← Industrial AI Podcast
Industrial AI Podcast
AI-Engineering Workflows
01:34:56
From data silos to autonomous agents—discover how AI is reshaping engineering workflows. Insights, stories, and the tools driving the next wave.
Loading summary
Transcript129 lines
- [00:00]
Stefan Zuvalak
This podcast is presented by nxai, your partner for time series foundation models and physical AI.
- [00:18]
Robert Weber
Hello everybody and welcome to a new episode of our industrial AI podcast. My name is Robert Weber and it's
- [00:25]
Peter Sieberg
a pleasure to talk to Peter Sieberg. Good morning, Robert. Did you arrive home, well, from our AI in the Alps?
- [00:33]
Robert Weber
Yes, I arrived I think at midnight on Friday and yeah, we spent three days at our event, AI in the Alps, in the wonderful Lech or Zug at Lech near Lech. So, yeah, it was amazing, I think. What was your. What is your opinion on the event?
- [00:53]
Peter Sieberg
Yeah, yeah, normally you want to know what. Who was the best or was the best. Right?
- [00:58]
Robert Weber
Yeah, exactly. That was my second question.
- [01:00]
Peter Sieberg
Now I'm going to start. I'm going to start. I'm just going to pick out one which was. I was impressed by the work that Bastian. Bastian Aud, he's the CEO of simcon. Right. And he's working with simulation for injection molding. And that came up during. That was my use case session with Marco. Right. That is Professor Marco Huber from. From the Fraunhofer Institute. As we discussed physical. We weren't certain if that was physical AI or more physics informed AI.
- [01:33]
Robert Weber
So it's for the buzzword bingo at the evening.
- [01:36]
Peter Sieberg
Yeah, was. Interestingly, that's. That's the way they were written down on the. One of the flip charts.
- [01:43]
Robert Weber
Right?
- [01:43]
Peter Sieberg
Yeah, yeah. Independent of the buzzword bingo. I was really impressed with what he's doing and for that reason I'm going to do a podcast recording with Bastian over the summer, I guess. So all of you can kind of learn directly after our break what it is exactly that he's doing. But basically, you know, it's about simulation for injection molding. So just really, really. He was using a very high number, improving the quality and quantity of the mold that you produce before you then can start producing, know whatever you're producing with the injection molder. Very impressive.
- [02:25]
Robert Weber
I think he does that together with Johannes Pantstetter, or am I wrong?
- [02:28]
Peter Sieberg
He has been. They have been working very closely together. Yeah. And there were some details that he was sharing also related to. Right. What they had been doing together. And then as Johannes Co. Got. Got sold, got bought, they had to kind of dis. Dismantle pieces. But yeah, they have been doing specific
- [02:51]
Robert Weber
stuff together and they also AI in the apps, I think three years ago or something.
- [02:56]
Peter Sieberg
Two years.
- [02:57]
Robert Weber
Two years, yeah, two years. Yeah.
- [02:59]
Peter Sieberg
I do recall two years ago that Johannes there was. Yeah, right.
- [03:01]
Robert Weber
Okay.
- [03:02]
Peter Sieberg
So that was, that was. That One best of the show award, which we don't have. But that's just a thought just to share that always at the end of the three days we always do this plus and minus thing and it didn't come up. But I just thought this morning like maybe we should do that. Maybe we should do that. And if we would do that and if we had done it this time, I would most certainly give to Zep. Zep, that is Professor Zep Hoge, inventor of LSTM27, 28 years ago. Meanwhile co founder, chief scientist at NXAI Now Zepp, who actually regularly reminds me of myself. He talks with his hands a lot and that is what my friends tell me or. But they joke about me when I do that and when I get excited as well. So there you go, Sepp. In his typical very impressive, exciting and convincing way he explained what he says. And if and not if, I mean I understand it like that, I'm going to share it with you. The revolution that is being brought to us by Tyrex2, which is the multivariate time series foundation model which you could all dear listeners hear about last episode. Yep here about like a week ago, a week or two weeks ago. And it was actually announced on the very first day, is that right that we were there on the. On the Thursday.
- [04:29]
Robert Weber
Exactly, exactly. Right now went live I think on Thursday.
- [04:34]
Peter Sieberg
Thursday morning. Yeah. Right. Now what is so revolutionary about it? You do not anymore need to train your model as we we is a little bit too strong as some of you know. And I've I've said before many years ago I was involved but it's manager never like as the scientist but I understood enough and I'm going to share with you one example there just to tell you. This was about a testing machine. A testing machine for one of the structural components of the ASML chip production machines. So I'm not going to go into the details as I signed an NDA a long time ago and but think about something really, really structural without which the machine would not function. And at that time that this testing machine was broken, it had been broken for whatever reason and we were brought in and that was at the very, very, very first time coming out of softing the industrial data solution or something like that. And now we received like a year we worth of data of about 150 different sensors. Right. And our job was to tell them if with the means of a dead time machine learning, you know, this could have been circumvented, you know, so is machine learning Better so to say, than until that time just looking at condition monitoring. Condition monitoring. You put a threshold level at whatever temperature and if it passes. No, that's what we did. Now, as soon as then a component of this testing machine, the new component arrived, we found out that our model did not work anymore. You know, there is a component, but it's kind of unique and of course it needs to be, it needs to have the same spec, but there are somewhere small pieces and the model would all the time come up with a false alarm. Right. So that's what we've. And that's the kind of story that we've been hearing all the time. I think today we're going to be listening, is that correct? To me, to me talking to Stefan. And he will be talking along the same lines. Why predictive maintenance has been so difficult, so cumbersome in the first 10 years. Now coming to the point, and I did share this example with Sepp as I was in the room with Sepp. So we always are there with around eight, nine persons attendees. And you and I, we go around a little bit and, and I did it for the very reason that. Yes, Zepp confirmed. Yes. Today you would have just plugged in the data into Tyrex 2 and after some time the model will start understanding what it sees. So no need. As we needed to do almost 10 years ago, we needed to wait for another three months or four or five months, I don't recall. And we needed to gather new data, we needed to build a new model and only then, so we lost almost like half a year. Now that is of the past and that is the revolution. So I'm not, I wouldn't suggest, you know, best of the show award. Yeah. Nevertheless, yeah, maybe there's a second thought, maybe that maybe that would go to Tyrex too. But it's really Sepp who, who I saw him, you know, doing a similar kind of discussion presentation four times in total. And it's just amazing that, you know, we can have this guy who is, I would say, one of the world's top 10 AI researchers right with us two days in the Alps and provide each of our 30 plus attendees one and a half hours in a small group interacting and as much time as each of them wanted. Right. You know, during dinners and during break time. So thank you very much, Sepp.
- [08:44]
Robert Weber
Yeah, perfect. So my, my highlight was a little bit different because we had a track on AI and engineering by Christian Heining. He was a newbie being a host at our event. And for me it was Very interesting because the focus shifted a bit to the question what will my organization look like in the future? So no, not a technical discussion about which tool, how to integrate CAD Cam. More about the topic. The team of one. How realistic is that? And how do I build the knowledge enterprise. And there are two. Memory plays a major role. And we will doing a separate episode on this topic I think because everybody was really impressed by this session thinking about the organization and how AI engineering organization look like in the future. Very impressive. Thanks a lot to Christian. It was very, very cool to have you on board and to be a host in the session. It was really interesting how the whole topic shifted from from a technical discussion to a, to a organization discussion. Right.
- [09:52]
Peter Sieberg
As when I was in, in, in that track which, which took place in the, in the cellar of the building which was on the kitchen cellar. Right. With a wonderful cool cl. Sometimes it was a little bit warm. Yeah, we had, we had a similar, I guess I feel like a similar discussion. I think the question was raised if the product management job would continue to exist. But I think the meaning was I felt like is what Andrew Wang now this time again in a positive sense because I believe that he was one of the first, maybe already like 6 months ago he was suggesting that because so what follows after. So you start with the designer of a product, then the engineer and maybe that goes together or the product manager will decide what it is that he or she wants to build and the designer and the designer will pass it on maybe to the engineer, maybe to the software engineer. And because the software engineering is almost like going three, four, five times as fast, you know, as in the past, you don't have enough product management or designing capabilities. And that I think that was the discussion. If you're going to be doing a podcast with Christine, a topic I'm very, very much looking forward to, I think the idea is. Right, so who, who, who is the originating person? That is the thought, I believe. Right. And you could be a one person, you know, show for example, in a big organization, maybe instead of 100 people. But you're moving capabilities to the front. That's what I believe it is. Right. So if you know what do as a one person company or as a ten person company, doesn't matter. You are going to be able to do so many more things that in the past people after you, you know, like the engineers would be doing and you as a designer, which maybe has never done any coding. And that's what I'm, what I'm going to be Talking in a couple of days with Scott Duncan. He is a guy in, I believe, chemical industry. He's never written a code and he's written an agent. So maybe that's an example. Looking forward to. Yeah, that was a great session.
