Agents, Molecules & the SaaS-mageddon

This podcast is supported by Siemens, your partner for industrial grade AI. Hello, everyone and welcome to a new episode of our industrial AI podcast. My name is Woodbeaver and it's a pleasure to talk to Peter Seiber. Good morning, Robert. Good morning, afternoon, evening to all of you, dear listeners, wherever you are. Good morning, Peter. Robert, greetings from Istanbul. I was going to say you're traveling. Have you been looking at what the blue mask or what I, as I recall? Yeah, the blue mask is very impressive. We went there yesterday and it's very sunny and 19 degrees. So it's very nice, a little bit of wind. So I'm sitting here outside of my hotel, looking from my balcony to the boss burrows and recording an episode with you. The boss burrows that's exactly what I thought of. So the last time I was there, I must have been whatever, 20 years ago, whatever I talk about is always 20 or 30. It's a good point. And I was working between, I was working for actually four Turkish airlines here. I do, I do recall that, Massibin was, I think it was in Istanbul and they had their head office there and we were recording. I was with a mainframe company actually at that time. They still existed at that time. I do recall, it was a very, very nice time. I used to say boss burrows. I was looking for the name of the river and the bridge over the river. And this bridge is actually the bridge between West and East, right? Europe and Asia. Right? Europe and Asia. And you can take a zip line and you can zip between Europe and Asia. We will do that today. All right. How do you do that with a, what kind of? With a zip line. A zip line. It's a zip line. You have the, yes, line and then you put something on the line and then you have a zip and you have to go through the, yeah, from the one side to the other. Like a cable connection or like a cable connection. Oh, it's, it's very funny. I think the, the kids will like it. Okay. So great. So let's talk a little bit about AI, not about tourism. Yeah, a little bit. Tourism podcast. Yeah, you can, you can do the touristic part. Let me start with two weird words. The first is sauce, spoke ellipse or the other term used here is the sauce, my garden. I'm not sure. You heard of them. You can, you can add your piece. So what they really stand for is S a S a S or software as a service, apocalipse or amigurant, right? So we all know we've all gotten used to sauce, cloud-based applications. So accessing applications through a browser, typically paid monthly yearly. We all do it, right? Sills for C a RAM, Google Docs, Microsoft, A, Do, Bs, whatever. Now what happened is that and tropic, they launched, that's already a month ago. Claude, co-work. So that's the workplace AI tool. It can read files, Jeff, documents, organized workflow, etc. But then two weeks ago, they followed up and they introduced plugins. Very specific plugins. Let's say tools that are capable of specifically doing things in these market, segments, finance, legal sales marketing, customer support. So on top of the typical sauce applications, now what happened is, and I was looking for the quote, but not sure who was saying, but there was like a 285 billion stock market sell off the data. That's what they called, sauce booker lips, sauce maga. Now why are we talking about it? At that time, hardest hit were the companies, of course. Today, the sauce companies in that area, Thomas Roychers, legal zoom sales force, etc. Why is that happening? Because the people who know about what's happening in the market, I mean, including us, people listening to us, and have people trading, they see what is happening that instead of us users managing, let's say, an example, managing our customer relationship, CRM, we use our higher agents. And today is again, of course, a lot about agents. And we want him to close tickets or transactions. And in the end, we want to pay them outcome based, rather than monthly. You know, I want to get something back. So that's kind of what's happening, I believe. Not only thing, and then why are we talking about it? It's like, you know, we need to consider, as in general, always, we always look at things happening within the eye space. And how can you, the listener, how can you use this information for what's happening? But there was a very specific one already. And maybe you know about a second one that you want to share. But there is one that is related actually to what was happening in the PLN market. So one company called DeSo, DeSo Systems, or DeSo C-STEM, a day dropped 21%. Now, I think one reason was that they had not been delivering in Q4. But on top of it came this, you know, AI narrative. Now, I think, and I don't want to go too far, I'm not saying that, you know, companies like Antropic Open AI, you know, sort of the general large language model providers that they will take over the industrial, in this case, PLN offerings from the SoCement SpeedyC and all the other ones. Because they have like, you know, multi-billion years of very deep knowledge. But it is important that we see that at least let's say the outer range. So where maybe the the Cloud Co-work, and that's a term I come to later as well, when it's about writing, reports, summarizing