E184 - Navigating AI in Public Sector - David Wallén

In Episode 184 of the AIAW Podcast, we’re joined by David Wallén, CEO and Co-Founder of Intric, to explore what it really takes to drive AI innovation inside governments and critical industries. Drawing on insights from more than 50 public-sector AI implementations, David breaks down the realities behind procurement structures, governance models, regulation, and organizational incentives that often slow down adoption. We discuss sovereign AI, secure infrastructure, Europe’s ability to compete in the global AI race, and what organizations consistently get wrong when implementing AI at scale. From practical lessons in public-sector transformation to the future of AGI and human judgment, this episode examines how governments and institutions can move from experimentation to meaningful AI impact. 

Follow us on youtube: https://www.youtube.com/@aiawpodcast

Transcript

Juggling Calendars And Priorities

Anders Arpteg 0:00

I'm gonna do this while I'm doing that, and then I will also do this. And then uh yeah, that's just how you juggle things. Um I mean it must be so tough to to be you know running a startup. I've been in a startup world myself, you know, so I remember how hard you have to work. But you said it were triple booked? Yeah. Yeah, that's that's great.

David Wallén 0:18

Tell me. Um so I was uh supposed to be at an event and talk to people there. I was supposed to have an interview and then I was also supposed to be at home to take care of my dog. So that was not actually a meeting, but that was a responsibility I had. So triple booked and you didn't cancel, so you navigated it. Yes, yes. So it was an interview, dog care, so to speak, and what yeah, and then the event like uh an event about uh the society of tomorrow and about AI in the public sector. Um so yeah, but uh I managed to.

Henrik Göthberg 0:50

But maybe that's maybe that's the core thing. And when you're an entrepreneur, is it more about juggling and navigating than actually mean being able to cancel things?

David Wallén 1:00

Yeah, I think um I think the or I I think that the successful people managed to like shape the role as the company grows. So I think um now I'm in a position where I probably should like say no to things and manage things a little bit better, like manage my calendar a little bit better. Um, but I think I've grown uh I've managed to pretty well mold myself throughout all the years, like from being two people to being 30 people. Uh I mean it's a constant moulding of the role, and um I think that's hopefully how the next five years looks as well.

Henrik Göthberg 1:36

But this is a very interesting topic. Um, I mean, like I'm sure you've heard that uh that sort of conversation. Uh Steve Jobs uh was famous for having a really high signal-to-noise ratio, and we talk about oh, he probably had 80%. All his work, all hours he spent was 80% signal, so he was managed, he was quite brutal with what he was cancelling out as noise, and then now they say the one that beats him probably has some Asperger and stuff like that. Elon Musk is famous for having almost 100% signal ratio, which means he's not a functioning social being, probably. But is that the truth? Also, to be really I mean, like because you are juggling, juggling is not really the same as 100% signal. No, so what's the way here?

David Wallén 2:25

I mean, that's a good question. I think the question is like, how do you how do you find signal or how do you prioritize? Uh and I think that that's actually analogous to the like our biggest uh challenge as a company for Intrick. I've said it throughout the years, is that knowing like what to prioritize? It doesn't sound like a game changer, but in the age of AI, like there's so many balls. I say it in Swedish, like choosing which balls we run on. Like there's so many things you could be doing with AI. So and that comes in from all angles. It's ideas from customers, it's ideas that we ourselves have, it's things you read on the news, it's things that you you know your parents tell you, or your relatives, or your peers in the industry. So you can do so many things, and um, yeah, that's the that's the struggle, like choosing what you build and what you build.

Henrik Göthberg 3:14

But but then it then it's two questions. It's like really picking out what you believe. Your belief is in signal. Yeah, what's your real signal, as in all the things that potentially could be signal, but turns out to be noise, and then brutal focus to stay with that signal and not jump around. Yeah, two different challenges, yeah. Actually, two different personality traits, even. Yeah, yeah, I think so.

Anders Arpteg 3:37

Do you have a day nor way to just prioritize your everyday calendar? Um get things done, so to speak.

David Wallén 3:45

Yeah, I I think I I do it in a um in a pretty like structured way. Um I mean every day I pick things that I like three things that I need to get done, like no matter what. I do it on the train on my way into the office. It's at the Team Ferris way.

Henrik Göthberg 4:00

Sorry? Team Ferris, a four-hour work week.

David Wallén 4:02

Yeah, I I'm not sure. I'm not sure. Maybe that's it. I don't know. I'm not familiar with that. Uh but start with the vital ones. Yeah, yeah, probably like that. Then then you feel good. Yeah. And you also have I also try to think of tasks as like urgent but important, urgent but not important, not important but urgent, and then not important buttons.

Henrik Göthberg 4:21

Yeah, exactly. Seven habits of highly thinking people. That's a good one.

David Wallén 4:24

So that one I like I've I put things in those squares, and then like um every day you pick three things from the like you know, the important and urgent thing. Uh and that that's worked pretty well. Uh, I've done it for like two years. Work works pretty well. But that's that's pretty structured, that's pretty good.

Henrik Göthberg 4:39

Yeah. How do you do it, Anders?

Anders Arpteg 4:41

Well, uh, I follow the uh weighted shortest job first kind of approach. I mean, right? I mean it's it's um you always have you know these kind of long-term tasks that you have to do, but then you have also short-term tasks. And uh I try to make sure that uh during the day you actually spend some time on short-term tasks, and it can be anything from answering some emails or whatnot. Uh so I do them usually first and ensure that I actually make them the short tasks first, um, and then I uh have some tasks on more long-term per day. So the mix and balance between uh short term and long term uh is is the major thing that I try to do.

Henrik Göthberg 5:19

And is that uh back to the getting things done framework? So we have three major frameworks here that is playing into this. There's getting things done. I I can't remember who wrote that book.

Anders Arpteg 5:29

So it was a book called something, Henrik, or what's your preferred way to do what I'm passionate and have fun with.

David Wallén 5:36

That sounds nice. I hope to get there one day.

Henrik Göthberg 5:39

No, I'm terrible. I I I have tried all the different I mean getting things done, try that. Uh tried uh Stephen Covey's seven habits uh and it taught it to uh you know coached it and stuff like that, even um, and then um uh what was the last one? You know, the the top three things per day. I think I if if it's anything I I'm adhering to is the logic of Tim Ferriss, the four-hour work week. So it's it's fundamentally the 80-20 rule and it's fundamentally about signal. So trying to uh suss out what is signal and basically not thinking about the big to-do list, but basically what items on the list do you if you've done that, is that the one that makes the biggest difference for tomorrow? And is that the one that makes you feel that you had a productive day? So that is kind of I I think I'm leaning into the top three uh agenda most. The other ones have I I I'm I have a hard time uh with my discipline, simply to to stay on course. Uh with the so I'm more principle-based than structure-based. I don't know. It's I don't I I'm pretty bad.

Anders Arpteg 6:53

Yeah, well, self-discipline is a good thing, I think. But yeah, it's hard, hard too, hard thing to do. Well, with that, very welcome here, David Valent.

Meet David Valent And His Path

Anders Arpteg 7:04

Thank you. Um CEO and founder of Intric. Co-founder, yeah. Co-founder of Intric. Um, but really looking forward to hearing more uh about what Intrick is doing and uh how we can also help the public sector innovate, navigate in innovation and also become more agentic in one way or for. And before we go there, perhaps you can just give us a quick introduction to who is really David? What's your background and how did you end up at Intric?

David Wallén 7:30

Sure. Um, so um yeah, the short and the short version is um I got introduced to programming thanks to computer games, uh, like I think many of us. Uh and uh I was also always actually interested in hardware as well. So that's also thanks to computer games. I needed a graphics card for my my computer to play a certain game, and then I had to figure out how to get one and make it work. Um, and um yeah, so that led me into technology, and um I yeah, I've always been a builder. Like after after high school or gymnasize, I I programmed for Ericsson. Um, and then my parents who are academics they are um uh thought that uh it's probably a good idea to do that, get a silve approval on this like knowledge that you think you have. Uh so uh so I did that. I I started robotics in in KTH, and um yeah, that was partly for my love of hardware. It was really tough. Like it was one of the toughest decisions I've ever made, choosing between computer science and like mechanical engineering, which could lead me to robotics.

Anders Arpteg 8:32

Um, so I I think Robocap thing as well, or no, what's that? Uh the the football playing robots.

David Wallén 8:38

No, no, I didn't. I didn't get time to that. Um but maybe one day. Um but it uh I think that we could have, in hindsight, I intrigue could have been maybe happened a bit sooner if I choose uh computer science, maybe. But um I think it uh like I believe somehow in that you find your own way, and like with with robotics, uh I mean I I was part of starting up another company while it was a robotics slash design studio while I was studying. Uh, and then um yeah, robotics, like at the end of university, I took the first course in machine learning, and then I was really like that blew my mind. This was 2018, and I was like, holy shit! Uh sorry for the language, but uh I was that was like this is the only thing you can work on as an engineer in the 21st century. So I was like, no more a thousand if statements, like this is how you really build smart systems, and uh so I knew I had to do something with AI, and I was just not sure about like what, and I started experimenting with language models. In the other company, we got access to open AI's early models, and uh like just after they started to be closed source or stopped being open source, depending on how you see it. Um uh yeah, tried experimenting with them, and then I started training language models in Swedish. So that was the that was when I started the company. That's what uh what I did. But you said you used KB Birth.

Henrik Göthberg 10:01

Yeah, or uh yeah, exactly, or uh yeah, exactly. So KB Birth was uh like from the Bert model that came out of Google and then uh Louwe, I guess, at the Kungliga biblioteket. Distilled down the first KB Bert model and made that uh fine-tuned Birth. Fine-tuned Bert. Fine-tuned Birth is a good better word for the Swedish data that he has in Kungliga biblioteket. And then that was then open sourced on you could download that on the hugging face already then. Was it Hugging Face already then? Yeah, yeah.

David Wallén 10:35

And uh yeah, exactly. And uh, this is also like um, I mean, what I wanted to do was to just bring something to market, right? Uh, like prove that I was good enough to build something that other people found valuable with with AI. Uh, and yeah, so that was what was available. And you have to narrow it down to just a very niche task if you're gonna if you're gonna provide something of value when you are very few people, you need to be a small very small problem in a sense. So the this problem was summarization on Swedish language, and then kind of uh I've been iterating on that ever since, like bringing something to market, seeing what problems arise with that, and then solving that problem.

Anders Arpteg 11:15

Um for personal reasons, or what was the use case you tried to do summarization?

David Wallén 11:19

No, no, uh, so that was like in the ed tech sector. Uh so like to take uh for various reasons, you wanted to take a long text and make it shorter. So there were companies who wanted to take the long introductions of the Wikipedia and just make that more understandable for kids.

Anders Arpteg 11:34

But you hadn't started Intrigue then, yeah.

David Wallén 11:36

No, exactly. That was the company that then morphed into Intrigue. So it's the set like it's the same company, but um, I mean uh I met Jonathan, my co-founder of Intrigue, like six to twelve months later after bringing this to market. So then he came along, we met, we both had this passion for AGI. Like it's it started as a we met in uh we met in um how do you say as uh like a student uh party? Yes, yeah. Someone graduated. My wife's uh uh little sister, I think, graduated, and we I was speaking about this company that we were gonna build AI, we just need uh AGI, and we just needed to bring uh product to market and make it good, and then we'll get us to AGI. Yes, yes. I mean, I still do.

Henrik Göthberg 12:19

I mean it's the same as going to Mars and having several use cases and business models on the way.

David Wallén 12:24

Yeah, I think that that's a long-term plan. Um, so and then, but then we had customers in the public sector, and we saw the problems of I mean, we saw two things, like me and Jonathan, right? We saw that uh I mean we saw that language models were becoming more and more um yeah, they were more and more investments into language models, right? We were doing a very small thing that was oh, like uh Google was investing 5,000 and 500,000 and 5 million into training runs. We can't compete with this long term. So we wanted to ride on this wave that we thought it would become.

Henrik Göthberg 12:58

Immediately you realize the trajectory of oh, it's gonna be bigger and bigger, and they're gonna invest more and more. So we need to build something as our product that becomes better when they invest more. Yeah, exactly. Exactly.

Anders Arpteg 13:10

And this was approximately which year?

David Wallén 13:12

I mean, this is like 2022. Yeah. And uh then the other realization was that um when we bring like bringing an API to market, and this is like a bit controversial as an engineer, bringing an API to market, you're not really solving the real problems. Like uh in come like engineers at companies, they do something that then solve a problem. And they do something for a domain expert that then solve a problem. And so uh in most cases studies. Uh and um so we wanted to take something and like build something for these domain experts where they could have this technology. So the the two insights is basically like ride on the wave that is LLMs and build an interface for non-technical people. Uh, so it resulted in a model agnostic platform uh for non-technical people. And of course, we've added a lot of stuff now that also makes technical people love it. But that led us into the public sector because we also decided to solve the compliance challenges that we see and uh like the security aspects that come with that as well.

Henrik Göthberg 14:13

That is logical because then you're going down into the problems that are not generic enough for the big actors to go worldwide on it, but it becomes like almost like a local Swedish problem. Yeah, and within that context now, the the the the dock pond is different. Yeah, exactly.

