Tobias Schlottke: Hello friends, this is the Alphalist Podcast. I am your host Tobi. The goal of the Alphalist Podcast is to empower CTOs with the info and insight they need to make the best decisions for their company. We do this by hosting top thought leaders and picking their brains for insights into technical leadership and tech trends.
Tobias Schlottke: If you believe in the power of accumulated knowledge to accelerate growth. make sure to subscribe to this podcast. Plus, if you're an experienced CTO, you will love the discussion happening in our Slack space where over 600 CTOs are sharing insights or visit one of our events. Just go to alphalist. com to apply. Introducing the AI All Stars Episode --
Tobias Schlottke: Welcome to the Alphalist Podcast. I am your host, Tobi. And today we have a very exciting episode I'm looking forward to for a while already. Um, let's, let's call it the AI All Stars episode. Meet the AI Experts: Insights and Backgrounds -- Tobias Schlottke: Today with me here in the studio physically is Jonas Androulis. Jonas. Hey,
Jonas Andrulis: good to be here. What are you doing? I'm the founder CEO of Aleph Alpha, ex Apple AI R& D, serial entrepreneur, that kind of thing. Jonas Andrulis: Okay, Tobias Schlottke: and then we have Rasmus Rasmus Rothe: Rother. Rasmus, what are you doing? Hey, great to be here. I'm Rasmus, co founder of Mirantics. We invest in AI companies and I have a background in machine learning.
Tobias Schlottke: Great. And then we have Johannes Schaback. Last but not least, Johannes, what are you doing? Johannes Schabak: Well, certainly the least, the smallest, if at all, star in this star round of AI researchers. Johannes Schabak: My name is Johannes. I'm the CTO of SumUp and I'm super excited to be here. AGI VS. Transformative AI -- Tobias Schlottke: Then I drop a more, let's say, general question to kind of ignite the whole round. Um, When AGI. What, what do, what is AGI to you?
Jonas Andrulis: Well, good question. What is it to you? So I think it's, it's a horrible term. And so I stopped using it. Jonas Andrulis: And the reason is that it means so many different things to so many different people. Um, what I quite like is the term transformative AI. This was, has been used by Open Philanthropy in a report a few years ago, where they say transformative AI is AI that can do. About 50 percent or more of what by then was the total amount of work on the world.
Jonas Andrulis: It is technology that will shift the global power dynamics, and it's technology that will create new knowledge by itself. And they expect this to happen at around 2030. And I'm fine with that. What to expect in Transformative AI before 2030 -- Johannes Schabak: What would you think are the next steps that we see before 2030 towards transformative AI?
Jonas Andrulis: Oh, agents, easy. I mean, that's already happening that we are, I mean, we are, like, dense transformers are pretty much scaled out. Jonas Andrulis: Um, many of the, um, improvements basically come through brute force scale in a way, That kind of is difficult to deploy. One of the Agents and Scaling Limits of LLMs --
Jonas Andrulis: reasons why GPT 5, which exists, is not out there is because it's so difficult to deploy and Microsoft is losing money with every customer as is. So that's kind of one of the reasons.
Jonas Andrulis: And the other is that many of the capabilities of LLM are basically created by labeling, and that's the reason why scale AI amongst others, like they're making billions in revenue. Basically by just having tens of thousands, hundreds of thousands of people that are creating, handwriting correct answers.
Jonas Andrulis: So, you already see when you look at that, that there's kind of an upper limit. These machines are mimicking the tasks that humans are labeling or creating. And that is already pretty powerful for a lot of what knowledge work that we're currently doing. But I mean, that's about where that stops. And so basically to get more than that, we need to embed these LLMs into larger systems that can do planning, that can do criticizing, that basically can be called agents.
Jonas Andrulis: Things with like a memory. Like a plan, like an can, can run multi-step processes and that's currently what's happening all over the place. We're also working on that, but many, many teams are currently looking at that. Johannes Schabak: I was about to say, this is what Aleph Alpha is about, I guess, The Joy of Unsupervised Learning -- Jonas Andrulis: well agents in a way like Alpha was we, I left Apple in 19 and uh, 2019 and I was.
Jonas Andrulis: It's inspired by, like my last company was doing supervised learning and supervised learning mostly in the computer vision field. And when you do that, um, you basically get, uh, a, an overdose of, of labeling and bounding boxes around pedestrians and that kind of thing. And so this, but this technology is fundamentally limited in the sense that it doesn't matter how much your data you collect or how much compute you throw at it.
Jonas Andrulis: It will at best, you know, it will at best, Uh, recognize 100 percent of all pedestrians. So I was bored by that. There's still legitimate use cases for that, but I was bored. So I was inspired by some of the work I've saw, um, within Apple, but also outside of Apple, uh, that was building systems that were self supervised or based on reinforcement learning.
Jonas Andrulis: So systems that did not require labeling for the specific tasks that they then would replicate. And so the hypothesis was that this is an era beyond supervised learning. I had with the, on the laptop, I found, founded this company with, I had like a big sticker on it that said totally unsupervised. Tobias Schlottke: Did you already know which, uh, or guess which, which kind of red ocean you're entering there or potential red ocean? Tobias Schlottke: Yeah.
Jonas Andrulis: So I had this feeling that we would underestimate this technology for a while. Um, and this was because, um, I was surrounded by brilliant researchers within Apple. And there were some results, some experiments where these AI systems were showing, um, capabilities that I and all the brilliant people around me were expecting to be impossible.
Jonas Andrulis: Like I'll give you an example of it. By now, we're not surprised, um, is language models can correctly answer logical questions that they have not seen in the training data. And this was, this was a big surprise. Um, but also there was a, there was world models. There was an agent that was playing Doom, uh, just by, by, uh, experiencing the game and navigating this three dimensional simulated world.
Jonas Andrulis: I had no clue that this would bring me within a few years to a point in time where I had a five figure interest income per day. So it's a little bit wild how fast that grew. The Growing Pains of all AI Applications -- Tobias Schlottke: Rasmus.
Rasmus Rothe: Yeah, I mean, like, I think, I think, like, we've seen this in different areas of machine learning already before in computer vision, like, at the beginning, everybody would make fun of the models because they would make all kinds of mistakes, and suddenly, it surpasses a certain, like, threshold in terms of performance, where it's just good enough.
