AIAW Podcast
AIAW Podcast
E136 - Sweden's AI Journey and Future Possibilities - Sverker Janson
Discover the transformative potential of AI across various industries as we engage in a thought-provoking conversation with Sverker Jansson, Director at RISE. We explore the journey from symbolic AI to modern machine learning and discuss agent-based systems, revealing how AI is reshaping fields like chemistry and pharmaceuticals. Our conversation shines a light on the challenges of securing funding for these pioneering AI initiatives, and Sverker shares his fascinating career insights, offering listeners a unique perspective on building specialized models and the complexities of traditional budgeting processes.
Dive into the heart of the Swedish AI ecosystem and uncover its strengths and areas needing improvement. We compare Sweden's AI landscape with that of the US, reflecting on how tech giants nurture engineering talent and startup ecosystems. From examining Sweden's successful companies to pondering government-led, large-scale projects, our discussion paints a vivid picture of the strategic investments needed to boost Sweden's global AI competitiveness. We also delve into the organizational shifts required to embrace digital transformation fully, emphasizing the role of data readiness and strategic adaptation.
Explore the future of human-machine interaction and the philosophical implications of AI-driven business landscapes. We ponder the potential for AI to create unicorn companies with minimal personnel, highlighting the importance of aligning human and AI agency within organizations. The episode closes with insights on AI adoption in Sweden, measuring investment's impact, and the innovative potential of transforming documents into engaging audio content with Google Notebook LM. Join us for a rich exchange of ideas that promises to inspire and provoke thought on the future of AI and its role in reshaping industries.
Follow us on youtube: https://www.youtube.com/@aiawpodcast
Bordering to creating specialized foundation models for different areas.
Anders Arpteg:Of course it's cheaper right.
Henrik Göthberg:Definitely so what we are talking about we had this conversation with others is like of course we're going to utilize some of these general models at the base, so this is a given. So when we say we're going to build our own language models, there's something underneath there that we can use. But to really take it into a more concrete application field, that is where you build your sort of narrow language model or your specialized approach. Right, right, right.
Sverker Janson:But everything isn't language or images, no, so we have chemistry.
Henrik Göthberg:Yes.
Sverker Janson:Of course. So we've been working a bit with that, specifically on toxicity of chemicals, specifically on toxicity of chemicals. And you know the industry, of course the pharmaceutical industry is running with this like crazy and it's so productive and it would be great if we could make that available to, let's say, the 200 or more small pharmaceutical companies in Sweden.
Anders Arpteg:And I think a modality is missing here, something like text, images and audio, to be able to do the more chemistry related applications.
Sverker Janson:You're training models on a different type of data on formulas and chemical reactions.
Anders Arpteg:Could the language model be a base for that? Do you think?
Sverker Janson:Might, but training on Gutenberg text is no good starting point for for chemistry, but that's because there is code yeah but couldn't you?
Anders Arpteg:fine-tune on chemical text or formulas. Actually.
Sverker Janson:I haven't even thought about it or explored how much chemical information there is just in the training. If I were met, I would have maybe cleaned that out of the training for, for, for llama, I don't know I mean, that could be a good example interesting reason for that? Yes, but something that's rice could help with we would love to do it and we know how to do that kind of thing. But uh, you can do the experiment of trying to find funding for that and yeah, so let's circle back to.
Henrik Göthberg:So it's like a key observation of is you know, what we discussed earlier is that we kind of have now more and more understanding of the vast opportunities we can do, and even in layman's are coming closer and closer to sensing these opportunities so, but but we are not there yet.
Henrik Göthberg:So we have started to redistribute funding and budgeting in order to basically go after it. So that's was a little bit your observation. That was sort of. Right now, the key hurdle is how do we organize budget for this? How do we secure the money for this, regardless if it's within rise or within a large corporation or stuff like that Exactly. This is something new that doesn't just easily fit into the normal budgeting process. Maybe, or is it too big, or what's the point?
Sverker Janson:I don't know. It's easier to present to an investment committee that you are going to build a house that has some kind of equipment in it and will serve some kind of function in I don't know, testing something, to invest the corresponding amount of tens or maybe even hundreds of millions for building a software infrastructure. That's totally different.
Henrik Göthberg:So I think this is a very interesting observation because traditionally, if you've got a large enterprise, they're dealing with this in a traditional budget, where you have your functional budget in different functional areas and then you have your IT budget.
Sverker Janson:And here all of a sudden.
Henrik Göthberg:Should it really go into IT budget? Maybe not. So this is a different beast altogether. So this is a different beast altogether.
Sverker Janson:I think some companies are perfectly on top of this and understand it and so on, but it's still hard to free up the funding and when you look at the numbers that are being thrown around by the big IT giants, it's a little bit off-putting. Maybe, yeah, let's invest $10 billion in new models.
Anders Arpteg:No, not really super interesting topic and let us get back to that a bit more. I think you know the ability for companies in sweden in general to obtain the funding necessary to do the transformation that may need or to start making use of ai properly is a very interesting topic. But before that, very welcome here sw Sverker Jansson.
Sverker Janson:Thank you.
Anders Arpteg:Director at RISE right, or what's the title right now?
Sverker Janson:My title. I have two roles. I am, on one hand, head of the AI center at RISE, which is formerly called the Center for Applied AI at RISE, and I am also what is called a unit manager. I have 20-ish people working for me on a subset of those topics, mostly scalable platforms, scalable data platforms. We'd love to hear more about what you do at the center shortly.
Anders Arpteg:That sounds super interesting.
Henrik Göthberg:And.
Anders Arpteg:I think I've known you, at least I'm not sure where we met the first time 2014 or something the Spotify days for me.
Sverker Janson:Yeah, something like that. Yeah, it could be. Yeah six.
Anders Arpteg:Uh, at that time right now, yes, a long time uh, how many years were you at six?
Henrik Göthberg:uh, should I really yeah, how many?
Sverker Janson:years were you at SIX Should.
Henrik Göthberg:I really reveal that. Yeah, should you really reveal that, oh dear.
Sverker Janson:So actually I started chatting with the people who were at the university at that time, when SIX was about to be formed. Oh, you were at the university and I was a student at the time and they were saying, hmm, they're going to start a student hotel, I wonder who should work there. So I had a summer job the first summer that existed in 86.
Henrik Göthberg:So you were there from the start.
Sverker Janson:Yeah, more or less. And then I started in 89. I was a PhD student at the time and La La Land and them.
Henrik Göthberg:they came in a couple of years later, right, yeah, I don't remember exactly when but probably early 90s.
Anders Arpteg:Cool yes it's a small network, sometimes a very small world, but please describe to us who is Rele Sverker. How would you describe yourself who is?
Sverker Janson:Rele Sverker. Sverker likes to make sure that good things happen.
Anders Arpteg:No, but I mean it's sort of a perspective on so I am very much into recent years, as we suggested. I've been at this for some time. And getting back to that issue we said the year no, quite a large no, I became head of a unit in 93.
Sverker Janson:So I've been there for over 30 years and seen many, many generations of research groups and several spin-offs and so on and so forth.
Henrik Göthberg:Were you talking AI then, or was it more? Data Six was formed on the basis of AI. That's one of the legs.
Sverker Janson:It was standing on. Definitely all the time, but AI at the time. I don't know I'm trying to see how old you guys are, but, uh, old enough maybe. Yes, um was symbolic, ai. Symbolic, of course, uh, and so I was active in logic. Program programming was the theme at the time, and also automated theory improving expert Expert systems. Constraints, expert systems for a little while. That sort of faded out fairly quickly in the late 80s. Cool.
Anders Arpteg:And how was the years, the few number of years at RISD?
Sverker Janson:Very few.
Anders Arpteg:yeah, exactly If you just, very briefly, were to describe the journey, so to speak. What do you start as and how did you end up in the position you have today?
Sverker Janson:Yeah, so I did my PhD on programming languages, that is, concurrent constraint languages, languages that were easy to implement on parallel machines, which we had at the time, that were good for symbolic processing. And when I completed my. Phd this coincided with the surge of internet and distributed systems, so we started looking at agent-based systems, distributed agents.
Henrik Göthberg:Coming in vogue again.
Sverker Janson:Yeah. And there were many concepts then that can be leveraged in a totally different way using language models, I think as the basic reasoning engine of agents that will change everything.
Henrik Göthberg:So agent-based approaches will come back in the big time now.
Sverker Janson:I would expect them to do, and many of the concepts being investigated at the time make sense. Make sense now even more. Then not because the agents were too weak.
Henrik Göthberg:Yeah, yeah, much, much, much To give them autonomy. Then was no.
Sverker Janson:But we organized an international competition that was sort of co-organized or co-located with major AI conferences. We had something called the trading agent competition. So we had competition formats where you would build agents that competed in one case for the components of a business trip, so you would compete with other agents in markets for air tickets, hotel nights, event tickets, things like that. That was one competition format. The other competition format was supply chain trading competition, so you would buy the components of computers and satisfy the demands of customers flows. And the cool thing about that was that the winning approaches so clearly combined machine learning and optimization of reasoning. And again we see that coming back the need for combining these two because machine learning is well. I should moderate that because with O1, we now see how super powerful machine learning can actually encompass also algorithmic aspects. But essentially you need some kind of optimization engine. For really difficult problems you need different techniques.
Sverker Janson:How you combine them is the magic source. That will be for any non-trivial AI application.
Anders Arpteg:I guess you mean some kind of reasoning or search.
Sverker Janson:Reasoning or research. Yeah, I mean, we've seen this over and over with AlphaZero, etc.
Henrik Göthberg:But I think this is so obvious we talked about this here If you can combine the power of what we could do with AlphaGo and AlphaZero and combine that with a large language the logical. Next step in research is how we combine these different strengths.
Sverker Janson:So alpha zero, of course, already leveraged machine learning or deep learning in understanding board states and sort of encoding the intuition of millions of games. The interesting thing will be to see to what extent the algorithmic part, the Monte Carlo tree search etc can be somehow picked up by models like O1.
Henrik Göthberg:Exactly.
Sverker Janson:And I'm super enthusiastic about this step. I think it's a major step in this class of models.
Anders Arpteg:This is a big rabbit hole.
Sverker Janson:Let's go to the profile, slip down it an hour later.
Henrik Göthberg:Maybe in an hour.
Anders Arpteg:I want to talk again, dick, in different ideas. Okay, and you did your PhD and then oh, about me.
Henrik Göthberg:Okay, yeah, okay just finish where we are.
Sverker Janson:No, no we were talking about things we did some time ago, but then I mean about soon. 20 years ago, you know, deep learning started appearing a little bit and step by step. From that time onward it's been on the radar and we have tried to build up capacity in groups and so on at rice and a few years ago we even managed to convince central management to fund center.
Henrik Göthberg:So here we are, and when did your journey formally become a RISE journey that was completely linked to the acquisition of SIX by RISE. Acquisition of SIX by.
Anders Arpteg:RISE then you were.
Henrik Göthberg:RISE, yeah, yeah.
Sverker Janson:So the Department of Computer Science at RISE is pretty much six, not exactly.
Anders Arpteg:No, but that's the origin, that's the origin yes, Perhaps it's also valuable to just describe a bit for people that don't know you know, what is really a RISE, and then also perhaps a bit more about the center you're driving.
Sverker Janson:But if you start with RISE.
Anders Arpteg:What do they do? We'll start with Rice. What do they do?
