AIAW Podcast
AIAW Podcast
E166 - AI Agents at the Government Offices of Sweden - Magnus Enzell & Peter Nordström
In Episode 166 of the AIAW Podcast, we sit down with Magnus Enzell and Peter Nordström from the Government Offices of Sweden to explore one of the most forward-thinking public sector AI initiatives in Europe. Together, they’ve helped launch over 30 AI agents inside Sweden’s central government—digital assistants designed to support civil servants with document handling, data processing, and smarter decision-making.
We discuss the real-world impact of these agents, the lessons learned from deploying AI in a policy-heavy environment, and the balance between innovation, ethics, and trust in the public sector. From boosting daily efficiency to reshaping the future of public service work, this episode offers a rare look behind the scenes of Sweden’s digital transformation strategy.
What happens when governments start building their own AI agents? Tune in to find out.
Follow us on youtube: https://www.youtube.com/@aiawpodcast
So 1634. So it's it's it's a strange mix of what we're doing. We're trying to apply today's technology to yesterday's reality.
SPEAKER_04:Oh, you mean you have a core business that's rather traditional in some sense, but very traditional. But then you apply really a novel technology on top of it. Yes. While still keeping the traditional kind of workflows that you're having.
SPEAKER_06:And also the traditional values and the integrity of the advisors and all of that. And then you're putting something really interesting on top of that. And that makes a strange mix.
SPEAKER_03:But it's a strange mix, but it's also a very important mix. Because somewhere here we need to start safe. So in a way, we understand long-term, it will reform the fundamental view of how regulation is done, how policy is made. That whole workflow can of course change, right? But where we are, it needs to sort of fit for this to make sense and to be correct.
SPEAKER_04:But getting back to the topic then, I mean, for one, it was fun to see, and I'm impressed, and I know um the government offices of Sweden here is very innovative, and it's glad to hear that you actually are hearing similar kinds of experiences when speaking to other agencies that we have in Sweden. But can you give some example perhaps, you know, when speaking to other agencies? What do they say if they compare what they are doing to what you're doing?
SPEAKER_00:So most of the times we would talk to other agencies, uh, it would be from a uh kind of generative AI point of view. Yeah. So what Magnus is saying, just know us going back to the 1600s, we could probably do what we did then without technology. I mean, we could do our job without computers or AI. Exactly. So we don't have any systems like you know, the tax agencies has the list of all the people living in Sweden, and we don't really have any services towards the citizens or the or the corporations in Sweden. Uh so a lot of what we're doing right now is in in generative AI. And when talking to other agencies, uh, I mean we want to learn from them because we started with this in in in the spring. Uh and uh we uh we found out that uh the kind of scale that we are testing and running it at is uh pretty fast compared to to others. So most of the times when we're you know reaching out, we want to hear what Pahuns Mindihatan or someone else is doing, they quickly want to know what we do, and that the conversation has a quick turnaround.
SPEAKER_03:And and if you try to contrast sort of the angle you're working on and the angle that usually what you met other people working on, has it is it a difference here that you've been going in the uh agentic or agent route or assistant route and then unpacking the problem from that angle and they coming more from IT or data, or what is the contrasting difference? Because I some of these, as you assume, they are great or maybe better at some of the other parts of this puzzle. So what's the what what have they been focusing on you different differently?
SPEAKER_06:IT, if anything, we've put the core business in front. They are driving the change. Very good. It's not IT that it's not a tech project. It's what what we find is uh cognitively mature civil servants in a very knowledge-intensive environment longing for this new tool. So suddenly when you when they get uh their hands on it, they take off.
SPEAKER_03:Uh and we sort of whoa trying to uh so this is very important when you're trying to push a technology from underneath that they don't even understand. Here we have the opposite scenario, really. Yes.
SPEAKER_06:Uh and it's really interesting environment. Um, Rianis Castle in that sense, because it's it's packed with uh intelligent, um, hardworking and loyal um people that is working under hard pressure, um, and they really need to deliver. So as soon as they see uh a shortcut in a way to do their job better and to deliver more, they will go for it.
SPEAKER_04:I mean it's nice, and it's also nice that even though you have very traditional um workflows, so to speak, and I'm sure you, as a history expert as well, will describe more what that really means, you can still keep the traditional workflows but simply empower them with technology or AI in this case, right? And that still makes a lot of value for the business, so to speak.
SPEAKER_06:Yes, it's it's part of the principles that we're working for, putting the human in the center. Yeah, and what we're doing is we're cognitively enhancing the civil servants to build a wiser state uh or a wiser government. So actually, the the focus is on other people to making them do better in what they're already doing. So we're not primarily changing processes. That's where that comes later.
SPEAKER_03:That comes later. Yes, but I think this is very important, and and um we can come into this rabbit hole later on because I think, especially when we go the agentic route or agent route, immediately you come into the fundamentals of topics like agency, you know. And uh the in a in a large enterprise where I have done a lot of work, that that is probably typically a problem that we have nice workflows that look good on paper, but it but as soon as it's unclear it was who's doing what or what teams should be doing what to serving who, you have a problem. So maybe this whole view of starting from the team perspective and what the servants within the team does, and and and how we build from there with agents in mind, is maybe a key recipe. This is something I want to explore today.
SPEAKER_04:Well, I certainly want to learn more how you did it and why other agencies may not be as successful as you have in making AI actually create value, which it sounds like it has. So very eager to get into that rabbit hole and learning more from you in this podcast. But first, very welcome to the AI After Work podcast, Peter Nustram and Magnus and Cell. Thank you. And if we start with you, uh Peter, um, I understand you work in a digitalization office, or what's the name of it in English? Did it digitalization strategy, or what would we uh reorganized uh last uh fall?
SPEAKER_00:Uh so uh um we're we're part of the uh it's the old IT department that has had a facelift, so we're now the digitalization uh department. So we work for for the government and serving uh all our uh offices, yeah.
SPEAKER_03:And uh broad scope from from these you know very cutting-edge uh agents we're gonna talk about all, but also more practical practical emails and networks and all that kind of stuff, or have you separated that?
SPEAKER_00:No, we do we do it all. Uh we we service and uh I was so impressed when I started because I came from the tax agency, where you're you know 10,000 employees with 2,000 people working in the IT department, and then I came here to the government, and we're you know roughly six and a half thousand people, uh both in Stockholm and uh in 100 countries with our embassies, right? And you know, we're under 300 people in the IT department, but everything is working, you know, the telephones, the emails, all the systems are online. So impressed by how how well you know we've been able to service everything.
SPEAKER_04:So it's not very efficient. It's really interesting that you came from the tax authority as well, because they are also very good at digitalization and also starting to use AI more and more in a really good way. But before we go into that and your current role in the government offices in Sweden, perhaps you can just speak a bit about who is really Peter Nström.
SPEAKER_00:Oh so I sent an email to uh to uh my new team yesterday, so I changed my my role and I was thinking about what would I what would I what would I say? And uh I guess uh uh I came up with uh uh I'm a nerd, so I love to play video games. Uh I still play the old ones, Quake 3, Warcraft 3. Uh love esports, love movies. Uh, but at the same time, uh I play golf. Uh I hang out with my uh my friends. Um, but most most of all, family man, you know, hang out with the kids.
SPEAKER_03:And how many dreams can Sliet knows you have a background in stand-up comedy?
SPEAKER_00:Uh not many at all. I think only one that's now knows it. Too few. But that was uh a whole lifetime ago. Like like AI lifetime, like last year, or you mean normal lifetime? No, normal lifetime. Yeah. Um I think it was 15 years ago when last was on stage doing stand-up. Which one was your biggest gig? Uh so it was uh I did something called Bungie Comedy, which was like uh the the idol, but for uh stand-up, so it will be you know uh tryouts, uh semi-finals, and finals, and I actually made it all the way. Uh, and really good comics came from there, Messiah Halby, Messiah came that way, yeah. Yeah, cool.
SPEAKER_04:I mean, I think it's impressive being a stand-up comedian is actually very difficult, and I think that's so valuable to have with the timing and the way you phrase it. Um, even you know, we are very much inspired by your your Rogan podcast, and he is a stand-up comedian as well, and I think he really had a lot of benefits from that.
SPEAKER_03:But I I need to uh take it back to one of your key rants on what is the most important skill for a great engineer, and then communication skills is maybe one of the highest because it doesn't matter how good you are if you can't convey what you're doing, and you are you you you need to put the punchline where it belongs when you do stand-up comedy. So I think it's a fairly cool uh connection here, yeah.
SPEAKER_00:So it's so hard though. I just want to say it's so hard because it's one thing, you know, going on on the stage and and delivering the jokes, uh, but writing them is so tough. People find this funny. I find it funny. You know, you'll try it out on a few friends, they think it's funny, and then you go on stage. And I told Magnus it it sucks to deliver your best joke, and a hundred people are silent. It's so quiet.
SPEAKER_04:How good do you think AI isn't writing jokes these days?
SPEAKER_00:Uh the little I have tried, not great, uh, but it's really good to get you started.
SPEAKER_04:Cool. And if you were to just introduce a bit more about the current role in um the government offices, what do you do?
SPEAKER_00:So uh I'm the head of a unit where we have uh architecture, high security, and uh innovation. Uh and it's the innovation part I've been working on since I joined in in uh January. Uh so we've uh we kind of uh act as an incubator where we take uh new ideas, new projects, and and uh try them out. Uh, if it works, we'll you know put it into the ordinary kind of uh line of business. If it doesn't, just kill your darling and go on to the next. So it's really cool to have like an incubator in a government office.
SPEAKER_04:Yeah, it's very cool.
SPEAKER_00:Did you call it incubator? What's the Swedish name for it? Oh, yeah, this is gonna be a linguistic challenge with all these kinds of translations, and I'm so impressed by your English, Magnus. You're so good. Uh but yeah, we actually call it an uh an incubator. Incubator in Swedish.
SPEAKER_03:Yeah, or you're cooler than that and uses the English terminology. We don't have many English words, no. That's maybe one thing that is interesting. Like in in government and stuff like that, then you have your Swedish or your your your your native tongue word for things. When you come to Scania, they they don't care, they they use the terminology like you know, like like in tech, right? It's so obvious.
SPEAKER_00:Can I can I tell you something really cool? We do have one thing with an English uh name, uh uh and we call it quick wins. The idea was that you could take something to the IT department that's you know uh everyday struggle, and it's something that we can fix in three weeks. For example, why is not everybody presenter in Skype for business? We can fix that immediately. So, in with that, we fix it. But when we went out to the uh, you know, and and meeting with the um like how do you say the the ordinary, like the vaccine. Yeah, the civil servant. The core business. The core business. We went out to the core business and you know asked them what are you kind of struggling with to identify quick wins. And we told them it's something that needs to be able to be quick quickly fixed. I mean three weeks. And the first things they kind of listed was, yeah, how to reply on all of all my emails. So the kind of expectations from the from the line of business on what we can do quick is so big. Okay, interesting.
SPEAKER_04:Cool, and Magnus Ensel, very welcome here, you as well. Thank you. And uh please describe your role, your change manager is that the proper way to describe it, or what yes.
SPEAKER_06:Normally, though, I work at Minister of Finance in the line, okay. Um, in the department called um digitize digitalization of the of the um um public sector. So I've been doing public policy for nearly 20 years um within the field. Um, but now I'm then um working 80% as a change manager for this project uh with a particular focus on AI, of introducing um AI agents or assistants, I should say.
SPEAKER_03:How did you come into that role? Because I think it's brilliant that you have you have the background, you have done a lot of different things in was it 2006? Yes. Yes, and and now you change manager in AI.
SPEAKER_06:Yes. I came in under under Spori and Mazzadel uh during the finance crisis. So I've been working there for for a long time. Um I was brought in to the to the reference group actually to the digitalization strategy and me and other couple of uh other um uh civil servants sort of we were critical to the early on um uh strategy because of very tech oriented. So we managed to sort of turn it around to make it more business focused. And I think maybe there um was sort of the natural step in for me to uh to um I I'm just curious how you got asked the question or how how how the how it was framed and pitched to you.
SPEAKER_03:You need to become a change manager in AI, or did it evolve?
SPEAKER_06:How did I No, it was actually um uh an offer I got from the from the director of digitalization. Um and uh they asked me, do you want to come in and do this?
