AIAW Podcast
AIAW Podcast
E145 - AI is transforming journalism - Hans Hjelm and Magnus Engström
AI is transforming journalism—how are newsrooms adapting? In this episode of the AIAW Podcast, Hans Hjelm and Magnus Engström from Bonnier News discuss the impact of digital transformation, the rise of generative AI, and the role of data science in shaping the future of news. From AI-written articles to deepfake detection and the newsroom of the future, we explore the challenges and opportunities ahead. Tune in for an insightful discussion on the evolving intersection of AI and media.
Follow us on youtube: https://www.youtube.com/@aiawpodcast
As an assistant, and that is I mean. Maybe we can touch upon that later as well. But if you have the AI as a separate assistant, that is something that you use that is disconnected to everything else. In that case, of course, you can write a very simple prompt and you can get something out of that, even if you spell something wrong or it's not completely logical what you're writing. But if you would like to use the ai as an agent in a system, right so it's a component in a like an ecosystem, for example.
Magnus Engström:Right, in that case, I feel that I wouldn't trust a prompt that is not constructed in a way where it's super obvious for me as a human, as a programmer, that, all right, I completely understand what is expected for the agents to do here, for example, yeah, and and so now this makes sense.
Henrik Göthberg:The evolution of prompting has really taking a second or third gear in the context of starting to build multi-agent flows and stuff like that.
Magnus Engström:Yeah, and I think that this is also what you see, if you look like, for example, when you get into really advanced philosophy, for example, from an academic point, in the end you end up with mathematics, right.
Anders Arpteg:You end up with mathematics, right, you end up with mathematics yeah.
Magnus Engström:So, for example, even if you do, for example, linguistics right, in the end, if you're writing a like a phd in advanced linguistics, it's gonna be equations. Right, it's gonna be algebra. So I think that this is true for everything, right? So, even prompt engineering, sooner or later, when you really really need the agent to perform something that is quite complex, you will need to turn into more structural prompting again.
Henrik Göthberg:Because this is so many angles. So if you start with the reflection on how has prompting evolved, reflection on where has prompting, how has prompting evolved? We did programming, we left programming and now we're getting back to pro-programming, like prompting. It's such a big topic because when we say how we can replace engineers and we can replace programmers, I've been trying to say that that rhetoric is really dangerous because we are augmenting the way we program. We're augmenting the roles that programmers have. That okay, maybe we need less engineers as a consequence, but the way to look at problem is not a replacement problem, it's an augmentation problem. So I think what you're saying is literally like oh, programming has never been more important.
Magnus Engström:No, and also going to the data science side of Bonnier Bonnier News, for example, you will. I mean, I see a lot of people who are prompting language models and are being very nice saying hello, can you help me with this? Thanks, and so on. Language models are being very nice saying hello, can you help me with this? Thanks, and so on. Basically the same thing as thanking Google for the search results.
Anders Arpteg:Do you say please when you write to them no, nothing, I don't say hello, I just write exactly what I want and I feel a bit bad when I do that. Yeah, and I never ask questions.
Magnus Engström:I only give instructions right. And if you look at, for example, from the data science side at Boninus, the prompting that takes place there is basically just very short but very clear instructions. We don't act as the language model actually is an entity. It's just a part of a process.
Anders Arpteg:Now we get into psychology of AI, which is actually a new field, I think, Potentially adding some kind of more emotional kind of statements like oh I like you very much, Can you please provide it? Perhaps it could actually influence the interactions and the output of the model.
Henrik Göthberg:Still right, but it's a deeper. You touched on it, but I think it's a quite profound message that you slipped in here. We don't ask it questions, we instruct it. That is some unpacking around that how we think about our agents and our two way to use responsibly.
Magnus Engström:Yeah, it's a sequence model. It's nothing else. It's a sequence model.
Anders Arpteg:It's more than that. I think it's a bit weird to say it's just a sequence model. I mean, it's like saying it's just statistics or a statistical paradigm. I think it's much more than that.
Magnus Engström:I don't agree. Okay, but I don't agree. I mean, one argument can be how humans act, for example. The thing here is, of course, that most people I think there's multiple studies on this most people don't know how they're going to end the sentence when they start speaking. Right?
Anders Arpteg:so saying that, basically, language models are a sequence model well, I could also argue for that, like the language center in the brain is also just a sequence, yes, yeah, and that would be a bad description of the brain, and then it would also, I would argue, a bad description of the brain, and then it would also, I would argue, be a bad description of an animal.
Goran Cvetanovski:This turns out to be an interesting discussion.
Henrik Göthberg:This is going to be a great talk. I'd love you to disagree.
Anders Arpteg:This is also awesome. Well, with that, I'd like to welcome you both very much here to the AI After Work podcast. Let's start with Hans Hjelm.
Magnus Engström:You are the product owner right In the data science team. Exactly Awesome.
Anders Arpteg:And Bonnier News. And you're also something which I don't know what it is, but I think perhaps Henrik or Goran knows You're a Krautrock guitar player.
Henrik Göthberg:Krautrock.
Goran Cvetanovski:Krautrock, I have no idea what that is.
Hans Hjelm:Craft rock is rock music made originally in Germany in the 70s. So think of music that's like craft work, but played on electric guitars instead of synthesizers. Cool, cool, very cool.
Anders Arpteg:But you also have a PhD in computational linguistic from Stockholm University like 15 years ago, right Something. Yeah, and you're also supervising another PhD student, a WASP PhD student, yeah, and that's something I'm looking forward to hear more about shortly. That's super cool and, before we get more into, perhaps you can describe a bit more who is really Hans Hjelm.
Hans Hjelm:Sure, I started out trying to become a musician actually, so I studied music for a few years. I moved to the US and I lived outside of Dallas, and I studied music for a few years and realized there were too many other guitar players that were much better than me, so I had to switch to something where I could make a living, and so I switched to the obvious choice computational linguistics.
Henrik Göthberg:That's so obvious, yeah.
Hans Hjelm:So I got my master's in computational linguistics in Gothenburg.
Anders Arpteg:Yeah, so I got my master's in computational linguistics in Gothenburg, and after that, I worked for a few years for some startups in Germany and moved back to Stockholm to do my PhD.
Hans Hjelm:And what was your PhD about? If you just briefly describe, it was before deep learning and but my topic was ontology learning and cross language ontology learning. So basically I was looking at what we now called embeddings and trying to see can I use the information in embeddings to create a taxonomy of words or terms and what happens when I do that in Swedish compared to English, compared to French? So I looked at a lot of texts from the EU and tried to learn these semantic representations of words and terms, basically from that.
Anders Arpteg:And this was before Word2Vec, I guess Exactly this was support vector machines.
Hans Hjelm:Using support vector machines. Random indexing was a popular technique back then.
Henrik Göthberg:Cool stuff. If you would have done your PhD today, how would you reframe the question and what techniques would you be looking into?
Hans Hjelm:Oh, that's a tricky question. I mean, back then you had to program a lot of the basics yourself, so I was basically doing like singular value, decomposition and all these mathematical exercises that you now get out of the box, so you could have just started from where I ended, pretty much, I would say.
Magnus Engström:Is computational linguistics still like the standard move for people?
Henrik Göthberg:guitarists or has that changed? No, no, no, no, it's social media, all right.
Anders Arpteg:It's still the same.
Magnus Engström:Yeah right.
Anders Arpteg:But at some point you started up on your news. How long ago was that?
Hans Hjelm:I started in 2017.
Anders Arpteg:17. Good, and if you were to describe your current role, what are you currently doing there?
Hans Hjelm:Yeah, so I work as the product manager for a data science team. We are five developers in the team plus me and we work with the whole organization. So we work, for example, with the advertising department, trying to create special segments for advertisers for different markets. So we have segments for the car industry or for fashion or for food, for example, cool. We work with the newspaper delivery so we try to optimize. Optimize the number of newspapers that get delivered to each point of sales, for each day.
Anders Arpteg:How long do you think you will continue to deliver papers?
Hans Hjelm:I think it will last longer than people think. When I started in 2017, people were saying isn't this going to end soon?
Henrik Göthberg:Actually, I went from digital to back to a paper quite recently.
Hans Hjelm:Just because I wanted the premium product on the weekend.
Henrik Göthberg:Right, I only got the weekend product then.
Anders Arpteg:Yeah, awesome, let's get a lot of topics there to get into, and how you work in Bonyin News and what they do, etc. But before that, I'd like to welcome here as well, magnus Engström. Let's get a lot of topics there to get into, and how you work in Bonnier News and what they do, etc. But before that, I'd like to welcome here as well, magnus Engström. Thank you.
Magnus Engström:And you're the head of architecture and data. Right, yeah, that's right. And at Bonnier News before that, I was head of data and some parts of the development team. Some of the development teams at Mitt media was acquired by bonnie news, so that's how I ended up staying. Yeah, yeah, that's how I ended up at bonnie news. How long ago was that? Oh it's 2019, I think right. So I've been at mid media for quite some time. I started off as a system developer, tech lead at mid media first and was and was Mitmedia.
Henrik Göthberg:You were working at Jävle then too.
Anders Arpteg:Yes, that's right, and you started as a software engineer or software developer. Yeah, exactly.
Magnus Engström:I started off a classical college dropout. I was actually aiming to be working with sports and like sports science, and then I quit that. That started as a web developer instead, because it was I had time off during the summer and then I just never went back to um, so that was my only time working, or in the academic sector actually. Then I took some certificates and so on later on. So but but basically that's how I started and then I worked as a web developer and less and less web and more and more systems, and then that's the power, and then and then data science uh, yeah, well, data science.
Magnus Engström:I don't know how much the term data science was actually a term, uh, back in like 2012 or 13.
Henrik Göthberg:so, but it starts then. It starts then, but it's not used very much in sweden.
