20VC: Why The Future of AI Is Open Not Closed, Why We Are Years Away From AI Being Autonomous, Why AI Founders Do Not Need to Move to the Valley & Why Founders Should Not Meet Investors in Between Rounds with Clem Delangue @ Hugging Face

0:00The DC and the mainstream interest is like a catch up on the reality. We're very far from a world where AI is autonomous and has conscience and is taking over the world and destroying humanity. But I don't talk to any external investors in between routes. Welcome back to 20vc with me Harry Steevings and our series with the best founders and and the one on the Jan Lakoon is such a special show. Wednesday we have Amadot's stability and today we're joined by Clem Delaing, co -founder and CEO at Huggingface, the AI community building the future. To date, Clem has raised over $160 million from the likes of Sequoia, Cotu, Edition and Lux Capital to name a few.

0:47And I want to say a huge thank you to Lee Fixle, Pat Grady, Brandon Reeves and Teebo else here. Some amazing questions, suggestions today. I really did so appreciate that. But before we dive into the show, Dave, let me talk about how Coda is the doc that brings it all together and how it can help your team run smoother and be more efficient. I know this because Coda helps me. I often have to scale my learning on a new space very quickly, so I speak to many people. Then I need one central location to bring all those nodes together. My team then add to those nodes and make amazing additions. And by putting all the data in one centralized location, regardless of format, it eliminates so many roadblocks that can stop your team in their tracks.

1:25And this is what really slows down productivity and collaboration, when there's siloed information in different tools. And so with Coda, your team can operate on the same information and collaborate in one place. Help your team run more smoothly, more efficiently with Coda. Get started today for free. Head over to coder .io slash 20vc, that's coda .io slash 20vc, and get started for free. And some EU tools we cannot live without, Angel List is fast becoming the centre of the venture ecosystem. So for startups, Angel List reduces the friction of capital management, banking and fundraising all in one place.

2:00Teams can focus on scaling and let Angel List handle the rest. Thousands of startups have moved their cap tables to Angel List in the past year. Angel List also supports large venture funds and their teams, with an automated software first approach and the best customer service in the industry. Fun Managers can focus on making great deals, while Angelist handles reporting, taxes, compliance and more. What's more, with the recent release of Angelist Network banking, for fun managers and investors, your deposits are secure, with the most trusted banks for maximized FDIC coverage and mitigated single -bank risk.

2:32If you're ready to scale, you'll start up or fund with the platform of the centre of it, visit angelist .com forward slash 20VC to get started, and finally, Brex. Since his founding, Brex has been committed to helping start up small and scale faster at every stage of growth from MVP to IPO. Today, Bryce is all in one financial stack, is used by one in four US startups and counting. I get to speed to founders all day and I know how crucial it is for them to have the right financial stack. Bryce gives you fast access to a high yield business account where you can safely store and move your cash while getting up to 6 million in FDIC protection.

3:09Lately, it's been all too clear how important that is. Plus, you get high -limited corporate cards, easy, expensive tracking and automated bill pay. To learn more about the all -in -one financial start for startups, visit brex .com -4 -20VC. That's B -R -E -X .com -2 -0 -V -C.

3:40Clem, I am very excited for this. I've stalked the shit out of you from Leafixle, Pat Grady, Oliad Datadog, Devamongo, Tibo Elsie, who told me about the very early days, so thank you so much for joining me today. Thanks so much for having me, I'm excited about this will be great so I want to start with a little bit of context hugging face where did the name come from what's the origin of the company founding in a short two to three minutes yeah when we started hugging face we joked we Michael founders Julia and Thomas so we wanted to be the first company to go public with an emoji versus of the three -letter ticker you know we felt like the three -letter ticker like on the NASDAQ and all that is scoring felt like It was time for refresh and to finally have emojis up there on the boards.

4:25So we absolutely wanted an emoji as a name. And the choice is the hugging face emoji, the one with hands like that was our favorite emoji. So we're like, okay, let's do that. Without maybe we would keep it for a few weeks, like for a few months at most. And then the community started to put it everywhere. You know, like on social media, on the clothes, like literally everywhere. So we're like, oh, maybe we're gonna keep it. And now it becomes such a brand, so popular that unfortunately it's going to be hard for us to change it. Listen, at least when you do go public, the ticket will be in emoji.

