Join the Age of AI Mailing List. You'll be in good company with CEOs, investors and innovators around the planet.
A few years ago, DeepFake technology appeared on the horizon.
It allows you to take a person’s video and voice recordings, and create a digital “avatar” that you can use forever. We saw synthetic videos of political leaders and celebrities saying things they had never said in real life, but which looked so real they were indistinguishable from the real thing.
While many were concerned, others saw a more ethical potential –making unlimited videos with “AI models,” by simply giving them a script to recite.
One such company is Colossyan, and in this episode I sit down with their CEO Kristof Szabo to discuss how they’re dealing with the technical and non-technical challenges of building such a unique, AI-first product.
This is an example of a product that would be unimaginable a few years ago, but which will be virtually everywhere in the future. AI celebrities (which do not exist) are already taking over social media and have scores of ardent fans.
Personally, I can imagine a world where 90% of actors in movies are not real human beings, and where almost every movie you see is “animated.” Not sure if that’s what will happen, but it’s a definite possibility.
Here’s what we discussed:
2:01 – Who are early adopters for a technology like this?
7:16 – Challenges in building the technology; what makes it so hard
13:08 – Why the “secret sauce” matters less than the product roadmap
18:00 – How to recreate body language for an AI avatar? (Also, Aman makes a non-European joke)
22:30 – Colossyan’s origin story
27:00 – Lessons from talking to customers
31:36 – Fundraising cycle and product-led growth
37:00 – How to keep research and product teams in harmony
42:23 – Crazy idea: doing away with the “ML team” in an AI company?
46:48 – (Allegedly true) story about Kayak
(Ethics Policy: These opinions are 100% my own as an independent observer and educator. I don’t own stock in guests’ companies or their competitors, nor do I get paid by them in any form for any reason at the time of publishing, unless specifically stated. Episodes are also not intended to be an automatic endorsement of any company or its products and services.)
(Scroll below for the transcript.)
Are you a “non-technical” entrepreneur or professional?
It might seem like the biggest success stories today belong to the tech tycoons — while you’re left behind as a second-class business leader.
But there is a way to leapfrog to the head of the race instead.
Become a “Tech Fluent CEO.” Read the book here.
[00:00:00] Aman (Host): Hello, everybody. Welcome back to the age of AI podcast, where we talk to companies using AI for real tangible business value. I’m a host Amman again and today I have with me another special guest guest. Is that right?
[00:00:18] Kristof Szabo: Yeah.
[00:00:19] Aman (Host): and, Chris, as I, as I will address him is the founder and CEO of Colossian, which is very interesting company.
[00:00:28] It’s they build AI video creator. So in a sense, if you wanna create a video with an AI avatar, which you can just give the words and they will speak it in an authentic human voice. And, with that, you can create many. Videos at scale. That’s what Colossian does. So it’s, it creates AI avatars for video presenters in a way.
[00:00:56] Right? Christoff, thank you so much for joining us.
[00:01:01] Kristof Szabo: Thank you for invite.
[00:01:03] Aman (Host): Yeah. And so we are gonna talk about how this technology works. Um, what were the challenges in building? What were, what it was like to launch this technology? I think Christoff, uh, recently came out of a beta launch for, for the product. So we’ll talk about his lessons learned along that process and. Other aspects of what other people, what their customers are already doing with video. Like what else is going on in, in the landscape of digital media creation with AI? So let’s start with the high level simple question, right? Who are your target customers, and what’s their situation.
[00:01:44] Kristof Szabo: Uh, well, it’s a tough question to start with, but, we have been, we have been trying to answer it for the last six months or so. And initially where we started my vision was to bring this technology to the people who want to create. Marketing videos who want to go into the direction of, creating videos easily for whatever purposes.
[00:02:05] So video content is, you know, omnipresent and it’s, it’s a must have for any business, whatever size it might be. Um, even, you know, the smallest freelancer based businesses to the largest multinational corporations video is a must for internal communication for external communication, for explainer videos, for product videos.
[00:02:26] So once I learned about this technology, and by that time I’ve been involved with video production for about five years. Um, you know, it was pretty exciting. Like you can create a video basically from nothing, or you can create a thousand or million videos, um, which just. Basically input from, from putting in a script and then connecting some sort of database that you can push out thousands of videos that are customized or personalized.
[00:02:57] So that was the grand vision. That was the, that was the. You know, the big idea that we wanted to wanted to come to this world, but of course, starting small, we had to build something like an DP, something that could be tangible, that could be used by people that could be built with, you know, a 5% team. So we started to address a problem that was basically creating a presenter led video.
[00:03:21] Uh, let’s go the spokesperson. That is a digital spokesperson that can, for example, tell about your business, that. Onboard, your new colleagues that can explain something about your product. If it’s a, if it’s a physical product or if it’s a software product, it can create an explainer video essentially, or a presenter video.
[00:03:42] So as we are standing right now, we are targeting small, medium enterprises. And we also have a couple of enterprise clients that are, of course, much bigger and they have a, a wide variety of use cases. But I think the. Interesting to the most. Um, no, the ones that they’re most likely to start with is, these kind of explainer videos, onboarding videos.
[00:04:07] Um, yeah, I think those are the, those are the ones that they have starting with. And, as I said, We even, you know, have just, one person companies that are creating, educational videos for their audience. We have, larger clients that are creating onboarding videos for their stuff. And, yeah, so it’s a, it’s a vast variety of clients.