David Wallén 14:30

Yeah, and then that's when you launched Intrigue or Yeah, and that's so exactly our 2023, because we narrowed it down to the public sector at that point and said we're gonna we're gonna solve all of we're gonna be good in these processes, like this, because uh the public sector, it's not one value chain there. So there's hundreds of value chain. A municipality, for example, they have hundreds of processes that they serve the citizens with. And so that is the perfect breathing ground for a general platform that also solves compliance and security things uh or problems, and then it we we just we just expanded from there, like we solved their problems, and then we realized that there's adjacent industries that also have this problem. So now we're actually building for I mean Europe's most important organizations is how I choose to word it. Um, so it's um I mean public sectors.

Henrik Göthberg 15:20

But but it's it's to find the to find the niche here, it's it's like public sector because it's it's the flavor of very very many different processes that ultimately is of high and low decisions and very operational that ends up on a poor little municipality or it ends up in one governmental department, and all that work needs to be done, and they have an underlying foundation which is generic around compliance and security. Yes. So if you do that really well and manage that smartly so you can go in and set up these unique workflows without redoing the hardcore stuff every time. Exactly. So you started intriguing.

Anders Arpteg 16:04

Where does the name come from, by the way?

David Wallén 16:06

Um yeah, so it's uh I mean it's um it's a kind of I think it's an abbreviation of intricatus, which means like um complexity or something like that. Uh I've never been uh like heavy on uh heavy on the names.

Henrik Göthberg 16:19

Uh like uh something that you could get a com or an SC address on.

David Wallén 16:24

Yeah, that that that's a good uh that's that's about the benchmark.

Anders Arpteg 16:29

Okay, so you launched uh Intrigue, and if you were to just mention what what's basically the the mission of Intrigue, you could say.

David Wallén 16:36

So the mission is this has been uh this is um

From Machine Learning To Intric

David Wallén 16:40

been the same the vision is from uh the same vision when I like found artificial intelligence. I realized that uh I want the vision is like we want to elevate a human species using artificial intelligence. And that's it it's a it's a big uh burden to shoulder, but I think it's a responsibility as an engineer. Uh you have to uh it makes sense to shoulder that responsibility because AI can have systemic impact, so I want to work for that it does. Um, so yeah, yeah.

Henrik Göthberg 17:14

You want to work for it that it does in such a way that humans is elevated rather than humans disappear.

David Wallén 17:21

So humans should be able to like basically like we should be able to jump higher, we should be able to see further, we should be able to build more, we should design more beautiful, like everything should just be better. And AI is the enabler of that.

Anders Arpteg 17:35

And then you start to build some kind of product as well, yeah. Right? And can you just explain what is the product and what's the main functionality it has?

David Wallén 17:42

Yeah, so uh I mean it's a platform. So a platform, in a sense, we call it actually a workplace assistant platform. That's the category that we uh we have defined. And so platform for um from the perspective of that you build things on top of it. Right. Uh and assistant because it assists you in your work, it doesn't, I mean it doesn't replace you, it assists you, and workplace because it's a professional thing. So it's for work, it's not for play. So and what it does is that it allows our customers to build assistants. We we choose to call them assistants still, we we refuse to call them in the GUI agents because I think that's a hope. I hope that's a hype word that kind of dies down after a few years and people just realize that AI is everywhere, it just assists you in work in various ways. So we don't want to go jump on the hype train too much there. And we also I also want to infuse like organizations with the idea of AI is not there to replace you. I think that if we're gonna survive as a species, or if people who work are gonna have a better future, they should adopt the mindset of that to use AI to assist you and not to replace you, because then you will come into a mindset of like, okay, how can I be better? How can I use this? And not from a place of fear, like, how will this replace me? Because that's just a bad mindset, right?

Henrik Göthberg 19:05

Like to have I I think this is very important because it's the whole discussion of replacement versus augmenting work. Yeah, and I think one of the major topics uh why this becomes so uh bad is that I mean, like if if we have drawn out roles where literally people in their current role job description is a cog in a machinery, is that even a humane job from the start? It's it's like treating a human as a robot, and from that perspective, uh when we are now elevating human work with assistant work, we need to elevate also how we define the role and what the human job is. And all of a sudden, we are not the machine, we are not the cog in the machine, we are the supervisors, operator, uh stewards of the machine. And all of a sudden now it's not the same process. We haven't replaced a person with AI, we have augmented the human cognition and work uh with assisting enabling technologies. So I think it's the the replacement sounds like you're gonna do exactly the same process and you're gonna replace one cog, human cog, with it with a machine cog, and that is the wrong way of looking at it.

David Wallén 20:21

Yeah, I totally agree. I totally agree. I mean it's um it's uh when when technology that's this transformative comes along, it's very naive to think that we're gonna work exactly like this, but just faster. Uh, but that's often how you start. That's often how the transformation journey has to start in order to build understanding. That's a linear thinking. We can't deal with exponentiality really. Exactly. Yeah, I agree. Uh, but you want me to circle back on maybe like, okay, this workplace assistant platform, what do you do with it? Um, thank you.

Anders Arpteg 20:52

So I was perhaps also thinking, you know, you need to build a product that you can sell. You're a startup in some way, right? Yeah. Did you take any funding by the way, or have you did it more organically?

David Wallén 21:00

We've done that recently. Um, so actually, the whole point of, I think the whole point of the business is to build something like very sustainable. And I looked at this goes back if we we go some some more history. Like I looked at the best companies in like 2020 when I started thinking about this, you know, like or putting it to work. I looked at the best companies in AI, and they were what they were doing is they were taking in external funding and then they were burning it in all sorts of directions. Like uh, I'm I'm speaking about open AI here mainly. Uh, and uh I was like, how is this long-term sustainable? And I wanted to take a different approach, so I wanted to bootstrap because first of all, I wanted to prove that I was good enough, and then to be able to fundraise or expand from a more long term or powerful position, if so to say. And uh so it led me to do that, and uh Jonathan really shared this vision. I think it uh uh played out great for us because then when you work also with the with the public sector or being Enterprises as well, they want to know that there's underneath a sustainable business model. So that that is um yeah, my approach to it.

Henrik Göthberg 22:09

So your reason for funding was not really seed funding, but it's scaling funding, yeah. Exactly.

David Wallén 22:13

Yeah, exactly.

Henrik Göthberg 22:14

So yeah, exactly.

Anders Arpteg 22:15

Okay, so how for other people that may be trying to start a startup and and they want to understand how to get started to get that initial funding. Can you perhaps elaborate a bit more how did you start up to to build some product that enabled you to actually get this funding? Because you you need to also have something to show and a trustworthy kind of yeah market going forward.

Henrik Göthberg 22:36

Yeah, well, what's your bootstrap all about? Yeah, see that.

David Wallén 22:40

So I think I mean the it's I think it's incredibly hard to bootstrap something into the public sector. But I think it if you go about it, I mean I didn't do this way, but there are there's grants or there's organizations that you can partner with, there's there's things that you can do. If you start with funding, there's Vinova funding, there's great things, there's funding to be had. Um, so you can start there, but you you fundamentally need to start with solving a problem. Like you need to solve a problem, and if you solve an important enough problem, someone will pay you money for that solution. And that was your core bootstrapping to try to figure out that narrow problem and get that sold. Yeah, that was the summarization of the Swedish text. I think that there's hundreds of I think the climate for starting new companies now um is uh is a lot better in a sense. I mean, but I wouldn't say maybe it's as transformative as the mass media says it will be. Like it's a market, so like, yes, it's easier to build technology, but that just means that more people are building technology, so there's more noise out there. I mean, you could probably go out on the internet and search thousands of AI platforms that might look like Intrick on the surface, but underneath there's nothing.

Anders Arpteg 24:01

So tell us, you know, why should someone use Intric?

David Wallén 24:04

Yeah. So that leads back to the workplace assistant platform.

Workplace Assistant Platform Explained

David Wallén 24:07

Nice, nice, nice job there. Uh so it's really about bringing the AI models close to domain experts. So I you you create what you create is assistance, and the use cases they fall into roughly three buckets, right? So it's for internal productivity. We can just say like that. If your company organization needs to be more efficient, this is your platform. And the assistance they fall into one of three buckets. The first bucket is like this uh knowledge or information distribution bucket. So that's assistance that help you answer an FAQ or HR document or a policy around uh, for example, how you treat someone, uh uh, how you help a like service worker or something like that. That's a class for your technical people out there, like it's a basic rag thing. And it evolves into the second bucket, which is some sort of analytics or decision support. I should also say that these assistants, they're not like they're not binary buckets, they're usually like they're you usually um open up exactly. Uh but the the second bucket is where most, I think most of the assistants in Intrigue today are like some sort of analytics or decision support. So that might be combine the results of this system with these policies, with this thing that my boss just told me to do, basically. You pull in information, you make some sort of intelligent document, and you make a conclusion for it, or you send it to someone. So that is that. Uh, and that that's the third takes us to the third bucket, which is some sort of process automation. So you can evolve these assistants, they can build plans that can automate entire processes. And they can be triggered by either an API call, incoming information somewhere, or by a schedule. So they can take information somewhere, they can look in another system and do something, and then they can output it either to you, to your mailbox, to Slack, or to another system. So that's the third bucket. And when you realize that when you realize that these three buckets, then you realize like, holy shit. There's a lot of work going in these buckets. Most of my work, even my work as a CEO, falls into one of these three buckets. The only thing that doesn't fall there is some sort of meetings, but we also have transcription in the platform. So that kind of falls into a bucket somewhere. Um, so yeah, that it's it's really up to the customer what they do with it.

Henrik Göthberg 26:26

So the key topic is that you find a way to generalize uh three fundamental assistant patterns. Yeah. And then from those fundamental assistant patterns, you need certain technology traits, you know, a compound AI compound system approaches. Yep. And within that, you can then go in and maybe 50, maybe 80-20, look at a lot of administrative, a lot of work we do in office that typically can can end up in one of those packers. So you have a good starting point to it, you know, from an infrastructure pattern there, what you have in your platform, which means you can go to the focus of actually uncovering the real workflow because the technology, the pattern is more or less fairly generic.

David Wallén 27:15

Yeah, yeah, I think so. Yeah. And um, I should also say that once you like acknowledge that this kind of patterns or this assistance, like when you acknowledge that this is how you can work or this is how you what you can build, then in our like our typical customers, it also comes stuff holding you back because this stuff has been around for a while, but there's many organizations who haven't even gotten started yet. And they're held back by, for example, most often like knowledge that you don't know where to start or how to do it, uh, but also security concerns, like our data is very sensitive, and compliance concerns. Um, so those the security and the compliance concerns and the knowledge, those are things we address in the platform as well. So here's where the model agnostic thing comes in. If you're a bigger organization, inevitably you have some information that you cannot leak to an American company for it for example, for the uh because of the Cloud Act or GDPR, um, or uh you have some internal secret information which basically the CEO doesn't want to leave the basement, for example.

Henrik Göthberg 28:23

So, what you're saying is that in a big enough company or in a sensitive enough environment, you always come back to that you need to have computational governance, you need to have policy steering that is easy to set up and automate as part of the system. So it's there, but you cannot expect the whole enterprise to know all the details of these policies, so it needs to be hidden from the masses, but the vital fuse can set it up easily.

Anders Arpteg 28:49

But I guess an easy way to frame this. You mentioned, I think, a number of functional uh uh capabilities that you can have, but then you can have non-functional capabilities that other companies do not want to build or have a hard time building themselves, meaning security, compliance, etc. Exactly. Can you just mention a bit, you know, you you mentioned the functional capabilities that you have. What about the non-functional uh capabilities?

David Wallén 29:14

Yeah. So, for example, one I mean, one thing would be if you I mean if you take the use of AI models as an example, like you might have a policy, or you definitely have a policy by now. Most kind of organizations have a policy where they say something like these kind of models or this kind of use cases, we are allowed to use an American or an open model, for example. But these kind of use cases we need to stay in Europe, or this we cannot use AI for at all for this and this and this reason. And we allow you in the platform to build out different rooms where these different, how do you say, policies can be adhered to? Uh, so that is, I guess, a non-functional capability. You can think about it as rooms in a house, and you can build a room in well, now we're sitting in a basement, so this would be the most secret, tightly locked down room where here we could only use on-prem model, for example, or even in our air-gaped version. Uh, so that that, allowing you to like transcend that. Um, and another thing.

Henrik Göthberg 30:16

This is painful to build that scaffolding and sort it out. Yes. But when you have it, it's a non-functional requirement that probably plays into many different workflows.

David Wallén 30:26

Exactly. And another thing that is really painful to do is to build adequate logging and transparency into the decision processes. Exactly. So we have uh beautiful legislations that are coming that demand dependent on how uh like what risk category you fall in, or dependent on also, like some organizations just choose to do this from their own free will, but you need certain levels of transparency, or you need certain levels of human in the loop, um, certain level of logging, different retention policies, and all of these things are super painful to build uh and maintain. Um, and so we build those things as well.

Anders Arpteg 31:04

So, okay, so policy you can choose and customize basically when you should do what and in what way. You also have some logging to be able to be more compliant and easily comply with coming AI acts at some point.

Henrik Göthberg 31:15

And documentation needs and all that.

Anders Arpteg 31:16

And I guess you have also some security features there, or thinking depends on what you mean with security features if you elaborate a bit. I mean you can take uh authorization if we start with that. Who should have access to what? If you do integrations to your, I don't know, your HR systems or whatnot. I guess not everyone should access everything.

