welcome back to another episode of the edge cherries front attack podcast with lutz finger and jasper thank you for joining us here for this live podcast so this is totally uncut now we have to cut it here and there a little bit and i’m very happy that you’re still here with us in berlin yeah it’s an amazing team it’s very much fun fun office so if you ever make it to berlin definitely good coffee is close to the office but definitely come into the office i don’t know about that but they have an awesome team and they have amazing paintings so come to the office cool today we want to share a little bit of insight so lutz and i we obviously discuss a lot in our podcast topics around ai new front attack we talk with a couple of companies in in this area as you can imagine actually quite a lot and then in the partnership and with the with the investment team what is happening in ai and what kind of thesis we might have where we should invest in and we we felt hey why not share it with you i actually think it’s super important to share it because i’m now a founder by myself right i i sit there and i kind of like look at ai and it was like and i understand that stuff but still how do you make it applicable how do you create value how do you do something good with it that’s not so easy so structuring the world was for me a super helpful exercise as well yeah we’ve both been founders bootstrapped our companies at the beginning so for us it’s always around hey how can i make money with that who would buy a product what is actually the product what’s the problem i’m solving and with ai it’s a bit different you have to get the data you have to run your models you have to see what’s the output so there are a couple of things that are different and challenging we talked about it in our podcast and that might lead to the fact that here at cherry ventures we have a different view on ai and what’s happening right now than maybe some other funds so just being transparent about it you know when you speak to us you get an idea we have a nice discussion we love that and and by the way we’re definitely not right as a founder when you start you’re probably also not right but you will get there and that’s where we try to help and it gives you like we will give you six areas and that helps you actually to saying well look jasper and lutz i actually have the seventh area i have actually something else or i’m in between two it’s just a good way for us to talk about your ideas all right let’s go first area the first area we kind of felt we should split it up between two main things that the new wave of ai is doing one is an evolution so you have workflows in progress your everyday life basically but ai can make it better so that’s kind of an evolution for us something is there but it makes it much better as we said in the last podcast it’s nothing new this ai thing right so yeah evolution is something you expect it can make things a lot better so the 10x we’re all looking for and the second part is really a revolution something that is completely new something that would change our lives let’s get into the first one so i spoke about it it’s your everyday life and you’re not going to be able to do it all the time you’re going to be working at it at the very getting better so we call this augmentation it’s a very nice word and the augmentation many of you have probably experienced when you use chat gpt or something similar or even mid journey which is an assistant or an interface you speak to you talk to and it improves your email writing it makes me an oxford english native speaker quite easily at least on the email level that’s interesting that’s interesting and i can’t even imitate that so so what’s happening there i think the first part is we have workflows in general we have workflows and workflows very often involve us as a human taking one piece of information remodeling and placing it again looking up a document writing a summary attaching something and communicating that and that is always this interface can become easier so think about example how about instead of going to an online booking service and selecting potential dates you can talk to it in normal language or think about instead of when you want to go on vacation then you probably have an sap system or you have a different erp system where you kind of have to register how many dates do you have left minus the dates you want to go travel then you have to send an email to your team saying i’m away and then you have to set an out of office reminder or something like that these steps are feasible for you as humans to do but wouldn’t it be better that you’re just saying i want to go on vacation and the computer is taking over everything sometimes it’s just taking information from one system to another you just as you said you remodel it a little bit you put a little bit of your own experience in there but it’s quite repetitive and some of you might now say wow wait that sounds like rpa robotic process automation your ipad automation anywhere all these companies or if you if you know how visual basic works in microsoft excel for example so it’s scripting it’s automating things that are extremely similar yes but it’s kind of the next level because we have large language models that understand language and can help you to prepare certain decisions so augment your work and say there are three ways how you can solve this there are three answers you can give to this email for example and now you can choose one or because you’ve chosen always this one answer i’ll choose it for you and the best way to explain this the automation like the robotic process automation is you really learn an interface but humans can do actually more and all of you have gotten a rental car a rental car that is not your car different surface but you still know how to operate it maybe the shifting of gears is slightly different maybe you have to press a button or something the dashboard looks different you still figure it out so i think the hope which we are having is that when we’re saying it is rpa but as a next step is to have knowledge of what we’re actually trying to do and