- [12:03]
Robert Weber
And then there was another discussion about KPIs and AI and engineering. What are the KPIs, is this token consumption? But we had no answer at the end. So we also need to record an episode on KPIs, I think. Yeah, yeah.
- [12:18]
Peter Sieberg
Well, I mean, I would say the tokens are on the negative side. What I say, I mean, in the end you need to, to do something like you, you produce something and needs to have a high quality and all these good things. Right. In relation to the cost. So the, the, the tokens are cost. And yes, I think we've seen since maybe a couple of weeks, only couple of months, only that yes, there is a price to be paid for high quality artificial intelligence language models, agents can be maybe, cannot be, you know, but that there is a discussion on maybe it's the part below the equation, right? Yes. And on the positive side, there is always this. In the past, we were looking at, we were writing codes, we said the million lines of code produced. Right. When we did it on the IBM mainframe or whatever, which is in itself, of course, never a means of quality. Right. You can write whatever you want. And I think we're moving a lot these days towards accomplishment. You need to, you need to, you need to do something which is worth worthwhile. Right. Which has value. So maybe it's the value, I guess, in the end relation to the. Yeah, but we'll, we'll find out. You're thinking you're going to join that discussion also with Christian then, or have I gonna do a separate one?
- [13:41]
Robert Weber
Yeah, yeah. Maybe I will find somebody else who can tell us a little bit about how to Measure what are KPIs, what are realistic KPIs. Do we have KPIs, how to develop KPIs, something like this? Yes.
- [13:55]
Peter Sieberg
So we cannot name all first the way was said.
- [13:58]
Robert Weber
But I have one more topic when it comes to AI and engineering and tools because I shared a picture of all our flip charts on LinkedIn. And then a colleague of mine who was not able to join AI in the Alps this year, he wrote me a WhatsApp and he said, wow, that's very cool. I already put it on my AI and analyzing your results. And thanks a lot for that. So we need to be careful at the end, Peter, only sharing the flip charts. And the people are using their AI to analyze the flip charts. But yeah, greetings to Hans Michel. Great approach, great idea. See you next time at AI on the volcano.
- [14:42]
Peter Sieberg
And of course, we're very much looking forward to these results that he has received from whatever large language model. Exactly. Now, I think the. Yeah, and I think it's perfectly okay to still share the flip chart, which is the only thing that we write down. We don't do PowerPoints just to remember. I think in the end, my feeling is the ultimate winners always. But I felt it very strongly in general are the attendees, I mean, especially I would say those from the small mid side companies, in addition to those of the big companies. Big companies like Boris, you know, Boris, who received a lot of interest for his book. But the guys and the girls, I come back to that there were two women. We need more women in the future. So you dear women, female listeners, listen to what we're going to be doing in the future. But those of the big companies, they have many colleagues, you know, from around the world and they, they can much more easily within their big company, you know, check where am I going, where are my colleagues going, where is the United States going, where's China going? Et cetera. Those of the smaller companies, I feel that they are so, so happy with being able to meet face to face in the sessions, but also in the breaks during the meals and talk to, you know, talk to the other people, talk to the other attendees and, and at the end they go home. And that's what we always say. That's your quote, right? Learn new people, learn new things. Right? And that's what they do. And they feel okay, and they feel like, oh, yeah, I'm spot on, I'm doing everything right. Or they say, okay, ABC is good, but I've learned, you know, XYZ in addition. And I need to do something else when I, when I. And they all go home with thoughts and ideas of, you know, what it is that they're going to be doing from now on.
- [16:43]
Robert Weber
Perfect. Because you mentioned China and the US we have had a very interesting session on sovereignty by Robert Josic from Schwartz. But sovereignty is exhausting. So it's really difficult. Right. So the whole discussion, what to do, I think we raised some question, we raised some ideas in the sessions, but it's a long way to go. But I think everybody was very keen to go this way, I think.
- [17:11]
Peter Sieberg
Yes. And it was the same when I was in Robert's session. The main thing I took away, and I believe Everybody agreed, is that you try to maximize sovereignty. So I think, because we were looking for, you know, what is it like, Maybe I'm looking not for rules. What is it like, Suggestions, you know, What a suggestion. 1, 2, 3, 4, 5. And we could come up with some. We talked about open source, we talked about fallback solutions, all those wonderful things. But what was very clear that none of us, and that's why I think we agreed, is that you cannot have 100% sovereignty. So you can at best strive to maximize sovereignty. So you try to go from whatever, 70 to 80% or maybe 80 to 90%. And that feels very much. Maybe that's the thought I have now. I didn't share it at that time. It's very much the same, like the OE, overall equipment efficiency. You know, there's a theoretical 100% what comes out of your factory. You know, 100 cars at 100% quality. You don't get that typically, you know, in, in what we do, producing goods, you know, maybe you get, you know, 90% what in process typically last 60, 70%. And maybe that's similar. You strive to get to the 100%, but you, you know that today you can't get 100%. Yeah, I agree. It's a rather complex topic. It's a little bit different from the things we've been doing before. And I guess we're going to come back to that topic in the future.
- [18:51]
Robert Weber
Yeah, but we had a discussion because my opinion is, sure, we lack, we lack balance, right? We lack what China and the US absolutely want, except asml, for example. And I think we need to think about what to get what, what should we invest and what kind of technology should we invent that the Chinese and the US absolutely want so that we get back as a balance, right? So the Chinese want X, we have X, we want Y. They have Y. The US guys have Y, B, and we want B. But they also want A from you, from US in Europe. So they get A. So in the moment, there's no balance between the, the three big parts in the, in the world. And I think maybe it's a topic for regulation, right? So we also talked about regulation, but maybe it's also a topic to invent something to invent. To invent. To invent in Europe what the others are missing. So I got some, some support for this thesis by Werner Faulhaber from AB Work. Let's see. But yeah, as I mentioned, a very, very difficult discussion.
- [20:02]
Peter Sieberg
Yeah, but there's two, at least two. There's Many more probably. But there's another perspective. So the one that you just talked about is the, you know, what shall we. You mean in this case, you mean Europe, I guess do. Right. Which is very political. And that's all about. And you talk about the three big ones. Well actually as far as AI is concerned, maybe there's only two big ones today. And we do a lot of, we invent a lot of stuff but typically we don't get it to the street. Right. The second perspective, which is what I was very much interested in, is giving our listeners, giving the people in the companies who, who struggle with exactly the same thing. Not about who am I going to vote when I can vote in democratic Europe and what is my politician going to do? As far as AI is concerned, that is one thing, but the other thing is what can I do today, what should I be doing tomorrow? So that is the approach that I was very much interested in. But maybe the other one we do, I believe anyway, I mean all the time we look at what is coming out of China. Completely different approach to large language models. Yeah. What's happening in the United States? Oh, suddenly we can't use. Oh in the meantime we can use entropic fable again.
- [21:25]
Robert Weber
Right.
- [21:25]
Peter Sieberg
What's happening there? So those things we talk about all the time and those are in the end they're becoming so political that it's. Yeah, I mean we can and we shall talk about it, but we cannot make changes today, you know, things that we're going to talk about today and in the next couple of months if that's what we want. They are maybe going to influence politicians in, you know, the next one year and then maybe in five years we're going to see any changes. I really believe so. The other thing though, which is I believe very, very important you listeners, you within your companies, more so even the smaller and the mid ones again because you don't have so many colleagues who you can talk with. You need to know and maybe you need to hear from us. Maybe we should do that also. Robert, at some point in time to tell them what can you consider doing? Open source is one example, but not absolute. Right. What is the advantage, maybe also disadvantage of going open source and all the other things of doing perfect.
- [22:27]
Robert Weber
What else do you have, Peter?
- [22:29]
Peter Sieberg
I also want to mention Joshi. Joshi, the owner of the Rote Wand, which means Red Wall Hotel, which you know since as a kid he showed us Friday afternoon a walk, the walk that we typically do, who brought us to see a water as energy based wood saw. That was really great to put nature in perspective to the technology that we had been talking about for a couple of days. Right.
- [22:59]
Robert Weber
It was a sawmill. Right. So from exactly 1797.
- [23:04]
Peter Sieberg
So woods or.