David Wallén 31:35

Exactly. Um so there we usually like uh we usually mirror what organizations have put in place already. Like if you just sync with the AD uh account, you can like basically what you can view in SharePoint, you can view your agents can view in SharePoint via intrigue as well. So it mirrors those capabilities. Um yeah, but there's for sure more exciting things to do there as organizations. Like what we try to avoid, or we definitely avoid, is this like uh I don't sure what it's called now, but the the I think it's called the Lethal Trifecta, uh, where you kind of give some, you give external access and you have no um no policies or no IAM, and then you have uh what's the last thing? I think the last part is uh that you just have an LM there basically. So it allows if you have that, if you if you don't think about this, you could end up in the lethal trifecta where someone could essentially inject prompt inject you to retrieve a lot of information from the outside of the organization. So that we are as far away from that as you could possibly think. Uh, and I think a lot of companies or a lot of organizations that rush into AI, they you could end up there accidentally without even thinking about it.

Henrik Göthberg 32:52

Yeah, I mean like so. Do you uh do we the that's the traffic in terms of access to private data, exposure to untrusted content, external communication keep it being. Exactly. Can we let's wrap up this up because I want to get to the main and I want to get moving to the themes quite soon, Anderson.

Anders Arpteg 33:10

Yeah, but uh yeah, okay, yeah, let's do that. But uh you know, a lot of companies have a problem with um I call it shadow AI, or whatever people call it that. And they you know just copy and paste a lot of data and

Security Compliance And Avoiding Shadow AI

Anders Arpteg 33:20

files into the public external services, and I guess this is something that Intric could help with potentially as well to fight that kind of problem, right?

David Wallén 33:28

Yeah, absolutely. Absolutely. I think a lot of shadow AI arises because you don't give you as an organization, don't give powerful enough tools to your um team members.

Anders Arpteg 33:38

But then you do have access to models underneath that is powerful still, right? So you don't have your own model, you you use the public one available, right? Or open source version, or what do you use underneath?

David Wallén 33:49

So we we our standpoint around models is that we actually don't. I mean, I think we do one thing really well is that we abstract away like the stress of following the latest LLMs, you know, like that. We we do some amount of serving and we put in models that we think actually move the needle, but we also don't weigh in on what model should be used where. So I think that's quite appreciated in in organizations where you have also ethical concerns and also where you appreciate stability, that all of a sudden it just doesn't change underneath your feet what you're using. So we add all of the models. I mean, we even add models from American providers just because the Frontier Labs that's where they end up first. So they end up first in the US behind US APIs, and then they come in Azure hosted in Europe, which under what I call like fake sovereignty, you know, still still an American provider, but it's in Europe. And then at last, like six to nine months later, an equally capable open source model is released either by China or some other company, and uh then we can actually download it and run it in Sweden or even yeah, on-prem.

Henrik Göthberg 34:55

But uh just to sort of wrap up uh the depositioning and understanding of intric, and I'm gonna do this in a in a little bit more pointed way. I mean, like so, in my opinion, a lot of AI investment when when people were rushing into AI 2025 was sort of like enterprises, or maybe even public sector, but I think this is even more enterprises. Oh, we need to get started with this AI thing, and you know they forgot about machine learning and everything like that. And and what we're talking about was different styles of generative AI. And they literally went and bought uh very you know large open AI enterprise licenses. They ended up uh switching on co-pilot in Microsoft everywhere and say, you what about now we have AI? So if you want to contrast what you are actually doing with Intric and how different that is to going down that route, because I think for some decision maker, it's not so clear that this is a quite big difference. How would you how would you sort of how would you meet me when I say, Oh, I don't need you, I have copilot? Yeah, would you how would you you know meet that discussion?

David Wallén 36:05

Yeah, so then I would ask, first I would ask if you're um at first I would probably ask like what uh what kind of data are you processing now? Uh and then I mean if it's just like summarizing a document with open information, then I would say, okay, uh yeah, sounds like that works there. And then I would ask, is that everything that you think you would do for in the coming future? Like, is that everything that your organisation can do, or do you think you could do more? And then is that real workflows or is this tasks of individuals? I mean it depends on what you reply, right? But probably it's like, yeah, once I understand that I can summarize a document, I also realize that I could be comparing this to our internal strategy, or I could be case handling uh like a building application, and very soon you start to get into either an integration with some bespoke system or treating sensitive information. And then you run into a problem with these kinds of systems. The large ones that are built for the entire world to use, they are not catered for the specific needs and they are not able to adjust to regulations and I would say like customization that different organizations have. I mean, they it took co-pilot, what did it take? Like three years. I still don't think that in the basic version you can attach a file, right? So like you've been able to attach files for years to intrigue, and you know, it's this kind of things. Like it when the giants move, they move, but it's incredibly slow.

Henrik Göthberg 37:39

But and isn't it all about having something that caters for the ubiquitous use case like compilot? Exactly. Means that it needs to be super broad in terms of uh versatility, which means it doesn't really have a harness, it doesn't have a framing that makes it, you know, all those things you can't use that in a specific workflow because then you need to put in the ifs and buts, you need to be able to put the scaffolding around, the harness around, the policy around, and how to do that, and it's not really part of what that system is designed to do. Exactly. So you you end up doing a lot of things to the co-pilot setup that the co-pilot system is not designed for, simply because it's it it solves a different use case, in my opinion.

David Wallén 38:26

Yeah, and that is like uh a little bit of the intricate sales speech. It's almost like you're working here part-time or something. Uh but because uh I mean you it it comes it comes as a natural follow-up question on what do you want to use it for? Then you all people, or most people, I should say, they have more complex things, they have more sophisticated things, or they have slightly different workflows that they want to do.

Henrik Göthberg 38:50

They need to live in the real world with the real world fucking messiness. Yeah, exactly. And and and and a universal ubiquitous system is not designed for messiness deviation. Yeah. So then you need to fix that yourself. Yeah, exactly.

David Wallén 39:05

And that we give the people who know what you need to do the power to do that themselves.

Henrik Göthberg 39:11

As long as they know what is the messiness and they know what to do, we make it easy to fix that part. So that is a sure that's the different types of feature set that Intrigue is coming with in order to do that. Yeah. That doesn't simply doesn't exist in a co-pilot environment.

David Wallén 39:25

Yeah.

Henrik Göthberg 39:25

Sorry, I'm selling for you. No, I know, but I think it's uh it's accurate. But I don't think people distinguish it's apples and bananas. I love apples and I love bananas, but I don't really want apples in my banana split.

Anders Arpteg 39:39

But but just to challenge that a bit as well, I mean, I think everyone is working on how to build agents these days, and I guess Intrigue is is categorizing its itself to be an agent builder in some way. And then uh we can think about you know, all the big enterprise companies are, you know, Microsoft R, Google R, and Azure and Amazon R, building these kind of agent-building environments. But how will I mean they have huge budgets, of course, and they have you know so much distribution, they have the users already, and the other tooling and integration into office and so many more things. So I I've been thinking I shouldn't ask this question, but I I couldn't stop myself here. But given the mode that they do potentially have there, there with how easy and well integrated it will be, how can what will make intric unique in that aspect?

David Wallén 40:34

Yeah, so I think the the main differentiator is like in the numbers, um, or numbers, I I can talk about that later, but the the real thing that we have managed to that we have managed to crack or seem to have managed to crack is that we can meet the people who are trying to adopt the AI in their work, you know, like so so giving people interfaces that they can work with and that they can mold, that seems to that's how you get usage. Because I mean in the world of software, anyone can build anything super quickly. So you you have to first of all build it very fast, but you also have to build the things that that um the end user likes to use. And um, I mean you can speak if you just take the fact that for a large a large part of the world, they can't even be using those systems to begin with. That is just starts to that that that sets the playing field a little bit for for compliance reasons. But then when you want to when you want to have when you want to have a integration to a specific system, or when you want to slightly mold it in a way where you want to like how how can I say this? Uh when you want to build a workflow and you don't want to build have a like very complex drag and drop thing that like engineers, uh you need to be an engineer or a PhD to solve, that is like the differentiator. So um, I mean when you look at like retention numbers, like the retention numbers on an organization that rolls out intricate versus someone who rolls out co-pilot, there's 20 times higher retention rates on intricate versus co-pilot. And that that's so this is something like it's bad to say on a podcast, but you can't really speak about it, you need to feel it. So it doesn't really work for a podcast. But do you do you understand what you do you understand what I mean?

Henrik Göthberg 42:32

In a practical sense, in a real-world sense. Yeah, is okay, but this is also something copy copy, but maybe this is also about how you went to a niche to really fully understand a very niche context of where you want to do. Yeah, so you're not trying to sell solve everything, you're trying to solve three buckets really, really well.

David Wallén 42:49

Yeah, exactly.

Henrik Göthberg 42:51

Okay, I'm gonna stop here because I you know we I really wanted to move into what I think is the main theme. When when we had the conversation, I was excited to get you on board. I mean, like knowing more about Intrigue, I think it's a super interesting product, and and uh and we were you know you are pre-Chat GPT. Was it 23 or 22 now? November uh 23, right? Yeah, so that's so so you saw this. Well, we we saw this in this board as well. We were we were full following the benchmarks on the GPTs, blah blah blah. Nice, no one else was. Um so this is

Why Copilot Is Not Enough

Henrik Göthberg 43:26

cool. But now I want to move into the sort of the the meat and potatoes, what I think is the main theme, because one of the key things you did early was to actually to focus on real problems which are very much part of the public sektor core fundamental policy problem working environment. So it's so we so we put the theme as navigating innov AI innovation in the public sektor. And of course, that's a theme that has two underlying main perspectives. I mean, like how. How did you go about navigating as a startup to find the customer, to get through procurement? My own personal view is like I'm not touching public sector because they are too hard to work with. I don't know how that works. And you are thriving in that environment. So that's sort of the startup angle. How do you navigate? And then more importantly, you just released a playbook. I think you've done more about more or less 50 AI journeys with different types of anything from municipalities to some serious big agencies. Government agencies. So 50 AI journeys of different projects in public sector. And by that now you start seeing patterns of what works and doesn't work. So you decided to publish a playbook on that. So that is the you know to explore your navigation, but then explore the navigation of the public sector unit that needs to do something or wants to do something. Could you start just elaborating? You know how did you actually succeed as an as a startup and and scaling up? Like what was your main approach going into finding customers? And what hoops did you go through in order to land your first customers? What you know, five, you know. So, what's how did you do it practically?

David Wallén 45:31

So, uh, I mean, I think the if you go back to the early days, it's just like uh pick an important problem and show up every day. Work extremely hard and show up every day, like seven days a week.

Henrik Göthberg 45:44

So you found a problem and you found some actor, you you found it a key player, key sponsor, yeah, that you started to talk with, and you you put a lot of effort to build the relationship and understanding. Well, what did you do?

David Wallén 45:55

Yeah, so I mean, um, yeah, that was just um like uh bringing that uh how how how can we say? I mean, it's just like deciding on something that you think is gonna be big. Like it's starting a company in general is a huge bet, right? So like you you have to just take a bet on some problem that you think will be important, and so that was what the bet was. Like text summarization, that's probably important, and then you find someone who is willing to pay some money for that, and then you jump through the hoops, like you just you just jump through the hoops. Okay, what does that mean? It means answering how you develop the software, it means answering how how the system is built, it means building the system, it means all of these things. And then once you have solved one problem in a good way, then if you if you solve it well, people will talk. So that's the beauty of um public sector. They will talk, they will tell each other about that. It's good, and um, I mean, it then you can talk to more people and you say I did a good job here. Do you have similar problems? And then then you just then you just continue like that.

Henrik Göthberg 47:05

Right. So to summarize, you you figure out the process together with one client, and then one of the key things is then reference sales over the community. But I I really want to stop at you know learning the first deal or the second deal. Yeah, how does that look like? Because to me, how can they buy you? Don't you need to have a public tender? Don't you need to do a lot of things? So did you does that mean someone loves your product and then you need to back up the Tate and do the LOU public tender? Or how does that practically work?

David Wallén 47:34

Yeah, so when you're I mean, uh there's what's called direct procurement, where if it's the contract value is below 700,000 per year, I mean you can do direct procurement. They can ask three people or uh or three independent actors, can you deliver the service? Whoever does it at the lowest cost wins. Um, so that's how you do it without a public tender.

Henrik Göthberg 47:55

So there is step one, as a startup, know about the direct procurement law and and and and keep smart on the 700,000. Yeah. That's that's a very rule of thumb, right?

David Wallén 48:07

Yeah, yeah, exactly. Uh and then but then you're asking how do you scale that and how do you grow? I mean, I don't think it's any different in the public sector compared to other industries. Like it's a boring generic statement, but like do a good job, work really hard, and uh good things will follow. The difficult thing comes, I think the interesting thing to speak about is like how do you how do you I don't know how do you how do you do that across more like how do you go from municipalities to government agencies to other regulated industries? That is something I think we are still we're still figuring out.

Henrik Göthberg 48:46

But we have grown to a but it but now you're getting to proper deals where the 700 is not really relevant anymore, so then you end up in a process where even if you build relationships, there there is a bureaucracy here that you still you need to play that game. I mean, like that's and and what have you learned about that to answer tenders in a certain way and and to have a team that does tenders or practicalities of of getting into the tendering approach. Are you there or do you know can you avoid it?

