This is the Data Science Conversations podcast with Damien Deighan and Dr. Philip Diesinger. We feature cutting-edge data science and AI research from the world’s leading academic minds and industry practitioners so you can expand your knowledge and grow your career. This podcast is sponsored by Data Science Talent, the data science recruitment experts.
Welcome to the Data Science Conversations podcast. My name is Damian Deighan and I’m here with my co-host Philip Diesinger. How’s it going, Philip? Good. Thank you, Damian. Great. So today we have a very special edition of the podcast. We’re going to be talking about the impact of data and AI in the steel industry. And we actually have four guests all from the same company, which is SHS.
They are the parent company that own Saarstahl and Dillinger Steel in Germany. We have with us Michael Schaefer, the head of AI and digitalization. Thank you. Ulrike Faltings, head of AI for R&D. Tobias Bettinger, head of Gen AI. and Anna Volker, who is the head of specialized AI. Hi, thank you. So if we just set the context in terms of your industry, guys, just explain. for people with no prior knowledge of the steel industry, what a modern steel plant actually looks like? I would say a modern steel plant is no longer just a mechanical facility, but it’s more like a digital ecosystem. The combination of sensors, simulation and automation ensures better quality and higher efficiency.
So in our case, a lot of digitalization, a lot of computer science, and of course, a lot of artificial intelligence runs in our steel mill and rolling mills. So I would say it’s a combination of technology and great people. So right now, this is also very much about a new mindset. You need people who are willing to come along for the journey and the change. what do you mean by high quality steel and how does it differ from commodity steel um oh that’s really a hard question the difference isn’t just better versus cheaper steel it comes down to composition precision performance and consistency you know you need to stay at cutting edge to remain competitive michael can you give an example where high-quality steel would fit in or would be used for use case versus commodity steel like Chinese steel that could not be used for such a case? You can say high-quality steel is used in automotive or in offshore wind industry where you need really good steel with great mechanical properties. And for some construction facilities, you can use commodity steel, I would say. So it’s a security aspect as well. Yeah, of course, sure. Interesting, yeah. So Damien already announced a little bit what SHS is doing, a little bit, you know, it’s a holding with lots of plans underneath and so on.
Can you explain a little bit what the mission of your team is in this context? I would say our mission is all about AI. So we started long before the AI hype. before November 22. So at the beginning, we started with three full-time people and one working student. And we focused entirely on AI for production. So industrial AI, I would say. And now we have grown significantly and have three sub-areas within the department. And what we would like to do is to run AI in production and to help the whole company with this AI to optimize our processes. And for that reason, our AI team is highly interdisciplinary.
So we bring together computer scientists, mathematicians, physicists, mechanical engineers, computer linguists, and so on. And that’s very important for us. So in the next couple of years, I can say that we run AI in almost every part of our company.
Very impressive. You mentioned that you have three AI sub-departments. Can you quickly, briefly talk us through what these are? In general, you can say, Anna’s department, so specialized AI, is mainly focused on projects in the steel production. While Gen AI is currently more present in, I would say, administrative functions. And Ulrich’s department, R&D, on the other hand, is running projects. across all areas and is doing a lot of research international and national projects. What size are those teams comparatively and when did you guys start the data function there? So we started in 2017 with a small group and right now we are 20 people. In Anna’s teams there are 10 people and in Ulrike and Tobias team five people. And you serve both Saarstahl and Dellinger? That’s right and all daughter companies.
Very impressive. So you talked us a little bit through the team setup now. How does your operational setup look like? I think if I remember correctly, you have more than 10,000 sensors that you’re reading out. You have lots of systems. You run your own compute center and so on. Maybe talk us through this a little bit. Oh, yeah. Very good question. So we are a company with a very, very long tradition. And that’s why we use the right range of technologies and techniques. That’s not always good. So we have to take care of various databases, ERP systems, MES systems, and many in-house developments. So we have a lot of in-house development applications. But that’s also part of what defines us, I would say. So it’s similar with sensors. Older plants or older facilities have around 500 sensors, while our newest continuous casting plant at Saarstahl has more than 10,000 sensors.
So it’s a very, very heterogeneous system. And what is the reason for your own data center?
So we have more than one data center. And one big reason is speed and data safety. So when you are right in production, you have no time. Everything is in real time. And when you’re running all your stuff in the cloud, it’s really hard to secure that you are always online. Very interesting. So, latencies would also play a role? Yeah, absolutely.
