Latest / Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™ / Casablanca: AI for Authentic Video Calls

Transcript

Casablanca: AI for Authentic Video Calls · Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™

• WEBVTT 00:00:00.000 --> 00:00:07.760 Music. 00:00:07.609 --> 00:00:11.069 Dot io your podcast and youtube 00:00:11.069 --> 00:00:14.749 blog covering the german startup scene with news 00:00:14.749 --> 00:00:20.589 interviews and live events hello 00:00:20.589 --> 00:00:24.649 and welcome everybody as you hear him laughing in the background today i'll 00:00:24.649 --> 00:00:29.929 bring you another startup interview today with caston carsten is an entrepreneur 00:00:29.929 --> 00:00:34.929 um co-founder of Casablanca.ai. We'll talk about it soon. 00:00:35.109 --> 00:00:40.969 He has been an entrepreneur since age 16 and also is the business angel of the year 2024. 00:00:41.809 --> 00:00:45.549 So-called the golden nose because you can sniff pretty good investments. 00:00:46.309 --> 00:00:52.229 But we'll get to this afterwards because this is part one of a two-part recording 00:00:52.229 --> 00:00:56.309 because usually we get through something like 20, 25 questions. 00:00:56.549 --> 00:01:02.129 And I was preparing for this interview, I was getting to almost 50 questions 00:01:02.129 --> 00:01:04.969 and I was like, that may be a little bit long. 00:01:05.129 --> 00:01:08.689 Let's cut this into two recordings. So, Karsten, welcome. 00:01:09.849 --> 00:01:13.789 Yes. Hello. Now from Zanzibar. I'm currently on holiday, but I'm very happy 00:01:13.789 --> 00:01:14.969 to do this interview with you. 00:01:16.109 --> 00:01:19.349 Oh, very nice. You're doing a sundowner at the Africa house? 00:01:19.549 --> 00:01:22.449 No, not the Africa. I'm somewhere on the south coast of Zanzibar. 00:01:23.849 --> 00:01:28.109 And I want to go diving tomorrow. So I'm going to do the sundowner. 00:01:28.269 --> 00:01:32.809 If we finish in time, I'm going to do the sundowner on the jetty here, which is also very nice. 00:01:33.469 --> 00:01:38.849 And tomorrow morning, I'm going to do some diving and hopefully also see some dolphins there. 00:01:39.489 --> 00:01:43.249 For everybody who's not been there, Sanzibar is strongly recommended. 00:01:43.369 --> 00:01:47.149 I spent a part of my honeymoon there with my wife. Wow. 00:01:47.789 --> 00:01:54.349 Yes. We also may tell that this interview is in association with the Business 00:01:54.349 --> 00:02:00.509 Angel event back in 2024 with the German Business Angel Association Band. 00:02:01.589 --> 00:02:07.149 Let's dive right into the interview. You built your first company with 16. 00:02:08.329 --> 00:02:14.669 What did that early experience teach you about entrepreneurship that still applies today? Okay. 00:02:15.712 --> 00:02:19.372 I think the most important lesson for entrepreneurs is let's do it. 00:02:19.752 --> 00:02:26.272 So don't get blocked by whatever rules, obstacles, something that you might 00:02:26.272 --> 00:02:28.332 not have and it would be better if you had it. 00:02:28.512 --> 00:02:32.512 Of course, it's always better if you have more, but if you don't start, you don't start. 00:02:32.792 --> 00:02:38.612 And so the most important thing is once you have a really good idea, you just do it. 00:02:41.312 --> 00:02:45.172 Not quoting an international apparel client here. 00:02:46.692 --> 00:02:51.092 Now, some people, because we do know, we're usually pretty popular, 00:02:51.592 --> 00:02:58.392 8.35 and up, some people will get nostalgic because you sold a company to Atari 00:02:58.392 --> 00:03:01.892 back in the days, very popular. 00:03:02.312 --> 00:03:07.292 And a lot of people will get, oh, yeah, I remember those games with the joystick and all that stuff. 00:03:07.412 --> 00:03:13.052 Oh, yes. Can you tell us a little bit the story of selling Omicron Basics to Atari? 00:03:13.772 --> 00:03:16.692 Because it was a major milestone in your life, right? 00:03:17.392 --> 00:03:22.932 Yes, it was, yes. So the key thing was we invented or I invented new architecture 00:03:22.932 --> 00:03:24.192 for programming languages. 00:03:25.192 --> 00:03:30.052 And so principally new structure, how to build a programming language interpreter, 00:03:31.072 --> 00:03:32.992 and built this together with two friends. 00:03:32.992 --> 00:03:38.112 They wrote most of the code and we were 70 times faster than what atari had at the time, 00:03:38.752 --> 00:03:43.392 for their atari st computer series so that was a new processor from motorola 00:03:43.392 --> 00:03:47.972 with a 68 000 processor very fast processor that was also used by unit packet 00:03:47.972 --> 00:03:50.792 and some other um high-end manufacturers, 00:03:51.392 --> 00:03:57.792 and atari and also the apple mac and atari launched a computer for about 1500 00:03:57.792 --> 00:04:04.412 at the time that was like one-tenth of the price of the unit packet and one-fifth of the Mac price. 00:04:06.392 --> 00:04:10.792 And we launched a new programming language which made it faster than all the 00:04:10.792 --> 00:04:12.932 other machines with the 6,000 to 8,000 processor. 00:04:14.192 --> 00:04:19.172 And we first sold it separately and then Atari offered to buy it from us and 00:04:19.172 --> 00:04:22.252 we had hard discussions and they negotiated very well. 00:04:23.412 --> 00:04:25.632 And in the end, they bought it from 00:04:25.632 --> 00:04:29.092 us and they bundled it with all their machines at the time. Thank you. 00:04:30.324 --> 00:04:35.084 And the processing capacity one computer had there at the time you now have 00:04:35.084 --> 00:04:36.344 on a smartwatch, I assume. 00:04:38.684 --> 00:04:42.384 No, smartphones are far beyond. So when you take a look at… No, smartwatches. 00:04:43.104 --> 00:04:47.244 Smartwatches, yes. So probably, yes. So my comparison actually is when you take 00:04:47.244 --> 00:04:52.864 a look at the computer that brought the Apollo 11 to the moon and you compare it. 00:04:52.984 --> 00:04:56.344 So many people have an iPhone. Now, don't compare it with the iPhone. 