E489 | Philippe Petitpont (Moments Lab) & Gökçe Seylan (OX): Moments Lab Raises 24M USD

0:00In a world drowning in synthetic content, one startup is betting everything on something radical, the truth. You need to have real content. And real content is not working exactly the same way as artificial content. But here's the problem no one's talking about. Video is a complex signal. And the human eye is not very fast to understand what's going on. While tech giants chase artificial everything, Moments Lab cracked a different code entirely. So agentic AI in our definition is the capacity for a user to interact with a massive amount of video. Is the result something that sounds impossible?

0:33An agent has the capacity to do 1 ,000 search queries per second. So it's like having 1 ,000 people working at the same time, probably with more quality and more expertise. For years, everyone knew video AI was the holy grail. But until now... Since the invention of AI, we knew video discovery, video indexing, is a very high-value AI use case. But simply the solutions until almost left didn't deliver. What Reuters discovered changes everything, and it's happening faster than anyone expects. Editorial decision is going to be something that will arrive way quicker than what anyone can expect now.

1:09Join us for an exclusive deep dive into the$24 million bet that's rewriting the rules of video intelligence and why the future of storytelling just got very real. This is the EUVC podcast.

1:32This is a union of values. Let's start acting. This show is not investment advice and the hosts of this episode may be invested in the funds and companies featured. Welcome back everyone to another episode of the UBC Podcast. I hope by now this is the go-to show for connecting and championing the European-Montarico ecosystem. And today we are zooming in on the intersection of AI and video. obviously something that's higher in our mind. We're paying far too much money to agencies helping us with that. It's a frontier that's not just reshaping media, but also defining how we tell stories, build brands, and monetize content.

2:10We're joined to do this conversation with Philippe Petit-Pont, and I'm struggling always with French names, CEO and co-founder of Moments Lab, and Gürtse Ceylan, principal at Ox. She's giving me thumbs up here for doing a beautiful Turkish pronunciation. They're fresh off of a$24 million raise to scale Moments Lab's authentic AI technology. This is a deep dive into one of Europe's most promising B2B SaaS scale-ups and the VC perspective on why now is the time for video AI. So Philippe, let's start with our origin story. Tell me, what was the problem that you and Fred set out to solve when founding Moments Lab and how has that vision evolved to what you're building today?

2:49So usually a lot of founders are starting their journey by the previous work where it's probably during working in a big company that you find the best pain points on the on my side I was working in big media and communication group tv networks on the during my last job I had as an engineer the mission to be able to help journalists and content creators from this big network to produce video faster it was back in 2016 and where AI was not really a thing yet, but the pain point was already there, is where you are, as a media, investing a lot in videos. And most of the time, these videos get lost and you cannot really find them.

3:33And the problem is that when you are a journalist and you need to build streaming stories, you need lots of video. And it was very difficult in my previous job as an engineer working in a media company to really build the tech. and it's not really the mission of a TV network to digital technologies. Well, actually, Fred, who is not only my co-founder, but also my twin brother, is also an engineer, but more in big data on AI. And during a family lunch, I shared kind of the pain point I had. Well, I'm struggling to help these folks working faster. There's not really technology existing to be able to do that.

4:10So we both quit our jobs on team that together are creating a team. on the, you know, you're starting to have an idea about where our goal is to make sure that people are going to build video faster. But actually the issue is not really this. The problem when you deep dive is that video is a complex signal. Video is at raw pixels and audio frequencies. And the human eye is not good. It's not very fast to understand what's there in a one hour video. For you and me, it will take like one hour to understand what's there in one hour video. So very early, we identified after bootstrapping that the main issue was to be able to have machines understanding video.

4:53So the pain point was really how to make sure that you can be able to understand like vast volume of video to be sure that then you can unlock new use cases. And of course, building video faster is one of these use cases you can achieve when you know exactly what's just inside one million hours of video. So let's take some raw number, but if you have a look on YouTube, there's 500 hours of video uploaded in YouTube each minute. So that means that you need to be able to do that, which is probably the most intense and dense upload rates in the world. You need to have a technology that is able to understand way faster than human, better than human, and add a speed on other code that is not possible to achieve by human.