SPEAKER_03:So you have some sort of solid and dotted line to the digitalization director and to your your home department, or how does it work?
SPEAKER_06:Um Well, the uh um I was part of the reference group, so we got to know each other. Uh and and um there was the question. And and from sort of a the formal point of view, I'm uh I'm I'm alone. Uh so I'm I'm on loan from the line to the section, the innovation section uh the other Peter.
SPEAKER_03:And this is my real question. How how how did you solve it? Because typically the to bridge tech and business and make them speak the same language and actually innovate together is hard in many organizations and practically.
SPEAKER_04:Perhaps you should need to add that to as a topic to the list of great topics. But before we continue too much, um I'd love to hear a bit more about your personal background as well. I know you have a PhD in political science as well. Perhaps you can just elaborate a bit more on what you actually researched at that time.
SPEAKER_06:Um I was doing uh constitutional history. Um actually with What does that mean? Constitutional history. Well, we have a constitution in Sweden, as America has. It's just that our constitution is less known and its status and authority in the Swedish political culture is a bit unsure. Um so I was investigating how that come came to be because during the 19th century it had an equally strong role as it had in as it has in America. So something changed. And I was researching how that change came about between 1907 and 1917. So there was a shift in the Swedish political culture, and that was my dissertation. But it's basically uh um a dissertation on conceptual history. I I consider myself as being a conceptual archaeologian. Uh I I dig up old concepts uh to see how they uh come about and how they work today.
SPEAKER_04:I'm so eager to go to hear what you think about the Cardinal Trump administration, but perhaps you shouldn't. Don't take me there.
SPEAKER_06:But are you a history buff?
SPEAKER_03:Have you always been you've always been a history buff? Um since kid or like interested in history?
SPEAKER_06:If Peter is a stand-up comedian, I'm the philosopher. So uh I'm more of a sort of philosophical take on things. Um, and part of that is being historically sensitive to whatever we're trying to understand.
SPEAKER_03:Know your past.
SPEAKER_06:Yes, past-dependent history making.
SPEAKER_04:I mean it's so impressive to hear, and I think the the role you have is so important and difficult to do. But I would love to jump into the meat, so to speak, of the discussion and the theme of the podcast as well to hear more about you know the different types of digital assistants that you have in the government offices. And if I understand correctly, you are having more like 30 of them, or how many exist today?
SPEAKER_00:Yeah, I think uh yeah, at least 30. Um we have 30 in in the spring, uh, and uh we've killed some of them that hasn't worked, and and some that has worked, we we've uh built upon. Uh, but we're making it pretty easy to create new ones and try them out. Um, we're also pretty quick to identify the ones that don't work or the ones that we shouldn't do because of you know uh high risk according to AI Act or something like that.
SPEAKER_03:Let me put the signpost right there. This needs to be a rabbit hole on its own. Park that question how to rapid prototype and get going with stuff and what works, you scale, and what doesn't work doesn't scale. Simple to say, really hard for most companies to make happen. Yeah. Let's let's really dig into how you did that.
SPEAKER_04:Because you do fail fast. I think that's it.
SPEAKER_03:Fail fast. The whole uh fail we're gonna be the best of failing fast. Spotify, yeah. Spotify, right? This is so cool stuff, right? And and uh so hard. And and I I want to know the let's park it, but later on, I really want to know the mechanics to make that work. Sure.
SPEAKER_04:And I'm thinking about it. I'd love to hear the story, how you get started with the whole digital assistant project, etc. But perhaps we can start by starting to speak about the specific use case. If you were to pick one assistant that we can speak about, and you just try to describe how it works a bit, and then perhaps we can go back a bit about how you got started in general, so to speak, with all the work of the different assistants. Is that someone that we can use as just a good example of how it's easier to follow?
SPEAKER_00:I I think I can give the background and you'll tell the example. Sure. So the first week I was at the uh Rearingskuncleet, there was this uh health day that HR uh put together. A help, sorry, what? A health day. Uh health, yeah, health, exactly. Yeah. So we had line dancing, we had uh yoga and different you know activities throughout the day. But we didn't have anything from you know the digitalization department. So I sent in, I want to do three walks, and I want to walk around Gamaston, and I'll, you know, everyone can come and I'll tell the one we're gonna do and you know, listen in. It was a total fiasco. Uh, the first walk, nobody came, and I had prepared so much. You know, uh, the first internet cafe in in Sweden was in Gamaston, and you know, did all this trivia. The second walk, nobody came, and I started to feel pretty down. But on the third walk, I met uh this guy called Birrier, uh, and uh we started talking, and he wanted to uh become a local admin on his computer so he could install Python because he had an idea of uh using AI and building an assistant that could put together uh loads of information to create what we think is gonna be our killer app, and I'll let you tell about the killer app, but that was that was the background, and also what's really cool Barrier is in our team. So I asked him, like, how long ago come come and work for us? This was in January, the first week. Really? Yeah. So now he's part of the team, and uh, that's kind of been you know the way we have been working. We find someone out there that has brilliant ideas that is wicked smart, and everybody is super smart. And then we asked them, Do you want to come and help us out? And uh most of the times they say, easy.
SPEAKER_04:Um impressive that you've done so much in such a short period period of time. That's really impressive. But perhaps, yeah, it could be nice to just have some concrete examples still to what is it?
SPEAKER_06:What's the killer app? Well, it's uh the company analyst. Company analyst.
SPEAKER_04:Okay.
SPEAKER_06:Not many people know this, but government is actually a big owner of companies. So they have a portfolio of companies, right? El Coabi and Post Noble. Who worked at Wattenfall? Yes, Wattenfall, all of those. So it's actually 40 companies and they turn around 850 billion kroners and they employ 122,000 people. So it's quite important that government is a responsible and forward-looking um uh administrator of that company portfolio, right? To be able to do that, you need to analyze a whole bunch of company information. So there's a whole department doing this, and what they were doing was downloading the the uh reports, reading the reports, sitting down, taking notes, and that took 24 weeks. Exactly. Um to come in a year, but right? 24 weeks. 24 weeks to do sort of an analytical cycle of those 40 companies. Okay, and they have different dimensions that they have to do to to to uh to go through, whether it's sustainable and whether you have human rights or or whether it's efficient or what have you. Yeah. So what they did was to make the analytical uh assistant uh of companies. So they prompt an assistant and downloaded the reports and uploaded the reports into the assistant, and they prompted it. And the prompt, when we look at it, is I think it's six and a half A4 pages long.
SPEAKER_03:Questions that you want to dissect from the data.
SPEAKER_06:It's it's actually their uh analytical and judgment framework that they used manually that they put into the system workflow. It's the same workflow, but they put it in the system prompt. And now they're doing the same analytical work with higher quality in six weeks.
SPEAKER_03:From twenty four weeks rather than six weeks. 24 to six weeks for a department. Yes. And oh, there's so many error, there's so many angles here, but um where to start even.
SPEAKER_04:I'm using uh perhaps you can just okay, how how how does it work underneath? I mean to start there. Um I'm a more of a techie person as well, so I love it.
SPEAKER_03:But it's interesting to understand concretely what what are the mechanics is it a rag? Could you describe it a bit more perhaps?
SPEAKER_06:Sure. Um so we're using a platform called Intric. Um, and you can build AI assistance in those. And then AI assistance is basic basically a package of information and prompts and um data that you pack together and send off to the language model to get your response back. And part of that is um a rag uh database with the knowledge that you want to be part of your uh analytical stack, whatever you want to call it. So when you when you build this assistant, you first take sure make sure that you structure your knowledge source sources, upload them into the assistant, and then prompt the assistant to do what you want. And what a lot of things what we've been doing is teaching the civil servants the craft of prompting. Right. So even if we change platform, thou shalt know thy prompting. Okay. Um, because it's it's you need to know how a rag database works, um, and then you know when to trust it and not to trust it.
SPEAKER_04:So you're saying that for a first-time user that have no idea how to work it, they will not be able to find value as easily unless they're trained for it.
SPEAKER_06:For simple cases, you will definitely find value. For more complicated uh cases like the company analyst, um, you need to be able to prompt it properly to get a proper answer.
SPEAKER_03:It makes sense, but you are essentially dealing with the system is the data part, is the integration part, all this stuff. So if you if you don't have any awareness of that, the system that you can deal with as a newbie is very simple. But the more complicated.
SPEAKER_04:So, how does it work then? Can you go a bit more? So, in you use Intrigue, did you buy that recently or did you already have that uh available in the government offices?
SPEAKER_00:Or no, we uh we uh purchased it before the the summer, so uh we did some tests with with other vendors, and we've been trying to stay as you know, we want to be language agnostic all the way. Uh so uh we what we enjoy with this platform is we can choose between uh language models from Anthropic or OpenAI. We also have the the open source models which are hosted here in Sweden, uh so Geometry and Lama, etc. Uh, which is really cool because that makes uh puts us in a position where we can start to see if we can handle more sensitive information depending on what model we choose and and uh where it's stored. Uh, but it's uh it's still you know, we're testing. Uh and even though we're testing it at big scale and we want to keep testing at an even bigger scale. What keeps me up at night is doing you know a huge investment in something that's gonna be completely outdated in one year. Uh so uh we want to use cloud services as as far as it's possible. Uh one, we don't really have the knowledge how to handle the GPUs in in in-house. Uh we also don't really have the personnel to do it. Uh so anything that's you know easily accessible, uh easy to to try and and get value from, that's where we want to be.
SPEAKER_04:And you use cloud services now? So you're running in trick on some public cloud, or is it some on-prem solution that you're having?
SPEAKER_00:No, it's it's uh it's a cloud service.
SPEAKER_04:Okay, okay, good.
SPEAKER_03:Was that tricky to get through? Secure device and all the things you know. We we keep hearing the public sector is in pain over cloud. And I think it's about knowledge and understanding of how to do it safely. And if you know what you're doing, it's it's safer, in my opinion, to in many degrees, in many ways. Uh so, but how did you manage that?
SPEAKER_00:So we look at it from from all the different angles, you know, from uh IT security, information security, the legal aspects. Uh, and there are a lot of new questions here. Uh, and and it's so hard because it's there are no answers to it uh as of yet. I mean, the the the kind of system prompt Magnus is talking about, is that you know uh public information or not? So we have the the Swedish law cream, you know, the information being uh made available. Uh so there's so many hard questions here. Uh but the approach in you know where we are testing and trying, it's not just uh technology, it's everything you know around it as well.
SPEAKER_03:You're testing and trying not only the take, but actually the legal boundaries, the regulatory policy, you know, all this. And I think this is this is one of those signposts that I was looking for. I'm gonna I'm gonna put a rhetorical question where I have an opinion and an answer to. Can you even get close to something meaningful in terms of results in answers to these questions without trying and testing and piloting? Can you get close to anything meaningful in in PowerPoints?
SPEAKER_06:I I I think we we've noticed this, and and I think this has been one of the um uh sort of most important results of the first pilot phase that we did was that when we started, uh the questions you could ask were so general that you can you couldn't get a straight answer. Oh, you frame it so well. Once we have the the company analyst in place. Place, you can start asking questions on the right level. And then you can get a yes or no, or maybe, or now maybe we should dive into this particular question and see where it takes us. So it needs to be a trial and error that takes you down to a level where you can ask the proper questions.
SPEAKER_03:And that to me is both in terms of the assistance working, but also in the bigger question on can we do it? Are we allowed to do it and all this? You cannot answer that until you've done it. And I mean, like it's a typical law perspective. Show me your data model and then I'll tell you if it's correct or not. Then you need to build it. Are you not allowed to build it? You know, so you need to go down this route safely, of course.
SPEAKER_00:But you have to be humble. And I mean, what we can do is make sure that the tech that we deliver is, you know, robust, it's good. We tried it, uh, we've done our pen tests, etc. But one of my favorite uh cases, uh, so we had um uh Erika, she uh got in touch and asked if she could get one license. I said sure. Uh, and and she wanted to try it, you know, just as a kind of chatbot. I think it was five or six weeks later. Uh, she uh told the story what she and the team had done. So each year, about 6,000 people called the government's office and they can you know call and ask uh what are you guys gonna do about you know uh the bank and and uh uh the high prices? When are you gonna ask them to lower it? Uh and uh they could ask uh you know what are you doing about the the the fishes in the in the sea or something like that. And at the end of the year, uh we uh take all these uh phone calls and we'll just analyze them and uh break them down, and this is uh manually done. Uh and uh she she said that she and her team had done it manually and then they did it with AI, and they came to the same results, so they stood on stage in front of the whole government and said, We're not gonna do it manually anymore. And we had no idea about them doing it, we just took part of the results, and that's pretty hard for an IT department to just okay, you do it.