- [23:07]
Robert Weber
Yeah, saw milk.
- [23:09]
Peter Sieberg
It's very good.
- [23:10]
Stefan Zuvalak
Yeah.
- [23:11]
Robert Weber
It's likely like the Cloister people.
- [23:14]
Peter Sieberg
There you go. As a maybe putting it all together, you and I, I believe, being very happy what, what we've been doing since. And I think you wrote on 7 years. I looked at it with the first podcast we did on January 19th in 2019. In that one we explained the format. We shared what our listeners quote, unquote, why AI is and will continue to be so important for the industry. That's what we said more than eight years ago. So in all humbleness, I think we were quite visionary. The AI in Acts gatherings, starting with Wurtzburg, they came later. We should also say thank you to Barbara, who has been, you know, one of the three of us that we started these gatherings together, was on site for a couple of years and I think we've provided astonishing value for the attendees and which have in the meantime then become an integral part of this network which I just mentioned, you know, just feeling, you know, I'm in sync with the other people in the industrial AI world now. I think there are still enough, more than enough industrial AI topics. So as far as I'm concerned. Robert, let's do some more. Maybe you want to share.
- [24:38]
Robert Weber
I have one more tool tip because I think Christian shared this tip with the others. It's called Zakana AI. It's from Japan and it's a multi agent orchestration and model connections. So maybe try out Sakana AI Sounds very interesting. I already did an interview with Jakub Tomczak on Zakana and he also said, yeah, very interesting approach was released in June this year. It's called Fugu the model. And yeah, maybe try out Interesting approach. Peter, thanks a lot. But what is on our event agenda in the next coming months? Because we planned something.
- [25:15]
Peter Sieberg
Yeah, but you know better than I. I think the next one is AI on the Volcano. Right?
- [25:20]
Robert Weber
Okay, okay. AI on the Volcano. Yeah. So everybody was asked us, what does it mean, AI on the volcano. And yeah, there is a volcano in Parkstein in Bavaria and there's a company based in Parkstein and the name of the company is Vitron Intro Logistics companies market leader when it comes to retail logistics. And they invited us in October to run our event there, AI on the Volcano. And then we plan a trip to Paris visiting the station F the startup ecosystem in France, there was also a discussion in our sovereignty track about France and what the French government is doing differently to German government. But this is another story. But yeah, if you want to join, apply, send us an email and we will see what is possible.
- [26:14]
Peter Sieberg
Was there a potential alternative to Paris discussed or suggested as well? Or we still want to leave that open?
- [26:21]
Robert Weber
Yeah, yeah, yeah. I'm not. It's a surprise. Maybe. Maybe we take another plane. Different plane at the airport.
- [26:32]
Peter Sieberg
Yeah, right. Yeah. There is enough places in Europe we don't have to fly to Japan. We can do this sometime as well. But we do still have enough cities, I believe, for the next couple of years. And as we're going, there's going to be, I believe, more areas, more conglomerates, maybe developing artificial intelligence.
- [26:57]
Robert Weber
Yeah, I got, I got the hint by a guy and he said you should fly to Varsha and not to go to the west, go to the east. Because there's also a lot happening in the east.
- [27:07]
Peter Sieberg
Yeah, yeah, yeah, yeah, yeah. In general, I would say on a. If you look on a European level where, where things are moving, I mean, there is a lot of stuff happening in the. Shall I say, more developed from the past cities. You know, if we look robotics or we say Munich, Zurich, just an example, there's another city at some point in time we can go to. But having said that, and not not only, I mean it more from the general development within Europe. The eastern countries are building and moving forward and becoming more important. Important in general within Europe all the time. So let's look at those as well. Thank you, Robert. Thank you attendees. Thank you dear listeners. Maybe you want to join us next time. Looking forward to.
- [27:57]
Robert Weber
Thank you. And now we switch to the main part with Stefan Zuvalak. He was also an attendee at our AI and Yups event. So yeah, yeah.
- [28:05]
Peter Sieberg
And the first thing you're going to hear is that I'm asking Stefan, because there is this story when in the past I would say, you know, we're going to be doing whatever software engineer, I'm going to be doing agents. And then you, dear listener, the engineer would come to your customer and they would say, you know, Robert or Peter said, you know, we're going to have agents. Can you let me have an agent? And I would typically use the word of Stefan. He would be not happy sometimes, I believe. And until a year ago, it was like, it was like, Peter, don't say things like that. Or maybe now Stefan has turned around 180 degrees in the time that is the question I'm going to be starting and Stefan will explain us why he has turned around.
- [28:50]
Robert Weber
Perfect. Thanks a lot. Bye bye Peter.
- [28:53]
Peter Sieberg
Talk to you soon. Bye Bye Robert. Today I'm going to be talking once more to Stefan Zuvolak. He's the co founder at Renumix and he and I today are going to be talking about who would have thought that possible agentic AI in engineering. Hi Stefan.
- [29:14]
Stefan Zuvalak
Hey Peter. Thanks for having me back.
- [29:16]
Peter Sieberg
You're welcome. Yeah, the one or the other listener may actually recognize your name or your voice. Last time you were here, that's about what do you think, how long ago?
- [29:28]
Stefan Zuvalak
Actually, I don't know, maybe like 24. Something feels like a very different word. I think we were talking about data centric AI and all the things and now like we have completely different topics. Very, very, very interesting.
- [29:42]
Peter Sieberg
Yeah, and a lot has happened in those two years here. So nevertheless, please introduce yourself to our listeners.
- [29:50]
Stefan Zuvalak
Yeah, I'm Stefan Zuvlak, I'm one of the co founders of Renumix. My academic background is actually in like medical engineering. And then we founded Renumix in 2017 with the goal of bringing machine learning and AI to engineering.
- [30:06]
Peter Sieberg
Very good, Stefan. You and I talking agentic. Who would have thought? So let me give a quick explanation why I say so. So please hang in there for a minute. So there was a time when I would look into the future, actually I always do that. I believe. And I talk about how we engineers, I'm one of those as well, we should eat our own dog food right by. And I think this one specifically was about having copilots support us coding and how in the not so far future person or business assistants in the form of LLMs, you know, they would have agents do whatever, you know, we would have asked them. Right. And then I almost felt guilty for being like a smart ass, I think. And I would say something like yes, I can easily talk and say what's coming while you out there, you dear listener. And then I would typically call out your name. I believe I would say, you know, like guys like Stefan, you walk into a customer and they will say, oh, Peterson Seaburg said we're going to get these agents who. Can I get one as well please? So I always really felt very sorry afterwards for calling out things that I believed and I still believe sometimes and I'm not always right, so be very, be very careful but that I believe that we're going to come right? And you guys girls out there, engineers, you know, you had to develop them. So also part of it. I wasn't always certain that talking about agents that you, Stefan, you know, were that convinced maybe, I don't know, a year or two years ago that agents were going to play a role or at least maybe that you did not always feel happy. So thanks for hanging in there. Now we are here again back. Great to have you and we're talking agents. Seems you have come around. What happened?
- [32:01]
Stefan Zuvalak
Yeah, I think that's probably a super, super long story. I try to maybe tell it as short as possible, but yeah, absolutely. I think one was about one and a half years ago. I think I was a very big skeptic and I think there are several reasons for that. I think obvious reason is back then the technology just didn't work well enough for like real workflows. I think at that time that was true. But I think what's even more important is that I think in the whole field of industrial AI we had, I think quite some rough years and maybe I can explain it a bit and give some examples. So when we started in 2016, 2017, I think this was maybe like the first year when Elon Musk promised everybody in, in six months there will be robot taxis on the road. Right. We distinguished between a weak machine learning or weak AI and strong AI. Actually you don't really use this terms anymore. At the back at the time we said, hey, weak AI is something you cannot just apply to a very narrow use case. I don't know, say if there are apples and oranges on an image. And strong AI is you really have an AI that have a broad knowledge. Right. And at the time we only had weak AI. And this was a big, big problem because AI was a kind of like automation tool for processes, but very narrow, very costly to implement, could fail easily. So. So I've seen, I think a lot of failures bringing good demos to production, so to speak, and very few success stories. So I think I was very wary about like the new hype and the new demo that might not fulfill expectations. And I think the biggest reason is, and we can go into this, into more detail, is now we have strong AI and I actually feel it's underrated. I think people underestimate like how strong this is, how much background knowledge it has, and so on and so forth. And I think that was the big change that I think I see very clearly now, but wasn't aware at the time.