Can you give an example where that is the case? When you want to control the temperature on a steel mill, you cannot wait 10 seconds or so. So, you need to result immediately. Wow.
And you mentioned that you started initially with three people. Can you talk us a little bit through the journey, basically, how you develop the team, how your role within SHS has also been shaped over that time to where you are now? It’s been quite a journey, actually. Today, AI is used very broadly across our group or across Europe, but that wasn’t the case in the beginning. So first we had to fight for acceptance.
But one key point was involving the business units and the domain experts. the start so if i had to summarize our journey i would say we started early we focused on collaboration and we stayed open to new ideas and that was a good combination and and over the time we did a lot of project in production we had a great impact and i think that was our our success but a very important part as people so People matters, I would say. Yeah, connecting closely with the business side. And you brought three team leads today, so it’s clearly a success story. If we look at it from the other side, what would the people in steel manufacturing say how people like you changed the industry or changed SHS? I think to some point, using AI made the people more brave. Because now you have AI that encourages you to go new ways.
You have suggestions coming from AI models that gives the people the confidence to test new things, to test new approaches. And I think in some way we got braver and more eager to go new ways in steel production. Because steel production is a very traditional industry. We’ve been doing things the same way for… over 100 years and AI gives us the courage to or encourages us to test new things and go new ways. Michael, do you want to add anything to that? Yeah, we are able to do things today which we are not able to do 10 years ago. So we can optimize processes with AI, we can do some hyper automation, we can replace physical models with data-driven models and these new technologies have great impact to the steel industry. So with AI we are getting a lot faster.
We can optimize processes, we can hyper-automate processes and we are able to develop online models. When you have some physical models they are very slow because you cannot do all the equations in real time. But with AI techniques you are able to do that.
You can implement models very, very fast. And that’s a big advantage when you’re using AI in a steel industry or in every industry. Anna, give us an overview of your team. What type of people have you got from data scientists, data engineers, software engineers, etc.? How is your team made up? We’re actually like the entire group. My team is also a mix of people.
We all have… the broad range of disciplines in our team.
So in my team, we also have like classic computer scientists. But me, myself, I’m a physicist. We have mathematicians. So that’s this wide range of knowledge, I think, that enables us to do the things we do. So we come from different perspectives, looking at the problems. And yeah, I think this really enables us to do the things we do.
Can you talk us a little bit through what are the operational problems that you are solving on a daily basis? Of course, we have, I would say, the classic thing like increase the yield, decrease downtime. This might be via predictions, via optimization. But then, of course, we also take safety issues. We use computer vision to alarm people being in dangerous areas. So it’s really a diverse range of things we’re doing. So you mentioned data quality.
What does bad data mean in a steel plant? Can you give some examples? Yeah, I could talk about examples all day long here.
As we said before, right, we’re a traditional industry, which means we didn’t plan around the data model. We started collecting data. We started to collect data in different ways.
And being honest, a lot of things here still work on Excel sheets, right? This is the way people tend to work and they still work like that. So bad data might be data written in Excel sheets in different formats. So you cannot just read out them like a database.
Bad data might be… missing columns and databases column a being written in column b handwritten data everything like that also like we have let’s call it data garbage in the databases like some maybe a manual entry that stays there forever and then we we collect it and wonder what that might be or what that could be A lot of missing values. You have a lot of points you can stumble over when you start to build an AI model there. I would also say the biggest part in building that AI model is always getting your data and getting it clean. That’s the hardest work we have to do when building an AI model. Yeah, that makes a lot of sense. How about the data that comes directly from the plant, like the sensor data? Sensor data, that’s always a good sign.
Having sensor data, yeah, that’s actually the best. Having sensor data that is written to a database, that’s a good start. Not meaning it cannot have any bugs in it. I mean, a sensor can fail too, right? But that’s kind of the perfect scenario, having sensor data written to a database. You have just one problem with sensor data, the data drift. So the sensors age, furnaces get relined, raw material change, and what was normal signal six months ago slowly shifts. And that’s a big problem for us with sensor data. But Anna’s right, that’s good data for us. So maybe one question, where does low data quality… prevent you from from building something engineering something that you would like to do like where does it hurt the most when the data quality is low i think it hurts in every project we’re trying to build where the concept is plausible like you would say having this concept we could build an ai model we could build an ai model that would have some benefit for the company but we just don’t have the data that’s disappointing right we don’t collect it the data quality is low and We cannot do magic, right? If we have no data, we cannot do data science. We cannot build AI models. We have our ways and tricks to deal with that, but there’s a limitation to what we can do. So, yeah, there is a point where data quality is so low, we just cannot do anything.