Sverker Janson:What do we do? We are the Swedish National Research Institute. We are 3,300 employees all over Sweden, dozens and dozens of sites. The origin is a large number of different branch, a large number of different branch institutes for different industry specific research institutes, so in a wide range of topics concrete 3D printing, etc.
Sverker Janson:Telco everything under the sun really. So RISE was a way to consolidate that and make it much more rational. It was really a mess with so many different little institutes and um today, uh, so from my perspective, it's a fantastic playground. So I am a super curious person, I love different areas and it's, I would say, the best possible substrate for driving AI that one could imagine, to sort of be able to have people who are experts in so many areas and who want to explore the potential.
Anders Arpteg:How would you describe what the Research Institute is?
Sverker Janson:Is it academia?
Anders Arpteg:Is it a cooperation? Does it work for profit? How do you collaborate between the two Excellent questions.
Sverker Janson:It's a non-profit Organization-wise it's one of the government's companies, so it's like Vattenfall in a sort of strict formal sense. We are neither university. It's a little bit tricky to sort of really get into the groove of understanding why, what the difference is. So people are very much value driven or curiosity driven who work at rice. They are not there to make the best, have the best salaries or be do an exit on a startup or anything like that. People spend some time at Rice and leave to do that, for example. It's not a university, but many of the junior people do PhDs, but as external PhD students to our partner universities.
Sverker Janson:Industrial PhDs yeah we call them institute PhDs actually.
Anders Arpteg:Really.
Sverker Janson:No, but I mean it's a separate form and has also separate funding from the Strategic Research Foundation, for example.
Henrik Göthberg:But is it fair to, from the outside, try to understand this as the bridge between academia and industry and somehow be able to do applied research or bringing academic cutting edge stuff into business. My take on this has been comparing a little bit with Teknália in Spain. We are here basically to be some sort of AI. It's really research driven to some point, but now it needs to land in real business, but it's still too immature to be really so.
Sverker Janson:could you frame it as a mission and a purpose from this bridge perspective, I suppose, to universities. Our mission is to strengthen the competitiveness of industry and the innovation capability of the public sector, and this is what we strive to do. One can use different mental models. In some ways it can serve as a bridge, but that's not precisely. I would rather call it a complement. So we take positions that neither the private sector or the universities do.
Henrik Göthberg:actually, so it's a little bit also filling a gap here. Definitely Lots of different shapes.
Sverker Janson:There are many different shaped gaps depending on the area. So for, let's say again, concrete or whatever, the gaps have some shape and for the computer science, it sector, the gap is different. So in our case, for example, our role has always been to, when we work directly with the industry on contract research, so to speak, we complement their cutting edge expertise. So they have lots of cutting edge people, but we are almost always able to have a couple of people who complement their expertise and can work with them. No, no, it's fine.
Anders Arpteg:Cool, and I guess we won't come to the center for applied AI as well. But just for people that don't know the RISE, I mean it's rather broad right, or what would you say? What type of sectors does RISE operate in?
Sverker Janson:Yeah, everything, everything.
Anders Arpteg:Everything Energy, transportation, business yeah everything.
Sverker Janson:Everything energy, transportation, production materials, chemistry, pharmaceuticals, bioeconomy yeah the some, some it's big now, yeah, yeah infrastructure and we often right, no, I'm but.
Henrik Göthberg:Samhällsbyggnad, yeah, yeah.
Sverker Janson:Infrastructure, and we often Right, no, but I mean. Built environments, built environments, built environments, smart cities, everything I mean. Just pick something out of the hat and we will have someone.
Henrik Göthberg:But if we now zoom in on the center, where does the center fit in? So you have a center and what's the organization org? Chart so to speak, or division for the center fit in. So you have a center and what's the organization, org chart, so to speak, or division for the center?
Sverker Janson:So I guess, it's in some ways a little bit different from some other centers. Where sort of the center is the thing that you talk about and the people are working in the center and it could have its own house and its own partners, et cetera, et cetera.
Sverker Janson:At RISE the center supports we already have a line organization with various units that already drive AI very actively, and the center supports certain groups to drive them even more accelerate what they are doing. So we both work with our cutting-edge expertise and invest in that, and we also try to find the forums to work across all of RISE, which is, of course, not super easy.
Anders Arpteg:Would it be fair to call like a hub and spoke kind of organization, where the center is the hub and then you have a lot of spokes outside in the rest of RISE to make them all be able to.
Sverker Janson:Yeah, you could say that.
Henrik Göthberg:So you have a mission as a center to strengthen the fundamental capability around AI, which you can apply for many different purposes, and then from this center you can support corporate clients in projects and stuff like that. But you can also now support and strengthen the other functions. You expressed it so well, I'm trying to, yeah.
Sverker Janson:No, no, no, exactly, exactly. That was very well put, yeah.
Anders Arpteg:Okay, awesome. Can you give some example perhaps, what the center is perhaps focusing on right now?
Sverker Janson:Right, so we have a fairly simple structure when it comes to the cutting edge research that we support. We have a deep learning group that also covers computer vision and AI for sound various modalities. Then we have a group that focuses on more the classical data analysis type of AI. So causal reasoning, things like that. And then we have people who focus on more compute infrastructure, so the computational Is that part of the center? Yeah, no, this is part.
Sverker Janson:I mean, these are topics that the center invests in if you put it like that so all of these groups that I mentioned belong to different units in the organization, but the center supports them.
Anders Arpteg:I see Super interesting. I'm thinking we should move to more like a general Swedish topic a bit and you say that RISE, of course in the center, then supports, I guess, primarily the Swedish society and ecosystem. Or is it also outside of Sweden?
Sverker Janson:I mean, we can work with partners outside of Sweden, and we do in EU projects and other international projects, but of course the government funds us to strengthen Sweden.
Anders Arpteg:So yes, how would you say the current state of the Swedish system, if you call that is? When it comes to AI adoption.
Sverker Janson:I would say, as you could expect A few years ago. As we all know, the government has been rather laid back in funding AI explicitly in contrast to some other nations.
Anders Arpteg:Including Nordic ones.
Sverker Janson:Including certainly Finland, etc. Etc who have been there, but we had fortunately the Wallenbergs who did WASP, and that was a great boon to Sweden, I would say. So that took care of in some sense.
Anders Arpteg:Just to explain, Wallenberg had the WASP program. So that is the biggest research program. It's 5.5 billion Swedish or something.
Sverker Janson:Something.
Henrik Göthberg:But it's very much focused on strengthening our research.
Sverker Janson:No, it doesn't do from my perspective, a lot for Applied AI, but it's great for building long-term really well-educated people and the universities also took, of course, Batten and started educating people en masse With that and because here again. Related to that. I guess some of the people they hired can teach the sort of but it comes down to funding.
Henrik Göthberg:At some point there is a big bucket of funding and now we can now start up different types of education and we can now fund PhD students. Yes, yes, so it's all related.
Sverker Janson:No, it always comes down to funding, and whereas the other types of funding, like Vinnova, that has always been solved within the existing budget, which is already fairly small compared to the needs.
Henrik Göthberg:But you ask a question here what is the state of the system? I think you can take a step back when we are talking about the Swedish AI ecosystem that is here to drive. Sweden's competitiveness forward. It's sometimes almost hard to grasp it. I mean, there are several key components. I think crisis is a key component. We know we're doing different things in terms of stuff. Ai, sweden is doing something, and then, of course, we have the universities doing something, and then we have different directives from government.
Henrik Göthberg:So when we say the state of the system, what do you think constitutes the key actors in this system?
Sverker Janson:Well, you mentioned several. I mean it's a very pluralistic system with so many different parts intermingling in so many different ways.
Sverker Janson:I think we're maybe missing a few, at least compared to some other countries. So what the US Silicon Valley has is the ecosystem of the companies that have been investing heavily in this, and what does this mean? It's not just the money, but it's the people who pass through that system and then out into the ecosystem yes, and people who know how to get things done. So we mentioned before our former colleague, lalle Lars Alvorsson, and I think he is an excellent example of someone who is a super data engineer, you know, knows exactly how to make things work at scale, and that is so valuable, and then he could move around and and help other companies and and he actually went through google and spotify exactly exactly so.
Sverker Janson:That's an example of exactly the ships that I and the ship's that I think, and maybe I don't know. I haven't followed in detail.
Henrik Göthberg:But this is a you are highlighting.
Sverker Janson:But that's just one person.
Henrik Göthberg:But it's a component in the system here, like if we contrast with other ecosystems around data and AI. I mean like if you go down the Silicon Valley route, you have very strong connections from Stanford or whatever. But they have a very strong startup ecosystem around the university and then they have a very strong VC. It feels like it's a machinery that sort of fits together.
Sverker Janson:Yeah, they have an order of magnitude larger investments in startups, et cetera, of course, which gives that engine a totally different level of fuel. But what they have above all is that they have people who have passed through Google and Microsoft and. Facebook and know their engineering practices around lots of things, including AI. When they start a startup, so they have a totally different starting point, of course, and we don't have any companies just like that I would say I don't know spotify has maybe been the source of people well, the only ones we talk about.
Henrik Göthberg:That could sort of not even come close, but they are, or our examples would probably be the spotify's and the clarinas and a couple of more and, in a a few years maybe, recorded Future. Which is Epidemic Sound. You have a couple of us, but they need to grow and become.
Sverker Janson:You know they need to have hundreds and hundreds of people to have offshoots to the rest of the.
Henrik Göthberg:The real hack to scale list that we can talk about now, I guess, is Klarna and Spotify. So just to summarize a bit.
Anders Arpteg:I mean we have a lot of fortunately scientific funding through WASP and Wallenberg.
Anders Arpteg:We get some from the government as well from the universities and RISE as well that do get money from the government, but it's a lot focused on the science part of it. We are potentially lacking a bit on the engineering side, if I hear what you're saying. And then in other countries like the US, know they have the tech giants and they have insane engineering skills and investments. And why are we missing those very simple question why don't we have a tech giant?
Sverker Janson:or more of them. Well, we have ericsson, I mean ericsson is I think it's a pretty good example. I'm sure that what they're doing at ericsson now they are investing heavily in ai and that will certainly have spin-off effects down the line.
Goran Cvetanovski:That's a good point.
Sverker Janson:But that's just one, and they are not like Google in how much they can invest in this type of project. So, yeah, I think AstraZeneca what they are doing is super interesting. They've been working at it now for several years. That's a super good example of leveraging AI and building expertise not just in Sweden, though, but still.
Anders Arpteg:But is there something missing? Then, if you were to speak to the policymakers here, what could we do to potentially improve? Why is Sweden lagging behind? And potentially, you know, some recent rankings say that Sweden is dropping down a bit in the rankings when it comes to AI adoption.
Sverker Janson:The government side in super ambitious projects that, um, uh yes, the the fighter plane development type scale of projects where, with several billion per year investment in some important like flagship projects for sweden.
Henrik Göthberg:So that's an interesting idea here that would.
Sverker Janson:That would put hundreds of people through the system of building something for real, which is important.
Henrik Göthberg:So we are talking about flagship projects or moonshot projects, not necessarily moonshot in the sense that it tries to do something regarded as almost impossible.
Sverker Janson:It could be AI for medicine, for example. Yeah, but you do it?
Henrik Göthberg:I mean, like if you contrast that with grants and funding from Vinnova, that is a totally different thing. There's a very, very small snippet.
Sverker Janson:It's so fragmented. It's so fragmented.
Henrik Göthberg:Which is a little bit like okay, it helps a lot of small startups to do a little bit. I'm a startup and I'm like I'm not going to be bothered with that. So what we Sweden need moonshot is the wrong word. I fully agree. We need ideas.