Rasmus Rothe: And we suddenly say as a human, Oh, wow. Um, it's good at recognizing these images or it's the text that generates is actually really good. Like before the early, earlier models, um, like before chat GPT, I mean, you know, there was like, still like a lot of mistakes, every other sentence. And so everybody would say, look like this is not usable in any commercial application, but suddenly when it could generate something that made sense and maybe would even allow you people were like, okay, now it's there.
Rasmus Rothe: So I think it felt, or it feels very binary. from a user perspective, but from a technological perspective, I mean, there's been a lot of continuous progress. I mean, both on the, on the, on the architecture side, but mostly as Jonas said, on, on scaling with more data and, and more compute and just seeing, can we make these models even bigger, throw even more data at it, um, until we are where we are now, um, but I think now it's also, as you said, Jonas, I think there's like a bit of a transition point. Move to Smaller Models --
Rasmus Rothe: I mean, sure, people will try to make these models even bigger, but, um, it becomes also then very expensive, both to train, but also inference that it's maybe not. Um, necessary anymore. So people are looking now at more approaches where you actually, you know, make them smaller again, make them more efficient because we see that also with some of our companies, um, that use LLMs, like when they have a lot of usage, um, you know, you need to start looking at the, at the, at the cost you have, and that's also at your margins and need to do all kinds of tricks to make sure you don't call the LLM too often, um, or you restrict the users and, um, they're obviously like smaller models again help. RAG and combining with Propriety Data --
Rasmus Rothe: So I think that's kind of on the, on the LLM side and they're obviously also. Um, people are trying to combine it also with, with like hard facts, with database, with a retrieval augmented generation, where you basically, um, not just have the LLM generate stuff, but you can also connect it with kind of Tobias Schlottke: real world data, real world
Rasmus Rothe: data, where you also know that whatever comes out there is being used properly and people are still experimenting with it.
Rasmus Rothe: Some also now try to throw the data directly into the context window, but I think there's still a lot of. things people figure out. And also we see for some of our companies, um, people are just using LLMs out of the box. And then for some stuff, you also still fine tune because there is actually, um, proprietary data available that not just throwing it in through your context, but on top of that also helps further fine tuning.
Rasmus Rothe: So, but that's application specific. So I think there's still a lot to figure out when you take what's there now, to all these kind of enterprise applications or consumer applications and I think everybody's still experimenting and it's not like converged. Johannes Schabak: That's super interesting. The Role of Data in Training AI Agents -- Johannes Schabak: Um, Trent, this idea of the, like, we're going into an agent based AI episode or it's like phase in the world.
Johannes Schabak: What does that mean for the data that we need in order to train those agents, right? Because you just said, um, Um, data plays a massive role. Obviously for LLMs, we had all these massive corpora of text and all the data that's readily available. But, um, you know, training more generally capable agents, um, or building simulation environments in order to create the data in the first place.
Johannes Schabak: What, what does that mean to this whole data industry? Is there, um, is that the next challenge? Is, so is it more, less an algorithmic challenge? challenge or say research challenge more than an, um, and it's more and more a data creation challenge. Tobias Schlottke: And, and, and if that is true, then like, wouldn't automatically like someone win per default, like Google, for example, because they have most data because Jonas Andrulis: they scrape most data.
Jonas Andrulis: So it's certainly an advantage to have the kind of data that a Google has or an Apple has, certainly that, or others in that space. I would say there's a naive answer and there's a maybe less so naive answer. Of course, data, is a part of the answer. And so what that means is we can use LLMs, and it's already happening, to plan and to judge future outcomes.
Jonas Andrulis: And basically the the simplest form of an agent is as a an LLM that kind of knows when to use certain tools, what what to do in terms of like action, retrieve a document, query an API, use a tool, and then evaluate answers. whether that's the desired answer, whether that's complete. I think that's already pretty, like we're probably pretty there. Jonas Andrulis: We can already have these systems. Dreamer Model Architecture --
Jonas Andrulis: And what is also very possible is that we will see um, algorithmics innovations. There is, there is an, an old, like a few year old model that's called Dreamer, the architecture that what, what it's doing is, this is an agent that has a world model and is using the world model to project them, the agent cell itself.
Jonas Andrulis: and the world with all its actions and random events and everything that can happen into the future and then you get hundreds of thousands of paths and potential future states and then it's using a value function to evaluate those states and the paths to those states and basically makes decisions with a, with a utility function.
Jonas Andrulis: This utility function could be be maximum expected value, but could also be a risk averse utility function. And so we've built these kinds of systems. And so we could have systems like that. Um, but, um, an agent that basically is LLM based and that is just helping me solve a certain task probably does not need a world model that also contains the agents itself, because things like self preservation, And the future of the agent, like the future state the agent is in, is probably less relevant for these kind of tasks. Synthetic Data --
Johannes Schabak: Okay, but if I was to build an agency for, in the healthcare system, if I was to build an agency, um, I don't know, for a certain disease or predicting how a certain medication would work out, acquiring that sort of data is so incredibly more difficult, at least the real world data, sampling the real world, and then that case for, say, cancer treatments or whatnot.
Johannes Schabak: Um, so that an agent based model seems exponentially more expensive in terms of data acquisition. Is that true? Or can I replicate and manufacture data somehow?
Rasmus Rothe: I mean, like, I think in many cases it helps to have the real world data. Um, especially if there is like, it's also of a different modality. So if it's not just text, but say some biological data that is in some other structure, not text, I mean, Um, you can obviously always transform data, but there's some particularities.
Rasmus Rothe: Um, when it's about like, we're, for example, one company that optimize protein materials that also is a generative model, but for protein data, very different than like normal texts. So I think there, it helps to, to have that data, to train on the data, to learn from that data, create new proteins and check basically based on the lab.
Rasmus Rothe: Like, is it a good or bad protein? And then use it as a feedback to optimize it. Um, so that data, I don't think you could like. Dream, dream off, or you might also not find in that, that capacity online. Where foundational models excel --
Rasmus Rothe: But at the same time, if you think about the enterprise world, a lot of what we do is ultimately, you know, writing documents, like if it's about like legal compliance, finance, marketing, for a lot of these bit more generic tasks, actually, the data that is in the web is a good like training resource, a good head start.
Rasmus Rothe: And you can, you know, already built pretty good system based on this rather publicly available data. And so just take one of the LLMs, um, and then, and then build an application around it. So I would say it depends. I think what we've seen from an investor perspective, I think like six to eight months ago, everybody was saying, look like these generic foundation models.
