Challenging the Average With Open-Source AI: Hugging Face’s Thomas Wolf

Today's guest believes open source and closed source models will coexist in the world of AI. Find out what he considers the opportunities and drawbacks to each, as well as how communities can make AI tools and themselves work better together. On today's episode

Hello, I'm Thomas from Hugging Face, and you're listening to Me, Myself, and AI.

Welcome to Me, Myself, and AI, a podcast from MIT Sloan Management Review, exploring the future of artificial intelligence. I'm Sam Ransbotham, professor of analytics at Boston College. I've been researching data analytics and AI at MITSMR since 2014, with research articles, annual industry reports, case studies, and now twelve seasons of podcast episodes. On each episode, corporate leaders, cutting edge researchers, and AI policymakers join us to break down what separates AI hype from AI success. Hey everyone, thanks for joining us again and welcome back to a new season.

Today I'm lucky to be talking with Tom Wolf. He's the co-founder and chief scientific officer of Hugging Face. Tom, great to have you on the show today.

Thanks, Sam. It's a big pleasure to be here.

Let's start with Hugging Face itself. Some of our listeners may not be familiar with Hugging Face. Can you give us a brief overview of what the company does and what you do?

Yeah, of course. Hugging Face is an open source AI platform. We give access to all the AI models that are open source, which means that basically these are the models you can download and run wherever you want. So when you use a AI model nowadays, you can choose either to go to ChatGPT, Anthropic or Google, they are the most widely diffused at the moment. Or sometimes you want to run the AI models on your own data center, or you want to run them on some specific hardware.

Could be like local hardware, or could be maybe faster chips because you need instant response. In most cases, you will want to go for open source AI model, which is a model you can basically just download. There is quite a lot of them on the Hugging Face. There is close to 4 million of these models at the moment. There is one new model being published every five seconds. Some of the most famous one are the Meta series, the Llama series, and one that I think got the most adoption and most visibility recently was DeepSeek, who was released in January and kinda crashed the stock markets at the same time that it was released.

And so over the past eight years, Hugging Face has been building this platform, growing it together with the community of people and teams who are both sharing and downloading models. This community is now roughly 10 million users and AI builders, how we call them, and we've expanded it as well beyond just model hosting to also host datasets, which are used to train models, to fine-tune them, to evaluate them, and more recently also what we call spaces, which are simple low-code way to test all of these models.

So there are a lot of people offering solutions here. And if I think back on the way technology has developed throughout the history of mankind, people came up with chips and Bell Labs and and Intel came along and built fabs for processors. None of that was open source. Why is open source important here?

I think open source has been always important in a way. The thing is open source is more often the long game in computer science. So if we go back to, for instance, the year 2,000 or pre 2,000, where basically Microsoft was one of the largest operating system, right? And Linux was somehow more for fanatics or geeks. And now, if you fast forward twenty years later, right? Unix is really the basics of all enterprise software and all enterprise cloud.

Basically, you almost always run them on some version of Linux, even Mac OS, which I think is probably the most widely diffused on, like consumer laptops nowadays, is one of the largest competitor to Windows based, is itself based on the Unix core. So there is this trend, which is open source has some advantages that make it extremely appealing in the long term. Obviously, on the short term, you can go faster with closed source, and that's also what we see with closed model.

You can iterate faster, you can raise large amounts of capital to train your models, you can try to grab, you know, the most expensive AI researcher and pay them pay them like a huge sum of money. We've keep pushing a lot for open science, and just this Tuesday, we published a new model that's called a smaller M three, which is an extremely smart model, but this is the best one. But uh, at 3 billion parameters, so it's in the range of size that you can run on your laptop and even on a smartphone.

And we've decided to share at the same time all the data, all the recipe, all the knowledge on how to build these models. It's fine for us because we don't make money out of these models, and we think it's very good because anyone who can who want to build a model based on this, or looks a bit like this, or want to extend this type of model now has all the knowledge they want to start. So we think open source can be defined in many ways in AI, but we think the most radical way is to say you share just everything, you share the data, you share the code, you share the recipe.

And we wrote even a very long blog post. We're gonna probably make it into a full-blown paper on all the nitty-gritty details on how to build this model. So we've kept kind of publishing all of this. We're even making a book right now, which is on how to train efficiently LLM on GPU cluster. But we think basically all of this should be really accessible. But I think, yeah, I think I don't want to give the impression that I'm a open source absolutist.

I think both of them just have interesting advantages and drawbacks, and I think both of them will generally coexist in AI. If you compare to hardware, it's an interesting, um, comparison you may, right? Bell Labs was in part mostly developing at the time. Software was probably much more niche, and software and hardware were much more tied together. I think if you compare to hardware, there is some difference, but there is also an interesting advantage and point to be made for open source hardware.

