Hugging Face and watsonx: why open source is the future of AI in business Malcolm Gladwell: Hello, hello. Welcome to Smart Talks with IBM, a podcast from Pushkin Industries, iHeartRadio and IBM. I’m Malcolm Gladwell. This season, we’re continuing our conversation with New Creators – visionaries who are creatively applying technology in business to drive change — but with a focus on the transformative power of artificial intelligence and what it means to leverage AI as a game- changing multiplier for YOUR business. Our guest today is Jeff Boudier [BOO-dee-ay], Head of Product and Growth at Hugging Face, the leading open-source and open- science artificial-intelligence platform. Before getting into the world of open-source AI, Jeff co-founded a company called Stupeflix, a video-editing software which was eventually acquired by GoPro. An engineer
by background, he has a self-professed obsession with the business of technology. Recently, IBM and Hugging Face announced a collaboration, bringing together Hugging Face’s repositories of open-source AI models with IBM’s watsonx platform. It’s a move that gives businesses even more access to AI while staying true to IBM’s long-standing philosophy of supporting open-source technology. With open source, businesses can build better AI models that suit their specific needs, using their own proprietary data, while browsing a ready catalog of pretrained models. In today’s episode, you’ll hear why open source is so crucial to the advancement of AI, how IBM’s watsonx interacts with open-source AI, and Jeff’s thoughts on why the singular, omnipotent AI model is a myth. Jeff spoke with Tim Harford,
host of the Pushkin podcast Cautionary Tales. A longtime columnist at the Financial Times, where he writes the “Undercover Economist,” Tim is also a BBC broadcaster with his show More or Less. OK, let’s get to the interview. Jeff Boudier: Hi, I'm Jeff Boudier,
and I'm a Product Director at Hugging Face. Tim Harford: So I'm immediately intrigued. Hugging Face—is this a reference to the Alien movie, or something else? Jeff Boudier: It is not. And it may be not obvious to a listener, but “Hugging Face” is the name of that cute emoji—you know, the one that's smiling with his two hands extended like that to give you a big hug? That's Hugging Face. So basically we named the company after an emoji. Tim Harford: Okay. And it is—I saw your website, and it is a very friendly emoji. So that's nice. So tell us a little bit about Hugging
Face and, and about what you do there. Jeff Boudier: Of course. Well, Hugging Face is the leading open platform for AI builders, and it's the place that all of the AI researchers use to share their work, their new AI models, and collaborate around them. It's the place where the data scientists go and find those pretrained models and access them and use them and work with them. And increasingly, it's the place where
developers are coming to turn all of these AI models and datasets into their own applications, their own features. Tim Harford: So like the Facebook group or the Reddit or the Twitter for people who are interested in, particularly, generative-language AI, or all kinds of artificial intelligence? Jeff Boudier: All kinds of AI, really. And of course generative AI is this new wave that has caught the world by storm. But
on Hugging Face you can find any kind of model. The new, sort of, transformers models,to do anything from, translation—or if you wanted to transcribe what I'm saying into text, uh, then you would use a transformer model. If you wanted to, then take that text and make a summary, that would be another transformer model. If you wanted to create a nice little thumbnail for the, this podcast by typing a sentence, that would be, uh, another type of model. So all of these models you can find—there's actually,
300,000 that are free and publicly accessible—you can find them on our website at huggingface.co, and use them, using our open-source libraries. Tim Harford: And so this is—this is fascinating. So, so there are 300,000 models. Now, when you say “model,” I'm thinking in my head, “Oh, it's kind of like, um, like a computer program.” There were 300,000 computer programs. Is that—is that roughly right? Or it—not really? Jeff Boudier: It's a general idea. A model is a
giant, set of numbers that are working together to sift through some input that you're going to give it. So think of it as a big black box filled with numbers. And you give it as a, as an input maybe some text, maybe a prompt. So you're asking—you're giving an instruction to the model, or maybe you give it an image as an input, and then it will sift through that information, thanks to all of these numbers, which we call, in the field, “parameters,” and it will produce an output. So when I told you, “Hey, we can transcribe this conversation into text,” the input would have been the conversation in an audio file, and then the output would have been the text of the transcription. If you want to create a thumbnail for this podcast episode, then the input would be what we call the “prompt,” which is really a text description— like, uh, “French man in San Francisco talking about machine learning.” And the output would be a completely original
image. So that's how I think about what an AI model is. And I think what we're starting to realize is that this is becoming the new way of building technology in the world. It has been—for the field of dealing—understanding, generating text—for quite some time. But now it's sort of moving across every field of technology. We have models to create images, as I say, but also to generate new proteins, to make predictions on numerical data. So every kind of field of machine learning is now using this new type of models.
