Aprovechando la inteligencia artificial generativa en Bosch

Episode transcript

Rebecca Parsons: Hello, everyone. My name is Rebecca Parsons. I'm one of your co-hosts for the Thoughtworks Technology podcast, and I'm here today with my colleague Birgitta.

Birgitta Böckeler: Hi, everybody, I'm Birgitta Böckeler. I'm a Technical Principal with Thoughtworks in Berlin, Germany.

Rebecca: We are joined today by Andreas, who's the CTO and Executive Vice President of Bosch Digital. Andreas, can you tell us a little bit about yourself and about Bosch?

Andreas Nauerz: Yes, absolutely. First of all, thanks for having me. Maybe, I'll tell you a little bit about my short history in Bosch because it's quite untypical. Normally, when you're talking to someone being the CTO or EVP of Bosch, they all have this typical Bosch leader, like being 20 years with Bosch. I can't offer you that. I joined Bosch in 2019 at corporate research and then, one and a half years later, I was then appointed the CEO and CTO of Bosch.IO, where we have been looking into Bosch AIoT strategy and where we developed AIoT and IoT-related products and solutions.

Then exactly as you have said, at the beginning of this year, I became the CTO of Bosch Digital, which is, at least as we call it, which is the digital powerhouse of the Bosch Group on its mission to drive forward the digital transformation. Before I joined Bosch, I worked for different software companies worldwide. I think the longest time for IBM, and there as a program director and a senior technical staff member primarily responsible for IBM serverless compute solutions. Yes, I studied computer science, by the way. Yes, that's me.

Birgitta: I'm from Germany as you are, so I am aware of Bosch who are based in Stuttgart but can you tell the audience who are not immediately aware of what Bosch does, a little bit more about the many things that you do?

Andreas: Oh, yes, and that was really hard for me when I joined. It's not an IT company, it's a company that is so diverse. I would call it a conglomerate if that translates well. We do a lot of crazy things and that makes life exciting. What is Bosch? The Bosch Group is an engineering and technology company. It's headquartered in Stuttgart, Germany — the Southwest — close to the Black Forest, some might know that, and close to Lake Constance, which is, in summer, wonderful.

Our strategic objective is to facilitate connected living with products and solutions, exploiting the power of different technologies like IoT, AI, and so forth. Our overarching goal is to improve the quality of life worldwide with products and services that are innovative and spark enthusiasm. And that's why we have our slogan: “We create technology that is invented for life.” This is our mission, right?

We operate in multiple sectors, so now we come to this conglomerate aspect. We are active in mobility, mobility solutions, industrial technologies, consumer goods. That's where most people know us from, right? You may have seen a fridge or a washing machine or whatever, or a trailer from Bosch already, and we are also active in energy and building technology. A lot of different things. And this broad spectrum is what, for me at least, what makes Bosch super exciting because every single day you can learn something new regarding to one of these sectors that I've just mentioned.

Of course, all these sectors and business rely more and more on IT and digital technologies, especially on software. I can bring in my background to drive all these different topics forward, especially now being in Bosch Digital when just the world of Bosch Digital, so that's an 11,000 people organization on acting around the globe. As I've said, we are on our mission to help the rest of the Bosch universe to drive forward the digital transformation. We develop and operate products, services and solutions so that all our internal customers can run their existing business and build up new digital business all based on our digital solutions.

Our portfolio ranges from infrastructure, workplace solutions, software development tooling so that the rest of the Bosch group can develop the software they need in their domain, even consulting services and products that we are trying to develop.

Rebecca: What we wanted to talk about today, actually, there is, of course, so much hype around generative AI, and we've got the apocalyptic vision of the end of the universe or the end of the earth as we know it, to, "Oh nobody's going to have to work anymore because everything's going to be AI and everything in between". What we'd like to talk about here is let's ground that hype in an organization that has the breadth of offerings and requirements and use cases that Bosch does.

Tell us a little bit about how this whole Gen AI revolution is playing out. Where are you looking at using it? Why might those use cases make sense? How are you approaching it? I know that that's broad, but let's just start with the basics of where you're looking at applying it now.