[00:04:29] I would say one third small company is one third, medium, and one third larger.
[00:04:35] Aman (Host): I see. And so I think, I mean, the basic value proposition is pretty clear, right? If you actually had to get a recording studio and have people sitting in front of a camera, take multiple takes and this and that. You know, it’s this whole overhead that you don’t wanna deal with. If the purpose is not to make an Oscar willing film, if the purpose is just to disseminate information in a video format, then it makes sense to use the, a technology like AI, right.
[00:05:01] Kristof Szabo: Yeah, yeah, yeah. Yeah, as you say, it’s an interesting way to put it, but right now, of course, the technology is not capable of producing Hollywood quality films. So there’s, depending on the, on the input, you’re not gonna, you know, you can get, um, very kind of like informative videos, where you have a presenter.
[00:05:21] So think of like news or weather reports. Those are something that are attainable with the current technology. And in the next three, five years, of course, there will be crazy. That will be possible, but for now it’s pretty, pretty much like informative videos, um, that are kind of used for educational. I think educational purposes is a big one where.
[00:05:43] It’s, it’s kind of like, like you, you can have a teacher or someone to explain a complex concept. That will be a little bit more boring if it’s text or audio. Um, additionally of course, the benefit of the technology is that you don’t have to go back to the studio. You don’t have to even go to the studio and you don’t have to stand in front of the camera for hours or.
[00:06:04] Then edit the video for hours or days. That’s a lot of money, a lot of time, a lot of energy put into it. Essentially what you can do is take a basic script. Put it into our system and pick an actor. And you have a video in a couple of minutes that you can start posting. And if you change something about, you know, the script, if, you have some updates or you have new episodes, then you can just change that in the script, create a new video and put it out after a couple of minutes.
[00:06:30] So that’s the main benefit of the, the whole technology and the whole pipeline of your building.
[00:06:37] Aman (Host): Yeah, that, that. A lot of sense. And so let’s talk about the current challenges of, this technology, right. Of building this technology. Right. So, um, the first thing I would say is, you know, if I just had to, if you just had to judge the final output, of course it, you have to define what is good quality and what is not right.
[00:06:59] So let’s talk about that. Like how, what was the engineering or the product development product iteration process. For you for building this technology.
[00:07:11] Kristof Szabo: mm-hmm so you can think of, you know, a lot of, lot of things with AI, the main issues, of course, collecting data, building a database, cleaning the database. So. It’s a lot of experimentation, which is sometimes annoying and it takes a long time, but that’s the beauty of it. And it’s also something that will provide a competitive advantage for any company that builds their own proprie system.
[00:07:35] So in our case, we have been building the technology for almost two years. We have been recording, you know, hours and hours of video and then converting that, training our machines and then putting out, the video. And of course, looking at. um, we always see that there are problems, you know, there are artifacts in the video that we have to fix, but I have to say, you know, that the, if we, if we show it to someone, they will look at it and they will, okay, it’s the video?
[00:08:03] What is, what is what’s new about it? And when we tell them like, oh, this is AI generated, they will go, whoa. Okay. That’s that’s crazy. So that’s the reaction that we, that we want to get. Um, But yeah, it, it it’s been an experiment. It has been, a long process to get to where we are. And that’s something that not a lot of people can say that they have this technology.
[00:08:26] And that’s kind of, you know, our advantage that we, we already at this level where we have a, we can create photo, realistic videos, the digital actors, and now we are getting to the stage where we can generate those video. Two times faster than before. And we can also, record people just a couple of minutes and then create their avatars already.
[00:08:47] So we don’t need to record, you know, half an hour video of someone, but a couple of minutes basically. And yeah, that’s, that’s kind of the speed of the technology. So it’s exponential and, I’m, you know, proud and happy to say that we also keep up with. With the exponential growth in the, in the, in this synthetic media technology.
[00:09:07] And we are one of the, one of the top companies there, um, among a handful of companies that have this technology in their possession. So that’s pretty cool. Um, and that’s kind of like the, the that’s, that’s what kind of the problem with, if you wanna put it this way, that it’s a, it’s a complex technology and it’s a hard technology and we needed a, you know, a big team.
[00:09:29] Relatively big team to, to make it happen and make it work. But that’s also something that gives us the advantage and gives us the, the that’s, what we build the business upon. So, um, yeah, it’s not an easy industry, but, but I think it’s very exciting one, so it’s definitely worth the time and the money put into.
[00:09:51] Aman (Host): I understand the goal is to, of course, to build an AI avatar, where people are like, shocked that this is a, not a human being, speaking to them. Right. What are the different aspects of building an AI avatar like that? Educate us about that, process. So we know why it’s that hard
[00:10:12] Kristof Szabo: to go in detail with
[00:10:15] Aman (Host): into an
[00:10:15] Kristof Szabo: with the.
[00:10:16] Aman (Host): but
[00:10:17] Kristof Szabo: Yeah, I wouldn’t, I wouldn’t even be able to, so there is no danger that I will, I will say something like that, but it’s, um, it’s a long process. So it’s basically to create one avatar. Currently. It would take about a week if we. You know, if you wanna rush it, probably it will be done in three, four days, but essentially we have to have the raw data, which is the recording.