Rasmus Rothe: They will. Basically can, can be applied everywhere. Now people see also that for some of these more specific tasks, you have models that are maybe fine tuned or better embedded, or have some, you know, like database that, that, that, that, that is included, that basically creates better results. Um, just because there's some very specific information that changes maybe also all the time, say it's like in the regulation space or so that that can be used, um, by, by the model.
Rasmus Rothe: And so for these applications. Um, I think you will have more vertical solutions. Johannes Schabak: Super interesting. Yeah. AI Sovereignty and Build/Buy/Pivot -- Johannes Schabak: Should we talk a little bit about Aleph Alpha? I would find that super interesting to understand, you know, what you do, what are your, what is your ideal customer? Jonas Andrulis: Size does matter with our ideal customer.
Jonas Andrulis: So we're currently focusing on, uh, biggest enterprises, um, And governments. And the reason is that, um, our, uh, advantages over off the shelf commodity, um, software, LLMs and, and, and generative AI is that, um, we give more control and more sovereignty to the user. And so basically what we're currently doing is we are, we are helping the transformation of huge enterprises, um, precisely in the areas where it matters most to them.
Jonas Andrulis: So I'll give you an example. What is already pretty commoditized is customer support chat. I mean, this is, this is not something where a huge corporation would say like, Hey, Let's say our investors like Bosch, for example, they would not say that their ability to answer maybe customer support tickets is necessarily where they have their unique advantage.
Jonas Andrulis: Although Bosch may be a bad example because they have their service organization, they do quite a bit of those things. Um, but, um, and this is already illustrating in a sense that, Um, what I'm asking my customers is, what is the part of your knowledge work that is the most meaningful contribution to your company value?
Jonas Andrulis: Like, what is something where you can basically just become a user of something somebody else is building? And what is the thing that you cannot give? to somebody else, where it's your data that is valuable, where you are in a unique position to make a great move and build an empire. And this is the answer to that is different for every, for every enterprise, but basically all enterprise have, and I think they should have the ambition to not just become a paying customer of tech that somebody in the U S or somewhere around the world is building, but also to build new markets themselves.
Jonas Andrulis: Um, but where. you should focus with that. This is really different, uh, for, for all the players. But we, we are there basically to provide a foundational technology they can build upon. Why your strategy matters more than using the hottest foundational model -- Jonas Andrulis: Because for these transformations, there's a lot more things that matter than who built the best LLM for that. Um, this is also important, of course, right?
Jonas Andrulis: But there's, there's tons of more. Example of detecting contradictory statements using explainable AI with positive and negative sources -- Jonas Andrulis: And one great example is we invented a method to make the results from LLMs explainable. with positive sources and also with negative sources. So what that does is it's a modification on the attention mechanism that allows you to see what are the observations that have been critical to create a certain factual output.
Jonas Andrulis: But you can also do this the opposite and kind of say are there observations that would contradict this output. So you can make results auditable but you can also say Is there like if you let's say you're writing a quarterly report as a big enterprise, so you can say, Oh, I'm writing a paragraph and I have a claim inside the paragraph.
Jonas Andrulis: Is there anything in the interviews my CEO gave to the news? Is there anything that would contradict that claim? And I can easily kind of see this one at a glance. So I think this is critical. quite powerful. And this is really what matters to a lot of our customers because, um, the, the progress of LLMs is so fast that once you start this transformation, it'll take you one year or one and a half year until you have this in production with like a, a DUX 30 company.
Jonas Andrulis: And by that time, all the LLMs will have changed. And so for, for that decision, it is basically irrelevant whether LLAMA 3 is now better than CLO 3 is better than GPT 4. 5. It matters little. What matters is, what is the strategy? How do you position yourself so that you are in two years, in five years, you're in a great position? Tobias Schlottke: Interesting. And how do you position yourself to be in a good position in five years?
Jonas Andrulis: Um, we are the, maybe the most profitable company in Gen AI. Um, we are currently commercially very successful, but it will, this will not be enough. So the, the challenge basically is. Can we get strong fast enough so that we can survive, right? Jonas Andrulis: Because even with 500 million, which is, which is an insane amount of money where I'm coming from, but even with that amount of money, this will not be enough to survive as a player that is a legitimate alternative.
Tobias Schlottke: Wait, if you, if you, if you're listening to Sam Altman, it's, it's not enough, right? Like, it's not, well, if you're listening to him, Jonas Andrulis: no number will ever be enough. Vendor Lock-In and Data Breach --
Johannes Schabak: Okay, well, just let, give me that, give me, let me help you get that straight. So you're saying that as a huge enterprise, I'm, um, you know, worried to obviously have my data, uh, been used in, say, a Azure hosted GPT 4 instance. It's, that's, I would have never let that happen. And that's where you come in, right?
Johannes Schabak: That's where Aleph Alpha comes in and would say, Hey, look, we not only give you sort of like a data vault, uh, that's super secure. Um, but also we insulate you from, a vendor lock in onto a particular LLM. So, but, so that means you're actually not even building your own LLMs internally? Is that what I'm hearing?
Johannes Schabak: Or are you, and you're building a massive abstraction facade towards other LLMs? Or are you saying, well, you, we do build our own foundational model? for, you know, for the language, or for the country, or for the domain that we're in. Um, and, or is it super bespoke? So I'm trying to understand the value, um, that Aleph Alpha adds to an enterprise that, um, you know, that comes with its own data, and does not want to use GPT 4, um, and where, you know, where you're trying to say, well, you know, we keep your data, but you're, at the same time, they want to have the access to the capabilities of a GPT 4, say.
Jonas Andrulis: Yeah, it's all of those. So, We have built our own model. We will launch the kind of next version of that this year. This will basically, from current tests, will be a pretty good choice, especially for non English. We develop technology to fine tune and to specialize, but we also basically have an end to end stack that adds innovation on top of LLMs, and you can integrate open source LLMs, even closed source LLMs as an API callback.
Jonas Andrulis: The method of explainability we developed, of course, our LLMs support those, but also open source LLMs that we integrated into our platform, they support this as well. So you can basically have a LLAMA3 result, auditable, In our stack, completely scalable, uh, you can run it on different clusters, like one in the cloud, one on premise, and we do have customers that tell us, if we see any of your components phoning home, you're in deep trouble, like they want to run this air gap, they want to run it in a secure environment, and it's not just the data, it's basically, and I can absolutely empathize with that, Why companies are pursuing Multicloud in fair of being 'milked' --
Jonas Andrulis: If you think about your, like, one of the most important parts of your value chain being exclusively in one cloud, that you can never move out of, in a behind the black box API.