Andreas: You phrased it absolutely correctly. Due to the fact that we are such a big conglomerate, we have many, many different use cases that we have to find a mechanism, what to really go after and what not, right? Maybe I start with a short description, how we identify the use cases we want to pursue. We go, I would call it a four-step approach, with the team that we have built, and by the way, we have built a team or set up a team that is cross-business units. That's something that Bosch is not doing all day or every day.

We have to break the silos between these different business units that we have, I would call it, a center of excellence has been set up. This team, comprised of people that come from corporate research, comprised of people that come from Bosch Digital, they usually follow by working together with our business units, our internal business units, which are our customers, they follow a four-step approach I would call it.

First, we talk to all of them and we collect all the use cases, all the ideas that they have, no matter if I talk to the power tools guys, or if I talk to the consumer goods people, we collect all the use cases. Then we analyze all these use cases. Because what we want to do is we want to go for those solutions that seem to be low effort, but cause max impact for a plurality of our customers.

If there's one use case where we say, "Oh, if we build that, and if we build it in a very generic way, then we can help not only the what customers, not only the guys that are working on fridges, we can help a plurality of our internal customers at once." We cause max impact with low effort, so to say. Once we have that clear, we cluster these things a little bit because when you talk to all these customers, what you find out very quickly is the needs they have, even though they are in different domains, there are clusters of things that they all need, and that is something that you can then build a generic foundation for to then implement their specific needs that they have on top.

Birgitta: As a consultant, I now have to bring up the word synergies, yes? You're looking for synergies across your business units. [laughs]

Andreas: Yes, that was the word I've been looking for. Exactly, that's what we want to do. We listen carefully, we try to detect the pattern that we see, and then once we have seen the pattern, reassess what is causing max impact and what is leading to max synergies to make sure that we help this organization the best way we can.

Then, of course, we check the technical feasibilities, or some of the use cases, we might find out, "Oh, actually generative AI is not the right thing to use here". We can talk about that maybe later because, currently, I also see this trend or more problem that everybody's trying to solve the traditional problems now with generative AI, that's something I would not recommend to do. There's this third step that we do. We look at all these use cases and then we try to check for technical feasibility, and if it really makes sense to tackle the problem with Gen AI.

Last but not least, we also have to do, I would call it a legal status check. We also check, if we do that with generative AI, that we do not violate laws, that we take care of data protection regulations, and that we do not harm individuals' personal rights, something like that. This is how we tackle the problem. Use case collection, clustering, assessing the impact, making sure that generative AI is the right tool to be used, and then checking that we do not violate any rules that we better do not violate—four steps.

There are also four clusters. We have been asking for, "What are you using it for?" I keep it first on a more abstract level, and if you want, we can go into more detail what concretely we are doing. The four clusters, and feel free to ask in between, of course. The first one is what I would call the search and summarization scenario. This is all about searching or allowing people to search in natural language, to analyze and understand large volumes of data in order to ease and fasten and consolidate access to information. That could be the headline for this.

In this scenario, generative AI enhances enterprise search and data summarization by allowing to search a natural language and keywords. We all probably have this internal search engine in our companies. I'm not sure how yours works, but the most internal search engines that I've seen are not the best ones I've ever seen on this planet earth. I think many people know what I mean.

Now we have a new tool here where we do not have to come up with the right keywords, and even better, we are not being presented with this list of links that we all have to open in a new tab and see if the information behind is really what we are looking for. Now, we get not only the opportunity to ask our question in natural language, we also get a concise summary of relevant information rather than that list of links that you have to go through manually and that makes it so much easier and faster to extract valuable insights and to make, at the end, an informed decision. That's the first cluster.

Birgitta: Maybe a question about that first cluster. It's a very common use case that we also see with a lot of our clients. What are your experiences so far with the correctness of results and also data quality? I think a lot of people are almost having this hope, there's this magical data lake now of unstructured information that we can ask questions about, but the data quality still matters. What are your first experiences there about preparing your data in a way that gives you added value?