Magnus Engström:No right, yeah it's worked a lot with advertisement, a lot of optimization. I think that that's something that me and hans have in common right, I also also worked with that, yeah yeah, I think that's actually, I think that's uh, I don't think it's a coincidence, I think that's a good start. I mean, it's huge data sets yeah and uh, you always it's very close to the business side, always you kind of exactly know what the goal is, right for something, yeah it's the objective.
Henrik Göthberg:Function is quite clear. Yeah, and now you work as a head of exactly know what the goal is the objective function is quite clear Now.
Anders Arpteg:you work as a head of architecture and data. What does that mean?
Magnus Engström:Yeah, well, if we build things at Boni News that we have intention of scaling within Boni News and Boni News is I lost count it's over 200 brands by now. So, if we're going to scale things, the pony news, not always, but usually I'm involved like from a central perspective then, and then actually I think it's official today, or it was yesterday maybe yeah.
Magnus Engström:So we have kind of restructured a little bit at Bonnie News. Now we have a domain where we have the data science team, the Atom team working with like the agent layers at Bonnie News Agent layers and the Bonsai team they're working with like smart journalistic tools for example, and that is a domain now and I'm the main lead for that and together with me in like the leadership of that domain also haunts person we need to dig into that one, because there was so many exciting agent layer and what you know decomposing how you, how is how you frame this domain or structure.
Henrik Göthberg:It was very interesting just what I heard yeah, that's certainly true.
Anders Arpteg:Should we perhaps give a quick introduction to what Bonnier News is? Yes, you started to speak a bit about that 200 brands, I think that's that just we can start there.
Henrik Göthberg:I don't think people know what Bonnier News is. I didn't know when you flip it like that.
Magnus Engström:Well, bonnier News is well by the name's. It's mostly news brands. It's not only news. There are some lifestyle brands and so on and some other other types of businesses within the borneo news company as well, but mostly it's like news brands. I think borneo news is, I think with quite a good margin, the largest publisher in Sweden. So, for example, expressen, dagens Nyheter, dagens Industri.
Hans Hjelm:I am a customer.
Magnus Engström:Yeah, all the local news from Mitt Media and the local news from Gota Media and Hall Media. There's a lot of local news.
Henrik Göthberg:So the Boni News is the group that owns a lot of different companies or brands.
Magnus Engström:Yeah, yeah, bonnie publications in denmark, huvudstadsbladet in finland, for example.
Henrik Göthberg:Then there's a lot of other these and these are more classical news yeah but there are other types of brands.
Magnus Engström:There are digital brands here as well, right yeah, there's some purely digital brands um, and some of those used to be um magazines as well, and some are still magazines, but, for example, with lifestyle brands. You have, for example, technique and svad um Altomat. I mean there's far above 200 brands. Did I miss anything? Any major ones? No, no major ones.
Henrik Göthberg:But several of these brands are in fact known publications like Technicus.
Anders Arpteg:Yeah, that yeah. And how is it organized? Are you then having a set of centralized teams, like the data science team, et cetera, that works across all these brands, or how does the organization look like?
Magnus Engström:Yeah. So more like we move towards a more of a I don't want to call it a centralized way, more like a collaborative way, like we have a product domain and within the product domain, for example, we have the different news products that are collected. We also have a commercial domain and you will find two legs of that. It's like the advertising, the sales part, and then we have like the subscription, like the business domain basically, and then we have like a central tech domain, for example, and there we have things like systems and tools and things that we see that will probably be used equally across all those domains. But then, of course, we have a huge part of bonnie. News is the journalistic part, like the journalists and reporters, and so on and they are, are they almost?
Henrik Göthberg:always connected to the different applications or are they also some parts journalistic role, part of Bonnie news, or is that more decentralized, so to speak?
Magnus Engström:How would we describe that? It's somewhere in between, I would say Mostly decentralized, just given the scale of the local news and the different content, like from cars to food.
Anders Arpteg:So a reporter is normally connected to a single brand, but you have a set of centralized-ish domains that works across the domain, or brands?
Magnus Engström:Yeah, and we try to, like it's from case to case. I mean, some cases are narrow enough that they should probably happen at the brand level, for example, and we also have business units to make this a little bit more complicated. Of course so business units is like a subset of brands, for example, so expressen is a business unit. Dagens nyheter also is a business unit, so kamrat posten, for example, is a part of the dagens nyheter business unit yeah, but I, I recognize, I recognize this in scania.
Henrik Göthberg:You, you have Germany and UK are their own business units, whilst the business unit Asia consists of Singapore. So to group them to get some sort of balanced view on business unit level is quite common.
Magnus Engström:Yeah, all of this will be in the test that will be given after the podcast.
Henrik Göthberg:I have my chat GPT, is that okay? You know how we start every meeting now.
Anders Arpteg:I mean awesome, but the teams that you work with is mainly in one of these domains. Is it called like a tech domain then?
Magnus Engström:Or what's the yeah, well, yeah, of course it's a part of the tech area. I would say we have domains and we have areas. Just to make it a bit more complicated, actually, it actually makes it easier, at least when you look at it from an organizational perspective. But in our case we have the data science and AI and machine learning domain yet to be fully named, because we haven't gotten to that part really yet. But in this case we have so strong bonds and dependencies between those teams that it just kind of made sense to structure it this way. But already, for example, the data science team has grown a lot during the last years because there's so many different parts of Bonnie News that are depending on the data science team. But that also makes it hard to place it in one domain. So we have enough different dependencies now to just logically make sense to put it in one domain.
Henrik Göthberg:But have you had the conversation? I mean, we have used many different words on this, on Hub, and Spoke this whole conversation on how do we organize platform teams that provide services for value stream or domain teams, and then we have in this book that we have. So we you know, so I work a lot with Scania and here we end up with engineers and data scientists both in the central platform teams and also embedded in the different. Is this happening also in Borneo News when, as you say, data science or analyst or data work is happening?
Magnus Engström:Yeah, both things. I would say we have a strong central presence of data science teams and so on, but I mean there's still, to some extent, things like data science. But I mean there's still, to some extent, things like data science. I mean, first of all, the term in itself is a bit hard to kind of frame, but I would argue that some data science happens in different parts of organization. But also I would say that data science is happening in the central analytics team also, because where do we draw the line?
Henrik Göthberg:This is an ongoing conversation on how to solve this.
Anders Arpteg:I think the organizational topics is really interesting and it's easy to get stuck here as well, yeah.
Magnus Engström:But I mean, if you want to get a little bit better understanding for kind of what is a typical thing for the data science, I mean maybe Johans could talk a little bit about, like, what is actually data science? In Bonini's doing Some example product Instead of talking about organization, because that will really not help to understand what is actually going on. Yeah, I love it.
Anders Arpteg:It's still interesting, but still products would be super cool. Perhaps we can give some examples.
Henrik Göthberg:I think it's great to talk about the real stuff and then how you organize around this concrete product or topic or use case. That's easier to understand for sure.
Hans Hjelm:So our current focus product, I would say, is personalization, and the idea is to have, let's say, one of our brands wants to get started on the path, then we want to offer them like a white label solution.
Anders Arpteg:White label solution. What does that mean?
Hans Hjelm:So we make one solution that is good enough for all brands, so it's supposed to be able to be used by both Expressen and the local news and the doggens industry. But if they want something that's bespoke, so that's specific for just one brand, then they have to build that on their side. Basically, so we're from the.
Henrik Göthberg:I mean it's not as clear-cut as that, but that's our, our main intention at least but it makes sense because if you're going from the complete newbie they haven't started, they want to start implementing then you get the generic personalization engine, Exactly. And then, when you get a generic personalization engine as a service, you start using it, you're maturing and then you realize you need to tweak it, fine-tune it, you need to add the next layer of features and then you might end up in a very brand-specific or very content-specific fine-tuning. Yeah, Is that summary?
Hans Hjelm:That's a very good summary. Okay, and what do you?
Anders Arpteg:personalize. Can you give some examples?
Hans Hjelm:Oh, for example, one of our biggest brands is Expressen, and if you read an article at Expressen, then when you scroll past the end, we show a list of suggested further reading articles below the starting article.
Henrik Göthberg:And this whole field is so interesting, how it has evolved. How dynamic is the personalization engine today? Starting article. And this whole field is so interesting, how it has evolved.
Hans Hjelm:How much is now? How dynamic is the personalization engine today? I mean, it's very much under development, I would say. But we have seen, just comparing to. I mean, we started out with some very simple algorithms like what's the most trending article currently, or what's the most popular in the last 24 hours, or something, and just comparing to that, we see a large lift switching to the personalized algorithm.
Anders Arpteg:I get the recollection from my old spotify days. You know we have playlist extensions, not that you can do. Can you speak anything about, like in general, about the techniques that you have experimented with?
Magnus Engström:yeah, I'm thinking that probably we can, and I also feel that there are some things that could be somewhat controversial, so let's talk about them.
Hans Hjelm:I love.
Anders Arpteg:that angle, that twist was awesome.
Magnus Engström:First of all, I would say that I mean how well does a personalization and your needs need to perform? What is the benchmark? Benchmark Right now, by the look of it today, the personalization first of all. Hans is very humble here, but it's rolled out at scale.
Anders Arpteg:Okay, so it's been used for some time.
Magnus Engström:Yes, yes, a lot, and by the looks of it, it's by far the best personalization we have deployed at Boninus, and it's performing very well. Even if you compare it to other types of content selections, for example lists or manual content selection and so on, we see good results. The thing here is, of course, that it doesn't really need to outperform anything. It just needs to be as good as a manual selection, for example, so it doesn't need to be better, it just needs to be matched. You mean to?
Hans Hjelm:reduce manual labor.
Henrik Göthberg:Yes, exactly To reduce manual labor and then be able to scale it in some ways infinitely then you have won. The rest of this is used, that is margin.
Magnus Engström:Now you're making margin.
Henrik Göthberg:Exactly and how much more you know. So you won but, how much are you going to win with?