and we’re working at it models and i mean the world is so full i worked for like i was the president of uh product and technology at more pay we are a third-party administrator meaning we adjudicate claims healthcare claims and the person adjudicating a healthcare claim is actually getting the claim from the doctor look up certain information look in a document whether that claim should be paid or not enters in the workflow that’s a lot of copy pasting but every doctor writes different types of claims so you need that human all of those things can be automated and when we discussed it it’s interesting many people directly jump forward and ask them oh how many human beings can we replace here don’t worry i mean dear auditors out there dear insurance claim processing people you don’t get replaced but your work will be much much easier faster maybe even because there’s more thinking involved there’s less copy and pasting involved there is yeah just a different kind of work involved we are kidding ourselves when we say we’re not replacing people yes we are i mean come on we are saying we’re supercharging people we’re saying the same amount of work gets done by less we are saying oh we needed 100 people who do adjudication of claims now we only need one computer that means those 100 people they have time for something else isn’t that awesome no it’s not because they just lost their job but you’re right it’s a good discussion because what i just said is don’t worry you just get augmented but obviously when now there are 100 people that all get augmented and there’s not enough work for these augmented people you need less now i would claim history has shown that there is always more work meaning we will have more workers society but we have to admit that those 100 people that might get augmented they have to change and change is hard and we also have to agree that the value creation that ai can bring that doesn’t necessarily get to those 100 people it goes to the owner of the company that now uses ai or to the vc who invested in the company who supports the owner of the ai so the reason why i i jumped in here is ai has the potential to replace people and it’s actually that’s what we call value creation but it’s not only ai any new innovation any new technology has changed the way how we conduct our business and that has changed job types so for us when we look at it we look at processes that are pretty repetitive so the rpa space obviously is very nice and it’s just not general rpa it could be very specific point solutions for high value process it’s quite repetitive required a lot of human in the loop decision making which can now be done by data so there has to be patterns obviously as we said for the llm so there is something in the audit area there’s something in accounting in insurance you mentioned it even in healthcare if it’s safe so all these things with recurring patterns this is interesting it will replace human beings but they are good news and that’s the next thesis all right so now we have number one number one is ai in workflow ai with an improved workflow interface what’s number two so this is number two ai makes everyone smarter even me so which is amazing because i can retrieve information much faster much better we discussed that it’s language models so they read language they can extract and summarize knowledge in that language based on the data that they have and they can then use it to create new data and they can request what i’m asking for so i can search much faster this is an exciting space and the typical word here is information retrieval or search and it is exciting because the internet has made data more available to everyone in democratized data now data still needs to be processed and summarized into chunks that we then can act on now llms allow us this chunking to make this more effective if you go to a movie you can watch for two hours the movie that’s the movie if you have watched it you can talk to your friend and saying this is the content of the movie very short and concise an llm can now do this for you meaning information the internet made information free and democratized information llms are now people saying smartness and i’m not i’m not not 100 happy with this word it’s it’s more the summarization and retrieval of information and i think the interesting part is here it’s really around accessing information in a very compact concise way think of it as a summary for a paper but exactly to that point and also to the movie it’s summarizing it in his own or her own way whatever the model is and as a human being i might still take a look at it and say thank you i searched for this there’s some part missing so i reprompt i’ve tried to search again because i need certain information because what you could now expect is oh well we don’t need journalists anymore because it just summarizes all the news out there and writes a nice piece of paper but then what is the journalist doing is extracting the right information putting it in a paper and then it’s summarizing it in a paper and then it’s in the right position in the text so i enjoy reading it and retrieve it in a good way for the specific audience because the llm as we discussed it it’s just a probabilistic distribution of data forecasting what should be in there next but it’s not addressing certain individuals i would say the journalist will use llms to be more effective like they use a keyboard or a computer now let’s talk through a little bit business models so definitely search so in anybody out there is trying to replace google the other part is anything where we aggregate information for example podcasts like us maybe you have a tool out there which summarizes us maybe there’s a tool which reads the news and automatically summarizes the most important news because we talk too long and you just want very people information from this one podcast and that exists already like like there are like sport news or stock ticker news all like that and then you those very often get written by a computer so that would be one way where we will see new business models upcoming yeah and and think about there’s