So that's actually something we have started to do very recently at Hugging Face in robotics. I don't know if it's something we want to cover nowadays, but we just acquired this year open source hardware company in robotics called Pollen. And I think there is some definite interest. The general idea is for me that a lot of the way you see hardware in the long term can be reinvented using software. So maybe we'll come to that. Don't want to, uh, to go No, I think too quickly, that's pretty fascinating, you know, crazy tangents.

though, especially actually with the hope junior and the robotics initiatives that you're involved with right now. I I guess you can go ahead and talk about that, but as you're doing that, I wanna push you a little bit and say, I don't even know how we define hardware and software anymore. You've got firmware, you've got so many layers in the software stack, and traditionally, even if we had open source, like Linux was the example you mentioned, it still may have built on top of firmware that was vendor specific and proprietary.

So, you know, even within those stacks, it becomes complicated. But so, yeah, talk about hardware. I I'd like to talk about hope junior.

No, I think this is very interesting. One thing, uh, it's slightly futuristic, but my job at Hugging Face is mostly to think about what's coming. So I do spend a lot of time thinking about the next year and the coming few years in AI. I think just like you're saying, the frontier between software and hardware is becoming maybe smaller again. We had a moment where there was really this huge. I mean, maybe at the beginning it was all the same because everyone was so close to the hardware that basically we didn't really have anything that we would call software, any like large abstraction.

And nowadays, the interesting thing is we see this tendency to kind of fill the gap and to go back to very low level. There's couple of, I think, trends. I I don't have a fully thought about, but what I see is, for instance, people using AI again to speed up the process of building hardware, both computer-assisted development, basically for mechanical pieces, but also hardware like chips, all these type of things. And saying, using AI, we can maybe reinvent and basically, like, lower or maybe digest all of this knowledge that you kind of need nowadays if you want to develop something in hardware.

In a form that's actually so helpful that, again, we could develop hardware a little bit like we develop software, so we could iterate much more quickly on it. We would have much less of an entry barrier of knowledge to basically being able to design things in hardware. So I'm very excited about this, and I think this is in part unlocked by the tools that we now have available.

If I go to Hugging Face's site, I'm overwhelmed. There's so much stuff. And and I think, I mean, it's great that everything's open, but how do we solve then this curation problem of there's a whole bunch of information, I'm never gonna get it all. What do I do? How do I get started?

Yeah, it's a difficult problem because there is like 4 million models, so how do you find the model you want to use, right? In the beginning, we tried to have some curation manual for us, but with one new model every five seconds, that's just not really sustainable. There is a couple of general guidelines, of course. I mean, if you're looking for a speech generation model, you can restrict yourself to this type of model. If you're looking for a text You're down to just model, you

a million.

can, you can filter exactly. And it's surprising because everyone think usually that they're that the category of model they are interested in is the most downloaded, but it's not. It's often like the most downloaded one will be like a speech model, and I get the LLM people or text people are very surprised, they're like, what? But the reality is AI is becoming a huge, huge field with many subfield, and each of these subfield is actually getting really large itself.

So the way we try to do that is we cannot really curate ourself anymore. Just like on the internet, you cannot really just try to curate yourself the best websites. You can have a couple of one you like, you can rely on search, you know, but the best way usually is to rely on kind of a social discovery. So you can go on Reddit, or you can find place where people like likes. If you're on Pinterest, maybe there is some likes for different things.

And so the way we're trying to do that more and more is to give social tools to comment on models, to make collection of models, a little bit like the internet itself, I would say, which is often guided by other people that tell you you should go there, and that's how you find a place. I think the most relying signal right now are two places on the Hugging Face website. One is the blog, we have a blog section that's very active actually, and that has really high quality content, so it's a good thing to follow.

And the other one is just the trending models that show you basically in the past weeks or two weeks, which have been the models that have got the most likes or the most interest among all the new models in the past period.

Yeah, actually, that makes a lot of sense, although at the same time, I can't help but want to connect it to what we said earlier in saying that, oh, it's only the weird thing, it's the not average thing. So maybe you should be looking at the bottom of the list instead of the top of the list, based on our. The problem is there's gonna be a lot there. But back in the original software days, you think about the huge monolithic IBM three sixty operating systems, they were giant and expensive to get going.

And and we've seen in software a great reduction in that cost by building on components. But we still have just a handful of chip manufacturers that again, because of the really billions of dollars that it takes to to create chips and to create hardware. And that may not be true of all hardware, you know, like robotics or whatever have a much lower entry barrier because of the the scale. How do we enable that? How do you push that?