SPEAKER_06:And and the point here, and I want to compliment that, was that um we're not trying to change processes, we're not trying to change uh forms. What we're trying to do is to cognitively enhance civil servants working inside these processes, and that makes things much better. Um, and also uh since we're working in the in the core business, I mean, for 400 years, we developed the practice of of uh knowing and judging how to use information. So it's it's an important part that the responsibility of how to use uh information, uh what kind of information that you uh can or are allowed to upload to an assistant, right? That responsibility is on the individual civil servant. And they're used to making those judgments because it's part of the of the normal practice of the government offices to work sense to work with sensitive information.
SPEAKER_04:And just to close the topic on the company analyst, I thought it was a great example there. Can you just elaborate a bit more? How did you get started? What was the first initial phases of how you tried to use it and uh how did you get value from it? If you just try to capture that in a nice way.
SPEAKER_06:I did not make it. It was Corinne Svensson that made it. Uh, she's actually the first data special specialist uh in in the government offices. We have one. But uh, and and uh she's part of uh the department that is responsible for all these uh companies. So I I couldn't tell you that story. I just know that the civil servant picked up this tool and built that agent, uh build that assistant. Uh she came back with no questions to us, I think.
SPEAKER_04:But you're giving access then to intrigue uh for public servants like her or civil servants like her, and they can then upload data and they can tune it and they can instruct it, and then they can let other people use it. Or can you, if you just elaborate a bit more how it works?
SPEAKER_06:Yes. Um, so building an assistant, you have to handle uh all that data and upload it and prompt it, and then and choose language model and say how deterministic or creative that interaction should be, and then basically you're done. And then you can start to interact with it as your own personal assistant. You can then, and you're working in sort of a working space that you can uh invite other people into it so they can start using that assistant as well. You can also uh create a widget so you can make sort of a window to that assistant for everybody else to work with. We haven't gone that far yet, so we're still keeping it very personal in the sense that who is using whose uh assistant. But what we're trying to do now in pilot phase two is to come up with a secure policy um of uh uh sharing those assistants, and and very important part of that is validating that assistant. So any other uh uh person that is going to use it, they need to know the uh level of precision that is uh uh that is coming out of the assistant, who owns the information and who owns this assistant. So there's a lot of policy behind sharing assistance. It's kind of easy to build your own and take responsibility for the information and the administration of it, but once you start sharing it, you need to uh become more organizationally aware of what you need to do.
SPEAKER_04:But the process, how does it work? So some person um starts to build some kind of agent and um then they add some data to it and prompts to it, but they can share it somehow, right?
SPEAKER_06:Yes, but we're restrictive with sharing as of now. You what what we've seen in the um company analyst is that people within that department is using that, but they're not sharing that with any other. Now that wouldn't make much sense because they're the only ones that is doing that kind of work.
SPEAKER_03:But you have other processes who don't look like that.
SPEAKER_06:Well, we have a lot of processes when it comes to uh uh analyzing government agencies. So we could take that assistant, turn it into an agency analyst, but that will be part two.
SPEAKER_03:And then you have more complication of how to share it. But let me see if I've followed you here, because I think there there we're there's an underlying story that uh uh we want to help you to convey here. And now you made a there was a principle in here. Agents, whoever we build it, ultimately needs to have very, very strong ownership within the department or the core business which is made. So the whole idea that an IT department can build agents for you or manage it for you is not the way you are looking at it at all. No. I fully agree with your point. Henrik Nieberg, who was on the pod last week, said the same thing. And but I think it's I'm not sure everybody's doing thinking hardcore like that, or it's like, oh, this is data and AI, oh, we send it to the CIO department to fix it for us. And you are saying the complete opposite, in my opinion, in terms of ownership.
SPEAKER_06:Yes, and and the core principle of any um sort of uh business operation within the public sector is responsibility principle. Someone needs to be pointed out to have the responsibility, as you say. So, what we've done in structuring the project is to have AI pilots, they develop uh the assistant, then it needs to be validated, and then you have an AI editor that actually um takes care of it.
SPEAKER_04:AI editor as an inhuman that does it, or what is it?
SPEAKER_06:Okay, so it needs to need to be human, that needs to be part of his portfolio. So his chief, his boss needs to tell him that you manage these assistants. I use sometimes use the metaphor of a stable. So you have these uh AI assistants that you when comes nightfall, you put them into the stable and you provide them with new data and you mock out the old data and to make sure that they're fresh to go work.
SPEAKER_03:You have the stable metaphor, Henrik Nieber used the intern metaphor for this to explain the same thing that, dude, if you have an intern, you kind of you need to coach them and you need to guide them and you need to tell them when they're doing right or wrong, and then you don't give the intern the hardest chore first, you let them grow into their role, but it's your job.
SPEAKER_06:Yes. And I think here is something that we speak about a lot that it's this is not something that the IT department can do or digitalization department, because what differentiates different assistants is the data more than the prompt. And you need to have a clear ownership of data, yeah, and that needs to be core business.
SPEAKER_03:And and the the logical, what's the role of digitization, and what is the role of the owner or the one wanting the assistant? I think Henrik Nieber makes it quite eloquent. I I'm testing you. This was last week. You can look at the podcast from last week. The digitalization agency, they have again technology competence and competence on how to prompt and technically what you think about, and then you need to co-create with the owner who knows the has the domain knowledge. So you this guy can't do it if they don't know anything about how to technically what it how it works in the beginning, and these guys can't do it because they don't know the problem. Together, you can set it up, and ultimately then the owner can think can write, can operate it.
SPEAKER_06:Yes.
SPEAKER_03:Is that a fair summary?
SPEAKER_06:Definitely. I was actually visiting Kungsbaka municipality uh to see what they were doing. Um, and they have um a nice metaphor, which they call the two-handed uh principle, and and it's a picture of uh steering wheel with two hands on it. So you have digital and you have core business, and they're driving this together. So I said, that's what I said exactly.
SPEAKER_04:So walk me through the process a bit more. So, for one, you have a large number of assistants now, potentially up to 30 or more. Um, and then someone is trying it out, I guess, on real data or real use cases and real workflows that you're having. What happens if someone finds that this is actually working rather well? Do you have a way to scale it up, to share it with more people, or what happens if someone finds out that this is actually a useful pilot?
SPEAKER_00:That's what we're gonna be working with now.
SPEAKER_06:This is part of pilot phase two and building that organization. Okay. So what we're setting up is um a center of excellence. Okay. Since we're having AI pilots, where we call it Top Gun, of course, where we bring in the best of the best, uh and they can then take the best assistance and turn them into uh government office-wide um assistance if you have general uh general functionality that could be shared but by others. And we will also have uh a capability uh council, which will then uh and that will be core business sitting together in the room discussing on what organizational level should that AA AI editor work? Should it be on the single department, on a minister level, or government office-wide, because that will differ uh and that will change over time. So we need to build an organization that already now sort of is capable of administering potentially hundreds of AI assistants in an orderly fashion with responsibility uh very clearly uh sort of as a target for the organization.
SPEAKER_03:And once again, try to think, try to cook up this in PowerPoint without having done anything, or try to sort, or try to start working, and then you see the then this problem leads to this scaling problem. This scaling problem leads to this organizational problem. This organizational problem leads to this policy problem. I I think it's so in impressive stuff. You're doing it the right way, I think.
SPEAKER_04:So one thing is to build like more like basic agents, we call it that. You can simply put the data and the prompts there and you can find value. But in some cases, I guess it's not sufficient. And uh perhaps you need to do some development around it to have the data pipelines working, or some additional UI, or who knows what, or actually integrate it into some current system that you're having so it doesn't have the normal kind of chatbot interface. Is that something you're doing as well, or are you mainly focusing on the like basic state chatbot uh interfaces right now?
SPEAKER_00:We do have now uh for one of our as we call them killer apps, we are gonna look at at more of a development project uh because it's something that's been done in in one department uh successfully, but it's something that we see that will work just as good for all the departments. Uh, and that's just gathering huge uh masses of information. Uh, and we want to find a way uh to to develop the whole approach of of validating uh because so the example I gave in in terms of of summarizing the the 6,000 calls, so you have um a situation where you've done the manual job and then you compare it to AI and you see that AI is is good enough, but you still have the manual work to compare it to. When they do it next year, they probably won't have the manual work because they're only gonna do it in AI. Like, how do we have the backbone of the validation process to because it needs to transcend with new language models? We don't really know how the new language models will compare to the old ones, so we need to find a framework of validating. So that's what we want to do now with one of our uh assistants to bring in a development team and see how we can scale it up and how can we uh find a validation process that will hold over time. Um, and it's also in part of finding these kinds of uh so we do have, I think at the moment, four or five different assistants that are all based on information from our uh internet. It would be very easy to just build an assistant to search the internet and you know, problem solved. Uh, but by breaking them down, we we get a better understanding from the people using the assistance. Uh, and that's kind of the PowerPoint approach. Like, we would like you know an AI assistant to be able to search all of the government's information. Like if I want to take a uh a vacation, what form should I use then? Or um I want to order a new computer, where do I turn? But by breaking them down, we we we get quicker tests and we know that we're not you know gaping for too much at once.
SPEAKER_04:But but uh I have a small rabbit hole, so it's continue, and then I want the thinking about the next step, then you know, and the most simplistic uh agent, if you can even call that an agent, is just an LLM that you ask questions and it can't really do anything more, it can't perhaps even reason properly, or it can't really take action or use tools in different ways. And and then you have you know the more advanced agent that are getting more and more autonomous and can reason for a number of steps and decides by itself now it's actually going to continue or stop doing something, and it's not controlled by some kind of manually written scaffolding around it, but actually it makes decisions by itself on what next steps to take. Is that something you're also getting into? Or uh how what are you thinking about that type of agents?
SPEAKER_06:Maybe the best example is how Intric have um uh integrated the databases of the Statistical Bureau and OCD or yes, SAB, yeah, and also Colada, which is municipal economical data. Um, and they're actually totally integrated into the platform so you can chat with the data.
SPEAKER_04:Oh, nice.
SPEAKER_06:Um, and when you chat with the assistant, it uh now produces uh a plan of progress so you can see how it plans its research and then you can change that plan. Otherwise, it goes off and dives into the database and comes down, comes back and then start analyzing that and then presents the uh the analysis for you.
SPEAKER_03:But uh let me I was on exactly the same question as Andes, but let me frame it a little bit differently because I really want to start uh almost like from a definition point of view. So I mean, like so you clearly now build 30 assistants, and and in a way now, in simplistic terms, how how do we contrast uh an assistant from an agent or something that is a gentic? And uh what you hear out there, and what I kind of subscribe to is the simple view of like if you are if you are having something that's sitting idle and reacting to your question and giving you back, this is great value, and you are you can sort of go through your flow and you can uh question it and you get the results. The next level when we start talking about is you know how do we give or should we give more and more agency to the system. So literally it has uh uh you you ask it questions, you don't prompt it all the way through, but it can actually go away and do things. And and now we get a little bit more techie because now we you might need to incorporate tool use and you know, whatever it has done, what happens next in the workflow, or we need to send that data to this server or to this system, etc. etc. So have you had this? So, question number one with that framing have you had the conversation on what is assistant versus agent and what is your ambition level? This is sort of more strategic, and then and then um what steps are you taking, you know, towards more agentic, I guess?
SPEAKER_06:Perhaps the most important part that is what we call assistant is cognitively enhancing people. Yeah, you're not changing agency. No. Um so we're now working with assistants, and that is kind of within the framework of the government office, kind of easy because then you have responsible civil servants that will take responsibility for what comes out of the assistant before they take it further in the process.
SPEAKER_03:They're still the co-worker, the servant, civil servant. So the assistance is behind the civil servant, it's not part of the flow, so to speak.