auditors tax advisors engineers so sometimes you have to do something new or something very different that is comes up maybe once every year and before you would have to read it up in all your documents or remember it now you can just retrieve the information out of many many documents so you’re much faster in your work and maybe even more precise because you don’t know how to do it and you don’t know how to do it and you don’t know miss out any of those documents and the good news is this is information retrieval is in every part of our society so there should be loads of good business ideas and we believe that this is not one size fits all let’s take education what is the best explanation for a given math problem or for a given language problem and that’s an information retrieval topic right because you might have many different potential explanations and you might have many different potential explanations and explanations you have only one student jasper so the question is from all of those explanations for a given math problem which explanation do you want to choose by the way nothing big different for a founder who comes to us and tries to explain the business model there are many ways of pitching your model and you’re trying to figure out what is the best way to actually pitch it now what’s the angle here and it’s not just text so for example i love pictures i love descriptions maybe that’s why i was a consultant at bcg so just sending me a text with a summary doesn’t really help if i have a diagram a matrix that’s amazing so multimodal yes probably help yes and like to find the right pieces information retrieval is the one we we had on the podcast legal os right legal as a typical case legal text find the right part in the legal document to actually tell it or we had on the podcast as well ultimate they recently got sold yeah it’s like very like was a good investment overall congrats again to reito and his team again you have customer care information pick the right information retrieve the right information and make it easy accessible so the whole space around information retrieval slightly different than being part of the workflow that this is one but that’s a huge space yeah and if you combine both of them and just picking up the ultimate ai example customer care is still it’s a huge industry it’s a multi-trillion salary industry so we expect a lot of business models coming out of there because just thinking multimodal voice can now be easily integrated and so forth so take a look we love it we like to see more business and it’s not only because you can involve an integrated voice obviously what’s the business need what’s the user story behind it but yes there’s a lot ongoing number three number three so as we see we see a lot of companies trying to solve very specific engineering problems computer science problems so these kind of tools you see them in the open source space you see them for software architecture infrastructure everything and last year same happened as happened seven eight years ago and the first kind of ai with that i at least experienced a lot of tools come out and this time it was much much faster that they got either be replaced by something from the big models for example or they were just not needed anymore still we think there is a new wave of really relevant ai tools that can become platforms for one mere reason it’s new technology and it’s it’s something new in our toolbox whenever you get something new into the toolbox and you need to figure out how to maintain it how to manage it how to set it up and as we saw certain companies hugging face-to-face and we saw that there was a clearly stepped up other companies civic ii launched relatively new into the space so i think we roughly in in that space we have quality and training ml like llm ops we had machine learning ops we have now large language model ops or image generation model ops or fund ocean model ops so there is the whole part of how do you control quality how do you train how do you retrain how do you keep track on your lower us weights and biases is a extremely good example and for us it’s a little bit and we had a podcast about it last year you can you can imagine ai sometimes it’s a bit like a wild animal and you have to tame it so you have to train it in the right way guide it in the right way especially when you come to enterprise applications so b2b i want a certain output i want a certain quality level so i want to make sure first of all i understand in what direction the model is going to be and i want to make sure that i understand the model in what direction the model is heading so observability but then i also want to have some kind of guidelines quality control quality gates that the output doesn’t do anything bad for my product or my customer now the other area for toolings is definitely around security and privacy we know that working with data makes business models vulnerable which data are you allowed to use how do you store it how you train it how can you take data from your customer’s and train the model huge space we see microsoft and amazon obviously doing a lot there to secure the data sets but i believe there is a wide opening of more opportunities for companies to develop stuff yeah and talking about data i mean still it’s a challenge to have good data to fine-tune the models even for your rack so just having working on the data quality storage we saw vector databases obviously which are not as good as the data that we have in the market so i think it’s which was a huge trend but just ensuring the quality managing that we hear that for many of the large consultancies the large enterprises that’s their biggest challenge to really apply ai now the next area would be chaining we talked in our podcast before about long chain long chain meaning you the large language model takes one decision and then you plug on for example a calculator you do not calculate it within the large language model you do not calculate it within the large language model because it’s a probability distribution it’s just a word continuation you better do this in a calculator for those different agents