Jonas Andrulis: This puts you in a horrible position. I mean, we see big German companies putting tons of money into a multi cloud strategy because, and one board member kind of phrased it like that in a conversation with me, he said, if One of those clouds know we can't move out, they'll milk us for all we got. And so I think that's important, and if you want to take responsibility for future value generation, for the position of your company, you need sovereignty, you need to be able to understand the technology, to run it yourself, to move from one cloud to the other, to move on prem, to move from NVIDIA to AMD, and I think this is really important, and of course then we're adding innovation like explainability or guidance where you can have like more transparency and control over model outputs and stuff like that.
Jonas Andrulis: I The Future of AI in Business and SocietyMarket of customising foundational models --
Rasmus Rothe: mean, like we, how we think about the market, I guess we have like two, two ways to think about the market. So on the one hand, we have a services company where we build customized AI solutions, um, for also governments, big fortune 500, um, also SMEs. Um, and that's kind of what, what Jonah said, that those are usually applications where kind of your horizontal solution that comes from like a big cloud provider.
Rasmus Rothe: Sure. is not sufficient because it doesn't, maybe you need to add proprietary data. There's certain integrations you need to do. Uh, there's maybe different models you need to use. So it's something where things don't work out of the box. Plus they sometimes also, as you said, Jonas, um, maybe, maybe are a bit cautious about things being hosted, uh, not, not in the same country and having less control over it.
Rasmus Rothe: So I think that's going to be a rising mark because market, because ultimately there is a lot of customization. It's a bit like in software development. I mean, there's products out there and there's like a huge market of like software development, which is bespoke. And that doesn't make always sense.
Rasmus Rothe: Sometimes now it makes more sense probably in a few years because some more things will be productized. But, um, for, for this regard, I mean, you are also kind of independent of what kind of model you use in the end. Like we see that also with the latest models, like Lama three and everything, the models, I mean, yes, there's a difference and there's some better than others, but, um, the gap is closing.
Rasmus Rothe: And, um, I think also the knowledge. transfer between these organizations as people hop from one to the other organization will, will kind of close this gap even more. And so, um, I think you need to just be flexible ultimately and also switch out models. Um, and a lot of the open source push I think is also helping, helping for this. Rasmus Rothe: Um, so I think that's on the, on the service side. Room for vertical AI business for industry specific solutions --
Rasmus Rothe: The other thing is, I think there, there will be, industry specific solution where you have enough customers that have exactly the same problem. Um, for example, we started one company in the litigation space, so workflows, it's always kind of similar. You send certain letters back and forth.
Rasmus Rothe: There's certain other data formats and databases you need to look at. It's a very specific workflow. Now, Microsoft will not release a product that will be like, Hey, we're The best litigation product in the world because they could do that for like thousand different verticals and they will not have the best product in all of this.
Rasmus Rothe: So I think that's where exactly then the room is to basically build like a vertical AI company that is solving this problem better than everybody else and maybe even sit somewhere different in the value chain. So, for example, not directly sells to law firms, but for example, sells to insurance that have huge amounts of claims they need to settle, where you directly say, look, like, instead of giving money to the lawyer, give, give money to us.
Rasmus Rothe: And then we subcontract the lawyer. And then coming back to your data collection point, then also starting to collect data from, I mean, all these like different claims, but also for example, collecting additional data on like what's what price to settle for with what for what case like this is data that is not available and that you only collect when you run this at scale which then builds a defensible mode because you are the only one know when to settle for what and so that's that's I think what we will see emerge so in that sense I think there will be a lot of customization more enterprise bespoke and then there will be um secondly like very strong like vertical solution and it's a matter of And I guess the time and how good the models are, what way to go at the current times. Message for B2B SaaS --
Tobias Schlottke: So that's actually a positive message for like today's B2B SaaS world, right? Um, that like you will survive guys, because, uh, you, you, you can apply AI to your use case and, and, and it will have a positive effect for your customers still. But you have, Rasmus Rothe: but you have to do it. I think it's, it's almost, you have to do it.
Rasmus Rothe: It's a bit like, as if you were a company. that didn't embrace the internet or computers in general, like it will just be very hard to compete in the future. So especially if you're a B2B SaaS company, you have a interesting like data record from your customers, you can add additional services where you automate some workflows, like almost like how your customer uses your SaaS solution for a certain, I don't know, like an HR solution, like you can probably automate a couple of those steps.
Rasmus Rothe: So you can learn from the data you already have from your users and then start to basically. Build more kind of agents in your solution and upgrade your solution. And then it's a question of, will you be able to do it as a legacy SaaS company, or will a new entrant come in that basically offers a much better solution? The New Pricing Paradigm created by AI Products --
Rasmus Rothe: And the interesting part, also coming back to the agent, it's also changed the pricing model, actually. So we've seen this, like, one thing is to sell software, but if you sell it more like a worker, it's almost in your recruiting budget. You're like, hey, you know, buy this, like, junior, like, sales. agent. It's, it's like half the price as if you hire somebody, but much quicker, much faster.
Rasmus Rothe: Um, and then people think, start thinking very differently about budget, which is a great opportunity for some of these companies. Exactly. How to attrach AI Talent -- Johannes Schabak: And that's, uh, talent is a fantastic segue into something that I always wondered and admired. about LF Alpha. It's amazing what you've accomplished. I have to say that, um, it's fantastic what you've built.
Johannes Schabak: It's really, really impressive. I have to say, not only from a purely German, but also from a European standpoint, it's really amazing. I have to say, this is super, super, super impressive. Um, and one thing that I always wonder how you do that in the beginning, um, how do you attract the best talent, um, in order to grow.
Johannes Schabak: Build a Salesforce, um, an enterprise sales with these massive long sales cycles, right? How do you do these deep tech integrations, um, bespoke, right? You don't only need to understand, you know, the LLM piece, the AI piece, but you also have to understand your customer that are, Big ships on their own. So how do you get the best talent on the planet, um, in Heidelberg? Johannes Schabak: Um, so this, this would be super interesting.