SPEAKER_06:True, true. So you're enhancing the human agent. What we see now is that with the integration of the databases, you see the first traces of agency coming in. So far, it's um perhaps more um uh empowering the assistant in terms of oh, the assistant can also get out to a database and start doing things in the database. But you're not in this case transferring agency from the civil servant. And that's I think that's that that's the most important distinction, especially within public processes where you have accountability. Um, you need to make sure that you stay within um the account the accountability process.
SPEAKER_04:But you can still have the human having the overall responsibility, but then letting the assistant or agent do more tasks in a like greater time scale or a greater set of tools than before without losing control of the output, so to speak. Meaning it could search the web or it could go to SCB database to do some things, perhaps even start to change stuff or to write you know some documents or to uh perhaps you know upload some report somewhere, or who knows? So it could be to have you know a set of different actions being taken, not just from a let's call it a research point of view or from a knowledge management point of view, but actually changing them some things. Is that something that you're also considering, or have some assistance starting to do?
SPEAKER_00:Well, we're finding the breadcrumbs for that going forward, but I think we're reaching a point where we really need to get the governance in in place to scale up. Uh, but we know uh where we will need to scale up and and we have an idea on on where we must go to you know reach that kind of uh position. Uh but it's so useful to to start small um and and build upon it. And don't be you know uh greedy if we can make our our uh civil servants 20 or 30 percent more efficient, of course, then we won. I mean, that's perfect.
SPEAKER_03:Maybe maybe assistant view is a very, very smart stepping stone, and because then those agency topics come to the forefront, and then you can, within the agency of one civil servant, you can give more and more tool use behind the back. Still you're accountable, but but you're straining together more and more tasks, right? And then and then when you understand how that works, we have a chance to have governance. That's okay, what's the governance now? Because now you're now you as a civil servant, you actually your your role need to shift to observe quality assurance within the role, and then etc. etc.
SPEAKER_04:I think we need to move to the very famous example of Ulf Christessen here, who you or he cognitively enhanced himself, right, by using AI and got a lot of uh blame for that, right? Can you give do you have some more insight into that? What really happened and give some perhaps a background for people that haven't heard about this?
SPEAKER_00:I don't really know. I was on vacation, but I'm happy. Uh I'm happy that the the leadership is is using AI. Uh definitely.
SPEAKER_04:I mean, as long as you keep it at that level, I mean who can be blamed for being cognitively enhanced by an assistant? That should be something everyone is doing, right?
SPEAKER_06:It was a very strange debate. Debate was and debate should be um you should be upset if it wasn't using AI. Exactly, yes. Um of course, you need to ask questions about the responsible use of it and and and uh knowing uh sort of uh the frameworks of it, and we don't know much about that, of course.
SPEAKER_03:Um but it was but it was so stupid because it's it was obvious that the person, I mean, I think the person who's initiated the debate and the people who jumped on it hurt themselves by their pure stupidity. There was an underlying very interesting debate: how should we use it or how should we develop in this direction, and when should we use it or not use it? That's a very good debate.
SPEAKER_04:But if and or not, that's that's just but I think it's highlights, you know, so many people are immature in what AI can do and uh what probabilities are, right? And that's something as you working as a change manager have to face, I guess, regularly, right?
SPEAKER_06:Yes, and and we shouldn't be I mean, we we also need to listen to those people that um is cautious about this new technology. Um, so it's understandable in one way that this debate comes up. Yeah, um one thing that we've been very conscious about in the project is to work that we we we single out and focus on early adopters. Uh that's our change actors, that's our AI pilots. Uh, and we actively seek them out and and and draw them in to the project uh because we believe that um you cannot convince people by PowerPoint. You need to make you need to make you need to show in practice uh the kind of utility uh that comes out of using these tools. If you can show that in numbers or just in practice in front of in front of people, uh you can win them over. We we do workshops with civil servants, okay, and it's really fun to see how they start off being a bit perhaps not skeptical, but sort of wonder where this is gonna go. And once you do a workshop and you let them work with the tool, you see how their eyes light up, and suddenly it stands clear to them that oh darn, I can use this in my everyday, and then they convert.
SPEAKER_03:I need to try some stand-up comedy. PowerPoint will never light up eyes. Never, nope.
SPEAKER_04:But I think you brought up an interesting topic here. And um, for one, of course, you have intrigue and we have a tech and we can build assistance, but then to get people to use it, we need some change management in place. And um, if we have the time, it would be fun to just think a bit more, you know, from a non-technical point of view. How do you actually scale value in the government offices in Sweden to make them really use it and find value? You you've been imagined already part of it with a workshop, etc. But can you just describe a bit more what's the process to really find value beyond the tech, so to speak?
SPEAKER_06:Um well, it's it's part of finding the right people that want to drive this development. It's finding the AI pilots and then supporting them and providing the platforms. How do you find them? Um, so far it's been kind of easy. They're calling us. But we've spent a lot of answer is good marketing. And and I can just give you a good example. Um, we have a commissioned investigation that that looks over the environmental process. You know, it takes 15 years to get the uh environmental uh permit to do something. And they're the government is is having a commissioned investigation to look this over, and they got 6,000 pages of investigation back, 300 pages of legislative proposals, and they had another 10 uh 4,000 pages with 200 pages of legislative proposals that they were sending out to, I think, 250 remittances, that is, other agencies and companies for review. For review. So they were expecting somewhere around 25,000 pages of information about these stand-up comedy. And they were two people working on this. Um do you do you do you wonder if they called us? Yep. Oh, really? Could you help us? And then uh we tried to help them. How do they know about you then?
SPEAKER_04:You have to have some kind of you know awareness building, so to speak.
SPEAKER_03:Yeah, what's the marketing here? What's the branding? What's the evangelization that leads them to call you? This is part of the story.
SPEAKER_00:Well, it's just feats on the ground. We're we're uh always uh saying yes to the meetings. Uh, if they want to know more, we'll show up, we'll tell them more, we'll listen. Listening is is I'd say the key factor. Uh so as soon as we hear something going on, you know, we'll we'll uh kick in the the door and and and try to join the conversation and see if we can can help. Uh but yeah, we spent a lot of time just hanging out with the with the different departments and talking about AI and listening to the so you go around and give like presentations that or okay walking the corridors and also just having the conversations wherever you can.
SPEAKER_03:But how how can how okay? This is this is obvious, but how to systemically organize it efficiently, you know, like so how have have you made an agenda that sort of we have a we want to have uh town hall meetings or we have a uh we have people who are dedicated to call, you know, how as a sales organization, or how how have you Practically sorted it out because this is super important.
SPEAKER_06:We had a range of workshops, prompt workshops. We have something we called innovation arenas, where we invited people to have a conversation about not tech but change management. Um, and within that framework, we had the history corner where we try to connect to uh Uksan Juana and the development of legitimate due process when it comes to uh uh having uh uh a good governance of the country. Uh and that it sort of is it's it's a package of understanding that we're in the middle of uh developing um how the country is governed, I think. And and once people uh take that in, they really become interested in the subject. Uh and then tech is one part of it, but uh, more important is how can I improve the job I'm put to do.
SPEAKER_03:Right. But here's another signpost, here's another principle. You didn't lead with tech, you put you put the tech in some sort of context of what they're really all about and what the real mission is all about, which is easy to say, but do we do it all the time? I think I mean, like if you think about instinctively or strategically, that's the principle.
SPEAKER_00:I don't think we would be here if we would have led with tech. We would be talking about how do we do this agency, how do we you know push the tech delivery further? Uh so we had the at the beginning with the first pilots, that was kind of word of mouth. People found us and we tried to, okay, what do you want to use AI for? Um, so when we had maybe I think 30 or 40 uh uh civil servants that that were interested, we you know put them in in a room and they came from different backgrounds and different departments, and then we just uh split the list uh in in half. So, Magnus, these pilots are yours, I'll take these. And then we you know gave some to the other guys and you know, give them a call once a week, listen to what you're doing. Yeah, exactly. Footwork. So instead of us spending time thinking about you know what can we deliver, how can we just listen to them, what they're doing, and and what do they want. So everybody got a few pilots per person.
SPEAKER_03:But that's another principle. We cannot be so completely preoccupied with our own agenda, so we don't have time to do the legwork or footwork. We need to have Slack in order to listen and to sell or to be there. That that is you know, because uh you know how it is. We are all inundated with work, and you're supposed to work with your digitalization, blah, blah, blah. Oh, I don't have time to meet the customer.
SPEAKER_06:I'm joking, but but not really. I think perhaps another, as as you say, I think listening is key. And government is funny because we we fly YOS in Libya with our right hand and we hand out student loan with our left hand, right? So government is very broad business. And if you listen in, most civil servants are doing very different things. So they sit with very different problems. It's very difficult to cluster them up, cluster them up and say, do this, do this, and do that. No, everybody has their own particular problems that you need to listen into and make sure that their problem is solved. Uh now, luckily, assistance, AI assistance is a very agile tool that made that it's it's possible to um comprehend more. Yes, to to uh you know solve their problems.
SPEAKER_04:Awesome. Okay, so what happened is someone reaches out to you. What's the next step? Do you how do you get them on board, so to speak, so they can start working with some pilot?
SPEAKER_00:Well, now we say let us get back to you when we're ready to bring on more people. But basically, we asked, what do you want to use it for? Like what's your what's your goal? What do you think you can use it for? And what do you think the results will be? Do you have to do some onboarding or show them how it works? Or right?
SPEAKER_06:It's it's and it's not that easy so far. Yet again, we're targeting early adopters, yeah, and we're not we want to make it a bit difficult to become an AI pilot. Okay, so you want to qualify to that.
SPEAKER_03:So we have readiness to commit.
SPEAKER_06:Yes. Um so what we're setting up now in pilot phase two is a nomination process. Uh-huh. That means that the different um heads of units will be able to nominate people, which will then hopefully be early adopters. Um, and then those nominations is is uh poured down to us where we can actually select what we believe will be the best AI pilots. Looking at their sort of qualifications and also their business cases that they want to try out.
SPEAKER_03:Feasibility, simply. Sorry. Also their readiness and the feasibility.
SPEAKER_06:And and also make sure that we don't double work or overlap in too much. Uh, so we're on the proper level of what we want to develop when it comes to functionality. And then we deliver the licenses and then onboard them and make sure that they have whatever they want to know and put them into user groups.
SPEAKER_04:Um, interesting that you gamify, I guess, this a bit by simply having people nominating other people and in that way turning it into a competition, I guess, in some way, right?
SPEAKER_06:Should be somewhat competitive.
SPEAKER_03:But I'm I'm hearing between the lines also that there's there's an emergent portfolio model around this. So we know we have limited funds, 300 in IT, and uh even fewer working with agentic or assistants. So you need to sort out um where to allocate your prioritizations. And so the nomination process is one part here, you know, readiness and feasibility. What how else would you describe your sort of portfolio view on how you sort of prioritize and act on these things?
SPEAKER_06:Yes, and and what we do is nominate people sort of way out in in the business uh processes, but we're look also looking at the more strategic processes, the um the uh government office-wide processes, and that's the ones that we take to the top gun, uh the center of excellence, and where we can work with those. So though those then we discuss with the um um expeditions chefer, um the head of the department, head of ministries, uh or or administrative heads of ministries, uh perhaps, and then um we have a conversation with them of which processes that we should prioritize within that function.
SPEAKER_04:And you have to describe a bit what is really the top gun, what is really the central excellence? What does it mean if you get uh nominated and elected to that?
SPEAKER_06:That you can participate in uh solving the most difficult problems that we face.
SPEAKER_04:So you help them out in that, and then yes.
SPEAKER_06:Yes, and it's not in place so far.
SPEAKER_03:Uh it's something that we're uh yeah, but but you're you're thinking carefully now, and I think this is a very smart of some sort of two-tiered approach where you are nominating people or we're nominating cases, and then we look at we which is sort of opportunistic, or basically this is uh the entry point or the inbox, so we can get the best cases for everyone, and then you strategically and surgically proactively think from a top cam perspective what are the broader impact cases, and probably also with the broader impact cases, it will require more DevOps, it will require more coding and stuff like that around it to make it work. Yes, and so you have a you have a portfolio strategy with the tow prong strategy where you have an inbox for for the greater masses, capture them all, and you have a strategic proactive view on with TopCal. I love I love that uh McKeinsey uh PowerPoint.