you need an orchestration for my own startup for example i take in question from my customers and then i make five questions out of it and then i ask the model then i do a rack then i control using cosine similarity whether this is the right thing to do then i merge it and then i send it to the other side of the model and then i do a rack and then i do a control using cosine similarity these are all different steps partially these are non-large language models that needs orchestration and i think we will see a platform emerging that not only looks at the quality and the version control and the data security and all of the good things from one model but and which model comes after which model and how do they plug into each other because we see every month a new model outperforming or an existing one outperforming the other models and different benchmarks so you want to be very flexible and orchestrate them and observability also comes into it and we see two directions one is having a more generalized platform for maybe certain areas or very specific applications of orchestrated platforms for example in user interface testing that might be something to take a look at where you just optimize you just control those agents a little bit by the way and a huge like very cool investment doing user testing but we just recently looked at a company where we used a very specific platform and we just said okay let’s do this look at finding compounds now that is as well a combination of different models because compounds have logic to connect as well as you can use the ai to iterate through the most likely compound combinations this is what we saw in foundation models or in transformer models being very very effective use of those you need to combine what is logical with the transformer model in terms of finding potential new compounds by doing that you essentially have a staging problem and for that we we expect there to see some platforms now let me summarize evolution first bucket in those first bucket we give you three areas area number one workflow i want to just make it easier to connect the digital multimodal setups workflow number two search engine and then the second one is the digital multimodal setup workflow and information retrieval a category by itself as much as google became a category by itself loads of things are happening there because information retrieval is so core to whatever we do in our knowledge work and area number three is we have way more models we have way more complexity we have way more structure let’s come bring all those tools together these are ai tools so these are the three areas in the evolution part all right so now we come to revolution revolution obviously with revolution it’s a little bit if you hope for it it might not happen but we think there’s a huge chance this could happen we had when we started our little podcast series with a big discussion in one podcast about the interface so we just enforcing this idea and we still see this happening and it can happen a lot out there and what does interface mean it’s a lot out there and what does interface mean it’s personalized access to information consumption of information in a much lighter more natural nicer way than you had it before if you compare it to two steps in the past the first computers were just a lot of you know you would pull levers you would press maybe even buttons at some point in time then you had the keyboard but then there was this mobile interface revolution where you would access information where everywhere you are i think you gave the example of google map existed before mobile and then you would print out google map yes run around with the printout so that wasn’t necessary so it opened up a lot of new business models at cherry we invested in inflix bus so you have these buses that you can just hop on and book a bus or we invested in zalando which gives you access to choose 12 things from your home place and you don’t have to go to the store anymore this could happen again now yes the question to all of you is having this capability what’s now actually feasible right as the phone was invented people actually used it initially to listen to music and only later people figured out hey we can do other things with the phone as well listening to concert halls and a remote setup but i think here as with mobile suddenly oh use cases only become usable because of mobile here having a large language model good examples you talked about zalando we saw that amazon is now doing it by using all the customer comments and aggregating them in the future we will have bots that help you with advice way better than anna from ikea which told you beds are in the bed section and kitchen are in the kitchen section no thank you that didn’t help but really kind of trying to figure out okay what is it what you need what what’s your circumstances there are different features i think you can do that but i think it’s a good thing people like you like the following feature so being a real sales consultant we will see that capability coming up and showing you what you need or what you’re looking for right now not bombarding you with too much information much more customized key here is planning becomes actually very important for this we talked last podcast that planning is still an issue but can be overcome or faked by a large context window i assume that we will see those companies coming up and building generalizable sales people and that might be amazing because now you actually can say i want to have advice from somebody who i expect to sell this is as much as we look for google to show us a blue link where we’re saying yeah that is meant for advertisement so we will go to a large language model but and you might have to set up your company in a different way that’s why there’s a chance that not just any company would adopt this but just going back to our friends from zolando it’s not just the web interface and you show the clothes on the website but also have the logistics in the background to deliver them fast work with the returns so think end to end for example you could because you search for a certain