And that’s always disappointing when you have a great idea in mind. So even not modeling from first principles or so, is there not an option? Of course, modeling from first principles, that’s something we do. If there is some database and say we have some missing values, if there are first principles you can model from, we do that. But building an entire AI model on first principles, I wouldn’t say it’s impossible, but I imagine it to be tough because we always need process data to… describe our process. Processes are very complex.
In a steel plant, in a rolling mill, everything is really complex. So model the entire thing from first principles, it’s nearly impossible. Imagine modeling a blast furnace from first principles. You would have to be a genius mathematician, physicist to do that. So I would say we always need data and can fill. missing data with first principles. That makes sense. Do you have any tricks up your sleeve for building robust data pipelines in your environment? Yeah,
you have to remember the process and mind the process and to talk to the main expert.
You cannot blindly build your data pipeline. You must talk to the domain experts to What are classic ranges of your data? What can change? What changes do we have to expect so we can model them? Then you have to implement guardrails to prevent your data drifting away to some areas you don’t know. So it’s collecting this domain expert knowledge to give your model and your data guardrails. I think that’s the… That’s the most important point. And then, of course, monitoring. Once you have your model running in the plant, you have to monitor it and see and check if it still performs the way you expect it to perform. And most of the times, there will be times where your model doesn’t perform the way you saw it in training or something changed and data is completely different to a year before. You just need to monitor and adapt. so if we come maybe we go a little bit into the core use cases that you guys are working on maybe you can talk us to two or three of those so so which use cases delivered most of the value in your work or in the environment that you’re in when we talk about saving money i remember a project where we assessed input materials not just by their price but through the entire process chain.
So an input material goes to the blast furnace. There it has some effects. Then the pick iron goes to the steel plant. There this former input material has some effects on the process we have to run
in the steel plant and so on and so forth. Then we have side products we sell to customers.
Having this change input material, this has some effect on the price we can sell our products. And considering this entire chain of processes, starting with the price of the input material, this had a great effect when we talk about money. Because price difference might be small, but we produce thousands of tons of pick iron a day. So even a small difference in the price of the input material might lead to… large amount of money that we save another project that comes to my mind was with a big customer of ours we were seeing defects in a product we produced so there was the risk of losing this big customer and we worked together with the domain experts to work out what was the cause of this defects and we were able together with the domain experts to find the cause and to turn it off so We could save our biggest customer, you could say, in this one regime. I mean, you cannot put a price tag on it like you can do when you have input materials and the effect and so on. But I think we can all imagine what it means to lose your biggest customer in one domain. Michael, anything you want to add? Anna told you a lot, but we have a lot, really a lot of these projects. So some other examples are our temperature model. We developed a data-driven or an AI-driven temperature model for our two steel mills, one in Dillingen, one in Fölklingen. And these data-driven models are performing a lot better than our traditional physical models. Or we can say physical tests. We are able to predict our oxygen level in the steel and we don’t have to do a lot of physical tests anymore, just a few to control us. we can say we save the most money in production. Very impressive. Anna, maybe another question to you. So you’re leading the specialized AI team. What kind of specialized AI is most important for you? What does it mean in your context? A specialized AI in our context really means it’s tailored to our company and not only to the company, but the specific process we’re looking at. Our processes are, let’s say, so special, we can hardly buy AI off the rack.
