Yesterday we all enjoyed the keynote of his co-found founder adrian lorca, and today we welcome the cto of megantics, the ambitious berlin-based venture studio, hes, also co-founder and board member of the german association of ai companies hes one of the masterminds behind hex zurich, the hackathon, the Biggest hackathon in europe – and he is among many other things – a
Robocup junior world champion, please give it up for rasmus voter thanks for the kind introduction um so im going to talk about tech trends and business model trends around ai, and i hope you take something with you. After this talk, i will talk about three different pillars. As part of this talk, but what actually qualifies me to talk about this so were a venture studio. Marantix were building ai first companies here in berlin. So what we do is we take founders or potential founders.
With us work with them for six to twelve months test and validate a lot of business ideas with them, then incubate a company with them. We invest even up to three million euros per company and then also help these companies to scale weve been doing this. For the last four and a half years and actually started a few ai companies, so, for example, vara in the space of medical imaging that is right now live with the medical product um helping in this in the case of breast cancer.
Diagnostics and also see a search that is giving a talk today here at rise of ai, which is focused on on data management, on handling large uh, unstructured data sets, and finally, also mirantix labs, which are where nicole also will give a talk today, which is around Uh really building customized ai solutions for clients, and recently we actually started two new
Companies uh kausa and cambrian kausa, focusing on explaining kpi changes, so the the ai buzzword around it is a causality and causal inference and then cambrium, which is at the intersection of biology and machine learning, and we actually plan to build 10 more companies in the next. Four years, and so weve just started two companies and we are planning to build eight more companies and that obviously poses the question. Uh
What areas are we looking at and, as i mentioned before, with these entrepreneurs in residence with these founders, we work six to twelve months. We test and validate a lot of potential business ideas around ai and so weve made quite some learnings over the last couple of years, but we have also looking forward a few areas. We are particularly interested and that’s. What i want to talk about, i want to share some of the things weve been looking at or we are looking at at the moment and where we believe
there could be a lot of potential for for building ai first companies i want to structure the talk actually in three parts so i will talk about surely technological achievements as a second thing that is maybe the most obvious so seeing where whereas research also um advanced but before that even focus more on kind of industry trends understanding what are industries spaces where it could be interesting to build a.
Company, and then the last part which is, a bit overlooked sometimes is also business models focused all around ai that are very particular to. To this and um yeah starting starting with the industry trends. Um i want to talk a bit about first, maybe also the last. I guess decades of innovation. So i think if you look at the startup landscape, weve seen a lot of these uh loan.
Dropout founder paradigms, so the idea, basically, that you have a group of people, two three people who maybe dropped out of college or dropped out of a company and then started started, started working on a company, um and so weve. Seen a lot of these kind of more execution, heavy b2c business models, where you know you need great execution, theyre very hard to build. You need to usually raise a lot of capital, but also
You need relatively narrow domain expertise. I mean many of those people um, you know come right out of university or just have a few years of work, experience and um. I think that’s that’s what weve seen in the last decades, especially also very much in berlin, and i guess one trend which weve also seen is that things get democratized. So if
you think about um how how easy it is now to build an app how easy is now to launch a direct to consumer brand you can get something live within a few weeks and um even even with an ai right like with all the open source and all the tool tools there you can easily get a machine learning model to work within a few few few hours even and um that has really like gradually pushed i guess the innovation um for these kind of more low barrier.
spaces and um what i want to talk about is is kind of the next frontier what we believe is that um there’s going to be next frontier of areas for the next decades 10 or 10 20 years where you can basically start companies which are in space that are much more complex in terms of their research which are touching much more regulatory hurdles um and where there’s still a lot of potential and which are often still very untouched and um this is the.
um you know taken from from from an article from from denny from techcrunch he called it the dual phd problem so basically areas where you not just need to be an expert in ai but you also need to be an expert in another domain um and its not about the phd it could also be 10 years of work experience but like its areas where you have the ai expert and you also have somebody who knows another domain like really really well and its something you cant easily pick up in a few weeks and.
um and for example healthcare is such a space right like if you if you’re great with like training algorithms on data sure that’s that’s awesome but like you also in order to be successful they really need to understand the healthcare sector regulatory the incentive systems um all the medical terms and so for that you need somebody whos been in that area for a few years that’s not something you can.
