How Hugging Face raised $235M

Full Transcript

Transcript

Nathan Labenz: (0:00) ML is the next programming in general. It's so massive and the people that do it are so important. Owning that platform becomes a strategic priority. Again, you look back at the list of investors here and Google and Amazon, given that they don't own GitHub and they want to compete with Microsoft in everything, I can see how an exit to 1 of those companies would be supernatural. Hello and welcome to the Cognitive Revolution where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz joined by my cohost, Erik Torenberg. Hello, and welcome back to the Cognitive Revolution. Today, Eric and I are talking about Hugging Face, which just announced a major series D fundraise at a $4,500,000,000 valuation from a long list of major strategic corporate investors, including Google, Amazon, and Nvidia. I had been watching for this story to drop anytime because going back to July, I had received a flurry of expert network call requests to discuss Hugging Face. So in this episode, we'll cover everything that I told the investors. Hugging Face is a fascinating company. Started in 2016 as a chatbot company more akin to Replica than the Hugging Face we know today, the company has gradually become both the leading online hub where people share and discover all sorts of AI models, datasets, and demos, and also a leading champion that the open source AI community counts on to advocate for equal open access to AI tools in the halls of power such as The United States Senate. I find Hugging Face and this deal in particular to be extremely interesting because while the Hugging Face team is absolutely stacked with incredible experts, and I've interacted with many of them, and the value that the Hugging Face platform brings to the community is unmistakable, I think it's still currently much less clear that the company will be able to build a business that supports such a high valuation over time. While the company is a clear leader in AI model discovery, it's currently just 1 of many options in the fiercely competitive compute market. And while demand for compute is growing so fast that it's likely to support growth for all players for a while to come, I do see reasons to believe that the company's commitment to open source everything could make it quite difficult to sustain moats and margins in the future. We'll get into all that in more detail. But zooming out from Hugging Face for a minute, you might have noticed that we've recently been doing more analysis episodes with just Eric and me discussing important news and developments in AI. I love doing interviews for this show and personally learn a ton from the process. But our audience data, imperfect as it is, suggests that you all might be getting as much or maybe even more value from this sort of analytical content. Our goal is not to cover all the news and probably never will be to cover all the news, but instead to contextualize key developments with in-depth analysis that I'm not seeing well represented online. As always, we value your feedback, and I'd particularly love to know if you'd like to hear more of these analytical episodes or if you still prefer the interviews. To borrow a meme, we're in the podcast arena trying stuff. Some will work, some won't, but we're always learning. So let us know how we're doing. Leave a review on Apple Podcasts or Spotify, leave us a comment on YouTube, or contact us directly via email at tcr@turpentine.co or by DM-ing me on the social media platform of your choice. Now here's an in-depth discussion on Hugging Face and the unique role it plays in today's AI ecosystem.

Erik Torenberg: (3:57) Nathan, we're here today to talk about Hugging Face and their most recent announcements and analyze the company a little bit. Why don't you give maybe a little bit of overview on the company and the product, and then we'll get into it.

Nathan Labenz: (4:09) Cool. Yeah. Happy to. I think this will be a fun 1. So context for this is Hugging Face just announced a big series D fundraise, $235,000,000. That is now more than half of all the capital that's raised to date, which is just under 400,000,000 total raised. And investors, some notable ones including Google, Amazon, NVIDIA, Intel, AMD, Qualcomm, IBM, Salesforce, and Sound Ventures with a valuation of $4,500,000,000, which is up 2x over the last year since the last round. And notably now valuation over a 100 times the company's annualized revenue. But I think they're doing something around 30,000,000 a year in revenue, which would put this at a 150x. Anyway, it's north of 100. So pretty heady valuations for a company like Hugging Face. And without disclosing too much, I've signed up for a bunch of expert network type mailing lists. And about a month ago, there were a flurry of, I don't even know what companies were scheduling these calls, but definitely all of a sudden everybody across networks was interested in talking about Hugging Face. So I think this is a really interesting 1. There's obviously been a ton of debate about where the value is in the stack, which companies have positions that are going to prove defensible. Is it about infrastructure? Is it about bleeding edge algorithms? And this 1 is kind of a unique 1 because more, and this is why I really wanted to talk to you more about it too, is this 1 is more about community than maybe almost any other AI startup that's got a high flying valuation right now. And it's definitely, I think, pretty interesting to kind of compare its value. And we can get into more detail on its product line too and the role that it plays in the ecosystem. But it's interesting to compare its value to something like a runway, which has raised over $1,000,000,000, but is much more focused on models or somebody like Mosaic, which was acquired, obviously, as we've talked about on a couple of episodes for north of $1,000,000,000, much more focused on software scaffolding to help you perform your own model training as well as an inference product. I think what's really interesting about this 1 is product head to head, point for point, feature for feature, and certainly revenue probably as well, doesn't seem obvious to me that Hugging Face is killing it. But what is special about it is this kind of place that it occupies that's at the center of so much activity in the broader ecosystem.

Erik Torenberg: (7:12) Let's get into that. But first, can you give a little bit of overview or deep overview on kind of its product line?