SPEAKER_04:Now you destroyed it with the McKinsey PowerPoint.
SPEAKER_01:Okay, I was just putting it into consulting charge of it's time for AI News brought to you by AI HW Podcast.
SPEAKER_04:Yes, so we normally take a small break in the middle of the podcast to speak about something else and uh then speak a bit about the latest and greatest AI news that we heard about in recent couple of weeks. And um, and then just go around the table and see if we have something that we'd like to share. Perhaps we start with you, Peter. Do you have something uh that you heard about recently in terms of um AI news?
SPEAKER_00:Oh uh no, not really. I've I've been so bad uh uh with my you know ears to the ground on on what's happening. Um but I did I did see a real cool uh uh from from a gaming point of view when they were using the new Unreal engine uh together with AI. And uh I think there's gonna be some cool stuff happening in gaming.
SPEAKER_04:For sure. And with the Sora 2 and everything that recently came out, you know, the way they can create videos, and then I'm sure it will come to games and the Unreal engine as well, and and yes, have perhaps you know real-time video generation, perhaps that would be amazing, right? Have you seen any of the videos from Sura 2? Um, not recently, no. Uh it's amazingly good. Magnus, do you have anything specific that you heard about?
SPEAKER_06:I actually stumbled across uh stumbled across some interesting news. I think it was Samsung of all companies that introduced a new language model that was kind of tiny. Yeah, I heard about that. But it sort of crushed all the big ones. And that was well, not crush, but they it was competitive. And from what I gather, and this is how I think about it, is they came up with a novel cognitive architecture that sort of complemented or empowered that little language model into something greater because it's iterated the uh the question and answer back and forth and made that little model work really hard to somehow look like a really big model. And that's really interesting because it questions this big is beautiful paradigm that everybody's running after.
SPEAKER_03:And there were some papers released if I've knocked on on that, and and I'm really speaking out of my depth here. So we are talking about different architectures, we're talking from the LLM architecture, and then we're thinking about uh actually we can get further with SLM small language models. And now I saw papers on new reinforcement learning-oriented uh small models, so TRMs. I don't know, I can't even remember what they were called. Do you know what I'm talking about, Andesh?
SPEAKER_04:Well, if there's something something one, yes, uh of Chris. And uh let's go there a little bit then. No, but uh, I I think you know that was a bit about reasoning, and you know, for something we know AI. I think a nice way to describe this is really that we know AI is good at some things, but really bad at other things. One thing is really good with what you are using it for is I would call knowledge management, meaning taking a large bit of text or images or data in some way and just being able to understand what it has and answer questions about it. What is worse is that is reasoning. And that's something that humans actually do much better. So by taking it in your approach and putting human in the center, let humans do what they do best, which is the proper reasoning and putting it in a context that really makes sense, and then let AI do what it's good for. Then you find the best of the two. What Samsung did here is trying to address the reasoning part a bit. And there is a benchmark called ArcHI by Francois Schulet, and he he tries to find uh clear benchmarks where AI of today really fails. And but it's super easy for humans. So, what this ArcHEI benchmark really tries to do is see that this is super easy for humans, but super hard for AI. So that's the case. Chat GPT has no chance to solve these, but this kind of tiny model that they had, I'm not sure exactly the name, is some kind of um yeah, some tiny recursion model or something, I think it was called. Um, and by having this super small seven million parameters, it's not seven billion, nor a trillion. I mean, the biggest one that we have today is trillions of parameters in it. It's a huge, huge number. Like a couple of billions is a small model. Yeah. Now we're speaking seven million. It's like not the thousands, it's a million part of the biggest model that we have. It's an insanely small model. You can run it on a whatever, yeah. It's about running on the edge. This is all about it. No, no, no, but this this is much more than that. You can run billions of parameters on a small on a phone. This is seven million. It's really small. But it's doing something else, which is it's much more iterative, it's recursive. And then it's focusing on the reasoning part. So apparently it does much better than the big one. For this specific task, it would be horribly knowledge-intensive tasks, but this kind of reasoning tasks, it's actually surprisingly good at.
SPEAKER_03:And a tiny recursive model, is that TRM then?
SPEAKER_04:Yeah, I guess so.
SPEAKER_03:Because I heard TRM, but I couldn't, I couldn't for my life remember what I think it's I'm not 100% sure, but I think that's. No, I don't, I'm not sure either. But uh the the um what I draw out of this that I think sort of pay attention to the trajectory, is that the world is a combination of things. Yes, it's a combination of LLMs for what they're good at, with SLMs, what they're good at, and now even now TRMs or something else. So we are talking ultimately as this grows, and to some degree, when you're using a platform, this is behind the scenes for you. It's AI compound systems where we are combining okay, contextualized data, guardrails, different types of AI techniques to solve different uh task lengths or different complexity of tasks. And I think that's it for me, this is very important that more people understand that or have a humbleness for this, that it's not one size fits all, it's not LLM for all.
SPEAKER_04:But I think I think it's also very important that people understand AI is bad at many things that humans are really good at. Yes. People are fooled, I think, to a large extent, just because they're so great and so much better than humans in knowledge tasks, that they believe it's really good in everything, and that's certainly not the case. And you could never replace a human today with the best of AI models that we have out there because they're only good, I would say.
SPEAKER_03:And this is so deceptive. This is so deceptive. Because it's sometimes from the outside really hard to make understand the difference. Was this did did the AI solve this because it has a vast amount of knowledge, or did it actually reason towards the goal? So it's it's if you have all the answers in in your head of all the different tests that's been done in the past, that's knowledge, that's not reasoning.
SPEAKER_06:And then you come down to the philosophical question: what is reasoning? What is intelligence? Do you have a precaution? No, but it strikes me when I when I see this, um it seems like uh a bigger part of intelligence, uh of what we call intelligence, seems to reside in language rather than in the brains, since we it seems like we can export it, we can export it into a language model.
SPEAKER_03:So and then if you put reasoning into that, also it becomes and then you hear me as a as the philosopher, it really becomes interesting to under trying to understand um because intelligence is because sometimes uh I think it's we can philosophi around um what is intelligence, how much is knowledge, how much is reasoning, uh, how much is knowledge acquisition rate, stuff like this. But then you can go another route. And then that is actually to start start thinking and framing it from what would be considered intelligent behavior in this context. So so as to make it more pragmatic, right? And intelligent behavior typically has to do with something doing it smarter, better, more efficient, co-adaptive in the context of uh of its environment, whatever that is. So something that doesn't work within its environment in a logical, in a reasonable way is not acting, behaving intelligently. So and I can take that to the school world.
SPEAKER_04:I would disagree with you, Dharm, but let's not go into that rabbit hole and define intelligence.
SPEAKER_03:No, but I'm not trying to define intelligence, I think it's very difficult. But I'm trying to understand that I'm trying to point to why don't we talk about intelligent behavior and reason about what would be intelligent behavior as a civil servant in this particular use case? And then you can get quite far.
SPEAKER_04:Let's get back to your news. I think it's another capital here. I would be happy to go there, but I think perhaps in the end, when we go more philosophical, I would be happy to discuss what intelligence, reasoning, knowledge management, etc. is. But do you have any news, Henrik?
SPEAKER_03:No, I mean, like so I I'm gonna do I'm gonna do the news angle from you know what of all the uh AI news reached the TV Fira so fun uh this week. And uh Paulina, who's been on the pod, she was uh you know, she's there sometimes to talk about different things, and of course now Sora too. And I mean like the news angle is sort of like um what fake news is one news angle, and then the other angle is like uh Hollywood, the the C C CAA or whatever it's called, and sort of the implications of what you know of how we do video and all that. And uh so there's the Sora news. We we talked about this technically, um two or three weeks ago, I guess. But but but already now, within two or three weeks, it has big ripples in terms of um what does it mean? Uh, how you know uh Sam Altman let it completely free, and then he needed to poodle on that and sort of rain it back a bit. Um it's available in the app in US right now, but not in Europe. I mean, like so I just wanted to point that out. I mean I mean stuff is happening, but um not really techie news, or not not specific, but uh it was just interesting that it it makes ripples all the way out in normal media.
SPEAKER_04:I think a related topic to that is that um apparently Hollywood in movie production and whatnot is making use of AI much more than they are saying. You know, it's had a real it's had a really bad rap because you know a lot of actors and uh screenplay writers, etc., is is really angry with AI and it's afraid about taking their job, but apparently under the bonnet that's been going on, they've been using that for for a lot of movie productions for a long time, um, but not really speaking about it because they're a bit afraid about the consequences and the yeah kind of interesting.
SPEAKER_03:Anyway, so that was sort of the the the the the bigger media uh angle today. What any tech news uh or any things you want to highlight?
SPEAKER_04:Tech news every should have a contrast in one. Um yes, uh so you know one of the things that AI actually also does really, really well is AI for coding, software engineering, and um especially cursor. Um I'm using it a lot myself, and the way it actually adds functionality on top of the LLM, you know, one thing is simply what the what the LLM can do, but if you as Cursor does, you know, the the real USP, so to speak, the the value-adding thing that cursor adds is the logic around it and the way it makes use of the LLM. And that's super hard to copy. Um they spent like millions and if not billions of dollars in just coming up the proper scaffolding, as they called it, to do it. And that makes cursor better than any other software development IDE, in my uh view. So they just this week added a new functionality called the plan mode. Um, so they had uh one for a while called the agent mode, and before that, yeah, basically out to complete. Agent mode is simply doing a lot of things, taking action and deciding on what to do by itself without you having to approve it, and it can change files and run, uh take actions and whatnot. But now we have a plan mode, and um, and that basically can can take a longer time to just come up with a big change and do the plan for it and go through the big uh code base in a way that it wasn't possible to do to do to do before. And I tried it out and it's really impressive. And AI for coding, I think, is a perfect metaphor and like a window to the future of what AI will be able to do for more or less any kind of business process or workflow that we have today. And in coding, you can see it now. And I'm just so eager to see with when this comes out for doing uh company uh analysis or for doing so many more things. If you can have all the functionality that already exists today for coding everywhere, it will be such a game changer.
SPEAKER_03:Yeah, and and the angle here is like within the coding space, there are there are also signposts in relation to the characteristics of the type of process and workflows and how you can decide how you can with code see does it execute, does it compile or not, and you get a very clear false or negative, or you know, does it work or not? So, but it also tells us a lot about which processes can we take the furthest next, and which processes is further and further down, like with the whole stack and technology and how we work, it needs to be a little bit more sophisticated because it becomes more and more human-like. Coran, did you have a news to you? No, but what about uh so I I saw Goran put up uh Claude Haiku. Is that what we're talking about here? Or is that something else?
SPEAKER_04:I think he mentioned the bioanalytics part, and that's actually a favorite topic of mine as well. Oh, okay, this is that one.
SPEAKER_02:But for me, this is funny because now bioanalytics, this is thoughts for 2015. Uh they are confusing bioanalytics analysis with uh like uh vibe coding, which is completely different.
SPEAKER_04:Yes and no, I would say. Yes, so actually it's something we've been doing at work quite a lot, and it's really impressive how AI for analytics have improved insanely in just recent months.
SPEAKER_02:True, but is it a vibe?
SPEAKER_04:Yes. So vibe is a you know term that's a bit controversial because a lot of people think it's a bad thing to do vibe coding. Um, I do not necessarily, but I can see the point. And I think a term like agentic coding is probably better, and so would agentic analytics be a better term than vibe coding. So, but I think the underlying meaning of it is very similar. It's just that people have a bad connotation to vibe. Anyway, what you can do, for example, in Snowflake these days, uh so Snowflake is really good in analytics, and uh they added a new functionality called Snowflake intelligence, basically building agents. So you can drag and drop and build agents in Snowflake, and uh what that means is you put in the data similar to how you build the assistance, very, very similar.
SPEAKER_05:Right.