change requests, drafting compliance documentation, that's what they can do today already. So they start to move into the market in which we industrial AI are in today as well. But this is interesting because then the big players like the SoCement etc, they need to find another sweet spot, right? So what is there? All of them, yeah, exactly, exactly. All of them. So moving back one step, the Salesforce and all the other ones, they have to. I think there was this discussion about the vertical horizontal sauce that was exactly spot on in the direction where you were saying. So whatever we call that horizontal vertical. Yeah, finding the sweet spot or the term that's always being used is this moat. Moat is, I think, is that the water around a castle in the past in the middle, either space, maybe that's what you see today as a tourist. I don't know if we if you have that in the area of Istanbul. But what is the mode of these companies because the general large language providers are eating into are kind of breaking up the moat. You know, they come from outside, they let the water out and then you know, they're going to come close to the castle and break down the stone, so to say. So yeah, and I don't have them. And that's, as you say, you know, all of us need to be looking for what does it mean to what it is that we offer and how do we maybe need to change our offering. Yeah, that's interesting. So I want to come back because you mentioned agents and do you know Peter Steinberger? Exactly. That's my next topic. But you start. I'll just add, you will add almost four or five hours yesterday. You start. Yeah, okay, because Peter is not Peter Sebert, Peter Steinberger. He's from Austria. He's joining OpenAI to drive the next generation of personal agents. And Sam wrote, he's a genius. There's a lot of amazing ideas about the future of very smart agents interacting with each other to do very useful things for people. We expect this will become core to our product offerings. OpenClaw will live in a foundation as an open source project at OpenAI. We'll continue to support the future as going to be extremely multi-agent. And it's important to use support open source as part of that. So that's that's as it all. When did you when did you read that now? It's important that for you listener, whenever you're hearing this, you may have heard of it, of course, in the meantime. But we are now recording Monday a couple of days before we go live. So and that's why I say it because yesterday as by the couple of hours, and even this morning as well. And I'll share. So what is your opinion? Well, and the first question then is, when I didn't know it yet, what you tell me now. So did you already hear this yesterday on the Sunday? I woke up to the morning and I saw that on my in my stream. I thought I was going to look for later on a day to find a time. Yeah, so I looked at a couple of interviews with Peter yesterday. And as we yesterday he was saying, yes, he was still talking to Meta. To the Meta guy, what's his name? Zachary Brook. Yeah, exactly. Was his first name, Zachary Brook. Mark. I am Mark. We're talking to Mark. That was a very nice story. He said he was talking to Mark on what's, but then Mark said, yeah, we can do that. talk now, but Mark was he was coding. That was very interesting. And the other story he did was exactly with Sam because so what so what Peter has been doing, he has this thing called Open Claw. Today it's called Open Claw. It started a couple of months back. It was Claw Bought, Maldbought. It's not Maldbought. It's another thing. It's a little bit confusing. All the stuff and the Claw Bought was then too close to Clawed. So had to change it. It's so interesting. In the end, what it is in his words, it's an AI that does things. So you communicate it with through messaging platform, WhatsApp, telegram, Slack, whatever you want to use. And then you connect to an external LLM, Clawed to deep seek or any of the opening eye. Now, and then in the end, I was going to say and we put it in the front because you already know. We now know that he's going to join. Is that the way that was put by Sam so Open AI? And then he did a call to Sam maybe a couple of days back when he came up with the latest name which was the Open Claw, which a little bit sounds like Open AI. Yeah, exactly. And Sam said, "Oh, what? Spalla to me, please?" And Sam said, "Oh, yeah, that's okay." So it's interesting to hear. Now, why is this going to be big? Well, now it's probably even going to be bigger, I guess. What I just heard you say is that it's going to continue to be an open source project. That's what Peter was saying. That's very, very important to him. So you connect through your, let's say WhatsApp or any other, there is, it's local. That's very, very important. So you decide yourself how much information you're going to share. Now, let's see how then Open AI is going to be dealing with exactly this information. So today, I think it's important to say that you need some base coding capabilities. You start with some kind of pip install for those people coding. That's a standard thing. If you're not coding, that's still too early for you. You need to be careful. There is still some general security