But what's interesting is that if you're,say, a product manager at a tech company, and you say, “Hey, I want to build a feature that does this,” a few years ago the approach would have been to ask a software developer to write a thousand lines of code in order to build a prototype. And a new way of doing things today is to go look for an off-the-shelf, pretrained model that does a pretty good job of solving exactly that problem. So you can create a prototype of that feature fast. So it's a new approach to building tech. Tim Harford: I'm not a programmer, but I'm aware
that there's this idea of open-source code, and now we have open-source models. So what does it mean for something to be “open source”? Jeff Boudier: Yes, “open-source AI” actually means a lot of different, specific things. It's the open-source implementation of the model. So if you use the Hugging Face transformers library to use a model, you're using an open-source code library to use that model. Tim Harford: Just to interrupt on the “transformers”—these are these, kind of, ways of turning a picture of a dog into a text output that says, “Hey, this is a picture of a dog,” or “This is a French text,” and—with the transformers helping you turn it into English text. Or it's doing all
of these things that you've been describing. That's—the, the transformer is the—kind of the engine at the, at the heart of that. Jeff Boudier: Yes, exactly. And we call them “transformers” because they correspond to this new way of building machine-learning models that was introduced by Google, actually, with a very important paper called “Attention Is All You Need.” And it was published in 2017 by researchers out of Google DeepMind. Tim Harford: Wow,
that's just six years. That's so new. Jeff Boudier: It is very new, and ever since, the pace of innovation has really, really accelerated. But it really started from this inflection point that came from this paper and its implementation in what is now called “transformer models.” The transformer—that has conquered every area of machine learning since. Tim Harford: Okay. Sorry to interrupt. So, so—you've got this library of transformer models, and they're open source, and that means, that means what? Anyone can use them for free? Or that, or that anybody can implement them for free? What does it mean? Jeff Boudier: So again, there's, like, lots that go into it, but the most important thing is for the model itself to be available so that a data scientist or an engineer can download them and use them. And also there are a lot of considerations about how you make them accessible. And a very important one is whether or not you give access to the training data—all the
information that went into training that model and teaching it to do what, what it's trained to do. That's, uh— Tim Harford: Have fed millions of words into a, into a language transformer, or I might have fed millions of photographs into a, into a picture transformer. Yeah. Jeff Boudier: Yes. And the accessibility of that training data is very, very important. TIM HARFORD: What's the relationship between the,
the Hugging Face libraries and, uh, GitHub? Which, if I understand GitHub correctly, it's this—the repository of open-source code, lots and lots of lines of code and routines and programs that are, that are shared and updated and tracked, and they're all available on, on GitHub. Which sounds similar to what you're doing with Hugging Face for AI. So what, what is the interaction or the relationship there? JEFF BOUDIER Yeah, I think you nailed it on the head. So Hugging Face is to AI what GitHub is to code, right? It's this central platform where AI builders can go find, and collaborate around, AI artifacts, which are models and datasets. So it's
quite—it's different than software, but we play the central, the central role in the community to share and collaborate and access all of those artifacts for AI like GitHub offers for code. Tim Harford: And that community must be incredibly important. I mean, the—open source is nothing if you don't have a community of people working on it. So how have you been able to foster— Jeff Boudier: Well, I think it goes to the origins of the transformer model, and, and Hugging Face rolled into that. So when the first, sort of, open model came out, it was called BERT,
and it came out of Google. The only way you could access it was to use a tool called TensorFlow. But it happened that most of the AI community was using a different tool, called PyTorch. And something that Hugging Face did is to make that new model, BERT, accessible to all PyTorch users. And they did it in open source. It was a project called “BERT's Pretrained PyTorch” or “BERT PyTorch Pretrained.” Tim Harford: So this is like being able to play my Zelda game on, on an Xbox or a PlayStation, right? Or am I not really understanding what's going on? Jeff Boudier: No, that's exactly what it is. And the thing is, everybody was using the Game Boy, and so it became very popular. And from there, the community sort of gathered to make all of