So everything we do is tailored to the data we have, to the process we’re looking at. And that meets specialized AI in our context. What kind of models or methods are you using? I think we use everything that’s out there in data science, right? When we’re in computer vision, we use YOLO models. When we’re working with tabular data, we use boosted decision trees. So everything that’s currently out there, we test it and it’s in use here. As I said, we have a wide range of projects we’re doing. We also have a wide range of techniques and methods we’re using. So if I understand you correctly, you are looking at the problem first and then you’re choosing the method that fits the problem most. And you adapt it also to your data, to your process and so on to make it work. Right. But there’s one hard problem with integration. So at the same time, everything has to integrate into an existing… highly heterogeneous it or ot landscape so we have legacy systems we have different vendors different generations of automation and that integration right into the production is often not just challenging uh it’s it’s more challenging than the model or the application itself in our case so that that’s really hard for us so the like the integration part can be more elaborate than the the model building part that could happen yeah because we we train our data with with some data warehouse data historical data and some real-time data and then i would say um it’s getting real and then you have a problem and then you have to integrate it very very probably in our processes and systems For building these models or these methods and integrating them and so on, do physical experiments play a role or calibrations? I would say not so much. Our usual approach is building a model from historic data and then getting that to run parallel to our process.
Running parallel meaning the model doesn’t, or the prediction of the model or whatever the output of the model might be, doesn’t control the process yet. but its output is deployed to the engineers in the plant or whoever is interested in that output. We at the same time collect the data, the process data and what our model is saying, what the experts at the steel plant are doing out of it. And then we can see what the model is doing and how it would behave if we implement it in the process. And you mentioned already that you’re closely monitoring and maintaining the models. How do you detect model failure earlier? Most of the times I would say the domain experts detect it because they’re working with the model on an everyday basis and they will surely contact you if predictions get… unrealistic if they don’t see any forecast, if they don’t see a picture or your web interface isn’t working. So we basically have our failure detection right at the plant.
That’s a big problem for us. So in practice, model failure almost never looks like a clean crash or an error in the system. It’s much more saddle. And that’s what it makes very dangerous for us. And you would need domain knowledge to detect it, actually, because what might look like a normal forecast for us might be highly unrealistic for someone knowing the process. And that might be because data has changed and is now outside the model’s data range, right? And we would never detect that. There are methods to detect it, but they are less reliable than having an expert looking at the model and its output. So it’s very important to still have a human in the loop. I would say we, almost in every project, we have a human in the loop and it’s a highly important feature. Yeah. You’ve talked a lot about the role of domain experts.
Obviously, they are in no way data experts. How have you seen them evolve to understand and accept the data, what it says and what the models say? Can you talk about that briefly? It’s actually a great evolution to watch because as you said in the beginning, there are no data experts at all. we often have meetings with them because we’re a project team, right? We really work closely together and they teach us about the process they are experts in. And we kind of teach them about data. So in many projects, you suddenly have a steel plant worker knowing about SHAP values, right? Because we talk so much with them and for them, they do understand what we teach them about data. And you really can watch and observe like being click and they understand what the data is saying. They have new questions, really interesting questions. And seeing that interest and also the data on their side, that’s inspiring. Also, it gives you a great impulse, often in the right directions, because they are asking the right questions to challenge your data and challenge your model. So, yeah, we are learning a lot about the process.
They’re learning a lot about data. I would like to take a brief moment to tell you about our quarterly industry magazine called The Data Scientist and how you can get a complimentary subscription. My co-host on the podcast, Philip Diesinger, is a regular contributor and the magazine is packed full of features from some of the industry’s leading data and AI practitioners. We have articles spanning deep technical topics from across the data science and machine learning spectrum. Plus, there’s careers advice and industry case studies from many of the world’s leading companies. So go to datasciencetalent.co .uk forward slash media to get your complimentary magazine subscription. And now we head back to the conversation.
So Ulrike, you’re leading the research and development basically into AI to support the other teams as far as I understand. Could you give us a little bit of an overview of what you and your team are doing? Sure. so like research and development sounds very academic but of course we’re still a steel mill so it’s applied research meaning it’s maybe more experimental techniques or maybe something we read a paper about and thought that sounds like viable for our processes and then we try it out and see whether it works but of course there’s like a chance that it won’t work but it shouldn’t be completely off so fingers crossed that it will do something useful And it’s also a joining research project funded by the EU or maybe the German government. We also try to work on topics that have an actual benefit for us, so not too academic. But of course, in research projects, you have no guarantee that it will work out. So that’s why it’s getting funded. Can you give some examples for projects or use cases that you’re working on? So, for example, one research project was an EU-funded project where we digitalized and partially automatized the scrap process in the steel mill, meaning we developed another technique for creating a data set for steel scrap, but then also trained an AI model that classifies different steel scrap varieties. And then further on in the process, we also trained another AI model that can predict the chemical content in the like liquid steel based on the type of scrap we used.