Magnus Engström:Yeah, and I think that this is like the thing with a lot of different AI systems that you kind of expect them to perform better, but they don't really need to perform better. But they don't really need to perform better, right, they just need to perform at the same level. I mean, this is like the hard problem with talking about self-driving cars, for example, is that, well, the bar is set to no accidents, right, but I mean, realistically, it only needs to perform as good as, like what we have today. Right, we don't have zero accidents at the roads today, right?
Henrik Göthberg:now you could. You could argue that I mean, like, in order to make sure that we are pushing it, it should be better. But you know it can. It can never be zero. No, I mean, I think you're trying to make a quite profound point here that we are. We are looking at it wrong sometimes, think you're trying to make a quite profound point here that we are. We are looking at it wrong sometimes when we are trying to make business cases and when we're having quite idiotic arguments in in type should we do this or not do this, or we can't use this because it's only 15? Yeah, but your guys is working on 30 percent you know, yeah, you know.
Anders Arpteg:Do you actually try to measure the business value or added value by removing manual labor and simply receive the same kind of level?
Magnus Engström:I mean, the purpose isn't to kind of roll out something so we can scale back. The idea here is, of course, to just kind of make sure that competent people don't spend a lot of time doing things they don't want to do, or maybe, like we would, very they put a lot of work in something that's given a little so you could they become more productive than they can yes other stuff yeah exactly.
Magnus Engström:They can have more high value, yeah, yeah yeah, but for them to be able to do that, the machine only need to perform as well yeah, but you could try to measure that right, yeah, I mean, we measure stuff.
Anders Arpteg:Do you have a measure for that, or is it more on?
Hans Hjelm:Not for this case, for the manual labor, we don't measure it.
Magnus Engström:No no, I mean mostly because it's not within the scope of what we're trying to achieve right. We are trying to see if we can roll out personalization that will engage users to a large extent.
Henrik Göthberg:I have a rabbit hole we should go into now, but I've been in a lot of different conversations on how do we put KPIs and metrics around value.
Goran Cvetanovski:How should we?
Henrik Göthberg:measure value and that that kind of conversation comes back and forth. But so this is a quite interesting topic because, as you're saying, in advertising is one of those areas where you are very close to the actual objective function. So then it's a lot to learn from you guys, in my opinion, from other teams that has has not as much experience with this. So maybe that's a good topic. How do we measure value of this stuff?
Magnus Engström:But not now, and even I think that, from a data perspective, the even more controversial take here is that most publishers and other type of content actors as well spend a lot of time with metadata right Setting, manual metadata, setting up structures for when metadata should be set and by whom and where. Like Full-length news articles.
Anders Arpteg:Right right.
Magnus Engström:And then when we do the personalization, I don't think that data science even look. It doesn't look at the metadata because the vector distance is far superior. Interesting.
Anders Arpteg:So the content is more important than the metadata.
Magnus Engström:Yeah, we don't need metadata, you can do vector distance.
Henrik Göthberg:On the content, simply.
Anders Arpteg:So many interesting topics.
Henrik Göthberg:Can we talk about that? So the hardcore ideas around personalization is vector distance to other content Could you elaborate a little bit on that.
Magnus Engström:Yeah, I think and that is not like a born-new-secret I think that that is basically how this is done.
Henrik Göthberg:But to the outside to learn how it is done, then so, generically Hans.
Hans Hjelm:I mean. The basic idea is to look at the reading patterns of our readers and represent them in an embedding and reading patterns, as in how long you stay on an article or what, which articles or. Yeah, exactly All of them. So we have a fixed time period that we study and then we pick the articles that we think the user was most interested in, and that's like a combination of how much time did they spend reading the article, how long is the article, and we do some heuristics, basically based on that.
Henrik Göthberg:So you need heuristics to actually approximate which of all the articles were they actually reading. Within this timeframe you can get close, but you can't really see exactly where on the page they are.
Hans Hjelm:No exactly, but if you're on your mobile app we know exactly which article they're reading, but we don't know which one they were actually most interested in. But we can approximate that by looking at how long they spend reading a certain article.
Anders Arpteg:So then you get a set of vectors, I guess per article, in some sense and you can weight them in different ways or how do you go? About doing that.
Magnus Engström:Yeah, I mean, it's the same thing as if I were to use a RAG architecture, for example, if I just want to vectorize a text string and I search whatever set of documents I have that are also vectorized embedded right like the cosine similarity search. Basically just give me the nearest neighbors to whatever that text string.
Anders Arpteg:So that's actually very similar to how discovery weekly works in spotify. Yeah, I mean, that's what'm saying.
Magnus Engström:I don't think that this is like any secrets, right, but the thing is here that you don't need metadata.
Anders Arpteg:Yeah, I mean that's super cool. But the problem then, you know, is if you create like a single vector, represented user this is my centroid, so to speak, of what I like as an article and then you simply say give me the most similar kind of articles. It can be that the user have a sports interest, it can have a tech interest and it can be a set of very distinct kind of vectors. If you move in the middle, you could actually end up at the point, which is not interesting at all Exactly.
Henrik Göthberg:Yeah.
Hans Hjelm:So we're experimenting with different ways of calculating that centroid and we also have experiments where we don't use a centroid but just use the articles that the user was most interested in and then do a re-ranking. So let's say you have these 10 articles where your top articles, then we do a search in vector space and get the most similar and then we do a re-ranking of the results from those 10 articles and present that as the so this is the evolution where you kind of, okay, this is the step, maturity level one, we got here, okay, what next now?
Henrik Göthberg:And then you can see these challenges and now you have a hunch we need to go the ranking way.
Magnus Engström:Yeah, and then of course, I mean, if you have vectorized data, you can do this, for example, like on the content level also, like can we easily find more content to kind of link that are relevant to this content, for example, yeah, you can expand on it.
Henrik Göthberg:When you have found the engine principles, you can expand on that.
Magnus Engström:Yeah, in one use case that happens, but in reverse. We started with the last case I gave you and then we moved into personalization. Super cool, are you?
Anders Arpteg:going to try to do some clustering, perhaps per user, to say that these last case I gave you and then we kind of moved into personalization. Are you going to try to do some clustering, perhaps per user, to say that these are the top clusters for each user? It's similar to the my Daily Mix in.
Hans Hjelm:Spotify, if you remember that, we have a long backlog of things we want to try.
Anders Arpteg:I'm super cool to hear. I think we have so many interesting topics to cover, so perhaps we should move over to another topic and start to speak a bit about journalism. I think in general in some sense, and we all know that social media had a big impact on journalism, and then you know it was a big fight to try to keep traditional media compared to how people consumed news in social media, etc. And I guess it's been going on for at least 10 years or something, or what do you think? Yeah, something like that, yeah, and, and if you were to describe it, what, what's your? What do you think happened, you know, during the, the social media rise, so to speak, speak, and how the traditional media were forced to adapt in some way. What happened then?
Magnus Engström:Well happened in what regard?
Anders Arpteg:Let me phrase it in a bit more controversial way, as you like to hear. I mean, I would guess that a lot of journalists were, you know, they had to downsize a bit, they were forced, potentially, to write articles with a bit more clickbait type of titles, which a lot of social media people were doing, and they were receiving a lot more traffic, and then you were losing it to social media and that meant that it became much more aggressive in terms of competition and you were forced, potentially, to downgrade the quality of an article to become to get the clicks and views that you needed yeah would that be fair or?
Magnus Engström:yeah, I also think that if you look at most brands, or most brands that are doing well, they kind of moved away from that again, and I think that is it's quite obvious why I chose preface this, where I'm not a journalist, hans is not a journalist either. So whatever we say is, you're basically just well, it's an afterword, so yeah so.
Magnus Engström:But I think the thing is that, um, and if you look at what's his name, if you look at the Hooked model, for example, you have this model to drive people into engaging more with your products you have a trigger and that could be me walking down the road and I see this plume of smoke and fire trucks and so on of smoke and like fire trucks and so on and up on that. Up until that point I didn't have any need for, like, looking into my local news. Suddenly I have a large need of doing that. So in that case, I will do need to do an action and that action will be to interact with a product, hopefully a Bonnie News product. So I will open my news app or use my browser or whatever, and I will go into the local news brand to see is there any information about what I'm like? Because I have this information need.
Magnus Engström:It goes back to information theory. You have a need for information and then you have someone need for information and then you have, like, someone that could provide what you need, hopefully right. So you have to do some actions to get that information and if you get that information, you kind of your trust for that product will increase. And that means that you can engage the user in different ways sign up for newsletter or enable push notifications and so on. If you enable push notifications, then basically the news product can create a need when you're sitting home at your in your couch because you get a push notification and you didn't know that you had an information need until this very second when you read the push notification.
Magnus Engström:Right, and what I'm going here with is that I mean, if you break this in, like say, for example, that you post something on social media, we say that we have this incredible list of something and you won't believe your eyes. What's like the eighth, eighth thing in this list, right? So you click on that article and you go through the article and all the ads and whatnot, and you just kind of end up disappointed. You're like you started with I have this need of knowing what's the incredible thing that is in this list of things, right? Right, and when you're disappointed, instead of like increasing user engagement, you just kind of lower the trust for the product. So I think that that is basically, sooner or later, if you build viral products, those products will meet end of life, because you kind of used up, because yeah, yeah but are you saying now that there's some sort of pendulum going?
Henrik Göthberg:yes yes, so it started with okay, we need to follow the money and we need to follow the clicks. Yeah, so we need to adjust to something way more high pace yeah in the end. Now you see that way, there is certainly a niche for proper journalism.
Henrik Göthberg:And you. We need to go, we need to find that niche and we need to cater to that very concrete information need and not stumble on what is real journalistic information need versus all the other bullshit information needs that we shouldn't, we don't want to be part of yeah, so that's the pendulum to finding your identity back again yeah.