the other models that were then published by AI researchers available through that library, which was quickly renamed from “BERT’s Pretrained PyTorch” into “transformers” to welcome, like, all of these different, new models. And today, that's—open-source library transformers is what all AI builders are using when they want to access those models, see how they work and build upon them. Tim Harford: What's striking about this field is that it's changing so fast. It's improving so quickly. So how do open-source models keep up with that? How do they get iterated and improved? Jeff Boudier: Actually, it's not so much that open source is keeping up with it; it's actually open source that is driving that is driving this pace of change. And that's because— with open source and open research, data scientists’, researchers can build upon each other's work. They can reproduce each other's work. They can
access each other's work using our open-source libraries, et cetera. So in a sense, it's not really that open-source AI is a new idea. It's rather the opposite. There's been a blip of time in which closed-source AI seems to be the dominant way, but it's really a blip. In fact, you know, none of the incredible advances that we're marveling about today would be possible without open source. We're standing upon the shoulders of 50 years
of research in open-source software. So I think that that's really important; if it wasn't for that, we'd probably be 50 years away from having this—amazing experiences like ChatGPT or Stable Diffusion, et cetera. So it's really open source that is fueling this pace of change—all, all these new models, all these new capabilities. To give you an example: so Meta released, uh, LLaMA large language model just a few months ago. And ever since, there's been this Cambrian explosion of variations and improvements upon the original models. And today there are over a thousand of them that we host and track and evaluate. So, yeah, open source is really the gas and the engine for that.
Malcolm Gladwell: Jeff just made it clear that it is open source, not closed, that sets the pace for AI innovation. If that’s true, then forward-thinking businesses shouldn't shy from leveraging open-source AI to solve their own proprietary challenges. But how? Businesses can face serious obstacles when trying to adopt open-source technologies, like complying with government regulation or making sure their customers’ data stays protected.
In the next part of their conversation, Jeff and Tim discuss how IBM’s collaboration with Hugging Face empowers businesses to tap into the open-source AI community, and how the watsonx platform can enable them to customize those AI models to their needs. Tim Harford: I just wanted to ask about the partnership between Hugging Face and IBM. How did that come about? Jeff Boudier: Well, it came through a conversation. A conversation between our CEO, Clément Delangue, and Bill Higgins, IBM, who's
really, really close to all the amazing research work and open-source work that's happening at IBM. And that conversation sort of sparked the evidence that we needed to do something together. We share a lot of values in terms of the importance of open source, which is fundamental to us; with the importance of doing things in a, an ethics-first way, to enable the community to incorporate ethical considerations in how they're building AI. And we sort of have a different audience to start with, which is all the AI builders use Hugging Face today. To access all the models we talked about; to use them using our open source, and build with them. And IBM has this incredible history of working with enterprise companies and enabling them to make use of that technology in a way that's compliant with everything that an enterprise requires. And so being able to marry these two things together is an amazing opportunity. And now
we can enable the largest corporations that have, sort of, complex requirements in order to deploy machine-learning systems and give them an easy experience—to take advantage of all the latest and greatest that AI has to offer through our platform. Tim Harford: Let's talk about this idea of a, of a single model or a variety of models. Because what I've been hearing you say—you've been saying, “Oh, there are lots of models, there are hundreds of thousands of models available on Hugging Face.” But you've also said “There's this single thing, the transformer, and they're all transformers.” So if they're all basically the same thing, why can't you just build one super-clever model that can do everything? Jeff Boudier: That's a really interesting idea, and very much a new idea. And the reason we have
over a million repositories, 300,000 free and accessible models, on the Hugging Face platform is that models are typically trained to do one thing—and they're typically trained to do one thing with specific types of data. And what became new and evident in the research that came out over the last couple years is that if you train a big enough model with enough data, then those models start to have a—sort of general capabilities. You can ask them to do different things. You can even train them to respond to instructions.