4:58All you need to do is get to that stage. In terms of company founding, why did you decide this was the idea that you want to spend 20 years of your life on? We actually studied with something completely different. The reality is that the company was formed because of some sort of professional crush between me and my co -founders, where we were like, we absolutely want to work together. Plus, our excitement about AI. It was seven years ago, so it was not obvious as it is now. Not enough people were talking about it at the time, but we were super excited about it, to get like a new paradigm to build technology, new opportunities and all of that.

5:34The first company, the first startup I worked for 15 years ago, was already doing AI. We weren't calling it AI at the time, So I had some software glimps of the capabilities and when we started hugging face when we started the company where like okay What can we work on that is going to be both Scientifically challenging we all have a lot of interest in the science side of things But at the same time entertaining so we actually started with some sort of tamaguchi AI or AI friend However, you call it some sort of theory in next out chat GPT except just for like entertainment not for the boring productivity.

6:11And we actually did that for almost three years. We raised our pre -seed and seed rounds on this idea. We got a product out from a couple of billion messages exchange between users and this Tamaguchi AI. But as it sometimes happens, when we started sharing a little bit of the underlying technology and the underlying platform that we are building to do that, we saw a lot of traction from the community, from the open source community, from companies using that, and so that basically made us people from this AI to the AI platform that we are now. Can I ask the thing that's really striking about what you just said there is, hugging face today and in the last year with the rises become a very prominent brand in this space.

6:55But seven years, it's a long time. So when you look back, to what extent is the hype and the excitement that you have today around hugging face? To all extent is that like hype cycles driven by investors and hype cycles versus true underlying development in AI given you've had this 15 years of experience in the space. That's completely true that perception and hype can be delayed or like very different than reality. My understanding of the current situation is that the DC and the mainstream interest in And it's like a catch up on the reality. Because if you look at usage of AI, it's been massive and it's been growing massively for the past three years.

7:41Even before, before Chatchy Pt, even before the new Bing, AI was used in Google for like billions of users every day. AI was used on Facebook to rank your posts. AI was used on Zoom to remove your backgrounds. So, in my opinion, current interest about AI from VCs and the public is just a catch -up on usage and on how it's carefully comes this new paradigm to build a product, technology or workflows. So I don't really define it as hype like a catch -up from usage with the perception of VCs in the mainstream. Have there been catalytic breakthrough moments in the last 12 months, which have taken it to the next level.

8:26Or is it just as you said, a continuation in a catch up from existing usage with an incumbents? Like I think Open AI and ChatGPT to the world brought this awareness. Was that a step function change? Or was that actually just the continuation of the catch up that you mentioned? It's really important and I obviously have an agenda there, but to remember most of the progress that we've seen in AI is based on open science and open source. It's because AI has been so open and that everyone is building on top of each other with this really interesting positive loop of feedback, improvement, experiments that we could move so fast with AI, right?

9:08With open science, with open source, without Google sharing their attention is unit paper, sharing their birth paper, the latent diffusion paper, maybe we would be 30, 40, 50 years away from where we are today. And then what happened, I think, in the past few months, is you started to have some mainstream breakthroughs, which are GPT, and also, like on the underlying technology stack, I think you've had better hardware, more availability of GPUs and better optimization of models and all these techniques to basically be able to run bigger models at scale for hundreds of millions or billions of users These were kind of like put it like the last missing piece for AI to go mainstream as what we're seeing right now So before we dive into kind of the granny like I do want to get into kind of different models So we can be approached.

10:02I do have to ask I've done now three deals in AI by any good ambassador has done Thank you, Club. And everyone they've asked, it's been with Valley VCs and they've said, AI, the heartbeat of it is in Silicon Valley. We have to move the founders to Silicon Valley. Do you agree that SF and Silicon Valley will be the center of this next generation of AI startups? What do you think that's bullshit? And it's actually a decentralized globalized talent network like we've seen over the last years. I think there's no denying that there's tremendous excitement activity there. Both of that I felt that because I was in San Francisco two weeks ago and I just tweeted that I was around and said oh we should do get together with like open source AI, fellow community members and in a matter of few days the thing blew up and we ended up with 5 ,000 people joining for a huge community showcase event that people started to call the woodstock of AI.