Magnus Engström:So if you talk to any editor at bonnie news or any other publisher, they would probably say that the greatest value lies in the quality, right and the like, the trust you can put into. And it turns out that is completely true, right when the smoke settles from like the viral sites and like the social media, like the clickbaiting. When the smoke settle, it turned out that that was right and I don't know if I understood it at the time, but but obviously people smarter than me realized that, yeah, this, this is just going to.
Henrik Göthberg:But I would imagine that, okay, the chaos is there, the smoke settles, but for me. I have no data, I don't know this, but my sense in myself is that the map was redrawn, so what we had as newspaper behavior before digital, and that we you know I was a paper boy for Dagens Nyheter when I was a kid.
Magnus Engström:You know I worked for a company.
Henrik Göthberg:You know that world is gone, so we are now in a different landscape that we now need to find identity? Yes, but the size of that market, or how that market looks like now, after this whole task is settled, yeah, I, it is probably different, right, and we are still trying to figure that out, or do we do? Is Bondi and you is trying starting to see the contours of this, because I think the whole media space has been in shambles to find their feet here?
Goran Cvetanovski:Yeah, well.
Magnus Engström:I think that finding the contours is a good way of framing it. I think that, in the end, we just kind of need to make sure that the product is actually of value to the user, right? That is the only kind of way to invest in product development making sure that whatever the need is for this user, we can actually provide, rather than using that need for, like, for example, looking at something completely different. If you look at, for example, online cas casinos, for example, like, the magic trick that an online casino needs to do is because everybody I guess I don't use online casinos, but, but I guess that the main reason that people would enter an online casino is because they want to make quick money, right, or they need, maybe they want to make a lot of money, so that is kind of what they feel that they would need, or they think they have fun.
Henrik Göthberg:Yeah, they think it's fun to gamble.
Magnus Engström:Yeah, yeah, I mean you start off with maybe I should join this because maybe I can get some money out of this, but, of course, if the casino, online casino will pay every user, whatever the user feels is a good amount to be satisfied, that would be a really bad business idea. So they need to make this magic trick, this illusion, right. So I guess you came here for the money, but you're staying because you never know what is going to happen if you push this button one more time, right, and if you leave now, you will never know what would happen if you push that button one more time, right. So that is what is driving engagement. It's not like the promise of you will just hold on, you might be rich any time now, right? That is not what is driving engagement. Engagement is driven by saying that you, if you close this browser window now, you never know what would happen if you have flicked that button one more time and that's the same in social media right you never know what will happen.
Magnus Engström:Exactly my point. Yeah, right sorry no, no, no, no, it no.
Henrik Göthberg:I mean that in a very good way, because so that consumption of media consumption is something else than to have an information need. It's the same as the gambler who is hooked on looking at pushing the button again has forgotten about the money. We're not looking at news anymore.
Magnus Engström:We are hooked at you know we're not looking at news anymore we are hooked as, yeah, at we are drugs. You know what comes. Next question yeah, yeah, yeah, and I think that the journalistic products of quality need to. Not trying to do that bait and switch right, so you?
Anders Arpteg:we really come there, then I mean, who is really influencing the society the most these days? I mean we have influencers that you know perhaps people trust more. And if we take the American election and Taylor Swift and how she could actually move the election by herself in some sense, I think before this, you know, the traditional media had a really, really big stake. Yeah, but not so much still today, or?
Magnus Engström:what do you think? No, that has changed completely, I would say. But it mostly is the means of distribution, right? So you're able to distribute information in other ways. If you go back I don't know 30 years, if you go back, I don't know 30 years, in that case, like the local newspaper at the editorial desk in the morning, you could have discussions about what should we decide? Is the opinion right?
Anders Arpteg:We could choose that.
Magnus Engström:Yeah, not necessarily talking for a Bonnie News brand, just in general. I think that when you have like monopoly on local news and on distribution yeah, and on distribution then you can also kind of drive opinion right. I think those days are over.
Anders Arpteg:But do you think it can come back where you know media institutes really have the similar or close to similar kind of distribution as well, or is it over?
Magnus Engström:Do we want it to come back? Good question. I think it's proven that journalistic products can thrive at this landscape as well. I mean, there are competitors of different kinds than before, right, so you need to be like on top of your game. But I think that's like to some extent for all products, not talking specifically about Boninus. I think that's always a good thing, because you can't resort to, for example, clickbaiting, because then you will kind of lose whatever edge you have right because then you will kind of lose whatever edge you have right, but can we go down in a little bit of a fascinating rabbit hole?
Magnus Engström:I thought this was a rabbit hole. Now we're doing the rabbit hole on the rabbit hole.
Henrik Göthberg:Sorry, we are completely off track. What are we talking about? I think this conversation about what is the new media landscape and how to find our feet in those contours are super interesting. And who wasn't fascinated of the American election? Who wasn't fascinated of the role a guy like Joe Rogan plays? Is he news? Is he, you know what? Is he really right? Is he social media? Of course he is. He's a YouTuber, but he's bigger than that. And today and this whole conversation I mean I don't know how much you have been following, but I've been trying to untangle who's talking shit about who, who's telling truth about who, from the democratic to the republics, you know and there are so different narratives about who, from the democratics to the republics, and there are so different narratives being explained from the different angles. So it's really polarizing, really confusing. You know, I don't think it's that bad in Sweden, but I think there are. We are seeing patterns that this complexity of you know who is pushing what agenda has never been more difficult.
Anders Arpteg:And I think it's a big shame that media outlets have lost a bit of their power, because I think there is a big need for finding sources of truth and institutions you can trust, because it's so much deep fakes out there and certainly influencer and other people that are not schooled in this and right for other reasons. But I would argue.
Henrik Göthberg:There's a big difference here. Um, I I don't sense that the swedish traditional media has been bashed. You know, in america it becomes almost like a religious war, you know where. You know where some of the big newspapers are fundamentally accused of wokeism and going a very, very specific agenda from one side of the political spectrum, and of course they don't agree with that, of course. But it's a very, very polarized view on you know, how do we understand traditional media? I don't think it's like that in Sweden. Do you think so? I don't think so. What is your take?
Hans Hjelm:I don't think it's as bad, but I think you see some indications of that here too. I mean Dagens Nyheter. They are not very popular in certain Age categories. Dagens Nyheter, they are not very popular in certain categories or right-leaning groups, for example.
Henrik Göthberg:So you can tell that the political color of the newspaper, so to speak, we don't believe what is. You can see tendencies also in Sweden.
Hans Hjelm:Yeah, I would say so.
Anders Arpteg:And we can speak about this forever, but I think we need to get back to AI in some way.
Magnus Engström:Yeah, but I mean it's on a tangent anyway, I would say, I mean you start with the social media. But I guess where we're going with this is that we are kind of adding more complexity to that question because yes, you went into the social media rabbit hole.
Henrik Göthberg:when should you have just passed that and say what?
Magnus Engström:will.
Henrik Göthberg:AI do to all of this so we have a very big context background story.
Anders Arpteg:now moving into the real question, yeah, and perhaps just connecting to the story, if we speak a bit about the LLM dilemma, the AI impact that we had, perhaps if we take the American election I think a lot of people thought it would be a big problem with deepfakes et cetera what do you think AI has had as impacts if we take the US election, for example, as an example, Was it more or less than you expected? For example, as an example, Was it more or less?
Hans Hjelm:than you expected? I don't. I didn't see big impact coming from fake AI.
Henrik Göthberg:This is our conclusion too.
Hans Hjelm:Yeah, and if you have big platforms like social media platforms, then it's enough. If you have a strong voice like Joe Rogan, you don't need fake content. You can just influence people's mind.
Anders Arpteg:I think it's an interesting example is the election in Argentina. It was one year ago approximately, but they had a lot of generative AI in that election, but it wasn't for deep fake purposes. It's more like it's a way to just communicate, right. So you create like a teddy bear around your candidate and you want to make him look like a safe and trustworthy person, and you have a vampire around your competitor, but everyone can see it's fake. It's not trying to fake, but it's fake.
Magnus Engström:It's a satire. It's a magical satire.
Anders Arpteg:And I think a lot of people were more afraid than they should or, looking back, it didn't have the negative impact that we feared a bit.
Magnus Engström:And to tie that back into the previous discussion, I think that the thing with deep fakes are, I guess, that most usually they are products distributed by sources that don't have a lot of staying power. Sooner or later they will turn out to be deepfakes and then they just kind of disappear from out of focus. And I think that that kind of goes back to what a news product needs to be. News products need to be things that we ensure have staying power, so you need to be able to trust, and I guess the same thing goes for things like Joe Rogan's to some extent, I guess always needs to be on brand because he needs to build up some kind of staying power.
Henrik Göthberg:So be on brand, staying authentic to his own identity and who the fuck he is.
Magnus Engström:So I think that I would argue that maybe the big impact of deep fakes are the exact opposite. It's not that we are exposed for a lot of things that are untrue. I think it's much easier to argue for the things that are true or not, like, for example, if anyone does any bad interviews, or, for example, they will always argue it's editing Right.
Henrik Göthberg:But maybe, if I try to follow on, someone argued that we will simply adapt that the baseline is that everything is fake and now you need to qualify truth that the baseline is that everything is fake and now you need to qualify truth. So as long as we adapt as humans to that, we can't take stuff on face value. You cannot take anything on face value in terms of what you see and read, in terms of how you can manipulate and and maybe it's even better now that we truly understand that maybe we've been faked with pictures for years that we didn't understand. So now we're getting to a point where this is the cat is out of the bag and now we now need to adjust and balance the way we validate our you know, the way we go about. What is information, what is entertainment?
Magnus Engström:do you see what I mean?
Henrik Göthberg:like so so there's a calibration going on here and if and if we get that right, I would argue that that opens up new data and AI-driven features in order to prove credibility and trust.