5:39So our goal was really to be able to develop that as really the first layer, like a foundational layer, to be able to develop application around that. And the application we've been developing now around really being able to build this video faster, relying on this approach. So one of the things you'll be working a lot on about release is an agent is able to take any text input from the users and to turn this text input into a video of already existing content. When I opened my JadGPT a couple of days ago, I saw for the first time Sora in there. I haven't even used it yet. Tell me just both from your perspective, Gautja, and also yours, Phil, why is it that this market is not won by ChatGPT and Gemini?

6:30First, there's two ways of building video. There's one way, which is synthetic video, and we're just very well-known players. And so, V03, there's Runway, Synthesia. when they are working about creating video that does not really exist, which is part of the use case when you're building video. But when you're building video, you also, in some market, whether it's entertainment, advertisement, tutorials for makeup, for instance, or even news, you need to have real content. And real content is not working exactly the same way as artificial content. From artificial content, you are taking a prompt on building something new that does not exist.

7:16When you are working in a vertical, you need most of the time to start from real video and to build a video from real video. On this, both the world would coexist together. I will never generate a video of my kids that will be synthetic. That would be point. I hope we'll never get it. That's a bit weird. So that's a very relevant question because when we started fundraising one year ago, It was like a common question about is artificial video is going to take the lead over real video? And the answer is probably more like, no, it's going to coexist together. So we need to have unified tool that can both work with real video and artificial video, I would say.

8:01I don't know if that was a question you had when we first met, but I guess, yes. Yeah, no, I actually wanted to chip in with, if we take a step back, I think the question Andreas you asked is very relevant, but it also comes up in every like vertical AI solution that you look at. Why is JetGPT not doing this or why isn't this a case for JetGPT? Whether it's legal AI, for instance, you have the likes of Legora, Robin AI, or whether it's industrial AI and you have an AI-powered maintenance solution. You can raise the same questions to chat GPT or perplexity and get somewhat of a working answer. And now they're also releasing video AI or image to text or image to video solutions as well, where you might think that's going to address the issues to a certain extent.

8:57But I think the real use case is we look for companies and founders who really understand the customer needs and can release AI products that can be adopted and used for specific use cases and their business use cases. so when we engaged with moments lab we were for instance incredibly impressed by the product depth and breadth that they built and their deep industry knowledge and how well they understood their customers and their customers specific use cases so for instance we spoke to customer after customer and we raised the same question to moments lab customers as well and they said if you want to, for instance, analyze a single file of video, yeah, you may be able to do that to a certain extent through maybe hyperscalers.

9:52But if you want to deploy it at scale and seamlessly integrate that into your ways of working and have a proper B2B self-solution that basically gets the job done for you and you don't have to deploy an army of engineers in-house to take care of that, then something like Moaman's Lab comes to the rescue. And I think it goes true for other vertical AI solutions as well. Phil, maybe you can just hammer home the point to everyone about how synthetic video creation, I guess is the word, is different from real video creation. Just be completely down to earth about how are these two different. And I also am sure with the use case, as you described, I probably don't want to create a synthetic video with my kids doing an artificial activity.

10:45We're not going to reminisce about that 10 years from now. So let's see which kind of use case you can use one or the other. So let's say your brand, you're building a 20-second advertisement. You need a pack shot at the end. There's no point of shooting a pack shot with your video. It's nonsense. sense. And by the way, that's interesting to see that one of the latest feature that these synthetic AI video startups have just added the capacity to import 3D models directly into the system so that you can take, I know, let's say it's a yogurt pack shot. You take the 3D model of your yogurt pack shot and then you start generating the pack shot with a prompt.