00:04:56.684 --> 00:04:59.624 Compare it with the charger of the iPhone. 00:04:59.764 --> 00:05:04.524 And the processor built into the charger that controls the voltage is about 00:05:04.524 --> 00:05:08.364 the compute power of the Apollo 11 computer. 00:05:10.364 --> 00:05:16.024 Okay, I see, I see. Getting to the next milestone in your life, 00:05:16.164 --> 00:05:19.564 FactFinder, scale to 2,000 plus online shops. 00:05:19.864 --> 00:05:25.324 What was the single smartest strategic move you made that led to this success? 00:05:25.324 --> 00:05:29.144 And can you tell us a little bit about what FactFinder is or was? 00:05:29.764 --> 00:05:35.264 Okay, so you mean an additional strategic move just outside, 00:05:35.444 --> 00:05:38.404 just building by far the best search technology at the time. 00:05:40.664 --> 00:05:44.524 Because that was also very important, I think, because else it's difficult to 00:05:44.524 --> 00:05:48.624 sell something if you're a small company and nobody knows you. 00:05:48.844 --> 00:05:53.364 So what did we do? We built a search technology in the year 2000. 00:05:54.044 --> 00:05:58.304 I actually built the prototype with an intern. 00:05:58.304 --> 00:06:01.204 So the intern coded I did the algorithms in six 00:06:01.204 --> 00:06:04.684 weeks and the prototype showed 00:06:04.684 --> 00:06:07.644 that we can solve this the question of 00:06:07.644 --> 00:06:10.664 how to find things that are misspelled better than 00:06:10.664 --> 00:06:13.624 anybody else it was the prototype was very 00:06:13.624 --> 00:06:16.584 slow and we had to make a lot of additional inventions which 00:06:16.584 --> 00:06:19.644 were made mostly by my best programmer armed at 00:06:19.644 --> 00:06:22.684 the time to have it fast enough but 00:06:22.684 --> 00:06:25.384 I could show that we have an 00:06:25.384 --> 00:06:29.864 understanding if that was also ai it was just not machine learning and understanding 00:06:29.864 --> 00:06:35.604 how humans perceive similarity and how we can represent this in the computer 00:06:35.604 --> 00:06:39.184 so that was the new thing in fact finally in the year 2000 two years before 00:06:39.184 --> 00:06:42.824 google launched its did you mean and did you mean from google. 00:06:43.718 --> 00:06:47.278 Had a totally different approach they just asked millions and millions of users, 00:06:47.778 --> 00:06:50.838 did you mean this and then if they said yes they uh 00:06:50.838 --> 00:06:53.778 they they saved the synonym so they saved 00:06:53.778 --> 00:06:56.458 like this is in in fact that and this is in fact 00:06:56.458 --> 00:06:59.138 the other thing and they were very often wrong in the 00:06:59.138 --> 00:07:02.098 beginning but then they improved with all the users 00:07:02.098 --> 00:07:04.958 and that was a totally different approach and we solved it 00:07:04.958 --> 00:07:07.678 from the ground up we didn't have the resources we 00:07:07.678 --> 00:07:11.038 didn't have the millions of and billions of users um so 00:07:11.038 --> 00:07:14.098 we had to do it with ai and we did in the 00:07:14.098 --> 00:07:17.298 year 2000 and we solved the similarity problem with 00:07:17.298 --> 00:07:20.238 for misspelling things so um what was 00:07:20.238 --> 00:07:23.438 the most uh important strategic decision besides the 00:07:23.438 --> 00:07:26.598 very very good product um i think 00:07:26.598 --> 00:07:30.058 the most important thing was that i started uh not offering 00:07:30.058 --> 00:07:33.478 it for sale only but also for rent because we 00:07:33.478 --> 00:07:36.918 were a small company with seven people at the start i had 00:07:36.918 --> 00:07:40.578 some other products before but then i started fact finder and we put a lot of 00:07:40.578 --> 00:07:46.478 resources in that was almost killed us um and um and then people did not trust 00:07:46.478 --> 00:07:54.458 a small company so um if we said it costs up to 60 000 deutsche mark to buy this so that 30 000. 00:07:55.258 --> 00:07:58.778 Um many were still reluctant at the time because e-commerce 00:07:58.778 --> 00:08:03.178 was not as big as now and things were not so important as now and not so expensive 00:08:03.178 --> 00:08:08.218 as now and um so people were reluctant i said okay i can offer this for rent 00:08:08.218 --> 00:08:12.458 you can you can rent it instead of buy it and you pay only one and a half percent 00:08:12.458 --> 00:08:15.538 of this amount per month and um. 00:08:16.379 --> 00:08:21.139 So, of course, so four and a half percent per month, four and a half percent per month. 00:08:21.399 --> 00:08:28.839 So after 22 months, this pays off the full purchase price. So you should switch after a time. 00:08:29.439 --> 00:08:37.199 And we offer until half a year, we offer you to deduct 75% of the rent you have 00:08:37.199 --> 00:08:39.299 paid from the purchase price. 00:08:39.679 --> 00:08:44.539 And so people said, OK, I can decide this. I can risk that and we will test 00:08:44.539 --> 00:08:48.779 it. and we just throw it out and cancel it immediately if it doesn't run in 00:08:48.779 --> 00:08:50.979 practice, if it doesn't run well in practice. 00:08:51.339 --> 00:08:59.979 And so that reduced their threshold a lot and made us also trust us because 00:08:59.979 --> 00:09:03.459 they said, okay, if they are offering it for rent and not for purchase, 00:09:03.719 --> 00:09:05.859 that means that they believe that we will keep it. 00:09:06.879 --> 00:09:13.559 So it was in both directions. It was good for trust. And because it was so good, everybody kept it. 00:09:14.299 --> 00:09:17.399 And some purchased immediately after a month and some actually 00:09:17.399 --> 00:09:20.919 tended to not buy it because they 00:09:20.919 --> 00:09:25.879 were CEOs who did not own the company and so they said it's better for my bonus 00:09:25.879 --> 00:09:30.999 if this year I don't have the full purchase price on my expenses and just the 00:09:30.999 --> 00:09:35.579 rent and that's how I kind of invented ZAS but we didn't call it ZAS, 00:09:35.699 --> 00:09:37.439 we called it software rent at the time, 00:09:39.979 --> 00:09:44.559 that was quite new Because software as a service was not usual at the time. 00:09:44.679 --> 00:09:45.759 That was very, very uncommon. 