Magnus Engström:Yeah, and also I think that it doesn't have to be content generation per se. I think that one of the strong aspects of AI and machine learning, looking at media landscape, is that you are able to aggregate everything right. Like a person with a with enough exposure in media, you will be able to kind of dig out and edit and, like present that person in in a light. You don't have to fake anything, right. What's at your power is the possibility to just go through everything, just having an AI go through and find every time this person says this word or does this gesture or whatnot, and then you can just edit it together right, so no deep fake needed.
Henrik Göthberg:Yeah, but I've really seen there is a new type of media product, so no deep fake needed, but I really see them. There is a new type of media product or news product that I think has evolved more in 2024. That is basically trying to build a product. How they are on different articles, rating them, triangulating them with different facts and even clicking. Here is the positive side of the argument. They are trying to do unbiased news and they are doing unbiased news not by presenting one news, but maybe they have one news story, but they are trying to build the full landscape on the topic and I think that is high integrity. I think this is product development of the real quality newspapers to show this, you know. Do you know what I'm talking about? Have you seen those types of services? I forget the example names. You know what I'm talking about. I mean, like there are these news sites that are literally about truthfulness or balancing different aspects of the same news.
Anders Arpteg:I guess we can. If we speak about the infamous filter bubble and the people get stuck in a certain get type of news all the time, then it's just inflating all the time your beliefs in that and you don't see anything outside the bubble. But perhaps using technology, it can actually be in help to get out of the bubble if you start to also have news that is a bit more objective in some sense. Sure, Right, so perhaps going back a bit. Okay, so AI can, of course, be used by deepfake. I don't think that's something that Bonnier does. I hope not, but can you perhaps elaborate a bit more? What does Bonnier use if we take generative AI for Can you give some examples.
Anders Arpteg:Perhaps Does the journalist use it on a regular basis. Is there anything you can talk about?
Hans Hjelm:So we use it in different parts of the organization. The one where our team is most involved is we have like a news aggregator. So, like we were saying, we have almost 200 bands or something like that, but we also have. You can buy a subscription where you get access to something like 50 brands. I don't remember A package, a package, yeah, exactly. Package, a package, yeah, exactly. And then the question is uh, now I have access to all this content, but, uh, I don't have time to read 50 magazines, so how should I get the value from, from all these 50 brands? So then there we have a site where all the content is displayed and also personalized for you.
Anders Arpteg:So if you like to read about a certain topic or you write, and that's for the journalists themselves, not for the users.
Hans Hjelm:It is for the users. Yeah, so if you buy this package subscription, you can access content through this site and is it like a chat content through this site?
Henrik Göthberg:Is it like a chat? Are you starting now to make a user interface so they can find and they can?
Hans Hjelm:prompt where they want to go deeper into this. Yeah, so our team is working on finding new stories that are written by several different brands and combining them to a single story and giving like a short overview of the stories yeah exactly, and also so there we're using some level of generative AI to write the summary, but then there's work also going towards what you were saying with chatting with the content, so to speak, but maybe Magnus is more involved with that work.
Magnus Engström:Yeah, so, for example, we have added what we call trigger questions, basically finding relevant content to the article you're reading and figuring out what is a good approach to kind of get the user to notice this article. So, for example, we generate questions that the user might have from the article in front of front of the user and then we say, right, so these are questions that you could dive into. It's the perplexity way. Yeah, exactly the perplexity way, right oh, which is brilliant.
Henrik Göthberg:Yeah, how many times have we not clicked on those questions?
Magnus Engström:yeah, exactly because I mean, um, talking about rabbit holes. So, um, I think that the best way to think about AI when talking about approaching end users or doing product development, you could basically say that All AI is UX.
Anders Arpteg:Okay, that's a strong statement, that's okay, let's go with this.
Henrik Göthberg:I love this.
Magnus Engström:I love this entry point. Let's say that I brought with me to this podcast. I had a black box, right, no interfaces, and I would just put it on the table. I would say that this is the most advanced AI ever developed, and we would sit and look at this black box for a while and then you will ask well, how do we?
Henrik Göthberg:What do we do with it?
Magnus Engström:Yeah, and I say nothing. There's no interfaces. There's no interfaces? Right, but it's the world's most advanced AI. We don't have any interfaces Actually.
Henrik Göthberg:I'm glad you said that you can't open it, but it's such a super intelligence in this. Actually, I'm glad you said that you can't open it, but it's such a super intelligence in this, but don't open it.
Magnus Engström:Yeah, right, so clearly AI is a UX question right Somewhere.
Henrik Göthberg:How do we interact with AI? I see your point. You're reducing the point to a very tricky topic.
Magnus Engström:Yeah.
Henrik Göthberg:Lovable is doing the same, by the way.
Magnus Engström:Yeah. So if you look at this from a news product perspective, right, we need to understand. If we're going to use AI to create better news products, we need to understand what is the interaction that is actually increasing the value of the news product. Here, like I, fully right. So, because we don't do that? If you don't do the ux part, well, I think that it's. If we do it in our product development.
Magnus Engström:It's about accessibility right, we make our products more accessible and that could be, for example, maybe you need the article in another language, or you need it to be shorter, or you need it to be read to you by some, like text to speech or whatnot, right. So let's say that if we do that in our product and we understand that this is the user need, then it's accessibility, right. But if we don't do that, then the end user will. Because I don't want, I can't access, I don't feel that this information is easy for me to access. So I will just kind of ask ChatGPT to go to this news site and read the article for me and just summarize it in points, for example In that case, we don't have an end user as a consumer anymore, it's the tool that is the consumer and the tool aggregates and then the UX of that tool is actually the user experience. So for us, ai needs to be UX, because otherwise we won't have end users, we will have tools, and those tools can do whatever. I mean maybe.
Magnus Engström:It could be a good value as well to have tools as a user Well the problem here is, of course, that I mean if you use ChatGPT well, or, for example, let's take DeepSeek, for example, let's say that I use DeepSeek and we write something about China, right, it's possible that the tool will not convey the original content in the way that we intended, right, but you can increase your distribution, perhaps thanks to the tool. Yeah, but in the end it's like a losing equation, because if you build a subscription product and you build for end use I mean in the end otherwise you will provide data sources, not a news product.
Henrik Göthberg:But I think I follow what you're trying to say here. So when you're doing a very, in a sense, provocative statement, reducing the AI I didn't even know that it was provocative. No but you're reducing all the brilliance of a large language model, and all this down to the UX. But I follow that because let me try another way to say the same words, but taking it out of this context goes back to how to make AI agents work effectively, and all this.
Anders Arpteg:I would phrase it differently. Can I try first, because you don't know where I'm going.
Henrik Göthberg:I can guess, but okay, no, because it's very simple that in the end, when we do a tool like AI, we are doing it to improve the flow of whatever work we're trying to do. In this context, we're reading news or we're interested in news. If I flip it to any other context, we have a workflow, we have someone working, someone needing to get a job down, and we want to do it more effectively. So here we now talk about oh, we have an AI here, we need to automate stuff and we can do that very intelligent, we can do that smart and we can put in an agent and do it automatically. And we use the word automation.
Henrik Göthberg:But what I think is missed when we are only looking at the techniques of the technology is that any time you have automation, inherently you're augmenting someone's workflow. And every time you're augmenting someone's workflow, and every time you're augmenting someone's workflow, ultimately, if this is going to work or not, it's about the interface between the human and whatever the AI is doing in that workflow. And so what you're saying now that if we want to have value from our AI, we need to secure that it augments the socio-technical workflow now, in fancy words, that we have people. That is human flesh and blood. It processes. You know the process how we consume news. It needs to fit into our workflow. When I use my, when I use my spotify. It needs to fit into my flow of how I want to use music and you can take that generic thought into every single.
Anders Arpteg:Let me disagree a bit here, because I disagree with both of you.
Henrik Göthberg:If you disagree with this, we need to go to the bottom of it, because you can't disagree with fundamentals.
Anders Arpteg:And you are fundamental, no, no. Okay, let me try to explain the techniques augments workflows is fundamentals.
Anders Arpteg:Yes, so AI can be about many things and it can have many purposes to work with AI. If you go back to the original definition of AI, it was the science and engineering of building intelligent machines. If you take the scientific approach of this, it has nothing to do with the value for the user. It has to do with the knowledge of understanding how things work, so that can be a very useful thing to do as well. So if you want to work with AI to produce more knowledge, then you're not considering the end user value at all. If you then work with engineering, you do work with a value, and I think what you're referring to here is when you say UX, you mean user value in some sense.
Anders Arpteg:Yes, I do too, so I would call that not UX. I would call it engineering, because it's more than UX I think is required to really provide value. Yeah, but sorry, continue. No, no, it's fine.
Magnus Engström:But let's stay at this topic for a while, okay, this is a good topic, all right. So let's say that you're working with building intelligent machines. Yes, you're working with collecting and understanding and, in different ways, handling information, because that is like the pure definition of AI. Yes, right, are anyone going to use this information?
Anders Arpteg:but there is still a good reason and to have the purpose of building knowledge, even if no one is going to access the knowledge. Yes, it is right, that's the reason for doing research.
Magnus Engström:You know, scientific research yeah, but scientific research still. You need, you like, need to publish in academics papers and so on, but not build products. Right, but I mean to be able to publish in, for example, academic papers. You need to show the text, the purpose, then, is knowledge, not products. Yeah, I will still argue that it's UX right.
Anders Arpteg:Okay, then you have a very broad definition of UX.
Henrik Göthberg:Yeah, I follow that the way in engineering, and all that. Of course, this is an engineering.
Anders Arpteg:But then UX becomes like a concept that encompasses everything.
Magnus Engström:Yes, yes.
Anders Arpteg:Then the topic becomes useless.
Magnus Engström:No no.