And these were also like more experimental things. We weren’t sure it would work, but we say we try it out because it’s a research funded project and actually the results were very promising. So yeah, nice.
And how do you work together with your colleagues or with the other teams? Is it that you build kind of an earlier version of the model, like a pilot or so, and then you hand it over to be
integrated within the running processes? Or how can one imagine that? Internally, it’s a very close collaboration. It’s not like strictly separated. But of course, like my team will put more hours into this research project. But then… Of course, collaborate with Anna’s team and maybe turning over the more mature model once it’s running and just needs supervision. But internally, it’s also like maybe if Anna’s team has a lot to do right now and our team is still waiting for the funding from EU, then, of course, we can also help out over there. It’s not strictly. separated which is i think very important because otherwise you sort of lose the feeling of what’s actually going on in the plant if you just do research so it’s i think it’s really good to like keep in close contact over all three of our departments and talking about your team what kind of talent do you have in your team i’m imagining a very diverse background probably as well with chemistry physicists data scientists Yes, yes. Just like in the other teams, it’s like a lot of, of course, like natural sciences, but also computer science. So it’s really very diverse, which is good, I think, because it always gives you different perspectives on how to approach things. So you’re basically working on the unsolved problems. It must give you a really great perspective on that topic. What are some of the biggest unsolved problems in industrial AI that you see? I mean like one problem which is like also Anna mentions of course sometimes you just don’t have the data you want because maybe it’s like some very subtle effect that’s causing problems at some point in the process and it’s maybe really hard to measure exactly this effect and you’re not sure where it’s actually turning up so you wouldn’t even know where to place sensors to measure it directly and maybe it’s like yeah like You can see it blow the noise in some other sensor readings, but it’s hard to find. And so maybe, I guess this is like the biggest issue we were always working on. Like how can I optimize this process where I don’t really have the data, but I want to like try to get there and how can I still model it or find other data that can help me. makes sense because i mean it’s like it’s a dirty like environment is perfect not like in a lab and then maybe michael back to you so from from what i see it’s not so common to have someone working on actual research in ai in a in a team of around 20 people what was your like motivation behind that behind the setup i think we would like to stay state of the art you have to stay cutting edge and to be competitive and you can really learn a lot from from other teams or other other cultures in europe and so on and that that is that what motivates us most i would think so you can discuss a lot of ideas with with with other with other steel plants or with other researchers or other data scientists and that’s that’s that’s really a great advantage for us so
you mentioned state of the art tobias you are actually working a lot on gen ai topics Can you
explain a little bit what it is that you are working on and what role Gen.AI plays in the steel mill? Yeah, thank you for the question. I think Gen.AI is the topic that most people are talking
about at the moment. That’s right. And all the people are saying just AI, but as you have seen, we have other topics regarding AI in our company. And yeah, in the Gen.AI aspect, we are working with mostly in administration at the moment so we have some problems that we try to solve in technical feasibility analysis of customer requirements which is a quite complex task in reality so our experts take most of the time two weeks in order to understand
all the customer requirements and the impact on the production processes and this can also be carried out by the use of ai models jai approaches and it saves a lot of time actually another project we are working on is the automation of customer inquiries so we also receive pretty heavy unstructured data there per mail and it’s also manual process at the moment so Yeah, all the employees of the sales department have to extract the information manually, and also GNI models can help there. So if I understand correctly, the GNI models are mostly used in administrative tasks, while other types of models are more used in the production environments. That’s right, yes. so jenny is notoriously known for problems like hallucinations and so on you know safety is a big concern guard rails and so on so forth are there some areas where you decide not to deploy it because of those issues potentially yeah yeah you got a point here um obviously obviously hallucination is a big risk if you take a look at the steel plant it’s also yeah area where safety matters And I don’t think that it’s a good idea to go directly with fully agentic AI systems into the steel production process. But also we try to mitigate the risks by the use of human in the loop in most use cases. So I think in the near future, there also might be a possibility to apply such systems in production processes. at this point in time we are focusing on on the other areas of our company yeah and you mentioned agentic are you using agentic workflows like orchestration between different agents at the moment or is it always single agent based there are use cases that where we need actually some some more agentic teams but in most use cases one agent is sufficient at this point in time and Yeah, also at the moment, we are just focusing on abstracting the workflow that the human teams already have detected or implemented. And then we can make use of this also in the agentic AI system.