Andreas: That's a good one. I want to be honest here. We have just started that, and due to that, I think it's more like a gut feeling what I'm answering now. I think people are, with the experiments we have been doing, quite okay, but we also see and not only we, we see some of the problems that are quite surprising. I'm not sure if you've read about that Stanford paper recently. That's very interesting. You see, for example, even in systems like ChatGPT, that they can even derogate, if that is the right English term. You see that, even so, an answer has been correct back in March, for example, if you ask the system again, it's wrong now because they use reinforcement learning and stuff like that. That means the system changes its behavior over time based on the feedback it's getting from the users.

In March, maybe there was a different population, only a small group experimenting with this technology. Now it's the entire world. I'm bored. I do not want to touch on the IQ of all people on this planet Earth, but obviously, systems are, for whatever reason, derogating to some extent because they get the wrong feedback.

Birgitta: Yes, or also not necessarily wrong feedback, but an answer might be useful to one person and not to another person, the same kind of answer. Then, was this useful to you, thumbs up or down? For one person might be thumbs up, the other person wanted something else. Correctness is also subjective sometimes. [chuckles]

Andreas: Absolutely. That's a very valuable addition that you just made. I think the important thing that you need to consider when using generative AI and that links a little bit to what I was already teasering a couple of minutes ago, it's a very unstable technology.

Due to the fact that it's using reinforcement learning, if you need something-- let's say you are working on a use case that is safety critical, I wouldn't use that because it's an instable technology and you might run into problems that you do not want to have and that then you better use one of the more traditional concepts that give you more security or certainty about the results to expect.

The second cluster I would call it the chat voice bot scenario. This is about adding natural language to chats to facilitate human-like understanding and responding in natural language to improve interactions. In this scenario, generative AI is enhancing chat and voice bots by improving the natural language processing abilities of these systems so that they better understand and respond to consumer or customer inquiries and requests.

It could be internal employees, so maybe we equip our internal IT support with such technology, but it could also be that we apply this in our service centers, for example, where our external customers are calling. This is, of course, leading to a better user experience, no matter if it's an associate or an external customer, and to making people way more happy. This is something we are really looking at at the moment. We are looking at how can we improve our internal IT ticketing system. We are looking at our customer supports and their agents, or how we can be a system.

It's actually funny. If you look at how our service centers work, some people think there's only a human sitting there and that human gets the call coming in. That's not the case, actually. That's nothing new, but now we have a new technology to make it even better. The voice is being split in two parts. Of course, the person you're calling hears what you're saying, but at the same time, there is an AI analyzing in real time what you're saying.

Of course, now you can put a model behind a large language model that has been trained on all the manuals, on all the problem tickets that you had before. While they are talking to each other, the agent sitting there gets in context in real-time hints what the problem may actually be. That's an assisting technology. That's something very concrete that we are currently-- not only we, you find that all over the internet. It's a very, very common use case currently, and that'll be very, very helpful.

Rebecca: The nice thing too about using these capabilities, one of the real frustrations with many of those voice systems as well is, no, you can't use the word speak to a representative, you need to say, speak to the pharmacy, or, what is exactly the word that is being looked for, and yet with these large language models, I can quite naturally say, I want to speak to a person, and they'll route me to a person, because whether they want you to say pharmacist or representative or whatever, all of those things can now-- It becomes much more natural to be able to correspond with these systems. Even if that's all you use it for is to improve that interface, from a customer service perspective, that makes a huge difference to--

Birgitta: Makes it fuzzier.

Andreas: Absolutely. I think we all remember these funny systems where they are telling you, if you want to X, Y, Z, please say one. If you want to do A, B, C, please say two. Oh, I didn't understand. Can you please repeat? This is a story of the past then, right?

Birgitta: Yes.