Henrik Göthberg:We're going into Mary Parker Follett, 1925. She states, before the times of data computation, that we cannot really separate the human and the machine in the work they're doing. She's examining and working and writing on this in the 1920s. I don't know how she says it, but the actions of the machine and the actions of the human cannot be, in the end, separated, because it's how they interact that creates the value and it's in this context. You know, if you reduce that down to the AI is part of a socio-technical system, which is then part of what we do, what we're supposed to be doing, what value we want to get out of Still thinking engineering. There is a big get out of, in this sense, thinking engineering there is a big purpose for science.
Magnus Engström:Yeah, but also it needs to be thought about how we're going to build upon that science, not when you're doing the science All right. The. Thing is.
Anders Arpteg:I usually argue for engineering so much more than science? Yes, but now.
Magnus Engström:I feel I need to for once to find science, which I think has its place as well. Yeah, and I mean we started in product development, so we're kind of far away from products.
Henrik Göthberg:So the argument here is fully an engineering topic, and from engineering all the way out to actually not only engineer the system, but to engineer the interaction between the AI and the human. That's what we ultimately talk about, because we need to engineer a human, physical, human, artificial workflow. We need a socio-technical workflow. That's what we are talking about, and then we are going to the exact interface where they are supposed to interact. How?
Anders Arpteg:that interaction is supposed to happen. I think engineering has much more than to do with the interface as well. So I think you know you can broaden that discussion and I would still disagree with the thing, that UX is everything.
Magnus Engström:Yeah, I'm not saying it's everything, but I'm saying that, okay, let's say like this applied machine learning is UX, right?
Anders Arpteg:I would say it's more closer to engineering than pure research, but it's still.
Magnus Engström:This is one of the most interesting discussions I've had.
Anders Arpteg:I think we could take this on in the after work.
Henrik Göthberg:I agree with you both, because it's great engineering to get to a good experience.
Magnus Engström:But to kind of return to where we started off and kind of get on track.
Magnus Engström:I guess one way to talk about this at some higher or lower level, depending on your program or not. I mean, let's remove AI and UX from discussion. Say that you have a person in a wheelchair and then you have this bus stop and the only way to get to that bus stop is some ground that is not paved, for example, and there are two ways to do that. You could either prepare the ground and pave it so it's easier for the person in the wheelchair to get to the bus station, or the person with the wheelchair can get I don't know larger tires for the wheelchair, or train the wheelchair can get, I don't know larger tires for all train.
Magnus Engström:Yeah wheelchair right, so, and some people might do that right, a lot of other people will start using another bus stop, for example. Right, and that. There you have, like the difference, like either you put the responsibility for the you on the user to apply whatever tools they feel needed to use the product in this case the bus stop or you prepare the way so you can take as many users as possible into your product. And there you have. Should the user apply AI to access the product or should the product use AI to make the product, or should the product use AI to make the product more accessible for the user?
Anders Arpteg:Yes, we're still in engineering land.
Magnus Engström:I can hear, I kind of turned it, because I felt that we were moving into the seventh rabbit hole. We had to go down to where AI actually is.
Anders Arpteg:Hans, I would love to hear it before we move to the next topic.
Magnus Engström:You're a scientist as well.
Anders Arpteg:Do you have a preferred definition of science and engineering?
Hans Hjelm:I can't say that I do at the top of my head. Can I try mine and see what you think about it?
Anders Arpteg:Yeah, yeah. So I think science and engineering very much overlaps. It's hard to do science without having some engineering skills and it's hard to do science without having some scientific skills. But I think the difference between the two is the purpose for the work. So the purpose for science it's very clear is to build some kind of new knowledge that we haven't seen before. The purpose for engineering is to build a product that provides value for some set of users. That is the big difference between the two.
Anders Arpteg:But if you want to build new knowledge for how you can add reasoning to an LLM from a scientific point of view, you probably need to build some prototype to try it out. So you need some kind of engineering skills not the same level as putting it in production, but some kind of engineering to do that kind of scientific work, to do the experimentation of it. If you then come from an engineering point of view and say I want to build a product like Chat-TPT or the extend or recommend articles in Bonnier News and you're thinking purely from an engineering point of view, you need to have the knowledge that is required to do so. But the purpose will always be how can I build this product to maximize the value, and that's the difference. In the case it's maximizing knowledge. Other case it's maximizing value. Other other case it's maximizing value yeah, well, I agree but, also
Magnus Engström:I feel, that I mean all type of science needs to be like building blocks that you will be able to stand on right. Yes, because if you don't pass knowledge along, this kind of, the point of societies is to kind of build upon the previous generation, right standing on the shoulders of giants. Yeah, yeah and then you can't put your science into like a black box and say, no, it's no way to interact with the scientific findings because we haven't an interface for that. So, for example, peer review is an interface.
Anders Arpteg:People doing research that also have to give lectures that's also an interface but my point is really that the main difference between science and engineering is the purpose.
Henrik Göthberg:Yeah, yeah, I agree, yes, yeah and I I like this argument because of course we cannot argue that everything is interface, but we are trying to make a point that this is a very, very critical part of the whole equation and if you get that wrong, everything else is lost, especially in product yeah, and I mean, if you move all the way back to like business value, yeah, in that case if you go into your reality of working with ai in borneo news. In reality, is it usable, is it adoptable?
Magnus Engström:yeah, yeah yeah, but I don't feel that that's maybe a point that we could disagree a lot on. So I had more fun talking about the scientific side.
Henrik Göthberg:No, we love it, and we had a little bit of rosy cheeks. It's always nice.
Goran Cvetanovski:It's time for AI News brought to you by AI AW Podcast.
Anders Arpteg:Awesome. So we have this kind of small break in the middle of the podcast to just take the discussion a bit differently and think about some of the top news that we have heard in recent weeks and try to keep it very brief because we are a bit over time. We normally we should have started the quarter fast, but we got stuck in a rabbit hole speaking about science and engineering and UX and whatnot. But let's perhaps do a quick round of news as well. Henrik, do you have something you want to start with?
Henrik Göthberg:Okay, very quickly, the whole DeepSeek debacle that came out. We were actually talking about deep seek in R1 on the news show last Thursday, so for us it's old news, but the whole everything went viral, happened afterwards. I was only going to make a remark that we've been talking about this podcast, about how do we invest and how can we compete with the frontier models, and I think the angle here that I picked up on the most is that there is clearly a different argument on what is the moat, or we can never compete. Well, maybe you cannot compete exactly on the same game, but there are more games to be played. So this was one angle. The other angle was to pick up that there was.
Henrik Göthberg:In the end, we were comparing apples and pears. There was very, very nuanced. You know what is the cost, how much smarter is it, but my main angle was that it's great, open source, open innovation. If you do good stuff, don't be scared, you can compete. That was my take in the end. I don't want to go deeper than that, but it was funny because we talked about it and then the whole thing blew up.
Anders Arpteg:Do you have any thoughts about DeepSeek Hans?
Hans Hjelm:I mean, I agree it's an impressive feat of engineering. I would say what I've read about it so far it didn't seem like they brought something entirely new into the world. They've done a very good job of applying already existing technologies. Agreed, but that in itself is impressive.
Anders Arpteg:Magnus, do you have any thoughts?
Magnus Engström:I think I just kind of agree with what you're saying. I feel that the obvious most interesting thing with DeepSeek is that it's open source. I feel that that is basically. It's all the way back to what you talked about personalization. It doesn't need to outperform, just needs to be as good, right, I think that's true here as well. I mean open source model that is as good as the OpenAI 01 model. I mean that's huge, it doesn't need to be better.
Henrik Göthberg:Maybe it's only better at one thing, but what it does, it can do.
Anders Arpteg:And perhaps we should just recap a bit, even though we spoke about it last week. It happened a couple of things this week which was rather astonishing, and I think actually we should blame media in this case. I think so too, but let's get back to it. American media, for one. You know, we had a huge drop for Nvidia. They lost like 17% of their stock market cap. More, more, what. I'm losing money Five minutes. I actually sold most of my Nvidia stuff, so I'm laughing my ass off.
Henrik Göthberg:I think I'm going to buy now. Now is the time to buy. Yes, I do think I will buy Nvidia stocks now.
Anders Arpteg:Anyway, 70% is like $600 billion.
Magnus Engström:It's like the Swedish GDP size kind of thing it's like the largest value loss ever in history, and the funny thing is that I think that Nvidia have made seven of the top ten highest value losses on the stock market. Nvidia is seven of them.
Hans Hjelm:Oh, really. Oh, that's interesting.
Anders Arpteg:I think, also that its media has been…. It's never a loss until it's sold.
Henrik Göthberg:Until it's sold.
Magnus Engström:Do you own NVIDIA stock?
Henrik Göthberg:Do you own?
Anders Arpteg:Pharmacia stock. I think it's been very misleading in media. You know how it's been reported and for one I think you know they spoke about. You know how much it costs and it says like five or 3% of the total cost. But for one they actually are using a pre-trained model, the V3 version, and then they doing the fine tuning above it and then they're doing the fine-tuning above it. And to compare the cost of fine-tuning to doing pre-training is simply wrong.
Henrik Göthberg:Yeah, there's a lot of apples and pears in here.
Anders Arpteg:And then, secondly, if you compare it to GPT-4.0 or even the O1 model, that can handle both text and images. This one can only handle text, handle both text and images.
Henrik Göthberg:This one can only handle text and images, is actually much more complicated to handle and requires a lot more parameters and training data, especially in size. So we can just argue it's blown out of proportion. It's interesting ideas, but then it's, isn't it just blown out of proportion?
Anders Arpteg:And then there are like a third thing saying basically they actually say that they had high quality data to do the fine tuning for r1 on where did they get it? Well, they did distillation from open ai. They even have for the base one, the deep seek v3, the normal lm that they had. If you asked it, you know what, what is the name of your model? It said I'm gpt. I mean they even did a poor job of copying it. I mean they should at least try to filter that out in some way. And obviously they use distillation. And if you use that, then it's much simpler because you get more high quality data which they brag about and they say, oh, we have such high quality data. Well, yeah, because you cheated and you used another model to obtain it. Right, of course, that will be cheaper and better, and I think it's a little bit polarized.