00:09:47.379 --> 00:09:48.359 I see. 00:09:50.019 --> 00:09:55.959 You obviously do a lot with Omicron Basic and FactFinder. 00:09:56.239 --> 00:10:02.979 You had two huge successes early on. And you also do have 25 patents pending. 00:10:03.879 --> 00:10:10.899 I would be curious. Some pending, some still granted, yes. Ah, okay, okay. 00:10:13.159 --> 00:10:18.559 How do you get to all of them? And if you could just pick one that would change 00:10:18.559 --> 00:10:20.859 an entire industry, which one would it be? 00:10:21.559 --> 00:10:25.859 So actually, I have lots of ideas which just keep coming when I'm under the 00:10:25.859 --> 00:10:30.259 shower, going through the park, when I'm sleeping. 00:10:30.499 --> 00:10:34.579 I suddenly wake up at night, have to write something down, then fall asleep again. 00:10:35.299 --> 00:10:39.179 And sometimes I make a pattern out of it. not all of the ideas are worth a patent 00:10:39.179 --> 00:10:42.499 because some are just rubbish commercially, 00:10:42.879 --> 00:10:46.359 some are also not good ideas technically, and some have already, 00:10:46.699 --> 00:10:51.439 many have already been invented by other people and just not put into practice by them. 00:10:51.779 --> 00:10:58.379 So I have patents, for example, a granted patent for a new, for a new online, 00:10:58.559 --> 00:11:03.079 sorry, on-screen keyboard, which will make significantly, 00:11:03.758 --> 00:11:10.098 less miscorrections. So very often you type something, you mistype and you get 00:11:10.098 --> 00:11:14.318 a correction and the correction is actually worse than what you mistyped. 00:11:15.198 --> 00:11:22.538 And I have a granted patent to significantly improve this by some additional 00:11:22.538 --> 00:11:27.798 measurements which I do because I get more of the intent the user has rather 00:11:27.798 --> 00:11:31.618 than just the characters he types, he or she types. 00:11:32.718 --> 00:11:37.478 So that's a granted patent where I have not actually started developing yet 00:11:37.478 --> 00:11:41.858 because I don't have the time to build up a new company right now for that. 00:11:42.598 --> 00:11:50.278 I have other patents also for understanding if a picture is reality or if it 00:11:50.278 --> 00:11:56.978 has been manipulated, which also applies for pictures which are in printed magazines. 00:11:56.998 --> 00:11:59.938 So you can use this on printed pictures as well. 00:12:01.538 --> 00:12:05.038 And there are several other things where I have not built a company yet, 00:12:05.078 --> 00:12:08.138 but they also might do a lot of good for the world. 00:12:09.818 --> 00:12:13.318 So which is the one pattern that can change entire industries? 00:12:13.338 --> 00:12:16.938 I think most of my patterns don't go for one industry. 00:12:17.678 --> 00:12:22.938 So like when you look at Casablanca, it has the potential to disrupt video communication throughout the world. 00:12:23.278 --> 00:12:27.478 So in all industries, but of course, video communication is just a small part 00:12:27.478 --> 00:12:34.518 of the work, mostly of the work for internal meetings, sales, recruiting. 00:12:34.838 --> 00:12:40.158 But it's not for, for example, for production, not for fabrication. 00:12:40.578 --> 00:12:44.738 It's not for programming. It's not for all these things that produce the things. 00:12:44.918 --> 00:12:49.578 It's just for the communication. But of course, communication is also an important part of our lives. 00:12:49.778 --> 00:12:55.718 So Casablanca can disrupt video communication, but it will not be limited to one industry. 00:12:55.878 --> 00:12:59.798 Same goes for things like this better on-screen keyboard. 00:12:59.998 --> 00:13:02.418 It's also not for one industry, it's for many industries. 00:13:04.358 --> 00:13:12.158 I see. Given that you have so many ideas, built so many successful companies, when? 00:13:16.643 --> 00:13:21.983 When do you decide, now it's the time to implement one idea or step away from 00:13:21.983 --> 00:13:26.863 something and sell it or give it to another leadership team? 00:13:27.223 --> 00:13:33.323 That depends. So in some cases, I have an idea and suddenly I find somebody 00:13:33.323 --> 00:13:37.643 who would be the right person to implement that idea. 00:13:38.243 --> 00:13:43.463 And if I can get hold of her or him, then I start a new company with them. 00:13:44.763 --> 00:13:48.143 So that can be one point when I start. 00:13:48.843 --> 00:13:54.283 With Casablanca it was different, so I decided I sell FactFinder in order to start Casablanca. 00:13:57.343 --> 00:14:02.383 I wrote the patent, I registered the patent, I still had not sold a FactFinder yet. 00:14:03.823 --> 00:14:08.143 But I knew I needed to shift my attention to that project because it's a huge project. 00:14:08.303 --> 00:14:12.363 It's a huge work also to get it done. And I underestimated it, 00:14:12.443 --> 00:14:13.863 even though I thought it was a lot of work. 00:14:14.323 --> 00:14:18.963 I thought it would be like one year for five people. And it took us four years 00:14:18.963 --> 00:14:21.903 with five researchers plus me, so six people in total maybe. 00:14:22.503 --> 00:14:25.423 But the researchers are doing more of the new patents than I did. 00:14:27.843 --> 00:14:33.043 So in total, we are 20 now with Casablanca, so including people like marketing 00:14:33.043 --> 00:14:36.143 and administration people and so on. 00:14:37.203 --> 00:14:41.623 And besides the researchers we also have developers now because we wanted to 00:14:41.623 --> 00:14:44.503 implement the software and not only build the technology, 00:14:45.703 --> 00:14:54.163 so sometimes I have to dedicatedly start a new company and sometimes I can just 00:14:54.163 --> 00:15:00.963 pick the chance and do something so I just started the biggest German AI. 00:15:03.443 --> 00:15:05.823 What's the thing in English uh, 00:15:07.773 --> 00:15:16.653 I just started the biggest German AI directory called alles-ki.com. 00:15:17.013 --> 00:15:23.433 And I had this idea several months ago because I thought the ones I had seen 00:15:23.433 --> 00:15:28.613 in Germany were all so small and so difficult to find things. 