would be careful with, with Twitter these days, something that I've noticed.

• 44:20Jacob Effron?

  Every day.

• 44:22Speaker 01

  Yeah, yeah. Uh, but it's, it's really something that I've, I've noticed in the past, past few months, kind of like the, the average quality and, and truthfulness of, of the takes in, in AI, in my opinion, has been going dramatically down.

• 44:38Speaker 01

  So I would be really cautious because I feel like now on Twitter, there are a lot of people talking about AI without knowing much about it. And a lot of it is going viral. And I'm kind of like scratching my head and like, what? What is, what is happening?

• 44:59Speaker 01

  You know, I think I love generative AI, right? If you think about it as kind of like the way to focus on like new use cases and the ability for models to generate text or generate images. What we're seeing is that most of the use cases are today on non-generative AI. So for example, the ability to classify, you know, text, to do like fraud detection, you know, like time series that we talked about, to do like ETA, you know, search engine are still mostly non-generative tasks.

• 45:40Speaker 01

  Social network moderation is done with like non-generative models. So I would encourage, yeah, people to not, not, not forget about all the non-generative stuff that is also super, super exciting these days.

• 45:56Speaker 02

  Awesome. All right. Flipping gears entirely from a, well, I guess maybe not because. That's the fun of the rapid fire, right? You get to completely jump around.

• 46:06Speaker 02

  All right, so AI, regulation and governments. Where should, in your opinion, governments and regulation get involved with AI?

• 46:16Speaker 01

  In my opinion, this should foster more openness and transparency. Because as I said before, there's no way to control something that you don't understand. Right? Or there are only bad ways, right? If you want to control and regulate something that you don't understand, you'll only kind of like take the wrong steps. And so if there is a way to create kind of like a foundational regulation for me, it's to foster more transparency and more openness. Because that's what's going to allow regulators everywhere to, you know, start building, understanding more what's, what's happening and then create kind of like the right regulation to mitigate the risks and make it a positive technology for, for humanity.

• 46:59Jacob Effron?

  Well Clem, this has been a fascinating conversation. I feel like we could go hours on any of these directions, but I feel like we'd be doing a disservice to the broader ML community to take more of your time away from Hugging Face. And so we'll let you get back to the day job. I'm sure a lot of our listeners are intrigued by you and what you and the team are building. What's the best way for them to go learn more about what you guys are doing at Hugging Face?

• 47:19Speaker 01

  Yeah, you can take the Hugging Face course. That is like a really good introduction to our ecosystem. So hf.co/course. You can follow us on Twitter, on LinkedIn. You can join our Discord to join the community. And then explore the Hugging Face hub. It's very much like GitHub where you can find a lot of things that people are not talking about, but that are very, very powerful in all sorts of machine learning, from like text, audio, image, biology, chemistry, time series. So really just explore the hub.

• 48:02Speaker 01

  It's a good way to also unleash your creativity about what's possible and not with machine learning.