So how can one imagine that? So is it that you are finding mechanisms to capture the knowledge of the human workers into like a knowledge base that then enables the agents? Or what’s your process for that? yeah it’s also the the same process that we established in the other projects which overall ai department is using so we are working in close collaboration with the people in our company we try to understand their domain knowledge strictly in detail actually and in order to being able to implement such an agentic AI workflow.
So without the domain knowledge that we have to extract in workshops, it wouldn’t be possible.
So your team is basically conducting workshops then with the domain experts. I’m assuming you also have some data from them and then you try to kind of build a knowledge base and then have agents basically develop based on that. That’s right. That’s right. You have to understand their daily work and then you can design the system that you are trying to develop. When you design such systems, how do you deal with establishing guardrails? Yeah, guardrails are also very important. So one of the biggest guardrails is something that I just mentioned. We have a centralized GNI platform in our company. And in this platform, we can make use of elicitations as kind of pop-ups. And this is the human loop mechanism.
So we are trying to catch the errors as early as possible in all the processes just to ensure that no bad data is entering the downstream parts of the company. Makes sense. This is one of the biggest guardrails and actually we are also implementing a test phase as Anna does with her team. So we have a test phase running together with the process owners and the domain experts and they are giving us the goal for going into production.
When you work with the domain experts, what are the biggest adoption barriers that you run into? The biggest adoption barriers are not related to the people we are working with it’s it’s more like the integration complexity that we are facing it’s something that michael has mentioned before so we are working in a very broad i.t landscape with different proprietary systems and all these things so it’s very complicated or complex to gain access to all the data structures, data sources, and different systems. That’s the main complexity. And maybe looking a little bit more into the future, I think we had conversations about self-learning AI systems that you’re interested in. Potentially, what’s your perspective? The perspective on self-learning agents or something, I think it’s a very interesting topic that we also should try out.
But it’s not something that we are focusing on at the moment. But I think self-learning systems could be used for detecting errors also in the Gen.AI pipelines, maybe. This might be quite helpful use cases for ourselves. It’s not something that I have and my team has in its mind at the moment, to be honest. running, are they commercially available foundation models or are you running local open source language models? It’s actually a mixture of both.
We are focusing on OpenAI models that we can use in Azure. So these models are doing the heavy lifting and in the broad sense of all use cases. And if we are working with sensitive data, Then we have to take the approach of hosting everything locally and we have to ensure that no data accesses the cloud. So we are running in a hybrid cloud environment and we can switch directly from cloud to local or that’s how it works for us. And what would you say is your biggest Gen.AI success story so far? I think the internal Gen.AI platform because it already helps our employees with common questions.
So they can make use of it instead of searching information throughout different sources and so on. And the Genii platform also takes care of the governance and security of our company, which is also very important for us. Good. Okay. So, yeah, if we move back to you, Michael, what’s the strategy for the future of your area look like?
I would say our goal is to make the entire SHS group AI powered. So AI everywhere. So we want to use AI across all areas where it adds real value.
So a real impact. Not just for AI. So we need a real impact. And in my view we are on a very good path and the individual pieces are starting like a puzzle and they come together. And step by step a clear picture is emerging so in the next couple of year we will use more ai and we will optimize and automate a lot of more processes in our company what advice if any do you have for data scientists or other data professionals that might be thinking about working in heavy industry oh that’s that’s a very very good question first i would say be curious about the process and not just data because in heavy industry the most important insights don’t come from algorithms alone they they come from understanding processes and understanding people And second, your most valuable collaborators won’t be other data scientists. They’ll be engineers, operators, guys from the shop floor with decades of experience in their domain.
That’s very, very important. And what would you tell your past self at the start of this journey if you could go back? When I could go back? I’m not sure. I would tell myself to slow down on the technology. and speed up on understanding the processes and the people and what they want. I would also tell myself to embrace imperfection. Nothing is perfect in the industry. The data will never be clean. The process will never be stable. And waiting for an ideal condition just delays impact. So finally, I would say be patient and play the long game. Good. This is a question for Ulrika. Given that your team is very much on the experimental side, working on these unsolved problems, how do you convince the business guys, the domain experts, to invest time in a project that maybe has a low probability of ever producing a business outcome for them? I mean like obviously if it’s too much time for the shop floor guy
then it won’t work so more work on our side instead.