- [34:07]
Peter Sieberg
Yeah, you're right. This dividing between weak and strong AI, that's really part of the history of AI, I would say when I would do maybe up to like 10 years ago when I would be giving this introduction and I forget the name, there is a. The terminology of weak and strong is related to one specific person. And in this case I believe it's not Jurgen Schmidt Hoover, which is. Yeah, many times it is. Many times all these things are related to him. In this case there is like maybe Canadian. It's not important. You're right, we don't talk about that anymore. So in the meantime, thanks for that. You even completely overhauled your engineering software, right, earlier this this year because the fully agending approach worked so well. I understand. Now we're going to be talking about that in a moment, but for now let me first ask you to share with us what you see as the challenges in engineering and the role that gentic AI can play there.
- [35:07]
Stefan Zuvalak
Yeah, I think obviously engineering, especially European industry, I think is in a very rough spot at the moment. And I think there are obviously many reasons for that. We don't have to go into all of them. But one really big problem is that in the past, right, we built really complex products like a car. And we did it by divide and conquering. We divided the car into smaller components, ECUs and all the components were developed independently. And now we see that doesn't work any longer. It's much more important to have the end to end experience. Obviously in automotive you have this software defined car, but in other industries the same applies as well. And now what we have to do is we have to develop end to end, but we have an engineering culture at the moment. We have data infrastructure that's super. We have like the whole system is basically built to have little data silos. And engineers are really like, at least like 1 1/2 years ago they were like most of the time building reports and like shifting information from one data silo to the other, but not doing actual work. And I really believe that agentic AI is the technology we need right now because it's amazing at breaking, breaking open these data silos, giving everyone access to data, to tools to have. Like every stakeholder gets the ability to see the full picture. And I think that's exactly what engineering needs right now.
- [36:41]
Peter Sieberg
Okay, sounds almost like the same problem we hear with all these other jobs as well. You know, like medical, you come from medical technology. I wasn't aware of that. We say doctors, you know, they're not doctoring so to say, but they're writing reports about what it is they've been doing. So in general then how does it work? Let's go One step further. I understand that you think in layers of which the agentic layer would be one. Can you, you share that with us?
- [37:16]
Stefan Zuvalak
Yeah, I think it makes sense to have like a model and at most people, I think you use it in some sort of others to have like basically like three layer cake, so to speak, where at the bottom you have your data sources, right? Your different databases, systems, but also tools you use. I think it's a known problem that at for example, the German automotive oem you have like hundreds or even thousands of different tools and databases and most often they don't really talk to each other. You have to get to know a new tool for specialized work and we have to standardize this and this is something that has been on the way for a couple of years now. But I feel with like the new technology also talk deeper about that like mcp, this will be standardized and on top of that you will have a new layer. You will have this agentic layer, you will have, have agents in a network if you'd like, or ecosystem of agents that can use the bottom data and tool layer to actually get information, execute work, do studies. And this can feed, and this is for me like the top layer to the different teams, the engineers, the different stakeholders in the projects to get them the reports they need, the studies they need. And this constitutes this three layer thing where the agentic layer is like the innovation, that's like the new layer we didn't have before. And that actually makes it possible to break these data silos where previously this was just impossible.
- [38:54]
Peter Sieberg
Very good. And that's actually, we're probably going to move into that for the moment, but let me check one more time. I like this idea of a three layer cake. Just need to put it the right way. There was this. Everybody these days is writing about robots, which we're not doing specifically today. If we do so, that's okay, but that's not the topic we have. And there was this. And then everybody's talking about the components of the robots and, and everybody at the same time seems to find out that the actuators have about 60% of the bill of materials, I believe. Right. And then there's this person writing, also making this cake, cake with 10 layers. But then the robot is at the bottom. And I say, I understand everything you do, but why do you put, put the robot at the bottom and the ace. The, the. The silicon is at the top and it's the other way around. Right. So I follow what it is that you're saying and I Understand you say the agenda is the middle one and then the user interface is the top. And I'm with you exactly in that thinking. And then in a moment we're going to be talking about the example that you have this onion for engineers. But let's make sure that somewhere today we're going to be talking mcp. I don't think it's the right point yet, but let's not forget that. Let's say a little bit with the data. You already said before, you mentioned a couple of times, data silos. Yeah, we've always heard about them and all the time. So how then in this time of you know, agentic AI, in this case for engineering how to not manage data and how to best manage the data.
- [40:35]
Stefan Zuvalak
Yeah, I think I'm super happy to include MCP here as an example where agents can, can, can make, can make this a lot easier because with this classical APIs you had like have to, if you change something, you have to rewrite a lot of integrations with mcp if something's changed, the agent or the LLM will pick it up and so the integration effort is much less. So in this case agents can really, really help with the integration. But there are some things agents can't help and this is mainly related to like how fast is the integration especially obviously, I mean we have customers, they have like hundreds of terabytes of fleet data. So, so it's really important to access this data fast in order for it to be useful. And also I think security is a super important point. We all know with agents security has become as important as never before. So to sandbox it correctly. And in this case I think it's super important that your data is in a place where it can be queried easily fast and it's governed correctly. And there are probably several ways like to achieve that. But we really see a big shift in the market where people gravitating towards the data lakehouse because it's like a very open layer and also to be very honest because databricks and Microsoft, they have good sales teams for the fabric and databricks products. But I think it makes sense in a way as a kind of underlying data infrastructure for a lot of data. And I think most engineering data like especially sensor data is there, it's the right choice. So we see with a lot of customers they're migrating away both from file based systems but also from like isolated time series databases.
- [42:30]
Peter Sieberg
If you say, if you use the word lakehouse, you know I've, I've been Used to the general term of data lake. But lake house for me is like in you. Is that like a typical way of organizing data in a lake or in a house of lakes or what? What is the understanding? Can you share that with us? Lake house?
- [42:53]
Stefan Zuvalak
Yeah, absolutely. Would really love to. But I think because I think it's a super important term, I think the origin it originated, I think it's probably a marketing term, maybe even by databricks, I'm not sure but what is meant by that. So in a classical time series or a classical database, what you get is you get a good access, fast access and you have with SQL for example, you can run complex queries. But the problem is in a typical database storage is very, very expensive and you have to have good data schemas. You have to clean your data to fit into the data schemas. Right? That's, that's one story. The other story is the data lake, you basically just have a blob storage which is very cheap. It's about like 10 times cheaper or only like 90% cheaper, whatever as database storage. But if you just put your files there, you can't use becomes a swamp. So that was not a successful concept. And the data lakehouse is you store your data on cheap blob storage but you put a SQL based typically can be other but typically SQL based engine on top that can query the data. And the engines are built in a way that they have intelligent caching and stuff. For example like Clickhouse is like super fast on that. And the last point I would like to mention is that the data lakehouse is like really dominated by open source architectures. It's Delta Lake which databricks uses and Microsoft fabric use users. But then there's also Apache Iceberg. So I think it's a good choice that at least like gives you a little bit of vendor independence and it's like supported by so many industries that it's like really mature and fast.
- [44:35]
Peter Sieberg
Very good. What about vector databases? There was a time that we said king minus man plus woman equals queen and we would find words in a multidimensional space that you know, I try to realize in my brain how to get there. Have they been forgotten? Did they not come through at all? Was that part of the hype cycle that did not pass the trough of disillusionment?
- [45:07]
Stefan Zuvalak
Yeah, I think so too. I think so actually. I think there's several factors at play here. So first things first, we still use vector databases. It's super important component for any agentic AI system. To find data, for example, to find the correct channel name out of like 10,000 signals. But we learned that this, what we all did in 2024 maybe even is like all these retrieval augmented generation rack systems that rag alone and just vector search alone is typically not sufficient. So you have to combine it with more agentic architectures that like on demand you use vector search, but also other types of search to find information. I think that's one reason why it's not hypy anymore. And the other reason is that for most applications it's actually fine, just use a reasonably good database. For example, we use postgres, so we use the postgres extension for the vector stuff and that works fine. We don't need some specialized high performance thing. And I think that's true for most use cases.
- [46:09]
Peter Sieberg
Good. One question before we then move on. So you mentioned the MCP example. We're not going to go into mcp. For those of you listeners that do want to know more, there's a couple of places that we have been talking about MCP in the last half year. I guess now MCP for me was kind of going the other way around. So it came out, I don't know, a year or something. Everybody was completely impressed. And then very quickly people started talking. You already mentioned already security, but you use it. Is MCP receiving updates? How does that work? Is the community, are you happy with the capabilities and the capability of using it because of security, et cetera?