issues, which Peter will work on, which then I guess to get it with Open AI. And they are more general. They're not specific to him. But it's been amazing what's been happening around Peter. Exactly. It's a little bit of hype too, right? I've heard like, was it 24? Was the Open AI? The chat GPT was a 24. Then came 25. Was the deep seek? I heard it say and 26. It's a bit early. Maybe it was going to be even bigger things. But 26 is going to be the open claw moment. From your point of view, what does Open claw means for industrial applications? Yeah, right. And that's exactly the thought important for us. So if we say what it does do today, it does millions of things, whatever you want. It does your email. It creates your calendar. It checks your flights, whatever. It does do home automates. There's only like 10 ideas that Peter shares. Home automates. If you want to have your Tesla, you know, moved in front of your door, that's what all these things that we've always been talking about as a person assistant. Yeah. I think that's the way to look at it and his words, as I said, as concierge service in the hotel. Yeah. We all are going to have our personal assistant. Yes. Right. And the positive was the concierge for the people who could afford it. And the future we're all going to have our, it's a person assistant. That's my word. Peter's word is as the AI that does things. And because it does things, there is this potential. Now it's going to check you in for your flights, home automation. And that's where you go. It's doing home automation today. Now let's leave the home out. Will it be doing automation? Sure it will be. Now if, if and then of course, we already talked about the security in order in automation, as we talk about, industrial, there can be zero. You know, we talk about industrial great. And we always think of Boris Scharinger, who came up with that terminology, which is 99.99, you know, security. We can't have any security potential. So is it going to be used in our world? I expect so. Yes. I can't say more. So the concept, you mean? Yeah. Yeah. I mean, is it only going to be this one? The very fact, I mean, I have not been in all these discussions on. But why do not we see something like from the big automation players, some, some, some multi agent industrial framework? Well, let's, let's keep the industrial outside. Why don't, why did not we see exactly this level solution from open AI from meta from Google, from AWS, from the big players? Why must it be that now we just learned that Sam is going to put whatever I assume a couple of hundred million, but I've no idea. He's going to put big money into what this guy in this Peter almost by himself has been being together. And at the same time, he is being almost not accused, but people saying, you're doing nothing new. He says, yeah, it's all, it's all there. It's all there. All the messaging services have been there. Of those, you and I, we use Slack, for example, we use WhatsApp. But there's many, many more. All the external alarms are there. And he's been programming those interfaces, basically. And that's what he says. Yes. Yes. I didn't do anything more than that. He seems to be just a real smart cookie, like there's many smart cookies, including, you know, those of you listening. But that's my understanding what he's been doing. And he was, he was kind of, when he talks about what he's been doing, he was kind of impressed himself. And suddenly, it started answering. And I thought, how can that be? I think he said he was in Morocco and in Marrakech. And he was communicating through WhatsApp because he didn't have the broadband connection for other services. And then his solution, whatever it was called at that time, came back with something very smart. He was kind of perplexed. Whatever I've been doing kind of thing. Yeah, it's amazing. I guess important is that today, as I said, you need some base understanding, coding, installing capabilities. But if now we know that this approach is going to be part of the OpenAI, I assume that means that at some point in time, I guess within the next couple of months, whatever, they are going to make it at the level that you and I as consumers and that our colleagues listening, making the industry our solutions are going to say, or maybe you listening will already be doing that today. But but I'm sure you're going to be doing that on your yes, local pussy that is not connected, right? To start. Or I wasn't, I've not been using like a VPN as a VPS, like a virtual private server, which you will run it on until at some point in time, you're going to say, okay, now you're going to make a sandbox solution with an industrial environment with sensors and whatever. And is it going to be the only one who knows no clue, but at the moment, the OpenClore has the big drive. Exactly. So let's move to the main part, Peter. Quick mentioning only. You're going to be talking to ferry Abel Hassan, not today, but now that you should tell, I didn't want to just move on without this. We have this opening of an industrial AI cloud by German Telecom in Munich. That's just 10 days ago. I just wanted to mention it. 10,000. Yeah, exactly. And VDH GPU. So without the Nvidia, they talk about sovereign European AI or