Andreas: The third cluster, and now we get more into the creation of content, I would call it. I think this is very powerful, and very powerful, not only what you may think of immediately like engineering and software development, but channel the content creation. I think using generative AI to support content creators, like for example, marketing specialists by assisting with the creation of high quality and maybe even personalized content at scale, this is a super powerful thing to improve efficiency and to reduce costs while allowing the creators that originally have been responsible for doing so can focus more on the strategic tasks.

One project where we see this currently is, we have a little tool currently in the PLC where we generate our texts for websites in different languages and so people are being relieved from doing that manually. We can do that in 47 languages. We can just let it be generated for what previously people have been doing by hand. For example, if you are in HR-- we are discussing with our HR colleagues, if you want to generate a job offer, a job advert, why do you want to write that by hand any longer? Why do you want to write that manually? Let it be generated by generative AI.

Of course, you need to be careful because we know there might be hallucinations in and also the system has not an ethical understanding. Be careful, there might be, I don't know, violent language, harmful language, whatsoever, in there. Please have a human being have a review on that. Still, it's better to start on this 80% as a basis than writing something on your own from scratch. That's the third cluster, the content creation cluster. Marketing is just one example, right?

Rebecca: Yes. In fact, one of our colleagues, I was talking to him about how he's used it for content creation. One advantage of course is you immediately get past the terror of the blank sheet of paper because you've at least got something and it's always easier to edit something that exists than it is to create from the whole cloth. What he also told me was there was one thing that he just never liked anything, and he tried two or three different prompts and never liked anything that he got back, but he got so irritated with the fact that why can't this model come up with something sensible that that got him over the terror of the blank sheet of paper and motivated him to go write something.

I do think that that's interesting but there are so many stories of when people haven't taken your caution into account. Please keep the human in the loop! We have lawyers being fined by judges because they put together case law that chatGPT made up out of whole cloth and nobody checked it, nobody went back to say, oh, does this case actually exist?

I do think, even as the technology gets better, we're still going to need the human in the loop for a lot of that, just to make sure that the tone is right. Although I do know of somebody, and I thought this was an interesting use case as well. He writes a lot of blogs. What he did was he instantiated one of the large language models with his writings. Now it basically answers in the same way that this individual would answer, and of course, that can apply to organizations as well, where you can get that tone of voice, if you will, by saying hone in on this aspect of the corpus.

Andreas: That's an interesting addition, to make sure that you fine-tune the model to make it even harder to distinguish that this is coming from an AI. It's not only coming from human, but it even sounds like you. It starts to mimic you.

Birgitta: I've been using GitHub Copilot actually for writing because it's like I get suggestions as I type. None of the other tools for writing assistance do that at the moment, at least not that I know of. I've been using it, for example, when I wrote some of the blips for the last technology radar, just getting suggestions and then editing always and double checking, of course.

Andreas: Wow, you're saying you're not using it for coding, you're using it for writing normal text?

Birgitta: Yes, for both. In a markdown file. For articles, blips, and so on.

Andreas: Wonderful. I liked the saying that Rebecca just made on the terror of the blank sheet. That reminds me so much of when I've been a student and I've been sitting in front of my diploma thesis, there's this blank sheet of paper. I know this feeling very well. Coming to the last cluster, and I think that's something that is very, very known already so probably doesn't need a lot of explanation. This is all about engineering and software development.

It's about automating the creation of documentation, or even code, or writing tools, or developing tools that help to detect box or security vulnerabilities that allows developers to focus on the higher level design and on problem-solving tasks, and accelerates the entire development process. Of course, we being Bosch Digital and we, as I've said before, in all our sectors that we are active in relying more and more on software, that's a very, very powerful thing.

We all know these studies that are being out there. There's this one study from GitHub that has been quoted very often saying, hey, if you use GitHub Copilot, you can increase our productivity I think by 55%. That's the number they've been using. What I found very funny is, when you look at a study in more detail, what they measured is 55% on productivity increase, but when they ask people what they think how more productive they were, they said a number that was way beyond that—80% or 88%, which shows that people are not only more productive they are also happier because they do not need to waste so much time searching for the information that they need to do the actual job they are supposed to do. That is something that I found way more interesting when reading through the study.