Sverker Janson:Super ambitious development project.
Henrik Göthberg:Ambitious development project, development, development, development, development project. You know not development.
Sverker Janson:Because we know how to do things we talked about before, that there are so many people out there that want to do AI. They see opportunities. We could identify at least one sector where the government could step in and just drive this, hopefully in a non-competitive. I don't know about EU regulations for supporting but. If it's a public sector, like medicine maybe we could do it.
Henrik Göthberg:Because then it comes to argument in what way do we want the government to invest and what should they stay out of? Should we have a government that becomes an incubator? Maybe not. We need them to fund, to build stuff. But I kind of like that simple idea.
Sverker Janson:They could order something. I don't know how it was done with the Jaws.
Henrik Göthberg:the airplane Jaws is an interesting example, because they are not trying to build the JAWS, they're ordering it, they're making a super ambitious project. They could order a Swedish cloud provider. Yes.
Sverker Janson:Let's order a Swedish cloud provider. Yes, that would be great, but it's far from enough.
Henrik Göthberg:Because a cloud provider is, I'll use an example, the distinction is we have a super ambitious project and we order it.
Sverker Janson:Or order the next generation of support for medicine.
Henrik Göthberg:For medicine, exactly Because it's not like oh, we need a new fighter jet, so we're going to start a department build a public sector fighter jet? No, that would be destructive, but you can understand how easy it is to fall in that trap with data and AI that, oh, we need a department to now build digital infrastructure for the whole public sector. It's a disaster. Right there are several traps.
Sverker Janson:Another trap would be to say oh, we need to support AI in Sweden. Let's build a house in Västerås and put 200 AI researchers there. That would be completely useless, I can tell you, Because what will they be working on?
Goran Cvetanovski:You see what.
Sverker Janson:I mean I mean AI needs to be used now to do things.
Henrik Göthberg:But the tricky point is do you know how close that is to reality of proposals?
Sverker Janson:I know that people are talking in these terms and it's exactly the wrong idea actually.
Henrik Göthberg:Why is it the wrong idea?
Sverker Janson:Because, you build a little blue sky house, where people don't understand what applications are.
Henrik Göthberg:They don't have the domain problem, they have no domain problem or domain expertise.
Sverker Janson:They don't know what to do. They will, in a blind way, go out and try to find their way.
Henrik Göthberg:It's like hen's chicken at things. It's not a focused effort.
Sverker Janson:No, no, no. That would be completely wasteful and they would probably diverge in dozens of different little tracks.
Henrik Göthberg:We need a focus.
Sverker Janson:The same house might be fine, but working on the. Ai for medicine and in a very structured and sort of project-led way. So large-scale engineering investments I would like to see that If it's possible and allowed, then it would make a huge difference.
Henrik Göthberg:But the key word that I also pick out of this is not research development. No, we have enough research.
Sverker Janson:I mean, one can always do more research I like research.
Henrik Göthberg:But that's not our.
Sverker Janson:No, we're talking about what would make a huge difference for the Swedish ecosystem, and this would change everything, I think.
Henrik Göthberg:Because if you do a large project which is focused on development, it's not only that you get something, it's the spin-off effect you explained with Lalle and all the people that come out of this, that then understand what real infrastructure means, what real AI compound systems mean Even if the big project almost fails.
Sverker Janson:Yeah, because as long as we're doing, research.
Henrik Göthberg:We're working on this stupid simple model. We're not building AI compound systems. Exactly, we need to build compound AI systems. All this stuff. It needs to be real right.
Sverker Janson:Yes, in every way there are so many other things to think about.
Goran Cvetanovski:Cool, by the way. Cheers from Lalle Hi.
Anders Arpteg:Lalle, Hi Lalle.
Henrik Göthberg:We stopped talking about Elon. We talked about Lalle instead.
Sverker Janson:No, no, Lalle is like a lighthouse for how you should what more people should be.
Anders Arpteg:Okay, and why should the government invest in AI and large-scale engineering projects potentially? How do you see AI and why is it important for Sweden's economic growth?
Sverker Janson:potentially to invest in AI related engineering projects. I think I'm not in a context where I need to convince everyone about the potential for both incremental and revolutionary change in just about any sector, so of course, we need to adopt AI.
Anders Arpteg:I mean everyone is but it's a bit unfortunate because you know, I think even today, this morning on.
Sverker Janson:TV4 News. You know, they said AI doesn't provide value.
Anders Arpteg:It's TV4, some kind of journalist yeah, probably a very creative person that is generative AI chat.
Henrik Göthberg:GPT and they are delusional that they think that's enterprise grade. I mean, like the main problem withGPT and all that is that they think this is AI and this has nothing to do with enterprise-grade AI, in my opinion.
Sverker Janson:ChatGPT. Yeah, I can make so much use of ChatGPT. You need to get used to what you can get from it.
Henrik Göthberg:Yeah, but I mean to put it in production is so much more than just using a chat interface. Yeah, but just for personal efficiency.
Sverker Janson:I mean, let's stop with that. I mean, how many people work with like office type jobs and write text and read text? Now I'm sure everyone can be. I'll pick a small number 10% more efficient by using this is a small number.
Anders Arpteg:I think it's higher in reality. But it will have a big effect to any company.
Sverker Janson:Yeah, I mean just out of the box. And then if people get used to using co-pilots for software development I'm sure you have discussed that or have heard the report there are people who try to downplay that as well. Yeah, the gain is just 20%. I mean, 20% is huge in itself. But when you talk to people who I'm not a full-time software developer, I try it a bit and it feels great. But they say, yeah, I'm probably two times more efficient, maybe three.
Anders Arpteg:So it's 100%. That's insane. Yeah, it's insane, it's insane what it means for.
Henrik Göthberg:And imagine now we have our old companies chasing budgets for 10% more efficiency or 5% more efficiency in the old way of doing things and now they are not investing in things that have maybe magnitude.
Sverker Janson:And this is only in an engineering discipline where we, for various reasons, have good models already out of the box by training on internet data and fine-tuning on even more codes. I'm pretty sure this can be applied to more disciplines as well. We mentioned chemistry, et cetera, et cetera, but lots of disciplines.
Henrik Göthberg:But I could go on into a huge rabbit hole here. Say then, then obviously there's something else wrong with this picture. It has nothing to do with with the fluffy understanding that this has potential. It has nothing to do with sort of if we have money or not, because it has to do about now know how, how to organize, how to do approach these things, because if you knew what the hell you were doing, you would have made those budgets. You would have made those budgets, you would have made those arguments and business cases a long time ago.
Anders Arpteg:I don't think it's obvious. Actually, Even if we knew how to do it, it's not obvious. You do the business case and it's not obvious for the government, apparently, that they do the investments that would bring so much, Apparently not.
Henrik Göthberg:Then there's a gap here there's an AI divide in front of us that we think this is so obvious.
Sverker Janson:It's stupid.
Henrik Göthberg:I'm just frustrated. Give me snooze so I can relax. And on the other hand, it's not clear at all. The story has not sunk in at all.
Sverker Janson:I think to be fair, there is still considerable uncertainty and things have moved very fast in industry terms.
Henrik Göthberg:We're in shock. We were in shock and denial.
Sverker Janson:Maybe to some extent.
Anders Arpteg:But I think there is a communication problem at least here. A lot of people don't understand the potential value that they can have. We don't speak about the opportunity cost that we have spoken about so many times Opportunity cost. If we also speak about medicine, as you said, not only can we make this staff potentially 10% or more efficient to do their work and help more people, but potentially it can literally save lives if we do the investment.
Sverker Janson:I think the impact would be mind-blowing on medicine, that you can diagnose things perfectly based on the knowledge we have today, as opposed to you know about all these people who have these tens of thousands of rare diagnoses and who need to go to the hospital I don't know how many times 30, 40 visits before they get the diagnosis, the proper diagnosis, the proper diagnosis.
Henrik Göthberg:That's crazy, that's real. I mean, like all the outlier diagnosis is like that I know.
Sverker Janson:I know, I know, I know and we could change that.
Henrik Göthberg:Quite easily.
Sverker Janson:Easily is almost the right word. It's hard work, but it's easy in a conceptual sense. In a conceptual sense, it's quite easy yeah just throw money at it and we can fix it.
Anders Arpteg:I think there are so many companies right now that are at least when they have the right management at least on board are thinking how can we make use of this technology for our purpose in the company, and they're trying to find the best product to get started with, et cetera. But it wouldn't be nice to do that on a country level, so to speak, is that, of course?
Anders Arpteg:Then, if you say that if you invest X amounts of money and you get to double back in a couple of years, it would be kind of obvious that you should do that investment right. And I think you know the DIG had a report back in 2021 sure, I think 2021. You know, saving basically they could save 140 billion swedish crown per year in the health sector. Uh, if they did this. But it didn't really help, it seemed.
Henrik Göthberg:But no, but it's strange, but it to me this is not strange. I'm getting to a sense where I think we're all approaching it wrong in terms of the consulting industry and everything like that, and I'm actually I'm a huge fan of Lala's way of thinking around this and he has a pet peeve. He's been driving around Fundamental gap in engineering of industrialized approaches with compute versus the craftsmanship we are doing here and where, basically you know, he says okay, I can help you accelerate on industrialized compute. But you know what the main problem is?
Henrik Göthberg:the cultural aspects, the organizational aspects so what I'm trying to get to here is that I think I get to a point where I meet people on very high level and we talk about this and, and you know they, I don't come through to them because I think you actually don't need to start in strategy place with this, because it's profound. You can go into, I can, I can probably improve any process in your company with this. So what's the strategy is simple. There is no fucking strategy.
Sverker Janson:Is you start using it and then at one end, start at one end and then do it right and then do it right then what is do it right?
Henrik Göthberg:you need to have the right team. The competence is the way of working and if you start focusing on that, the strategies will come something important there.
Sverker Janson:You said you need the competence. Yes, so that gets us back to the page of the ecosystem and we need to funnel more people through. I don't know like it's not a bootcamp, but it's sort of a bootcamp to work at a big place that already know how to do it.
Henrik Göthberg:But the interesting thing is then it's not enough to do that on an engineering level. You need to do the same on the middle manager level and executive level. So you need the bootcamp of data readiness across the the board and let's, let's take one step back.
Sverker Janson:Yes, um, because it's easy, when you start talking about, yeah, let's do ai for x, y, z, etc, etc, the whole alphabet, and uh, ai for medicine might, or for the government, or for automating whatever might give us.
Sverker Janson:You know, yeah, we have a little AI solution here, provided by startup X and one here and one here and one here, and it's all I don't know.
Sverker Janson:Feature hell with all sorts of little things that are not connected.
Sverker Janson:I expect that what actually needs to happen at the same time as we start developing concrete solutions is also rather turning companies inside out and making them digital at the core, and that the key, the core processes and state, et cetera, is sort of a digital factory, a digital machine that takes in and maybe I'm losing you now, but that takes to go to a future where all organizations, public sector or companies are digital machines in this sense and that has some kind of management, of course, to oversee and organize and structure and so on, and maybe a few operators and maybe some people that complement still what this machine can do. But this is where we need to go and I use my. It's probably not entirely correct, but I think about how staff talks about what recorded future does as a sort of a blueprint for this, because they have this knowledge graph and they have these pipelines of information coming in and services being based on that and so on. It makes perfect sense to me. That's probably what whole companies and organizations should look like.
Henrik Göthberg:I want to stop and say I think this is profound. This is what I used to call a goosebump moment. I haven't said that for a while.