00:15:28.793 --> 00:15:33.873 And also the American ones were not so easy to navigate and not so easy to find 00:15:33.873 --> 00:15:35.293 the solution you actually need. 00:15:35.293 --> 00:15:44.033 And with alles.ki with our german ai directory which is also usable in english actually, 00:15:44.773 --> 00:15:50.513 we have built everything on an ai stack and when i had the right people i started 00:15:50.513 --> 00:15:59.453 that was in october and i did some pre-tests before if my ideas work my ai ideas 00:15:59.453 --> 00:16:02.373 but but then i had the right people together. 00:16:02.653 --> 00:16:05.653 And we started with two people. Currently, there are three people in there. 00:16:05.833 --> 00:16:10.553 Besides me, I only put very little work into it. Mostly it's the other people who do it. 00:16:11.533 --> 00:16:16.053 So I mainly had the idea. But so sometimes it's this way. Sometimes it's the 00:16:16.053 --> 00:16:18.653 other way. It depends on the idea. 00:16:20.133 --> 00:16:24.893 Sorry for our audience. We have some very bad internet connections here. 00:16:24.993 --> 00:16:30.013 And actually, I can only understand all of the stuff Carson is saying after 00:16:30.013 --> 00:16:32.253 his file is uploaded. so not during the video. 00:16:32.413 --> 00:16:38.893 Really sorry about that. But my next question is, many startup founders focus on scaling fast. 00:16:39.053 --> 00:16:43.093 What's the biggest mistake they make when trying to scale quickly? 00:16:43.913 --> 00:16:47.553 So actually, I think there are some people who do it excellently. 00:16:48.013 --> 00:16:52.673 And I'm just a mediocre scaler, I would say. 00:16:52.713 --> 00:16:57.653 So I'm not the worst, but also by far not the best. There are people who really know how to do this. 00:16:59.275 --> 00:17:03.175 And so in terms of scaling, I would love to learn more. 00:17:03.575 --> 00:17:08.735 But some mistakes that people make at the very beginning is not being dedicated enough. 00:17:09.615 --> 00:17:13.115 Also, not starting quickly is often a mistake. 00:17:13.395 --> 00:17:19.795 If an idea is somewhere already somewhere in the air, and then you wait until 00:17:19.795 --> 00:17:25.655 you have more resources, more something, more something, then somebody else will just do it. 00:17:25.655 --> 00:17:28.555 So you can save the most time 00:17:28.555 --> 00:17:31.195 in the very beginning of an idea when you're when 00:17:31.195 --> 00:17:34.455 you're the first one to do it of course you will get people 00:17:34.455 --> 00:17:38.175 who copy you and that's there's 00:17:38.175 --> 00:17:41.155 you have to prepare for that and maybe the idea is 00:17:41.155 --> 00:17:44.435 just not working and somebody else does it much better but 00:17:44.435 --> 00:17:48.195 generally it's rather like start fast I 00:17:48.195 --> 00:17:51.095 think that's the that's a big mistake that people wait 00:17:51.095 --> 00:17:53.935 too long and until they start and then 00:17:53.935 --> 00:17:56.835 of course you have to be able to make mistakes and you have to 00:17:56.835 --> 00:17:59.795 understand that everything you do when you build when 00:17:59.795 --> 00:18:03.515 you scale a startup everything is an experiment and that 00:18:03.515 --> 00:18:06.455 means many many things will go wrong not just. 00:18:06.455 --> 00:18:09.995 Some things and it's not so error 00:18:09.995 --> 00:18:13.355 tolerant work in in german companies normally 00:18:13.355 --> 00:18:16.515 is seen as yes i accept that sometimes people also 00:18:16.515 --> 00:18:20.555 make mistakes this is not what we're talking about I'm 00:18:20.555 --> 00:18:23.535 talking about people do experiments every day 00:18:23.535 --> 00:18:26.675 and many of those experiments don't work so 00:18:26.675 --> 00:18:33.015 they don't just sometimes make mistakes they consequently make mistakes and 00:18:33.015 --> 00:18:37.155 you have to accept that many things go wrong so in marketing especially but 00:18:37.155 --> 00:18:42.015 also in lots of other things you just have to accept this what you don't have 00:18:42.015 --> 00:18:45.415 to accept and you shouldn't accept is mistakes out of negligence. 00:18:45.615 --> 00:18:47.675 Mistakes out of laziness and so on. 00:18:48.155 --> 00:18:53.235 But mistakes out of you think this is right, but you were wrong and thinking it's right. 00:18:54.175 --> 00:18:57.315 That's mistakes you have to learn to accept when scaling company, 00:18:57.475 --> 00:19:01.975 else you cannot scale it. So bottom line is do mistakes, but learn from them. 00:19:03.815 --> 00:19:08.195 Yes. But be prepared that many, many things go wrong when you start something. 00:19:09.915 --> 00:19:13.495 Aside from Casablanca if you 00:19:13.495 --> 00:19:17.495 were to launch a startup today what market or problem would you go after? 