thing the ai could make up clothes that are similar that don’t even exist and then you click on them and so i start a bit predicting i can show you more relevant stuff helps with the conversion then advise you to do that and then you can do that and then you can do that on size and whatnot and then also again to the interface one only show you what you exactly want to see and want to take as a decision so i don’t show you all the buttons coming to the sap example and we love sap don’t get us wrong it’s an amazing company do we pretty clunky interface and maybe i don’t need all the buttons at the same time or adobe photoshop and i just need two or three so just understanding historic patterns of the users understanding the different user personas it’s personal and then i just need to show you all the buttons and i just need to show you so i can personalize i show you what you want to work with right now not more now this the the example you bring it’s a good one because i would expect that we see different social networks appear think about how how social networks looked like right we we had friends and then we shared cats and baby pictures and everybody got excited and after a while they didn’t and then we got excited about the kardashians liking different social networks and then we got excited about the different cats and baby pictures and we used replaced the friends with the kardashians for our news channel or whatsoever and now the computer makes up those pictures and selects them for us i think we saw a trend in the home feed which is ai selecting what we want to look at first our friends then the kardashians then the ai now we will see the same thing in the generation like the generative part meaning first this was the pictures from our friends cats and then it became the more professional created pictures by the kardashians and now it becomes the more professional pictures created by ai and importantly we see this trend coming in consumer first for several reasons and there will be podcasts about it but then going into the enterprise because usually those interfaces happen more with consumers at least what we saw in the past because and this is a like you you might wonder what’s here the workflow discussion the workflow is enterprises integrate into the workflow because most of us will people suffer through workflows because they get paid essentially so making a workflow smoother is one thing however being able to directly communicate in natural language with the computer will change the way how we consumers act and let me make this very clear because this is a distinction i see very often people getting mixed up in my view mixed up and i don’t want to judge anybody but if you come to us and saying i have an interface where you can talk to your favorite charting application look at tableau whatsoever you you write it and out of what you write about your text i want to see cell figures from the last three years compared to the temperature rise on the north pole and we convert this into sql and then we send it over to your favorite application be it tableau looker you name it then it’s a workflow because why would this be so amazing that you go from words to sql and then to the application which is part of your workflow just put it into the workflow directly right if we’re talking about a new types of social network we used to watch facebook then we all went to tiktok and now we have this new amazing network which is your idea whatever it is that’s the difference between workflow and utilization of this new capability for consumer interface and and in simple terms for the workflow part you can keep it ugly but you just make it much more efficient and for the interface part it should also look quite nice and different and easier let’s talk revolution two which is creation music or image creation or whatever creation like at the end of the day it’s super charging innovation and creation and you know you might say no well but i’ve seen that right i’ve seen the pictures i’ve read the text i heard the music amazing by the way that we love the music part we had a whole podcast about becoming the future becoming more in the year or runway which creates videos so right amazing yeah but it’s it’s not just the creation it’s super charging everyone making innovative things creating something for audiences because again the ai is using patterns probability distributions to predict something which could be a picture but then it has to be the right picture for the right occasion so you as a creative out there you can actually use that to iterate much faster what is in your brain to put it on paper on your computer or on on the internet and then tweak it in the way that the audience might like it you might like it or as an artist so to actually make sure that the ai which cannot reason and plan you reason and plan for the ai yes working together to actually have them as a buddy took me a while to work with this dude right and but like i figured it out now and he did as well so same for you you colleagues and whatsoever that is an interface which we now created an ability to have somebody who has constantly new ideas and we had a discussion about what’s the cost of a movie production 500 million for a movie is not uncommon and we are talking about future is in the booth no what is becoming way more easy to is to generate many ideas and what it needs is good tools to manage this the best chess player is a human with a computer that’s the best chess player team not a computer alone not a human alone now a human together with a computer and they need to have an interface can that interface being smarter and more effective so that human and computer can go better together yes there’s a lot to come so what is so super charging is obviously the creation part so i get a lot of people who are not familiar with the idea behind it where you come in you come in you come in you come in you right content for the right audience. Because coming back to this 500 million movie budget, yes, I have kind of an idea what should be the movie. But then I build it and create it. And then I release it to the audience.