Nathan Labenz: (7:17) Yeah. It's interesting. So Hugging Face has a lot of surface area, put it that way. I first started to pay attention to it as just a way of finding new cool stuff. Early 2022, as I was looking very actively for Waymark for what's the best model to caption an image? What's the best model to pick an image that corresponds to a certain bit of text? This stuff was just starting to work at that time. And over and over and over again, it seemed the demos were starting to pop up on Hugging Face. That was just kind of the place I kept noticing everybody was posting their demo experiences to. That has only really accelerated, but I'd say that's been true now for the 18 plus months. From there, I got engaged with them, and I don't even know if they still offer this program, but it's definitely something new where I learned a lot from it in multiple respects. I believe at the time it was called the Expert Acceleration Program. The basic offering there was, okay, you have a lot of questions. We have a great team of experts that is expert in everything, and we will help you figure out what the best models are. You can kind of ask us a certain number of questions each month. And so for me as a 1 person AI R&D department, I was thinking, I could really use that. There's definitely some times when I have questions about feasibility. Could I fine tune this or that? How does this work? What is the state of the art? And that's 1 thing I've always kind of gone back to Hugging Face with my feedback. I'm saying more curation, I think, is something that they're going to have to continue to develop. Because basically right now, it all starts for them with an open source library, which is transformers. They now have a diffusers library as well, but they created just a canonical open source library. And then they created this concept of the model hub on top of that library. And it's basically all the models that can run on this open source execution code becoming, I wouldn't say the standard, but a standard as kind of an underlying library and then increasingly playing for the standard of the place where you would post your models to. And then elaborating on that has kind of created this over time, this very diffuse web experience where there's sort of pages. There's a whole section now where it's all the papers almost trying to play kind of publishing platform in a way vis-à-vis, I mean, arXiv, of course, but they're mirroring all the papers on the hub and they actually have the weights. When people open source, you can download. And if it's built with their software, then you can easily spin it up into a little demo environment. This is what they called the inference API for a long time, just kind of real simple. Hey, if a model is here, we can load it up for you. You can ping it a few times. Then they have the spaces, which is where the demos come in and they've acquired a company called Gradio about a year ago that makes it easy to do these demos. And so they host those as well. And if you buy a pro account, you can get bigger servers. It's a little bit Replit in that sense where you have your own little coding environment that you can kind of create the demo on top of the model, because the model itself often needs a wrapper of what are the inputs, what are the outputs, how's that all going to look? And so they're kind of building this community first approach. This is where people come to share. This is where people show off their demos. This is where people can fork and recombine models. Whole datasets are published as well. It's all organized across all these different tasks. So if you want to go look at what's, show me everything that captions images versus show me everything that is text to speech versus show me everything that is speech transcription. You can kind of see all of these different lenses on everything that is going on in AI. And this has all kind of been built. I think the company and the product in some ways kind of reflect each other. It's a remote first company with people kind of spread all over, very broad set of expertise. And all these kinds of experiences have kind of been built up over time as well. The team is really strong, right? So you get them with all that stuff going on and with all the kind of different ways you end up getting linked to Hugging Face. I was thinking, I know there's a lot going on here. I feel like I want to make sure I'm making the most of this platform, even if that just means being able to clearly figure out what is the state of the art that has been open sourced at any given time. And for that, it was worth it to me to become a customer of the Expert Acceleration Program. And from that, I really learned, yeah, they do have a really awesome team that has a super diverse set of different backgrounds and expertise. They actually were very useful to me, especially in those kind of early AI obsessive days where I was most focused on just making a product work well, building an app. I'm thinking, I need the best image caption that I can get. Fast forward to the end of this research question and tell me what the answer is. Can you help me with that? Now, it's not always super easy, right? There's a lot of, as we've covered in many different ways, there's a lot of surface area. There's a lot of trade offs. It depends exactly on your use case. So it's not always to say that they could provide definitive answers, but I did come away from that engagement super impressed with the team and feeling, yeah, there's a lot of expertise here and you can kind of see why they are a center of activity. And I think they're still honestly kind of figuring out how do they actually turn this into a business? There's so much activity, but this ratio of the 4,500,000,000 to the 30 some whatever million in revenue is super high. And it does, I think, reflect that there's a lot more latent value than they have been able to monetize so far. As I understand it now, it seems the biggest way that I think they're looking to monetize is by access to compute. They want to kind of say, we've got all these different things, whether it's models hosted or datasets that you can run training on or increasingly more and more different things. What do they all have in common? They need compute to run. And so in addition to the way that you find all this stuff will also be the way that you run all this stuff. And that's where products like the inference endpoints come in. And some of these numbers too, the download numbers on these are kind of amazing. Sometimes you look at just a random model where the actual weights have been trained and now uploaded and open sourced on the platform and how many people are downloading these things. And some of them, just right now browsing around the website, 35,000 downloads for some particular fine tuning of CodeLama, 34B. I mean, that's pretty niche in some respects, and it has 35,000 downloads in the last week. So there really is a lot of people coming here to discover stuff, and then they want to just take that next step to say, okay, now you want to just serve that in an application, click a couple buttons here, and we will handle everything from there. And again, the fact that most of these models have been built on their software library allows them to spin that up relatively easily. But it's still kind of a jungle. And I think they're very much still working on figuring out exactly how to make that awesome for people. 1 of the complexities, of course, is there are literally thousands, if not tens, there are tens of thousands of models. I don't know how many are really actively used these days, probably hundreds that are doing any significant inference endpoint work. But still, that's a lot. And it's very hard to optimize across all those different possibilities and all the different workloads that customers bring to the table and so on. So it is very interesting because it is 1 where community drives so much. The promise is there. The product set is definitely still in the kind of phase of maturing, but that has not stopped them from raising a handsome amount of money.