Birgitta: By the way, this 55% that was in their study that they did, there was one task they gave to people which was to create an HTTP server in JavaScript, and this is what that is based on as one example. It depends on the situation, like how much faster you get, or on lots of factors.

Andreas: Yes, thanks for adding that, Birgitta. I just want to make a similar comment. It's just that there are not so many studies out there, and I would be a little bit careful with this 55%. If you start Googling around and trying to find a number for a real-world project that is running since a while already, with really a diverse set of tasks and not this one example like you've just explained, you don't find a lot. We have been looking for that the last few weeks. Let's see if we do another podcast in two years, maybe we will know a little bit more what do we hear about that you can expect actually is right.

Birgitta: We just talked a lot about exploring the problem space, like collecting the use cases, and what types of use cases are there and so on. What about the solution? How do you, at Bosch, approach that? Buy versus build, how to roll things out, what are your experiences there in the space at this point?

Andreas: I think I touched a little bit on that. We have the center of excellence, I called it. We really have this interdisciplinary staff team where people from legal, we have people from purchasing but also the IT guys are in.

We bring together people with the digital knowledge and that are experts in fields of AI and software development with the people from the business units that are our customers that have the domain knowledge, they know everything about the fridges, the cars and so forth. Then we bring them together and the to-do things.

First we as Bosch Digital provide the right technology. That can go into two directions. It can be that we onboard tooling that they need. For example, maybe it's a very simple problem — they want to use generative AI to generate video material, for example. Maybe we then just onboard a tool like Synthesia, for example, because this is able to do that. Then, of course, we need Purchasing and Legal to see how to do that the best possible way.

We try to do that in a very coordinated way. What we do not want to end up with is 1 million tools for the very same problem. We do not want to end up with a slew of tools for the same use case. That's the one thing that could happen if we just onboard tools. The second thing that can happen is that we provide them with a technology stack, where we, in-house, in our own data center provide the entire stack that you need maybe up to the point where we have our own foundation model or large language model or whatsoever.

This is not the standard case. This is something we usually do. If the data we need to train is IP-sensitive, then we go that path. If not, we usually use the technology being out there being provided by the big players. At the very end, so onboarding tooling, providing the right tech stack be it on the public cloud or be it in a private cloud environment, and the last thing that we do is we assist them during development. We really have our people with their software experience coming in, and jointly with the domain experts, we really coach the problem and try to come jointly to a solution. That's how we work.

Rebecca: We've talked a lot about the interdisciplinary, cross-functional nature of the teams, but how is this all being received within Bosch? Are people excited to use it because it's going to make their job easier? Are they scared of it? What's your sense, so far, in how the people within your organization are responding to the potential of GenAI?

Andreas: We see a lot of excitement. We see that the interest is really super high. Almost everybody in all the different parts of the company is starting to do something. Even trying to take over or claim ownership for the topic, even that happens here and there. That has a risk, of course. That has the risk of doing a lot of duplicate work, and even things a little bit quite risky, because it's being done maybe without Purchasing or Legal being part of what is going on.

What we think or what I think is very important, and that people are enthusiastic and want to try things out it's something I like very much but still, we need to be careful because of all the risks that are associated with the technology, like the legal risks, the ethical risks, the commercial risks, and all of that. Some governance is being required to get everything that is currently happening under control.

The point here is that we need to find a good middle pass so that you allow individuals to have some freedom, while at the same time, not ending up with a slew of developments and with a lot of redundant work and with a lot of duplicate work being done. Then the risk is that things are reinvented over, over again leading to massive inefficiencies. The goal must be really to allow for a reasonable degree of, I would say standardization and homogenization, and especially reuse without ignoring the individual needs of different units and forcing them into borders. Maybe that just describes it.

Birgitta: I also think this is a space that requires a lot of exploration, and so you also need to keep the barrier to experimentation low so that you can actually get people to play around with this in their context, but in a safe way and in a way that doesn't create too much waste. The typical challenge of innovation and experimentation in a good way.