10:57I really felt thanks to this event obviously the energy that you have in Silicon Valley right now with AI. But at the same time if you look at the full stack especially outside of just early stage startups, if you look at AI scientists, if you look at ML engineers it's heavily distributed. One of the point is for example Lama which is arguably one of the best open source model that came amounts for Meta, 10 out of 13 authors of flammar actually based in parts in the Meta AI lab that they have there that is huge. I think there's a lot of energy in Silicon Valley. So for all AI founders being told you need to move to the valley as the founders you need to be in the valley.

11:39Do you say that's fair or do you say no you don't? I don't think you do. I think you have to be in Silicon Valley sometimes. But at the end of the day I think you can build a company from anywhere. The most important thing and I'm sometimes like calling bullshit on founders saying oh I need to be there I've taken like a very strict decision to move there for my company at the end of the day I think it's important for founders to be happy and leave their happy They can build a great company and so the most important thing in my opinion is for founders to find where the happiest and Leads their happy London they should build a company there I do the same approach to my career.

12:19Everyone told me, I've been doing this eight years, Clam. I've been, ever told me, you have to be in the valley if you want to be in venture, Harry, you have to be in the valley. I love London. And so I've stayed and made it work. Listen, I want to dive into really the two models really of the world, so to speak, and they give me the effects or credit for this. Here's essentially posed the notion that there's one model which is your kind of open AI view of the world, which is one model to rule them all. And then there's the other idea which is an open source model with many models before we get into like opinion, just for people to understand how do these approaches differ?

12:50They differ a lot in where do you allocate AI builders? One model to roll them on U -Bets on down -flake models getting bigger and bigger with more and more generative capabilities and the builders of these models being concentrated in one or few organizations. In the model where you think there are a lot of different models, you bet on things being more distributed on the fact that all companies are going to actually build and train models. And that comes from the thinking that in the simplistic terms, the model or AI is like a code base. It's a bit different, but at the end of the day, it's a code base.

13:36And so it's silly to say, oh, this code base is better than this code base. or there's going to be one code base that is going to roll all code base. The truth is, the code base is good or bad depending on your use case, depending on your constraints, depending on what you want to do. So if you're Facebook, you have one code base that does what you want to do for your users. And it's the same thing in my opinion for AI. If you were a company that wants to do consumer products, you need to build AI models that are optimized. Whereas this use case that are going to be faster, cheaper, more efficient, and that's how you differentiate yourself.

14:18Okay, so say with that consumer idea, we have the consumer idea and we need to leverage AI models to build what we want to build. We then have the choice of whether we choose many models and the world of many models or leveraging one model to rule them also to speak. How do we know which one to choose and why do we choose which one? It's a tough question, especially because it involves a lot of shorter versus longer. The reality that sometimes today using one model behind an API is faster and easier at the beginning, but the challenge in the long run, you have more risks because then you don't really internally build the capabilities to actually do AI yourself.

14:58You can't optimize this model, so they're going to be inherently more expensive and you risk being in competition with others, not really differentiating yourself. The analogy that I sometimes like is that on the early days of the web, you could use something that would create a website for you, right? You could use the equivalent of a square space of the weeks. And that would make you feel good, right? Because really quickly you would have a nice website up and you can start experimenting. That's the equivalent of using an AI API for me. But the reality is that if you really want to differentiate yourself, build your capabilities and really do something that is specifically catered for your use case, for your users, and be able to optimize that, the same way you need to write lines of code to build a technology product.

15:51In my opinion, you need to train, optimize your models in the machine learning paradigm. So I agree with everything you say, but I'm also aware of enterprise buying and enterprise education levels. And you sit in Paris, I sit in London, we know how slow and bluntly ignorant large enterprises generally have been an art in new ways of innovation and actually if you can offer them a bundled service with a blue check mark and it's verified and safe and it's easy then they'll go for it. Do you think we'll see a three to five year period where they go for the bundled solution because it's easy before they realise the need to embrace the multi -load.