- [46:54]
Stefan Zuvalak
Yeah, I think it's super interesting to see how this develops and I think there are a lot of lessons to be learned and I try to be as honest and candid as possible. So right now we actually don't have use cases that really use it in production. And the reason is, and we can talk about this later, that for most use cases, generating code is a lot more efficient than tool calling. So for example, we use SQL and we generate SQL code and then I can attach my database very easily through this SQL interface. So on this side I would say like it's still. There are many applications where MCP still has tools to prove its worth. But on the other hand, I'm absolutely 100% sure at the moment that MCP will be a key factor. And I will give you an example. Just the other day I was on a trade fair and I saw super cool examples by Vector. Vector is one of the market leaders in can bus technology and automotive. So for everyone, everybody in automotive knows
- [47:56]
Peter Sieberg
Vector and don't sell vector databases.
- [47:59]
Stefan Zuvalak
No. Yeah, no, just a very. I Think it was founded like, like in the end of the 80s or so. So very forward looking name and a very cool company. And, and, and what they, what they are doing is they like building MCP interfaces for all their products and it's just such a big enabler for example for our workflows when, when we have some data analysis and you want to jump deeper and then you want to have your vector tool that has all the bells and whistles that they've been using for like the last 20 years. Now with MCP you can easily, easily do it. And just seeing how like I would say it should be that these legacy engineering companies, so to speak, like the software companies, they are already introducing it and I think that will give such a big push that MCP I think will indeed be a very important integration layer that makes sense not for everything but for a lot of use cases. And you will see it in a little bit of a delay because obviously in order to be useful, for example for our system, we need the integrations by vector. Right. So it's an ecosystem play. And this has, this will develop slowly, but it will develop very good.
- [49:08]
Peter Sieberg
Let's move to Onion Renumix Onion Agentic AI Assistant for Engineering data analysis. Is that correct? Please give us a first introduction.
- [49:19]
Stefan Zuvalak
Yeah, if I only have a few sentences, I'd say like Onion is the best AI harness for engineering data, specifically for analyzing sensor data like test data, fleet data, simulation data production data. And what we really strive for is Onion to be very intuitive to use. So it's not like a specialized tool with a lot of bells and whistles. It's a very general purpose tool that you don't need like a handbook to learn. And that's very fast. People often say how is that so fast that you can, can pull data very quickly. See it like even very large amounts of data where like I would say like very general purpose systems like for example Power BI or tableau or whatever just don't have the performance. Right. And yeah, that's like the Onion in a nutshell.
- [50:16]
Peter Sieberg
Very good. Yeah, we're going to get into the details anyway. Now harness is a relatively new term. There's yet another term. I mean I just talked about it. Robert and I talked about it just two weeks ago because I read about somewhere there's this new job, I think it's the artificial AI harness engineer, whatever. And I've got a little story to say here as well, but I'm going to do that just in a moment but and ask you to share with Us what, what is the harness element of onion? What does that do? What does that not do?
- [50:50]
Stefan Zuvalak
Yeah, so yeah, that's very interesting. It's a new word. And it's also hopefully now we can see sometimes define a little bit better what we actually do or what onion is good at. So maybe going back to the three layer cake right now. Onion is a full stack system. It has connections to data sources. In particular, you can also like upload files and have a good integrated data management. You have this agentic layer which is the harness. I'll talk a little bit more in detail about that. Where we have agents that are very good at dealing with time series data problems. And then on top of you have the UI layer, because there was no UI layer before that handled that with the sufficient speed and usability that engineers need. So as a system, it has all these three components, but they are modular. And I expect in the future that if we think of how the agentic ecosystems will play out, that every company will have modular agentic ecosystems. And one, the middle layer, the agentic layer in this you will probably have different agents that are specialized for certain tasks. And why do you need to specialize agents? It's because the LLM is only one factor. It's, I think the biggest factor. But you need to control the context obviously around the LLM. Right? For example, just take an example. You have some, some test data, you want to investigate some anomalies in the charging cycle of the car. Then you need to give the LLM information about what signals are in the data, what are maybe even typical anomalies and maybe you can also introduce some information from previous analysis. So there's a lot of descriptions you have to put in and you have to give. And that's typically the other big ingredient, specialized tools to the, the system, for example, search for channels very quickly. Right. And providing this context also with input from user, but sometimes automatically providing tools and increasingly providing sandboxes and governance. This is the task of the harness.
- [53:05]
Peter Sieberg
I do recall there was this piece of research a couple of months back. They looked at whatever code I think it was code by cursor or the other one, whatever.
- [53:14]
Stefan Zuvalak
I think it was cloud code. I'm not completely sure you know what I'm referring to.
- [53:19]
Peter Sieberg
Yeah, they said 90, 98. Well, 2% only if you look at, I guess the, the number of code lines or words. I have no idea. Only 2% was like related to AI, maybe large language models. And 98%, they said is like the rest. And maybe the rest is like is for the control, right? Is is the rest maybe what typically what you could then call the harness or.
- [53:43]
Stefan Zuvalak
No, no, I think the cloud code is the harness and I'm not sure like what exactly these numbers are referring to. So obviously in the harness you don't have code for the LLM, you have just code around, it prompts how to call tools if you have sub agents, stuff like this. And obviously there is a lot of work currently you have to do in the harness to make the model behave good and to optimize it for your things. So obviously if you look at the model, the code for cloud code, it's like basically the harness. But I think there will also be like code and Tropic has around the LLMs. I think that's one point actually that's also where I honestly have to say might be dangerous for a company. For example like renewmix where we say hey cool, we are building the best harness for engineers. But what can happen and what we see happen is that Entropic, for example as an LLM provider and cursors also with their own models, they train the models specifically for their harness and that's actually a very good strategy. So the model is already pre trained to fit exactly in this context with exactly these tools and things. And I think it's going to be very interesting to see how this play out. How I would say like will we all still have more specialized harness like stuff we built in the next 10 years or will everything be just the entropic layer, the predefined harness that works well because if you look at how good for example cloud cowork works or cloud code works for many applications, it's always amazing.
- [55:27]
Peter Sieberg
But then the joke I would make is as long as you are still allowed to use it, right Claude?
- [55:33]
Stefan Zuvalak
Yeah, yeah, that's right. No, no, absolutely. I think sovereignty is an important point but also like governance is an important point. I think these are things and I also believe like especially like in our case right now, usability is also a super, super important point, right? As an engineer, for example, cloud code or a cloud cowork might be if you would like to use it, it can answer a lot of stuff. It just goes on a loop and comes back with a reasonable answer in 10 minutes or so. But as an engineer you first of all care about that. It's like very robust and repeatable and I think their specialization can help tremendously. And another thing and I think that's really important is also efficiency. Actually just yesterday I read that the Average token spend in the US is I think, I hope I'm not misquoting this because it's like $2,000 per developer, which is more than the median salary in India or something like that. So we are close to the point where you spend more on tokens in America than developers earn in India. And I think the customers see this in software development already that like, efficiency is key. So we believe that as long as we have like the most efficient harness and typically we need like at least 50% less, sometimes far, far more or less token than general purpose harnesses, even if the general purpose harness can get those result, that this is like an important factor in deciding for like more specialized harness.
- [57:05]
Peter Sieberg
Yeah, I think my answer to, you know, are we going to depend on one? I think the answer is always. We never want to depend on a single solution, Right. So we always should be supporting, you know, several solutions. You know, the majority can go for the best one, but we always need a second option. I believe I want to, I want to share with you a little story. It's from autonomous driving, so a little bit related, but not directly though, that you mentioned. Can bus, for example. I already told the first piece when I talked to Robert. It's because it's about the Munich Center Brewer Day just a couple of days back. And that's where I saw these horses, you know, these big horses who carry these carts and the beer and stuff like that. I'm not going to redo that. But when I was preparing with you, and then it was about the harness and I came to the horse again and. Because that's what typically. And then I learned what the horse typically have. Even if, if consumers ride a horse, that's. I believe that's a harness as well. So the, so for humans, I believe, to steer the horse. Right. And then I came to the story of the origin of the autonomous driving with my. I told my fellow marching band colleagues who normally, you know, they are not really interested in engineering and autonomous driving. But I told him I have to tell you. So I share it here. I have shared that in the past because it's a story that Robert and I learned about seven years back where in the meantime, Ernsteta Dickmans, who meanwhile has been recognized and reported by the German government for inventing autonomous driving. 1980, 1990s, right? And as a young man he wanted to build a machine because that's when the tractor came on site and he had been used to sitting, you know, on the bench with his dad behind the horse. And when his dad would fall asleep or he. The horse would drive back home itself. The horse would know the way, right? And then he decided, now we've got this strong mechanical horse, a tractor, but it cannot find its way home. And that was the reason he said, that's what I'm going to do as an engineer. Isn't that a wonderful, wonderful story? And then he did, he did actually really come up with the basics of autonomous driving. Now, I want to play this a little one step further then, because if you're going to have a system running autonomously, is that ever going to happen? I mean, is the question is more like how, how loosely can you deal with the hardness? Maybe that relates to the things you were just saying, if you have one or two, or are we going to be dependent on one? And I'm not sure I have. I have that clear in my brain where that's moving.