Sverker Janson:Do you have goosebumps?
Henrik Göthberg:No, you are too experienced to get goosebumps okay, okay, yeah, for real, and what you said. Now I've been preaching for years. You know, if you look at the digital natives, so to speak, they have a dna where this is part of the dna how to do data?
Henrik Göthberg:how to do so it makes sense, and then they and then they figure out what techniques and what they need, and we're still living in a paradigm, in my opinion, where the enterprise companies are still analog at the core and sugarcoating this, and they have not really figured out or understood themselves that we are 50% software company in the future, like we have a domain. We might even have physical assets, like Vattenfall, but in order to run this, we need the digital infrastructure at the same level of focus.
Sverker Janson:It's not just a software company that develops software. The company is software.
Henrik Göthberg:Yes, that's the difference.
Anders Arpteg:You see what I mean.
Sverker Janson:No, but I mean, it's a system. The human organization needs to be downplayed, that needs to be organized around the digital organization that produces the services, that takes in the information, if you look at an enterprise as a complex adaptive system and you have agents, artificial or human.
Henrik Göthberg:They need to work with perfect feedback loops, all of them as one big, complex adaptive system and in this context, now it's all socio-technical, but it's also all digital, if you like, and this is a mindfuck in terms of traditional functional division of labor.
Sverker Janson:Companies try to sort of how should I put it? Re-engineer what they're doing. You buy a number of systems previously not so much AI but various systems and then you retrofit an enterprise architecture, sort of an organizational architecture, and you try to put it together and make it work and in the end we have Conway's Law.
Henrik Göthberg:So everybody you know we organize in such a way so we get shit out of it. It's obvious, your words, but maybe no, but I mean like it's obvious that organization matter and how we understand this at the core matters, because then we are framing the problem differently than what we're going after.
Sverker Janson:It's a reframing, so maybe the already software native organizations will simply eat the rest step by step. That's a perfectly viable scenario, perfectly viable.
Anders Arpteg:So how do we fix this? How do we make people understand this? I mean, I think you touched it partly before. If we were to go to a company and say you should start becoming more AI and data driven, then they can, you know, put the technology at the hands of the people or do something, but it won't be sufficient, I guess, and we need to have proper change management in place to make the company really, you know, make this work. Is that potentially what we need on a more like country level to make that happen? Or what is how do we make the same kind of transformation happening on a larger scale?
Sverker Janson:I don't know. I think I mean change is well known to be difficult.
Sverker Janson:Yes, I'm not going to say it, and so maybe the solution is not to focus on changing existing companies so much, but try to lead by example. Much, but try to lead by example and have at least a few very shining examples that show how this should be done. I don't know if this is compatible with what I said about the government investing. When I think about the government investing, that should be in the order of several billions per year, and even if that means increasing the national deficit, I mean it's well worth the money. Yeah, it's nothing in the long run.
Anders Arpteg:But what I mean? I mean some people at least say that. You know, if you want to have a proper like digital transformation of a company or an AI transformation of a company.
Anders Arpteg:You know the tech is one part, of course, and the org could be something, but change management could be like 60 plus percent of the work. To perhaps not change the organization per se that much to start with, but at least change the processes that you have and the understanding for the people how they should be using services that exist. Would you agree with that? It's a lot of work to make that happen.
Sverker Janson:Yeah, I mean I've seen the inside of Rice undergoing change and I have super big respect for the challenge of changing anything, so I would probably I don't know what I would do if I had a big company, but I maybe think in terms of self-cannibalization to actually start a couple of core activities that are just done in a new way. And then you try to sort of transfer step by step the old activities. To that it's hard to change, you need to create something new, you need to containerize.
Henrik Göthberg:Yeah, containerize, it's a good word In order for this to work. Also, to containerize risk of your experiments. Yeah, I think the chopping up of the problem into containerized bits is a key strategy here. I think somehow it's a revolution in know-how, so it starts on a practice level. It's a revolution in know-how, so it starts on a practice level. So if I had all the money in the world and I was a CEO and.
Henrik Göthberg:I said you want to become data-driven, I would spend the first millions on a fundamental overhaul of practices and training in new practices. So basically, if we want them to act as a software company and they have never seen a software engineering practice they don't even know what it is. They don't know what domain design is, they don't know what agility DevOps is, blah, blah, blah. You need to start there, otherwise they don't have the reference or frame of mind to think what they should be doing differently. I think it's a major practice overhaul and then containerize. I mean, like what I'm saying is containerize, change practices here and let that spread.
Sverker Janson:To get the chance to think deeply about this, about some concrete case, would probably be very intellectually satisfying. Because it's so many different problems, it needs to be done incrementally right and do you do that In an agile? Way, I guess. Agile, whatever I mean you can't just do a big bang flip. Agile, whatever I mean you can't just do a big bang flip from one day to the next.
Henrik Göthberg:I spent the last 10 years on this intellectual problem.
Henrik Göthberg:And I think this particular problem and I think there are so much research done to lean into on how to think about this. I mean, like you take innovation adoption cycle, the work from Josef Schampeter to Everett Rogers to Jeffrey Moore basically lays out a fundamental view of innovation to adoption. What that is all about If you take that simple view set and then apply that into a company and you understand that any transformation is used as long sequence of adoption events, many small events, compounding effects. This is all really in the books but no one is applying it in terms of but are these?
Sverker Janson:theories or experiences applicable to disruptive change of the nature that this might evolve.
Henrik Göthberg:I think so Because, if you take Jeffrey Moore, crossing the Chasm right, it's literally written from a startup to hyperscaler perspective. So how do you start from scratch? Spotify? You know you build a product, you beta test it, you get the beta clients, innovators. You know the major shift is how you do the Crossing the Chasm, how you go from innovators early adopted to early majority, late majority. We've heard this before. Right. If you think about that, it's more applicable and written on the macro market scale, how you build companies. But if I take that logic and put that in Scania, completely applicable in terms of intrapreneurship instead of entrepreneurship.
Henrik Göthberg:My hypothesis.
Sverker Janson:Yeah, cool. Maybe, our listeners want to hear something about AI, but this is what is necessary to make AI happen. Probably this kind of thinking that you are talking about.
Anders Arpteg:Should we take a little bit different topic. Perhaps we started speaking about funding a bit, and that's a big problem that we are lacking, at least in some aspects of the funding to do that I mean there is a funding here and there, but it's it's.
Sverker Janson:It's too fragmented compared to what sweden might need to accelerate. How does it work, can you?
Anders Arpteg:celebrate. How do you raise money, so to speak, for different projects or for collaborations with other companies? You get money from companies, right From the state. Yeah, yeah, yeah, we have some. Just describe how it works, yeah.
Sverker Janson:I'll tell you. So we have a bit of basic funding, which is maybe 20% of the total. That is needed just to have the freedom to direct what we do. Otherwise we are completely funded by. So we have a fairly large part of RISE is from the former SP, so it's Testing, inspection, certification activities, stolten Spruning Stalt, stolten Spruning Stalt. That is some part of what we do. Otherwise we have commission projects, public funding, the whole shebang.
Henrik Göthberg:How much is EU funding versus Swedish funding?
Sverker Janson:I don't know the numbers and it's so different in different types of departments.
Henrik Göthberg:I'm sorry. Yeah, yeah, that is, you're right.
Sverker Janson:But maybe it's. I don't know. I'm just picking a number out of the air. Maybe it's EU might be 15-20% in our department. That's just an imagined number.
Anders Arpteg:I guess you during your few years or extremely large number of years at Rice have been part of one or two EU funding projects, as well or applications right. What do you think about the application process for EU?
Sverker Janson:The application process. I started my journey at then six by jumping headfirst into a EU project and working that. That was great as a social process and getting people to get to know each other, and I think that was a major objective also of this program. That was a major objective also of this program. And then I co-wrote, or at least organized, the proposal just after that. I think the process was still fairly simple then and after that and then I um I sort of semi-led that project as I was too junior to do it formally, but I did a lot of coordination, as it's called.
Sverker Janson:After that, I have been able to stay away from being directly involved.
Henrik Göthberg:In writing. But maybe connecting this we had Tor here as a guest from HPC and now we have coming up all the works and grants and funding around AI factories. Exactly how will you? Because then we have HPC and the competence center around HPC and now we have the competence center around AI Will you collaborate If you have the title AI factory? This is quite interesting, right? Because you have a lot of strengths here in RISE.
Sverker Janson:Yeah, I mean we are part of an application being prepared.
Henrik Göthberg:Together.
Sverker Janson:Together with others, yeah.
Henrik Göthberg:Of course. So how does that work then? Because then basically, you have one center, but in order to write that application, you need to now find the. Should I use it as a consortium, or, you know, you need to find the right ingredient in that proposal with several actors, I assume. Or can you do it on your own?
Sverker Janson:No, no, no, no, you work with several in Sweden and otherwise. Yeah, yeah, yeah Of course, and the center as such is not a body in this.
Anders Arpteg:it's RISE which is a partner Cool, and I guess most of the projects are a bit more engineering focused. Or are there some projects that are a bit more like base research or foundational research kind of projects as well?
Sverker Janson:I would say we have the full spectrum of projects and you know for PhD students that we have. We of course need to give them fairly base research. It's not always applicable for computer science, as you know that you do something, you develop a new cool thing which is sort of basic research, and then you just turn around and start a company using that technology. I mean so it's not the traditional TRL levels. I don't apply it straightforwardly.
Henrik Göthberg:I was going to ask you about the TRL levels.
Sverker Janson:Some people use that term. I haven't used it so much.
Goran Cvetanovski:No.
Anders Arpteg:Okay, but you do some. Or the PhD product, I guess, is a bit more basic research, kind of yeah and that is at least for us within computer I mean.
Sverker Janson:So I mentioned before that at Rise we take the necessary positions that complement what other actors do in different fields In computer science that means being fairly cutting edge in our research because we have lots of good consultants in companies et companies etc. That take other positions.
Henrik Göthberg:So we are applied but with bleeding edge knowledge but in order to also humor the real ai folks. Yeah, um, could we, could we use elaborate a little bit on what you would consider your key lighthouse projects or research Things that you're really proud of or you think is super interesting, super sexy. That is selling your house, so to speak, selling my house.
Sverker Janson:So yeah at any point in time. We did some kind of inventory. We found that we are probably running around 150 projects with more or less AI content at rice. So I mean more or less so, some less, some more, and my ability to keep up with what everyone is doing is limited, but I'll mention a couple of examples that I have at least close to myself on my radar in front of my nose's probably the best metaphor here.
Sverker Janson:No, so we are, um, we are working with ericsson. I can mention that I asked if I could not for this specific, but more generally, uh, they are building a co-pilot, um for their own proprietary hardware and software. You know they have their own little environment that is not well covered in its sort of the type of architecture and the programming models that they use by the public co-pilots.
Anders Arpteg:So we're doing that and have a just elaborate a bit more. Is it more a type of like internal chatbot kind of solution, working for their kind of?
Sverker Janson:No, it is about co-generation, co -generation, co-pilot, co-generation for telco-applicable coding of radio-based stations.
Henrik Göthberg:Yeah, roaming and all those.
Sverker Janson:I can tell you that if they succeed with that, with our hopefully valuable help, that will have a huge impact.
Henrik Göthberg:So they are making a custom co-pilot applicable for radio-based stations.
Sverker Janson:Every company which is in a similar position should do that, of course.
Henrik Göthberg:Of the scale and size required to think about that, but for them, of course. For them, of course, interesting.
Anders Arpteg:Can you share any technical details about that?