00:19:18.975 --> 00:19:20.095 So if I would, 00:19:21.331 --> 00:19:26.931 launch something just because I want to launch something, not because I have 00:19:26.931 --> 00:19:31.051 a specific idea, because if I have a specific idea, then I would do that. 00:19:31.591 --> 00:19:37.611 So, but generally I would build something on the basis of AI technology that 00:19:37.611 --> 00:19:43.791 is here and is going to be available easily for the next year. So like LLMs. 00:19:45.431 --> 00:19:48.591 And the problem is of course that everybody does this. 00:19:48.851 --> 00:19:52.931 So what you need to do, and I wrote this in an article also. 00:19:53.071 --> 00:19:56.691 What you need to do is you have to find some niche where you have more expertise 00:19:56.691 --> 00:20:01.891 or more knowledge, more data or something else that other people don't have 00:20:01.891 --> 00:20:05.211 or more market access where other people don't have this. 00:20:05.531 --> 00:20:09.471 So when you build something new on the basis of, let's say, a large bandwidth 00:20:09.471 --> 00:20:15.251 model or some other AI technology, it could also be like object identification or other things, 00:20:15.811 --> 00:20:20.771 but you build it on top of an AI technology that's easy to use and there are 00:20:20.771 --> 00:20:24.471 many AI technologies right now which are easy to use so they have APIs, 00:20:24.811 --> 00:20:28.551 it's easy access, you can use them without much difficulty, 00:20:30.131 --> 00:20:35.171 and then you just go for some niche where you have the better market access 00:20:35.171 --> 00:20:39.391 than anybody else or the better knowledge or some data which nobody else has 00:20:39.391 --> 00:20:43.851 and so on so what I yeah okay so this kind. 00:20:46.331 --> 00:20:50.831 Of company is not a world changing company in in all industries this is going 00:20:50.831 --> 00:20:52.631 to be something that changes a specific, 00:20:53.371 --> 00:20:57.451 industry or a specific topic probably because that's the thing where you have 00:20:57.451 --> 00:21:03.711 more knowledge so um if you are the best person for let's say um uh, 00:21:04.983 --> 00:21:08.363 spices or so then you build something built on 00:21:08.363 --> 00:21:11.223 your knowledge of spices that nobody else has and maybe 00:21:11.223 --> 00:21:15.883 your data on spices that nobody else has and your market access to spice suppliers 00:21:15.883 --> 00:21:22.523 and maybe spice users which nobody else has and spice could be a big market 00:21:22.523 --> 00:21:29.843 but it could also be something very specific like the pumps for fire sprinkling, 00:21:30.083 --> 00:21:36.643 for fire watering or something. So it could be something very specific. 00:21:37.423 --> 00:21:41.963 But you build it fast and you build it in your area where you have better market 00:21:41.963 --> 00:21:44.443 access, better knowledge than anybody else. 00:21:46.043 --> 00:21:49.563 And we just talked about mistakes, doing mistakes. 00:21:50.543 --> 00:21:58.123 And I was curious, looking back, what's the biggest bet you've made that didn't 00:21:58.123 --> 00:22:01.823 work out, and what did you learn from it? You're talking about learning from mistakes. 00:22:02.163 --> 00:22:05.763 Yeah, so there are probably thousands of small mistakes I made, 00:22:05.783 --> 00:22:08.803 and there are a few big ones. 00:22:09.063 --> 00:22:14.363 And I would say the one that cost me the most money, where I went in the wrong 00:22:14.363 --> 00:22:17.023 direction and stopped it completely. 00:22:17.023 --> 00:22:20.443 So not just did a turn or something is 00:22:20.443 --> 00:22:28.103 fact finder travel so the idea in the year 2011 was that we have sold fact finder 00:22:28.103 --> 00:22:32.903 at that time to i don't know like a thousand something e-commerce shops 53 of 00:22:32.903 --> 00:22:37.263 the top of the top 100 in germany and um, 00:22:37.959 --> 00:22:43.259 There were no travel sites on it. 00:22:43.999 --> 00:22:49.439 And the point is when you search for, let's say, an iPhone charger, 00:22:49.719 --> 00:22:54.639 you type in iPhone charger or maybe iPhone charger for iPhone at that time, 11 or something. 00:22:56.239 --> 00:23:05.039 And when you search for holiday and you type in Christmas holiday with my little daughter on the beach. 00:23:06.919 --> 00:23:09.779 Then what you 00:23:09.779 --> 00:23:15.859 would get with a standard keyword search is something where Christmas appears 00:23:15.859 --> 00:23:20.339 in the text daughter appears in the text also beach appears in the text and 00:23:20.339 --> 00:23:25.839 Christmas so the thing is keyword search brings you the wrong results because 00:23:25.839 --> 00:23:27.839 that also might be something like. 00:23:29.479 --> 00:23:32.979 Clothes on Christmas and the word Christmas is still in there, 00:23:33.619 --> 00:23:39.059 so you would not want that so what you need is you need understanding of the 00:23:39.059 --> 00:23:43.359 text and we produced the world first semantic search for the tourism industry 00:23:43.359 --> 00:23:47.879 which really worked so we could find those things so we thought okay so beach 00:23:47.879 --> 00:23:50.139 probably means swimming, swimming means warm, 00:23:50.359 --> 00:23:54.819 where is it warm over Christmas and we reasoned these things at the time so 00:23:54.819 --> 00:23:59.679 that was over 10 years ago and launched the product and the industry actually 00:23:59.679 --> 00:24:01.639 was enthusiastic about us So 00:24:01.639 --> 00:24:09.459 we got six customers within a few months who built it into their portals. 00:24:09.699 --> 00:24:12.959 And what happened was the end users did not use it. 00:24:13.679 --> 00:24:19.999 98% of the end users just typed in one word, one word like Mallorca or Caribbean or something. 00:24:20.879 --> 00:24:24.519 And they did not type in long sentence because at that time they had just been 00:24:24.519 --> 00:24:29.699 trained by Google that the best thing you can do for searching is type just one word. 00:24:30.979 --> 00:24:37.339 And so they didn't understand what it was about and none of our travel portals, 00:24:38.279 --> 00:24:41.079 really trained the users to change their behavior. 00:24:43.174 --> 00:24:49.354 So we were only participating in success, so things that were booked at a higher rate than normally. 00:24:50.214 --> 00:24:53.794 So if it was just as good as normal, we would not get anything. 00:24:54.454 --> 00:24:59.594 And in the tests, we made tests in the usability studio. People were much happier 00:24:59.594 --> 00:25:01.934 about our search, but after we explained it. 00:25:03.674 --> 00:25:07.874 And so in Effect Final Travel, the end users didn't use it. 00:25:07.914 --> 00:25:12.414 We got a number of prizes for the software. So we won like the innovation prize 00:25:12.414 --> 00:25:16.754 of Deutsche Börse for the most promising startup, 00:25:16.934 --> 00:25:24.914 which is going to be an IPO soon and so on. 00:25:25.314 --> 00:25:29.574 So we got a number of prizes for it, for the innovation we made, 00:25:29.614 --> 00:25:31.654 because it was really disruptive innovation again. 