Andreas: Absolutely. Now you're touching on what I wanted to mention as well. The Center of Excellence has, on the one-hand side, the task they already explained. Helping all our business units to develop the solutions that are Gen AI-powered, but they also have some additional tasks. One is about really identifying the risks, really identifying the potential mitigation options, and really informing the rest of the organization, so there's also the educational task behind, about these risks and providing the rest of the organization also with training material and with guardrails and giving them, or providing them with a kind of certainty and security what to do and what not to do.

This is something very, very important because, especially we at Bosch, we are 420,000 people, I think, but not all of the 420,000 people are IT experts. We need to make sure that everybody being part of our organization understands the potential, but also the risks, and is being assisted a little bit and not left alone to understand what to do and what not to do. We are really working hard on developing training material. We do a lot of publishing of articles in our internet.

We have so-called sofa sessions, which is our internal YouTube where I'm sitting on the sofa, that's why it's called sofa session, and try to explain things to everybody being part of our Bosch family. That also helps to address something else, namely fears.

Even so, there are people that are quite enthusiastic, and I think that's maturity. There are also people having fears. I think most of them, they have understood that this year is not about the launch of Skynet, and it's not about achieving technical singularity, but some are, for example, worried about, "Is this replacing my job?" We know it's affecting more than just IT. We have talked about the examples. That it's affecting people in marketing, it's affecting people in HR.

I think the important message, and that's also part of this educational thing that we are after, is generative AI should be seen as an enabling technology or commending human capabilities rather than displacing them. We have always experienced human history. We have always seen new technologies automating tasks to free up time for other higher-value mental work, let's call it like this. What is key, in my opinion, is that people learn to exploit the power of this technology in a meaningful and responsible way because if you don't do this, you will fall behind from a competition point of view.

The analogy I use very often is think of two doctors-- when I'm asked this question. There might be one doctor that is not using, in the future, anything related to AI, and it might be one that is using this technology. The second one might be able to find the best diagnosis and the best treatment much quicker because he can use the technology and search through all these databases the way I've described it with our internal search a couple of minutes back. He will be more competitive. He has an advantage, a competitive advantage over the first doctor.

It's essential, if you want to survive with your business, that you really learn how to deal with this technology in a responsible way. That means, for Bosch, for example, we are working on our AI codex where we really define the rules of the game, what to do, what not to do. We even provide our employees with protection mechanisms.

We have a tool that is called AI Shield that really checks the prompt before it's being routed to an external system, to see if we really want to let that go out. Not to control our employees, but to protect them from ending up with a case that we have all seen on the press from another company a couple of months back where some engineers thought it's a good idea to have IP-sensitive data in the prompt. That's things that we do.

We see that even on the political level currently. Out of Brussels, we see all these thoughts around the AI Act being driven forward because we all see that there are risks associated, for instance.

Rebecca: At the beginning, I mentioned the range of opinions between the world ends tomorrow and the tech utopian dream. Where do you personally sit on that? You've mentioned that you don't think Skynet's going to happen tomorrow. What you just said tells me that you are seeing this as a valuable way that we can enhance the human experience as opposed to degrade it. Is that a fair characterization?

Andreas: Yes, absolutely. If we use it responsibly and if we use it right. I already touched on that a couple of minutes back. They are also use cases where I really would encourage people to not use it. For example, if you think of a traditional classification task, there you wouldn't use generative AI, or when you think of non-generative task, where it's not about generating new content based on existing one, there you better use our reinforcement learning or traditional machine learning. Generative models do not make sense.

Maybe you want to look at these things from a cost point of view. Training these models and running them, at the moment, this requires a lot of computational power, so it's quite expensive. If you want to have real-time or low-latency applications, well then again, it's probably better to use one of the more traditional technologies. Maybe you don't have a lot of data.

Generative AI requires a lot of data. Look at ChatGPT and large language models, it has been drained on the entire text corpus of the internet, actually. If you don't have enough data, you can't use this technology.