In particular, if we talk about chips, which is a field I'm quite interested in recently, I think there is two converging thing. The the first one is AI, in a way, is quite a simple technology. So having AI chips is actually, uh, most AI chips are much simpler than CPU or GPU. We don't have all of this like complexity we needed to build on top of them, all the branch prediction, all the like the complex things.

We need to make sure we make really the best use of these chips in very generic compute workflow and setup. And a, yeah, an AI model itself is extremely simple. In a way, it's just really this series of matrix multiplication, a couple of nonlinearities, and you have the attention blocks. And that maybe slightly scary the first time you see the equation, but actually that's really that simple, right? If you compare it to really the huge and cumbersome system we had to design to actually make a general computing system efficient in all of these edge cases, just being able to support one forward pass is a baby task in a way.

So this means in a way we can really reinvent how we make this computing architecture itself, and we can make it much, much simpler. So that that's one thing. And if you project in the future, and if you think that basically AI might become the dominant form of compute, by which I mean the dominant form we use energy for compute, this might be the thing we might actually want to really over index on and say actually it's not like this.

The GPU is on the side helping the CPU, but that's actually the reverse case. Now, it's this AI computer is the central piece of what we are building. And so the first thing is, I think there is a way to redesign our computer architecture in a much more simple way, which is kind of a a funny gift that the AI revolution brings us, which is it's much more simple, and it's also much more powerful. Because a well-trained LLM and LLM in the future can simulate extremely complex things, and you can even ask it right now, which is, uh, funny thought experiments that my friend, uh, Steven from Lambda Lab was doing was.

You can ask it to simulate any type of software, and you could even tell it now behave as a spreadsheet thing or behave as a website, and so it behave as a very complex type of software pretty well. And if you look at the core operation, it's just all very simple forward pass, just you have a lot of them. So, yeah, so that's one aspect, which is we have this very simple comput- compute architecture. And the other aspect is we can use this AI software increasingly as helpers to simplify complex task we had to do.

And basically, we can offload a lot of the cognitive load we needed to have to kind of remember. Oh, yeah, I need to check this part of the software. I know where it's on the dock, but I need to check how this work. And there is this thing I also need to be careful about with the development of this type of AI agents, there is some hope that we can really automate a lot of this design part.

Just like nowadays, even when you use like some kind of CAD design software, you actually use something that's extremely powerful. So basically, with just a couple of parametric line, you can design something in three D that used to be extremely complex to design before. And the same thing is I could see us, just like we do a little bit of parametric shape on, uh, on shape, or one of these design software. I can see us designing very complex system by just giving a couple of points on the basic curve and hoping that the AI system will just actually connect all the things and make sure this all fit nicely.

So this is the other thing, not that we are building this system, but we are using them to help us build extremely complex system.

That seems great. You mentioned on shape, my son will now listen to the episode. Because, yeah, I mean, it's amazing. You can take a, like, a two D picture and turn it into a three D making models of himself, and just by taking a frontal picture there, and it's pretty amazing. I only push back on a couple of different things you got going there. One, I think the chip growth is something I hadn't really thought about, but you know, we originally had CPUs, and then we developed GPUs for graphical processing units for processing images on the screen.

And then we had this aha, that, hey, those matrices are the same matrices inside of machine learning models, we can use those GPUs much more efficiently. But what you're pointing out is that we didn't make those chips for that purpose originally, and we might have different design constraints if we did make those chips from the start. And if you combine that with your ability to make chips cheaper, make hardware cheaper in general, then that's a nice combination that might get started.

To be honest, the GPU is is increasingly also an AI optimized chips, for sure. I mean, tensor core, and you see a variety of chips, you see cerebras chips that we work with, a variety of them. I think it's very interesting to follow this field and to see how basically competition push people to explore. You know, just like you were saying, maybe reinventing this, and cerebras is this example of let's be able to host the full model on just this very large wafer scale chips.

Grok, that I was just seeing on the news, is also an interesting case of, you know, let's try to push, maybe for the low batch, the small batch, let's try to push the token per second to the max. Maybe the driving force here, if we step back a little bit and take more like a business view on this, is for the first time. One of the main metrics we have is a very low level metrics, and that's basically the cost per token, and the cost per token is an interesting metrics.

Because it's both something that basically almost a CFO level person could take a look at when you use Gemini, that's the first metrics, you know, people tell you. And when you compare these values per rider, that's maybe the most cost related metrics you take a look at, but it also a metric that's extremely low level. Because if you think about that, that's really just a series of operation, and if you think about that, there's just you can even link that to almost, you know, how many transistor will I activate?