So with the same model you can say, “Hey, summarize this paragraph, translate this into English, start a conversation in French, and then pivot to German.” And so these are general, sort of, language capabilities. And I think when ChatGPT came online, and the world, sort of, discovered these new capabilities, there was, at least for a short period, this sort of idea, this sort of myth, that “the end game of all this is maybe one or a handful of models are so much better than anything else that exists, that they can do anything that we can ask them to do. And that's the only model that we will need.” And I, for one—I think it is a myth. I don't think it is practical, for a variety of reasons. Say you're writing an email and you have, like, this great suggestion of text to, sort of, complete your sentence. Well, that's AI. That's, that's a large language model. That's a transformer model that does that.
So there are a ton of existing use cases like this, and these use cases are powered by specific models that have been trained to do one thing well and to do it fast. If you wanted to apply these, sort of—all-knowing, powerful, Oracle type of model, you would not be able to serve millions of customers through a search engine. You would not be able to complete people's sentences, because the amount of money that you would need, the number of computers that you would need, to run such a service—it just exceeds what is available on the planet. So one reason for which it's not a practical scenario is that it's just very expensive to run those very, very large models. Tim Harford: What I'm hearing is, it's like, “Look, if you want to screw in a screw, you need a screwdriver.” You don't
want an entire toolshed full of tools if the task is to screw in—a screwdriver. And sure, you could bring the toolshed. There are all the tools; there's a screwdriver there. But it's not necessary. It's incredibly expensive. It's incredibly cumbersome, and that cost exists even though, maybe, as the user who's just typing in a, into a prompt box, the user may not see it, but it's still very real. Jeff Boudier: That's right. And then another one is performance. So, taking the screwdriver example, so—and by the way,
like, we're not quite there at this moment where we have this all-knowing, powerful Oracle; that is still sort of a sci-fi scenario. But—we have screwdrivers, but we also have the Leatherman, right? The, the multitool—a Swiss Army knife, and that's, sort of the moment that we are in today. But now if I'm trying to open up my computer, it turns out that it requires a specific kind of a screw, like these tiny little Torx screws, and having a Torx screwdriver will get me much further than trying to use my Leatherman, where maybe I'll get the knife blade and it'll, it will mess up the screw, and maybe eventually I'll get to what I need—but my point is that if you take a very specifically trained model for a particular problem, it will work much better. It will give you better results than a very, very generalistic, big model that can do a lot of things. And so for things like search engines, for things like translation, for things that are very specific, companies are much better off using smaller, more-efficient models that produce better results. Tim Harford: That's really interesting and
presumably then being able to know which model to use, or being able to know who to ask which model to use, becomes a very important capability. Jeff Boudier: Yes, and that's what we're trying to make easy through our platform. Tim Harford: So tell me about how this works with IBM's watsonx platform. How do you see Hugging Faces’ customers benefiting from that? Jeff Boudier: The end goal is to make it really easy for watsonx customers to make use of all the great models and libraries that we talked about—all the, the, the 300,000 models that are today on Hugging Face. And to do this, we need to really collaborate deeply with the IBM teams that build the watsonx platform, so that our libraries, our open source, our models are well integrated into the platform. If you're a single user, if you are a data-science student, and you want to
use a model, we make it super easy, right? We have our open-source library. You can download the model on your computer and run with it then. But in enterprises, there is a vast complexity of infrastructure and rules around what people can do and how the data can be accessed. And all this complexity is, sort of, solved by the watsonx platform. Tim Harford: This season of the Smart Talks podcast features what we're calling New Creators. Do you, do you see yourself as being a creative person? Jeff Boudier: I think it's a requirement for the job. I mean, we're in such a new and rapidly
evolving industry that we have to be creative in order to invent the business models, the use cases, of tomorrow. My role within the company is really to create the business around all of the great work of our science and open-source and product team. And by and large the business model of AI within the whole ecosystem is still something that companies are trying to figure out.