SPEAKER_04:You just say that here here is the tables of data models that you want to use, and you give that to an agent in Snowflake, you give it some instructions saying this is how you should answer these kind of questions and what the you build a semantic view, as it's called, and then you let ask let people talk to it. So, vibe in the sense of talking to the AI is very similar. So you then can ask um a model like uh why has the sales gone down for product internet light in Denmark in the recent months? And then it figures out I should look at of all the thousands of data models we have, these are the ones I should look at. Then it writes the SQL scripts or code necessary to extract and transform the data. Then it takes action in building some visualization and reasons about you know what is the real reason the sales went down, and it comes back with a concrete answer saying the reason that the sales went down in Denmark in these months is because we had a lot of uh incidents and the competitor came out with a you know lower price point, and this is the rent the answer. This works surprisingly well today. It didn't work, we tried it out like half a year ago, it was horrible. Today it works much, much better. This is moving so fast, and this is a perfect example of something moving from by coding or agentic coding, agent mode coding into another process, meaning analytics. So now we're starting to see agentic analytics uh being actually or starting to work, not as well as coding, but surprisingly well. But it's interesting, right? And that will just continue for one workflow after the other in coming years.
SPEAKER_02:So decision intelligence, yes, right. So uh isn't it the same? Yes, what do you mean decision intelligence that we spoke about like uh two previous use of business insights, yes.
SPEAKER_04:So in that sense, it helps you inside as well. Yeah, yeah, you you have it helps with the decision making for the business purpose. But I think the point is simply to find insights from data in some sense, right?
SPEAKER_03:And and it's interesting now because I mean, like on the one hand side, you have the uh agent that works on company data on more sort of analyst market intelligence perspective, and on the other side of it, you know, you have coding cursor. And here all of a sudden we have business intelligence or more data wrangling, which if you think about it, it's a combination of understanding this part, but it's also a combination of coding and SQL or you know, be able to extract it. So it makes sense, right? And and it it it is it this is not rocket science or science fiction, it's simply a matter of the product becoming a little bit better and we becoming better at understanding and framing the problem correctly. And if you're gonna frame a BI problem, you kind of need to know a little bit more about BI, right? It's that simple. Telling going back to your principle, whoever domain owner is needs to own their own agent in relation to understanding the work they will give to an intern.
SPEAKER_04:Yep.
SPEAKER_03:Is that simple?
SPEAKER_04:Yep. It's a lot of challenges though. Yeah, and it's not as simple, I think, as that. But speaking about not as simple and the challenges, it would be awesome to hear. Uh, what are your experiences in building all these 30 plus agents or assistance? Sorry. So assistance is probably a better term. What has been the main challenges for you in finding value from them?
SPEAKER_00:Well, one thing that we did uh during the spring was uh we wanted to to uh go on vacation where we gathered all the the pilots and and their assistants and they would tell each other what has worked and and not. But it was so interesting to listen to. So we actually invited the whole uh government and and booked the the big uh aula and thousands of people. No, but uh a couple of hundred at least, yeah. So so uh it was really cool, and we invited some guys that have been on the podcast here, Magnus Gillen and Juan Falk. So uh some of the pilots they they got five minutes each and and and said what they had done. And uh one of the feedbacks that was so cool was they said that what you guys what you guys have done here is very believable because we didn't, you know, we just put the ones that worked, we took the ones that didn't work as well.
SPEAKER_04:Oh, that's a really good point.
SPEAKER_00:Yeah, because you get some credibility then, because everything is not sunshine and rainbows. Uh, and um, I think that is has been really good for us as well. Uh and and also in in form of transparency towards all the other ones. So uh we want to every time we we get something that works, we want to share it with all the other pilots. And also when we get something that doesn't work, we want to share it as well. Uh so the kind of transparency we have going on, uh you know, that the prompts and and the kind of uh system prompts that uh the guys are building, we upload it to a place where where it's shareable with with everyone else. So you can learn from each other in that way. Yeah.
SPEAKER_04:Okay, but but what were now you're still speaking about the success stories here, I believe, but what would you say the challenges were? Some surprises that you found out along the way that perhaps the um legal uncertainty around information management.
SPEAKER_06:Um and it's not a super easy environment to navigate in um when you're working in the public sector. Um and what we we are discovering uh is that some of the the uh the documentation, the logs, um, the system prompts could become a public uh document. Um and um and also So the the the the way that you need to um become more granular or more professional in your information management skills to be able to do what you want to do. And that even though you have that in among the civil servants, because we handle these these uh things every day, is that then you have to level up that knowledge for everyone to to um use the assistance and the information management on uh on that uh professional level. I think that's that's that was perhaps the biggest um challenge. And the way that we try to solve that is everybody learning from each other because that's the only way you can solve it.
SPEAKER_03:One challenge there is the actual learning curve, the knock-on effects. So when we start going in this way, we want to use assist and it's great, but all of a sudden you need to level up on information and data governance or information data encoding accuracy and all that because it's really now there's there's no room for sort of a human to wing it. So so I I would argue when these are blind spots that we have lived with for years that we don't really think about, that now becomes transparent, and we need to now we shouldn't hide them, but we should we should be very uh vigilant to spot them as soon as possible. And then what are the mitigations and how do we make this into process or workflow now? So your workflow as a as a civil servant actually now entails information management to a higher degree, more explicitly. The difference with doing it this way is that it's concrete. In what way do you need to, you know, oh, you need to do data governance? Yeah, what do you mean, dude? You know, now it's m in relation to your agent, you need to give him this information in this way, otherwise it won't work.
SPEAKER_06:Yes. So you sharp you need to sharpen that skill set all the time. And and what you also notice, of course, is that Minister of Culture do things different than Minister of Finance, they do things in different ministers than defense. So it's not uh you you can't find one answer, everybody needs to sharpen their skills within their environment to find their own answers.
SPEAKER_04:I think the legal uncertainty is an interesting topic as well. And um and an argument, let me just try this argument on you. I I don't believe it myself, so don't think that I believe this is the correct way to say it. But if you were to use a digital assistant to go through company data that may or may not contain personal information, either you do it manually and you read through the thousands of pages of documentation that you mentioned, or you use an AI that you query about the information. Should the legal requirements be different in the two use cases as you see it?
SPEAKER_00:Oh, well, definitely. I mean that uh if you do it by hand on pen and paper, it's it's uh not you know stored digitally. Yeah. What if it is in a PDF that you read through manually? Probably not. I mean, depending on on where you store it and and and how you use it, uh I I think there's a good chance for us to to use pretty you know sensitive information in AI, just as long as we have the the right storage, the right delivery, be it on-prem or be it, you know. Yeah.
SPEAKER_06:It it sort of boils down to accountability to the one that handles information. If you have full control on information, as you have when you have it as a PDF or a paper, you have full accountability of using that information. Once you start spreading that in different environments, it becomes trickier.
SPEAKER_04:Yeah, it's also the point, you know, DDPR has special rules about you know, if you have automated decision making, then it's the additional requirements, etc. But I think in your case, you are always having the human in the loop, so to speak. You're always having the human in the end that is accountable for everything, so it's not automated decision making at all in your case, right? And that should ease up things, yes, hopefully.
SPEAKER_00:It's not scary though with automated decision making. Yeah, I mean, the the tax agency are doing you know 100 millions of decisions per year, and 95 million of those are automated, so we can do it, and we will definitely do it uh in the government as well.
SPEAKER_04:Um and I think that's a very important point. And um, I think you know, Scatteracket or the tax authority is a good role model in that sense, and if you do the proper compliance work and do the proper documentation and risk analysis that you should do according to the law, then you can do automated decision making. It's not it's forbidding, forbidden anyway, it's just that you have to do the proper, you know, compliance work, right?
SPEAKER_03:And I I think we talked about this quite a bit on the pod. And I, for myself, is a very strong uh proponent of uh compliance or risk management by design. So you need to ultimately, and I think that's ultimately what Skatterwagt has been forced to do, is to basically build a DevOps view, build a software development lifecycle or an use case lifecycle, where we as early as possible understand the risks, understand the compliance work, and actually build that into fundamental definition of our requirements of whatever we are building. And the problem becomes when we are trying to sort this out as an afterthought and bolt that on top afterwards we haven't done the work. I don't think it, I don't think that scales can work. So I think the more knowledgeable we are on those things and get it into our fundamental process from ideation, and ah, we fail fast. This is not a good idea, we shouldn't do it. I think it's easy, as you say, but it's more work, of course.
SPEAKER_00:Yeah, and and uh we knew going into this, we would have a lot of discussions in terms of information security, in terms of GDPR. Uh, but we're stumbling upon uh new challenges that so GPT5 was working slowly a couple of weeks ago, so then we sent out an email. GPT5 is working slowly, changed to another model, and uh most of them changed to Cloud4 Opus. And when the invoice comes, so cloud four opus is you know exactly, exactly. And we didn't have you know the slightest idea that that was gonna happen. Now it happened, okay. New challenge, we gotta work with this, and that's something that's uh pretty interesting having those discussions. I mean, should we go out now and and tell the civil servants don't say hi to the chatbot, don't say thank you, because that's gonna you know render a lot of tokens. How do you use opus? Jesus Christ.
SPEAKER_04:Okay, another topic, but cool. You get surprises though. Oh, definitely. Yeah, yeah.
SPEAKER_03:And and that is, I mean, like so maybe for the listener, who was anyone interested, or even even other agencies. So to understand the 360 view of challenges or questions, step number one, you need to start in order to understand to uncover that they even exist, right? And now we see here now, okay, finances, total cost of ownership or cost of running it. It matters how you build it, it matters which models you use. What what what what okay, and what did you learn? What mitigative efforts can you put into this now, like moving forward?
SPEAKER_00:Well, it's it's it adds to the list of things we have to you know get in place before we can go into production. Uh and uh we we gotta figure out okay, how are how are we gonna invoice it, you know, uh handle it internally. What can we do?
SPEAKER_04:Can you take it yourself in your department right now?
SPEAKER_00:Now we're doing it, yeah, yeah, because we're we're testing. Uh so it's part of the we we cover the cost for the for the test, but it's not something that we'll be doing moving forward. And we've also noticed that you know, a kind of way that we think we want to deliver it is okay, we'll have some form standard, you know, small model with with uh great pristanda as the kind of default one. But for those that want a heavy workload, okay, you can switch to that model. But you know, it's gonna if you want to take the cost, that's fine.
SPEAKER_03:This is another principle we can draw out here. So when you go from testing or piloting or experimenting, you can have a central budget. Maybe it's the smartest way to use remove those hurdles. But as it scales, even if this is internal and you're a fairly small uh IT department, you still kind of need to have a product-centric understanding and treating this as a okay, how am I gonna go with this so it becomes more of a self-service, so we don't need to be there all the time? How are we gonna cost this out? How are we gonna price this out? How we or internally we would call it cost allocation keys, but it's nothing else than pricing and service management. So you get to the same, even if you're not a software vendor, you kind of need to know fundamental, or you're getting dragged into fundamentals. Would that be fair summary?
SPEAKER_00:Probably, but I think we'll we'll we'll notice it as we as we go along. I mean, so we were looking at how can we use genai to scrape websites for you know, kind of working with some sort of open data and and looked at the development project for that. And then a few weeks later, this is standard in all the deliveries from the language models. That's also interesting. So I think that we're we're always there thinking, okay, let's start a development project, let's hire some really good uh coders that can help us. And then the market just delivers. And that's so because listening to you guys uh when you were talking about what's happening in the scene, uh, a couple of years ago, all these releases they were you know at big conventions. This is you know going to be released and we're gonna wait three months. No, I mean, we are delivering without the the the super tech uh employees in in the office. I mean, we are delivering the latest technology with just a click of a button. When you know Cloud 4 or 5 is released, we have it the day after.
SPEAKER_03:Uh it's a different, it's a different game because once again, then okay, so we need to think differently. It's it's you need to have the product mindset, but at the same time, you also need to have another mechanism that keeps the eye on the uh or the ear on the relsen to understand, you know, let's not build this internally because it's gonna come with the next release.
SPEAKER_04:But I think that's a dangerous mindset to have as well. I I think you know, we we should, of course, understand that AI will improve in functionality in many ways, but we need to still develop things that is specific for our workflows, and that will never come specific in in a general model. So, to understand what you need to build yourself versus what will come in general solutions, I think is one of the key things that we need to figure out. And I think we also, of course, we know AI works with some things today, but it will work with other things tomorrow. And we need to have some understanding of what it will actually be able to do tomorrow and start preparing for that today.