Erik Torenberg: (15:50) Yeah. That's a great overview. Hey, we'll continue our interview in a moment after a word from our sponsors. Let's get into communities for a bit because communities on their own, they're not defensible. If you look at Digg, remember Digg? Digg, pre-Reddit, Digg was 1 of the most valuable communities on the Internet, the front page of the Internet where so many people went to see what was happening. And then the challenge with communities is sometimes what makes them special is their exclusiveness. It's for a certain kind of person and it's not for another kind of person. And what often happens is as communities scale, you sometimes have negative network effects where the bigger the community gets, the less special it gets. Think of any sort of nightclub or party or any sort of group of people.

Nathan Labenz: (16:38) Clubhouse. Erik Torenberg: 16:38 Clubhouse.

Nathan Labenz: 16:39 Exactly. Where there's a special vibe to it. As soon as everybody comes in the party, it's no longer as fun of a party anymore. Now, different communities get at that at different ways. The bigger it is, you have to be getting some utility to the community. People always ask me how big should we make our community? And I would say basically, I mean, it depends on the goal of the community, but often every additional member should be making the community stronger. And so it's less about size and it's more about solving that specific question or the community goal stronger. And often, if the goal of the community is the relationship between its members, smaller is typically better. But if the goal of the community is to aggregate some sort of data or insight, the more can be better. But just to close the Dig example, there was a bit of a rebellion among its users, and people were mad about something, and then they just switched to Reddit almost overnight. And Reddit became the new Dig, and Reddit has become a massive company ever since. And what that just showed is that community on its own is very difficult. You in order to make communities defensible, you have to get the community to do something on the website that it would be hard to get them to switch. Right? On LinkedIn, for example, you have all your connections, you've built up that connections over years, you have your identity, you have all these recommendations, you have these references. People could hate LinkedIn, but there's not going to be a rebellion onto a new platform because the social cost and the coordination cost is just way too high. With Dig, it's just a Reddit. You just create a username. That's it. That's all you do on the site. You have your post. I don't mean to downplay it, but with social networks, LinkedIn, Facebook, we'll get into GitHub in a second, you've created state on the website. You've input data, you've created connections such that the switching cost and coordination cost is too high to leave, and that's why people say LinkedIn is this incredible business. Everyone's trying to disrupt LinkedIn. They can't. They've been trying for 15 years. They can't, and people don't even like LinkedIn. The users don't even like it. They're like, I hate this website, and yet they're on it every day. They can't leave it because, and so that's an example of a company that started as a community, but built defensibility via creating places to own identity and connections. And I've been on both sides of that. I mean, I was at think of things like Product Hunt and Hacker News. It's not just defensibility. You also have to create a way for value to be exchanged. What LinkedIn does is it not only brings people back every day, you get information on LinkedIn that you can't really get anywhere else. People's detailed professional history, what people think about each other, who's viewing your profile, that's pretty valuable and ability to get in touch with people ultimately. What GitHub does is it and maybe this is where it gets a Hugging Face, is it's not just a community of engineers, it's basically a vertical LinkedIn for engineers in terms of people aren't just coming back. They're inputting all this valuable data that makes the website valuable even independent of people coming back every day. That is something that sites like Hacker News or Product Hunt have not done. Hacker News and Product Hunt have to reinvent the wheel every single day, kind of like Dig. Now they've maybe done a better job on community or branding where people aren't throwing rebellion on them, and there's some moat. People are coming back every day, but these are not big businesses. They're not venture scale businesses because they haven't really figured out they haven't created a ton of identity or state on the site itself. And every day it's kind of a rat race to create this value, but there's no compounding value. The way to do that, and we tried this and they're actually trying it right now. God bless them. Product Hunt could have turned into a G2. For people who don't know, almost like a Yelp for products. It would have just been not what's great today, but almost Wirecutter style. Here's a review of products historically in a way that would have people coming back to look not just at today, but historical. And so I think that's the big challenge if you're a community business is one, what can you use the community as a wedge to create? What data set can they create that will make the website valuable independent of people coming back right away because that will also make it difficult for those people to leave. So what do you think that could be for Hugging Face or any reactions to that analysis?