Sverker Janson:For two reasons I am blessed with a lack of knowledge of technical details and I couldn't tell you anyway.
Anders Arpteg:That's usually the excuse.
Henrik Göthberg:We stick with that excuse.
Sverker Janson:Good, let's mention another one. It's a pretty cool one which we often bring up as an example. Not one, it's a pretty cool one which we often bring up as an example.
Sverker Janson:so we've been working for a very long time on monitoring ships, ship movements in initially in the baltic, together with the swedish coast guard, and they put a system into operation. So you have radar and probably also what is called ais information and so on and so forth all the information you have, and that reason was initially to help sort of talk to ships that seem to be maybe running aground or something like that. Recently this has been expanded into subsequent EU projects and to the Arctic seas more generally and now with a focus on also monitoring potentially illegal activities, so companies that might be doing nasty things with underwater infrastructure, which was, as you know, became important recently but illegal fishing, dumping, you know, tanks, waste, etc. And this is a very cool application and although, again, I am blessed with the lack of knowledge of exactly what they do, I know that it combines sort of data-driven and more rule-based methods, since it's a very complex undertaking?
Anders Arpteg:What kind of sensor data do they make use of?
Sverker Janson:Yeah, I wish I knew exactly the excuse comes again.
Anders Arpteg:No, no, no.
Sverker Janson:But it's a sad truth actually Okay.
Anders Arpteg:Yeah, I mean, that sounds super cool and important and apt for this time.
Sverker Janson:I can mention more examples if you want yeah, please so rather under. I mean we talk about sound. I mean the term in the context of language and music, but there are so many sounds and AI for sound has at least the same potential as any other modality, so we're working a lot on that. An example is monitoring wildlife and there is commercial interest, so companies like, again, vattenfall that I mentioned our sibling, which is just bigger than us needs to monitor. The English term for this is capercapper cayley, do you know what that is?
Sverker Janson:okay okay, so a swedish large foul bird or not just swedish northern hemisphere? So apparently they need to monitor those around wind farms and understand the impact of wind farms on wildlife and also they monitor fish. We are not into that, but using computer vision etc.
Anders Arpteg:I'm trying to understand why and what the impact of wind farms could be.
Henrik Göthberg:I used to work at Vattenfall.
Anders Arpteg:You did yes.
Sverker Janson:For many years.
Henrik Göthberg:There is a lot of cool projects going on on lax Trappor.
Sverker Janson:Yeah, I've seen pictures of that. Cool stuff. That's great. Okay, so that's one thing. General area and more generally also monitoring computer vision for satellite images. Lots of work on that for various applications.
Henrik Göthberg:I have some friends at R drive that's been in projects quite for some time with exploring things with Boliden and Scania in transportation and mining. Oh nice, there are so many things, there are so many things.
Sverker Janson:I mean if I could maybe bring a list and go through them. I'll mention just one more thing, because I think I don't know this sort of tickles, my own imagination, the role of AI of this kind in research, in scientific endeavors, and so one very cool project we have is collaboration with very nice materials researchers in Uppsala. So we have a PhD student who is working on methods to accelerate materials development.
Sverker Janson:Ooh, To speed up the exploration of potential materials for solar cells. It's more general actually. It's sort of you use I think it's called a sputtering machine and you layer atom-thin layers and produce cool materials.
Henrik Göthberg:What's the data and AI angle in this?
Sverker Janson:Again, I'm blessed with a lack of details, but I think it has to do with optimizing the experimental design.
Goran Cvetanovski:We have a question, so I'm just going to read it. This is from Sverker. We have a question, so I'm just going to read it. This is from Sverker Do we need more in-house applied research and development with budget in enterprises, or does Sweden need smaller, fast-paced outlets, startups making building blocks that those companies can integrate? And the second question is what is the adaptation rate of AI ML in Swedish enterprises versus the rest of the Europe, nordics? I think that we can take the first one, the second one.
Sverker Janson:The second one is very difficult. I have no numbers about that at all, so I couldn't say but yeah, I guess both are needed. I mean, it's hard to seriously source AI components and expertise if you don't have it in-house. So whichever way you need a combination.
Henrik Göthberg:I think that's the answer, because if you don't have any critical mass internally, you have no direction.
Anders Arpteg:Direction or relationship with who to collaborate with you know, the combination of the two is good.
Sverker Janson:I mean you can't see snake oil. What is snake oil? What is real Exactly? There's so much snake oil? Yeah, right now it is.
Henrik Göthberg:So I would actually start from the internal to some degree, but then think about what should we really be good at internally? And how should we collaborate? But if you don't have anything internally, you cannot spot snake oil. I think is a good summary. Yeah have anything internally. You cannot spot snake oil, I think it's a good summary.
Anders Arpteg:Yeah, like should we do the other one about, I guess, the sweden's how is ranked or compared to other european countries, or what was it?
Goran Cvetanovski:it's more about the adoption of machine learning yeah, adoption of machine learning and ai in sweden I have the rest of Europe, nordics, but I think there is actually a ranking about this and there was there's a report actually which is done by Silo, which is the AI in the Nordics. And then there is another one from the European Commission that is going out where it ranks. Basically it's not this one, though I can try to find it actually, so we can come back to that a little bit later on.
Henrik Göthberg:But I think this is a very, very tricky question, because even in the large enterprises, the main problem to me is that you can be super mature, but it's in pockets.
Sverker Janson:So it's really about.
Henrik Göthberg:Can you find cutting edge? Yes, what is the average on enterprise scale adoption? So you get very, very different understandings and views. Of any company in Sweden of a large size today, I would argue.
Anders Arpteg:So it's not widely spread in my opinion still, but it was something published recently and it actually, if you just take the trend at least, is that Sweden is actually trending down compared to other countries. You mean in relative terms, not keeping up no, compared to other countries it's apparently being ranked lower in terms of AI adoption, if I recall correctly, Ouch that's bad. But yeah, if you find something that would be interesting.
Goran Cvetanovski:I know that there was a European commission was doing what is called last year. I know that for a, a European commission was doing what is called last year. I know that for a fact. Whoever did this? Let me just share the screen as well for the viewers. Let's see this is the ISOC.
Henrik Göthberg:This is a tricky one, but this is a very strange.
Sverker Janson:I mean, what are we measuring here? What is the structure of the sizes of the companies? How many employees do they involve in blah, blah blah?
Goran Cvetanovski:What techniques are we talking about here, Exactly, it's tricky because they don't actually measure that across the organizations. Usually when you find statistics like this is how much the European countries are investing actually, and that can be in research, that can be in a governmental sector and etc. But it's very hard.
Henrik Göthberg:I find this what you pointed out there's so many dimensions in here to compare apples with apples, so it's really, I think, the investment part is not that bad of an indicator.
Anders Arpteg:I wish we had more, but just saying that the company or private investments in ai is going there was another one called the ai index.
Goran Cvetanovski:I think that's stanford stanford happens.
Henrik Göthberg:I don't think they haven't come I can steal menu on on. You know why? Why investment is not a good metric. I mean, I, I, I, I think it's one metric that is useful to look at, but it's a little bit like politically we've we have always used this in our political debates Are we investing so much and so much in whatever, it's crime or whatever? And then you, okay, please could we understand what you are doing? Is it any useful? Is it a wasteful? You know so. You know in the terms of being green. You know, when I was at Vattenfall, someone actually did commission a very, very expensive report to understand from the different budget propositions from the different parties, which one has actually the biggest CO2 impact. So then you need to not only understand how much they're investing, but if they're investing in scooters, electric scooters- in Stockholm for 2 billion, or if they're doing something else.
Henrik Göthberg:And then, if I put it like this, miljöpartiet was not the most aggressive for CO2 reduction.
Anders Arpteg:Obviously you want to measure the value. The problem is really you can't get it. So I mean, if you don't have anything, else it's not a useless metric.
Henrik Göthberg:No, I agree.
Anders Arpteg:Not useless.
Henrik Göthberg:It's not useless, but how useful is it?
Anders Arpteg:True, we wish we could do it better, but if you don't have anything, it's better than anything.
Henrik Göthberg:It's better than anything.
Anders Arpteg:Yeah, I agree or nothing, I'm thinking, but you actually mentioned some project before. I think you played around with Google Notebook LM, right?
Sverker Janson:I did quite a lot actually.
Anders Arpteg:Perhaps you can start by just explaining what is it.
Sverker Janson:Perhaps you can start by just explaining what is it I haven't gotten into it.
Sverker Janson:It seems to be originally conceived as a kind of service or experimental service where you can upload various documents and then have a range of services. You can get sort of a study guide, frequently asked questions, summary, blah, blah, blah from that material and I guess that's useful enough. But that never came on my radar. But what really caught people's attention was this new feature called audio overview Was that a rather strange name of that feature that takes the same material and produces a podcast.
Henrik Göthberg:I think it's a podcast feature.
Sverker Janson:It should be called a podcast. And it is a podcast and it's immediately mind-blowing how great it is, how good it sounds, and so on. What I discovered, of course, is that if I asked ChatGPT to produce a script for a podcast, it does actually a quite good job just in itself. So I think the most impressive feature of Notebook LM is the very natural sounding sound and how they mix like you did. Now a little bit of intervention.
Anders Arpteg:Just to get people understanding how it works. Yeah, please, please, please, you just upload some kind of document, as you say a PDF or some kind of piece of text. I tried actually, just my CV. I listened to that cv.
Sverker Janson:I listened to that, oh, I listened to that as well, and as I replied when you mentioned it, this shameless self-promotion is what you produce but you know an american flavored?
Anders Arpteg:yeah, extremely bloated, kind of had him at leave, exactly no no, this is the great story of no.
Sverker Janson:it didn't say like that.
Anders Arpteg:It did Sort of sort of, but in short, basically you upload some piece of text and then you get basically two people discussing in like a podcast, setting that piece of text.
Sverker Janson:A man and a woman.
Henrik Göthberg:Yeah, a man and a woman.
Sverker Janson:With very peppy American voices.
Henrik Göthberg:Yeah, and they are excited. Excited and enthusiastic, and they are enthusiastic and they build up a cliffhanger and you know so they're doing it like and make jokes. They make Amazing, it is amazing.
Anders Arpteg:It is not just a text-to-speech kind of traditional AI voice that operates one word at a time. It literally actually has a lot of inflections and way of speaking. Fantastic, that is completely.
Henrik Göthberg:And the filler words and the storytelling. I mean they take it to. I mean, like I think they squeezed out like seven or ten minutes podcasting out of two pages of one page resume. You know that's amazing. You know bullshit machine, you know that really strings everything together and it's not bullshit in the sense that it's wrong or anything like that, but it's the way it's storytelling.
Sverker Janson:And what surprised me a bit was that it doesn't only take the text into account but also the images in the document. Did you notice that? No, no. So I think I sent you the link to our State of AI by Rice report and on the sort of second page, there is me I should mention. The point of this report is also to use imagery that is AI generated and a little bit tongue in cheek, you know. So it was a picture of me and with a cat that had been edited in, and they commented on this and in a very sort of humoristic way, and okay, great.
Henrik Göthberg:But it's amazing.
Sverker Janson:It's really amazing. And now you can also throw in and I tested this today YouTube videos.
Henrik Göthberg:Yeah right, youtube users they make a podcast and they make a podcast.
Anders Arpteg:We could put our podcast into the Notebook 11 and have them have a podcast about the podcast.
Sverker Janson:Yeah, A meta podcast. That is so great. This is a meta podcast and maybe people will be more eager to listen. They will be more popular.