Jonas Andrulis: So I'm pitching Heidelberg to our international candidates as the German Tuscany and it kind of works. Um, so since we, and this is really where um, our last funding round helped us a lot. Like we, at the very beginning, we started as maybe one of the first Gen AI companies or, like, where we built these models when there was not a lot out there.
Jonas Andrulis: So we were able to, um, vacuum up a bunch of those researchers that wanted to work on these, um, but for whatever reason, maybe they couldn't get into open AI or they didn't want to move out or something like that. Uh, we were able to, to hire a bunch of those. Investors as Partners First --
Jonas Andrulis: Um, and then as the, the kind of, um, funding picked up as the more and more companies entering that space, um, I was in a situation where I felt we have to now position ourselves somewhat differently.
Jonas Andrulis: I felt that if we, um, cannot build a meaningful differentiator, then we will just be left behind. And this was also the reason why it took me about a year to raise the last round. Um, and. probably I aged a decade in that time. And we had a tons of offers on the table from institutional investors, from strategic investors, from Asia, from the US, from Europe.
Jonas Andrulis: Um, and really what we went with is quite a unique structure in that we found, uh, investors that were partners first and investors second. And we partnered with companies like the Schwartz Group that is currently investing about 2 billion into this AI center in Heilbronn. So it's really, there's a lot more strategic alignment there than basically just having them as customers.
Jonas Andrulis: We are closely working with their internal experts, um, to, to help them, of course, master this trans, transformation, but also they, they are helping us understanding their markets, understanding their verticals. And with that funding. We were able now to, to hire phenomenal people, like hire people out of Google, out of Palantir, in like leading positions and bring them to Heidelberg.
Jonas Andrulis: We also do have a Berlin office, but, um, headquarters is still, is still Heidelberg. Um, and we have all the senior leadership team there. So I think this was, this was amazing. Keeping one's ego in check as CEO -- Jonas Andrulis: And I'm now. incredibly happy to be around this amazing group of people. Um, but yeah, this is far from the end. Like we, we, there's still far where we have to go.
Johannes Schabak: Where do you see yourself, um, personally, but also the company in, say, 10 years? Jonas Andrulis: So personally, I mean, there's, there's always, I try to be, um, not to be too full of myself. Um, I am in a lucky position that I had my personal, um, part of, or my personal time of hubris, uh, during the time at Apple. Like when, when I joined Apple, I had this kind of feeling that I am the king of the world, right?
Jonas Andrulis: Nobody I'm invincible, right? And, um, and then in Apple, uh, corporate politics, uh, brought me down a few notches, which was probably for the better. Reinventing Roles and Embracing Change -- Jonas Andrulis: Um, and now I'm thinking about, like, On a regular basis, I have to reinvent my own job. My company's changing all the time. I am always critically asking myself, am I the right CEO for the next 18 months?
Jonas Andrulis: And there might as well come a time where the answer would be no. Um, and this is fine for me. This is fine. Um, let's talk when, when, when I actually make that call. But for now, I think when I think about it, it's, it's fine. Um, and. What I can also say is that the next five years cannot, on a personal level, cannot be like the last five years. Jonas Andrulis: This is not sustainable.
Tobias Schlottke: I can imagine. Rasmus's 5-year plan -- Tobias Schlottke: Rasmus, does the same apply to you? Rasmus Rothe: I Tobias Schlottke: don't know what else Rasmus Rothe: to do. I guess I'm just, I'm just really excited by, by, by, by like bringing more smart systems into the real world. Um, that's all always has excited me. And, um, Yeah, I mean, I've been doing this now for a couple of years, and I want to continue this.
Rasmus Rothe: So I think there's the good thing about our setup is that both with the services company, but I guess also as we invest and incubate new companies in AI, both incubate, but now also start to invest externally. There's always like new areas of applications coming that I've Not thought about before. So I think, um, it's super interesting to go after these, you know, different job profiles that that will change or do something in life sciences or do something that has maybe like a positive climate impact.
Rasmus Rothe: So also working on things like, I wouldn't say we're like Like an impact fund. Um, but I think we're generally looking at every venture case we're investing in. Is it like, does it make the world like even slightly better? Like, even if it's just a small bit. And so I think that's, that's super motivating. What is the Merantix AI Campus --
Rasmus Rothe: And I think on the talent side, I think we're also super lucky to work with very talented, uh, leaders and all our companies. And I think one big advantage has been the AI compass in Berlin where, um, ultimately creates a community because for a lot of the questions I think we discussed today. Like, nobody knows, like this, the technology is changing super fast, you know, there's regulation, there's like investor sentiment, there's moves of the big cloud providers, this is all like impacting like strategy of the business, and we all have a hunch of where things are going, but like, Things can also move very quickly and suddenly, I mean, I was, I remember I was with B Amer who was, you know, you know, stable diffusion, you know, discussed like how long it will take until we can generate videos.
Rasmus Rothe: And that was like in December. Then Sora came out and generated pretty good videos and he was like, probably takes a bit longer. So even the people that are probably some of the best in the field, um, you know, are always surprised about progress. And so I think we are trying, um, to create this community on the campus where.
Rasmus Rothe: You at least are in the room with other smart people and so can kind of a bit extrapolate what's coming next. What is the next exciting trend? And I think that that helps to attract talent. Um, and that, that has been ultimately, I think also one of the big advantages of the big tech companies that you are surrounded by enough other smart people to make sure you stay On top of what's going on and Tobias Schlottke: Yeah, Rasmus Rothe: and,
Tobias Schlottke: and others pay for your learnings, eh, Rasmus Rothe: Yeah. And like, yeah. Yeah. I mean, like you all, I mean, everybody I guess puts like a puzzle piece in. Yeah. Yeah. Like there's, and there's lots of interesting learnings. Like some are obviously super technical, but some are also like more on the business side, on the fundraising side. On the hiring side, it's like, it's not, I would say like one specific thing that is most useful.
Rasmus Rothe: It's a combination and everybody brings something to the table. Johannes Schabak: Yeah. Hiring The Right Fit AI Reseachers -- Johannes Schabak: And I would love to know how you set up your. company cultures. Because in the case of Aleph Alpha, right, it needs to be a super strong sales mindset, enterprise mindset, but also a super strong R& D mindset. And also engineering Jonas Andrulis: product. Jonas Andrulis: Engineering
Johannes Schabak: product. So how do you bring both worlds together? How do you set up this feedback cycle of, you know, the research, the researchers being super excited about the customers, um, and vice versa? Um, how do you build that culture in, in, I don't know, one year's time? Hire Jonas Andrulis: the right leadership. Johannes Schabak: Okay, Jonas Andrulis: so we have now as a co chief research officer, Yasser, who was with, built the AWS team in Triebingen and built Bosch AI research.