11:28Perfect use case, similar use case in, I don't know, Hollywood movies, you need to generate B-rolls. You need to show or to generate like artificial background. That's really perfect use cases. When you want to drive emotion, there was, let's take the champion sling final. You cannot really have the same emotion if you're using artificial content. So if you want to have impact and drive emotion, real video will be forever the best media to use for sure. But what But we see that if you want to... Sorry, is that because the thesis there being we will continue to have a stronger emotional response to something we know is real than something that's synthetic?

12:17Absolutely. On one part, you are like telling a story with scripted content that know exactly like character doing some stuff. In the other hand, we're not speaking about character. We're speaking about people, real people, about humans. and that, of course, a human is driving more emotion. An athlete, a journalist, politics, any documentary stuff, you want to have access to real content. And I think real content, because it's going to be rare, actually, because artificial video is very easy. So on social media, there are going to be a lot of artificial video. It's already the case. People are asking, is it real or not?

12:59and when you see that it's artificial, the impact is not as good as if it's real content. So there's a lot of speak about that on kind of institutions that start to authenticate real video versus synthetic video. But that's for real emotion on impact, real video will remain forever because it's not about video. It's about an event that is currently really happening. And if we jump a bit to some of your giant customers that everyone would recognize. One of them is Reuters, which I imagine is very much driven by the fact that, well, we see every week, I think in some major outlet, someone has made a blunder putting in something that was made with Genitive AI, not a picture, but most often it's facts or it's background stuff that's just not correct.

13:55And it's clearly because it was done with Yana Tubei. I guess that's why someone like Reuters is a key customer for you and a key use case. Yeah. So I cannot detail officially exactly everything about Reuters. But first, Thomson Reuters is really a big company. And most people see Thomson Reuters as an agency. It's one of the biggest agencies in the world. But actually, a big part of the business is not only this agency, but there's also part of the business. let's say or not i can generalize to uh this kind of of big player on on big agencies whether it's afp or ep which are really some big players in in this industry they are here to relate facts whether it was if it was all facts and i think the oldest video from thompson reuters is uh the coronation i think of uh the tsar alexander the second or something like that like probably at the end of i would say it's the 19th century so it's it's very very old at the 18th century sorry so it's it's it's about how do you navigate into a hundred years of video when you don't know what's inside the video i i can't mention more accurate details but some of our customers they have like that hundred years of archive and people and that seen this events are not there anymore.

15:18We have a customer in soccer, in football, women football from the 1920s. You want to build a documentary around women football in the 1920s. No one is still living today, know what happened at this time. So you need a way to kind of build a documentary around that and you need to be an expert around that. What these customers are looking for is a way to discover, is a way to create new revenue streams based on this understanding of what they have into their legacy, into their IP, whether it's very old archive like that or fresh content from a game that happened last Saturday. I'd love to ask you, because everyone, whenever we talk about generative AI, and especially now with something like yours that so heavily draws on real stuff.

16:12How do you solve the ownership part of the equation? Because I guess there's a lot of property rights. It's always that it's on their own, or do you run it as a platform? Do you have a plan to run a platform? That's a very relevant question, and that's a lot of the question we have from journalists because they know that it's a nightmare. Rights management is very, very complex. So for us, it's quite easy, actually, because we are not B2C, we are B2B. So when we are deploying our product, the customer most of the time already know what kind of rights level he has on his content. And of course, it's not like that easy saying that I have the rights on all the content.

16:51It's never as easy like that. Even on Getty, you know, there's like lots of different rights management possibilities. But we integrate natively this rights management level directly into the platform so that you can query, I want only sources where I know I have the rights. Or you can think about even... Kind of like, you know, from Google, where when you do a Google search, image search, you could go in and either say, I want something that has, you know, free property rights. Yeah, absolutely. Yeah, it's exactly the same mindset, but here it's kind of a closed pool of media assets. So we have this knowledge.

17:30And sometimes, of course, lots of right owners, they don't know what kind of right they have. And so for them, it's I am taking the risk, yes or no. A lot of them sometimes are taking the risk because they know, actually, if I don't know, I don't see how someone could know when I would publish this video. So most of the time, it's not really an issue, but that can happen. To be clear, you know, this kind of rights management level that no one doesn't really know if we have the rights or not. So that's not really our issue because we are more like a technical solution. But it's more like the end decision for a human to say, OK, I am clear on the right.