00:25:32.314 --> 00:25:38.694 But we made €50,000 revenue and we invested about €1 million within three years 00:25:38.694 --> 00:25:41.154 and didn't get the market running. 00:25:41.514 --> 00:25:46.334 So we tried with some press releases and the press people just published the 00:25:46.334 --> 00:25:51.174 funny stories but never published like, this is a game changer for you. 00:25:51.614 --> 00:25:56.434 And we didn't have the money to really do TV advertising or so else we might have changed it. 00:25:58.294 --> 00:26:04.934 So that was hard. and I stopped it after spending roughly $1 million in three years. 00:26:05.994 --> 00:26:13.274 Okay. Talked about stopping that funding, talked about big decision management decisions. 00:26:14.434 --> 00:26:18.554 What is your go-to decision-making? 00:26:18.814 --> 00:26:24.594 Is it gut instinct? Is it data? Or do you have a certain rigid framework where 00:26:24.594 --> 00:26:25.454 you make your decisions? 00:26:25.834 --> 00:26:35.154 Yeah, all of them. So I do have two kinds of frameworks and one sentence to remember. 00:26:35.414 --> 00:26:40.714 So if you do a big decision and you have thought about it for a while, 00:26:41.074 --> 00:26:49.714 then the one sentence is never without your brains, never against your gut. 00:26:50.254 --> 00:26:53.574 Yep. So that sounds reasonable. if you 00:26:53.574 --> 00:26:56.634 cannot find a plausible reason so like when you calculate 00:26:56.634 --> 00:26:59.254 it so that you should do it you shouldn't do 00:26:59.254 --> 00:27:02.454 it so you have to have some reasoning why it 00:27:02.454 --> 00:27:06.854 should be done and if it still feels bad even though you have the reason then 00:27:06.854 --> 00:27:13.914 don't do it because your the gut feeling takes and takes in a lot more um factors 00:27:13.914 --> 00:27:19.314 and you take the factors uh which you you can't you cannot reason over 20 factors, 00:27:19.614 --> 00:27:25.454 but you can actually take your gut feeling for 20 or 50 factors because it will 00:27:25.454 --> 00:27:26.754 take all this into account. 00:27:27.014 --> 00:27:32.234 We'll be back after a short ad break and then we'll talk about Casablanca AI 00:27:32.234 --> 00:27:36.154 technology, AI differentiation and your future vision. 00:27:41.686 --> 00:27:46.606 Okay, guys, thanks for sticking around. Welcome back. Now we are talking with 00:27:46.606 --> 00:27:48.466 the Business Angel of the Year 2024, 00:27:48.746 --> 00:27:56.046 Carsten Krauss, currently in the first of two interviews about his most exciting 00:27:56.046 --> 00:27:58.686 startup right now that he does himself. 00:27:58.686 --> 00:28:05.986 It's called Casablanca.ai and it claims to be 20 times more efficient than Microsoft or Nvidia. 00:28:06.306 --> 00:28:08.446 What does it do and what's the secret? 00:28:09.806 --> 00:28:14.166 Okay so what does it do is simple so when 00:28:14.166 --> 00:28:17.046 you do a video call normally don't look at each other you 00:28:17.046 --> 00:28:19.786 normally look somewhere down on the side and so 00:28:19.786 --> 00:28:22.706 on because you look at the other person on screen 00:28:22.706 --> 00:28:25.906 you want to see their reactions but the camera is 00:28:25.906 --> 00:28:30.746 somewhere else it's not behind their eyes their eyes where it should be so what 00:28:30.746 --> 00:28:34.826 we do is we put a virtual camera behind the eyes no hardware it's just software 00:28:34.826 --> 00:28:38.846 we put a virtual camera behind the eyes of the person you're talking to and 00:28:38.846 --> 00:28:45.506 our software produces the picture in a way as if the camera would look from that perspective. 00:28:46.526 --> 00:28:50.966 And what's interesting for me, it also enables you to do some eye tracking, 00:28:51.006 --> 00:28:56.386 which you can also do with other competitors always to what's the camera. 00:28:56.686 --> 00:29:01.866 So for recording, it's pretty interesting, but we have to admit it's made it 00:29:01.866 --> 00:29:04.346 a little bit more for saving. 00:29:05.466 --> 00:29:08.606 Bandwidth so therefore it's not for a 00:29:08.606 --> 00:29:12.386 high for for high resolution recordings 00:29:12.386 --> 00:29:20.706 like we're doing now it's not made for that yet okay yes so generally it's the 00:29:20.706 --> 00:29:26.266 the Casablanca model is a we need to keep the model small in order to make it 00:29:26.266 --> 00:29:32.346 fast enough that you can use it without an extra GPU on a normal current computer. 00:29:32.586 --> 00:29:40.526 So an Apple M1 or a Pentium starting from end of 2022, beginning of 2023 is enough. 00:29:41.526 --> 00:29:47.586 Or an AMD from that time also. So that's why we need to make it small. 00:29:47.986 --> 00:29:56.306 And that means the resolution is not good enough for a high-quality video. 00:29:56.566 --> 00:29:59.726 But in video calls, you always have bad resolution anyway. 00:29:59.746 --> 00:30:03.186 Way so you don't see the difference so when 00:30:03.186 --> 00:30:05.946 when you do a normal video call Casablanca is good 00:30:05.946 --> 00:30:09.286 enough and it's fast enough to run on a normal computer so 00:30:09.286 --> 00:30:12.566 why why are we so fast it's not only because 00:30:12.566 --> 00:30:20.406 we have we we save time on resolution so it's also because we built a very very 00:30:20.406 --> 00:30:25.886 efficient model and we could do that because and we build a much smaller model 00:30:25.886 --> 00:30:30.826 than others do and we could build a smaller model because our model understands the faces better. 00:30:31.703 --> 00:30:34.523 And when you have more understanding, you can compress better. 