Erik Torenberg: 21:05 It is still an odd mix. It's not obvious that discovery and inference naturally go together. If the model is basically this, and I think this is kind of the core thesis right now seems to be pretty apparent. All the demos are going to be on here and you're going to find the models on here that power those demos. And you're gonna spin up your own on our infrastructure, and then it'll be convenient, and you'll pay us for that. And we'll make a margin over the cloud providers that they use. And just based on the investors here, Google, Amazon, NVIDIA, Intel, AMD, Qualcomm, AMD, IBM, Salesforce, Sound Ventures. I mean, obviously a lot of cloud infrastructure there. So they've got the right partners and they can presumably make some margin. That product has this thing on its own. It has to be excellent, because it's not that hard to take that model because, again, the Hugging Face premise is that the model itself is open and you can just download it and you can go do your own thing. You could take that direct to Amazon. You could take it direct to Google Cloud. Notably, not on there is Microsoft. You could take it to Microsoft as well. All these demos are also open source. The sort of code sharing on Hugging Face is very much, I mean, it's basically the same thing as Git itself, except that they lean more into providing a running environment. When you go to GitHub, you look at a repository. This is starting to change with code spaces where you can say, okay, I want to not just look at this code, but run this code in a way that's not going to take me a ton of setup on my local computer or whatever else to actually get started. They're making progress on that. But I would say Hugging Face is definitely ahead in that the demo is there. You can kind of run it. You can fork it into your own running environment. It's a little bit more replete in that respect. But the competition is pretty fierce from all these different angles because the code is open source. The model is open source. If I think that their inference endpoint is too expensive, I may love the community. I may love coming here and prototyping stuff. I may be willing to pay my I think it's $9 a month or whatever for the pro account that gets me the slightly souped up demo environment so I can do my prototyping more effectively. But if that inference product isn't awesome unto itself, then I'm going to make that decision pretty independently, I think, in most cases. Certainly, we talked about this a little bit in the last analysis episode too where what makes the OpenAI 3.5 fine tuning so awesome in addition to the fact that the model is great and it's easy to use and the fine tuning is quick and cheap, which fine tuning in general is getting pretty quick and cheap. But it's basically infinitely scalable and immediately available as soon as you do it. So if I do a fine tuning and want to make a hundred calls to it immediately and then never call it again, they allow me to do that, and it's so easy. The inference endpoint in the Hugging Face case right now is not as mature as that. It does not scale up. They do have auto scaling, but it's a little rougher. It's just not as mature. You kinda have to have one thing running all the time and then you can go turn it off. But it's not so immediately turn on, immediately available, just totally priced and structured to your amazing convenience like the OpenAI one is. And so if there are things that I don't like about that or I just don't like the margin that they're taking in, if I'm really hitting some scale and it's like it's time to actually look for cost savings, then I think the lock in that the community creates to the inference product is actually very small. It's more of a lead gen type function for an inference product. But if you're a sophisticated app developer, there's really nothing preventing you from taking all that stuff and going somewhere else. And I think there's also some interesting challenges that they're likely to face too that you alluded to with the Dig example, which is that they've kind of developed a community that you can't characterize. The Hugging Face community is huge. I mean, it is a big community. They've really closely tied the company's identity to open source and open source kind of being an inherent good. It honestly is kind of ideological at times. The company's positioning is kind of ideological. The community that is most excited about it is kind of ideological. It's definitely in opposition to more AI safety hawk type people who are, open sourcing this stuff is not necessarily a good idea. Now, I mean, typically, people, in fairness to them, are not so worried about the current generation of models being open source. But just the general pattern of making stuff, releasing it before we even know what it does, especially as we continue to scale stuff up, does seem like there's gonna be some real unexpected consequences from that. The Hugging Face discourse doesn't allow for a ton of recognition of that because their whole kind of thing is access, a sort of egalitarianism. It goes so far as to reflect in a leaderboard product that they have. And again, this goes back to I tell them, curation has been part of your value to me. I wanna hear what you think about curation too because that also seems like a tough business certainly to create lock into the high dollar products with. The leaderboard that they have, they have an open LM leaderboard. And so it's not like it's fair enough. You call it the open language model leaderboard, but it's not useful to me because they don't show the best models. You literally go to the Hugging Face language model leaderboard, GPT-4 is not there, and 3.5 is not there. Nothing from Anthropic is there. I think basically even though Llama wasn't fully commercially whatever initially, I think they kind of round that into inclusion. But it's like, if you want to make a leaderboard that shows me what the best models are or how the best open source models compared to the truly best models, you got to show me how good the actual best commercial models are. And the reality that they are beating the open source models is not one that, again, sophisticated customers are going to fail to realize. So at times, it feels like there is a sort of ideological kind of nature to some aspects of the company and some potential at least. I wouldn't say this exists, but certainly the potential for a certain kind of audience capture where if they ever want to make certain moves to you could imagine they might say, well, we're gonna train our own models and they won't be open source, then we'll serve those. I think they're gonna have a very hard time making those kinds of deals. There's any number of licensing deals they could do where folks might publish models on their platform and then have some sort of rev share depending on how much it gets used. And we're in very early days, I think, of business model development here. But the fact that there's such a hard commitment to open source so early does make, in my mind, for kind of a hard strategic dynamic in some cases. Because people are used to just everything on Hugging Face is supposed to be open and free. That's the model. Increasingly, that's the datasets. It's definitely the demo code that runs it. And all that freedom is awesome. In the end, the inference product still has to just earn its keep. It doesn't get that much, I don't think, of an advantage for all those nice public goods that Hugging Face creates. If you don't win head to head, that lead gen is not worth potentially that much.