18:05I am taking the risk or am I paying? I asked you the ChatGPT question before, which is why is ChatGPT not going to do this or SORA do this? We have a very clear answer to that. Now I have another question, which is in a similar vein, and you kind of touched on it. YouTube is a huge player in this. No one, like 500 million hours of content uploaded every hour. I would imagine that having a tool that would very quickly allow them to scan across everything, both when it comes to empowering their search, but also in terms of allowing creators to tap into it and so on and so forth. Like you can imagine so many use cases here for someone like Google.

18:47Is this something that you know that they're working on? Is this something that you're already talking to them about? So we have a lot of discussion with Google. And I mean, one of the holy grail in video is being able to access the YouTube database to build new videos. Yeah. Doing video understanding. The main issue that Google had here is a rights issue, actually. and they don't have the rights on all these videos. They belong to content creators. Content creators won't be very happy to see that Google is trying to build new videos based on already existing videos. So we could think about, you know, business model where there can be some revenue share or stuff like that.

19:29But that's not really what's happening today. You could imagine a blockchain model on Google where if your video was used as a substrate of something that's created using Moment Slap or similar, then you'd automatically get whatever small part of the revenue that is made with that video. Well, and that's totally a fair point. You know, lots of our customers that don't have lots of technology around their video production tooling, they are storing their video on YouTube. And so when we are deploying our technology, we are scrapping all these videos from YouTube to recreate and do exactly what could be doable directly into YouTube.

20:09So the main issue was kind of this messy rights management process that makes it very complex. Blockchain could probably deal with that. And actually, there are a few startups that started to do it. It's complex. Jalismax started to do something similar. It's a French unicorn that was really specialized into content creator market on a really great company. The problem they had actually was that they were missing a critical part of the technology that was video understanding. Be able to understand which kind of video can work and which kind of video cannot work. But what you are, I think, describing here is something that will happen in the future for sure.

20:49No doubt about that. Incredibly exciting. I think you guys are doing something that's so exciting. Let's talk about the identic AI in the context of Moment's Lab. Could you just describe completely clearly what is it, Identic AI, in this context? And why do you believe that this is going to be content or a game changer for content teams? So, Identic AI, in our definition, is the capacity for a user to interact with a massive amount of video. So, it's like me asking my knowledge base of video, let's say, I know, 1 million hours, 1 ,000 hours, whatever, the capacity to start a conversation with all my assets, my video assets.

21:31So for us, it's like customer or user asking, build me a three-minute video about Luis Enrique on the PSG during the Champions League final. And so there's answer from the prompt. It's a prompted experience at GPT, but that is taking all the information from videos. because we've been able to index video in a massive way with very, very accurate details. And the prompt then is able to say, okay, let's see on the web what was the last highlight about Luis Enrique. Okay, it just won with the PhD as a Champions League last Saturday, okay? Here's some highlights. And here's which highlights that are going to be present into your video archive, whether it's fresh, live, or very old.

22:21then the prompt on the agent is able to directly find all the moment that I will need to build this video. So most of the time today, it's human work on people searching when there's a search. But most of the time it's like, OK, there's a three hours game. I'm going to have a look to where I can see this in the week. Doing some instruction, you know, to the team on being able to understand what he's saying. So here it's like going way faster to sort the content, which is 70 % of the time of a content creator to build this kind of video. That's the first thing. So it's about like building video faster.

22:59It could be extracting insight faster. Anything that I can prompt, I can have an answer. Either a video answer, either an answer about, I want to see where the likey logo was seen, how much time. So the agent has this capacity to do that. The second thing is being able to do what humans cannot do today. Let's say you're a financial media and you want to build a documentary about the evolution of Apple stock over the last 30 years. To do this today, you will probably need an army of people, financial experts, analysts, financial journalists, senior journalists, because you need to pass a lot of data.