David Wallén 49:15

No, I mean we answer a lot of tenders, uh, and I think I can speak about that. I actually I was just at a at a hearing with the Minister of Public Affairs about how can we improve tenders? Because I think like tenders in general, I'll speak about this. I think it's pretty interesting. Tenders in general are perfect for when you want to buy things that are very static, very static and very feature, very explicit. So it works perfectly for buying chairs for a school or buying pens or buying hospital beds. It's a perfect system for that. But for buying software that should evolve over time, it's really a terrible thing, right? Because like you, if you if you put too many demands or if you put wrong demands, the people, especially buying AI products, how do you it's very hard to figure out what you're gonna do with AI in the first place.

Henrik Göthberg 50:06

I put the use case or the platform on in the tender, right?

David Wallén 50:09

Yeah, or do you put requirements about specific things? Like it's it's super hard to uh to procure AI. Uh and that I don't have a solution for. I just know that when I look at a lot of tenders, I'm like, oh, if we answer this, that means we need to build stuff that is really bad for like the good of the company long term, so we can't answer that. Uh, but I think also bigger enterprise deals follow the same patterns. Like, if you want to land a big deal, there's always requirements on the other side. So fundamentally, it's not that different, it's just that um I mean it's a little bit more rigorous.

Henrik Göthberg 50:45

Is that really geared for innovation? No, it's not investment. That's not that's the main problem because innovation means that you're installing a tool in order to figure out the new doing. Yeah, so the new doing is invention is the inventing of tools, innovation is the new doing of things, yeah. Different, right? Innovation investment is super hard to put in a tender. Per default, you need to figure it out together. Yeah, so that that needs the fundamentally different process. So, and I think AI innovation ends up in that bucket more or less every time. That's why I'm so uh impressed that how you navigated this because the tendering process doesn't make any sense.

David Wallén 51:27

No, I mean, especially especially if you want to build new things, but I think that's also very unfortunate because I mean uh it leads to that you if you want to do new things, it leads to often buying consultancy services, or it leads to companies having business models that incentivizes incremental improvement or over-tailoring, and that is um I think really bad for uh us as a nation long-term, right? Um so yeah, that's uh it's not something for me to figure out, but uh the government. I think they're on it with the dynamic in-shop system and stuff that they're putting in.

Anders Arpteg 52:07

You know, speaking a bit about the highlights that you did find, you know, what did you really come up with that uh you published in the report?

Selling Into Public Sector Procurement

David Wallén 52:15

Yeah, so I think there's one really interesting thing. Uh, that um I mean there's one model. So the the the main finding that I I'm not sure how controversial it is to you guys or to the world of AI, but it's definitely revolutionary to the public sector, I think, is that the I mean uh AI adoption doesn't cost start as an IT project. Um, so you need you need people like the key theme that we've looked at all of the customers, I mean it's biased because they've used our platform, right? But we looked at all the journeys who've adopted the platform in a fast way. So that we try to mitigate for the bias of um you know um uh that we they used our platform and just looked at which ones have done it the fastest, which ones have gotten like uh time to value fast. Uh, and the pattern there is that you have some sort of like um horizontal layer, you involve people from all parts of the organization in a tight core working group, and they together decide on a framework how you build AI or how the AI goes into their specific uh part of the organization. So you could call it an ambassador group or an AI task force or something. When you have that, and when you involve key players from also who are usually blockers, like if you involve a lawyer there, or if you involve a DPO, uh CIO, then it goes a lot faster. And because then everyone gets buy-in from the start, and you can weigh in on the legal matters before they become an issue, or before something is rolled out into production and they want to pull the brakes because they weren't involved.

Henrik Göthberg 53:54

So we what we did now we switched gears a little bit, moving from the practicalities of navigating as an AI startup to actually go into the uh report, you know. What was the playbook? What was the game with the playbook, and then we continue with that? What was what is the playbook? You released the playbook.

David Wallén 54:15

Yeah, I I did it or we did it because we want it's really important for Sweden as a nation to embrace AI. It's one like the government pointed out that we should be the top 10 in the world in our use of AI. That was published uh in February this year. And if we're gonna be serious about it, like we need to do, we need to do a lot in a mindset shift. So um because the technology is there, and um now it's about putting things into practice. So we, of course, who had done these seen this journeys, we wanted to share how it's done. Because a thing of our like some of our most important, or important I shouldn't say, but some of our biggest customers we're not even allowed to talk about. And they definitely don't talk in public about what they do with AI. But if we can share the common pattern in a totally abstract way where we can still give value back to society and to other organizations about how you succeed, that is a thing that it's just a responsible thing to do, it's a reasonable thing to do to like push the nation. It's like to elevate the species.

Henrik Göthberg 55:22

And how did you how did you structure the playbook? And I'm sure we can download it soon. Yeah. But but how what's the major setup of the playbook? How you see the couple of chapters, and what are the key chapters that you are looking into?

David Wallén 55:33

I mean, the one one of the things the themes are how you organize yourself and how you approach. So this is what I said, like the ambassadors and so on. And then it's about how it's actually based on um, I mean, uh summer model, it's about how you adopt new technology. It comes from the ed tech sector where it's about like uh one theme is like that you do things in uh four levels. So, like the first level is about um like taking, substituting uh or augmenting existing. I shouldn't, you should just actually read it. I think I will I will we don't get any reads if I uh if I spoil that one. But it's a four-step process that starts with just incremental improvements and ends in changing the process altogether, and you can look at the intermediate paths yourself.

Anders Arpteg 56:19

I like that because you know a lot of people then uh say that ah, you shouldn't use AI to improve existing processes, you should just do transformational. And I think that's very wrong. I think it's actually good to start to improve existing ones, right?

David Wallén 56:31

Nice, yeah, I totally agree. I totally agree.

Anders Arpteg 56:33

I think that's a big learning.

David Wallén 56:34

Yeah, nice. We can have you in the co as a co-author on the next one.

Henrik Göthberg 56:37

But it's isn't also a learning then that in order those four levels exist for a reason because you need to go about the process, the work is slightly different. Yes. You know, improving is one type of work, reworking uh the fourth level is you need uh it's a different, and maybe it's easier to evolve, yeah, but it's ultimately it leads to different things.

David Wallén 56:57

Yeah, exactly. And I think the why it works is that the like what you I think that if you we haven't done this experiment, but it it would be a very interesting experiment if someone could like go into an area with with where they think they will adopt AI, and then they hypothesize about how that process will look like from like you do it in day zero, and you say we're gonna change it to be this, and then you start incrementally changing and you start learning, and then you see and you follow, you do it incrementally. I I can almost guarantee you that what it turns out to be the augmented process in the end, once you have iterated through, it's gonna be totally different from what you predicted in the beginning, because you have learned stuff. I mean, your organization has learned about the models, you have figured out that this API didn't exist, you have figured out that it was too costly to run the query like that. You have figured out that the error rates compound after the like various agent calls. So all sorts of things you have figured, or you have figured out that this wasn't a valuable business like uh process to automate in the first place. So all sorts of things happen along the way.

Henrik Göthberg 58:02

Um do you are you are you saying that it's actually smart to start small and going? Yeah, of course. Even don't don't don't be shy to start with improvement because if you do it right, it can lead to the really reimagination. Exactly. Yeah. An iterative evolution. Yeah.

David Wallén 58:20

And it's super easy to get stuck in discussions if you're gonna transform the whole process. It's super easy to get stuck in discussions around budget, legal things, but I think it's really dangerous to start with transformation.

Anders Arpteg 58:31

Yeah, it's tricky, it is, it takes a lot of time and usually get stuck. So actually finding short wins or early wins is even more important, and then you actually learn as well what you can use uh or do with AI. Exactly. But it but it's much more strategic to do it in that way.

Henrik Göthberg 58:46

Exactly. But what we are talking about here is not thinking about a moonshot or you know the extreme reformation as uh we will brilliantly figure that out once, but we are thinking about the super fast learning curve. What do you think? Is isn't isn't this like this the way it evolves, even the risk.

Anders Arpteg 59:07

I think a problem is you know, most organizations, especially in the public sector, is super slow to change. Meaning people have their way of doing things, meaning you have internal workflows and processes that is you know even defined by law in some cases, and and actually changing them is extremely hard. Yeah, if you have to wait until they are changed, you will not find value from AI for a very long time.

David Wallén 59:28

Yeah, exactly.

Anders Arpteg 59:29

So I think it's very easy to to see that you need to do it this way.

David Wallén 59:32

Yeah, exactly. I mean, I mean you you worked with the AI at the police as well, Anders. I'm sure you didn't start by changing everything, you know, on day zero.

Anders Arpteg 59:41

That would be very, very bad.

David Wallén 59:42

Yeah.

Henrik Göthberg 59:44

Because you get into the topic, then if you have now completely potentially reformed a business unit, and it can be amazing, but it doesn't hook up to the other processes anymore, then it's a real problem. The real change is then super hard because everything else needs to change as well.

David Wallén 1:00:00

Yeah.

Henrik Göthberg 1:00:01

I get it. What else can we talk about? If you want to summarize a little bit about what what we think you have seen as the key I mean, like we summarize like what what are the key recipes of things of physics? The horizontal team is one of them. Yeah. Do you have any other like standouts that you think this this pattern shows up?

David Wallén 1:00:23

Yeah. Another thing that that would be um, I mean, you do the horizontal layers, you do the like incremental improvements, and then the third thing is like the courage from the leadership. It's not to be underestimated because I mean, some of these things will not work out exactly the way you want it to, and then it's very easy to if you don't have support from the leadership, or if you don't have courage in the leadership that builds a team that fails sometimes, then you will stop or you will not start at all. So the courage doesn't have to come all the way from the top. Um, but that is a very important to be able to figure out how you're gonna do it because it will take slightly longer time or you will take different turns. So the courage in the leadership and another thing uh is like that doesn't it doesn't say expect explicitly in the playbook, but to navigate these transformation journeys with like a compass rather than GPS coordinates is something that I usually like to like to say. So uh, and that means more practically, like, don't tell your team exactly what tool to use, exactly what library, exactly what models, how you're gonna do it. Like, say we're gonna use more of this technology to achieve this result, you know, like set the direction.

Henrik Göthberg 1:01:39

Um, so because some of these things you can't really detail and plan them, it's not a rollout, it's a navigation of uncharted waters, and then you you have a compass direction, you don't and then you need to adjust exactly. Yeah.

Anders Arpteg 1:01:55

Awesome. So you have started to write this playbook, or what's the current status of the playbook?

David Wallén 1:02:00

Yeah, no, it's released. So you can download it and um yeah, read you get some. If someone wants to find it, what should they um I mean, yeah, you should uh you should go to our website, intric.ai, and you should uh you should press on the news, I think, or look on look up look us up on LinkedIn and there you can find it. Otherwise, I can share a link link.

Henrik Göthberg 1:02:19

Uh it's I think it's uh organizing AI in the public sector. Exactly. Yeah, is the name of the report. Exactly.

Anders Arpteg 1:02:27

I asked Gemini to is it up uh can you exit it uh online?

David Wallén 1:02:31

You need to put in your mail, I think. Yeah, yeah. We should probably ungate that content actually. I'll uh circle in with the team.

Henrik Göthberg 1:02:38

Um but should we have a new section, by the way, guys?

Anders Arpteg 1:02:42

Oh yeah, right.

The Playbook For AI Adoption

Henrik Göthberg 1:02:44

Yeah, let's do a quick one. Yes. It's time for AI News, brought to you by AIW Podcast.

Anders Arpteg 1:02:57

So we usually have a small break, not so small all the time, but uh small a break in the middle of the podcast to just reflect on the latest AI news that we heard, and uh just go around the table and share some short stories about what we have seen recently.

Henrik Göthberg 1:03:13

I can start. So Intrigue has a report which is not implementing ahead.

Anders Arpteg 1:03:20

That's that's a great point. Implementing AI in the public sector. So that's something that yeah, everyone should look at. I think Daniel, do you have any some news that you recently seen in least in the last couple of weeks or something?

David Wallén 1:03:34

I mean, yeah, sure. Um, so I learned, I mean, I saw that um like Cloudflare started like creating AI agents that like navigate the internet, started to create accounts and like deploy apps and build stuff. So I I I'm not sure what the like and purpose is. Um, but uh I'm like my reflection on I I'm a little bit concerned about the internet and information in general. I think it's probably not a novel uh standpoint from people who understand LLMs, but uh I mean yeah, I'm wondering what the internet would look like in five years if we will like when it's mostly LLM produced. Yeah. Yeah. So that's garbage on garbage. Yeah. Uh or like if you build a source service, like do you even know if you have real users or like Twitter? What's the or what's the state of Reddit? Like, is it only bots or something? Like I don't even I don't know anymore. Like it's just scary.