Nathan Labenz: 29:13 Curation is interesting, right, because Product Hunt curated the best products every single day, and that's what journalists do. And it turns out that's just not a massive business. It's important. It's a business, but it's not a massive business. It's not a repeat. It's something that you have to start over every day, and the value there just isn't that high. It's not like there are core business problems around what's the hot thing today in a way that they can easily monetize product on. Now, with something like G2 or there are other sort of companies where they're basically reviews of enterprise products, these are decisions that have a big spend to them. And all people who are building companies have to use these products, they have to determine which are the best one. And they go to a site like G2 that helps them determine that. That's something they're willing to pay for. And that's an enterprise version, but a consumer version is something in a different way, something like Glassdoor, where people have to determine which companies to work for, and Glassdoor is a review site of companies told from the employee perspective, not just how they're doing, but what it's like to work there. Another example, consumer example, obviously, is Yelp. Right? People need to decide where to get dinner tonight, and Yelp helps us determine that. And that not just restaurants, but for all different types of services. And so what's cool about those examples is that you don't have to start over every day in the same way that you do with Hacker News or Product Hunt, so there's some compounding value to that, but then also they have real spend to them, especially in the enterprise version. It's also it's not that you want them to be evergreen forever, there's no turnover. If there is high turnover and high spend, that's also valuable because you need to continuously get new reviews. The thing I would ask a Hugging Face is okay, is there a firm of your creation that has a high spend to it that has some turnover but isn't reset every single day that you can add compounding, that you can get compounding value from continuing to have reviews and thus people willing to pay or businesses willing to pay Hugging Face in some capacity. Nathan Labenz: 29:13 Curation is interesting, right? Because Product Hunt curated the best products every single day, and that's what journalists do. And it turns out that's just not a massive business. It's important. It's a size business, but it's not a massive business. It's not a repeat. It's something that you have to start over every day, and the value there just isn't that high. It's not like there are core business problems around what's the hot thing today in a way that they can easily monetize. Product Hunt, with something like G2 or there are other sort of companies where they're basically reviews of enterprise products, these are decisions that have a big spend to them. And all people who are building companies have to use these products. They have to determine which are the best ones. And they go to a site like G2 that helps them determine that. There's something they're willing to pay for. And that's an enterprise version, but a consumer version is something in a different way, something like Glassdoor, where people have to determine which companies to work for, and Glassdoor is a review site of companies told from the employee perspective, not just how they're doing, but what it's like to work there. Another example, consumer example obviously, is Yelp. Right? People need to decide where to get dinner tonight, and Yelp helps us determine that. And that's not just restaurants, but for all different types of services. And so what's cool about those examples is that you don't have to start over every day in the same way that you do with Hacker News or Product Hunt, so there's some compounding value to that. But then also they have real spend to them, especially in the enterprise version. It's also not that you want them to be evergreen forever. There's no turnover. If there is high turnover and high spend, that's also valuable because you need to continuously get new reviews. The thing I would ask Hugging Face is, okay, is there a forum creation that has a high spend to it that has some turnover but isn't reset every single day that you can add compounding, that you can get compounding value from continuing to have reviews and thus people willing to pay or businesses willing to pay Hugging Face in some capacity.

Erik Torenberg: 31:29 Yeah. Certainly, the change is happening fast enough that you do need to keep the content fresh. I don't know if that works necessarily for or against the kind of staying power of the business because there's always something more to discover. On the other hand, if it ever stops becoming the place where the best kind of discovery happens for a month, the prior content on their ages pretty quickly. And nobody's really at this point concerned. I mean, as many models as they have on there, right? And it's, again, probably in the tens of thousands, if not hundreds of thousands maybe of raw models that have been uploaded. Any of those that are more than 8 months old at this point, anything before ChatGPT, who cares? Right? It's all kind of dated. So that'll continue to happen presumably. They have to, I think, maintain that kind of focal status above all else or everything else seems to be pretty quickly at risk. The analogy to some of these other businesses you're describing is tough because, for one thing, none of those are super huge businesses. Right? I don't think any of G2, Yelp, I mean, what's Yelp's market cap right now? It's got to be lower than Hugging Face's. Right? It's basically an advertising model at the end of the day. Right? Because you go here and there's all this credible content, then it kind of creates promotional opportunity alongside that content. And that's not the Hugging Face model as it stands today. Right? I don't know what the analogy would be. If you took G2, it would be almost as if it was just reviews of only open source software projects for whatever. And then at the end of the thing, can either host it on their site or you could go download it and run it on your own server. That would be very certainly very analogous. I don't know that there's the space to monetize Hugging Face in that way or the scale because there's definitely always room for some advertising. No doubt. But

Nathan Labenz: 33:26 Let's just dream big here. If Hugging Face is a $50,000,000,000 company someday, what is sort of the biggest version of what Hugging Face could turn into?

Nathan Labenz: 33:26 Let's just dream big here. If Hugging Face is a $50 billion company someday, why could it be? What is sort of the biggest version of what Hugging Face could turn into?