- [59:58]
Stefan Zuvalak
Yeah, I think that's super. I mean, in the beginning we established, like, how wrong I was in the past. So difficult to make predictions about the future. You mentioned it. It's actually quite interesting. Interesting that when we started off with this topic, genetic analysis for engineers, we had this big use case for fleet data where we build a system for an OEM that can answer questions like, show me the typical charging cycles, after which how many kilometers are our cars typically charged? And previously they had to do it with power bi and was like a long process. Data analyst had to do a dashboard to answer the question to took days. And now we had the system where you could just ask. But the system was constrained in a way that it's more like, hey, show me this chart, show me this histogram. Right? It's not really free and agentic, but we did it because the humans have a lot of control over it. Because in this case, the querying the data could take like minutes or even longer. So if the agentic AI makes a mistake, it's very expensive. So we guardrailed it super, super hard. And then about half a year ago, I think that was like very, very. I mean, where we had this open claw hype, right? And so there were like many things where people. I think for me it was really with opus 4.6, I normally don't call out specific models, but this was like really the point for me where stuff suddenly got good and it got like so good that it could like write SQL queries very easily, could. Could go on its own, could do longer workflows, and I was completely blindsided. What is going on? I remember this because Just six weeks earlier, I was with a customer and they were asking me how about text to SQL? And I was telling them, I know it doesn't work. And then just six weeks later was like, wow, this stuff works so well now. Unbelievable. And so, yeah, that's just to say I think it's difficult to make predictions. I think there are some studies where they say, hey, the tasks agents can do, like how long they can work uninterrupted on a difficult task, it doubles. I'm not sure if you know the figures. It doubles every six months or so or every year. So it's a kind of Moore's law for agents. Yes, and I think there's some truth in that. So I think right now the story is that at least in Onion you can do a lot of stuff interactively and, and you can what typically do. It's also maybe an interesting technical aspect. So we actually have two phases. So typically the agent goes on its own, writes code, does some analysis, anomaly detection for example. And then once the agent has the result, it generates charts to show the results to the human. But the charts are not code gen, they're tool calls. So they are just like some filters it configures. So we are always sure that the charts are correct. So you have this phase where the agent does a lot of stuff on its own. And then you have this phase where we consolidate the findings in a way that's like super transparent to the. That super. That humans can actually like understand it, have a good, good look at it. And I think that's the best strategy right now. Like where will it go in the future? Will we still have human interaction? Where will we have it? Probably very difficult to say. Very difficult to say.
- [63:19]
Peter Sieberg
Yeah, it's always difficult to look into the future, but Peter does it. So yeah, you already kind of moved in and I was going to ask you there. So who is the typical Onion user? And maybe you want to combine it with a use case and is that, is that person the same engineer that he or she was, you know, a year ago or then Thinking about new young engineers coming to the market, are they doing things at different way? And maybe if you take the use case, maybe you want to look at that. How would you have done it? The engineer have done it a year ago and how will he or she be doing it? By using Onion?
- [64:02]
Stefan Zuvalak
Yeah, I think maybe I'll start with like some example use cases that hopefully a lot of people can relate to. And the interesting thing is that they come like from all fields of Engineering we already had the fleet data analysis use case where I want to have some information about how do users use my product, like when do they charge or do I have more errors within your software version? Stuff like that. Then we have classical testing for example. You have endurance testing. Stuff is on a test bench. It's like tortures for months. And I want to maybe investigate if there are some error locks at some points and go through some event where stuff happens to understand what I want to improve. And maybe I want to correlate that with simulation data or tune my simulation so it better fits the test. And the last big big area is production. For example, I want to do an energy report. Maybe I have like a bill and I say oh why is the bill so much higher in March than in February? And you have to go through all the data and it takes days maybe. And with Onion or with agentic AI you can do all these tasks actually in a few few minutes. So these are the typical use cases for Onion. And in the past it's interesting that we had like these two worlds in engineering. One was the big data world. So we had a data analysis teams and they were working with databases like an InfluxDB or increasingly with a lakehouse there databricks. They had like Power Bi or Tableau and would make or Kafana is used a lot in production environments for example. And they were like building these dashboards. And then you had like engineers who had like more engineering tools. For example there's Diadem is one like very specialized tool. I was mentioning the vector tooling, maybe some handmade tools or often in testing actually some very custom tools by the manufacturer of the measurement equipment. And it was really difficult to answer questions because engineers sometimes for the big data stuff they have to go to the data teams and if they really want to investigate deeper they actually have to pull raw file data loaded into their desktop based engineering applications. And this process typically like took days easily, often weeks. And this is like independent from my eye, the big thing we actually saw many times in our projects with testing data that we have to kind of bridge this world. We have to make sure engineers can super quickly access like, like the broad data comparing a lot of runs, a lot of testing runs, a lot of trips, which actually engineers in many companies don't really do they want but they in in their tooling they can maybe compare two runs, right, but not seeing like the patterns and hundreds of thousands of runs. So you want to do that, but you also want to give engineers an easy to use Tool where you could actually see signals, for example. It's super easy but customers always love it. We have a very good functionality to zoom into the line charts really fast and it's very accurate and they just love it because they can actually see the details on what's going on. And bridging this gap, I think was an important part of the Onion mission. And it's also I think where a lot of the value currently is from people that use it.
- [67:22]
Peter Sieberg
Let's stay with the use cases and maybe one that I've been been actively involved in. And now we're talking 10 years ago when everybody and since then in a production world has been doing anomaly detection. Now how is maybe ONION based but in general agentic anomaly detection different from, you know, what we were doing then? Let's say classical machine learning anomaly detection.
- [67:51]
Stefan Zuvalak
Yeah, that's interesting. I'm not sure if you know the story, it's because it's in two years back or maybe not one and a half years back. Also when I had like my, my skeptics face, so to speak, I actually saw that anomaly detection was really hard and was like really had the impression, I don't know, a single like productive system that does anomaly detection with machine learning. And then I went to AI in the forest with you and it was really. And I talked to people and, and, and they were saying yeah, there are on things, yeah, typically it doesn't work
- [68:21]
Peter Sieberg
or really what are you saying that.
- [68:23]
Stefan Zuvalak
But yeah, yeah, it was really an eye opener for me. They said, I thought most be more specific. Right. They said okay, yeah, it works, but it's not worth it. It's just too costly to implement, to maintain and the use cases are not there. And, and from that point I actually had to say okay, I have to go to this group again because finally there's some real talk going on and not everybody is just doing marketing slides on oh yeah, we now do anomaly detection, blah blah, blah. Right. And, and, and why, why didn't it work? It, I mean obviously how you would frame it. I would frame it like this. Typically people were like drowning in the false positives. Why? Because it was just. And we can go back to what we discussed at the beginning. It was just narrow AI. So if you had an anomaly detection, for example, on acoustics, right. On a machine and somebody just dropped something near the machine, like you immediately get an alert because it's just different. Right. And the system doesn't know why it's different. It just sees out of distribution. And now with the context Knowledge with a strong AI, it's just so amazing. The way we do anomaly detection right now is typically just as a data scientist would do. It goes and looks at histograms. We can actually also incorporate like classical machine learning and now the new time series foundation models to do it even quicker. But the bottom line is once you have some candidates for outliers or candidates for anomalies, the system understands. It says, for example, hey, we have here a speed of 1,000 kilometers per hour. That's unrealistic for a car. It has to be a sensor error and stuff like that and what it says and how it relates it to the use case. It's mind blowing. I often sit in a session with a customer and they say, yeah, well, I mean the model cannot have the knowledge that I have about whatever like this specific test, this specific product. And when you see what the model actually knows about it, how it says, hey, maybe you should Fix magnet number three in your 10 state configuration or whatever, it's like what? So yeah, and I think that's, that's the main difference. You have the global context.