Because this model has this size and and one token is just one forward pass, you could link that to exactly how much, you know, billion transistor you will need to activate for this price. In the past, we didn't have that for the price we were paying for all our compute. We never said, Oh, you will pay actually this amount because you do 10,000 operation. We never had this connection between this cost metrics and the extremely low level of one single operation on the chip.

And so since this metrics is now the main one we focus on, there is this natural tendency to to lower this cost. And that just come directly to the low level hardware, which is I want to actually optimize the cost to do one forward pass on these things. It's very strongly driving the optimization of these chips in this direction.

Yeah, I think metrics are a big deal, and we respond to those. I hadn't really thought about it, but before we talked about computing hours, well, that was hard to understand what a what an hour did for you. At the other end of the scale, we had floating point operations per second, which have much less of a way to relate that to anything I want to do. Versus that the token actually makes a lot more sense then, and then people optimize on it.

It's a dollar per token, that's the thing. We never had the dollar per gigahertz that would maybe, uh, push you to make faster CPU, or like, I don't know, the dollar per flops that would make you make faster GPU in computer graphics. The connection was really wide between these two universe. And so I think a lot of people are saying that's pretty true, that the limiting costs of intelligence would be the cost of electricity. I think that's, uh, for instance.

And the last person I heard saying that was Patrick Collison, but I think a lot of people view this this way. I do agree, but there will be a multiplying factor between electricity and intelligence, and this multiplying factor will be exactly this. Like, how much intelligence can your chips give you per electron, per energy? And that's where you will want to squeeze this as much as possible.

You mentioned cognitive, and of course, a big deal is what we're going to do with these chips. Let's say we've got all these chips and we've got them cheap. What are we going to do with them? And you know, there's a report that I think you reacted to about a country of Einstein sitting in a data center. And, you know, I think that's pretty appealing. The idea that we'd have, oh gosh, we get all these models out there running in data centers, and we'll just suddenly not have one Einstein, but we'll have zillions of Einsteins out there.

Just think of the progress. But today, I mean, tools in general and AI specifically, of course, they can be a head start for people to get to average quickly, so they don't have to spend time getting to average cognitive output. But at the same time, they can also be a way that people just learn to depend on these tools. You know, without practicing skills, we don't get better at skills. How can we go beyond average? So can we have a country of Einsteins instead of sitting at home?

Can we have a country of Einsteins sitting not in data centers but at homes that are using AI tools to provide a head start?

Yeah, for sure. So in this case, started by this essay from Dario Amodei, the CEO of Anthropic, basically saying, Yeah, it's a beautifully written essay. It's very optimistic. It's called Machine of Loving Grace, and it's basically saying that AI will enable us to do extremely important scientific breakthrough. And what he was taking as example was really this Nobel Prize level breakthrough, and where I kind of agree with him, which is, I think if you summarize scientific progress, you have a lot of incremental progress.

And I was guilty of doing a lot during my PhD and postdoc, and that's basically that was also the maximal thing I could do, which is basically do your tiny piece on this little aspect, extending a little bit the frontier. And then you have this massive change of paradigm that usually are the one that will typically be awarded the Nobel Prize. And can be general relativity, can be CRISPR in biology. There is a couple of them in every field, and they usually also create new fields in themselves.

And what I was saying and what I think is AI will be extremely useful for all the incremental innovation. AI is very good at exploring many, many things around the status quo, but AI is extremely bad at challenging the status quo itself. It's very easy to get ChatGPT to agree with you on anything. It's very hard to get this model to disagree with you on something actually and challenge your view of the world, which is quite a problem in some case, and in particular, I think, in scientific research.

Two weeks ago, I had the pleasure to meet again, uh, one of my former professor called Alain Aspect, who got the Nobel Prize ten years ago, I think, if I'm not mistaken, for basically proving this disagreement that Einstein had with quantum mechanics, where basically it was disagreeing with the core idea of quantum mechanics. That if you project the waveform function, you just basically have a random output, and so you can map that in one experiment, and he did this optical experiment.

And if you talk with this type of researcher, they don't want to please you, they have strong opinion, they have strong ideas. And I think it's what actually, you know, led them to make strong discovery, because they were like, I don't think this is right, I want to prove this wrong. And they they don't try to do this type of sycophancy that the LLM will do, where they actually want to please you. I think that's a strong missing point for AI models nowadays is they're really trained to, I mean, first, they are trained to predict the most likely next word in a sentence, which means they will miss word that are unlikely.