Erik Torenberg: 33:37 GitHub for AI is kind of the elevator pitch by analogy for Hugging Face. Obviously, GitHub was worth a lot of money, and I think Microsoft doesn't regret buying it at all. So they've done well there. I would guess that the ratio of the GitHub acquisition price to its revenue was also extremely high because at least the things that I have usually paid for and I think that most developers pay for on GitHub is a handful of dollars a month. And the main thing it allows you to do is keep your code private. You can use GitHub for free, but you gotta publish all the code. If you're gonna have your private stuff, you gotta pay a few bucks. Hugging Face has a similar thing that's $9 and you can kind of have your private spaces and a little bit more horsepower behind your public demos or whatever. So so far so good. What is Microsoft really valuing in that? I think they're above all just valuing they have hooks into tens of millions of developers around the world in some way, shape, or form. And we know that that's a very valuable profile and hopefully we can sell a lot more stuff to them later. And you're starting to see that product develop with Codespaces and with Copilot. And presumably, will continue to mature, but they honestly didn't really need ROI. Right? It's a little bit more of a defensive arguably position from them where they're just like, we're a $2 trillion company. Let's go pick this thing up and we'll make sure that we continue to be super focal and almost inescapable, unavoidable from a developer standpoint. Hugging Face could maybe pull off a similar thing, I think. If it were to sell for $20 billion in 2 years, it would seem or in 3 years or whatever, it would seem the most likely scenario would still be that the revenue is very much kind of trailing the value. And it would be just a broader sense that, hey, ML is the next programming in general. It's so massive and the people that do it are so important. And increasingly, so many decisions are made at that level that owning that platform becomes a strategic priority for somebody. You look back at the list of investors here and Google and Amazon right off the bat are kind of like, yeah, I can see why given that they don't own GitHub and they want to compete with Microsoft in everything, I can see how an exit to one of those companies would be natural where they would say, we don't really care how many people you're charging $9 a month. And we don't even really care if your inference product is truly top notch. We want a community layer to our mega stack. And that's kind of that. If you're buying in at this valuation, I find it hard to see a different story aside from that right now. I mean, I guess the other candidates would be they go replant and just and by replant, I just mean world class execution, project after project. And Hugging Face does well on stuff, but it's not quite that sort of thing. The community nature of it all comes a lot more messiness and they don't have the same kind of super clear, super visionary product vision as much. They're more like, oh my god, we have all this activity, we have all this sharing, we have all this discovery, we have all this energy. What do these people need? Then we can kind of figure out how we back into offering them something. It doesn't seem like they're going to take a huge bite out of inference in the same way that I think a Replit, because it's so different and because it works so well, seems like it's a lot closer to taking a real bite out of just cloud in general compared to what I've seen from Hugging Face. And I love the vision, by the way, of the inference endpoints. I'm actually on full disclosure. I've aside from being a customer of the company, I have no financial stake or formal relationship whatsoever. But I did participate in the inference endpoint beta. And I do love the vision from a customer standpoint of being able to say I always tell them the task page is where I think you guys should really make your front door. Because if I have a task, if I'm that far along, then you can show me the best options to solve that task. A task would be, again, image captioning, text to speech, transcription from speech to text, whatever. They have probably 20 different tasks for which they have kind of something like a leaderboard. And the leaderboard doesn't even need to be against a benchmark. A good leaderboard these days would have a number of data sets that all these things are benchmarked against and you could see relative performance quantitatively. They don't necessarily have that for all of those. They don't need to have it for all of them. They can get by to a significant degree with the likes and the number of downloads and just the kind of community commentary can be pretty clarifying. So I do love the idea of just going to a task page, looking at the first 5 things being like, okay, here's the 5 coolest demos that are out there to solve this problem for me right now. Let me go mess with each one. They're all kind of running. They're ready to go. Now I can I mean, because this is a huge contrast to what came before. Right? I mean, to give the credit where it's due, the alternative to a Hugging Face demo, and I'm old enough to remember, I mean, open source has made incredible progress into the ML community in terms of just what the norms are. If you go back a few years, pretty normal would be we publish a paper, we share our results, we maybe publish a dataset, we maybe give you a little script that you could use. But it was more of kind of a reproducibility notion. The reason they were publishing code out of a lot of academic environments was we want you to be able to kind of reproduce our results and therefore, by putting that out there so you, in theory, can, we demonstrate that we actually did the work and that the model can actually perform we say it's gonna perform. But very few of those things were ever intended to be broadly used more than a couple of years ago. So it was often through this kind of, we're giving you the tools to reproduce and validate that our work is real lens, not like we're trying to set you up to actually go use it in a real practical setting somewhere. So that has changed dramatically. Code is now very common. Actual trained models is increasingly very common. People on kind of Hugging Face part of Twitter are kind of like weights or didn't happen. And even demos now, especially, again, they acquired this company, Gradio, which makes the demo creation process really easy. So it is incredible how much again, just in the last 18 months, it has gone from, okay. Here's an interesting paper. They're not refusing to allow me to reproduce it. They're not keeping secrets, but the way that they've made it available to me is hard to evaluate, hard to just hard to get over the hump to even see is this something that could work for me in whatever context I'm working in. With the demos on Hugging Face, that is so much easier. So you can fast forward instead of it might have taken you a week before to be like, okay. Paper, code. I have to maybe retrain the model. I gotta figure out how to do this. What libraries does it use? Oh my god. Blah blah blah blah blah. All just to get to the point where you could be like, yeah. They did it all in their academic setting, but I have some real world inputs. Does it work for me or not? A lot of times, it doesn't necessarily. So that's a lot to put in. Hugging Face today, you can get that down to 2 hours. You can go to the task page, try the demos, run a little script, try a handful of things against a handful of different demos, see if anything's really compelling, and then pretty quickly move over to your inference API as well. And that is super smooth and definitely great value to the community. Again, then the question just becomes if I go through all that and I'm really hitting scale, what lock in is there to keep me buying from Hugging Face? It either has to be the best on a pure performance value basis kind of independent of all that discovery or it's pretty easy for me to switch. So, yeah, don't know if a curation it is super valuable, but how do you it seems hard to capture that value. And then you can imagine that they could have a sort of more marketplace type of thing. Again, think about a G2. It's like people can pay to advertise there because they're offering proprietary stuff and they can get an ROI back. Nobody's really gonna advertise their fully open source model. So at some point, they may if they wanted to go a route where they might say, we wanna help you discover both open source and commercial stuff, and we wanna monetize the commercial stuff, that would make a lot of sense to me in some ways. I think it honestly would be more valuable to users. It'd be more valuable to me if Hugging Face would put the commercial models on their LM leaderboard and kind of do that across the board for all the different tasks. I may care depending on the circumstance. Right? But I'm definitely open to commercial things that are not open source. I'm interested in open source as well. Commercial has a certain appeal because I can generally assume that they're gonna handle more problems for me. Hugging Face tries to bridge that themselves by saying, we'll handle a lot of the problems for you of doing the open source stuff, but I want the complete set. It seems like to me the curation value right now is a little bit undermined by certain commitments and kind of expectations around commitment to open source where as an app developer, I just want clarity. And I want the full menu of all the models, how well they perform, ideally what they cost, what the rate limits are going to look like. If I want that whole thing and by kind of pretending that a certain amount of it doesn't exist in the way that they present stuff, that's definitely less value than I could be getting. So it's interesting to me that it feels like certain strategies I don't know if they're closed off, but it feels like they would have a hard time pushing back against the community. And there's again, there's going to be fearsome competition from every direction. Right? If they take one misstep, GitHub is right there to say, you know what? Guess what? We got a live demo thing. And who has plenty of disk space for all those models? And honestly, they might even be able to just mirror it directly. I mean, some of this stuff is the legal regime at all. This is very much uncertain. But going back to LinkedIn, right, there's been some court rulings that you can scrape LinkedIn and it's not illegal for you to do so. This information is kind of out there. And again, there's some fine points around that. But, basically, you can scrape LinkedIn. Something similar might be the case with Hugging Face where it might be you can download all the models from Hugging Face. They can't really prevent you from doing that. And if all of sudden, GitHub mirrors all of Hugging Face and has a compelling thing and also has commercial models that are kind of on that menu, I could see I could see some ways that it could become pretty uncomfortable for them to try to keep everybody happy. How do we keep that open source ethos, but also have the full thing and a lot of stakeholders. It's not easy to manage such a diverse, many sided marketplace as they're kind of trying to develop.