- [70:39]
Peter Sieberg
Yeah, very good, very good examples. And that is, I mean it's like the age of kind of really putting, putting a stage before the human looks at it. Right. And then even many times even knowing more than a human. I could come with examples, but I don't need to, where I can then very well imagine. But let me first understand, and that is typically foundational model knowledge that you have to enrich with the data of your environment, of whatever it is, you know, your cars or your production or your, your client or what is that? Where does that knowledge come from that the agent has access to?
- [71:20]
Stefan Zuvalak
Yeah, interestingly, I think that's a very interesting point that the knowledge about the specific domain in the sense how is a certain test structured, what is tested, and so on and so forth. The model knows basically everything. And actually there's a big difference between like the small models and the large frontier models. I think it's exactly in this area where the model always knows a lot. But what we have to give is like how is the data structured and where to find stuff, how to call tools very efficiently. And I think again it's a very interesting note that efficiency is super, super important. So if you use a cloud frontier model and you let it run for as long as you want, like you can actually solve most problems in a simple loop. It's amazingly good. But you just don't want or cannot afford to run it for like 30 minutes. When the task is so simple, yet if you know the right stuff, you can do it in 10 seconds. And then that's again part of the harness to provide this information that you can run it in 10 seconds. But we don't really have to give it information or a lot of information about the specific domain.
- [72:33]
Peter Sieberg
Yeah, or your finance person comes in and believes that the bill for Entropic or whoever is an anomaly, but it's not really.
- [72:42]
Stefan Zuvalak
Yes, absolutely.
- [72:46]
Peter Sieberg
It's suddenly true.
- [72:48]
Stefan Zuvalak
Yeah, absolutely, absolutely. I think it's a, it's a super big point. We are at or close to this point where if you had unlimited tokens, you could basically build whatever you want, but you don't have unlimited tokens. So you have to think about how to spend it wisely. And using a specialized harness like Onion I think is a good step to spend it wisely.
- [73:11]
Peter Sieberg
Okay, back to the harness. General. How does Onion deal with deployment security?
- [73:17]
Stefan Zuvalak
Yeah, that's a super big problem. It's probably also right now, I mean, we see like a lot of stuff going on in this realm. Maybe you also followed these different supply chain attacks. Obviously, because everybody uses agents now. It's much harder at some point to make sure you have rigorous processes in place that obviously people are doing more and more. We obviously are too. There's a lot of certifications that also mandate to do this. I think it's super, super reasonable. And then this is like where we, I think, have to make sure that in the development process already we are taking care of all these things. And then in the architecture, obviously the agents can do a lot of stuff. And I think the analogy is quite fitting that some people say, hey, an agent is a super, super intelligent but slightly drunk intern. And it would go on to delete your entire production database. Maybe it can make your code like 30% faster, faster. But just to test it, it would just delete everything. It has no problem with that. And I think it's true. So the only after work intern.
- [74:28]
Peter Sieberg
Okay. Yeah.
- [74:29]
Stefan Zuvalak
And I think everybody who codes with agent has these stories for themselves. For example, you have a secret, right? And you hide the secret from the agent with some mechanisms where it can see it in the ide, but then the agent can come up with other tools and the ways how it tries to access it anyway, it's like, like it's really strange or really, really alarming. So the only way you can really make this secure is to sandbox it really heavily. And that's what we're doing also in Onion. So you have to have Good sandboxes around everything to make sure even if for example, the agent builds a really bad query, the system cannot break or it cannot break out of the container, stuff like that.
- [75:14]
Peter Sieberg
So how does SAM box work while still you need to have access to all the interesting information because without having the access, the agent cannot be smart. So isn't that a contradict?
- [75:32]
Stefan Zuvalak
Yeah, absolutely, absolutely. There's definitely a trade off. And I also found it. I actually think about this a lot and I found it very interesting that to me currently the best abstraction is just you have a folder, that's how typically IDEs work or also how Claude cowork it has a project, but it's also just a folder. And in the folder you can basically do whatever you want like that, but you cannot access for example the Internet to prevent leakage. And in Onion we have a slightly more abstract and powerful concept. We call it a data container. And what you can do is with these data containers you can control what the agent first of all sees how the agent can understand the data. But also the security aspect I think is super important. For example on a database, super simple, but you make it read only for a certain data container. So I can make sure it can't mess with that. Whereas maybe you have another data container where you can write stuff to. And having sent boxing on this level, I think is the only way to make sure that the system is secure.
- [76:41]
Peter Sieberg
Very good. I read about this option of bringing your own LLM. I think I can imagine what that maybe is. But maybe you can answer us and combine it with talking about your license model just to get an idea of maybe interested parties. How does that work?
- [77:00]
Stefan Zuvalak
Yeah, absolutely. First of all, we really strive for Onion to be I would say like even I would say cheap. So it's something like every engineer can use. So it's very general purpose and a buy seat classical buy seat license where we also feel like hey, you pay by seat, but then you save so much on query time, token usage and so on so forth that it's like not a big investment. And the actual like LLMs in our case typically from our customers. And there's several reasons. So if we typically deploy to the cloud, for example to Azure and there our customers have their cloud environment and they use their endpoints like GPT, 5.4, whatever, it's available there. And that's for I think engineering software, the best deployment path for several reasons. I think there's a lot of sensitive IPs in the system, so it's important for the customer to not have a cloud based system that's controlled by the vendor, but that's controlled by them. I think it's also important for them to have it like from a German company. And that's by this bringing your own LLM, like hosting it in your own cloud tenant is so important. It's also important because when we do data analysis on like hundreds of terabytes of data, data locality is everything. So you definitely have to have your data in the very same a compute center where for example Onion or the harness runs. Otherwise you had to transfer data which would be like super slow and super expensive. So that's why bring your own LLM is very important.
- [78:40]
Peter Sieberg
Very good, thanks for that. Now one thing I was thinking of like what is your Renumix business model? So you said this is something that you will provide to engineers. They would be typically customer engineers. And when I say customer, so are you providing services to your customers? Like so you're being paid by the hour, by the week, by whatever you do. Is that a combination of two things that you provide to the market? And where do you see that moving? Do you see that changing over the next year and then we're going to be coming back to that topic at the end of our talk.
- [79:22]
Stefan Zuvalak
Yeah, yeah, absolutely. So I think it's also very interesting question because traditionally we've done like a lot of bespoke custom systems for our customers and we increasingly see that there is some standardization happening. We also see that actually, which is super cool, that more and more engineers or domain experts start to wipe code, something. Right? So I think there will always be the need for something, some kind of customizations that we gladly do, we love to do. I think that's also where a lot of innovation happens, right? All these standardized software, it's only developed at Renumix or in other places because you first have projects where you do the real innovative stuff and see the new path forward. And that's what we are happily doing with our customers. And then there's actually one area where currently there's a lot of demand. And, and we also feel that it's important to help people, which is like exactly to prepare this engineering data infrastructure for agentic AI. So currently a lot of people are moving from old on premise systems, Firebase systems to the lakehouse. And then it's always the same management says hey, move everything to fabric or whatever. And engineers have to figure out how do we do it? Like how are the data models? And there is a lot of stuff, stuff that we did wrong in the past, quite frankly. So you can do wrong and if you do a couple of things wrong then it's really difficult. For example, if you have the wrong partitioning and it gives you three times or you slow down three times and you make two other mistakes, then it gets multiplied and at the end you have a system that's 10 times slower than normal. So I think that's a big area. And then I think, think the overall market in custom software is like, yeah, let's see where it goes. Because like you have the wipe coding stuff, maybe that will fade. But then you have experts like in our team who are much quicker in developing solutions. Yeah, let's see where this goes.
- [81:26]
Peter Sieberg
Yeah, we're going to come back to this anyway. Now we talked about bring your own LLM. Share with us your experience with small language models in your engineering environment in Germany.
- [81:38]
Stefan Zuvalak
Yeah, I think it's also something where I think a lot about and maybe where I can also be like very wrong. But right now we can actually see that the frontier models, large language models are much, much better than the smaller ones. There are some tasks where the small ones are good enough and they are a lot cheaper and a lot faster and that's really important. So maybe the idea of having like these sub agents with smaller models is good but at the same time we actually see currently that you just have one flat loop with a big model you use a lot of token caching to make it efficient. The model sees the whole context. You don't have to put like a smaller context for the smaller agent, smaller model. It works just so well that yeah, I'm not sure we will see if we actually will really incorporate these smaller models a lot or if it will center on the, on the larger ones. What I especially maybe what, what I'm really, sometimes I have to admit I really have to get angry. I think there's this story in Germany that yeah the large models, they are so expensive and so compute intensive, energy intensive to train and obviously we cannot catch anymore but we have this highly specialized knowledge and we can train it into small language models and then we win. And I think at least right now, at least right now that's sorry to say it's just a bullshit thing. Especially in the domain knowledge stuff. The front large models are just the best because there's the strongest, they have the strongest AI. Everything else will is a to come back to that more narrower AI and it's just definitely right now the lesson to be learned is like the larger you, you know, more you you have the better context. This is winning. And, and we should not, I think just feel, hey, let's build a lot of small language models, then we have a chance. I think it's a little bit, I think that plays a lot of desperation into that.