SPEAKER_03:Ah, but you you didn't contradict me, you sharpened the argument. And the argument is to be super smart on what is the trajectory of the fast-moving inventions and what are fundamental slow-moving dimensions worth investing in, like data, data contextualization, whatever is your core business, they won't fix. You can go nuts developing here. But but as you said, techniques or stuff that is sort of ah, that is very close to the LM, probably gonna be sold. So I we so I think you're right. We we should not stop developing, but at the same time, we should not try to do what OpenAI is gonna do next month.
SPEAKER_04:I think what Sam Altman said is very true. It's never been a time when it's better to build a startup than now. And I think also the same can be said for enterprises, meaning there has never been a time where it's bigger possibilities and opportunities for innovation than right now. Meaning now is a perfect time to start building. Now it's a perfect time to actually start to improve the efficiency quality and transform processes and people. So I think you know, actually the opposite. I think you know now is the perfect time to really get started.
SPEAKER_03:I didn't say that I didn't say the opposite. I say the same. I used to say be smart on where where you put it.
SPEAKER_04:Let's leave it. Cool. Uh look the time is flying away. Looking a bit ahead, perhaps. Um, you had come a long way. You moved into phase two, it sounds. Can you just elaborate a bit more? What are the next steps for the government offices in Sweden to continue to scale value from the assistants and agents of the future?
SPEAKER_06:So maybe the pilot phase two was very much focused on the technical platform and whether we got utility and in the in the line. Um and uh we found it did. We found we got some good numbers on on the efficiency and also the uh mental relief. Um, or or we didn't find any cognitive distortion. We found a good balance of of um thinking versus uh allocating thinking to that assistant. So we we we thought this was a good tool. Now, what we need to do is to build the framework around it and also scale it. Um, so far in the pilot phase one, we had a fairly small group of people working with it. Now we need to make sure that each ministry has their own development group. Uh, we need to build that center of excellence, we need to work constantly with the different friction areas that comes up on what level should we put the functionality and um and and have the change manager within each ministry coordinating what happens there and between the ministries. So there's a lot of new things that we need to build and learn from. Uh so that's the next step.
SPEAKER_03:Is regering uh connected with the AI werkstan? Are you supporting here? Because Oskat de Verkis is there and a couple of others are like the versionskassan. Um I was just we I talked to Dante uh uh at at um Don Norel. And I'm just I'm just curious of how that works, or you know, because we have we we could be.
SPEAKER_06:You could be, but I guess so, and and I think maybe that Top Gun Center of Excellence that we're thinking of, that could be a partner. Um as could we we could suddenly probably if we would like to, but we don't know how that looks like.
SPEAKER_03:No, because uh without going bananas in imagination, AI Wexpan is trying to figure out how can we improve in the public sector around use of AI. And maybe they have many different uh facets of what that means in terms of data and big data, and but some of that is sort of a little bit old school, and here, you know, maybe there are pockets here which is sort of spot on what you guys are working on that should be what AI Bax then can help with. I don't know. What do you think?
SPEAKER_00:I think the old school is gonna have a really good place here because they uh both for Shark and Scusan and Scatterback, they are superb at information management and building robust big uh uh and also in in terms of of uh sovereignty, they're really good at building uh uh data with uh continuity. So we're super happy. Uh I mean we have the the public cloud services that are easy for us to use. We have the Swedish suppliers, the Nordic suppliers, the European, and now we have uh state suppliers as well. So we can, you know, we can choose.
SPEAKER_03:No, but my angle here is a little bit like there are two different value propositions that marry up really well. So they are working on sort of the the thing that scales big blah blah blah, sharing data across different uh organizations and all that, which is a really, really important fundament, and that's that's quite heavy to make that work across many different ministries or something like this. The problem with that is that is it's it's not as easy for the civil servant to get wrap their head around what am I going to use it for? Am I gonna use it? And the distance between using those sort of infrastructures towards what the municipality can do is quite large. If I flip it and like, let's think about what is the killer app that is closer to the civil servant, that then, of course, needs to connect to an infrastructure tissue in some ways, then you have a quite much interesting story. Here is the story that reaches the civil servant, public, you know, whatever, and here's the fundamental infrastructure that we need underneath it to govern and do this well. Together it's a good, it's a pretty cool story. Both of those stories is lacking in in one part, if if you want to scale it up.
SPEAKER_06:But but you can also see sort of a layer of of uh startups and and uh other companies coming in and working that space in between. Yes, of course. Delivering that service labor of what what what we will see come out of it. So there will be an interesting dynamic coming out of that Phillshop.
SPEAKER_03:It's just a matter of having the know-how of the different layers. And maybe you have strong focus here, but we have a blind spot here. You have done work here, and that adds clarity what that is all about, and then you it helps us fill whatever blind spots we have. Yes.
SPEAKER_06:And I I think perhaps the the most important part of that is get going. Yeah, get going. And then start experimenting.
SPEAKER_04:And I think you should be really proud because I think you are a perfect example of how a lot more organizations and authorities should be doing, you know, put it out there, start experimenting, see what works, and fail quickly if something doesn't work. So I'm very glad to hear that you are doing that. So congrats on that. What do you think potentially will have the biggest, you know, more transformative impact in uh the government offices in the future? What kind of digital assistants or agents do you personally believe this will have a big impact? Do you have any favorites?
SPEAKER_00:Oh, transcribing is one of my favorites uh because we have so many meetings and uh people are sitting writing notes a lot, exactly. And when you're writing notes, you're not paying attention, and and uh you want to pay attention, uh, be part of the conversation. So I think transcribing is going to be really big. Also, you know, taking a photo of your notes on a paper and just send them in and transcribe them, that would be really cool. I think that was and I think in terms of products, that will probably be you know one that we will deliver, but it will change and evolve and and and so. But that will probably be a standard product for everyone.
SPEAKER_04:Yeah, cool. Do you have any favorites?
SPEAKER_06:Tying back to I guess what you said early on. Um I one experience that we have is that if you give the assistant too much of a mouthful of work, it will become uh sketchy. So, what you need to learn is how to divide that process that you're working with into edible pieces, right? Um, and you can build one assistant per piece, right? But then you need to connect them also. Yeah. And then if you can find the language or protocol for tying those up, then the civil server can back off a bit and then start managing that as a knowledge-driven process. Um, and if you get to that, then it's the important part is also that if you if you build it as one big chunk, you lose transparency and you lose accountability. You need to have between these process steps accountability and transparency, and then that could be the part of the civil servant. You need to you don't need word as your primary work instrument. That could be produced somewhere else, but you need to have an eye on it to make sure that it's uh properly.
SPEAKER_03:I 100% agree with this uh vision. I mean, like so the all the agentic AI and all that, it builds a hive mind. It doesn't, it's not a big monolith. Uh I the big monolith for the in transparency and for the actual understanding of what is happening and from the complications of trying to build that. All the reasons why not go that route, but rather go small and then understanding orchestration. But but because but orchestration becomes a heavy topic here, then that that becomes interesting as well.
SPEAKER_04:It seems like all the major AI lab providers are building some kind of this kind of graphical UIs where you connect agents to each other in a nice workflow. So it certainly seems to be the future in some way, how you can do that in a easy way.
SPEAKER_03:If we look at the implications, what is gonna be the main difference in being or acting or you know, succeeding as a civil servant if we follow this trajectory? You know, what's the implication of work? What is the implication for the people, for the humans, and the leaders, and the leaders, and the prophets.
SPEAKER_06:The way I tend to talk about it in our historical corner is that we have a long heritage of independent, high independence and integrity as advisors to the government, and that's the values we need to keep, bringing that into the age of AI. That doesn't need to be, oh, I have to really fast typing in Word. No, I need to make sure that the government gets the best uh knowledge-infused decisions uh to make. Um so so, in terms of having a high level of cognitive maturity and able to uh judge whether the products that comes out of these machines or assistants or agents that will be still um the key um the key competence that will define civil service coming out.
SPEAKER_03:We talked about this actually on the last pod. I'm gonna test a a theory now that it's surprising to see that some super senior coders they completely excel and go can go do so much with with Cursor and with all these tools, and others has not really succeeded as well. And then we came up with a theory that actually what happens now is that even if you're the best coder, a super senior, but you never led other people, you're not really good at leading, communicating, sending or giving instruction. So the whole topic of communication, leadership, clarity becomes an essential uh stepping stone here. And I think that goes for any job. You need to be able to lead and delegate to your assistants, especially if you have hundreds of them.
SPEAKER_06:And I guess if you take that to the AI stated agentic world, you need to be able to assess what kind of cognitive capability you can transfer to an assistant or agent and what you need to keep in your own head.
SPEAKER_03:This is leadership, also, how should I delegate what should I delegate?
SPEAKER_04:Yes, I guess so. Are people in the cabinet offices afraid of AI or excited about AI?
SPEAKER_06:Um, as always, they're a bit like, huh, what's this? But once they get their hands on it, they become curious and they really get excited. Uh, I haven't seen any other reaction to it.
SPEAKER_04:Um that's afraid that it will take their job or something.
SPEAKER_06:Of course, they ask tricky questions about it, not about uh taking their jobs. I usually describe the government offices that you have um some 800% pressure of what you need to deliver. And with hardworking civil servants, you deliver 120%. So if you get AI in there, maybe we can deliver 240%. Perfect. We still got 800.
SPEAKER_03:Yeah, the demand is huge. The mountain is huge. And yes, and if everybody could understand that in the whole public sector and the shortage of staff or nurses or anything like this, it's the same everywhere.
SPEAKER_06:It is, but but this is an important tool in order to increase the efficiency and the productivity and quality. But also the cognitive height of what is delivered.
SPEAKER_07:Yeah, I fully agree.
SPEAKER_04:But are you a bit annoyed as I am sometimes that media portray AI as something else? Of course.
SPEAKER_06:And actually, we we tried to switch when we we talked less about AI than language models because language models is more concrete, it doesn't sound magic. Um, it's more, oh, this is a new tool I can work with.
SPEAKER_03:Maybe that's also an interesting topic to to start decomposing stuff and not using the the word that doesn't mean anything anymore. Like, yeah, what does it mean today, right? Uh and and we have many examples of that where we actually are using lingo that blurs rather than sharpens.
SPEAKER_04:Cool. Uh yeah, so we're moving towards uh the end here, and uh we are getting more and more philosophical, and I would know we have a philosopher here as well. So perhaps we can start to just consider that. You know, what is the implication and ramifications into our society when other parts of the society starts to do what the government offices is doing and actually employing agents at scale? Will that be positive, or what do you think will happen at society uh if we were to do more of your type of massive scale, massive scale everywhere?
SPEAKER_00:Well, as long as we can take care of it, because that's what we have to do. We we we need to know and learn. I don't think we're if we're gonna be in a good place if if uh uh we send out something and we get back a response that's made completely with AI and we use AI to to decompose it. I mean that's probably just a waste of time.
SPEAKER_03:So we that's the meme, right? Like someone is pointing, someone is uh three bullet points. Can you make an essay out of this? Sending the essay, and someone is oh, it's an essay. Can you make three bullet points out of this on the other side? Yeah, that's the joke, right?
SPEAKER_00:Exactly. But I but I think it's uh you said it uh before, Anders. Uh this is different than other deliveries in in the past. I mean, everyone is using it. Uh, our our moms, our our our families and friends, everyone is using it, and everyone is actually getting pretty good at it pretty fast. Uh, even people that uh I never would have thought would use AI is using it, and they're using it for you know at above average level.
SPEAKER_04:But it seems some some organizations are not using it. I agree very much what we're saying, but I know just the other or this week, you know, we had to collaborate with the university. Uh, I won't say the name because I'm going to be harsh. Uh, but they are forbidding the use of AI in the university and saying that that's really hurting education or whatnot. Do you can you see that happening, or what do you think about organization having this kind of mindset?
SPEAKER_00:I think that they are definitely using it. They forbade it and and uh they are not delivering it, but yeah, I'm pretty sure they are using it.
SPEAKER_04:Yeah. It's strange that so many still have that mindset, isn't it? So I think it is actually more than people think. So I mean I know companies that forbid using it as well. And uh I'm glad to have another example here of an organization that actually do, in a successful way, make uh use of it and find value from it. But it's surprising how many that don't.
SPEAKER_00:But it's tough. Two wrongs don't make a right, and it's it's really hard to, you know, you can have the better the devil, you know, approach. Uh if we don't deliver it, they're gonna use their you know, private phones or private computers or or whatnot. I mean, we we are you know abided by law to to deliver services that uh that follow the law and have the information in a good and controlled way.