Goran Cvetanovski:So if OpenAI does this, then it's not cheating, then it's innovation. But if the Chinese are doing it, then it's cheating, right? You remember when OpenAI said, like how did you train your data? Well, you know it's over-trained, it's good and stuff like that. They were stealing from your post and everybody else, right? And now Sam Altman is complaining that.
Anders Arpteg:Chinese are stealing his data which is not his data. It's a big difference.
Magnus Engström:I read a comment on Hacker News yesterday that OpenAI is losing its job to AI.
Goran Cvetanovski:That's funny. No, but I think it's a big difference here.
Anders Arpteg:If you have human written content, it its job. To AI, no, but I think it's a big difference here. If you have human written content, it's one thing, but they are having AI written content here and they're stealing from the other model. If open AI wasn't available, they wouldn't be able to build this model.
Goran Cvetanovski:If Google was not available, open AI would not exist. What's?
Anders Arpteg:the difference? The difference is that human data are available, and that is what OpenAI have used. So it's certainly much easier Imagine how much crappier it is in human data, and OpenAI had to deal with that. This DeepSeek does not.
Hans Hjelm:So it's a big difference. I think you could almost liken it to industrial espionage, right? They're stealing the Open AI technology and say look, we made it much cheaper than.
Henrik Göthberg:No, it is not espionage, because we have been sitting here saying you know what? The way we should compete in Sweden is to distill out baby models from the large frontier models, to try to do everything from scratch. But it should be done in a legal way.
Anders Arpteg:This was illegal. They have even in their terms of service saying you're not allowed to build competing products from this. They have literally broken the rules for doing that.
Anders Arpteg:So it's illegal, but China doesn't care about that. But, that said, they did a number of, I think, innovative things for R1, especially being able to output Python code along with the text code when they do this synthetic generation for the reasoning part and then use that to verify if it's an accurate solution. So there were some really cool innovations as well. But then there is the whole aspect of national security, of geopolitics, and when I tried it myself and wrote you know what's happened in Tiananmen Square, you know, and then first I actually started to output rather factual, correct statements, but after two seconds it's just bloop, removed the content and said sorry, I can't speak about this or something. And this is actually rather you know, you can think about.
Anders Arpteg:Why did they open source it? You know why and this is actually rather you know, you can think about. Why did they open source it? Why did they give it away for free in that way? Well, it is a nice society. Well, it could also be some kind of other hidden agenda behind this that they want to spread some more of their kind of propaganda. What do you think about that? Is that potentially what they're trying to do here?
Hans Hjelm:Sounds plausible to me at least. I mean, I can't read their mind.
Magnus Engström:Obviously they don't own Nvidia stock at least.
Henrik Göthberg:No. And then, and as Sona said, we said it many times before open source has never been about charity. It's always been about the business model in some sense, and this is a business model as well. What that exact business model is, I don't know, but open source is this viable business model period.
Hans Hjelm:But you see it also among the US large language models, right With the XAI versus OpenAI. How much should you correct for bias in the models, or should you never correct for bias?
Henrik Göthberg:or should you steer the model in a certain direction? So there is a spectrum here. It's not black and white, it's a spectrum Interesting, okay.
Hans Hjelm:Should we stop there or should we do some more?
Henrik Göthberg:No, that was this sort of-. Did you have some other news?
Magnus Engström:I don't know if you heard about this deep seek.
Henrik Göthberg:I mean, we haven't commented on Operator.
Anders Arpteg:Yeah, should we go there and we haven't commented on O3 Rumors. No need. O3's been open for a long time, I think.
Henrik Göthberg:But what do we need to say about Operator? Are they?
Anders Arpteg:better than Anthropic or not. In some sense I would say so, but it's more limited to just the browser.
Henrik Göthberg:I think you know, claude, computer use can be wider right, but it's a nice.
Anders Arpteg:I think the engineering not the science, but the engineering around the operator is better. They have a web browser in the cloud and you don't have to start your own container and blah, blah, blah. It's easier than using just the API that Anthropic provides.
Magnus Engström:Right, I love how this conversation kind of ended up, that we need to make clear which part of the divide between science and engineering we are talking. Now I'm talking on the science side. Now I'm talking at the engineering side.
Henrik Göthberg:But I saw one.
Anders Arpteg:Did you ask for?
Henrik Göthberg:anthropologist.
Anders Arpteg:Matt, let's call it that.
Henrik Göthberg:But one of the key things that someone commented on that I picked up on that makes sense, that resonates with me was that I mean, like were you seeing the very first levels of these computer use or web browser use? And the core example was like like today, when we are now sitting there and literally watching the computer doing our manual steps, like we would do it in the same amount of time as we were doing it, it doesn't really add any time value. I literally need to set the whole job up and then go to the gym and then go back and use it. Well huh, is that really useful yet? But then we can anticipate this will be solved. This will be solved.
Magnus Engström:This is probably the first time where we actually are hitting things that are in our notes.
Goran Cvetanovski:Yeah, I think that we actually wrote something about this.
Anders Arpteg:Perhaps we should explain what the operator is.
Magnus Engström:Not everyone probably have heard about this news, so if you could, if you have some notes, Well, operator, the short version there is, of course, that you can make OpenAI act upon your own behalf, provider of the web service or the website also kind of fulfilled. Then I can ask OpenAI operator to do it for me, right, so, but it's quite, I wouldn't say it's narrow, but I think that a lot of those things maybe will kind of break apart if you move outside of the verified websites, for example.
Magnus Engström:I don't know, we'll see. That's what's not what was in our notes. I think that what we wrote in our notes was that a good way to think about AI as an assistant is to think about, let's say, that the tools I use in work, for example, the number of tools, is a constant, right. So if I add something, something else has to go, yeah, right. So we can't add like now we have, like these, 28 tools and now you also have the AI tools, now we have 29 tools, right. That is not going to increase productivity.
Magnus Engström:What we can do is that we can build smart, like the tools you already use can get smarter or replaced, right. Or you can end up with fewer tools, I suppose. But as long as AI, as long as operator, is its own thing, right. Then you just add more and more tools, like you have more and more different things that you need to switch between. It's basically using the chat GPT app. In that case, if I'm going to use that for translating and things like that, I will take my text, I will copy it, I will move over to ChatGPT, I will interact with ChatGPT, I will copy my text and I will move back to whatever application I was in before. So we are kind of adding more and more abstraction layers.
Magnus Engström:So I feel that the best way to think about AI's assistance is that we need to kind of think about number of tools as, if not a constant, at least a number that should decrease, not increase.
Magnus Engström:An operator can help to decrease it, hopefully, but it needs to do it perfectly. I mean, that's basically the same thing as the Google versus Tesla in self-driving, like when Tesla says that we're going to work ourselves towards 100% right, and sooner or later we will end up in a situation where you don't have to use the steering wheel, and Google says that we're going to build cars that don't have steering wheels, right, so in that case, to say that we remove one tool and we add the self-driving capability, while in Tesla you have the steering wheel and the self-driving, and I feel that I mean, as long as you have to touch the steering wheel from time to time, you're not like you still have to do your duties as a car driver, have to do your duties as a car driver. I mean you won't be able to get from point A to point B without the driver's license, for example. You could try. You could try, yeah.
Anders Arpteg:I mean cool and I think it's also. The engineering behind it is so much better in OpenAI and it's basically integrated into the normal ChatGPT product as well. Not just the API they have there, but also it seems to be rather immature still. It easily gets stuck in loops and doesn't get out of it. But then you have the nice engineering, the UX feature of actually you can take control at a certain point right, we're at the engineering side.
Magnus Engström:Now we have to be clear Not science.
Anders Arpteg:They did have some scientific stuff as well, I think.
Henrik Göthberg:What was the hard problems they needed to solve? What was the engineering or research feat in Operator that they wanted to have?
Hans Hjelm:I don't know, I mean it basically has to take control of the mouse and keyboard right. I mean it basically has to take control of the mouse and keyboard right, so going from text input to actions on the screen.
Henrik Göthberg:It's actuator work right. That was what they needed to solve. Can you elaborate?
Anders Arpteg:Anthropic has done it before. I think they copied a lot of the stuff that they have been working with. But package is nicer but also it's kind of hard because it's such much higher frequency of the actions. You need to see what happened on the screen very quickly and then interpret what happened on the screen and actually take actions. And it's still a bit slow, I would say. And then it's not so much science but it's more engineering aspect of how to make it that fast, since you can't wait for a number of seconds because everything becomes super slow then. So you have to be really innovative to make it work, in this case for actually interpreting the screen, good luck. I shouldn't say that Actually, deepseek has launched their image generator as well recently, so they probably will get into this space as well soon their image generator as well recently, so they probably will get into this space as well soon.
Anders Arpteg:But it is so much harder to interpret an image than it is to interpret text. And do that very, very quickly and then generate an action like in a very high frequency is really hard, and that is kind of Cuba model as well. Computer use agent that they call it.
Henrik Göthberg:Anyway, but this is just the beginning. It's just the beginning, good, I think we stop it here with news.
Anders Arpteg:Yes, okay, good, let's see we have a bit too many topics here, so I think I'll jump down a bit here. I would love to hear a bit about the scientific aspect of the PhD works you're doing with WASP. So, hans, you are now the industrial supervisor for a PhD student. Can you please describe a bit more? What are you and he going to do there?
Hans Hjelm:Yes, so I don't know, does everyone know about WASP?
Anders Arpteg:You can give a quick introduction A quick introduction is good.
Hans Hjelm:So it's a Wallenberg-funded research project that started in 2015 and is supposed to go until 2031. And it's the biggest ever in Swedish history. I think it's 6 billion crowns 6.5. In total and they are focused on AI in different ways and it started out maybe more towards security and autonomy, but the part that we're interested in is media and language.