So creativity is really, really important—to really have the conversation with companies, understand what they're trying to do and then build the right kind of solution. So that's, like, where creativity comes into play. Tim Harford: And one of the things that you've, you've been talking about is just this, this growing number of models, this growing number of capabilities, this growing number of use cases—enormously exciting. But also, I think, completely bewildering for most people who are trying to navigate their way through this, this maze of possibilities that is, that is growing faster than they can even learn about it. So how are you helping people navigate and make choices in that environment? And how does the partnership with IBM help with that? Jeff Boudier: Hmm. Well, as I said, like, our vision is that AI machine learning is becoming the default way of creating technology. And that means, like, every product, app,
service that you're going to be using is going to be using AI to do whatever it is better, faster. And I guess there are two competing visions of the world coming from that. There is this vision of the Oracle all-powerful model that can do everything. And our vision is different. Our vision is that every single company will be able to create their own models that they own, uh, that they can use, that they control. And that's the, the, the vision that we're trying
to bring to life through our open-source tools that make this work easy, through our platform where you can find all those pretrained models that are shared by the community. So we really want to empower companies to build their own stuff, not to outsource all the intelligence to a third party, and the watsonx platform from IBM gives those tools to enterprise companies. So that's—you can use the open-source models that Hugging Face offers. Then you can improve them with your own data without sharing that data to a third party.
And then you could do every—all of this work in compliance with whatever governance requirements you have for your company. Maybe you’re a finance services company and you have a specific set of rules. Maybe you’re a healthcare company and you have very strong privacy requirements for patients’ data. Maybe you’re a tech company and you have your, your customers’, your users’ personal information. So you need to be able to do this work, respecting all of that.
Tim Harford: Jeff Boudier. Thank you very much. Jeff Boudier: Thanks so much, Tim. It was fun. Malcolm Gladwell: To create the AI models of the future, we are going to need open source. That means there’s a place for business in the open- source community to harness the game-changing potential of AI innovation. Like Jeff said, businesses face unique challenges they need to solve at scale. Without proper support systems, tapping into open- source AI at enterprise level is daunting: finding the right-sized model for the job, fine-tuning its purpose, all while addressing governance requirements around data, privacy, and ethics.
So for businesses, IBM’s collaboration with Hugging Face is a mark of progress because it signifies that business can tap into open- source AI while preserving enterprise-level integrity. Businesses should embrace the open-source community and the AI future, much like Hugging Face (and its emoji namesake) suggests. I’m Malcolm Gladwell. This is a paid advertisement from IBM.
Smart Talks with IBM is produced by Matt Romano, David Zha [JAH], Nisha [Nih-sha] Venkat, and Royston Beserve, with Jacob Goldstein. We’re edited by Lidia Jean Kott. Our engineers are Jason Gambrell, Sarah Bruguiere [Brew-Ghare (hard G!)], and Ben Tolliday. Theme song by Gramoscope. Special thanks to Carly Migliori, Andy Kelly, Kathy Callaghan [Calla-Han], and the EightBar and IBM teams, as well as the Pushkin marketing team. Smart Talks with IBM is a production of Pushkin Industries and Ruby Studio at iHeartMedia. To find more Pushkin podcasts, listen on the iHeartRadio app, Apple Podcasts, or wherever you listen to podcasts.
2024-05-03