Henrik Göthberg:So when we go to an American market, shorter, shorter.
Sverker Janson:And more perky.
Goran Cvetanovski:More perky word. Oh, this is crazy. Actually, this is a go-to market for america. I actually tried that tomorrow. But just to very that thing about like, uh, sweden versus the the rest of the european countries, if you, this is from 2023, from european parliament, so, uh, there is no, there is no ranking there, but it seems like sweden is a number five. Uh, so we're talking about 62.5 billion invested from US. China is 7.3 billion.
Henrik Göthberg:This is private investment.
Goran Cvetanovski:Yeah, this is Sweden 1.8 on fifth place after Germany and United Kingdom. And we are talking about private investments in AI by country.
Sverker Janson:But this is also this is last Not normalized by population, so we are in fact probably yeah that's surprising, almost per capita.
Henrik Göthberg:Would be amazing.
Goran Cvetanovski:Would be amazing from, from my perspective now traveling to all of these conferences, uh, and etc. Doing them. I think that, uh, when we are, when we're talking about maturity of ai and the topics that we're discussing in Sweden and on data innovation, etc. I think that we are actually quite good. Yeah, quite good, because if we remove the silicon value from this equation, the discussions that they are having in the UK and US it will be pretty much similar to what we are having here we are a little bit slower in decision-making process because we have this FICA decision-making process right, so it takes longer time for us to basically get into the innovation process.
Goran Cvetanovski:But I think that we are quite good.
Henrik Göthberg:But I think I can flip it and if this could also potentially be a dangerous, uh oh, we are good. Pat ourselves on the back when, in reality, are we getting adoption from these investments or we getting it into the ground of a really changing our companies, or is this innovation theater? We don't know that when you look at that.
Goran Cvetanovski:Yeah, but the interesting fact is like uh, because I had a news prepared for day. If you look at the, it's not actually only in Sweden you have in majority of the countries. You have basically a lack of delivering value with AI as a main problem. It doesn't matter where. It is right, because, according to this research that Accenture actually did, it was less than 2% of everybody who has invested in generative AI are actually using it or they have a full stack generative AI installation infrastructure and everything else. So it's not that we are that bad at all. I don't think so. We just basically need to live for a little bit more so we can be even better. Good.
Sverker Janson:But just a reflection here. There is also a matter of the distribution of the funding and how many of these units of funding that are sort of at a critical level to allow that enterprise to really move forward. I think this is a key point.
Henrik Göthberg:Because if you look at United States 62.5 million it looks like US is fantastic, but if you look at that per capita, if you look at that per state, you will find that there's a couple of states where the majority of this money goes to and there's a couple of companies where the majority is.
Anders Arpteg:But it sounds like a small number it's a billion, 62 billion euros Sounds wrong.
Sverker Janson:It sounds completely wrong. I mean, just sam altman is tossing around bigger numbers than that. Yes, okay, not necessarily in realized.
Anders Arpteg:Uh, if we include all the you know, compute infrastructure be invested in.
Henrik Göthberg:You know that by itself would supersede that, yeah, so it's something wrong with that number.
Goran Cvetanovski:No, so exactly it is from where you just stand for university, so it shouldn't be the tricky one is you need to really dig into the numbers.
Henrik Göthberg:Obviously, compute infrastructure is not in that number. It can't be. I mean, like, just from the numbers we look at, it cannot have. You know the NVIDIA market. If you look at investments in chips from NVIDIA alone, you know that is a bigger number than that.
Goran Cvetanovski:Anyway, okay, one company For sure we can look at more numbers later perhaps, but it's interesting.
Anders Arpteg:I was just thinking to continue on the Notebook LM and just speaking about some more cool like AI applications.
Sverker Janson:But let me say just one more thing about Notebook LM. So pretty much as I've done with ChatGPT and everything I use it to digest material.
Goran Cvetanovski:You do?
Sverker Janson:Yeah, I've used it several times, this rather boring druggy report. You saw that I threw that in and listened to 10 minutes, as you said, perky summary of the main points and that was delightful compared to having to wait through it, and I think the main points came through anyway. Of course, as you know, since you've tried it, it's sort of slightly superficial.
Henrik Göthberg:But still, you can then basically digest a lot more reading, being out, jogging and running and having something.
Sverker Janson:And in an entertaining, more entertaining way.
Anders Arpteg:I get it you don't have to ask humans right in an entertaining way but can I ask you?
Sverker Janson:can I ask you?
Henrik Göthberg:as experts, do we have any understanding of the techniques applied to get?
Anders Arpteg:to there. I think the next topic will come to get to where to if you look under the hood of this brilliant, of this particular brilliant.
Sverker Janson:So what are the techniques? Because, once again, it's not just one thing, it's a compound. You know, if you look under the hood of this brilliant, of this particular brilliant.
Henrik Göthberg:So what are the techniques? Because, once again, it's not just one thing, it's a compound.
Sverker Janson:I think my little. It's too little to be called an experiment. But asking Chachi Petit to generate a humorous or a nice sort of bantering script for a podcast that showed that, yeah, it does as good a job with the actual Transcript. Actual spoken, the transcript, the spoken content. But the voice generation, the voices that was at the next level.
Anders Arpteg:There are, of course, a lot of text-to-speech, but this is not text-to-speech.
Sverker Janson:This is kind of semantical kind of speech.
Henrik Göthberg:No, it's something else, and so, basically then, if you want to pick apart the techniques here, okay, the text side, the transcript side, we get to fairly fast because it's something similar to what we have seen in the LLMs the voice stuff. What is happening here? What is really going on there?
Anders Arpteg:Yeah, could be something connected to the diffusion approaches for generating images and using it for audio, but who knows? But I'm coming there with the next topic actually. So, if we could, because it's also connected to the notebook element, you know we have a number of new applications coming out from both open AI and Google, which is the advanced voice mode that that open AI has. And now you know Google actually released in Europe even I think I tried it, it worked for me the Gemini Live. So to just describe briefly, you know, if you use the advanced voice mode, which is not released in Europe, similar to Meta's models, it's not released in Europe because of regulation issues, but if you use the advanced voice mode, it is insanely good. Have you seen?
Henrik Göthberg:No, I know about it.
Sverker Janson:Chachapedi advanced. Yeah, I tried it. As I said, I managed to get it to work via VPN two days ago but, then that road was closed. I don't know.
Henrik Göthberg:So elaborate a little bit more. What is the experience? You can perhaps try it.
Sverker Janson:I didn't try it in depth. I I asked it to speak in different voices and, uh, there's a lot of image or videos out there.
Anders Arpteg:So I mean you can do stuff like you can ask it to please help me to learn chinese, and then you you say the chinese text and it corrects you.
Anders Arpteg:Please phrase it like this instead, and and you can really hear you know how you are, you know phrasing things and say does it pronounce it correctly? You can do hear how you are phrasing things and say does it pronounce it correctly? It can do stuff like we asked it to do a bedtime story for my kid and while doing the bedtime story in style with very calming kind of bedtime story, also add some kind of sound effects like whooshing, if it's like water, and stuff, and it can not only do speech, it can really also add like whoosh and stuff like that, and it can sing and it can do so many other things. That is not just pure speech as well, and it sounds amazingly good. It's real time. You can interrupt at any point and start to have a discussion with it and I'm really impressed with what that can do If you go to Gemini though it is not as impressive, but at least it is released in Europe.
Sverker Janson:So I mean we have to make use of that, but that doesn't have the.
Anders Arpteg:It's very clear that Gemini Live, so Live is basically you can press a button and you speak to it and you can interrupt it and you can ask questions and you can have a dialogue. It's actually a very nice kind of user interface where you don't have to text anymore. You just speak and interrupt and have a discussion and that's nice. However, that's very clearly from text land, so the whole text space. It only reads the words. It cannot, you know, change, you know speak in this tone of gothenburg voice or something, um, and if you ask it to sing, it won't do it. It says it will.
Anders Arpteg:I'm singing now and uh, but it's not. And you can ask it to do like a bedtime story they. There is a person on on YouTube who did the same. It compared basically the OpenAI advanced voice mode to live. I mean, if you ask to sing, it starts to say singing voice, blah, blah, blah, now doing this. And if you do it in OpenAI, it properly do it. I mean it can do more than text, but Gemini Live is pure text, it's very clear.
Anders Arpteg:Even though it thinks it's singing, but it's not. Anyway, it's kind of amazing, and this, of course, will be a very useful kind of interface for humans to machines in the future. Imagine sitting in a car or speaking to Spotify or whatnot.
Sverker Janson:Imagine sitting in a car or speaking to Spotify or whatnot. Should we be surprised about this? I mean, it's just a straightforward extrapolation. I'm being a little bit tongue-in-cheek with this because it's great A notebook LM what it does is great, but if someone asks me, okay, so could you do that, yeah, just throw a bit of money on it and you can do that.
Anders Arpteg:I think sometimes people I get a bit annoyed when people are not excited about the amazing AI progress that they have. It's so simple AI today you just put data into it and you have a result. No, it's not that simple, and it is surprising, I would say, that these kind of large language models and these large foundational models that work with nModality actually works. That is surprising in itself, at least to me.
Sverker Janson:In some basic way it continues to be amazing and surprising, but given that we have accepted the progress that we have already seen, at least I am perfectly ready to extrapolate rather a long way from what we have today.
Anders Arpteg:If you take kids or something Just investing, or people that haven't seen the whole history of.
Sverker Janson:AI.
Anders Arpteg:Now they assume that if it doesn't sound perfectly, if it doesn't generate perfect text or do whatever, they are annoyed with it and disappointed with it. I'm still amazed when I see this kind of improvements happening.
Henrik Göthberg:But when I reflect to this conversation here, I think that both these two standpoints are super valid. This is amazing stuff, but I think the profound you know, frustration or question that comes out of this, if you don't, if you see this happening now and people are just doing it, experimenting and it's this good, you need to understand that every single you know computer interface should be questioned within the next five years, and not just computer. Every single interface can.
Sverker Janson:Or a vacuum cleaner. I mean, there's nothing to argue against everything being voice controlled.
Henrik Göthberg:And this makes me wonder why we have such a hard time innovating. I mean, because the innovation here is doing, is done on techniques, you know, and clearly the techniques works and now it's up to us to put them into systems and into into real products and use it. And and for me, it's, it's, it's, it's not if it's only questions of when that is left now and if you want to be part of that game or not. And I I still don't understand why we have such a hard time then as a society, starting to use it more and more. Maybe it's used the ketchup effect here that I think yes, uh, catch, perfect, I mean.
Sverker Janson:So there is a substantial amount of investment being made in in providing these components that are then very easy to use. So I'm sure you are following all these tinkerers out there who toss together cool prototypes and doing this and that and using replitcom to just you know. Yeah, I made this AI application just in 15 minutes. Yeah, I think the ketchup will come out of the bottle to some extent.
Henrik Göthberg:Because it is a topic now that this will seep into everything. I mean, it's obvious that it will, but it's more about how our normal companies can, which?
Sverker Janson:is of course, not exactly the same thing as a fundamental AI transformation of the company's products. I mean, this is lipstick on the pig a little bit if you only sort of adopt speech interfaces, for example.