And I can’t really test much in between. But with AI, I can take individual scenes, it’s not as perfect as the final movie. But like with cars, I can show designs, I can test a lot of things, I can de-risk that my movie will be successful or my game or my picture. I actually wonder whether you can have something like AI-led or probability-led AI-led A-B tests, right?
That you kind of test storylines with your large language model saying, where in this probability distribution does it fall in terms of good or bad? Like, what would you give? We are not there yet, because those models would need a completely different training set if this is even feasible. But I’m just throwing out ideas.
We have a new interface, the same . . And we love to see a lot there. This is really, really cool.
All right, number three in the revolution stage, we think there is a comeback of hardware. And we’re saying that from obviously a VC perspective, where it’s always hardware, I don’t know, it takes a long time, it’s difficult. What happens now is that AI, the new way, of AI can actually make hardware really, really smart. So one thing is, and we will see this happen more and more, there are iPhones coming with AI, Samsung phones, we see chips on the edge where these models that have been trained get compressed so they can actually fit into those chips.
And why is this important? Because again, as we said for the workflow, these discussion, you can now augment decision making, you can have information retrieval in a very efficient way. And isn’t that amazing if you think of robots walking around and understanding their environment? I mean, yes, like the whole idea of measuring on the edge and operating on the edge.
I had, like I used to work for CERN, the high energy physics laboratory between Switzerland and France, an amazing place. It creates so much information that you actually on the edge have to condense the information and find information retrieval, the right information. Now, we have a similar thing if you use your like normal watch and you kind of measure pulse activity whatsoever. That’s a device, you shine light on your skin, you take on sensoric data and you’re trying to predict and move.
They already put it into rings. Yes, they put it into rings. I know the aura ring, but you can use it for way more. For example, you can predict now blood pressure by just looking on your face.
I like Samsung actually started to invest in a company having this in their screens. Now, you could actually monitor sitting in a health care environment, people staring on the screen all day long. Now they know when your blood pressure goes up or not down. You could do all kinds of very interesting things with it.
You can start putting it on electrodes on your head and start predicting dreams. We saw a few papers out there. Much information can be collected and very, By the way, if anyone wants to measure our blood pressure and put it in the comments, feel free to do so. It would be interesting.
But again, this enables application, very different hardware application. It could be in production. We said robotics. Cars are interesting.
Planes are interesting. Drones. So just make everything smarter out there that is based on metal or aluminum, carbon. So here they are, our second set of three theses.
So it’s a whole bucket of what we call revolution because it’s a leap jump. And what we had is a different interface. A different interface might create services that are not evolution, that are not an extension of a workflow, that are only enabled because now suddenly, you can talk to a computer like we talk to each other. Second part is that we believe the whole creativity is supercharged.
And that needs a management between the computer and us. And the third area is because we have now this new processing power, we believe there is a whole new suite of applications which will work, on the edge with metal, to trying to improve decision making. That doesn’t have to be generative AI, by far not, right? Many of those cases which we discussed here in our six areas are not necessarily generative AI.
It is a combination of AI, foundation models, fast computing that we can use in order to make the world a better place. So thanks everyone for listening. We will give you a write up of all the theses. We will have some content pieces around it.
So you can also read this up. It was amazing to meet with you in person. Next one is either New York or San Francisco, Mountain View. Maybe London.
Maybe not. Oh, London. We have a very nice office in London. Feel free to come by.
We see you around. Talk to us. Let us know your ideas. We count on you.
Thank you.