If you know, they will tend to regress, just what you were saying, they tend to bring you to the average. They're very good, you know, average thinking or average designer, or creative process if you use them for image designer. But they're quite bad at really challenging the average and going this crazy idea that might challenge some of their training data in particular. And so my point is they they will be very useful research assistant, but they won't be the one really that could lead us to extremely novel breakthrough.

So I think I ask myself a lot this question, in particular when I see my kids, my son using AI, which is how much should they use AI to automate their thinking? What is the remaining part that's very human that we should keep, that we should build? Like always, I think it's it's probably a bad idea to just say, don't use AI. We'll need to find a way to teach them how to use this tool and to remain very conscious of what's the missing part.

You brought in something that I think about a lot is, you know, both with my own kids and also just the university students that I contact all the time is well, what exactly should their relationship be with these technologies? And I think, you know, I I feel like because I'm at the front of the room or it's around the dinner table, that I ought to be having some opinion about this, and it's it's really difficult to know.

It sounds like you're pushing your kids to use these tools and to embrace them to some degree at least.

Yeah, I think you had to. This sounded maybe quite critical of this tool, but I think also this tool are a huge way to unlock creativity. I mean, let's talk about, for instance, vibe coding, I know quite well. This idea that you can prompt a website into existence, and a quite complex website at it. I think it's very fascinating because it used to be quite complex to code a website for sure. And a lot of people just, I think, self-censor themselves and say, Oh, I have this idea for something, but it's so complex to build.

I don't know HTML, I don't want to. You go and you have some no-code tools for sure, but they have all their kind of quirks, the limitation what they can build. Some of them don't have databases and this general idea. I can just ask for these things to exist is is quite new. And so one month ago, for instance, we organized, so my son is twelve, he's still interested in what I do. I would say, luckily for me, for a couple of years, maybe I don't know how long this will last.

But so I managed to bring him and a couple of friends and kids from other friends to a little hackathon we organized with them, where we selected uh, one vibe coding tool, lovable, that that I found very, very easy to use, very nice. And we explained to them a little bit like the design process, that it's better, for instance, to formalize a little bit your idea. So we asked them to draw the website they had in mind, to think about the idea, how it is, and then to prompt it, and we tried to organize a bit their their process.

But the thing we saw is they very quickly grabbed this tool and they started to create much more different app than we thought. So we thought they would just have one idea, but basically they had just ten ideas. And very quickly, each kids was experimenting with four or five different websites at the same time because they wanted to create, you know, this thing to connect scouts with football player, this thing to connect cats owner with second hand cats.

So that was very crazy to see. And then you imagine this, they were between nine and twelve, and to imagine they will grow with basically just deciding that if they want to create a website, it's like just a couple of prompts away. That's just something they can do in a couple of hours. I think it was very beautiful to see. And then you even see them morphing as a little entrepreneur. So my daughter was building this website to connect cat owners with these second hand cat people wanted.

And and she was thinking, oh, maybe I could also ask them to pay when they want to meet each other, because then they need to give you address. And so you see them like starting to ask question that I think just because the technical part is so easy, they start to think to project themselves a lot more in how this would be in real life. So, yeah, that was one recent example that really striked me as an unlocking of creativity I had no idea could exist.

And yeah, September and October, we want to redo this type of hackathon everywhere in the world. We want to see what happen if we do that at the bigger scale than just basically our neighborhood. So I'm quite excited. Maybe when this podcast will go out, we'll have this, uh, worldwide kids vibe coding hackathon going on.

Uh, I may have a couple of kids we can add to that, but I th- I think that ties together well with what you were saying before, which was you have this yes, this tool that always says yes, and will always do what what you want and try to do it quickly. Using that to your advantage and to take advantage of the fact that this tool will, in fact, do everything you ask and try very hard to accomplish that.

And that seems like it ties well with your framing of the tools as good assistants.

Yeah, yeah, I think in a way, we are quite lucky. I'm much more unimpressed, uh, by all the stories of AI, you know, f- freeing itself from its chains and decided to take over humanity. I think the way we are building these tools, and really that's both their advantage, but also their strongest limitation. In a way, is we are really building this tool as assistant to what we want to do.

This has been fascinating. I've really enjoyed talking with you. Maybe by the time this podcast comes out, we'll have some hackathons organized for kids all over the world. It's been fascinating talking to you. Thanks for taking the time.

Thanks a lot, Sam. It was a pleasure.

Thanks for joining us today. On our next episode, I'm joined by Angela McAlembe, engineering program director at YouTube. Please join us for an insightful conversation about trust and safety in the midst of an influx of AI generated content.

Thanks for listening to Me, Myself, and AI. Our show is able to continue in large part due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple Podcast review or a rating on Spotify and share our show with others you think might find it interesting and helpful.