Jonas Andrulis: So he has a tremendous amount of experience building phenomenal research teams in corporations with like a customer value focus. So he is, he is doing a lot there. And of course, when we're interviewing researchers, I always challenge them on. What if their goal is to bring value in the world, or if their goal is to have a high citation count, or like in Europe's paper, and if your goal is just to have a cool, like, be on conferences and have a high citation count, then you're probably not the best fit for what we are doing, but I mean, this is, hiring is one answer, and then we're building our whole organization so that we're somewhat decoupling these teams, so that It almost never happens that a researcher gets significantly pulled into a customer's problem.
Jonas Andrulis: So we have the research is allowed to think about foundational problems with like different time horizons. And then there's sometimes researchers that are saying we would love to talk to customers, we would love to see what they're doing, and then we'll make it possible. Balancing AI R&D + Product + Enterprise Sales -- Jonas Andrulis: But of course, we are now have brought in people that have tons of experience with enterprise sales.
Jonas Andrulis: And you can imagine, those are people that have a very different mindset and culture and way of going about things. And this has caused some friction. And when we build up the sales organization, there was, there was, um, there was fear. in our tech team, they basically said, Oh, now we are becoming a sales driven organization, right?
Jonas Andrulis: Now everything that will happen is basically just what sales can can sell to any customer, right? They'll sell whatever a customer wants. And then we have to be the one that the team that delivers. And I told them that this is far from where we're where we're taking this, like we need to sell, of course.
Jonas Andrulis: But this is an innovation driven company, like we cannot survive it. by basically just having a great sales team. And so basically, it took a little bit of explaining and now kind of learning and experiencing this new culture. But certainly, it is a challenge and it is also what some smart people told me in the very beginning would not work with the setup of Aleph Alpha.
Jonas Andrulis: They said, you can build a company. You can kick ass AI R& D organization, but you cannot also build a great product organization and a great enterprise sales organization basically as one Pivoting an AI-Driven Product Due to Tech Teams closeness with customers -- Jonas Andrulis: team. Rasmus Rothe: Yeah, I mean, like, I think we are also constantly learning how to reinvent the organizations and set them up differently.
Rasmus Rothe: What's interesting, one of our companies, Brink, they They started in the ESG reporting space, tried to automatize that, and then basically when the LLMs came, they realized, okay, we can actually do much more for the customers using these LLMs to basically do much more elaborate document processing for those when it's about analyzing supply chain data and basically all the reporting, and it's not just like Small reporting piece they can automate.
Rasmus Rothe: And they suddenly, like last year, kind of pivoted the whole product. And that actually innovation came from the technical team because they said, look, with these new LLMs, we can actually do 10 times more than what we can do with a kind of a more rigid, traditional, rigid, traditional models. And so they're actually, it was really good to have the.
Rasmus Rothe: The tech team close enough to the customers that they would see what other pain points there are. And then they basically just shipped features in like two, three weeks to the customers and suddenly got so much pull that that suddenly became the direction of the company. And so I think the point is like, we actually, Try like, cause the technology is moving so fast and obviously the tech team is closer to that.
Rasmus Rothe: We want to make sure that they also translate that to what the customers needs because the customer often also doesn't know what they want because they don't know what the technology can do. So having a technical person kind of helping with that discovery actually has helped us to move quite fast.
Rasmus Rothe: And then the other part is being even more rigorous about just throwing stuff away. You've done before and not, not trying to vent everything in the house, but basically trying to plug as much as possible together. And also, you know, when you, when your stack changes, it changes like can change all the time.
Rasmus Rothe: And I think that's, that's also the right mentality you have with people, which sometimes when it's a more R& D focused organization is challenging because people obviously love what, what they're building and try to try to optimize it until the last minute. That's Jonas Andrulis: an interesting challenge, right? Because if you're.
Jonas Andrulis: on one end of the spectrum, then you're basically an Accenture or a consultancy, because this is the business they're in. Like they're assembling state of the art technology, best of breed, and then they're building something, and then they're kind of handing it over, and then off they are. But, uh, you also have companies that are building, like, own IP, right?
Jonas Andrulis: So it's always like a balance in the sense, I'm absolutely with you, right? You can never, you know, kind of fall too much in love with what you've built in the past, what may not be the way to go in the for the future. Or you can also like this, kind of this fallacy, not invented here fallacy, right? Where you say, the best thing, of course, is the one we can build ourselves, right?
Jonas Andrulis: This is, of course, a big danger. But on the other side of the spectrum, you're, you're, you're, uh, pivoting or you're, uh, transitioning into a services company, which you may also not want. So I think that's an interesting, Rasmus Rothe: uh, balance. I think ultimately from for the product companies, I think you need to understand what the technology can do, what the pain was of the problem is now doesn't make in the product roadmap.
Rasmus Rothe: I'm not sure. But I think these running these experiments have proven very powerful. Um, we had we had a similar thing with our customer. The LTI, like computer vision and manufacturing application. They suddenly also started to have just an LLM interface for their customers and tried it out. And suddenly customers loved it.
Rasmus Rothe: And they're just, you know, somebody just hacked that together in the weekend. And I think trying to, to, to try out these new things and figuring out how to integrate in the product is It's good for obviously AFS companies, but also for, to your question earlier, Tobias, to like SaaS companies to see how people can quickly hack something together that adds additional value to the customers. Creating a Loonshot Team to Nurture Innovation --
Johannes Schabak: That is called a loonshot. It's called a loonshot. It's basically a, like a very small team, one, two, three people, max, that do something entirely crazy. that is totally face separated even, even from the rest of the companies. They're totally insulated, totally on their own. They can do whatever they want.