16:32Maybe and that's why there's a huge opportunity for new companies to disrupt the cutbacks because they're going to go for the easy solution whereas other companies that are more like AI native are going to be going for the more disruptive approaches and that's with a lot of startups, right? We see that with a lot of startups that are using packing face if you look at models closer to runaway email or stability AI or like photo room in Paris. You see these AI native startups that are actually building training their own models and how they can in my opinion build much better things than the ones that just use APIs.

17:12What you're describing is a great opportunity for AI native startups to disrupt the incumbent for your solutions even if it's not sustainable in the wrong way. I brought you a fantastic tweet from Jan LeCune who said that the biggest obstacle to the open model so to speak is actually the legal status of the training data. How do you think about that? Is he right as that the main obstacle and do you think that's fair? Yeah, he has a point. I would argue that it's a challenge for the corporate area approaches to because they're also going to get challenged by that. I don't know if you've seen but I learned most tweeted that it is going to sue open AI for using tweets in their training for GPT4.

17:52So I would argue that it's a challenge for AI in general and it's gonna be good this year I think because we're going to start to have legal clarity about how do we consider for use? What are the regulators Expecting from AI companies to respect in terms of growth. So I'm excited for more clarity on regulation this year I think it's gonna be a good thing for the field as we mature. Can I ask what do you think happens with content access? We've also seen Reddit here with now -solvings talk about how they're going to need to monetize their content and access to it. How does the relationship between models, whether open source or closed, and content providers play out, do you think, in the next six to 24 months?

18:34That's a good question. I hope we get into a model that works better for everyone, for their content creators to keep incentivizing them to create good content and AI companies. is alike. We have some initiatives on the topic. For example, we've been training a really good code model with something called B code and we trained for the first time. I think we were the first organization to train on a fully opted out data sets where developers could just remove themselves from the training. We're working on the topic. There are some interesting things. We're still scratching the surface in my opinion, both looking face while we're on this topic and before we dive into business, what do you think happens with the Elon Musk and OpenAI case?

19:20I saw it and I was like, why does that end? Yeah, it's a good question, they have very different approaches. Musk obviously is like a character that is going to say one thing and he opusits almost on the same day. But I think he has a point in the necessity of openness for AI and how important openness and transparency is for AI, but also for a society. So I appreciate that he's putting at least this part of the conversation in the spotlight. I do want to ask you, I speak to many of your investors before the show. All of them said to dive into business model and how hugging face makes money, speaking of the relationship with content providers and publishers there.

19:59How do you respond to how does hugging face money? What does that look like in the long time do you think? So our model is simpler than what people think. As a platform with a bit of usage, We can't like photo kind classic premium model. Most of the companies using us are using us for free We have 15 ,000 companies using us now and then a smaller subset of companies are actually paying us right and for us It's 3 ,000 companies and the reason why they're paying us is for premium features underpriced features Single sign on premium supports when they need help to use our tools and premium compute, right?

20:39For example, they want to use a gig face, but they want to upgrade to faster GPUs, then they're going to pay us. So 3000 companies are paying us, including metal, including Bloomberg, including Grammarly, and companies like that. Adi Charter, is it on a seat basis? Is it on a volume of query basis? What's the pricing model aligns that business model? It varies. I think we haven't really figured out and optimized for maximum more of a new because our main priority is more like adoption usage as a platform with network effects. Clam, would you get pissed off when people ask you how are you going to make money?

21:19Do you think it's the wrong question to ask? It's not the most important question because as a platform with network effects, The adoption and the usage is like the number one KPI course, especially it's an assumption and like a position that we took very early on with a gig face, especially coming also from like more like consumer backgrounds. Well, it's really obvious that Facebook or like a Twitter needs to focus on adoption usage first and this Assumption that usage is delayed revenue especially on the domain like AI where you expect Companies to be ready to pay for AI So if hugging face keeps being the number one platform that companies are using to be of the AI It's obvious that we're going to be able to make a lot of revenue out of that and build a good business But at the same time even if it's not the most important question It's an interesting question and the way I see it is that with monetization We have to take it as Staping stones and almost unlock some learning progressively you start with six figure revenue You learn from that you see how it works than seven figure, revenue, eight figure, revenue, nine figure, revenue, and you learn at each step, especially on AI because the underlying technology is moving so fast that probably the way we make money today is not going to be the way we make money in three years or in five years.