Anders Arpteg 1:04:30

I guess like the normal roads that we have, you know, today it's human drivers, but in uh 10 years probably the main drivers will not be human anymore. I'm looking forward to that. Uh as well. Yeah, cool stuff, Henrik. Do you have something? Uh I'll go last. Okay, so I think a big event was the Google I/O. It's basically the big developer conference that Google had that was happening this week. And um they of course released a large number of things, right? You know, everything that the new 3.5 flash model that is much faster than others and still be performing similarly to what the biggest opus models and and latest pro models are using, probably great for intrigue to use as well. Um but you know, and and of course, everyone, I think all the big labs are really focusing, I think, on two things just making AI work for um development, and for agentic workflows. And they're building a lot of tooling around that to to to add, you know, both the agentic feature but also enterprise functionality around the agent themselves. But I think one thing that I'd like to mention that they had it, I think is a bit strange, and it may not sound like a big thing, but I think it is. And they basically call it uh agentic search, meaning if you just go to the number normal Google search and you search for stuff, now it actually will of course use AI summaries as it's done for a long time. But now they are actually building software in real time. Meaning if you uh ask a question like how does uh Protein work. It not only summarized that with AI, it actually builds a software in real time for that specific search to show that kind of interactive piece of content and then throw it away. Yeah. This is actually much much bigger than people think. And let me try to elaborate why. So Sam Alton, even back in 2024, said there will be a time in the future where when we will have on-demand software meaning software that you build because you need it here and now and then throw it away one minute later. This time is actually here now. We actually can so quickly and easily build software that you can do it in milliseconds in a single Google search for a single user and then just throw it away. This is actually a very, very big thing, I think. So we literally have on-demand software now being deployed.

Henrik Göthberg 1:06:57

We joked about this uh thesis and trajectory on this podcast maybe a year ago, two years ago. And now we're actually seeing it. What do you think it enables? It's a different paradigm.

Anders Arpteg 1:07:11

I mean it's so many things. Um, if you just think okay, if I can take uh one minute perhaps to just elaborate a bit more. Uh three minutes. Oh, okay. No, but it's much bigger than people think. If you just elaborate and try to, in Elon style, think in the limits, uh, thinking if it is that easy to uh generate code and build code, meaning um you can do that quickly and then just you know, so cheap you can throw it away. What does that mean for normal software? Some people say, you know, software as a service is dead, I don't buy that, that's a bad thing to do. But what it does mean is that the role I would say between a developer and a user gets merged, meaning that users actually start building software and that normal users, the normal kind of interaction you will see in the future is that users change software. And software will not be the static thing that someone else does, but you will do it as a user. When you drive a car in the future and you use like the like the parking functionality, and you see, I didn't like the way it did that parking. Then you tell the car, you know, go and change yourself, change the the parking functionality so it does like this, and the car says, Okay, I'm done that now. Uh the software is updated and now it works like this. And it's just for that car and just for you, or potentially it can be uploaded so it starts to you know roll out for all the cars. This is a very big thing, you know. The merge of the developer and the user, when the user becomes a developer, is something that we are starting to see. And and I don't think people appreciate the change that that will have in what we consider software to be in the future. So yeah, I I think this is a really big stuff.

David Wallén 1:08:58

Yeah, I mean it's impressive if they served at a scale at scale as well, like that is next level, right?

Anders Arpteg 1:09:05

Yeah, of course, not everyone is building software, but still. Yeah, it's cool. It's cool stuff. So I think this is will yeah, we will see a lot of uh stuff happening here, and what software really means will change because of you know what we're seeing here.

David Wallén 1:09:20

Yeah, do you guys think um do how long do you think it will take before nobody looks at code anymore?

Anders Arpteg 1:09:27

I mean, in some applications where you're there, of course, Lava has been doing that for a long time.

David Wallén 1:09:31

Um but I'm a nobody now, like really nobody.

Anders Arpteg 1:09:34

Yeah, but it's like will no one nobody ride a horse. I I still ride a horse, even though I know I can drive a car or be ridden by a car. But some people will always love to code, right? There will always be some niche use cases, but the 99% of the use cases will not mean that humans will write programming languages and at least, right? In not too far of a distance, I would say. What do you think? Yeah.

David Wallén 1:09:59

Um yeah, I think um I mean I also agree. Like um you could you could sometimes when I try to predict the future, I just think about the past. Like, uh, I mean, still somebody somebody still writes compilers, right? So uh like it just I think I mean your court as well, like there will be some people who still look at code, I think probably forever, but it will be the same, probably the ratio of the people who write compilers uh to the people who write normal software will be similar to how many people look at normal code or look at code at all versus someone who just prompts.

Henrik Göthberg 1:10:34

There's always some nerds, yeah. Yeah, a couple of quick more news. We have a deal on the AI Act simplification. Did you did you follow that? Yes. So it sort of says uh a couple of things. It came through, I think, last week. Uh, from 2nd of December 2027 now, for EI system with a high risk use case, is one of some of the new things will apply. And some things from the 2nd of August, I'm not gonna go into details. And then not only did they sort of clarify and let us go on certain things, they actually increased and put more clear bands on what is called nudifier apps and stuff like that. So it was one thing taking where it was unclear, leaving giving us more time to get it sorted, but at the same time adding stuff more sharply that wasn't there before.

David Wallén 1:11:27

Yeah.

Henrik Göthberg 1:11:28

What do you what do we think?

David Wallén 1:11:30

Good work? I mean, I think it's a bit emblematic of uh the EUA Act. Like this uh this omnibus uh is what it's called. Yeah, uh, I think it's emblematic to the EU and regulation, right? Like we I mean, it's a little bit like I am pro, I should just say my stance first, I'm pro some amount of regulation because I think the the purpose of regulation as I've interpreted the UA Act is to ensure that there's no dystopic stuff, right? Like this really is where the foundation comes from, so that there's no citizen surveillance and so on, and that is good. But the way unfortunately it was formed during the merge of foundational models. So they had to slap on a bunch of things that made it super hard to follow and even harder, like to hard to understand, and even harder to like implement and live up to. So I think it's emblematic. Like uh, I mean, when I looked at the first stab of it, I was like, how should you able to how should we be able to comply to this? Like copyright, you're you were supposed to like give copyright attribution to how you trained LLMs. Like I see you smiling. That is just impossible to comply with.

Henrik Göthberg 1:12:40

But it for me, this is always back to the topic that the regulation and the uh to have regulation is great, but we are really, really bad on execution to work on. Exactly. I mean, like step number one, which we have talked on this podcast, is not a very good idea when it's only lawmakers and policymakers, no engineers, no one who knows what it means in practice. How to when you separate the policy making completely from the practicality of what to do, it won't fly. So, this is one problem that it's simply the wrong mix of people. We say horizontal team, you know, is it's a good idea. It wasn't horizontal and broad enough, whoever came up with it. This is number one. Number two is it's very problematic when we come in uh uh with policy but no harmonized standards. Because harmonized standards is the main thing. I don't care about the regulations. Tell me what I'm gonna do in order to be okay, and then I will do it. But when you can't define that, you have a problem. Sorry, that's my banging because I'm annoyed. Yeah, so it's it's back to the execution, right? How you can have such a distance between policy making, naively think you're done, then comes harmonized standards. Same problem in GDPR. We should have learned better. So it's a more like you know, it's kind of a package, dude. You know, you you do the first thing, you do the second thing. It's like there's two sides on a coin, you can't really release a coin in the market if it doesn't have two sides. It's that simple to me, right? So it's the way how you are completely neglect neglecting shit or the most important things to adhere to the law that makes that the first part use fluff, innovation theatre or regulation theatre. That's that's sort of my so the angle here, this is great. Omnibus for me then used to have the conversation, used to put the clarity, you know, put the spotlight on this blind spot. It's what dude, you haven't thought it through. That's what Omnibus did to me, which I thought was great.

Anders Arpteg 1:14:37

I think it's important to to uh to real make really clear that you know regulation in itself is a good thing. It's good we need and the intention of the AI Act and GDPR is good. It's good. It's more the way to execute it and make sure that we have a fair chance for companies to comply with it.

David Wallén 1:14:53

Exactly.

Anders Arpteg 1:14:54

But I guess also that's one of the USPs that Intric can help. Yeah, right.

David Wallén 1:14:58

Exactly. I mean it's very like um it's very it's a tough situation, right? Because I mean the the essence of them are good, the follow-up and uh but I think I also have to give some credit to the AI office who is supposed to like follow up on this because one of the main things I said, I was visiting Brussels when they were stipulating this. I was said, please hire some technical people. And they hired a bunch of technical people, like there's actually really good people in the AI office now, and uh, but of course, like there's a lot of momentum on this, so it's a like uh it's hard to catch up now. Uh, but I think the omnibus is emblematic of like we did some stuff and now we need to okay, now we need more time to figure it out. And I think that it may it's really good that these things get through and that it goes reasonably quick to like realize okay, there's not enough time to understand how we should implement it, and uh if anyone's gonna comply, we need more time. So that's good that that happens.

Henrik Göthberg 1:15:50

But it's is it it's because okay, the intention is great. Is it is there a different way of doing this? Because what we're dealing with here is also the difference that you said before, like we talk about procurement. This is not a fixed share with a very specific frame on it. I mean, like if you take a lot a lot of EU regulation about how you put standards in electricity and stuff like that, it's a very clear and framed environment that is then you can go about regulation in one way. When it's a very uncertain and fluid environment and highly rapidly evolving environment, I don't think you can use the same fundamental process. So this is maybe another thing.

David Wallén 1:16:29

Yeah, I disagree. Because like I think it's just a pro it's um it's just a consequence of how well we understand the technology. Like if you compare electronics, like anyone, you you can have a certain standard for like what current rate, what rating is this cable rated for, right? And how much EMI can this circuit board emit. And since we understand that relatively well, maybe not EMI, but like at least how much heat current through cables develops, it's very easy to say, of course, you cannot run 20 amps through a like 24 AWG cable because it's gonna heat up. But like when you didn't understand the current, when you don't understand current, it's hard to say how many amps can run through it. So now we're in a state where most people don't understand AI. So the standards that are forming around it is like unclear. But when we understand it, it will be more clear how we should regulate it.

Henrik Göthberg 1:17:23

So I then I disagree to disagree with you because there are two versions to this answer. Yeah. First, you have the storytelling. Step number one of the problem: people who don't know enough about the core underlying technology is trying to regulate something. I concur with that, that that would sum up that it's actually you can disagree. Then we do full circle with the innovation speed of doubling rate and capability, and where the trajectory is going and which one is now happening, and now AGI pops up, generative AI pops up, then it's a very fluid market. So then who can know? No one can know. Yeah, but then who is knowledgeable? Yeah, no one is knowledgeable. So, how do you then regulate? Yeah, because then then your argument that we need to find knowledgeable people flips because there is no knowledgeable people who can anticipate where we're going. So, once again, it's a very fluid environment. So it's not only about knowledge, yeah. Or is knowledge at such a rapid pace that no one can keep track? Yeah, I I mean I guess So that to me is like how we need regulation, right? But what's the process then? I I don't have an answer. I think it's difficult.

David Wallén 1:18:36

I mean, my only statement is that once you it's hard to regulate something that you don't understand. And when we do understand it to a better degree, then we'll be able to more properly regulate it.

Henrik Göthberg 1:18:46

But is that the point that we should not regulate until we understand it? Because that sounds dangerous, also. That would be one statement. Don't try to regulate something until you understand it. No, I think that's also crazy.

David Wallén 1:18:56

Yeah, that's it's crazy. I think we should try to regulate to a small enough degree that we don't limit our capabilities to innovate and compete. Um so that is and where is that limit? It's super hard, it's very debatable. How do you see this as well?

Anders Arpteg 1:19:10

We need uh AGI to solve this. I think we don't need to worry about it. Touche, touche. Perhaps we should move on. We move on. We move on. And um

AI News And The EU AI Act

Anders Arpteg 1:19:20

one question that is interesting, I think, and rather hot topic potentially is the idea of you know sovereign AI. And then there is a lot of definitions of what sovereign AI really could mean. Is actually using you know Google Cloud on-prem, which you can do still sovereign or not? Or if we start there, you know what do you mean when you use the term or or yeah, hear the term sovereign AI?

David Wallén 1:19:48

Yeah, so um, I mean, to me, sovereignty over a thing means control over it. And so when you take that in its purest form, um I mean control data sovereignty then means control over your data. Uh, and so when you put that, so so that means if an organization should have data sovereignty, they should have control over the data. And that means who can access it and what can they do with it. And uh, when if you ask the follow-up question, is Google Cloud on-prem sovereign or not? I mean, then it becomes very it becomes very hard question to answer because anything like you have the Cloud Act, for example, that stipulates that at any point when the American government suspects terrorism, they can say, we need this data now, and they have to comply. So that means that Google has to do everything in their power to give the government that data. So it doesn't matter if it's on-prem or not, or if it's on a cloud that is uh in the EU, if there's an American company that has some amount of some probable way of getting that data, they can get it. So then it's in that rawest form, it's not sovereign. You could, of course, argue that I mean if you air gap your system, if you unplug everything, the Google employees cannot come and get it. But then also it's not a way to have a functioning system because then you don't get security updates or you don't get any patches or improvements to your systems. So that that's what sovereignty means to me control over a thing.

Anders Arpteg 1:21:25

Well said. And I guess it's different levels of sovereignty as well. Uh you can argue that as well. So okay. But anyway, I guess in uh in the European aspect in the public sector, especially, this is an important question. And what I mean there is a problem, also, of course, that if we need to have uh an application running uh that is perhaps dependent on intric, and and you need to interact with or in uh uh integrate with a lot of other systems and they need to run on the latest type of functionality, then of course the big cloud providers have great functionality there. So there in some sense you could argue that there is a there is a balance between what quality of functionality do you want versus what is the quality of control, sovereignty control in some sense. Would you agree?

David Wallén 1:22:17

Is that yeah, of course. It's a it's a big it's a it's a trade-off for sure. Like you could you could reason about it in if that if you are some organization what want total sovereignty, then you should do everything yourself. You should not work with any vendors, and you should hire people to do everything in your business from uh A to C. And that is, of course, not possible as a people in functioning business, that's not possible. So it's a it's a question of choosing it's the build versus buy problem, right?