[00:10:38] So we record good quality with good frame rate, good lighting and everything. So we really have to set up a studio environment with the actor in the middle. And then, we have a professional, videographer who will record it. We also have makeup artists, um, stylist and so on. So. Of course, the input video has to look perfect because the output will be as good as the input.
[00:11:03] And later on, we cannot improve the input material. So we have to make sure that it is as high quality as possible. And once we are record about half an hour of footage with this actor, then we will be able to use that raw material to build the model, essentially the algorithm that we’ll be able to learn how this person moves, how they are moving their mouth, how they’re moving their.
[00:11:25] And Tong and you know, everything and, and we will build a 3d avatar of this person. And then once we essentially have that, then we will be able to take, a base video, a background video of the person, essentially just, standing there. And then we will be able to. Move the mouth, according to the input, audio.
[00:11:46] So once we have the model ready, we will take an audio. We will take a text from the user, generate an audio. And based on that audio, we will, um, essentially, you know, control the face of the person. And, that will drive the, the. Model of the, of the actor. And once we have that, you know, we have a, a video output.
[00:12:10] So it’s a process that, takes a couple of minutes. We are currently at this stage where one minute of video can take five to 10 minutes to generate. So it’s a pretty heavy process. Um, and then of course we are providing a very user friendly interface where the user. Add other video elements. So they don’t have to leave the platform to create a video they can create, put in the script, choose an actor, and then add in the extra media assets, background pictures, animations music, and, any logo, intro, whatever they desire.
[00:12:43] So we are also building that platform continuously, but, so far we have gotten really good feedback from our, from our current users.
[00:12:51] Aman (Host): I see. And, you know, because I’ve seen a few other companies, you know, so let’s, so we talked about the, how the product works. Right? So, thanks so much for sharing that. Now, looking at these different technology blocks, right? There’s the. You record the actor, you know, straightforward, you record them, you record their voice, you record their video, you train models, which can take an audio and put control the mouth in a certain way for that actor, for that particular actor.
[00:13:20] So you, they can pronounce the different syllables. And you can convert the, a written script into the audio as well. So it’s text to speech, speech to video, right. To put it very crudely. Of course, it’s much more complicated than that. If I look at the landscape for, let’s say video avatar tool, I know, for example, one other company, which is trying to make it a PR thing as well.
[00:13:44] Right. Where does the IP lie? Not what the IPA is or like, and how it works, but where does the IP lie for a technology like this? Is it just about I do it better and I don’t tell anybody how I do it, or is there more about like, okay, if we have a patent and nobody else can do this, like we do, how does it, how does it work in the landscape of competition?
[00:14:08] Kristof Szabo: That’s a very good question. Um, so since it’s an early industry and an early business, and we are still, even though, you know, we have, companies that raised 50 million in this, industry, we are still really early. So it has been not three years. And I would say, you know, it started out with. Somebody taking some research, putting it into practice, bringing it out of university or bringing it out of their dungeon.
[00:14:33] As the deepfakes I thought started, then universities pick it up. And, now, you know, there are several startups, trying to build a business out of it. So it’s started with, you know, not even I do it better, but I can do it. And I have these digital avatars and it’s pretty, you. It’s, it’s a unique thing that, we can do because we have researchers and there are only.
[00:14:55] You know, um, a handful of, researchers in the world that can do it. Um, so it’s not even, you know, because deep fake you can take, an open source code and you can face who up, or you can do whatever. This is not such a, such a technology. So it requires a lot of complex, models and algorithms and then the right amount and right combination of them.
[00:15:15] So in the end it will provide, you know, this very complex IP where you. Essentially you have a technology that, of course, you know, this is the, the main, the D IP that, investors will ask about and we’ll make sure that it is in the company that they’re investing in. But then of course the next step will be differentiation, building a better product, building a better UI.
[00:15:37] And building, you know, a PR around the whole thing, which is great, which is, you know, bringing a recognition to the industry, which is going to increase the valuation of companies in the industry. Even though it’s, it’s, it’s the other company doing it, but we can tell like, you know, this, world famous superstar data.
[00:15:55] A synthetic video for this campaign and so on. And we also have those campaigns coming up. So, um, I really think that we also put a lot of, IPN value in the company that we have those contracts and we have those actors in our, so that’s the other, the next step I think, is to have those actors in our disposal and, and be able to offer it to brands like, Hey, you wanna make a video with this influencer?
[00:16:20] Here you go, these are the rights, this is how much you’re paying. And this is how many videos you will get. This is the kind of campaign you can run. So that will also bring IP to the campaign that you have, those kind of contacts and network. And yeah, so I, I, it, it’s a process and I think it’s, it’s really coming to shape right now.
[00:16:38] So it’s really figuring out, you know, what are the use cases and what are the industries that will be first to adapt the technology. And then, how do. How do we build for it? How do we, what do we focus on? Like what’s important for the user what’s important to, um, how do we differentiate our product?
[00:16:56] Because essentially technology now is not enough to compete with the, with the starters. So you have to have the right product that you can sell and you can have the customers to buy it. So that’s what we are focusing on right now.
[00:17:10] Aman (Host): Yeah. I like that answer because, you know, I like the transparency that, yeah, it’s not a technical. War like technical race. It’s more like, an execution race. Like whoever can build a better business, with that technology and implement that technology properly is that’s what the whole thing is about.