Nathan Labenz: 45:50 I wanna return to something you mentioned, which is your articulation of the Replit strategy because, yeah, some people believe that Replit is also way ahead of their revenue and that they will eventually sell to Google or one of these major players at a really nice markup. But I'm curious what you think the path is for them to be an independent massive company. What could that look like? Nathan Labenz: 45:50 I want to return to something you mentioned, which is your articulation of the Replit strategy because, yeah, some people believe that Replit is also way ahead of their revenue and that they will eventually sell to Google or one of these major players at a really nice markup. But I'm curious what you think the path is for them to be an independent massive company. What could that look like?

Erik Torenberg: 46:12 Just continued world class execution at everything. I mean, it's that which is not easy. I wouldn't put that strategy forward for anyone who hadn't already demonstrated strong ability to actually make that work. Because that's like, I couldn't really go recommend that to Hugging Face. What you guys need to do is just be the absolute best in all of your product execution always. Tough. And tough for Replit probably to sustain as well. But at least in using that product, I do feel like it is on the right track. How high can they go up in terms of real scale? And I'm trying to post some pretty impressive little demos and tests over the last couple of days just around auto scaling, his own personal website. What happens if I send a quarter million requests at this Replit hosted site all at the same time? Does it get overloaded? Can it handle it? I mean, obviously, I'm sure there's some limit that he hasn't where he may have hit a limit and didn't share that number. But at some pretty impressive numbers, he's showing stuff that works amazingly well. People in the comments are just, dude, that is sick. That level of execution, is it sustainable? I mean, again, it's going to be hard to sustain, but they do make things so easy that I think where they can really crush it is in the sort of long tail of apps. I doubt that they get to the point where major engineering projects that are high scale, certainly that already exist, would migrate to Replit. What I think they can do is create something that this is the best way to create, which I think they're already right there at the top, with something like Cursor and Copilot. They have rivals, but they're right there at the top of the best easiest way to create. Then they're also the easiest way to just turn that into a live app by just hitting publish, amazing. How much business they can retain as those bazillions of small apps grow and mature is probably the biggest question for how big that business can really grow. Because if you get started on a platform and you don't have any pain points, and the markup isn't that huge, you're probably not switching from the platform, especially if it would take a lot of work to switch into something else that's not nearly as user friendly. So how far up that curve can they go where they're still, hey. I'm not hitting any problems. Cost seems pretty reasonable, and I'm just happy to keep growing with Replit. I would say that's going to be the key question. They have this deployments product set that looks super promising for that and seems like it's already probably pretty well there for kind of small to medium sized apps. What will be really interesting to see is if people start to and most, surely start to happen where what happens if you start to scale to millions of users on Replit? Are they ready for that? Is the cost profile such that they can keep you as a customer? If so, sky's the limit for them. If they can't handle that, then that puts a cap on how big the business can get, I would think. And that's probably the number one thing that I imagine that they're trying to figure out how to do. They've got the community. They've got people building small stuff all the time. Can they get people to build more ambitious stuff and can they keep those things as they really hit tipping points into millions of users? I think they'll be in great shape. Yeah. They also have a whole other thing too, which is the dynamic. I mean, that's conventional analysis. Then the next kind of mind blowing analysis with Replit is they're working on the AI developer. Again, it's a much more integrated product than anything you'll see from Hugging Face. Hugging Face is not going to give you a single copilot-like experience for all of their community. I don't see that on the horizon at all. I would literally be shocked if all of a sudden that drops. They have some of this stuff where it's talk to Hugging Face. It will help you figure out what the right models are for your use case. But just, again, it's just so diffused. It's so kind of everything's changing all the time. They don't have the same focus on stuff like that. It's all kind of more again, we talked about with Llama 2 last time. Right? They train it. They get to some published point. They release it, but they didn't really beat it up iteratively over and over and over again. That's kind of the vibe that, even though most kind of meta conceptually interesting projects from Hugging Face seem to look like so far. Whereas with Replit, they're just absolutely lasered in on, we want to make an AI developer. We want it to be able to do stuff for you. We want to be able to do more. We're going to chip away at that. You're going to have this actual pair of programmer for you. Then they've got this whole vision of around a dynamic economy. They've got the bounties. The bounties can be potentially served in some cases by their AI developers in the future. So there's a whole kind of I think they are set up to try to execute an economy on platform in a way that Hugging Face is really not right now. They have this kind of users can pay users to do stuff. Users can actually pay AIs to do stuff. Users could pay humans, which, sub delegate to AIs to do stuff. They have visions of AIs delegating to AIs to do stuff, which is where it really starts to get crazy. But it is all kind of coherent with this idea of we want to be the easiest way for you to build and deploy stuff. And, we'll make whatever kind of dynamism or commercial model underneath that is needed to support it. They're not as ideological about any certain commitments. And I think that does help them drive toward that execution. Not to say that's all it takes. Obviously, it takes a lot to execute at a super high level. That's but at least one thing that I see kind of getting in the way for Hugging Face that is not there for Replit as much. Another interesting story too. They Hugging Face, it has been a while, but they had a Bloom, an open science project where they created 175 billion parameter models kind of analogous to GPT-3. In some ways, it seemed like the point was to prove that an open source community could do this and do it in an open way. And they did. But again, it doesn't really work that well. Even at the time, it was not nearly as good as GPT-3. And it was kind of, what is it that has been missing from this process? It was, we brought a lot of people together. We assembled a huge dataset. We agreed on a bunch of decisions. Honestly, pretty phenomenal achievement in terms of just demonstrating that you can even pull a project like that by kind of an open consensus. So super cool, but didn't really pack the ultimate punch that its rivals did at the time. And so mostly it's kind of a footnote of they did this thing, it was interesting how they did it, but they didn't really prioritize performance and they didn't really iterate. And so it kind of became an artifact that doesn't really have a lot of usage today. In contrast, when Replit's doing their own code models, they're fine tuning those models on their own dataset, exactly how their platform is used. And it's just a lot more integrated tightly kind of coupled, which I think drives the performance that I'm talking about.