Johannes Schabak: Nobody asks for money or timelines or whatnot, and they can do whatever they want, but you can only sustain maybe two or three of them, because if you were to do it too many, then obviously that would actually add to the cost, um, too much. So that's, that's why it's striking this balance of having sort of like a budget for crazy ideas. The Challenge of a Loonshot Team on Company Culture --
Johannes Schabak: How much do you. Do that, um, in your companies? How do you, um, do you have sort of like a budget or, um, how, where does innovation, like these singularity, almost ideas that then, you know, take on like a wildfire? Jonas Andrulis: I mean, like, this is what got Steve Jobs fired from Apple, right? The fact that he basically did something like that with the first Macintosh team, where he isolated, selected people, into a separate building.
Jonas Andrulis: They hissed a pirate flag over the building. They really did this. And they were able to do something wild and unthinkable. But of course, this caused the rest of the organization to, to be mad about it. Tobias Schlottke: The ivory tower. Jonas Andrulis: Yeah. And, and I mean, like, what, why can they do all the fun stuff? They get all the honor and all the budget and all that, and I'm like in this boring delivery organization.
Jonas Andrulis: So I think that's also a challenge here, and we have, um, we are lucky in a sense that we have this research team, which is a true research team, in the, in the, academic sense. Like we're publishing, we had two NeurIPS papers, or last NeurIPS, so we are really out to solve things that nobody knows if they will succeed.
Jonas Andrulis: And nobody knows how to do them really. And this is a luxury that most startups cannot afford, having a team like that. Um, and so I think we're, we're in a somewhat fortunate position here, but still, even within our company, there is always this, this, um, trap you could fall into where people are basically saying, oh, yeah, the most sexy place to be is research.
Jonas Andrulis: Um, and basically everything else is basically just, just the boring stuff that comes afterwards. And I'm, I don't get tired of telling people that productizing that is really difficult and is as difficult. And then selling something to world's best enterprises and closing like an eight figure deal is really difficult.
Jonas Andrulis: Those are different skills. But they're all essential to what we're doing. And it's not just that the most sexy and amazing and cool things are only happening in research, but certainly we structure the organization so that we can make that possible. Tobias Schlottke: Johannes, how do you, Apply that to your daily business? Johannes Schabak: You what do you mean? AI or waste management?
Tobias Schlottke: Yeah. Well, uh, uh, like do you, do you have an ivory tower that everyone wants to be in? And how does it look like, are, are you in the every tower yourself or? I'm pretty much Johannes Schabak: in, in an ivory tower, I think. Um, SumUp of is a portfolio of various products.
Johannes Schabak: So we do point of sale banking, payments, consumer. And, um, so within each of those verticals, we have our own little hubs of innovation, but SumUp has this cultural element of a loon shot also engraved in the culture. And we even have a special budget for it that are crazy ideas. And obviously those ideas aren't so crazy far off that we would now go into the energy sector or healthcare sector.
Johannes Schabak: So it's always something where we, there are still. a value for small merchants as, you know, they were to build up their, their business. Um, but it's something that would be, uh, almost science fiction, um, as to, wow, like a social network or a different payment method and, um, something unthinkable, but often over time and sometimes faster than you may imagine, it catches up. Loonshot: Can ideas come when isolated ffrom customers? --
Johannes Schabak: Um, that's how we do it. Um, but, uh, yeah, always very limited. Um, because we also feel that the greatest source of inspiration are actually our customers. And, uh, I think earlier to your point, Rasmus, right? I think if you ingrain in your culture, the, um, the value of driving innovation from a customer's needs perspective, that can also yield fantastic results, um, and come with best both of both worlds of, you know, there is a demand on the customer side, but it's also exciting from a technological side.
Johannes Schabak: Um, that's obviously where the place you want to be. Pitching to the Innovation Board at Aleph Alpha -- Jonas Andrulis: We, we created something we call innovation board, and that is where leadership and smart people, like senior people from the organization sit, and everyone in the organization can pitch an idea to the innovation board. And, uh, the innovation board is then able to.
Jonas Andrulis: supercharge your idea and put significant resources of the company of all orgs. So it's really like changing the course of the company, not just within your team, but basically saying, okay, we have to realign our resource allocation because we want to make this a reality. Uh, so this is really how we want to enable these ideas, uh, beyond just building a kind of research prototype or building something just in product or, right.
Jonas Andrulis: And I think that's really, really exciting. Uh, but it's also, of course, risky. Guiding Companies Through AI Transformation -- Johannes Schabak: How do you, by the way, look at your customers, right? Um, they are in a transformative process. It must be scary, but also exciting at the same time for them. When you come in, um, helping them transform their company, what are the types of emotions and how do you help them make that transition?
Johannes Schabak: Because it seems you're probably far more than just a technology provider. You're probably also sort of like an Agile Coach 3. 0, um, and helping them how to use AI. Jonas Andrulis: In a way, like, in a way, our product feels like the Industrial Revolution itself. Um, and so I am in a amazing position to now be in, like, really deep conversations with CEOs of world's best enterprises.
Jonas Andrulis: And that is, to me, also, amazing. Absolutely amazing because I'm really surprised and amazed by the quality of thought that's already there. Johannes Schabak: Interesting. So, what do you see? The majority of CEOs actually embrace AI and are fearful of losing out or missing out on AI? Or would you rather say they're excited about the opportunity? Johannes Schabak: Like, what's the Is there a more German angst or are there more?
Jonas Andrulis: So there's probably a little bit of a selection bias in the sense that the ones that are talking to me for for a few hours are probably not the ones that are hiding under the desk. And so I have amazing conversations and then during these conversations I'm asking them questions and I'm saying like, what do you think is the empire that does not exist yet that you are in a unique position to build with generative AI?
Jonas Andrulis: in the next five years. And then we're thinking about this and then they're saying, what is, what, because like new, Like, in this, with this new technology, a new piece of land has been discovered, and nobody owns this yet. And this is value that can be generated. And tech companies, tech companies are coming at this, and they're saying, we are going to own all of that.
Jonas Andrulis: And when you look at the development of price earnings ratios on the market, the market kind of seems to agree. Like, the price earnings ratio for tech, gone up, but for all other sectors, it was basically flat. So, the market thinks that a lot of the value that's being created. will go to tech. But Microsoft alone cannot revolutionize manufacturing.
Jonas Andrulis: They have no clue about that. For them, it's all just data. But we have in Germany, we have phenomenal manufacturing companies that have this knowledge, and they have teams, and then they can do things that nobody else in the world can do. And so, You always have to combine the vertical knowledge, the vertical data, the vertical distribution with the technology and basically what I'm, what I'm trying to inspire people to think about is what are the unique moves I can do?