Germany stack with SAP with 10,000 is not so much. But it's all relative. It's still, I think they said they doubled the, it's the, yeah, what is it? The industrial AI public capability on the German market. So, and I think you're going to be talking to the CEO of T-Systems in March, sometime in details. Exactly. I think in March. Yeah, sorry. So let's move on to the main section, as you said. Yeah, the main section. So it's not about agent. It's a totally different perspective when it comes to AI. I talked to Stanislav from molecules one. It's a startup in Poland and Jakob Tromszak said, if you want to learn how the whole pharmaceutical industry, medical industry is working in the future, you should interview, do interview with Stanislav. It's a very interesting approach, what they are doing, how to find new molecules, how they produce new molecules, really produce. And that's very interesting. And our industrial AI sector also means chemical from a pharmaceutical. And it's quite interesting to learn something from this area. Right. I think one When you first mentioned it, I asked, is it something similar like the Alpha fold and use that maybe that direction with the main difference? As you just mentioned, it's like they produce, they do not just design but they produce actually new molecules. They produce and they, you will share the whole, let's call it supply chain from finding new molecules to produce new molecules and it's very interesting. And I'm very much looking forward to, I was very bad in, is that called, that's not physics, it's called chemistry I guess. It's chemistry, right? I was very, very bad. This is the worst I was in college I guess with chemistry. Nevertheless, I'm very much looking forward to listen and specifically with the idea of as we typically do and as you mentioned, you know, how can we maybe then use similar approaches or just use the approach and then turn it in such a way that we say, okay, if that's what we can use also in an industrial or even already we are, you are talking about the industrial setting. I'm looking forward to. Thanks a lot. Bye bye. Robert, have a great time in Istanbul. Take care. Bye bye. Thank you. Bye bye. My name is Woodweeba and it's a pleasure to talk to Stanislav from Vasha. Stanislav, welcome to the podcast. Hi, thank you for inviting me. Yo, more than welcome. Before we start talking about pharmaceuticals, chemical companies, process industry, please introduce yourself briefly to the listeners. Sure. Yeah, very happy to be here. Stan, co-founder, now a city of the Antiques, scientists of molecule one. So molecule one is a company at the intersection of deep learning chemistry and drag discovery. And what we do mostly is we help others make medicines faster. And what is your background? Sure. So foremost, I am a scientist. So yeah, my overall kind of passion has been since beginning kind of the scientific discovery process. So I am really very interested in trying to contribute to automating it, which is now a big, a big topic, especially with like from the perspective of deep learning and that's my background really. So I did PhD in the earlier days of deep learning, then postdoc at NYU, which is a lot of fun. Then something called more like a formal degree in Poland called habilitation. Yeah, in my scientific part of my journey, I was fortunate to write a few highly cited works on the foundations of training of deep networks. But then I was at this crossroad, right? So I was after a postdoc and I was like, should I go work at an overbicta, continuing this direction, but I decided to go to Poland to also to lead molecule one. So that's interesting. You have a AI background, not a background in chemistry, right? That's true. Yes. As you know, it was like very utterly accidental where I ended up. But yes, my background is not in chemistry. So folks at molecule one are very patient at explaining and entertaining my questions about chemistry. Yeah, we do a lot of work at, you know, like frontier of chemistry and automation of chemistry. You mentioned, is your field drug discovery? Is it the overall topic, right? Or M&M wrong? Right. So mostly what we do is we synthesize small molecules. So any drug discovery project really starts with the question, oh, I need to make physically the molecule to test it for biological activity. So yes, this is our primary field. But you know, molecule one starts from this perspective that we really need to discover novel ways to make medicine. So we not only automate how you make molecules, which I can go into more details, right? But also discover kind of novel chemistry. And that's why it's really interesting to me because there is this aspect of as automating scientific discovery, right? There's also the topic generative, right? To generate something new. Right. So, yeah. So discovery, I think the main thing to understand is that there is a very big gap between what you can make in the laboratory and what you want to make. Exactly. So your generative models can come up with all sorts of crazy designs and people, you know, do that. But all of them have to be made in the laboratory. So we have both generative models and, you know, discriminated models. Generally AI to tell what can be really