Nathan Labenz: 53:51 I thought this was a great overview in terms of really assessing what makes Hugging Face interesting and what are some of the key questions and risks looking forward in terms of as it tries to grow into its valuation or perhaps more likely, find the right buyer at the right price if it's less likely to be an independent company.

Erik Torenberg: 54:12 Yeah. It'll be fascinating to find out. I think one just looking back at our perplexity Q and A thread that we used to help collect some of the facts for this and some of the comparables, I think it is interesting also to contrast against companies that have also generated a lot of buzz. So something like a Runway, something like a Character, these are products that people are wowed by. My teammates at Waymark are creating the Frost 2 with wherever the latest model is from Runway. And they rave about it. It allows them to do super cool stuff that they couldn't do previously. We just couldn't make a film like this with the resources that we have without this frontier model. Similarly with Character, people just love the experience. They love going there. The creativity of all these different things that you can spin up, it's an awesome, very novel thing. It definitely has a certain kind of product market fit. Those companies, whatever they're valued at a billion-ish or something. I do think their position is in some ways even harder to defend than the Hugging Face position. If I was purely interested in financial return, I think I would for all the difficulties that we've covered, I think I would bet more on Hugging Face because it does have the community and that is undeniably super valuable and discovery is valuable and being a trusted source for curation is valuable. In contrast, a Runway and a Character, you have to bet that they continue to be at the frontier of models right now because that's the thing. And what happens if somebody comes out tomorrow with a strictly better model that just does a better job of generating video clips than Runway's latest model? You could see a dramatic drop in usage. If something really is strictly better, the switching costs are almost zero. You look back at what was the state of the art in image generation a year ago? Well, we just passed the one year anniversary of the original Stable Diffusion being released open source. And so that was pretty much it at the time. It was DALL-E and it was that. And obviously, we've had an insane flourishing of all sorts of different things since then. But, nobody's really using that original Stable Diffusion. It's been through 10 generations since then, and it's we've had multiple handoffs of leapfrogs of, oh,

Nathan Labenz: 56:49 look,

Erik Torenberg: 56:49 you know, DALL-E, there's another DALL-E version that's improved and Adobe, Firefly is a thing. And Midjourney continues to do a phenomenal job on their model, and has proven that they can maintain their position on the frontier. But again, for all these companies, the switching costs are extremely low. The stickiness is extremely low and the sort of month, maybe not the minute, but the month that Adobe comes out with a strictly better thing versus the Runway is so easy to switch. All the Runway customers are already Adobe customers. They all already have Adobe licenses. So credit card is depending on what the business model is, they might just start to get it for free as part of their existing purchase. Maybe it becomes an add on. But again, the credit card is already on file. So the stickiness there is super low. I think in general, this whole discussion does kind of reflect that. It's still not really that obvious where in the AI stack massive value is going to accrue and, the sort of gravity of the incumbents, the infrastructure owners, the ones that the Hugging Faces and the Runways are built on, it seems like those end up capturing a lot of value here. And it's got to be a stressful job right now to be the model team at Runway. You're, man, we've had a couple hits in a row and we've run that up to a billion dollar plus valuation and global minor fame, certainly in the AI world, global fame. But that other competitor could drop anytime. What's our timeline to get our next one out? How do we how do how does it go if we get leapfrogged even for a little while? That's a tough position to defend. And so, yeah, I would the relative valuation in Hugging Face's favor, I don't think is crazy by any means. It's just that, even in their position where they do have so much going for them as, a real hub to use their word, it still does seem like a very difficult hand to play strategically.

Nathan Labenz: 59:06 Totally. That's well put. Let's wrap on that. It's a great episode, Nathan, and, until next time.

Erik Torenberg: 59:11 Alright. Cool. Thanks, Eric. Always a pleasure. It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.