[00:17:29] Right. Just having the nicest research that, you know, you can look at a video and say, wow is not enough,
[00:17:35] Kristof Szabo: Yeah, exactly.
[00:17:36] Aman (Host): I mean, that that’s a lot in itself, but it’s not like the
[00:17:39] Kristof Szabo: Mm-hmm
[00:17:40] Aman (Host): factor as to whose technology is better. Okay. Just a side note, by the way. Um, when we talk, um, when we watch a video, it’s also usually the body language. That is a part of the, the action, right beyond just the, the mouths, the mouths being faced. So
[00:18:02] Kristof Szabo: Yeah. Yeah.
[00:18:03] Aman (Host): just as a, I’m asking a nerdy, technical question, you know, out of my own curiosity, how do you get that holistic? I mean, of course it’s still under research and development in general, but tell us about that problem. Like how, what are the different approaches to solving that problem? How people think about solving that problem and.
[00:18:26] Kristof Szabo: Yeah. It’s a, it’s a tough one as well, but I would say, um, one way to approach it. What we have done with the face and the mimics is really to, to teach the model, teach it, you know, what kind of words, what kind of, you know, maybe if you’re asking a question, then you will move your hands apart. Or if you’re discussing something, then you will, you know, put your hands together.
[00:18:47] And then I think those that’s something very natural. That comes to people very naturally. If we see something that’s a little bit off like, oh, why is he moving his hand? And that’s a bit weird, but, it’s a problem that we can approach with, you know, training the model for the hand movements and then generate the hand movements.
[00:19:05] So it’s a little bit, a little bit different, like, because on the face, you know, you have very, you know, specific points and the hand synthesis is actually really complicated because you have, you know, many sites and many movements. It’s really complex. So we have to approach it. Sign of like, like two dimensional first, like what kind of movements you can record, what kind of movements can you tie in to the speech?
[00:19:29] And then once we have that, I think that will be already good enough. But then the next step will be to really train the model to learn the hiring movements and, and really kind of also. Yeah, it’s really, it’s really weird because in different cultures and different, different people have different movements, so different things that you can do with your hand.
[00:19:48] That means different. You know, if you talk to an Italian, it’s really different than talking to an Norwegian. Um, so that will be the next step to really experiment and figure out like how we can make it work best and how we can make it most natural and come across as, you know, um, a person that. That is, has the right kind of movement of the hand.
[00:20:09] Really. Um, another kind of approach is, is really just, you know, video editing. So take the person and, cut them in half head and arms and then, kind of tie together. Somehow. I also see some potentially in that, on the, on the short crime, but on the long run we are, you know, we definitely have to go into that and definitely train the model.
[00:20:32] To, to learn the right hand movements. There is no way out of it.
[00:20:36] Aman (Host): Mm. Yeah, because there’s also the, there’s the hand movement, which is one thing. There’s the body, the, the body movement, there’s the head movement. Right. And none of them happen at the same. Speed as the mouse movement, they all have different paces, right? The speed at which you speak, say, it’s not like you can match word to hand movement to head nodding or whatever.
[00:21:00] Right. When you are speaking with emphasis. So it’s a tough technical problem. You have all these multiple things going on at the same time at different, levels while communicating.
[00:21:10] Kristof Szabo: yeah. Yeah. I mean, it’s not a huge problem for computer to do these parallel calculations, but to, to kind of like also with the, the lip sync, you know, it. It’s hard. Like sometimes, you know, you have the, the wrong kind of calibration and then the mouth is not moving the way that you expected to. And then people, people are very, they, they are, it’s easy to notice because it’s just human to notice that something is off like, you know, um, so we are working on that and, but I think the next, the next step will be that actually like the bigger, let’s say the next step of this technology will be to change the kind of the camera.
[00:21:48] So that you can have like a side angle and you can have like, kind of some movements, I think, that’s what we are gonna see next.
[00:21:54] Aman (Host): Maybe you can put a mask on your avatar and that will be a quick
[00:21:58] Kristof Szabo: would be easy. That would be so easy. Yeah.
[00:22:01] Aman (Host): for the V2, the V3 of the protocols, you can say, Hey, we, the actors are wearing masks.
[00:22:07] Kristof Szabo: would be so easy. I wish I only wish I, but yeah, so, but, I don’t think the market would reply to that very well.
[00:22:15] Aman (Host): Uh, okay, cool. So then, you know, let’s learn from your beta launch, right? And you already talked a little bit about that, that you’ve learned a lot about your market, your market as you from six months ago. But tell us, tell us more. What? So what did it take to build the first product? Right. What was the first product?
[00:22:37] How did the product evolve,
[00:22:39] Kristof Szabo: Mm-hmm
[00:22:40] Aman (Host): um, some more history about the company, you know?
[00:22:43] Kristof Szabo: Cool, cool. So if we go way back then, as I mentioned, I worked with video from about 2015 and in 2018 was when, you know, deep Fs became the new sensation and that’s when, I first found, you know, about, you know, face swap and deep defects. And I was like, Hmm, that’s interesting. But how is that gonna, you know, affect, fake news and you know, that elections in, in a couple of years, and, it was such a big hit in 2016 with all the fake news that I was expecting something crazy, from 2020 and deep fix.
[00:23:17] So with two of my friends from DTU, which is the Danish technical university at the time I was at the Copenhagen business school and I was just finishing, we have set up a company that, was, you know, essentially the goal was detecting the effects and then it expanded to detecting fake images, fake videos.