pick up in a few days and so um these you need both both people coming together um and also understanding each other which usually is the case i mean talking very different languages right and so i think making that work and and going after these dual phd problems is super interesting and that also actually poses yeah some i guess different view on how you build teams in this space so if you think about again the lone founder dropout paradigm.
usually a group of people they say hey lets start a company then they look at a lot of ideas and then you start that and that works for for areas where you dont need too much deep domain expertise but um for this dual phe problem actually you you should maybe think about it a bit different and that’s that’s actually also how weve been doing it at mirantix so the way it works with us is that we actually just work with an individual person with an individual potential founder.
then ideate with that person so look at a lot of different ideas and based on the idea that person ultimately will build as a company we actually also assemble the co-founders so um right now for example we have we have one co-founder who will start a company with us and then the space of natural language processing but like hes looking at a lot of different areas.
and depending on where hes going to build the company he will have a very different co-founder and so i think the the way basically ideation or team assembly should work um should be very different this is you can you can also look at this maybe not just in a startup context but potentially also in a corporate innovation context so um yeah and so out of this basically you can also then derive like areas we are currently actively looking at so we are looking at a lot of areas where there is this other hard science you can.
combine and cambrian one of the companies we started this summer is in the space of synthetic biology so trying to create new biomaterials using also data and machine learning to to optimize those and that’s a typical area right like you like with with your like middle or high school biology knowledge you wont know enough you need somebody whos really deep in that field but you also need somebody who works with data and like making them talk together and that was.
also an interesting process we saw internally it takes some time until they start speaking the same language but that’s that’s where there is a lot of true potential and then you can also look at other areas like in healthcare medicine neuroscience there are a lot of interesting errors in quantum space um geoengineering so there are a lot of these areas and so if you if you if you’re interested i think i would really urge you to pick one of those areas and dive really deeply there that’s the.
first part now lets talk about technological achievements so far most of our companies have been started more focused on computer vision but we are now actually beyond the computer vision efforts actually also looking at other areas and and actually text so if you think about yourself or and myself including like we spend so much hours reading and writing text were basically like glorified text processing machines if you want to put it very simple like the average worker spends 11 hours a week just on emails reading them and writing them and that doesnt even.
include you writing documents being on messengers communications and all these things so um a lot of the human work is actually nothing but getting text in processing it and then writing text again and so um were super excited about that as a like overarching paradigm and um im just just giving you a recent example this is you know like one application where its about you got a received email you want to draft the response you write.
a few bullets and bomb then it directly presents you like a fully written email that is um you know might have some things you can still optimize but its its crazy what you can already do with with natural language processing in this space and this has really to do with also some recent achievements in natural language processing if you think about the previous paradigm how natural language processing worked and this is also how much of other areas of machine learning.
worked you took a lot of data um you trained some neural network but then you basically fine-tuned it for a specific task so depending on what you want to do whether its understanding i dont know financial statements or whether its understanding something in the insurance space or um other other types of documents you would then fine-tune it towards your your text and your specific use case and that’s that’s not super scalable and that also.
means you need to have sufficiently sufficient amounts of data for your use case um and it kind of limits how flexible your application are and you know as you as many of you see like oh may i publish gpt3 um in early 2020 where they basically had a new paradigm they said okay we we gonna just train this model on all the.
data that is out there in the internet like unlabeled basically all the text you can find there and then this model is generally applicable to many different types of applications sure you could still fine tune it but already with this has basically seen most of the text that is out there and that just is really mind-blowing and i think is is very interesting for looking forward so just showing a few applications so this is something where its around felons filling out a balance sheet um the model contextualizes whats whats happening there and then.
actually automatically fills in the spreadsheet or a second application um you know its about like inferring some facts these are all things which are basically in the model and you can easily then say okay whats the founding date of alaska and then it will directly pump up this number so its its crazy whats whats possible there and another one is around code so you know you just say hey i want this web page it generates that and you can see on the left it even generated the entire code for for this.