We already touched on things like trading instability. The system can change the behavior, and if it's changing behavior, it's also not suited for safety-critical systems. Still, I think generative AI will continue to accelerate in the context and in the scenarios and in the examples that we touched on a couple of minutes back. I think it'll accelerate even more because I see what I call the democratization of AI will continue.

What do I mean by that? Democratization of AI, for me, refers to that the AI technology becomes accessible and affordable to a wide range of people, not necessarily being experts. This happens because there's a simplification of AI tooling and everything you need. A couple of years back, you really needed to know everything about the algorithms behind and so forth. Meanwhile, you can download a model from Hugging Face. You can deploy it on some AI platform. That's it.

The second thing is, which will even make it accelerate is, I'm very convinced that the compute power being required to train and run will significantly decrease over the next years. Then even more and more open source models will come up, which currently is, due to the associated cost, not happening so much. I think it will be interesting, and probably we won't find the answer to that one today, is I'm very curious what will happen the next months and years when looking at this ongoing war between the really large channel purpose and the specialized small models. Who will win? That's a very interesting question, in my opinion. Let's see in one or two years from now.

The second thing that I'm very convinced of here, which also needs to be proven, is from an end-user perspective, I am totally convinced that generative AI will change IT systems as we know them today. I'm totally convinced that every system or tool will come with a natural language interface that allows you to verbally describe your problem or the task to be performed. This is what we already see in GitHub Copilot, but I'm talking about every tool.

Think of Adobe Photoshop. You will just write, "Can you please edit the photo that I've currently loaded like the following? Can you please exchange the sky and make it a blue sky instead of the gray sky I'm currently seeing?" The tool will understand what you're talking about and will do it. This will happen when you use PowerPoint. You just write some text and say, "Can you create me some slides that deal with the following topic?" Even though it might not be perfect, you don't have the terror of the blank sheet as you call it before. You have something to start with and that will help.

These trends, I think, will continue. The democratization, the less compute power being needed will accelerate this technology, and we will see this war between the large language models, the small ones, and we will see how, at the very end, this will cause a huge change for how IT systems work and how they are being made accessible for non-IT experts by using this technology.

Birgitta: This democratization and having this natural language interface and all applications, that will mean great power for the users. With great power, comes great responsibility, right? Technology is not just happening to us, we can decide as users and as people how we're going to put up guardrails and what we're using it for and whatnot. We'll have to see how well we will deal with this responsibility for the good effects or the bad ones or somewhere in between.

Andreas: Yes, absolutely. What I think is not a good idea because this has never worked out in human history is this-- I'm not sure what you think about that but this idea of stop research, right? Let's not continue with that. Someone will do it. It's super naive, right? When I've been asked what I think about that, I usually use an analogy saying, it's like the Tour de France, like the big cycling event taking place in France every year. Everybody uses stuff he or she not yet should better not use because they're all afraid somebody will use it so I also need to use it right because otherwise, I'm not competitive anymore.

The same thinking will probably happen when you look at this technology. Somebody will do it because somebody will think "If I don't do it, the others may do it so I better do it. Maybe I do it in my garage. Nobody finding out." I think it's very naive to say, "Okay, let's better stop research on that until we exactly know how it works, how we can control it, and for what is allowed to be used or not to be used." This is the idea I personally find very naive.

Rebecca: Well, that answers the next question I was going to ask you, which was what your position on that is, so thank you. Well, Andreas, this has been a fascinating conversation. It's great to hear how this is actually playing out with people trying to run a business and take advantage of this technology. Thank you so much for joining us today. We'll make a note to come back in a year or so and say, "Okay, how did these things turn out?" Because I do think we are going to see a lot of advances.

We're going to see a lot of uses that we haven't really even thought of yet because the barrier to interacting with this AI is so low that democratization, I think, is going to be a source of tremendous innovation as people who are not traditional technologists, traditional computer scientists, are looking at this technology and, "I would see. This is what I can do with it." Thank you, Andreas. Thank you, Birgitta. I hope our audience has enjoyed this exploration of generative AI within a context.

Andreas: Thanks a lot for the invitation.