[00:23:34] And we started to build algorithms and, on the technical side it was developing nicely. But after even, you know, working with bigger media companies in Germany, we just saw that there was no business. There was no willingness to pay because they were not convinced that it’s such a huge issue or, they didn’t see how we can, you know, solve it.
[00:23:54] I’m not sure, but, but we were just not sure that that was a business. So I think the summer of 2020, we decided to switch, focus and actually start creating video content and. That, you know, we already had the expertise. Of course, we knew how these videos were made. We have done that, you know, before we created these videos, we had the basic technology for it and took a couple of months to develop our own technology.
[00:24:20] And the original product was essentially that we would be able to upload a video and, then we would find the face on the video. And after that, the algorithm would essentially, you. Um, make the face move to the, to the audio already. So the quality was pretty low, but, it was a, it was a pretty solid proof of concept.
[00:24:43] And, we could start talking to investors and we closed the seed, seed round. And then since then in the last, eight months, we had built the product that we have now. And. Now we are going through, another shift where we are making, you know, the product more resilient. It’ll be faster to create videos.
[00:25:01] It’ll be easier to record, base videos to train them. So it’s a couple of changes taking place. We are also looking into, changing the app, changing the brand and kind of, you know, implementing what we had learned in the last couple of months, from user feedback, from our advisors, from our investors and, and we, you.
[00:25:23] Getting ready to, to be a big business, hopefully. Um, But yeah, in the, in the last two months we have been running a closed beta where we had about a thousand users on the app, creating their videos. And, we have been looking into how they are using the app. Where are they coming from? What industries they are coming from, what kind of, what type of videos they are creating.
[00:25:43] Um, we have gotten a lot of feedback that we could, that we could work with and we could really, you know, figure. What we should be focusing on and what people find the most useful, the most interesting about our business. And, and based on that, we are building our strategy right now. So it has been pretty insightful and I can, provide a lot of value and a lot of insight for us to, to be able to build solid, foundation to, to grow from here.
[00:26:09] Aman (Host): were some of the, since you mentioned it, what were, when you were list getting. Impact from users. Right. And this is a very, like you said, it’s a very, there’s a very high, uh, standard for what is natural of and what is unnatural. Like it’s just the, we have a very low audience for something that looks very inhuman, right?
[00:26:28] Unless it’s by default in human, like a cartoon, we have a very light audience for something that’s, not believable. Um, Tell us about the feedback that you got from customers and, what you learned, something surprising unsurprising,
[00:26:42] Kristof Szabo: Mm-hmm
[00:26:43] Aman (Host): what was more important than you thought or less important than you than you thought?
[00:26:49] Kristof Szabo: Uh, so whenever I look at our product or videos, it’s, you know, for me, it’s always, I only see the mistakes and I only see the problems. Um, and we did have, you know, a very strict policy on, on. What we show the customers. So we have probably about, you know, 50 actors that we have, trained and we have ready, just waiting to be added to the product, but we still see, you know, very minor problems, little artifacts and so on, um, that we just, um, you know, decide not to put it in the app yet.
[00:27:20] So I think that’s one thing that I I’m learned in the, in the last three months, maybe that, you. I kind of, you know, shifted that to the, to the, to being more allowing with what we show the customers, cuz we need to get feedback. We need to see what they are using and we need to really track and analyze how they are re reacting to things, how they are acting and how they’re using things.
[00:27:46] So. Put, you know, always put it in front of an audience, do AB tests. I’m a very big fan of, you know, experiments and just putting things out there. And if it’s not working, just take it back. You know, we are still really, really, really early and, and it doesn’t harm the business. It doesn’t harm the product to show what we have built, just because we are not sure that they rely because we are very, um, judgemental.
[00:28:09] We are. Fast to be, you know, this is not good enough, but if you put it in front of the customers, they might be like, oh, this is actually cool. I can use it. And we have gone on, you know, Um, really like surprisingly, a lot of feedback from customers that, were very happy with the product. They were very excited and that’s always good to hear, but then I’m just always, you know, my second thought is always like, oh, but they don’t see that this is not working yet.
[00:28:35] And we didn’t add that yet. So it’s always,
[00:28:38] Aman (Host): example, give us an example of something that you really thought was like, you know, not good. And then the customers were like, yeah, that, that should be fine. What’s what’s the
[00:28:46] Kristof Szabo: Yeah. Yeah. So we a good one is really like, we have an actor that is this, older man and. A lot of customers just love, love him. And also the team, when we are creating videos in house, they’re just always using Kim and we have a, we have a text tope voice. That’s like a very deep voice and it just works so well with this guy.
[00:29:07] And a lot of customers are using Kim to, to do presentations about, you know, very, I mean, you know, products that I wouldn’t personally use, but, they’re just pretty cool, like, you know, agriculture and, and, and, and, and, and manufacturing. So it’s, it’s always interesting to see these kind of videos and how people are using, our actors in those, scenarios and situations.
[00:29:29] Um, but we have, you know, a lot of, a lot of cool, a lot of surprising videos that our customers are creating, and we. You know, love to see what they can come up with because it’s a creative product and you, you know, if they, if customers come there and then they have a creative idea, it’s always really cool.