made. And I think where the things are getting really exciting, we also kind of broaden what can be made. So if you make new discoveries in chemistry, then actually what is not what is a hallucination and what is not, you know, it starts to be something you can, you know, expand, right? Like because we expand the space of molecules you can make to close this gap between what you can make and what you want to be, right? Exactly. We're coming back later there, I think, because I have one more question when it comes to your company. Is your startup, so your idea, your approach in a way comparable to startup like Maxwellings Casp AI? So it's not Casp AI is working on new materials. You're working on drug discovery, but this is a new field in AI or M.M. Wrong. Right. So Casp AI is in inorganic chemistry. So less drug discovery more materials, right? We really focus on synthetic chemistry, organic chemistry. So things, you know, things of life. This made of carbon that, you know, can become medicines. And here in our case, which is not the case, I think at the moment that Casp it might become, we have a very big physical component in that we are, we have operating highly automated laboratory where we attempt these synthesis. It's actually very, I think, important strategically to have like this vertical integration if you will, right? So we have both AI, then AI is trained on data from your platform. And then you kind of use the platform to solve your client's problem. And in the process of so generate data that you can then feed to train better models. I think it's really a good kind of fit for this podcasting that sense, right? It's like industrial AI in the sense of really vertical integration of both digital and the physical. Is that your USB from your point of view? Right. So maybe, maybe taking a step back, right? Like what is what is molecule one and stuff like that. So as I said, really the problem is that there is this gap between what you can make and what you want to make. I think this is like crucial starting point to understand. And I think we see that also like tragedy between times of hallucinations, like it's kind of similar, you know, people like models hallucinate molecules that we cannot make. And this is important because there is, from my perspective, a bit of a crisis in track discovery in that the molecules that we design, they don't pass clinical trials. And one of the reasons often is attributed to the fact that those molecules are, well, there's this gap, right? So for example, Mother Nature, let's say when she is discovering medicines, right? So natural medicines, things you can find in plants and stuff like that. Those are usually much more complex, structurally than things you find FDA approves at her year from small molecule drag discovery campaigns. So that's the part that you want to enable creativity. So like that's the overall goal. Now we have went through three phases and actually the company has been repeatedly at the frontier for the last eight years. It's been operating for a while. First we're a software company. So we're designing pathways using our software using AI that was actually the first deep learning based software for synthesis planning. But then we realized that to really address problems of our clients, we should also make them molecules as a bit like dog food. So like we decided that if we trust our design plans, we should also make them, right? So that's what we created Maria. So Maria is our synthesis platform and our main product. Maria stands for, it's only the name or is it an acronym for something? Right. So Maria stands for Maria Skoto Skakiri or is named after Maria Skoto Skakiri. So she was double laureate, Polish born chemist. And kind of just name always reminds us that our goal is to not only automate chemistry. So the platform is for making molecules, but we want to not only automate making molecules, but also discover novel ways to make molecules. So that's our inspiration. So Maria is a microliter chemistry platform and we can talk about this, but basically we make very tiny amounts of molecules just more than enough even to do it by your chemical tests, right? So we're kind of going to, I think, your question, which is what do we do for clients? The typical workflow would be that let's say a drug discovery company has an idea for medicine, but needs to tell us, let's say, 500 different designs. Let's say those designs might have come from AI, right? Then what we do is we help design these 500 molecules, so that we can make them with the platform. How do you design or how do you build these molecules? That's very important. So the problem with, you know, like advances in AI has always been enabled or preceded by large data sets. By scaling, yeah. You know, it's like in biologic, there has been PTP in GPT that's internet in chemistry. There is no public source of data, which is very interesting. Like it's a very unique field in this manner. So we have generated, let's say, internally using the platform, the largest, as far as I know, on the planet microliter chemistry data set, meaning data sets of experiments about which molecules can be made and which cannot be. It's a classification problem. So we have trained models on the