00:30:34.723 --> 00:30:39.883 So if you have no knowledge of a language, 00:30:40.623 --> 00:30:44.603 then you have just to listen and you have to record the audio and you have to 00:30:44.603 --> 00:30:48.423 remember the audio by heart if you want to repeat it. 00:30:49.043 --> 00:30:52.103 That's very hard. If you understand the language, 00:30:52.303 --> 00:30:57.323 you can compress the audio to words and that's much less information bits you 00:30:57.323 --> 00:31:04.523 have to put and you can just put much longer texts into the same kind of sequence. 00:31:04.663 --> 00:31:08.203 And our model understands faces better than any other model. 00:31:08.443 --> 00:31:15.983 And that's mainly because we have built a self-supervised approach and we have 00:31:15.983 --> 00:31:20.763 built a new way how to build face models. 00:31:21.383 --> 00:31:27.843 And thus we can make a very small model and this very small model also allows us to make it very fast. 00:31:28.763 --> 00:31:32.503 And we will also launch higher resolution models, which will also be faster 00:31:32.503 --> 00:31:33.943 than all other high resolution models. 00:31:34.103 --> 00:31:39.283 But they will probably require a GPU if you want to use them on your personal computer. 00:31:46.883 --> 00:31:53.883 Your AI technology also enables a realistic 3D phase reconstruction using minimal 00:31:53.883 --> 00:31:56.163 data. How does this work? 00:31:56.723 --> 00:32:00.203 So that's part of our small phase model. So it's a number of new inventions. 00:32:00.423 --> 00:32:04.663 So we also made a new GAN, which doesn't need perceptual loss, 00:32:04.743 --> 00:32:09.763 which makes it faster and at the same time better in quality, actually. 00:32:10.803 --> 00:32:14.283 And we made a lot of inventions and registered a number of patents. 00:32:14.463 --> 00:32:16.803 The first one has been granted. The others are still pending. 00:32:18.723 --> 00:32:25.223 However, so to sum this up, what we do is we can compress a video picture of 00:32:25.223 --> 00:32:32.503 a person. So all their expressions, everything they do in a vector of less than 300,000. 00:32:32.927 --> 00:32:41.047 Data points so 300 dimensions of the vector and so if we have 25 of these vectors per second. 00:32:41.747 --> 00:32:45.687 That would be seven and a half thousand data points and that's 00:32:45.687 --> 00:32:48.427 fast enough for making a video call with this 00:32:48.427 --> 00:32:51.707 compressed data over an 00:32:51.707 --> 00:32:55.727 edge communication so there's a 2g communication that 00:32:55.727 --> 00:32:58.607 will be it will be fast enough for that if both 00:32:58.607 --> 00:33:01.627 sides have casablanca and it would be a direct 00:33:01.627 --> 00:33:04.567 communication so it's currently not integrated into 00:33:04.567 --> 00:33:07.427 this kind of communication into zoom or teams or anyone 00:33:07.427 --> 00:33:10.747 but if they would integrate it they could enable video 00:33:10.747 --> 00:33:15.587 calls also in areas like where I am now where the internet connection is not 00:33:15.587 --> 00:33:19.307 so good or Germany where the internet connection is also not very good outside 00:33:19.307 --> 00:33:24.867 the big cities so very often you only have edge connection and you always say 00:33:24.867 --> 00:33:28.147 oh I can't do anything on the internet but now with So Casablanca, 00:33:28.147 --> 00:33:32.947 if both sides had this technology and it would be built into video communication, 00:33:33.227 --> 00:33:36.907 you could do video calls even with only this batch connection. 00:33:37.127 --> 00:33:42.667 Every major tech company and non-tech company is investing in AI, 00:33:42.907 --> 00:33:46.107 especially talking about AI-driven video communications. 00:33:46.467 --> 00:33:48.907 What makes Casablanca stand out there? 00:33:49.427 --> 00:33:55.547 Well, the thing we are solving, we solve best in the world. So making authentic 00:33:55.547 --> 00:33:59.567 communication with a full face turning and eyes turning. 00:33:59.827 --> 00:34:04.927 So all the competitors currently can only correct the eyes, but Casablanca corrects 00:34:04.927 --> 00:34:10.207 the full face so that you have an authentic communication as if the camera would 00:34:10.207 --> 00:34:13.007 be behind the eyes of the person you're talking to on screen. 00:34:14.067 --> 00:34:19.647 So we're the only ones who can do this at the moment. That's what we do better. 00:34:19.647 --> 00:34:25.127 But then how do we do it better by a lot of inventions again some have been 00:34:25.127 --> 00:34:29.747 made by me but most have been done by my by my team of researchers and that's 00:34:29.747 --> 00:34:31.007 actually one of my talents, 00:34:31.527 --> 00:34:34.327 so i start with an idea and then i find 00:34:34.327 --> 00:34:37.107 people who understand the details even better 00:34:37.107 --> 00:34:39.827 than me and i can understand if they 00:34:39.827 --> 00:34:42.887 do or they can't because mostly the ceo 00:34:42.887 --> 00:34:48.807 does not understand enough technology to to realize if the person who claims 00:34:48.807 --> 00:34:55.907 to be an ai expert is just okay with ai or really good and i'm one of the few 00:34:55.907 --> 00:34:59.987 ceos who can do that and i've done this all my life so all i've, 00:35:00.727 --> 00:35:03.427 mostly i have made some part of the invention myself, 00:35:04.166 --> 00:35:10.366 uh where you needed a lot of things to come in together but for um that that 00:35:10.366 --> 00:35:15.866 does not complete the product so also in fact finder as i told you the um we 00:35:15.866 --> 00:35:17.726 started out with this very slow, 00:35:18.406 --> 00:35:24.186 um implementation which already showed that the idea of how to how to represent 00:35:24.186 --> 00:35:29.766 similarity in search is really understood well but in order to get the speed 00:35:29.766 --> 00:35:33.366 somebody else had to do it and there had to be a lot of additional inventions. 00:35:33.606 --> 00:35:39.446 And I find those people who make the details, who it's not a small detail, 00:35:39.606 --> 00:35:41.706 it's an important detail. Speed is essential. 