[00:29:46] Um, but yeah, I mean, small things as well, like, you know, music, images and video input, it’s just, you know, we had customers that were like, oh, now you have a music integr. Then I’m gonna subscribe. Okay. So it’s, sometimes it’s really surprising, what kind of requests the customers have and what kind of ideas we have.
[00:30:08] And then when the customer comes, we always take them first. So even though I want to build, custom asset management inside the enterprise accounts and the customers come and then they just say, oh, I just wanna upload my logo. All right. Then here you go. Or I just wanna, I just wanna add my custom font.
[00:30:25] I’m like, oh, okay. If that’s more important for you than something so futuristic that you can actually, you know, automatically AI generate video templates and they would just go, oh, I just want my logo. Okay, here you go. And they are happy. So it’s also to, to kind of prioritize and understand their needs and, and that’s, that’s been a major focus on learning, that we have been going through, but, I think we.
[00:30:50] We are really confident. And I can, I can say that I wouldn’t, you know, be like very cocky about it or very confident about it because I always, as I say, always see the problems that we still have, but we generally, we are getting good feedback. We are getting also, very good feedback from the team and they are really happy with what they see from the customers.
[00:31:07] And in general, I think, um, it’s, it’s a, it’s a good basis to, to move forward. Yeah.
[00:31:14] Aman (Host): that’s interesting. And how has, um, How has the fundraising, how has the team been built? So you said you started with three people at you and a couple of your friends. Um, what did your team look like? And, how was, how did the fundraising cycle go?
[00:31:34] Kristof Szabo: Yeah. So. so I think 2018. Yeah. We started with the, with the, with our detection startup, and then we already had some key people that are still with us. So for three years now they have been on our side, um, on the engineering part. So. We have a small hub in Budapest where we have most of the members of our product team, seven, eight people that are working on the product and they have come joined us after, we raised our seed brand.
[00:32:05] So that was around August and September that we recruited them. And, I would say we were lucky enough to have a network that we could just, you know, pick these. Really, you know, high quality and, and, and hardworking people out with the recommendations. So that was, surprisingly easy, but also we had, um, Dominic, my co-founder in Salton working hard and really introducing, an interview process where we really could filter the candidates and understand how they not only would.
[00:32:36] Fit the team on a scale level, but also in a, in a more, you know, fit in the team as a person as well. Um, so that was, I think very, um, very lucky for us that in a couple of months we could build a core team that was able to actually deliver the product, um, early January and, and, you know, not only deliver the product, but also every week come with new.
[00:33:01] You know, buck fixes and, new features. And, and it’s really been impressive in the last, three months or so that every week we are introducing new things, we are introducing fixes. We are adding new features and, and, and, and implementing customer feedback. So I’m really, you know, um, confident with the team and, that we can, you know, build, build on them and rely on them to deliver this product and deliver, new updates.
[00:33:27] Every, every. And on the, on the business side, we have been working with a, a couple of people and very much we are trying to focus on, on PLG strategy. So we are looking into how the product can lead the sales and the sales, the product. So, you know, we are talking about, windows 95 when you received it in, in a box, and then you, you installed it.
[00:33:51] And, that was, you know, 25 years ago. Now we are looking into. Changing from the SARS business to actually, you know, having something that is free to, to use for the users to a certain extent, or at least free to try so that people can get a feel. So they’ll just, you know, spend a hundred dollars on something that they maybe have seen a video on YouTube, but they can actually sign up and start using.
[00:34:18] Create videos, share it with their friends and see how it works. And then if they decide that they need a bigger package and they can move forward then, but until then they are as if, you know, promoting our business, they are kind of, ambassadors that are taking our videos, showing it to their friends because videos are meant to be seen by people.
[00:34:37] So it’s an amazing vehicle of marketing where we don’t have to move our little finger. We just let people use the platform. They will share it with their friends. They will put it on social media. And what people will see is. An amazing AI video. That was by the way, a little border mark created by Colosian, um, and VR, that’s kind of the logic by what we are building our growth strategy, working with, with people that are experienced in growth.
[00:35:05] And we are really looking into, you know, what kind of campaigns can support it, what kind of channels will bring us forward? So, yeah, we’re looking at a couple of very, so yeah, stay tuned. We have a couple of very exciting, campaigns coming in the spring where, where it will really like benefit from this kind of organic growth that people will share and people will, show it to their friends, show it to them at work, share it on social media.
[00:35:31] And then what people will see that it’s an amazing tool for delivering content and delivering value. And by the way, you know, it was created by Colosian. Um, So that’s, that’s why we have not built huge business team, but we are trying to build, a self driving vehicle that is based on the products.
[00:35:49] That’s why we have, the biggest team right now is the product team. And then we have a research team that we have built. Um, yeah, they, they also have been with us for almost two years. Now that I’ve been building this, the algorithms, the technology, and really, you know, experimenting every week on, on what works.
[00:36:07] It has been, you know, endless experiments, but, it brought us to where we are and what technology we have right now. So yeah, we have the, as I said, the core team, um, about, 14 people right now, and we have a lot of, satellite teams that are doing, you know, different things, either user research, content creation, growth management, and so on.
[00:36:30] Aman (Host): One thing that I’ve, noticed since you, since you mentioned it, one thing that I’ve noticed among teams, which have a product team as well as they have to do research, right. Um, how do you keep. Let’s let’s talk about the operational per process of running a machine learning company, right. Where machine learning is the core of the product itself, right.