webpage so um you can imagine think about this model as as one model that basically consumed all the text that is in the internet and you also see that in the model size right like this gpt3 has 175 billion parameters its huge it costs millions of dollars to train the next model will likely cost tens of millions of dollars to train so its really expensive it needs a heavy computation but once youve trained you can basically apply.
it to a lot of different use cases um and if you think about it the capabilities you can you can buy these models so there there are plenty right there is text classification there’s information extraction semantic analysis semantic search um which is more on the understanding side so basically taking a text and understanding it you can also you know take the next level of conversational layer you can start summarizing text i mean think about how many people spend so much time summarizing content into like a more condensed.
format language translation especially if you think about languages that are more in the long term so basically not english code translation so translating code between different languages and then also on the generation side so you can also think about writing text and that could be um kind of serving more the long tails so think about like writing 100 000 product description right like if you have an ecommerce.
store like right now a lot of that is done still manually but given given some of the structure information you could automate that so there are tons of applications and um just to list a few were basically you’re right now looking at a few of quite a few of them um we are also running already some pilot with with customers with first versions of a product in one and two areas so this is really like one space where we will generally um go quite big on in the next one one to two years and you you can look at the.
financial markets you can look at um you know hr you can look at sales you can look at code generally there’s there’s plenty of opportunities now the last one is the the business model part so this is this is very important and as i said often often very overlooked um one thing we actually did with morantix together with the university of san gallanhasky is to also further develop the business.
model canvas so we took the you know very infamous business model canvas and said okay what do we need to add to also make it applicable for ai because a lot of things are similar but also some things are different right for some things um like data like safety like ai lifecycle are very specific uh elements you need to think about when you when you build an ai application and.
so this is also way how we internally think about different areas to look at before we start incubating a company so feel free to check this out we published this a few weeks ago and we will further develop that in the coming months no but like what i really want to talk about is standardization and um types of oblique applications so you know one one one part of ai application is is the vertical products that’s kind of what vara or medical imaging company does breast cancer is a.
global problem of course its relatively standardized globally so you can really build one vertical product that serves the entire market and you can go market by market and the product is basically the same so you have a very high high degree of standardization then there’s the other end of the spectrum which is basically what mirandis labs also does so any consulting customized solutions where you basically say okay.
you have a very specific problem you want to get a solution for that specific problem and that’s it but whats really interesting and is is this middle ground which is more kind of the the self-service solutions where you say okay um i agree that not everybody will have the exact same problem um but i also dont want to build everything from scratch every time so you’re thinking about like solutions that are um to some degree productize so in our.
case see a search its a data management solution its a product we ship code um like vara in some sense but in the end the customers do very different things with it so one company is using that for their autonomous driving use case another one is using that for their um remote sensing application which are very different applications and we dont care so where we basically say okay there’s a product but like everybody can customize it and that space is super interesting and if you think about how at least we internally saw uh.
thought this the things would look like we thought five years ago we thought okay you know either you build a product um and that’s it and you scale it like a normal you know vertical solution or you say okay you fully customize it um and its its kind of a consulting solution and anything in the middle basically its just like okay there is stuff where its neither nor but like.
whatever lets ignore it um but like what how we see it today and this is this is a bit exaggerated but um i think for a lot of product applications of ai you realize that um its maybe not that standardized across customers than you would think or then you imagine and sure medical imaging maybe autonomous driving there are a few where this is the case but like for many especially when you have to do also with more business processes you realize okay every company is slightly different and or its a manufacturing every factory is a bit different and.
then it starts to become difficult and you’re like okay yeah trying to build a product but also i dont want to customize too much then it doesnt scale and so i think we believe that this middle ground is actually super interesting where you say hey we ship a product but then we give give people a lot of degree to customize it and i think this boundary will even be pushed more so we see that also now with mirandix labs where.
we run right now like individual projects with clients and then after running three four projects with different clients and very different industries and we realized okay wow actually eighty percent of their problem is the same even though their business process looks very different and so especially in language processing where its around processing documents we see okay um actually if you would have given them a self-service solution they could probably customize it and so i think this is this is very interesting um obviously you need to be a bit careful when you when you think.