Henrik Göthberg:But I think the profound topic is that if you have been around the block for a couple of years, we understand that this is now happening on so many different levels in terms of compute infrastructure, in terms of how we manage data. You talked about knowledge graphs, you talked about recorded future example. So all these different components are now coming there to do something that is completely revolutionizing every part of it, and the proof in the pudding that sort of is shown in media and that everybody can relate to is the end user's interface. But make no mistake, everything is happening on all levels here. So if we can understand that, then it's not putting makeup on a pig anymore. But if you only do that part, yes, it is, of course, but I think this is the what is holding us back. You're circling back to that question, right? If it's the money, if it's the know-how I don't know, I always send that back to why aren't we doing more? Why isn't it going faster? For me, that's the fundamental question now.
Goran Cvetanovski:That there's this divide.
Anders Arpteg:It's already going so fast. I mean it takes time for something to come out as a product and to change.
Sverker Janson:It might be that it's an exponential process that is just under the radar.
Henrik Göthberg:Most likely change.
Sverker Janson:It might be that it's an exponential process that is just under the radar and most likely that could be and that will just say boom, in n months or years.
Anders Arpteg:Yeah well, it will be an exciting time coming years for sure and if we were to just go a bit more, like you know, thinking a bit further ahead of what would happen in coming years and the advances we are seeing with, for example, 01, etc. And perhaps you should go authentic as well. I think you're both very engaged in that and so am I, but perhaps we could start with just the O1 kind of direction of AI models. So what's your thinking there?
Anders Arpteg:I mean, my way to usually describe this is usually to say that we have the traditional language model at least, that have a huge amount of knowledge, much more than any human ever can.
Anders Arpteg:And then we have the traditional reasoning models like AlphaZero, alphago, that can beat any human easily in chess or Go and games like that, but they are very narrow, so they are not general and doesn't have any knowledge. So they're really good at reasoning, really bad at knowledge. And one other interesting part of this, I think, is that AlphaZero, for example, trained without any human data at all, so it's all generated data by itself. Now, if O1 is this kind of self-taught reasoner kind of approach, which are generating their own data, this is actually going exactly this direction. So that means that now the models are generating their own data can go beyond the human knowledge because it's no longer dependent on human data to be trained, and then start to add, you know, combine, the reasoning power of traditional alpha zero to the knowledge that language models had. Would you agree with that description or do you think that's a good way to describe it?
Sverker Janson:Yes, but I'm not sure that I'm convinced that. Yet O1, self-taught reasoning. I don't know exactly what O1 does. I mean that's speculation, but that self-taught reasoning that I don't know exactly what O1 does. I mean that's speculation. But that self-taught reasoning that you can bootstrap it entirely on just the human data, I mean I think it's a super important step from the underlying language model.
Sverker Janson:What does it do? It extrapolates text. It's just so happened that if you prompted this correctly, it might actually answer questions as a byproduct of extending text, and then we trained it specifically to answer questions so as to do that more reliably. Okay, so then we have ChatGPT, and that leverages, of course, lots of data. That sort of has this relationship between questions and answers and so on, and similarly we started exploring prompting techniques that did chain of thought and sort of reasoning which just happened to be reflected in the text we have already trained on, and then we can polish that further with various techniques I think we need to which is great. I think it's a sort of a quantum leap to a completely new type of model or a way of thinking about the model. It's sort of we are stepping on to new abstraction levels of what we're doing.
Sverker Janson:So I think we need to. Yeah, some reasoning patterns have logical correct answers that you can produce synthetic data, for example. But let's see how we need to train these models going forward. I'm not sure you can just extrapolate from the internet mass of text the reasoning patterns there. I think we need to add more.
Henrik Göthberg:But let me see if I can. I want to test a a line of thought here with you what the trajectory is here. So you know, looking around the corners, seeing the implication chain and what we're talking about here. So to me, this year we started to see different agentic workflows and you saw Andrew Ng doing simple experiments with 4.0 and getting much better results by taking an agentic scaffolding around the same large language model agentic and now we see if it's self-taught reasoning approaches.
Henrik Göthberg:But what I think you said is the correct analysis is that we are moving up in abstraction level in terms of how we are working with problems and how the machines can automate and how ais can help us. So for me, when you move up in abstraction level, it simply means we're going from one simple task to now we put goals or problems that they can be decomposed into steps. First question is this the general trajectory? Can we assume that this trajectory? We have just seen the very starting point that our systems will be becoming more and more advanced. So we are moving in abstraction levels like this. Will this trend go away? I think it's a rhetorical question. Yeah.
Sverker Janson:I think I used the term abstraction levels and I don't know if it's exactly the correct one. I mean, we are sort of implementing reasoning in a substrate that is a language model. It's a generic prediction engine, and if we implement reasoning, yeah, I don't see what the next step is, but on the other hand, I didn't see O1 coming exactly either. So I think we need additional components to climb up towards.
Anders Arpteg:Do you have any thoughts what that could be?
Sverker Janson:We need more mental state, I think, somehow More memory. I mean more short-term memory or whatever it is that we use to maintain sort of thoughts. We need world modeling but that I think to some extent can sort of the basis, for that is part of the general transformer model, so to speak. But we need to keep in our heads what is the current state when I turn around.
Henrik Göthberg:Oops, sorry yes, anders is still there and so on observe, orient, decide act.
Anders Arpteg:You know maybe, but mental state in a more general sense and I have to add you know I like the yeah maybe, but mental state in a more general sense. And I have to add you know I like the Jan LeCun kind of JEPA approach. Have you seen that Okay?
Henrik Göthberg:we don't need to go in. I think we've spoken about it so much. No, but if I don't go into the techniques, but I want to understand the general trend.
Henrik Göthberg:So I then can build a mental model for what do we need to think about when we organize business, when we organize, build systems, and I have then, like a bet, data. What can we observe? What is the insight here? What is our beliefs and bets? So I put up a belief that I think we're going.
Henrik Göthberg:Whatever techniques we are figuring out, we are on a trajectory that humans will be working on a higher and higher spectrum, labor on bigger problems, and the systems and machines will help us. So there's a very interesting phenomenon happening then, because the fundamental human-machine interaction is different. You very soon get to a principal-agent relationship with your machine. You're the principal, you're the prompt, he's the agent, he's doing the work. This is a very different type of relationship to how we have had relationships to our systems in the past, and if this is where we're going, that has fairly big implications for how we define agency of the AI and understanding agency of the human or the teams and how they are aligned, or the potential risk when they are not aligned. So I'm on a trajectory here. What fuck-ups? What do we need to change in terms of thinking around this?
Sverker Janson:It's still the case that you need to know your stuff to use, let's say Chachapiti, efficiently.
Henrik Göthberg:Yeah.
Sverker Janson:You need to recognize good output from bad output, so you need to be an expert already. So you are not in fact, at a higher level of abstraction when interacting with Chappity, it might appear. So you can, for a little while, get by saying something at a higher level and it produces something. Let's say code, and you run that.
Sverker Janson:But if you aren't an expert capable of verifying that code is okay, you are in problem this is interesting to see, because if we don't get over that threshold so that we can actually fully trust the output of the AI, we will enter an interesting situation where the people who are already experts in whatever it is software development will put everyone else out of the job who is on a career track of actually learning to become experts, and so in a little while we will have no sort of future experts who can handle this beast. This is just sort of imagined.
Anders Arpteg:You can think, I think, in two ways here. One, some people say that okay, in the future, when we have this kind of amazing AI system that we can have doing the advanced stuff, I need to know less, and I think that's the opposite. Actually, currently, definitely, and that means basically that we have to learn more than we did in the past to be able to be a useful person at least. But with AI, we can learn more as well, but it still means that we need to actually raise our game as humans, and we can with AI's help. But it's not that we can reduce or remove that need for us to learn more. We need to learn even more in the future. I would say.
Henrik Göthberg:Let's take a very concrete example, because I think this is quite profound.
Henrik Göthberg:So let's say, now I want to build an AI in a function in an organization let's call it sales and I want to have some sort of AI helping me plan my. You know, I'm not going to go into detail, but what you're saying now is that, on the one hand side, the team has only had sort of sales competence before and ultimately, our domain experts know this, and now we want to do math or whatever to understand with, with data and AI, how we can automate the route planning, whatever. But what you're saying now, in order to, for the foreseeable future, we actually need to have more competence. Maybe we need to be a team. There needs to be some AI experts overseeing or quality checking and understanding what the AI is doing. So for me, this is a huge topic around the agency of what you know. What is the scope of the more now complex AI system doing in terms of planning and sequencing in relation to a human in the loop, using that AI as a team member and within that team that is using that as a team member?
Sverker Janson:you need to have the people that might be the users of that data, but you kind of need to have people in that team actually understands a little bit more under the hood if something is completely haywire and you need the long-term plan for making sure that, as a company at least or a big company that you always have experts of this kind who have that deep understanding, who can have that role, and that has traditionally been that you're a junior team member who would do what the AI now does. I remember a comparison here. I think it's correct that when car companies started having robots paint cars, at least some companies put in a production line with human painters just to maintain the understanding of what's actually going on and who can program the robots in the correct way, exactly?
Sverker Janson:Who can program the robots? Yeah?
Henrik Göthberg:So if we're going on this abstraction journey now and we think that's the future techniques will come and we have only seen it improve, so we can assume it's going to be sold Then how we organize teams and teams composition and the agency around them to be able to work with these machines. This will be impactful because if you have a business organization today where you have a functional division of labor, where you only have one set of competencies sales over here, more technical competence or data competence sits over here they are not really in a state to have agency around what the ai can do. So you kind of need to now what we have learned, like how do you build a cross-functional team, how do you build product teams? That becomes kind of important in the core business if you're going to go the software route.
Sverker Janson:Depending on what happens with AI. So I mean maybe in 10 years? The AI is perfectly reliable. You can trust it as a human expert and it's fine.
Goran Cvetanovski:And then you don't need to be's fine, and then you don't need to be an expert, and then you have really reached that fundamental step yes.
Sverker Janson:And that should happen, but the question is when.
Henrik Göthberg:But we will have a transition stage. For sure, for sure.
Anders Arpteg:It could go in the direction that we hear. You know sam altman speaking about you know it's they are. I've heard someone saying they have a bet, saying you know what year will we have the first single person unicorn company coming along?
Henrik Göthberg:interesting thought experiment actually we.
Anders Arpteg:I mean there's been unicorns with, like, I think, 10 people or 16 or something.
Sverker Janson:I think Instagram was like 16 or something, I don't know, but anyway a single person, I guess some kind of power law or something like that will suggest that there will always be sort of unicorn companies.
Anders Arpteg:But I'm just also seeing a future where so many people are doing so many things that you know who can attract that kind of attention to become a unicorn, but call it a 100k company, then but anyway, the point the point is that humans take on a broader set of tasks and they perhaps more deal with the overarching like goal rather than the specific details which is delegated to a set of AI agents instead. But I think.
Henrik Göthberg:I think this is back to what we discussed. There's a transition phase where even you can build a five. For me, it's plausible to build a three-man or five-man unicorn today, someone who has the technology. We need one CTO, we need one CPO and we need one CIF founder sales so there are some key tasks to be done, right, but you can probably almost get away with that right now if someone is really proficient in technology but I think this is still, you know, uh, aiming too low at one point.
Sverker Janson:Why not a zero person?
Anders Arpteg:yes, I know no, but really absolutely yeah, I mean you invest some money and then you tell the agents to go and do it it's a bit more unethical perhaps. No, no, no, I don't think so. Now we're getting these.
Henrik Göthberg:Philosophically it has a very interesting one.
Sverker Janson:The only people left are the investors. They could be automated.
Anders Arpteg:I hope they can. You know who's going to be automated.
Henrik Göthberg:First the investors or the unicorn?
Sverker Janson:or the people with the money. At least I don't know the owners very meta.