[00:36:53] It, and machine learning is not something where you can predict with a scrum cycle or something like that. Hey, like, yeah, we have these shipping and continuously improvement or something like that. Right. How has tell us about that lessons on the operational side, where your product roadmap. Is constrained by research, essentially, right?
[00:37:16] Kristof Szabo: I can feel you there. Yeah. It’s how do I start this? So I was feeling that, you know, since I’m on the business side, I always have to keep an eye on the technology and see like where we are, what can I sell? What can I offer the customer?
[00:37:29] Aman (Host): Yeah. Hmm.
[00:37:30] Kristof Szabo: But now we are pretty much, you know, business, um, or now we call it growth and the product and technology or research and development.
[00:37:40] So now it’s three kind of units that we have to keep an eye on and they have to move par, you know, in harmony basically. So we, they have to be aware of what the others are doing, what they can deliver. What the customers are requesting. I think we are finding the, kind of the, the, the, the, the right strategy or the right mechanisms by, you know, revisiting to the customers.
[00:38:05] Um, whenever we have a bug report, I can show you that in the next 24 hours, we will have a bug fixed out there. Um, so it’s really, you know, putting the customer front and the product team. The one that is responsible for getting the feedback, collecting the customer needs, collecting, you know, what is not working, what is working.
[00:38:27] And then they they’re really onto it. And as I mentioned, we have weekly, weekly, versions coming out on the app. So every, every Friday we have a developer version and then that we will check on on the weekend. We will see what
[00:38:40] Aman (Host): but that that’s for the product, right? That’s for the that’s for the, that’s for the tooling, which is around the main, the core machine learning stuff. Right. That’s for the that’s for the wrapper. Of like how people use the us user experience and all that. But how do you keep that? I mean, of course, a lot of changes you have to make to a product are not related to changing the algorithms.
[00:39:01] Right. But rather changing their, the way it’s packaged, of course, like the UX and everything, but. On an operational side. How do you keep these teams in harmony where you don’t have, like, what, what have, what are some tricks or some tips that we could, that other entrepreneurs could use on making this
[00:39:21] Kristof Szabo: Mm-hmm so I can think of a couple of things. I don’t know
[00:39:27] Aman (Host): work or, or, or, or maybe let’s start with some problems that you had to solve for? Yeah.
[00:39:31] Kristof Szabo: Okay. Yeah. Yeah. Um, Yeah, I have a lot of those, so still, still yet to be solved or already solved. But, so I think communication, especially like when we are working as a kind of semi remote or a hybrid company, um, we have the product team in one room usually, but all the other units are, some, you know, in, on the other side of the world.
[00:39:58] So I think communication is really key and I think that’s something that. We are managing very well. And we have pretty, you know, transplant communication every day on what, at the start of the day and what at the end of the day people have done and are doing. So I think daily updates are, are essentially, you know, even if it’s not company wide, but at least, you know, the product team is up to date every day with what the others are doing, what problems they’re solving, how they can help each other, and then also being available, I think.
[00:40:30] Aman (Host): How do, how do machine learning engineers feel about the daily updates?
[00:40:35] Kristof Szabo: It’s it’s really, you know, it’s a startup that they, they put down two paragraphs of what they’re gonna be working, what they’re gonna be testing, some links, some example videos. And then at the end of the day, they do the same. So we have weekly and biweekly calls that have very specific topics. So we have learned that, you know, keeping the agenda is important, especially if we are talking, you know, six people, seven people at the same time.
[00:40:58] So. It has been a problem that some people talk too much. Some people talk too little in a meeting like this. So keeping a strict agenda was important. Also, not only should be heard and to hear what other people have to say, but also to not go over, you know, hours and hours and, and take up somebody else’s time who might not be.
[00:41:18] Important, their listening and so on. So I think that was, we never had a complaint, but we also don’t do, our long meetings or even daily meetings. So we keeping communication in text, as much as possible, but we have weekly schedule calls where we have specific agenda topics that we go over and usually are done.
[00:41:38] You. Never longer than an hour. If it’s a really long call, then it’s gonna be an hour. But, the one and a half and two hour calls are always with the, with the management team where we have, you know, long, long agendas and a lot of things to talk about. I think keeping it short, having the strict agenda, having everyone review the agenda before the meeting and arrive on time on the meeting, having somebody leading the meeting is always important.
[00:42:02] But, I also notice that video calls usually, are finished shorter than, in person calls. Yeah. Something
[00:42:08] Aman (Host): this, this is something that every company these days is facing, you know, and the problem has of meetings has been just there since, since the, since before meetings were invented, right. In corporate, Right. One thing. So I’m gonna throw, I’m gonna throw a wrench into the works in a way,
[00:42:22] Kristof Szabo: Mm-hmm
[00:42:23] Aman (Host): So one thing that I believe now, and I’m just discussing you since we are peers. One thing that I believe now after, you know, talking to so many entrepreneurs and also other projects is that no company should have a machine learning. Engineer team like the machine learn, no company should have a machine learning team.
[00:42:43] And I believe that machine learning engineers should be part of the product team and they should be almost like product engineers, where they report, they report directly to. The product team as a way, like they’re part of the product team, not like, Hey, you, you guys do your science stuff, we’ll build the product.