SPEAKER_04:But if we take a specific problem like hallucination, I mean a lot of people speak about that, and we know AI really wants to say something, and it's sometimes just you know, in the strong urge to say something completely makes some facts up. Are you afraid about that? Or what's your thinking in your use cases in the government offices about AI hallucinating?
SPEAKER_06:It's something that you need to deal with. Um, and it's a combination of prompt craftsmanship that you build. The correct prompt. You manage your data in a proper way so that it it focuses on on the right things. But then all in the end, also you need to validate uh the precision or what comes out of these um these uh assistants. Um but I think it's unfair to compare these assistants with humans that would be 100% correct all the time. Would you miss be that? No, it would be unfair to to compare them with uh people that are 100% correct all the time because we make mistakes as well. Yes, that's right. Um the difficulty is to find um measurements or metrics that make sense of that, that we're not perfect, right? Because we always think that I had this conversation with um uh Minister of Justice, the the the um one of the departments there, and um uh one of the participants asked, well, if I have to uh read through what uh the memo uh if if it's correct all of it, uh well what what's the use of it? And then we ask them to back back to them. Well, if you write uh uh a memo in Word, do you send it off right away or do you read it through before you send it away? Right? So you you don't even trust yourself when you produce things, um, and we have to find ways of dealing with that discrepancy between 100% and 98%.
SPEAKER_03:But let me test an angle here. One way of looking at this is that we've been very good at uh and we have a legacy where we are good at understanding deterministic systems and we build guardrails and processes that work according to what is fit to have deterministic systems fitting into a workflow. All of a sudden now we enter in probabilistic systems. So I think it's quite logical that we need to uh figure out the guardrails and the workflows in order to have probabilistic systems work. And the point is this with probabilistic systems you have some huge benefits, what you can do, but you have some other mitigations or risks that doesn't exist with the deterministic. So each horses for courses a little bit. So one way for me to look at that is I think you solved it, is like, well, we need to know our probabilistic processes, they can hallucinate. We call it with jargon. This is not hallucination, it's probabilistic behavior. And then therefore, our processes need to account for that in order to maximize the value from it, but also minimize the risk. It's that simple.
SPEAKER_06:And you have transparency and accountability at each step of the process, not building those big chunks of the process.
SPEAKER_03:Once again, why it's tricky with with big chunks because then you cannot call check out the probabilistic nature of what happened. Yes.
SPEAKER_06:So it's it's a it's a craft, um, and it's doesn't it's not one size fits all. Uh, you need to look at each process and then build accordingly.
SPEAKER_03:I it's going a little bit back again, but but I'm I'll let's try to make a philosophical question out of it. I mean, like, how much have you needed to reform? I mean, like, so I the practical question is because all the stuff you're saying now, okay, we that means that has made implications. How does the onboarding process look like? Onboarding of a, you know, how do we educate, how do we do change management and all that? Very practical. But if I if I zoom out of that question to a philosophical uh to a more macro level or philosophical level, you know, where do you see the trajectory of of how we need to organize work or how we need to drive change management? That is that you can see in the small things that you need to do differently now. Uh, you know, you we need to educate people around this stuff uh on the job. Um, I'll give you an example. Learning and development as a function was usually used to happen in a different area, and they had a big course catalogue, and it was all quite generic. Philosophically, that that won't fly with with the context of I need to relearn exactly what I need to do, and I need to know uh generic techniques, but then I need to apply it in my context. So all of a sudden, philosophically, you kill the old learning and development function in its old capacity and it needs to get to a new capacity. You know, can we elaborate a little bit about the there's I think there are quite big implications of what that you can actually see in the small things that this shifts a lot of stuff, how we learn and how we change.
SPEAKER_06:I think it's just falling back on we need to become a learning organization. Um in different respects, that we will see continuous innovation. Um, that means you need to pick up the new stuff and integrate it into your workflow. Um but also in another sense, what what we see is like I've been working in that government offices 20 years. So if I would leave, so 20 years of experience would leave the house, right? If I train my AI assistant, 20 years of experience would stay at work. So my successor would come in and interact with that. So that means the organization would become also more of a learning organization. Um, and building structures for that, building ways of administering those assistants so that would stay relevant over time, these are new problems that we need to dive into.
SPEAKER_04:So keeping the knowledge, you know, not just in humans, but actually in the AI in future could actually make the organizations more of a learning organization. Yes. Yeah.
SPEAKER_03:But but there's a key word here, learning organization at hard. And we need to reassess every single process and how we do things based on that fundamental simple statement.
SPEAKER_06:We we sort of start off the premise that we don't change or uh processes necessarily. Um I usually think that culture eats organization to breakfast. So once we build an innovative culture with the help of AI, you'll get many things for free.
SPEAKER_03:But then maybe a learning culture. Yes. An innovative culture that then adapts the organization accordingly. Is that a better way of putting it, maybe?
SPEAKER_06:Yes, I think so. And once you got that, you will get incremental organization organizational change as well.
SPEAKER_04:Adapting, adapting. Yes. I'm just talking about it.
SPEAKER_00:And adapting, I just gotta uh tell you there, Henrik, it's really cool. So adapting and changing to the times. Uh so the assistance that we're that we're building, we're getting the portfolio. Okay, where should the portfolio land? Who is gonna govern the portfolio? Exactly. So we have a library in in the center of Stockholm, and uh they came uh to us and said, you know, people are not renting books anymore. We can take care of of uh governing that um so you would go to the library and ask to you have this book instead. Now you would go. I need this kind of AI uh AI workload.
SPEAKER_03:Is it Louve some uh Horme? Is it Luve? Is it cool we're talking about? Or what do you mean with that? No, no, no, our library.
SPEAKER_00:Yeah, so no, and this is a library with only legal books and finance books, but that's really cool because that then they are coming in. Okay, how can we how can we join you guys? How can we be relevant? Exactly the point, right?
SPEAKER_03:So the learning culture leads to learning organization, leads to organizational reform.
SPEAKER_06:And and you might also see that what happens with AI is that the knowledge intensity of the organization increases. So you build a more knowledge-based organization. And if you talk about library, AI Assistant is like a book, right? It's a knowledge product, but it's more of a dynamic book. So you need to manage that somehow in an organization, in a new way. And that's one of the things that we're exploring.
SPEAKER_04:And humans actually are really, really bad at knowledge management, I would argue. So with the help of AI, we can improve humans so much. Okay, and let's move to the last and perhaps even more philosophical question. So, for one, um, if we at some point in time in the future get AGI, or can we start with that? Do you believe that we'll get AGI at some point soon in the future? Uh, if we start with Peter? Yeah, eventually, probably, yeah. The last invention. In uh two, five, ten, fifty years.
SPEAKER_00:Hundred years. Um leaning more towards 10 than 50.
SPEAKER_06:Minus. I'm more a proponent of ASI. Um artificial specialized intelligence. Specialized intelligence. I I'm not a great believer of the business model because behind AGI. You don't need AGI to run my process at my desk, but I need an ASI to run my process at my desk, right? And I will not pay for the tokens of the AGI to run my little process here. I love this call. I will I will pay the price for the ASI. So it might turn up. Will it be profitable? We'll see. I think there will be loads of ASIs popping up that will specialize in mining, traffic, health, whatever that will at least put up a good fight against that AGI. Good point. I think it's a great point.
SPEAKER_04:Similar things for many times here. So we probably will have some very, very huge models, you know, speaking not trillions, but tens or even hundreds of trillions of parameters in them, but no one will be able to use them. They're too expensive. Yep. So then we'll be a large number of uh SLMs or smaller language models, or ASIs as you call it, and they will be the one we actually make use of. But yeah. And also potentially more than humans. If that happens, uh, we can imagine two extremes. Uh, one being that um AI will do the very dystopian uh version where they try to kill all humans, and it's the matrix and the terminator version of the future of the world. Or it can be the more utopian and the other extreme version where AI will help to solve all the challenges we have and cure cancer and solve the energy crisis and create basically products and services that's where the cost goes to zero and we have a world of abundance. Where do you think we will end up in that spectrum?
SPEAKER_00:Most of the the the larger organizations that are doing good stuff with AI now is putting putting the humans in the center. And I think with that approach, I'm I'm uh I'm hopeful. Um yeah. I I I I don't believe it will be a dystopian AI end. Okay.
SPEAKER_06:Glad to hear it. Becoming a bit philosophical, then I turn to the theory of extended mind.
SPEAKER_04:Okay, interesting.
SPEAKER_06:Which basically is yes, it basically tells you that um if you have two persons, one has a bad memory and one has a good memory, and they've got to go buy milk and what have you in in the store, right? One needs to have a notebook to scribble down what he needs to buy. That is from a cognitive point of view, the same thing. Yes, you just uh extended your mind to something external to your mind, and this is what we see now, right? So we're extending the cognitive capabilities to something outside us. The human is still its center and we should put the human, we should keep the human in the center. Um I'm not a big believer of the magical part of intelligence in these language models. I to me they're still just numbers and probabilities of the big calculator. If something magical happens in there, um I wouldn't know. I I I still believe that um the human will stay in the center of this um development. Um what we see is an extension of our minds rather than a foreign mind.
SPEAKER_04:I like it, it's very similar to actually Elon speaks about this, and um, he speaks about the limbic system, you know, the uh more reactive kind of um part of our brain uh that really moves and thinks in an instinctive way, and then you have the cortex part that is the more conscious part of the brain that actually do reason about the things we want to do before it actually takes action. But then we, as he called us, have the tertiary layer, the third layer, meaning the mobile phones we already have in use or the computers we already have in use, and that's actually part of the human. And if we were to remove it today, so many people would be extremely challenged because we are so dependent on our tertiary layer already, but we still have the limbic and the cortex system at heart, right? And uh then the question is you know, what happens if AI goes better? And then Elon's answer is uh Neuralink, right? To integrate even physically with the AI. What about that? Would that be something you would consider?
SPEAKER_03:Going extending the cognitive is that Neuralink next, you know, if you follow that trajectory, you end up in Neuralink then, because then the keyboard is too slow to interact with uh your tertiary layer.
SPEAKER_06:For for certain applications and especially for disabled people, um it's definitely a promising future, I think, um, for normal people. Um I'm not sure.
SPEAKER_04:What if you can tenfold your memory capabilities and you just import that brain could use that?
SPEAKER_03:But because but you your philosophical uh trajectory is cyborg. Yep. Um not not good or bad, but just technically it's it's cyborg. You know, when you take it when when you extend that far enough, in in in in the end you have a bandwidth problem or or a communication or memory problem. Or you know, so so that so the the extension of the mind it leads to cyborg logic, right?
SPEAKER_06:Yes, I'm not sure I want to turn into a cyborg, I want to stay human.
SPEAKER_00:Peter, you want to have a memory ship? Uh no, no, definitely not.
SPEAKER_03:Um but I heard that story, you know, of the fundamental divide of the haves and have-nots, of the riches and the poor, of the guys with oil. And in the in the end, you can imagine a society where a third is cyborg and has extreme capabilities in terms of what they can do, and the rest of the world is not. Then you get to an art, then you get to another answer that's some of our podcast guests has said, and it's like, oh, we will have both. We will be on both sides of the spectrum. Why wouldn't we? That's how it looks in the world today. We will have uh dystopia, we will have uh we we we will not be able to distribute it equally to have to end up in one spot.
SPEAKER_06:My point here is probably I we shouldn't fall into the trap to think that intellectual capital is all there is. Uh, people would with emotional capital or social capital, that might even be more worth, right? From a human point of view. So extending our minds in the intellectual way to build more intellectual capital, yes, that's interesting. Um, is that from a human point of view the most valuable thing we can do? Or the whole story, even. Oh, the whole story, maybe not. Maybe not. So I love it.
SPEAKER_04:Cool. Well, let's wait with the membership and perhaps stick to the digital assistance of the government offices in Sweden and keep up the great work that you have been doing. And I think you are truly a role model for how inspiring more agencies should do this, putting the technology in the hands of the people and letting them explore it and empower themselves. Thank you so much for coming here, Peter Nordström and Magnus Ensel. Thank you.
SPEAKER_03:Thank you.