23:40If you're a financial media, you already have all this existing data, and it will require an amount of time that probably won't give any return on investment on this documentary. That won't be possible. Now it's possible to do it. An agent has the capacity to do 1 ,000 search query per second. So it's like having 1 ,000 people working in the same time, probably with more quality and more expertise. So there's the capacity to build a new story that may not be possible before. And the third thing, and the last one, is when you have this amount of video, video is knowledge, right? It's probably the best way to start.

24:18Yeah, that's a good point. You have the capacity to transform this knowledge into insights. And it's about giving you a new meaning to your data, extrapolating prediction, for instance. So that's further on, but probably what's going to happen in next year. I have one crazy curveball. Surveillance. You're smiling, you're looking away. Have you ever thought about that use case? When you start working on computer vision, you're being approached very quickly by companies that are doing that, Palantir, Airbus Defence, Dassault System. Especially in France when we started, there's a lot of companies like this.

25:04And sometimes we are, you know, I would say institution that, how do you know I exist? How do you know I'm doing that? Wow. Anyway. And so we started to have a look at the very beginning of the company because, wow, there's probably a tremendous market to do border control. And, you know, and the thing is when you're working with content creators and journalists, it's very difficult to have also a market in security and defense. So that's the first thing. And it's on the same audience. but also it's not exactly the same use case. In the market, we are media entertainment, sports, brands. Our customers are considering video as an asset.

25:44In defense, in border control, you don't consider our patient monitoring, for instance. You don't consider video as an asset, but that's just a way to exchange information. The kind of AI you need to analyze a border, for instance, is not exactly the same. In video, for asset and creators, you need to avoid too much noise. So you need to concentrate on what's important. So you don't care about like a small bottle that is at the back of you in the background, which is totally the opposite when you are working in security. You cannot afford to just miss a small detail in the video. So the way AI is working is not exactly the same for these different use cases.

26:29One is you need some noise to be sure that you don't miss anything. In the other one, you need to be accurate, but to filter only the data that might be interesting to tell a story, for instance, to tell a design insight. Goethe, I'd love to ask you about the whole journey of scaling from Europe to global and focus on product market fit and so on. I want to ask you a bit about the magic that you felt when you first saw the product working in front of you. I'd love to ask you, what was it about this experience that made it an absolute must invest for you guys at Ox? I actually did have a jaw drop moment when I first saw the product demo, but tried to keep it cool in the meeting.

27:13No, it was super impressive. I think in B2B SaaS, it's relatively more rare to actually see the future unfold in front of you. And I thank Bill for letting me have that moment. No, but that's a very good and valid question, I think, especially now that every pitch that a VC sees is the next big AI company that is transforming something. So how do you separate the signal from the noise? And I think what really stood out for us when we saw moments that was the product depth and breadth and a founding team, so Fred and Phil, who have such deep technical knowledge, but also such deep understanding of their customers and their core use cases.

28:01So they've really built a product that is fit for purpose, that fits all the requirements for customers. And I think this goes back to the why CHPT can't do this question. And at Ox, we always look for companies that have nailed product market fit and have emerging signs of go-to-market fit. And Mom's Lab checked all the boxes. and what does clear evidence of product market fit looks like. So they had a very well-defined customer profile and a very well-defined use case. Companies are using Walmart staff to run their core businesses and they're experiencing a clear market pull. So they have a very strong cadence, a consistent and accelerating cadence of new customer acquisition and also excellent retention metrics which is actually hard to see in high-growth AI companies these days because you won't even have two years of history to test retention.

29:05We also spoke to so many customers and why they are using Walmart's Lab. Interestingly, every one of them said, since the invention of AI, we knew video discovery, video indexing is a very high-value AI use case. but simply the solutions until moments lab didn't deliver they didn't deliver on accuracy they didn't deliver in terms of cost or scalability they were addressing a need that customers already knew about so moments lab doesn't need to educate the market or explain to the customers like the potential return on investment from this from this product the customers already know the pain already feel the pain.