[00:43:04] And you are like an internal customer, you know, where, and the reason was for that is because I’ve seen that when you’re like three, two or three machine learning engineers in our team, it’s not exactly a team. You kind of just do sitting together and doing your own stuff. Because machine learning is unless, except for the data collection and the data cleaning.
[00:43:27] It’s a heavily individualized sport. Checking out the right algorithms, trying out different, you know, reading papers and whatnot. And usually for startups, at least, I mean, not talking about Google, where you have five people crunching papers independently and working on the same problem, but in startups you have one person trying to fix one problem.
[00:43:49] Right. And, uh, what I believe is that putting these people who should be ideally in a product team, Putting them taking them out of the product team and putting them into a machine learning team or data science team. I feel like it’s, it’s very counterproductive for most startups. That’s, that’s a belief system that I’ve slowly developed as I’ve been in this space.
[00:44:10] What are you, what are your thoughts on that? Or what are, what have your, experiences been on the site?
[00:44:16] Kristof Szabo: Oh, that’s a really, that’s really interesting. I will bring it up at the next meeting, but, um, now that I think about it, I think we are very transparent in the sense that, we don’t have like a slack channel for machine learning. We have research channel, but we have, product people that are part of that channel and they are following the channel and we have a very open communication between.
[00:44:43] The lead in the engineering team and between the, the research team. And they are always up to date on what’s going on, what, you know, what new models need to be integrated and so on. But it’s also part of it is that they are doing the, the, I don’t know, I, you know, I haven’t worked in other AI companies, so I don’t know how it works there, but I imagine we have a very kind of special, technology that needs a lot of experiment, a lot of, tests, a lot of, you know, Something that the, the four engineers, the data engineers, they would sit down and then they would look at it and then they would do five experiments and see which one works best and show it to the team and then get their opinion and then continue the experiments on that line.
[00:45:25] And then come up with something in the end that might take, you know, three months. And then we have a better model that can be integrated by the engineering team. The engineering team, all they do is, you know, take the all day. Yeah. Take the algorithms that are done by the research team and then add it to the, to the, to the UI and through the, the app.
[00:45:46] So it’s kind of, you know, a, a, a ready package that they have to implement that they have to add, and then the users can access it and the users can start using it. So that’s why it’s, I don’t know if it’s a, kind of a special, case for us, but, I see the logic behind having, kind of a separation between the two not to over, over, over.
[00:46:07] I don’t even know the word, not to, over excite the engineering team with a lot of machine learning questions, but, This is kind of the logic, how we developed it, and then we have transparency. So, both teams know what the other one is working on. And, we have weekly updates from both sides. So, so far it has been working just fine.
[00:46:26] And while we are, you know, under 50 employees, I guess it’s, it’s, it’s easier to communicate, but once we go and hopefully start growing, then they might have to change it. They might have to be. Yeah. Um, adapting to, to putting, having the, the machine learning guys, sitting with the product team. I will keep that in mind.
[00:46:48] Aman (Host): Yeah, there’s a, there’s an interesting, story that, that I got reminded of just now where, do you know kayak. Kayak the, the flight search, you know, I heard this interesting story about the founder of kayak, right. And what he did was I think he had a red telephone, right?
[00:47:07] One of those old, old telephones, he had a, he, he had a red telephone and he put that telephone in the middle of the engineering team. Right.
[00:47:16] Kristof Szabo: Mm-hmm
[00:47:17] Aman (Host): and every time a customer had a problem with the product, the call would be transferred to that phone.
[00:47:25] Kristof Szabo: So, yeah.
[00:47:25] Aman (Host): the phone would ring in the middle of all the engineers, coding and everything.
[00:47:29] And somebody would pick it up because the phone wouldn’t stop ringing. And we’d like, okay, what is it about, okay, this is the problem. This is the product. So the customer would, would tell them, and. The idea of like, why, why would you, why would you direct like customer support queries about the product to the engineering team?
[00:47:45] Like why would they bother them instead of just having some other? And it was like, whoa, after the phone rings a couple of times, you know, the problem gets fixed
[00:47:54] Kristof Szabo: For sure.
[00:47:54] Aman (Host): you know, so
[00:47:56] Kristof Szabo: very
[00:47:56] Aman (Host): it’s the best way to solve, focus on the right problems all the time, you know, and not
[00:48:01] Kristof Szabo: Oh, yeah.
[00:48:01] Aman (Host): work on the
[00:48:03] Kristof Szabo: I don’t know. I don’t know if it’s the most efficient way to, for bug reporting, but I can definitely see how that makes it, makes the bug get fixed faster, but
[00:48:13] Aman (Host): Yeah. Well, cool. Thank you so much. Chris, this has been an ex very interesting conversation and I learned a
[00:48:22] Kristof Szabo: absolutely
[00:48:23] Aman (Host): uh, where can people find you? Mm-hmm
[00:48:27] Kristof Szabo: well, they can find me online. We are remote company, so go to Colosian or find me on LinkedIn. I guess you can have some notes for the video and the audio and yeah. Happy to happy to, you know, Product is online available. Try it for free, give it a shot. And then let me know if you, if you find anything to improve and yeah.
[00:48:50] Keep, keep your eyes open. And we have a lot of, interesting updates coming in the next couple of months.
[00:48:56] Aman (Host): Awesome. Thank you so much.
[00:48:58] Kristof Szabo: Absolutely. Thank you.