Henrik Göthberg 1:22:49

But it it's also around understanding what your real value is, and understanding how to maximum your value, and what is your uniqueness versus, you know, you know, what do you outsource in general, etc. etc. Exactly. So it becomes a more fundamental topic of what is most important and kritical to you and to your fundamental business.

David Wallén 1:23:13

Yeah, exactly. And my I can tell you that my firm opinion is that whenever there's a technology shift, I mean you can see that whenever there's a technology shift, I think there's a tendency to think that this new technology is so transformative, so we have to do it ourselves. I mean, if you do too much yourself, you will lose, you will it's an opportunity cost of improving your processes or making sure that you work on applying this new technology. So you could uh I I think it's fundamental, like the businesses who succeed are the ones who figure out how to use AI in the best way and to like really solve the build versus buy problem. And I think there's very, very few organizations out there who should build a lot when it comes to AI.

Anders Arpteg 1:23:58

I I think that that's a different thing with AI, that it is surprisingly difficult to do, and of course, building foundational models is very hard for most or almost everyone to do.

David Wallén 1:24:08

Yeah, exactly.

Anders Arpteg 1:24:09

But but what's your experience then? Having worked with public sector organizations here, uh how how has that been? What's their thinking if you were to share about you know sovereignty?

David Wallén 1:24:21

Yeah, so um, I mean the the fun the fun perspective that I think don't not so many people think uh or anticipate is that there's not really a big consensus on how you should do it. Uh like I work with companies of all sizes, you know, across um across Sweden, and there are small organizations who go full on-prem, and there are big organizations who go full, you know, uh American clouds, you know. Uh so there's I mean, but I think there's in general, if you compare public to private, it's more skewed towards doing more things yourself or being more on-prem to that sense. Uh so that is that is the movement I see. But I I and there's been also a tendency to fall into the trap of building a lot of things yourself initially. But I I know that also I myself as an engineer, if I worked at one of these organizations, how can you not want to work with this technology? So I think it takes really like a discipline for the leadership to realize like, where should I put my technical people? Uh, and so that is, I think, I think I see it evolving now. I think a lot of organizations have built some simple rag stuff, they've gotten into they're seeing the pains of maintaining that, or they're comparing their simple rag stuff to buying a system, uh, and they're seeing like holy shit, we can't keep up. And uh, yeah, that that is that is a shift that I'm uh seeing now.

Anders Arpteg 1:25:52

I mean I think that's good that you say that because some people that are here they just keep talking about you know, oh, it's just a rapper about an LM. And I think it's such a wrong way to phrase it because that is the product is the thing around the LM.

David Wallén 1:26:05

Yeah, exactly.

Anders Arpteg 1:26:06

And that is much more difficult than people think. Yeah, and to keep up with what the big players are doing here is super hard, yeah, right?

David Wallén 1:26:12

Yeah, totally. Yeah, totally. Yeah.

Anders Arpteg 1:26:15

And and I think you know, sovereignty, you can also think about you know what is really the the the dangers of not being sovereign here. Uh, I mean, of course, we can think about the economical impact. If Trump suddenly increased the tariffs, of course, that could be a really bad thing. It could be that you know someone is saying, Oh, we actually went into war, they actually turn off the cloud in some way. That could be really dangerous, of course, as well. It could be the cloud acting, and then people actually access your data or even try to steal your IP. And there's a number of things that could happen here, you know, if if uh we don't handle sovereignty properly. But okay, so what's your what's your recommendation, I guess, in the end? And then what should we do? If you were to be more of a policymaker here, trying to guide how public sector should operate, of course, using Interk is is given. But besides that, where do you think the proper way to manage sovereignty should be?

David Wallén 1:27:14

Um to manage sovereignty. I mean, I think that um in general, I can speak, I think I speak for all of Europe first in that I think it's a really it's a bad pattern if you're if a thing like one thing that could happen is like a scenario. I I spoke to the head of the army uh like uh a week ago, and he said that one thing that happened in in Ukraine is that I mean when they there was political pressure, they turned off Starlink and they start to look like lose the war, you know, and I mean not ex explicit those words, but like that is when you're dependent on one person or one organization for things that can move the needles of nations, that is I think a bad pattern.

Anders Arpteg 1:28:00

So I think organizations and say that you know the one thing Ukraine did when that war broke out was turn to the cloud. Yeah. Yeah. And the first thing the Russian was done did was trying to sh you know to bomb all the data centers that they had on-prem.

David Wallén 1:28:12

Yeah.

Anders Arpteg 1:28:13

So the on-prem was completed insecure. It's no way to have the robustness.

David Wallén 1:28:17

No, exactly. So my my point is not like go on-prem, definitely not, but more like about how many points of failures you have. So you should try to have not so many single points of failures. That's also like I have an amateur interest in aviation, and you should say that you should minimize the amount of single point of failures. And so with the sense of sovereignty, like you have to, but you have to take it to you can't take it to the extreme. Because if you take it to the extreme and say we should have no external vendors, then you will have the most inefficient slash dysfunctional organization ever. So you need to find the balance there. And it depends on the balance depends on how big you are as an organization, what capabilities you have, what teams you have, but also the amount of services that you provide. So if you provide the diverse set of services and you don't have technical people, of course, you need to buy a lot of things.

Henrik Göthberg 1:29:05

But I think there's also something here about to sharply invest smartly in this layered cake of sovereignty. If you think about real control, right, geopolitical of energy. If you don't have energy, and then it goes to the chips, semiconductors, and then it goes to the infra hardware, and then it goes to the operating systems. And somewhere here we have LLMs, like foundational models, and then we, you know, and then we have the data, our data and control of our data, and then we have you know, whatever we can build in the application layers in our local models and stuff like that. So isn't it also to think very, very carefully how and where we put our money to keep maximum control and where basically we work smartly with other technologies? I mean, like an example, you can build, you can, in some ways, when you build your own software, in the process of using software, you can have in your dev environment, oh, I'm using clawed code, I'm using American tools, blah blah blah blah blah. But when I run and compile my code, I take that home, that's in control, I put that in an open, you know, in a Swedish data center. So were you completely sovereign? No. Were you sovereign where it mattered? Most likely, right? So you see what I mean? So we cannot we don't have the investment capacity to be sovereign across the whole cake. I think that's fundamentally impossible.

David Wallén 1:30:38

Yeah. But I think and I think that humanity, like, we should thrive to collaborate and we should use tools that we build on in other parts of the world. So, like, on a if I speak to my humanity, the human side of me wants to say that sovereignty shouldn't be an issue because we as humans should collaborate. And that is what has made us like a good as a thriving species, is that we can have narrow professions and we can use currency to trade goods and services with one another. So, like on a fundamental level we shouldn't have to care about it. It's tricky.

Henrik Göthberg 1:31:13

It's very tricky. Agency is better. Like to have you know, have autonomy and agency to each other, but okay, but let's but maybe there's some places where you want to be really really sovereign. Yes, but where? You know what is your view? You know, is it is it data sovereignty that is the most important? Is model sovereignty important?

Anders Arpteg 1:31:34

Even no, I think you know what we should focus on mainly is thinking how we can scale value in Sweden and in Europe. And then you have to weigh the risks of not using the greatest functionality versus what the real risks of sovereignty could mean. In most cases, I think it's worse to not be able to use AI than being afraid about potential abuses of sovereignty resources. Yeah, so I'm actually rather strongly in favor of being able to build and stand on the shoulders of the giants and actually form partnership like with Intrigue and other companies and avoid building stuff yourself. Yeah, so that that I think is something that most so many companies do wrong.

Henrik Göthberg 1:32:16

And and what you're saying now, innovation speed matters most in some ways.

Anders Arpteg 1:32:20

I mean, like so to be able to innovate with speed and collaborate and I'm gonna focus on the core functionality of the organization, do not build a new data platform, it's that's stupid. Sorry, frick, which everyone is trying to do. Do not build a new rag system, you know, that's what everyone is trying to do. It's stupid. We have great partners and functionality that they should use. So I think that's that's the important point. I agree, I like it. But continuing that, you know, and and of course we have different model providers that you have to use in Indriq and that all the organizations have to choose by, and then we know the best one is coming from US, but then you want potentially to reduce that kind of dependency, and then it comes to what we have in Europe and not at least in China as well. What's your thinking about how it's how the the performance and and the if we should use or not the European and Chinese models compared to American ones?

David Wallén 1:33:18

So, I mean, first of all, the the standpoint we have at Intric is that we want to like we want to be totally transparent and we don't want to weigh in on what models to use to a large degree. So we want to just enable anyone to use what they see fit. And some organizations where it's it's hard to speak about different models when it comes to without bringing up like ethics and so on. I mean it's pretty I think it's pretty common knowledge now that some of the Chinese models have been purposefully altered to forget certain historical events and therefore like produce in like in infactual uh responses. And um so so therefore it causes a lot of ethical concerns, but you can also raise ethical concerns about American actions. And our standpoint as a provider is that we're not gonna weigh in on the ethics, we're we're just gonna let the organizations figure that out. We just want to build the best platform, and for one organization, the best platform can mean screw the ethics. We need an open source model that performs at the top of the benchmarks, and then they should do that. And some organizations might end up in the news if they even if they use an air-gapped version of a Chinese model because the Chinese model could leak data, you know, like um that's so we we should we don't weigh in on that.

Anders Arpteg 1:34:41

Um that's a really good thing from an in-trick point of view. But if you then think of a public sector point of view, um and of course, uh still, what would you recommend uh public sectors to do? Do you think it's a good idea to start using open source ones or closed models?

David Wallén 1:35:02

Or so now you're actually asking me to weigh in on it. Okay. Yeah, I feel like I have to do it then. I I think that you don't have to adapt. You don't you really don't have to weigh different like it? I think I will refer to a statement like this. I think that leaders should try to be brave. Like this is what this is what Sweden needs. We need brave leaders, and every leader needs to assess their own organization and figure out what brave means to them. Um, because that's what I think is really like that makes for prosperous long-term um yeah, good things for our nation. So we need more brave leaders. That's all I uh can say.

Anders Arpteg 1:35:36

Brave can mean many things. But if we just elaborate a bit, and and of course, you mentioned that you know Chinese models, as well as American one and potentially European, can be biased in different ways, they can choose to phrase their values in different ways, and of course, we know that to be true. It can also be like cybersecurity potentials or risks that you know, if if, for example, Chinese models were to be biased towards uh not protecting when using uh AI for code generation, certain type of security holes that could be potential risk, uh and so forth. But then what should we really do here? If we think if we take the other approach, if we forget these kind of things, the bias thing and the security risks thing, I mean that they do exist, but we can also say that from an ethical point of view, we know that the Chinese model, for example, has been distilling from the American ones. So we had a big report from Anthropocriting, you know, 24,000 accounts, fake accounts was created to as soon as they released Opus 4.7, they just hammered it and then they come with a new amazing kid model and minimax model and keep sec model. Oh, it's working so well. Well, no shit then distilling from the American ones. And you know, that's that's still good for European reasons if you just want to have a cheap one, yeah, yeah, right? And you don't care about the ethics, you just say I just want a well-working model, yeah. That is, you know, I don't care why it's working, it's so long as it's cheap and well working, then that could be okay. Um I'm not really sure what the question is here. I guess that the question is should we have the ethics here to choose or avoid choosing models that is you know built unethically?

David Wallén 1:37:22

Yeah, so I think um there's probably a theorem for this. I think that the the the close my view on it is I think um you should take a total uh what what's it what's it called? Utilitarian perspective on it? I think that's a framework to reason about it. Like um I think it would be unethical if if a certain model that is like minus one on the ethics scale, if you use that, it could give you plus five on the good ethics side. Like, for example, like can you give better health care? Could you save hundred babies if you just use a slightly unbiased model? I think it would be like you would want to save a hundred babies. Do you get what I mean? So you should try to make those assessments. It's about how much better of a service can we provide humanity versus how unethical is it? I I'm I'm really bad at these kinds of subjects, but this is the framework I would use to reason about it. And of course, if you sit at a government agency or a big company, then it's an impossible KPI to answer, you know, like how how many babies are we saving in that sense. But that's how I would reason about it.

Anders Arpteg 1:38:34

Well said, I think. And I think we have said so many times, you know, people have been asking not which models to use, but you know, if we should AI or not. And then you know, phrasing the question: if we use it in healthcare, and if we use AI, we save this many lives, and if we don't use it, we we kill that many people. Then you know, I think that could be a well-working type of reasoning.

Henrik Göthberg 1:38:55

It's the utilitarian reasoning, also about how ethical. I mean, like, I mean, like, so it's like how perfect should you be? Like, because nothing is perfect, right?

Anders Arpteg 1:39:08

Then the risk of non-action is something that people should consider so much, right? Exactly.

David Wallén 1:39:13

Yeah, and I think because the non-action is not ethical either. No, I mean it depends on what you're not acting on, but I think, especially in terms of running our country, improving our country, I think we have a history of unaction, and it's a pattern of uh passiveness has been more prevalent. So I think that um that is we need to lean into action. Um, that it's not a very crisp statement, but um I think that's where broad leadership comes in.