largest data set of this kind. And those models classify us that help us in the generative process of designing this library. And we can, in this way, achieve. very high success rates and make design interesting molecules. But really it was a crazy bit. It was three years ago, we decided we need to scale up data generation and kind of design from ground up everything in the platform to enable scaling when it was very weird in some sense. And now three years later, fortunately, we are in a position where we have a data that can support our clients. Can you explain in a few sentences where did you get the data from? Yes, yes. So importantly, we have generated it and that's very important. So have designed the platform that it enables running, well, tens of thousands. But you need a lot of domain specific knowledge to generate molecule data or M&Rong. Yes, no, it's completely true. It was a very interdisciplinary effort and many things in the platform have been designed together to enable generating large scale data. So we have 300,000 over individual experiments, which in this area, that's the largest of its kind. And it's normally a chemist would do maybe 100 experiments or 100 to 1000 years early. And our platform does like 10,000 weekly. And this discrepancy is not, let's say, magic. It is coming through to the design of the platform in particular. We use very small volumes of liquid or volumes of experiments to drive down the cost of reagents. I guess an a very specific example. So yes, indeed, that was a big effort to generate this data. And I think all fields that are like in the broader sphere of industrial AI will go or have come through this like revolution, evolution of generating domain specific often generated in a vertically integrated company data to then enable AI. Because generic AI, like it generic AI in chemistry, you know, it's really good for like ideating designing overall ideas, but not very good. If you want to like, I want to make like better this very specific molecule that no one has ever made on the planet, right? It's always another problem because there are just so many potential molecules. That's that's very, very interesting. Can you please explain a little bit what kind of models are machine learning you using in the back there? Oh, yeah, of course. Basically, there are two kinds. There are models that can generate a reaction of any kind in the sense they ideate how you can make a given product. So there are different reaction classes. So those are generative models. Those are models we started with. And this was actually interesting because when we started, it was the first deep learning based solution commercially for synthesis planning. And those models were hallucinating when there was no word for it. So kind of, there was an interesting moment, it was like 2019 before a GPT, but we had the same issues with like, we had to have this like discriminator that was helping with so that they went crazy. Yes. And it was very important because when chemists see is a hallucination in the in the generated reaction, immediately loses interest in the platform because it doesn't trust it. Still, the trust issue in AI is a big topic, right? And then the second type of models, those that we more often train on our data are discriminators. And simply put, they take in the whole experiment plan, meaning what reagents you will mix at what temperature, at what reagents, and then they predict whether or not at what efficiency the reaction succeeds. It's a test that humans often have to do themselves are pretty poor at it chemistry is very hard to predict outcomes of. But because we have very specific data to specifically our platform, we have models that are very robust and generalize very well in general, because they have a very narrow task. We don't ask them to predict any chemistry. We've asked them to predict chemistry of the kind that we have generated data for on our platform. So these two mainly models. You already mentioned at the beginning that you also automate the process, right? Can you share a little bit more details when it comes to automation? Oh, yes, of course. So we attempt synthesis through a process I would call that where critical parts are automated and some parts are not automated and we kind of go there. So one critical piece of making a molecule is mixing reagents, basically. It's in general organic chemistry is not that far from cooking. I like to use this analogy. You know, it's maybe sounds funny, but it's actually very accurate. You mix things, you know, making tea at wrong temperature will not be a very successful process. Things are that sort. So what is automated is in particular reaction mixing. What's interesting about our approach is that we use commodity robots. So there are approaches to automation that go through very expensive, very big buildouts. You know, I'm talking 40 million dollars, 200 million dollars for the space equipment and things like that. Our laboratory is much cheaper. So I cannot share exactly how much cheaper, but it is much cheaper thanks to the fact. Yeah, it's much cheaper because we first of all, we naturalized everything in the sense that we do reactions in small scale and then we kind of tuned the approach