00:35:42.506 --> 00:35:46.986 And I find those people and I can decide if they are really so good or not. 00:35:47.366 --> 00:35:51.166 And I bring them together and I make them work together and I make them work 00:35:51.166 --> 00:35:53.626 with me also, but also work together very well. 00:35:54.286 --> 00:35:56.866 So I think that's my secret sauce here. 00:35:57.866 --> 00:35:59.966 I've done this all my life in all my companies. 00:36:03.566 --> 00:36:10.526 Apple's FaceTime now offers an eye-powered gaze correction. Why is Casablanca a better solution here? 00:36:11.086 --> 00:36:14.666 As far as I know, Apple only corrects the eyes again. 00:36:15.126 --> 00:36:20.526 So that means it works well for a small angle, if you have a very small angle. 00:36:20.666 --> 00:36:24.426 So you're like on a smartphone, normally you're like this far away, 00:36:24.646 --> 00:36:32.586 and you talk, and then the angle between the eyes and the, 00:36:33.486 --> 00:36:39.426 And the camera is very small. And if you correct only by a small angle, just the eyes. 00:36:39.606 --> 00:36:41.886 So I'm trying to simulate this now. 00:36:42.286 --> 00:36:46.566 So if you do a small angle, it looks good. If you do a big angle, 00:36:46.726 --> 00:36:49.166 it looks crazy if you just correct the eyes. 00:36:49.526 --> 00:36:55.966 And so it looks like this is a different emotion you're showing. Yes, yeah. 00:36:56.406 --> 00:36:59.486 Yeah, so yeah, maybe that. But even if they don't appear on your head, 00:36:59.626 --> 00:37:05.826 if they're just correct, But just looking at somebody from the side is like, 00:37:06.126 --> 00:37:11.966 oh, he's giving me a side-eye view, so what does he want to say? Is something wrong? 00:37:12.746 --> 00:37:18.766 Is he not believing me? That's also a mimical expression when you look from 00:37:18.766 --> 00:37:21.066 the side. And you don't want this. 00:37:22.246 --> 00:37:25.426 You want the true mimical expression. That's the essence of communication. 00:37:25.426 --> 00:37:31.266 We humans are so fine in reading all the small movements of the mouth, of the eyes. 00:37:31.506 --> 00:37:38.566 If you move the lower eyelid half a millimeter down, people say, oh, now he's skeptical. 00:37:39.566 --> 00:37:44.806 Or he's ironic when you move it further down. So there are some things which 00:37:44.806 --> 00:37:51.526 are very small mimic changes, making an expression. 00:37:51.846 --> 00:37:56.546 And you want to avoid this. And Casablanca changes or corrects the full face 00:37:56.546 --> 00:38:00.706 so that you don't have this involuntary eye expressions. 00:38:00.866 --> 00:38:04.826 And it works well when you just correct the eyes if the angle is really small 00:38:04.826 --> 00:38:08.626 and it does not work well when the angle is bigger, like when you have an additional 00:38:08.626 --> 00:38:10.826 screen or even just a big notebook screen. 00:38:12.746 --> 00:38:19.606 I was wondering, seeing that you put so much thought, so much energy and so 00:38:19.606 --> 00:38:24.646 much future potential in this company. What is your ultimate vision for Casablanca AI? 00:38:24.966 --> 00:38:28.226 Is it an IPO, an acquisition, or a long-term standalone company? 00:38:29.084 --> 00:38:34.084 So the first vision is like in five years, nobody will remember how video calls 00:38:34.084 --> 00:38:37.804 were before when you had no eye contact. 00:38:38.484 --> 00:38:43.704 So I want to improve the world here. So, I mean, people will probably remember, 00:38:43.744 --> 00:38:46.424 but they say, oh, yes, long ago it wasn't like this. 00:38:49.744 --> 00:38:56.684 And so that's my first vision. And then I think I will go with what is a good 00:38:56.684 --> 00:39:03.924 offer. So at a certain size of the company, I think other people can run it better than me. 00:39:06.104 --> 00:39:11.384 And so right now I'm open to everything, but we are currently in a seed phase. 00:39:11.404 --> 00:39:12.944 So we're taking up angel investors. 00:39:13.264 --> 00:39:16.064 So far, I invested about two and a half million euros myself. 00:39:17.224 --> 00:39:22.824 And now we're currently gathering one million for the seed phase, 00:39:22.904 --> 00:39:27.224 but we also want very good angels in there. so people who can also advise us 00:39:27.224 --> 00:39:30.144 and also open us some doors and so on. 00:39:31.184 --> 00:39:36.504 And then probably like around the end of the year or so, we will go into a VC phase. 00:39:36.604 --> 00:39:41.484 And then probably it's more the VC who decides where we are going because they 00:39:41.484 --> 00:39:45.764 know better how to make the most money-wise out of it. It could be all of them. 00:39:47.584 --> 00:39:53.464 Anybody who's now interested about working with Keterblanke.ai or investing 00:39:53.464 --> 00:39:56.944 into it, they can reach out to you on your LinkedIn profile, 00:39:57.164 --> 00:39:58.924 which we link down here in the show notes? 00:39:59.844 --> 00:40:03.164 Yes, please do. So we already have some people we're discussing with now, 00:40:03.204 --> 00:40:06.604 so it will not be endless the time we will have. 00:40:06.764 --> 00:40:11.944 But we are also very interested to have international people because internationalization 00:40:11.944 --> 00:40:17.444 is one of the topics I personally would love to have somebody who advises us 00:40:17.444 --> 00:40:19.984 on this much better than I know myself. 00:40:21.364 --> 00:40:25.284 I see. Awesome. Closing words to our audience. 00:40:25.564 --> 00:40:29.624 What questions do you think we missed? Let us know in the comments down here. 00:40:30.424 --> 00:40:34.004 Karsen, thank you very much for the first interview. For everybody, 00:40:34.324 --> 00:40:38.144 it will be one of the next interviews published. 00:40:38.464 --> 00:40:41.964 Your interview number two. And we only take like five minutes break. 00:40:47.384 --> 00:40:56.564 That's all, folks. Find more news, streams, events, and interviews at www.startuprad.io. 00:40:56.904 --> 00:40:59.064 Remember, sharing is caring. 00:40:59.760 --> 00:41:12.646 Music.