Henrik Göthberg 1:39:41

But I I would like to sort of sum up now, you know, a little bit wrapping up also. We had a theme here: how do we innovate

Sovereign AI And Real Tradeoffs

Henrik Göthberg 1:39:50

and navigate innovation with AI as AI innovation in public sector or in general? And and now we are coming to sort of the you know core complications here we have we have explored together with you that we sort of sovereignty and all that, cloud, all that. I would like to think about now what what do we think is the biggest blockers or hurdles or things that we we we would like to sort of see change to increase the speed of innovation, do more, you know, and let's start with public sector, but I bet you it's a similar types of things we need to fix in enterprise as well. But if you if you take a step back and sort of lessons learned, and where are the things that sort of makes us muddle through or makes it go slower or yeah, yeah, what would really help? What would what would make a difference?

David Wallén 1:40:45

I I think that um I mean communication, if I speak on for Sweden's perspective, I think communication, this is might be a bit unorthodox, but I think communications, if organizations communicated a little bit more what they did and how they did it, I think that would help others um adopt as well. Uh because I mean, taking the public sector, it's like I said before, the most exciting things that people do with our platform, we can't speak about, and they won't speak about it either for probably years and years. And that creates it's that in combination with the pattern of like we're gonna look around and see what everybody else is doing before we do something, that creates an environment of inaction where although underneath the surface you have a couple of key players who are really pushing boundaries, like really pushing boundaries, but they don't speak about it. So that and that means that a lot of organizations are just looking around and seeing nobody's doing anything and until years go by and they actually some organizations have been really transformed. Like uh we have some municipalities saving millions in comparing invoices versus contracts, and that doesn't really get out, and I think that is uh that is a pity uh for for other municipalities or other organizations because I mean um yeah, so so that I would like to change. I would like the nature or the approach to communicate around communication to change. Um, that is much more needed than any like legal perspective change or technological shift.

Henrik Göthberg 1:42:29

But but let me try at the data innovation summit a couple of weeks back, there was a couple of key themes that was coming back in relation, and it was quite interesting because it was very much we would got more into how we practically put AI and agents and you know safely into enterprise environments. So the whole dimension with context, the whole much more uh mature organiz uh view on organisation and agency. You know, how do you actually install this? So if if we go into practicalities like the way we have organized uh funding, the way we have business people in one end and IT people in the other end, the way we have sort of not really an innovation steering in place. We have normal business or you know uh verksamhets styrning? Ja. Innovation styrning. So are we really geared to really absorb innovation? Is the funding modell and is the organisation gerade for this? So when you say we want to have horizontal people, does that mean a huge kommiter från people who ser completely different and have no clue about each other? Isn't that indikering hint hint? Are we organised the right way for this?

David Wallén 1:43:48

Yeah, I think a lot of organisations are. They just don't know it yet. Like you have digitalisation offices like CDOs and so on. But many people maybe take the approach of not forming these horizontal layers or forming these groups because they say that we need to figure out compliance first, or we need to build a data lake first, or like a data layer, data platform, whatever you think you need to do with your data first. You don't need to do it. You can grab really easy wins that saves you hundreds of thousands every month on a small team by just simple rag stuff. And you should get those claims next month, not when you have tagged up your metadata uh with consultants next year for millions. Uh, so um that is um I think the the organization structure is in place. We we started with digitization when the internet came about, so that exists, and just those people need to work.

Henrik Göthberg 1:44:45

But but they they are I think you're telling them to do work that is outside the fundamental operating model. I'm not really they are supposed to digitize their organizations and AI is a digital. Do they follow fundamental product steering approaches? Are they really funded according to product and product operating model? I mean, like these are the fundamentals that this guy's been doing for years in Spotify. He doesn't even recognize the problem until he's worked in in a different place where you have run span and project span and you have a line organization, this and that, and and where the fundamental steering model, funding model, is not product in the same sense. So it becomes really hard to innovate then because whatever innovation you do, you kind of need to treat as a product. So those kind of things I I don't think public sector has. I mean, like in some places, maybe like Skaterwerket and different places, they they have this because they have a DevOps environment. Do you know what I'm talking about, Anders? Because I'm trying to reflect back to enterprise view. The pro, I mean, like I think product steering as a quite important view of innovation here. And those things I'm I'm not so sure we have that in the municipality, and that is making it a little bit trickier to muddle through.

Anders Arpteg 1:46:00

I don't know. Well, I mean, the way you're working, of course, I think is a super important point. Now we're moving a bit you know away from perhaps from the topic, but still, uh yes. I mean, if we take the technology that AI can help with is great. And if you have rag solutions or you have AI that can help with managing knowledge and data in different ways, that's great. But uh, even if you have that, the big a big problem that a lot of organizations have is that they don't have efficient enough you know way of working. Exactly. If we take you know what Tesla, for example, is doing, and I've been you know looking and working in Spotify and a Spotify model, of course, they they have a great way to do it. Uh the the big tech giants, you know, the Google, Microsoft, and Amazons have great way of doing it. But when I started to look into how Tesla is doing it, I was shocked, right? Tell them more. Uh yes, we we can do that. But but I'd just like to say that you know, what to your question? Then if you think about you know enterprises and perhaps even more so public sector, there is a lot of things you need to do besides providing the tech, which is related to the way you're working. Yes, that's my point. And um, and I think you know, I wish this should be a science topic. This should be something that is uh teached in in school. This KTH should do this. This is what I'm doing.

Henrik Göthberg 1:47:15

This is what I'm trying to put up as an engineering topic, uh, equally to data science and data engineering. So I fully agree with you. Yeah, and and and there are patterns, yeah, and I don't think those patterns is really understood in enterprise and even less in public sector. And this is also what you know, hearing you know, on data innovation summit, oh, people are so excited, blah blah blah. And then you have a couple of you have a whole stage on the Friday for public sector, and the main topic is more or less oh, as as the general perception of the public servant or engineer working somewhere who maybe has not worked with you.

Anders Arpteg 1:47:54

And just to give an example of this, you know, in public sector, they usually blame and say that yeah, I know the tech companies are doing this, yeah, I know the startups can move fast, but they don't have what we have. We have strict security requirements, we are working with humans in this and that way, and and we cannot work like a startup. That's bullshit. That is actually bullshit, yeah.

Henrik Göthberg 1:48:13

Absolute bullshit.

Anders Arpteg 1:48:14

And and just an example, if you take Tesla then, if you take SpaceX, for example, if you take like rocket science, literally building a Starship rocket, which is launching today, you know, the V3 version, I'm I'm waiting for that. They are working in a super super regulated environment. They are working with hardware development, not just software. They are you know working with oversight committees and other you know, government agencies that has to approve what they're doing to put rockets in space and still manage to do it in an agile way. And I think that's and I wish you know people understood how it is possible and don't use excuses like well, we can't operate like a startup or like whatnot. It is possible.

David Wallén 1:48:58

Yeah, but that's where the brave leadership comes in.

Henrik Göthberg 1:49:01

Yeah, but brave is not enough, right? Because I mean, like Helena Hornebrand was on to this, like it's a it's it's a different view of what is enterprise when you're continuously innovating, and it becomes something else in terms of how you set up steering and organization. So it's it's it's it's down to the point of how we have with a command and control structure for hundreds of years, from university to how we educate people in schools to be cogs in the machinery doing their work, actually not taking decisions, not taking action, but rather being part of a process, right? And now when we are getting to this perspective with AI, and even you're gonna build your own AI, so you're gonna have an assistant, you're gonna delegate, you're gonna steer. So this whole thing of how to think and act like that and have the agency to do that, that's what Elon Musk is giving out in a different, completely different way, right?

Anders Arpteg 1:49:54

And then there's a lot of stuff just to do the disclaimer it's a lot of stuff that Elon is doing that is not something we should copy. No, but it's but it's it's I just put it now.

Henrik Göthberg 1:50:03

This is not an an an an analysis of uh of Elon Musk, it's an analysis of ways of working in very large organizations who apparently moves with speed in some of the most highly regulated markets you can have: automative, automotive, and rocket.

Anders Arpteg 1:50:18

Sorry, back to you. No, but if we move into uh intricate here, and and you mentioned inmates, your mission was uh elevation human species. Yeah, right. And then, of course, also you mentioned about AGI, and and

What Blocks Faster Innovation

Anders Arpteg 1:50:30

we will continue to you know of uh uh improve uh AI, of course. And it's do you have a date in mind when AGI potentially will happen?

David Wallén 1:50:37

It was quite a few months ago since I thought about that, to be honest. Uh that's um that's uh sad in itself. But uh I I I don't think I am uh I don't think I'm in a position to like I haven't trained an LLM in years.

Henrik Göthberg 1:50:53

Um your opening statement was that uh you're gonna build AGI, so tell us what I will tell you the plan to do it.

David Wallén 1:51:01

I can tell you the plan to do it and the time scale. So like I mean with Intric we're building this assistance, right? That uh that is basically allows you to augment or improve human work, right? So the hypothesis is just that at some point the system, I mean, AGI is something that can do any tasks as equal to or better than a human, right? So at some point, to me, language models can now reason and they can like semi-reason. Semi-reason, yeah, sure. Uh but you can let's say like this the trajectory of their ability to reason is upwards, of course, yeah. And provided enough time, that upwards will at some point become that of a human brain, and then it's just a question of like the amount of interfaces that you can connect to an LLM if we still call it that at that point, or if it's some multimodal um thing. But so that we will just follow along this journey and connect more and more stuff to things in Intrigue until we sit there at a point and say, Holy shit, we got AGI. And and that's not a very crisp plan, uh, but it is a plan, uh, and it's much more cost-efficient than uh burning money on compute and letting someone else do the compute. But uh I will say this I think that in five years, it's very likely that we will sit and we say, Holy shit, we have AGI, because we will we don't have the visual input as much. Like we don't we haven't solved that like the moving video, but we have images, like we have images and lms now. That is just has just happened. I think over the like uh past year and a half, it just happened, and the world barely paid attention, but we just need moving videos, and unless Jan Le Kun is really right, uh then we just need moving video into LLMs, and then we have it, right? So uh uh I mean we we'll see how it gets made, but um that is um the thing about Jan Le Kun by the way, is he right or no or wrong? Uh yeah. Uh I I think he I think he's wrong. Um so I think I mean it comes it's like a physical philosophical question at this point. Like, how do you define intelligence? Uh I think you land in a you land in a philosophical um question.

Anders Arpteg 1:53:24

But do you think the current transformer or the GPT decoder version of the transformer is the architecture that will stick when we have AGI?

David Wallén 1:53:35

I mean I would have to then I can say no. I mean, because you need if you define any task as going beyond like text, then of course you probably need other things. And if then I will say you could argue that if you just output things that control other modalities, you could you could say that we still have that architecture, and in a say, in a sense, I I could say also yes, but it depends on a little bit where you draw the boundaries of the system. I think so. I in a way it's a vague answer.

Anders Arpteg 1:54:07

I mean, I think you know when someone tried to reverse engineer mythos, you know the the not released model from Enthropic that uh created all these cybersecurity problems uh or possibilities. Yeah, but uh they in the reverse, it's called open mythos if someone wants to Google this, but then um what they did uh in the reverse version was basically they they had a forward pass that they normally had through the transformer decoder, but then in the middle of it, they basically do a loop so they 16 times just loop through the latent or the middle of the layers inside the forward pass. Okay, meaning basically it does latent space reasoning. Okay, meaning it doesn't output a token and then send that token back and do reasoning in token space. It actually does reasoning because it loops inside the model before it reproduces a token. Um, so that means it actually does latency reasoning, which is what Jan Likun wants to do. Jan Likun doesn't want to remove transformers, right? Or what's your view of what I mean Jan Likun's business?

David Wallén 1:55:12

Yeah, yeah, that's good. Uh we need to land in that first. So, my view of what he wants to do is basically he's saying like text is not enough. We need to take video into account as well to build a world model so that I don't infer any like neural network architecture in his statements there. But uh maybe that's a follow-on fact.

Henrik Göthberg 1:55:33

But it's but we have we have uh Karim Nora on the pod six uh superintelligence and a Swedish startup who basically thinks you you cannot get to real-world intelligence from language. That's the logic, that's the fundamental logic that you you can find. There is some sort of logic on intelligence that is related to language, but real physical intelligence to learn how to stand up or walk or stuff like that, even robots. Um that it's it's a different type of learning that you need to do trial and error. So rather than starting from the language paradigm, start from the paradigm of you know moving boxes and building a robot arm, completely deep learning and and learning like a child walking. So it's it's so it's it's a so it's an argument that there is a fundamental, not the language paradigm that takes us to that the language as is language intelligence actually, maybe not.

Anders Arpteg 1:56:36

That's another argument. What data is used for training? It could be language only, but probably that's not sufficient, and you need something more, at least you know, images, video, audio, and uh perhaps more. Um but the other thing is really the architecture as well. And you know, what Jan Le Kun is saying then is they want to have the JAPA, the joint embedding predicting art architecture, which means that uh it should um embed into a joint embedding latent space and then do the reasoning there, right? So today we have the GPT that is actually always producing one token at a time, and you have next token prediction happening, which is you know, in my view, really inefficient and and actually not what's happening. If you take any kind of image generated today, all of them basically have an autoencoder around the pixels so that it's not really you know reasoning in pixel space. They first auto-encode it into a latent space and then they do the diffusion transformer inside of that. So already we are seeing uh more and more things moving towards uh some kind of latent space reasoning happening. So I don't think it's a question that the direction of what the Jeppa and Jan Le Koon are saying is happening already. Um Mythos is potentially another example of that. There's a lot of papers moving in this direction. Then you know, if um if auto-regressiveness is is uh is bad completely, that's another question. If reinforcement learning is bad or good, it's another question. But in any case, I think in a lot of ways the direction he has been saying has been already happening.