to be able to use cheap robots. So in particular, coming back to your question, the reaction, reaction mixing is automated through a robot that's very cheap, like few thousand dollars already called open-trons, right? But it is fully automated and repeatable in high quality. Then another component after you mix reagents is to analyze the outcome of your experiment. And that part is also automated through auto sampling, meaning taking automatically the outcome of the reaction from the vial and going through the machine and then very importantly, analyzing it using proprietary data. What is not automated, though, it will be in the future, is moving, let's say, plates and reagents between those two and other stations. So that's how we approach it. We have been all, we are a startup, so we have been always solving the biggest problem. So where's the human time spent mostly? And it's not spent exactly at moving plates from place to place, but more like at analyzing results or performing reaction mixing and things, things of the sort. Coming back to your AI models and your AI approach, that sounds very, let's say, on Earth, right? That sounds not like an alpha-fold solution, big, big, big, big, big, big models or MM wrong. Yeah, I think I think this is a good way to put it. We have been in deep learning, you know, before it was this case and we kind of used the tool, you know, like very, very practically speaking. One place where we are increasingly both users and C potential is using LLMs for designing experiments like, you know, research hypothesis, that's super exciting. And also I mentioned mostly deep learning models that we develop in house. So there is a big role, of course, of using generic LLMs for overall company automation. So, that's also worth keeping in mind. But yes, those are like, you know, not billion parameter models and trained, you know, on 300,000 let's say, reactions. So it's not like internet scale dataset, of course. Let's talk a little bit about limits. What are the limits of your technology of your approach? Well, in a way, it's not that easy to answer because right now, because we really see a lot of potential just need to scale up. But right now, we have like two many things to do. But yeah, of course, there are limitations. We have three more ideas to scale up. Yeah, but like, two first, maybe, two question about limitations. One thing which is worth mentioning, and also I think illustrates better the companies that we work in this microliter scale or microliter chemistry. So just a second about it. So we usually perform reactions. And I think it might be interesting to the broader audience, even though it's a chemistry background, because it exemplifies something that I think also is the case with like, let's say, manufacturing of chips and other industrial AI, where you have a process that's like extremely fragile. So in that case, we do reactions in 30 to 100 microliter, something like that depends can go a bit above can go below, but this is an amount of like a few drops of water. Very small generally again, depends, but like it's really, really small. And at this scale, predicting chemistry becomes even more difficult. Models are clearly better than humans. And because of that, you know, it's also a limitation in the sense that we need to really be careful about our setup. And we really need to train those models and have this data. So one analogy that I think is really here helpful to think about is manufacturing of chips. Right? So it really now became pretty crazy with, you know, nano meter scale. And they had to do a lot of like innovation around the process, but also it's the challenge, right? Because now there's like one company that does those machines, right? The one in Amsterdam or someone. Awesome. Yeah. Exactly. Right? So it's also an issue when you were your process becomes really, really specialized. Of course, we are not like them, like they're, you know, way, way, way more advanced, but like this is this kind of direction, which I think is an interesting trend in industrial AI that you have this platform, which is pretty specialized. And you need to take very custom AI for it and very custom data. And that becomes your mold, right? Exactly. And where you want to scale in the future. Right. Like a few, a few things. I mean, one thing I am personally excited not to just kind of body with the overall company strategy is making big discoveries with the platform, right? So not only being, I mean, it's not only it's like very important, but not limited to being a provider of molecules, but also being able to discover novel chemistry, novel ways to make molecules like the something you could publish. And then kind of working on that in partnership. So that's something we have been doing more. We have some project success already in this category, some hopefully soon starting. some already in the process. I wonder this public like and talk about this really, really interesting. So and again, I think really kind of something that I think there would be more of both in our domain and overall in industrial AI. So our partnership with a company called WRGRISE. So no one has probably heard about it on the podcast, but this is the basically the biggest