22:51It's interesting to do this monetization and revenue learning as we do, I think. Can I ask Clan, if you said about adoption there, I always think about Alex This Ram Paladangerie's and who said the question in company building is will the incumbent acquire innovation before the start up requires distribution? And you mentioned adoption there and it leaves me to think about who gains from this next wave most predominantly. And my thinking more and more is incumbents who are fast moving like Microsoft who are incorporating in powerpoint like Adobe who are incorporating into all of their suite of products.

23:25they will take 90 % of the gains actually because they have the distribution and they're moving fast. Do you agree or do you think actually start -ups of the ones who will accrue the most value in this best generation? I think if you were thinking about AI as API's, I agree. If you're thinking of AI as a more radical paradigm switch to build technology, right? And if you think about an AI startup as a company that is actually training models, creating new architectures, optimizing models themselves. I think it's a different story because this is really hard to do for the incumbents. And so I think the new startups have like an opportunity there to do things 10 times, 50 times, 100 times better than the incumbent.

24:13Why is it hard to do the incumbent? Sorry, I'm naive here. Why is that difficult for them to do? Because it's a completely different way to build technology. It's a way where you have to have scientists, for example, working for six months on a new architecture, on a new model before releasing it. It's a different paradigm of how you build software. Different enough, in my experience, that it's hard to do for bigger teams and bigger companies that are moving slower and that have started with a very different paradigm. At least that's what I'm seeing on the field. It's hard to predict the future again, but I think there are many opportunities for AI first startups and really startups who are not just using AI with APIs but really building AI themselves, building new architecture, building new models, optimizing their own models.

25:05What are the biggest challenges or barriers those AI first startups face? Is it data model access? Is it hiring? What's the biggest challenge you think this next generation of AI first companies? I would say hiring, probably right now, like getting like the best people and getting like this hybrid profile because it's science plus engineering hiring and getting the right set of Confirmers early team members is the harder thing Especially because there's a lot of competition with others companies have raised so much money that's for the really good people the salaries are like insane. That's in my opinion the biggest thing.

25:44We've seen some monster funding rounds very early for some of these companies, 100, 200, 200 ,000 ,000. Do they fundamentally need that funding for data model access for something specific or is it a case of the demand is there and so raise what you can? Does it cost more money to build AI first companies than all generation of product companies? It does a little bit, a bit similarly, in my opinion, to how you would build an internet company or like a software company 20 years ago because Compute is like more expensive for AI than it is for a traditional software. Because the team members that you need to hire, as we mentioned, are more expensive than if you need to bring it to our software engineers, it does cost more money to build an AI first startup then a regular kind of like software startup.

26:39One thing that I am not really sure of is this model of not needing just a little bit more money but much more money is really the right approach or not especially because we're starting to realize that more compute for models is not necessarily the right thing or at least that it's not enough and the return on investment on training larger and larger models is starting to go down. So it changes a little bit the perspective on raising more money and having more money stands on compute as some sort of a motor barrier to entry. But like all technology cycles you have companies taking bets if investors want to take the same bets.

27:24I don't think it's a negative thing. I think it's a good thing. doesn't mean that they're all succeed and that it's going to work. It's an interesting risk to take an interesting kind of company to build in my opinion. So I'm excited to see some of what these companies are going to build and do in the future. Can I ask before we touch on your fundraising, which I had some great stories about, by the way, but can we, uh, those touch on Elon has said before that, unlike most regulatory environments, AI, you cannot wait until it's in play for you to regulate it. Once it's in play, regulation will not be effective.

28:00Do you agree with that? And how do you predict the regulatory landscape to play out in the next six to 24 months? I don't really agree with that. My point of view as you've heard, I'm extremely excited about AI. I think it's a new paradigm to build all tech. At the same time, we're very far from the world where AI is autonomous and has conscience and is taking over the world and destroying humanity. I think this is more of here that is very sci -fi driven that we're very far from. And so I think when you take this point of view, you realize that regulation is necessary because it's a new way of building technology and it's going to create some challenges.