David Wallén 1:58:03

Yeah, and I'm super glad he's pursuing it because for a lot of years he was just complaining about the LLMs and he was also funding them at the same time or like progressing it. So I think it's I'm I'm super excited to see where it goes, right?

Henrik Göthberg 1:58:18

But let's use intricate as a view on the philosophical view on AGI. Okay, because your statement was let them figure out the the underlying models, yeah. And what we are doing, we are building the right harnesses, scaffolding, uh, context all around, and all of a sudden, holy shit, we will reach AGI. So, this to me also says something about architecture. It says something about what is AGI is some sort of AI compound system. You know, if we can get all that inside the LLM, or if this is actually an architecture between what is the world model, what is the perception space, what is the harness space, blah blah blah. Whatever you, I don't know. So I I don't see it as an LLM topic of an LLM reaching AGI. I see it as some sort of architecture as an AI compound system reaching AGI. What do you think about that? And I and I think that's what you're telling me when you're saying yeah, well, that's partly what Jeppo is saying as well.

Anders Arpteg 1:59:19

So it's not a single model doing everything, it's one doing perception, it's another one having the world model, which basically is trying to understand is this good or bad if you make a certain prediction. And there is something a memory model, meaning more for long term and short term, short-term memory. So you have a set of modules basically operating together, and I think also that is basically what OpenClaw is moving towards. So it's actually adding memory by simply having markdown files here and being uh updated and changed and a sole file and whatnot. So we are starting to see that it's actually the surrounding logic harness, and I hate the term scaffolding or wrap her. That's not really it's cheap, right?

Henrik Göthberg 2:00:01

It's I I believe it's A compound system, was uh Berkeley Air Research launched a couple of years back. Ali Godzi was part of writing an article on that, trying to sell data bricks, of course. But the whole point was really if you think about it, I mean, like if if if I squint, what you are doing with Intrigue, you're providing a platform in order for people to put in the right long-term memory, short-term memory, call it whatever you want. When when you build a rag, you give it context, you give it information, you give it policy, you give it all these things that allows it to then act in the real world, which you can't really find in the generic AGI LLM because it doesn't have the eval criteria that fits for that purpose. So you need to be able to work with those things. You see what I mean? So, in my opinion, what you're doing is you're creating a platform that allows people to get to an AI compound system without being super engineers.

David Wallén 2:00:57

Yeah, I think that's a super fair way to put it. I mean, if you compare what a super powerful LLM or other type of neural network can do, I mean, it's still I mean, it's still some person in organizations, you still have processes, you still have customers or citizens to serve, and there's a there we will fill the gap between this super powerful thing and the job that needs to be done. And there will be always be people uh making sure that that gap is filled, it's just a question of what do you fill it with.

Anders Arpteg 2:01:36

I'm moving to that, and now we're the time is flying away a bit here. But if we move a bit more philosophically and I'm thinking, you know, as the journey towards AGI happens, then um we can see at least uh if we take AI for coding, for example, we can see that you know today people are writing less and less lines of code, but they are instead directing agents that are then you know doing most uh a lot of the work, but humans are still in the loop, right, all the time. It's no question that we have humans then just moving up in some way in the abstraction uh

AGI Timelines And Model Architecture

Anders Arpteg 2:02:10

pyramid or or levels and then directing a set of agents. How do you see humans as we move closer and closer to AGI being part of that? Or do you think when AGI happens, humans actually will move out of that loop?

David Wallén 2:02:27

I think going back to my pattern of like the thought pattern is like analyze how technological shifts have happened in the past. That's the framework I draw upon. And I think any tech abstraction that has come, or like um yeah, civilization abstraction has not led to people working less, sadly. I mean it's sad, but it's the conclusion, it's the it's the how it has always been. So I think that I would I would really like that AGI leads to UBI, like some sort of universal basic income where everyone can lay on the beach.

Anders Arpteg 2:03:00

Yellow call it human or universal high income.

David Wallén 2:03:03

Oh yeah, that would be great. Um but but I think there's certain tendencies of us that are that's what makes us human. Like we are we are hunters and we are gatherers, that's why made us the apex predator, right? So we will always, as a species, want more. And so even if we have uh UBI or UHI or whatever, you will always look at the neighbor and be like, man, if I just did something else, I could have one more of what he has, and then you will end up with work again.

Anders Arpteg 2:03:37

But I mean it's the same, you know. I get often a question, you know, sh will you even work as a software engineer anymore? And for me, you know, if you compare just five years ago or something, when I programmed them, I could do this much. Today I can do 10 times more, if not even more than that. Does that mean that I work 10 times less with engineering? It's it's the opposite. Actually, I work more with engineering because I can do so much more, and it's actually so much more fun.

Henrik Göthberg 2:04:02

Are you guiding a whole engineering team?

Anders Arpteg 2:04:04

Literally, yes, if you look at it. I spend more time with it.

David Wallén 2:04:09

Yeah, I agree. I mean it's the same pattern, I think, in all people who want to achieve more, when you get a lever, you pull harder, right? Like uh Jumatan, my co-founder, like he starts his day with writing like three instructions on the bus, and then when he walks to the office, he has three PRs ready. Uh you know, uh, so there is there is more, yeah, like you work more. Uh it's a Jebon's paradox.

Henrik Göthberg 2:04:33

Yeah, this is Jebon's paradox, and I'm I'm not sure everybody knows about Jevon's paradox, but it's in summary when something gets drastically cheaper to produce, and so it lowers the threshold. So what you you you can utilize that technology for, and therefore you you put it on into more things. Yeah. And and and and so this is one uh you know, another thing that was sort of I think a way of looking at this is that we have been part of a sort of the industrial revolution, of a of a sort of uh the assembly line view of work, the scientific management, command and control, tailoristic management. We are cogs in a machine, right? And obviously, right now, those cogs are the things that the AI most likely can do better. So I said I said on the key keynote, we need to stop driving change and live in the machine room and start living on the bridge. So, like the machine, it's just the machine that sits in a vessel, and you need to steer the vessel, you need to guide the vessel, you need to operate the vessel. So it's the machine room and the machine metaphor of an organization shifts to a vessel of opportunity, moving ahead, and then there is a new breed of work that is defined based on a vessel and the bridge metaphor rather than the machine room metaphor that has been prevalent for the last hundred years. So I think a lot of work and the way we look at process and the way we look at task fundamentally shifts. Uh because we we don't need to be in the machine room in the same sense. We shouldn't. We it's it's at it it's now clocking in at a different speed. It's the same as abstraction level, but it's a different metaphor to show it's it's not simply going up doing the same thing, it's taking a different role. What do you think? The fundamental of work and how what is work, what is human work is changing that. Steer, guide, judge, set criteria.

David Wallén 2:06:38

Yeah, maybe. I mean, uh you you can you can do more with less. I mean, that is uh the essence, or you can share you can shift your mind in in like more cognitively heavy work, you don't have to information gather as much, or you don't have to, you know, uh but it leads to that you have to like fact check a bit more, maybe. But I think we will get past that as well.

Anders Arpteg 2:07:00

But but if we think in the limit here, um and and think that okay, imagine that we are past AGI. Imagine that we have AI that is a thousand times more efficient in memorizing things and using working with a lot of of data, in reasoning, in actually taking actions properly, which is really better today. But imagine it can and can do that you know a thousand times better than any human. Should we then remove the human from the loop, so to speak? I can think in some cases, absolutely yes, like driving a car. Why should a human do that if an AI can do it a thousand times better? It just will be so much more dangerous to have a human driving it, but perhaps not everywhere. I don't know. What do you think? Do you think you know in some cases we should leave humans in the loop? Or what's your thinking here?

David Wallén 2:07:49

Yeah, I mean absolutely. I think it's very hard to answer that uh question in a binary sense, but absolutely there are processes where the human should be kept for sure. Uh I mean, not only from the compliance perspective where it's it is actually binary, but also like philosophically, you know. But I think that what will happen is that the boundary will shift. Like, no one no one is asking now why are you converting your Java to like a binary? You know, no, no one is asking why you're doing I mean that that's that was never maybe done manually, but you know what I mean. Like no one is lower levels of abstraction, no one is questioning that.

Anders Arpteg 2:08:28

So that can be automated without human synergy, yeah, no problem.

David Wallén 2:08:31

So it will just be the boundaries, will just the abstraction will go up, and um, but I think um so so but but there's there's certain things that I think where you will always like you will always question should we really go away from that? But I I can't really possibly like you would have to be I think a philosopher or or or ethics expert to say really like how will we progress in this dialogue? But I think it's like a it's not a question for a company or an organization, but it's a question for society, it's like a societal debate. Like, where do we want to be involved? So that is something we have to like decide together. Um, yeah.

Anders Arpteg 2:09:13

Yeah, I mean, I can see if you take healthcare or something as an example, and and at some point you may need to make a decision, as humans have to do today. Should this human be you know given help or not? Or how how should we, if we have only this amount of resources and we have to prioritize and say, now we have to help that person, we can't even help that one.

David Wallén 2:09:35

Making these kind of decisions, perhaps that there are examples like this where Yeah, I mean, there there are, we have agreed, there are frameworks where I have agreed, and I think we'll reach consensus, like in medical, you have the quality concept where it's like if it costs this much, if you are this old and it's cost this much, then it's not worth it. They have the doctors have these frameworks, so I think they will be developed in other areas as well. Um, and we as a society just have to agree.

Henrik Göthberg 2:10:06

Difficult questions. I was gonna launch into it, but I yeah, you have medical persons well. So I think some of these topics is also about rethinking where where is the judgment needed, you know. So there's the other argument: like when to build something, the cost is zero, right? Literally, throw away compute, throw away apps. Then the core question is not if you should build, but what you should build. And maybe those kind of questions then become more thinking about shaping society. What should you build and what should not be built? And maybe those kind of questions then are the real questions that we need to own up to create the society that works for humankind, sort of thing. So we simply we should not build that stuff even it's completely doable because it doesn't you know from the from the human society perspective, what am I for? And what is AI for? What is AI working for? You know, and and I think those kind of topics, they look like super big topics, but if you think about it, if we don't live in the machine room, but rather on the bridge, those kind of decisions are the micro decisions we are doing without thinking, even in the lowest job. Should I clean, uh, should I take away that flower or that flower? You know, I'm I'm trying to put it in the context of a very uh pragmatic work. So there's there's decision making or the judgment making in in every single job. Even you know, the judgment of what do I put on when I go to work? The judgment, what do I put on when it's cold outside? The judgment do I walk over the street when it's ice, whatever. So, what I'm saying is that those things around judgment, should this be done or not? What is the evaluation crisis? Criteria on this task, those are maybe the core things that we will be very preoccupied with as a species and and in in every kind of jobs, rather than doing the actual doing.

Anders Arpteg 2:12:12

Yeah, perhaps that's a good future.

Henrik Göthberg 2:12:16

Because it it it it it it tells us humans were never meant to be cogs. Humans were meant for cognition, for reasoning, for argumentation, for discourse, and for creativity and for argument. And what is then the jobs around that? And I think that's beautiful.

Humans In The Loop And Closing

Anders Arpteg 2:12:38

Maybe Daniel, if we no, sorry, when Intrick has built AGI, um what do you think will happen? And we we can think about this in two extremes, thinking in the limit here again. Uh, one extreme would be that that uh machines will kill us all. It will be the terminated matrix of the future and um rather horrific dystopic future. Or we can think the other extreme, the other limit, which is basically that that AI helps us cure cancer, it helps us to fight the climate crisis that we're having, and we have basically no energy needs at all, or basically the supply is infinite. And uh we basically have a world of abundance where cost of products and services goes to zero. Where do you think we will end up?

David Wallén 2:13:31

Yeah, it depends on if we solve the alignment problem or not, right? So uh I really I mean I I think we will we will solve it, and um I mean you have to believe you have to believe in good, right? I I I think it's a very depressive world if you don't believe in good. I think um we as a species we've we've come super far. We've had a few turns, we've had a few points where oof, it was close, uh you know, you know, um and we have we've shown that we as a species can handle other potentially dangerous technology. Like we have nuclear technology that we have handled to sometimes not so well, but we've handled it. And this and that is a worse technology because that has less upside, you could argue. Uh and I think for good. Yeah, I mean you can use it to like power, you can have nuclear energy and so on. So you can you can but you you could the downside is also so big. So uh um, I mean you could argue I I would argue that AI has way higher upside, I think. Uh and uh I think we will we will handle that well as well. I I choose to believe in the good. I think uh the alignment problem is on the radar. I think there's funding allocated to the AI safety. There's some serious companies that are working only on this, like uh I mean, you know, uh uh Iliand and like they are they're they're pushing hard on this. Uh so I I root for them, and um I I think we will we will solve it. So of course, like yeah, of course, it will be um the the good version.

Anders Arpteg 2:15:09

Yeah, sounds great. David Valen, thank you so much for coming to the AI Afterwork podcast. It's been a pleasure to have you here speaking about philosophy and so much more. And um, I really look forward to when Interreg will fix AGI. That will be uh a day to celebrate for sure. Thank you so much for coming here.

David Wallén 2:15:29

Thank you so much for having me, guys. It was uh the pleasure was all mine. Thank you, David.