manufacturer of compounds or like chemicals in the US. So very often, let's say, when you buy something at the pharmacy, there is a chance that they have made it or you know someone like them. And this is a super complex problem that now there is more movement to reshore to US and Europe. Like US and Europe have limited capacity to do this and like most of production is in China and there is a trend to reverse that. And our collaboration with WRGRISE is in particular about using our platform to discover other ways to make important building blocks for certain category of drugs. So it's really exciting because we might be able to go beyond what is done through like managing AI with platform and like with amazing industrial partner. So we hope to do more of that. And I am very excited about this because really, and I was want to see the impact of the platform in a very immediate way. So that's one one scale direction that I'm really hoping for this year to grow. Coming back to your customers, how do you onboard your customers on the platform? What does the customer needs to do? Right. So right now it is a bit like, I don't know if you know these forward deployed engineers in the sense that we do pretty custom work with clients. There is not like glass of paint software that you click and it just orders molecules. But we want to move in this direction. Right. So right now it's more like we have a conversation with the partner. Let's say what he or she wants to synthesize is a library. So coming back to the simple workflow. And then we do maybe one or two or three back and fours agree on the design. And then we can move it to like our platform for attempting synthesis. So mostly it's kind of like this, I would say, on the front of synthesis, then there are these bigger partnerships. So those are, you know, like just projects or this is like just working together. And then we have clients for our software for synthesis planning where there is like basically no onboarding. It's really beautiful. You why you just go on and you are off off to go. We already talked about the outlook. But are you hiring? Are you looking for new guys joining your company? Just at this moment, we have one position open which is very interesting for business development person of chemistry background. In general, we are opening probably more positions in the second half of the year. So so mostly right now, BD. But yeah, I mean, if we if we open these new partnerships and continuous scaling, there is there is ton of work and in particular, I think now with LLMs and things like that, what we really need are these cross-disciplinary people who can think in different categories. I just started saying that I don't have chemistry background. But you know, to some extent though, I think people in my company would laugh a bit here. To some extent, I was able to understand some of chemistry, right? And I think this is really important to so we have, when we hire, we look for such people, right? That can both experts in some domain, but then can quickly learn other domains. I think this will be a trend in general in AI that more such people are important as we, as we attempt to speed up problems. At the end, I have one question about Washa and Poland, because Mr. Wieszko is coming to Washa. And there are a lot of investments in Poland, Czech Republic. So is there a new horizon in AI? Because a lot of talent is in Poland and Czech Republic and in mid-European, then let's say compared to Paris or Amsterdam, where everybody is going now at Zurich, is maybe Poland the next big thing when it comes to AI and to an industrial AI domain-specific AI. Interesting. For sure, we have amazing talent, right? So I'm extremely kind of I'm amazed by what folks like Leven Labs have achieved. Open AI was confronted by Polish people, or like Polish or rigid people. So the talent is amazing and this is a good question where it comes from. Maybe the background is very good, because from my point of view a lot of AI is mathematics at D&R, right? So it's math and the background in Poland is very strong when it comes to mathematics. It is. I think that's part of it. Yeah, it's part of it. So do we become a hub? I don't know. I mean, there are things that hold us baggage are interesting. So for example, in Poland, so I don't know, I'm just to select a very people there can think much bigger if I had to be honest and in Poland, where yeah, we see problems and I think now it's a problem because now, now we really also need to think bigger, I think. We try to do it more like you want. Sometimes better, sometimes worse is to try to match these two cultures to be ambitious in thinking but disciplined in analyzing. And it's very, very difficult. Usually people fall on one or the other, right? I would say often there is, say people have hard time with me at the company because I propose something that usually doesn't make any sense, right? Then they have to kind of criticize it. But yeah, I think there are things that also hold back the region. So you know, whether it becomes a hub is a good question. I don't attempt prediction, but I'm very excited for Poland and for the further region. Stanislav, it was a pleasure. Thanks a lot. Thank you, pleasure is mine.