Jonas Andrulis: Like, what am I in a unique position to do? Where is knowledge work for, for my future position in five years and 10 years, what's the most important position to be in? And I think those are incredibly interesting conversations and I came out of conversations with CEOs, um, really inspired by the ideas they put on the table, uh, in these conversations. Tobias Schlottke: Guys, we slowly have to come to the end. Um, was really great. I have a last Johannes Schabak: question.
Tobias Schlottke: I, I also have a last question, but maybe you go with, with your last question first, , because before I go with the last question, How can CTOs get the rest of the company onboard AI -- Johannes Schabak: um, I, as our audience is very technical, right? I, um, think it would be super interesting to hear from both of you, what would you recommend, similar to the conversation that you just described, uh, with CEOs of, I don't know, DAX companies.
Johannes Schabak: Uh, what do you recommend CTOs who may be in a company that, um. Well, either totally embraces AI and is super excited on the one end of the extreme. And what would you recommend to those CTOs that may not be finding themselves in a place that is super excited? Tobias Schlottke: Wait, wait, wait. You just stole my last question, Johannes. Tobias Schlottke: See? Johannes Schabak: Great minds think alike.
Rasmus Rothe: I think first of all, it's about training. So you need to have your whole organization try to understand AI to a basic level. So I mean, there's a lot of courses out there. Even if it's your marketing team or it's your, you know, sales team or customer support, like give everybody like a basic training for AI because they, they're ultimately your internal customers. Johannes Schabak: What would you teach? What would you teach? A bit more concrete.
Rasmus Rothe: I think like the general understanding of these AI models, what kind of data. Prompt Johannes Schabak: engineering. Rasmus Rothe: What, how does prompt engineering work? How do they, how do they maybe basically internally work? Like what are kind of some of the data privacy things you need to be careful about?
Rasmus Rothe: What are some cool applications that maybe others have built in your similar domain? Um, so I think that's, that's one thing. Encourage people to try things out -- Rasmus Rothe: Um, second thing is, um, embrace people to just try things out. I mean, there's some companies who say, look like we, we want to stay away, but the majority, I think should just embrace the people to try things out.
Rasmus Rothe: Use like maybe LLM and some of their workflows or try out some AI tools, maybe even bring it up to like your internal budget department and be like, Hey, I want to try out this AI tool. Like, can I, can I use it? So, so the bottom up people start using it and are not scared of it anymore. The Role of the In-House Champion for AI --
Rasmus Rothe: And then the other part is, I think you need probably like one person in the company that, you know, is kind of a head of AI or really at the forefront, a champion to, to have an overview of also what's going on outside and kind of translating that it could be the CTO.
Rasmus Rothe: But like, you might not have the time for this. So then maybe getting one person who's really dedicated and has an overview of like, you know, where, where, where somebody I already use, what's happening outside in the market, what should we build in house where there's maybe solutions outside, like to have this kind of moisture cheek overview and either hire that or ultimately, you know, promote somebody internally, if there's somebody talented, do you agree on us? 3 Levels AI Company Maturity --
Jonas Andrulis: I kind of agree. Um, I would add to that, that if you are, if your strategy is just, like I was on a stage with somebody, an AI AI leader, and they said that they're currently giving everybody in the organization a chat GPT training, I think that's great, but if your strategy is basically just to training your people how to use things that others have built, then you shouldn't be surprised that no future value creation will end up on your desk.
Jonas Andrulis: So what we are doing is we have like a three levels of maturity. Level one is prevent an irreversible shift. So make sure that you're not losing out, that you're not losing essential data to somebody else, that you're not losing value chains so that you cannot easily get them back. It could also be like collecting your data set, changing your, your, uh, your way about going with data.
Jonas Andrulis: Second thing is, um, the. business transformation. So quick wins, generate value, like Bosch already has a positive return on investment of our tech. So that's important, right? Create value fast. A lot of this is about efficiency. Position your Team to Lead the Next Industrial Revolution --
Jonas Andrulis: And then the really sexy thing is the third, which is positioning your team so that the next industrial revolution, that the next transformative idea will come from your team, or can come from your team.
Jonas Andrulis: And I think this, requires a little bit of setup. How can you enable your internal experts so that they can come up with transformative ideas? Um, and I think this also requires a lot of like organizational building and, uh, bringing the people into the right, uh, kind of context and all that, right? Because you don't always want to be, the trying to catch up to a departing train.
Jonas Andrulis: You want to for for kind of one or two things, you want to be the leader globally for this transformation. Bottom Up and Top Down -- Rasmus Rothe: So in that sense, it's like bottom up and top down. I think like it's bottom up empowering everybody and like, you know, starting to use it and fostering innovation and like some units where you maybe never thought that people would use some of these new technologies.
Rasmus Rothe: But then also I think like creating, like, as you said, with your Bosch example, like one lighthouse case in your organization that is more pushed top down, maybe, maybe based on this information where you are also putting really resource behind where you say, look like this is not just some fancy AI demo, but there's a real ROI.
Rasmus Rothe: We used it here. We invested, we built something. There's a positive ROI. It increases sales. It increases efficiency. We can put a number on it, uh, like really commercially. And then also have that as a case study to then again, inspire more people like bottom up. And I think what we see a lot is especially larger organization.
Rasmus Rothe: There's a lot of hype around this. And so people try to do what looks fancy is visual, but not maybe necessarily what maybe, you know, creates crazy efficiency or had, you know, can increase your sales. Um, so sometimes it's a more boring department, so to say, where there's the biggest, biggest gains. Tobias Schlottke: Thanks a lot guys.
Tobias Schlottke: Great discussions. Um, really enjoyed it and, uh, hope to see you soon. Thanks for having us. Thanks. Thank you. Thank you for listening to the Alis podcast. If you like this episode, share it with friends. I'm sure they love it too. Make sure to subscribe so you can hear deep insights and to technical leadership and technology trends.
Tobias Schlottke: as they become available. Also, please tell us if there is a topic you would like to hear more about or a technical leader whose brain you would like us to pick. Alphalist is all about helping CTOs getting access to the insights they need to make the best decisions for their company. Please send us suggestions to cto at alphalist.
Tobias Schlottke: com send me a message on LinkedIn or Twitter. After all, the more knowledge we bring to CTOs, the more growth we see in tech. Or, as we say on Alphalist, accumulated knowledge to accelerate growth. See you in the next episode.