- [83:40]
Peter Sieberg
Now Robert and I, we've been, we've been talking about industrial foundation models about five years I guess. Right. And, and mean, meanwhile we have seen and you know, I'm not involved in, in, in, in organizations that would. But I've heard that there's many organizations now doing exactly that. So that is one input and the second input is that's the other story about. Yeah, and that's, that's then typically about Europe, I mean Germany, but also Italy, France, all the other more, you know, industrial oriented countries also saying yeah, we don't need to have the, the models, you know, that train has, has left the station, but we sit on all this industrial data. It's, which is, it's the combination of both. Right. And maybe if I combine those two in my brain, I'm interested in your view on that. It's like, you know, three years ago when Sam came with his large language model, he had been scraping the Internet and nobody knew for a year or two I guess. Now how has that, that worked in the meantime? That, and that's also the robotics story. You can't scrape industrial or can you? Industrial data, maybe you can scrape, but the area of physical AI for engineers, you can't really scrape that kind of thing. So it's, it's a little bit of a bigger thing. But do we need this industrial data or do you say. Well, your experiences, the, the huge, the big foundation models, they seem to have this kind of knowledge somehow in them anyway.
- [85:14]
Stefan Zuvalak
Yeah, I think it's a good question and I definitely have to say like I don't feel like I'm a big expert on this. So take it, please, please take it as an opinion. But I, I really think that, I mean what do the frontier models or frontier things have? They, they have like all the books, so they have all the book knowledge that's available. And then they have like I think very clever reinforcement learning where they have like a coding puzzles and all these things where they have a very good general purpose problem solving capability. Right. And like to replicate that is like super, super difficult. And then obviously if you have some very company specific knowledge how exactly this test is run and all these things, you can just put it in the context. You don't have to put it in the model. And I Think in these combinations. I just don't, I just don't see how this can help. And, but, but it's, I mean, I mean, what should we do? Like, should we say, hey, there should be no initiatives anymore? I think that's also wrong. So I don't want to like downplay it. I think it's important to try it to also maybe be on the topic, have people here who do some kind of meaningful research. But I think it's super difficult and it's, I think the narrative should be, should we try it? Yes. But is this the future? Don't know.
- [86:39]
Peter Sieberg
That's very important. Yeah. So at least you know, it's not like a given that you know, doing this whatever industrial foundation model with a handful or 20, I mean similar like the OPC foundation has been doing in the past, you know, 20 robotics companies and do an OPC information model. The same way we could sit together with 100 or 1000 automation companies and make this industrial foundation model. Maybe something like forgot the name now. The daughter company of BMW, I saw them again yesterday, they're doing. But then you say, yeah, maybe it works. But that's not a given. That's what I hear you say.
- [87:24]
Stefan Zuvalak
Yeah, I think actually if you look at how these frontier labs are structured, I think it's what you don't need. 10 different stake 100 different stakeholders with, I don't know, hundreds of people. It's more like you need 50 people who know exactly what they are doing and billions of compute. So yeah, I think it's, it's, it's especially in this consortium and so on and so forth. I think it's, it's difficult and maybe, I mean obviously don't, don't do nothing as also no option. But I really feel like making sure, for example, what I'm a big fan is really making sure that the data infrastructure layer and all these things, we have sovereign stacks, we have good best practices, we become data driven fast. I think these are the types of things that especially the German industry should gather around and should build stacks that are independent from the big American companies. And I think there, there is a lot of open source, there's a lot of cool new stuff that we are also using internally and where we can actually build on the frontier. So build stuff that's better than the other stuff and at the same time open source and sovereign. So I would say let's put this more in front and center than the LLM things.
- [88:53]
Peter Sieberg
Right. You started giving part of the answer to my final question. I always want to know from people I talk to like you. In this case, I was going to ask Fable and kind of you already given maybe the answer to that. You know, do the open source and don't be dependent on. So let's talk about this future of engineering around, you know, around your offer, which is, I mean you, you, you see or you believe to know how as good as anybody else, where the market is moving and, and one, one way of reacting or acting is that you put a certain solution into the market. Now let's start very specifically with. Do you see Jeff Bezos going to be a competitor with his Prometheus. Prometheus. Prometheus.
- [89:45]
Stefan Zuvalak
Prometheus. I'm not sure. Yeah.
- [89:46]
Peter Sieberg
Do you understand what it is that he is, he is trying to do? And do you, Are you. I'm not going to ask if you're scared, but do you, do you see him or maybe you say, oh no, I see him as a potential, you know, partner or I don't know, do you have a feeling for what he is doing in our market? He's going to be doing.
- [90:06]
Stefan Zuvalak
Yeah. So I don't, I know nearly nothing about it. So. But there is actually this general trend in Silicon Valley to not just build AI software, but just provide the service. For example, if I provide like a tax consulting service, for example, I don't build software for tax consultants, I run a tax consultancy. But internally I build all my processes with AI. So become like the service and not build the software. I think is something that, it's like a big, big mantra currently in Silicon Valley because then you capture more of the value and you have everything internally. You can iterate faster. Right. All these things. And I think to me it seems like a similar play in the sense that it's not a software company, but it's maybe like a company that actually I think they're planning to buy older or like established industry players, but then revamp them with AI. And I think on paper it's, it's probably a very interesting idea. Like, I don't want to discount it, but obviously it's also from players who have like little experience in the field. And I mean, just looking back at the history, for example, I mean maybe you remember Elon Musk's the Boring company where he thought he could just replicate very quickly or, or maybe only so what Google has been doing in the robotics space. I think the ideas are probably very good, but it also turns out like execution is really hard and there's a lot where the domain knowledge Experience the value networks are coming in. So for me personally, I'd say we try to deliver one puzzle piece. Super, super. Well, that's this data analysis, especially time series data analysis for engineering and really try to focus on the puzzle piece. And there are a lot more puzzle pieces that are solved by other people. And if Jeff Bezos wants to do it with a lot of billions like in a very end to end fashion. Yeah, let's see where it goes.
- [92:16]
Peter Sieberg
Okay, thanks. Yeah, yeah, thanks for that. It's always very difficult to be looking into and I actually maybe I thought something similar that he was going to be buying by as I just said, you know, as you could scrape text from the Internet, you can't scrape physically, maybe more like that engineering a little bit, but not really. And I was thinking as well that maybe he was going to be buying but then coming back to the discussion we just had before that he was also going to be buying industrial data. Now I believe I hear you say that maybe he's buying more the knowledge, the capabilities of the people that are inside of such potential companies. And I hear you say that maybe he's going to then reinvent him becoming AI first then.
- [93:05]
Stefan Zuvalak
Yes, yes, I think that's the plan. Buying like real entities, everything, not just some data package but basically the whole thing with all the people, all the processes. At least that's, that's what I, how I understand it might be wrong like disclaimer.
- [93:17]
Peter Sieberg
But yeah, let's close. I think we passed the hour. Robert is going to be very happy with us with me again. Tell us quickly if you want to about your team, where you base and maybe if you're looking for talent just in case if everybody, anybody would be interested in joining you.
- [93:34]
Stefan Zuvalak
Yeah, thanks. Yeah. We are based in Karlsruhe. We are, we are 10 people so we have this good size of what in software engineering we call like the two Pizza team. And if people want to do like some internship for example, we are already is open for new colleagues. But it's also interesting would be another podcast that at the moment because AI coding and everything is like making us so so fast. It's interesting that we despite living having a very good growth with Onion, are not at the moment looking for new full time team members.
- [94:06]
Peter Sieberg
Oh, that's a very interesting finish here. And yes we have talked about that very topic before and we will continue to talk about it before. We wish you very much success in working that way. Stefan, thank you very much for having shared with us how you see the market, the solution that you're putting into the market for the engineers. Thanks. And see you in the Alps soon.
- [94:32]
Stefan Zuvalak
Yeah, very much. Looking forward to that. Thanks a lot for having me, Peter. And very sorry for talking so much.
- [94:40]
Peter Sieberg
That's me. I talk much. There you go. Thanks. Bye. Bye.
- [94:43]
Stefan Zuvalak
Bye, bye.