about the characteristics of these self-service solutions so you need to still make sure that there’s some generalizability across organizations you also need to make sure that kind of what what cloud providers and i guess open source is pushing in the market is is you know it fits nicely in with that because the cloud providers have also very much noticed this and so they put a lot of self-service solution out there whether its computer vision natural language processing recommender system and so on and so they they also pushing very much here and there’s also more and more open.
source which is easier to use so um you know when you when you start trying to build a company in this space you need to make sure that you are somewhat more enterprising probably than or than open source or make sure you add a bit more value but you also maybe not competing directly with the cloud providers because that’s also obviously something where you need to be be a bit careful um you also need to make sure that the capabilities that the customers are are.
there to customize the solution and that depends really also who is there on the customer side so it could be it could be developer it could be somebody like like for zero search data management um were mostly selling that to developers so they can also code they can use the api its a different interface for other applications you might rather choose something like in the no code low code space where you say okay this is.
something were also a business person should be able to customize it you need to make sure there’s data available the customer because differently to vara where we dont need to train again at the customer because its a globally standardized problem in this case i mean because you’re customizing you also need to make sure the data is there in the right quality and then there’s still always this vertical horizontal trade-off in the end.
where you say okay do i add a bit more functionality to my product but then basically can potentially serve a smaller market or do i keep it a bit more shallow but then um you might not provide enough value to the customer um and so i think there are lots of areas where you can look at i mean quality control is a great example where you say okay every factory produces different products so yes you want to ship a visual or acoustic quality control solution which is a self-service solution.
but then the customer can basically adapt it automatically train it with their specific products or predictive maintenance very similar anything around documents i think there’s you know documents are very different in different companies and so you also need some some adoption um and then anything also in the data management ml op space where you say okay we provide tools to enable others to build machine learning solutions so these are all areas were looking at.
business intelligence i mean we just started um causer and then data management ceo search and document understanding um were also looking very very heavily at right now and so there is this id like to wrap up um if you want to join us we have lots of open positions if you want to found a company with us if you want to build algorithms if you want to work on the business or sales.
side um all based here in berlin both full-time and internships and were also opening an ai compass in q1 2021 so with this lets start with some questions thank you very much and weve got some questions already the first one is is it possible to invest as a private person in mirantex unfortunately it was not possible for me to find mirantix equity um we dont take investors into moral.
takes right we have a fund so through that private persons can also also invest but they need to be above a certain threshold so yeah all right and the other question is also about money your total fund how big is it is it 25 million euros it says here yes all right uh and will this be enough for your ambitious projects that’s my question now i i mean right like right now like we are the thing is were a very.
operational model right like we are not were not writing checks but we are actually building those companies from scratch and so with the current setup we believe we can basically start eight more companies in the next three and a half years right but we couldnt build much more and then um so the number of companies with our current setup is a bit limited.
and then the other thing is from series a onwards we are also taking in external investors and funds into the companies so then for the follow-up rounds for example for vara we did a series a earlier this year we also took a few investors from the outside to further capitalize the companies right and i wonder how you pick the companies you want to you want to build up i mean you showed the overview of the topics.
that are particularly hot are interesting at the moment are you looking for companies in these certain sectors or do you see what people offer you and then just shows choose the best ones even if theyre all quantum computing companies for instance no we try to diversify it right so we also and we sometimes even yeah pick founders and residents based on certain areas we want to get in so for example biology was a space we were very excited about and so we looked very focused for a founder who would would start a company in that.
space with us so sometimes its also been targeted we would not start five companies too close to each other because also we want to we want the companies to collaborate very openly with each other and if they would be in competition there would be also it would be a bit harder to collaborate and so that’s why there needs to be like a certain distance between the companies right and this uh this next question might be.
a little philosophical maybe but you uh you talked extensively about natural language processing and there was this great really great example of a long email coming in with business proposal and then you have like four little answers like yes no we only give half of what you want or whatever and then its turned into a big email again and i bet on the other end if this future comes to pass would again be like an ai dumbing it down then expanding it shouldnt we then skip the whole procedure of writing long letters and.
just say 94 million or 940 million or what it was that that might happen it might be that like you just end up sending like structured information bullets and then in the end you have some ai or some model on your side in the end that kind of presents it to you in the right format you want to read in so if you feel like you want to read the full email then you should write as an email.