Where there’s more experimental problems it’s usually like the people the engineer maybe has a problem they have been trying to solve it for years or even decades and it just never worked out because you didn’t have the techniques to do it so usually it’s them like asking us is there anything you could do like to help us and actually solve this unsolved problem so this means they have the motivation and otherwise it probably wouldn’t work as you say so it usually comes from them these like very these hard problems. Well, you’ve clearly built up a high level of trust with the business for them to actually come and volunteer in the first place because it’s usually the opposite dynamic in most traditional sectors. Yeah, that’s true. I mean, I think in the past years, we’ve proven our value and we’ve shown the engineer side that we can work with them together. We share also the glory in the end when it works out. and like we we will honestly tell them no this won’t work like just okay don’t waste your time but we will also honestly try to make it work and i think that’s what the other side has like experienced from us in the past in like maybe more obvious like easier problems and this helps now with solving these difficult harder problems this like basis that we’ve found in the past Otherwise, I think it would be very hard indeed, as you say. I think the final thought,
and whoever wants to comment on this, is I think the future for heavy industry is somewhere at the intersection of traditional data science machine learning and OT.
And we’re headed somewhere also where physical AI… as they now call it, is also somewhere at that intersection. Do you have any comments on how that future plays out?
Is that somewhere you see the steel industry heading and your team’s role in it?
I guess there are use cases where physical AI might also be interesting, especially if it’s also a security thing that you say people shouldn’t have to be there. And it’s extremely interesting to be able to avoid that risk for human life. So I think it might be something in the future, but not in the very near future. So it’s maybe a few years ahead still. Are there any new types of models that you’re looking forward to that you’re excited about being released? Yeah, I think Chennai is strong at the moment, but… The release cycles are very short, which is quite interesting. And it’s also hard for us to keep up with that pace. So, yeah, I think I’m just excited about what’s next.
So I don’t have some special thing in my mind, but I’m pretty excited about the next models,
what they are able to do. I think we are seeing a big change right now from really big and huge large language models to smaller and more specific ones. And I think that’s very interesting. So in the next couple, I would say weeks, not months, you can run very small large language or large multimodal models on your PC, on your graphical unit, on your usual PC. In what way do you guys use Gen AI in your daily work?
Do you co-work with AI agents? i think coding assistance is already adopted also in our it department in general people are using it And yeah, I think the boost, you can recognize it in your daily work. But yeah, you also have to be cautious.
You have to take a look, what is this thing doing? And you have to know what you’re talking about to the Gen AI model. So I think you have to be a good software architect in order to give the model the right instructions. But yeah, sure, we are using Gen AI in the daily work.
Maybe one last question from my side. If tomorrow morning someone would unplug every AI model at SHS, like what would happen? Like what would break first? Would there be things that might not get noticed even or what would happen? I think people will be angry because they have to type texts again or brainstorm by themselves. That’s I think the biggest thing now.
What are you guys thinking? What is happening in the other departments? We get a lot of phone calls. Oh, yeah. Yeah. If we unplug every AI model, I don’t think the middle will stand immediately, right? But people will notice and we will get a lot of calls saying this doesn’t work. We don’t get access to this and that, right? Yeah. I think the most… So we have production relevant models like the temperature models that will get noticed first. It wouldn’t mean we stand, we have downtime because that’s also a big point. We always have fallback options. We try not to solely rely on AI, but to have fallback options, right? So we wouldn’t cause downtime, I would bravely say at this point, but people will notice. and they will have to get their meal plan from the intranet again instead of asking our gpt unfortunately that brings us to the end of today’s conversation michael anna ulrica and debias thank you so much for joining us today it was an absolute pleasure talking to you all thank you that was fun thanks thank you thank you very much thanks for having us yeah thanks for having us you’re welcome thank you also to my co-host Philip Diesinger and of course to you guys for listening before we leave you i just want to quickly mention our industry publication the data and ai magazine it’s packed full of insight into what’s happening in the world of enterprise data and ai and there will be an edited version of this conversation featured in one of the upcoming issues you can subscribe for that free at Data Science Talent.co.uk forward slash media do check out our other episodes at data science conversations .com and we look forward to having you with us on the next podcast thank you