Henrik Göthberg:You need a pension owner class no, the pension fund and their algorithm and their startup machine.
Anders Arpteg:Yeah yeah, the AI is investing in other AIs and that becomes a market, I guess, at some point. Oh, crazy idea. Okay, time is flying away. I thought we would start to round off a bit.
Henrik Göthberg:We are already entered into the more philosophical realm here, so we have already started going. This is an interesting topic to extrapolate. Where are the potential scenarios?
Sverker Janson:How speculative do you think it is what we were talking about? I mean, are you saying philosophical, because it's a mind game, or are you saying that? Do you believe in the scenario of the, let's say, zero person?
Henrik Göthberg:Let's go here I like it. I mean like for us, if we have had, I mean like if a typical roundup and the sort of the standing research question that we are building our statistics on is where people think about. You know, how do how? Where do they put themselves on the spectrum as an ai doomer versus an ai utopia?
Henrik Göthberg:boomer, yeah, boomer you know what I mean and then and then you know a little bit like do we understand this as a singularity? Do you understand this as a very fast takeoff stuff like this? So, speculating on that, very seldom when we talk about this are we discussing if it's more a when discussion and a dystopian Will we fuck this up on the way, Is it really AI singularity we should be scared of?
Henrik Göthberg:Or is it we screwing this up as humans on the way? You know we couldn't regulate it, we couldn't fix it. We got, you know, an AI divide, economic divide that is leading to revolution. You know, there's so many scenarios where you know.
Sverker Janson:Or the perfect oppression scenario. Yeah, the perfect, you know.
Henrik Göthberg:So, but I think going back then going so, that's what we usually do, but let's go in, so going in that direction. Is this a philosophical question? It's a speculative question in relation to where we stand today, but I think it's a very important conversation to understand a little bit. You know, what does it mean to become data and AI ready? What does it become? I think sometimes we have a we talk about AI ethics or AI regulation, but we are not really talking about how the companies should innovate or organize themselves or think about these topics. Because if you think about these topics, you can kind of look at what you're doing today and say like you are not organized for dynamic adaptability, you're not organized for modular you know navigating an opportunity landscape You're organized for efficiency. In one old game, you knew that's the main problem.
Sverker Janson:But it's not just a matter of reorganizing a company as is. I mean that will get you nowhere fast. No, it's a fundamental you need expertise, experience, talent, not least talent to manage the challenge of doing this transition.
Henrik Göthberg:I fully agree. So for me, reorganizing is not about moving around the points that we have Shuffling the organizational chart. That will go nowhere. You need to reimagine what this enterprise is all about, and then you need to staff it and organize it accordingly, which is something else.
Anders Arpteg:What's coming up for RISE, you would say, given a zero person unicorn company in the five years.
Sverker Janson:Ignore that comment. Rise in general no RISE and the center of applied AI.
Henrik Göthberg:How should you be working on the right stuff? What is the right stuff to be working on if we take this more long-term view?
Sverker Janson:It becomes really tricky if we take the long, long term. That is sort of speculative today, although I think it's certain in some sense, in a futuristic sense, but it's interesting.
Henrik Göthberg:We are used to doing in large corporations strategy, five-year strategy, five-year plan, five-year outlook, three-year outlook, one year like this. So let's take a five-year outlook on where we are on the trajectory, but it's much more volatile than that.
Sverker Janson:That's what makes it so different. So so I, quoting jeff finton on on numbers, not for any specific question, he said, yeah, in a few years, or at least within 20 I mean uh? That's. I think that's. Those are the numbers we're talking about for many things. Really big things could happen in just a few years, given the pace of investment but so then we are talking about volatile, uncertain, complex, ambiguous we have all the different scenarios.
Henrik Göthberg:Yes, so then you need to go into another type of strategy work I mean like even scenario work like vattenfall. We do, you know, we do scenario work right and we have three scenarios. This is something else. So to strategize on how to build a company here and now that can go into an opportunity landscape of that breadth, how do you do that? How should we think about that? I can't even imagine.
Sverker Janson:I think maybe you need to organize for being very agile and adaptive.
Henrik Göthberg:I think that's the core topic. We're going from economies of scale and for learning. Yes, we're going from economies of scale to economies of learning. We're going to from extreme focus on dynamic adaptability versus efficiency. Efficiency is dead as a metric. It's about adaptability. I don't know well.
Sverker Janson:I mean, with a little luck, you will be very efficient if you manage to adopt these technologies.
Henrik Göthberg:Economies of learning beats efficiency of scale, because economies of learning means magnitudes more efficient.
Sverker Janson:And then it is more efficient, yeah, but if you stare From a different short-term or long-term perspective.
Henrik Göthberg:If you stare at the efficiency metric. You're going to do a point percentage improves If you look at economies of learning and dynamic adaptability. You're looking at innovation that fundamentally blows efficiency.
Anders Arpteg:Efficiency of the current process.
Henrik Göthberg:It's still efficiency, Of course it's efficiency but when you use the word efficiency, people think about efficiency of the current process and dynamic adaptability is efficiency or profitability of the current business model. Please don't do that. Do something else.
Anders Arpteg:So what about the future of central, overplayed applied AI? You don't have to think about a zero person unicorn company, just anything specific coming up that you're excited about for zero person, you know, uh, unicorn campaign, just you know anything specific coming up that you're excited about for the center of applied ai so this construct is not rigged for very disruptive change of course I mean it's an incremental uh activity at the sort of a fairly small scale compared to the big numbers one can toss around.
Anders Arpteg:Let me make a statement here. I would say that the center of applied AI will look more or less exactly the same in 10 years 10 years?
Sverker Janson:No, because it's only planned for another three years. Okay, no, no, no, I don't think so. I don't think so at all. I think within just a few years, the ketchup is out of the bottle, and when it comes to understanding the profitability of adopting these methods, this will just become painfully obvious. So I would be in 10 years. I would be surprised if rice will probably exist in 10 years, and I would be surprised if the corresponding activities will not be 10, 20, 50 times bigger in 10 years, in just a few years. No, it's the usual thing that you tend to exaggerate sort of short-term change, but long-term change, I think, will be profound actually.
Anders Arpteg:Do you mean for the center or for the site?
Sverker Janson:Not necessarily. We currently have a certain. It doesn't have to do with center. It has to do with RISE as an AI organization and we already have a sort of line organization with groups driving these areas. The center is just a program to accelerate this, but for RISE as a whole, I'm very sure that AI will be more than 10 times a bigger phenomenon.
Henrik Göthberg:It would be strange if it's not. But the argument then is, if you imagine now the AI part of RISE, how you organize that, if you basically now okay take a simple stance on this.
Sverker Janson:Okay, we need to go big on this.
Henrik Göthberg:You would probably start sorting this slightly different than today.
Sverker Janson:I think we still have a corresponding back office, sort of super experts in methods, but I think AI will have started to become integral parts of the way of working of all the parts of Bryce. Within 10 years.
Henrik Göthberg:certainly there may be one or two areas left With the model, the way you approach this now, would that scale tenfold, or where would you need to adjust to scale it tenfold? I think so.
Anders Arpteg:I mean you already are doing that. I mean you have more of a hub and spoke kind of starting point so that will expand and that, I think, is clear starting point.
Anders Arpteg:right, so that will expand and that, I think, is clear. The point I'm trying to make here a bit is and I would like to see if you agree or not um, I think we all agree that you know, the technical innovations that we're seeing will be really, really fast and, you know, in a couple of years, three, four or five years we will have amazing ai that can reason and be very agentic and do a lot of things that we can't do today. That will be extreme. My point, then, is that societal and organizational changes will take much longer time.
Sverker Janson:They always do.
Anders Arpteg:Yes.
Sverker Janson:And I mean, even if we say no, no, no, it's going faster. Yes, some things are going faster. Uh, just look at the adoption of internet and whatnot.
Anders Arpteg:Um, so I mean, so many people are afraid. Oh, and if we have an agi now coming in x number of years, we will lose a lot of jobs and everything will change.
Sverker Janson:And I don't believe that's that's.
Anders Arpteg:That's a bit my point. I think you're right. Uh, um, the the losing all the jobs will take longer time.
Sverker Janson:You're right Losing all the jobs will take longer time because it requires other technology revolutions, in terms of robotics, for example, and 3D printing, things like that.
Henrik Göthberg:And then history has shown us that when we get new tools and we work on different abstraction levels, not necessarily has the job disappeared or some has, but a lot of them has changed and other jobs have come up instead. So the factory looks very different.
Sverker Janson:But is it the natural law, I wonder Natural law, I mean it's a totally different ballgame now with the technology that actually can replace cognitive and humans what humans do. Yeah, so the core.
Henrik Göthberg:The core question is is this a different ball game altogether? Because it's a cognitive ball game.
Sverker Janson:We are not replacing muscle power, we are replacing brain power, augmenting brain power, and that might maybe fine motoric skills will take a little longer, because this is apparently more difficult than developing ai, but that will happen as well of course.
Anders Arpteg:So cognitive first, but then also to more physical systems as well, but changing this site and companies, I think will take decades. Yeah, okay, should we end there, perhaps with the with the you know standard question yeah, let's, let's.
Henrik Göthberg:I.
Anders Arpteg:I hinted to the standard question, but let's ask it properly, like you always do well, assuming that that point will happen, then we don't only have a cognitive agi, but potentially even a physical agi. Um, what and? And you can think about two extremes here. One is that this will lead to the destruction of humankind, a very dystopian future, with the Terminator movies and the Matrix coming to life. Or perhaps we start a new war that is being driven by AI that kills us all in some way. Or the other extreme, of course, is a utopian one, where we have a future of abundance, as many people say now, which basically means that we have sold the cancer, perhaps the post-scarcity society of.
Anders Arpteg:EAN and banks.
Sverker Janson:I love the culture novels. They are so good. Yeah, the post-scarcity, yeah, but they require also this, you know motor skills, robots.
Anders Arpteg:Yeah, yeah. So this is the mission. So where are you on that scale between a utopian and dystopian future?
Sverker Janson:Yeah, I believe they can, more or less. I mean not the full-blown dystopia cannot exist, coexist, but I mean today people, and today people are living in abject poverty and misery and some are very rich, and this will continue, so we will have a mixture of utopia and dystopia for a long time to go.
Henrik Göthberg:You answered this in a new way. Do you think there's a scenario where this is actually not either or no?
Sverker Janson:I mean, some people will kill other people with AI robots, as they are killing them in other ways today.
Anders Arpteg:Will there be less poverty in that future?
Sverker Janson:There could be, of course, but the question is who owns the productive technology? Who controls it? Who cares enough? Who distributes the wealth? Who distributes the wealth? Who cares enough to distribute the wealth? Sam Altman, I don't know. Maybe if someone will tell wonderful stories about him after he's dead, maybe he will.
Anders Arpteg:Yeah, I'm eager to go to the OpenAI story. We didn't do the news section. We forgot about that. No, we should be. Ah, no, it's not, it's too late.
Henrik Göthberg:No, it's done.
Anders Arpteg:Cool, okay, that's interesting.
Henrik Göthberg:That's an interesting angle. That was a fresh angle because we asked this question at least 50 times. It's an interesting thought experiment and to hear people's answer to this Yours was very nice, very interesting.
Anders Arpteg:We take the news off camera, so with that, I'd love to thank you very much, sarko Jansson, I hope you stay on for a bit of an off-camera discussion, but it was a true pleasure to have you here and thank you very much.
Sverker Janson:Thank you so much, thank you.