if you’re like hey i dont want to read all this like text right just summarize it to me in bullets then you could just get the bullet from it so i think and if im a romantic person i could say make it nice and then it adds additional exactly okay that’s that’s very interesting isnt this also for tech savvy people like us if i for instance um rent your solution or use it as a service isnt that a bit cheating i mean for people who really put in the effort.
write the email or think oh man this daniel hes so nice he writes me a long letter yeah i i think we will get used to it i think like right now it feels very weird probably in like 20 years its so normal that you know that like a lot of the text you read is not written by a human like that’s true that’s that’s like a mental concept in our brain that like sure text is generated by humans but like why should it be right like.
sure and if you what do you because yesterday we were we were talking about um like artificial persons and actors maybe that can be put in film do you think if i have a particularly uh good or distinct writing style that this would we will be patented there will be like an umberto echo ai and the heirs will have the rights to his particular style do you see this coming i i think there there will be ways where.
you say okay put that into a specific style of writing or like put that into my style of writing because i agree like everybody has some like a certain style how they write emails right and so its its distinct but i think you can also learn that from your historic email so if you if you would give me all your historic emails we could probably build something that sounds very much like you based on that data so are you using prototypes of this already yeah were working on this in turn its.
more internally not to the outside no okay okay im not i was not right i didnt want to go paranoid like israels email legit no but how do you use it i mean do you just play around with it or are you kidding around like were we have a few prototypes right now were you know running a few things with customers i mean this is the space um the founder residence was looking at this um he started this summer so he was um.
the first first employee at the legal tech startup near the open ai ecosystem so heres a lot of experience in that space and so with them we will now double down on something there ariane told us yesterday that for every um for every job in the company he has like for 50 people applying for it and and he told us that uh now he can because theyre all talented he can choose the nicest people right and i just wanted to get your perspective as a cto are you also.
looking for the nicest people or is it sometimes just the guy whos the most brilliant coder or no i think we look for both i mean probably adrian mentioned that yesterday were doing on site days with everyone so everybody has to come two days on site and work for us and with us um and so with that you really get to know the person this is like because like.
for all we have like online tests and like you know things which are more like quantitative and also in the interviews but like in the end you also want to see like can you collaborate with this person um because as i said with the dual php problem they all have to work with people who have very different backgrounds and that needs to work because otherwise you will not.
build a successful company right are there still things at the moment that you cannot do or cannot find are and are desperately looking for like a certain person with these are that qualifications or a certain startup which is doing something where you where you just desperately want to get something and dont have it yet i mean we are always hiring founders and residents i think that’s that’s the thing where we have like a lot of applications but.
its in the end its still like hundreds of applications of which only very few get hired so um i think because also the founder dna is is very mixed like some have like super technical background come right off of academia and have the drive to start a business some have domain knowledge in a very interesting domain someone started a company before so people have very different backgrounds and that’s right that that makes also very interesting very wide the funnel but sure in the end.
we need to drill it down sure like obviously you’re starting next year in april youll open your doors in berlin i think um this will be the beginning of something great tell us about your plans for the next five years or maybe not plans but visions like what do you think in 2025 what will we look back at and and how do you see mirantix and but also your many companies then yeah i think i mean the next the next three four years are really about basically deploying the fund and.
building the next next 10 companies and i think that will will then be really the foundation to scale it and i think scaling you could do in different directions right like you could either say okay now lets see how we can build even more companies at the same time or we say hey how can we kind of do that with even more capital even more human resources and maybe then the five.
10 years after we start similar amount of companies but like with more capital or so like that’s still up to a discussion and just of course like just at the moment um do you think there’s a good chance that you will stay in berlin yeah yeah we have we have i mean we have all our people on berlin weve been here we will stay here uh we think berlin is.
great for for attracting talent internationally and will i mean its also getting even better um with with with everything happening around the world i think berlin is very well positioned and so um we we believe very much in this like in the end like water cooler chat and this like exchange of people which obviously no its a bit of an anti-covet thesis but like um i think especially in this early phase is super important and so that’s why we want people to be close to each other although its.
rather the mio mio mata crate were very glad to have you in berlin thank you very much and good luck.
Leave a Reply