Henrik Göthberg:Sorry, Media and language. Media and language. Media and language. Media and language.
Hans Hjelm:So they have different research areas. They call them the one we're involved with Wasp Research Arena.
Anders Arpteg:Thing Sorry, Research Arena Okay.
Henrik Göthberg:And it's the Research Arena. It's called Media and Language. Yes, it's the Wasp Research Arena, exactly.
Hans Hjelm:So our PhD student, lukas Borgen. He is looking at specifically LLMs and how can we make them useful for Swedish media companies.
Anders Arpteg:This is nice because Wasp is traditionally very scientific and very little engineering.
Hans Hjelm:But, now with the Wasp Research.
Anders Arpteg:Arena. They're trying to get more in the engineering side, which is nice Sorry. Magnus, it's just very funny.
Magnus Engström:We're never going to get out of that tic from now. No, no.
Henrik Göthberg:I hope this stays on for every podcast going forward. It's beautiful, it's beautiful.
Magnus Engström:We're never going to get out of that tick from now. No, no.
Henrik Göthberg:I hope this stays on for every podcast going forward.
Hans Hjelm:It's beautiful, it's beautiful.
Anders Arpteg:Okay.
Hans Hjelm:But obviously the starting point was just noticing how much text we have available from Bonnie News. I mean, possibly we could start from over 100 years ago with Dagens Nyheter text, but we're not going to go back that far for this project. But we have large amounts of text from all the Bonne News brands and the idea is can we make the LLMs more useful for people who are working as journalists or editors or basically the journalistic side of things?
Henrik Göthberg:Are there any sort of research, theses here or angles that you're pursuing?
Hans Hjelm:I think they will crystallize as we go along. As Anders was saying, it's very engineering focused. But one research topic I think will be just around evaluation. So how can we measure? Did this get better or not? So that will be tricky, Like the quality of the content of an better or not.
Anders Arpteg:So that will be tricky Like the quality of the content of an article or what.
Hans Hjelm:So I mean there are benchmarks that we could use. There's one called, I think, scanned Eval, and there are different benchmarks, but maybe we would need a new benchmark to measure usability for journalists or usability for editors, and especially for Swedish. I mean there are lots of English benchmarks that we could maybe tweak or auto-translate or something like that.
Henrik Göthberg:So evaluating if new techniques improves journalism and then finding a way to measure that and benchmark that Exactly so.
Hans Hjelm:We have already tried to use LLMs for certain journalistic tasks. So, for example, can we use it to generate ideas for what's a good headline for a certain article. For example, can we ask an llm to generate five different headlines and then just let the the winning one take over once we publish it, or something like that. But what the journalists have told us in feedback is that they don't capture the tone of whatever publication. Is trying to use this exactly.
Henrik Göthberg:So that's a deeper. The brand thing is yeah, you know, that's a deeper, much deeper topic.
Hans Hjelm:Of course, so then the question is if we continue pre-training the base model with all this will the tone differ? Yeah, can we then? Access the tone, can you?
Henrik Göthberg:access your tone.
Anders Arpteg:Yeah, I like it and tone us in for the whole brand, or is it for the specific journalist that you want to have an article written in that tone of the journalist, or what do you think?
Hans Hjelm:so we have been talking about it from the brand perspective. I think the journalist's perspective would be interesting as well, but that's not in scope right now awesome and how.
Henrik Göthberg:How does it work? So you, you work together with us. You, you have been part of. How did you get? If someone's curious, I mean like how does it work, how did you get involved? Or, and how did you get the a student to help you out? This is awesome, so how do you do?
Hans Hjelm:yeah, so actually lucas was working as a member of our team, but we were already involved with the research arena and the language and media.
Henrik Göthberg:So that's where it started.
Hans Hjelm:Yeah. So they asked us for some data sets to start with, and we were able to give data to them to use for their students wanting to do experiments on Swedish media data, and that's how we got involved with the research arena. And then Lucas wanted to apply for this research position through WASP.
Anders Arpteg:Was Lucas working in Bonnier before?
Hans Hjelm:Exactly so. He was a member of my team.
Anders Arpteg:Oh, okay, very cool.
Henrik Göthberg:Super nice.
Anders Arpteg:And is he sitting and working mainly with you, or is he also sitting at the university and working partly, or?
Hans Hjelm:yeah, he's moving between the two which university is he? Linköping the best one, yeah, but there are some things that or other aspects that make this interesting, I think. So one is the possibility of hosting this LLM internally, like we have a lot of journalists that work on secret projects where we don't want anything to leave the building, basically projects where we don't want anything to leave the building, basically.
Henrik Göthberg:So then, having, like, our own bespoke LLM that we can host inside of BonneNews, and have you started to look at this architecture then, how you want to build it on an open source or that you fine-tune, or can you talk about that?
Hans Hjelm:So the starting assumption is that we will use one of the LAMAs and then continue pre-training the LAMA.
Henrik Göthberg:To distill out your own, so to speak, model.
Hans Hjelm:Right. So I mean, there are then questions like because when we do the pre-training we will not have an instruction tune model. So then the question becomes once we have the new pre-trained model, how do we add on the instruction tuning on top of that without losing any quality of the llm? And that's, that's another kind of interesting topic, yeah I would say a research area super important one. Yeah, you know.
Anders Arpteg:Some of the less good ideas is to build something from scratch. I think it's much more interesting and valuable for local companies like Bonnier's if you know how to adapt it and fine-tune it for your needs, because that's what I think more or less every company needs to do in some way. And then what we have been speaking about the big frontier versus the smaller baby models that we will have this is really what we, if speaking about, you know the big frontier versus the smaller, like baby models that we will have. You know this is really what we if we get that science right, we can get amazing engineering out in the end.
Magnus Engström:I mean, I think it's absolutely crucial for Sweden to have projects like this ongoing Right, and I also feel that we talked a little bit about this on the way over here that, um, like when we talk about data science, um, as a team at bonnie news, we usually apply data science to kind of solve a problem that is oftenly very business related right. But sometimes, like going back to the science discussion, we sometimes we kind of need to do things, take a step back and say, right, this is important for bonnie news, but it's also important that companies like bonnie news are doing this right and that in that case we we really like to do that more in on academic side, for example. So I think it's super important, not only for Bonnier News. I would hope that Swedish companies that have the capability to do things like this should be doing it, because otherwise we're already kind of behind.
Henrik Göthberg:So this is a profound topic. So why should we do it? And why should we do it open? And it is to stay competitive or stay relevant, and to build a critical I would argue, build a critical mass of talent and engineers. Who has the elite know-how?
Magnus Engström:for for this to diffuse out in society yeah, yeah, I mean, in the long run, what it boils boils down to for bonnie news is that we need to be like protective of a democratic society and like an open society. Yeah, and mostly this is through the journalistic work, but I mean this is also an aspect of that. This is through the journalistic work, but I mean this is also an aspect of that. This is like in the line with the vision of Boni News too, and I mean there's a lot of other companies with similar visions and mission statements and I think that it's important. Things like this is taking place.
Henrik Göthberg:I mean like I don't know if you've been following the pod. So we have been following this topic on the Swedish path to language models way back, so we've had Lov Börjesson here. Are you friends of Lov at Kobi Labs? I?
Henrik Göthberg:know who he is yeah, and I even used he helped me even do my keynote at Data Innovation Summit three, four years ago when I was talking about economies of learning and adaptability and use his idea Like he starts from a BERT model and then he develops many different KB models, which is adapting what is already there and where I say, like standing on the shoulders of giants, economists of learning, that's how you stay competitive. And then we have also had, you know, one of the main engineers who builds SWE Geppetir, who works in Magnus Eugren's team.
Henrik Göthberg:And we were sort of questioning and pushing him why should we build a Swedish frontier model? And should we really build it from scratch? And in the end, you know, after the podcast, the real conclusion is the real rationale is to learn. Not that it's useful, but to learn is the useful thing.
Anders Arpteg:And I buy that, even I don't actually, because it's actually very open how you do it. Yeah, and it's open.
Henrik Göthberg:There is no new knowledge actually because it's actually very open how you do it. So there is no new knowledge. And this is just full circle back on, when Anders is now saying that we are learning how to use a LAMA model, fine tuning it, adapting it and figuring out those technicalities that we're talking about here, that is actually to me that statement from Anders was much, much bigger, because it means also how should Sweden be competitive? What should we be doing in all this? And you know which path should we go? And we applaud that path. Personally, it sounds, and I applaud the path that is sort of the mini model. You said it before, even before the port started we're going to go to the edge. Model is going to be more. You know, the Apple way will win.
Magnus Engström:Yeah Well, I don't know if it's a serve-sum game, but I mean I think it's a good strategy to kind of if you was it Alan Kay, I think that Steve Jobs quoted him saying that if you're serious about your software, you should build your own hardware. And I think it's an Alan Kay quote, but it's used by Steve Jobs and I think that that is very true. I mean, if you own, if you have the possibility to develop hardware and you're great at software, as obviously Apple is, in that case it only makes sense to kind of try to see how far you can push the envelope of putting the model into hardware.
Anders Arpteg:Wouldn't it be super ironic if actually Apple turned out to be the best one to find value from AI, when they have been the worst to use AI?
Magnus Engström:even to this day. That's the case I'm making.
Anders Arpteg:I think that we're a little bit.
Magnus Engström:we're kind of rushing to judge my turnout. Same thing with Meta as well. I think that Meta, but they have the most.
Hans Hjelm:Are you saying that Bonnie News will move into hardware?
Henrik Göthberg:Yes, you know, you saw his. Now I saw your ambition, I saw your real ambition. But it's a great topic, right?
Magnus Engström:Basically, of course, Bonnier is already in the hardware business.
Henrik Göthberg:Like, we print our articles on papers.
Magnus Engström:And then we.
Anders Arpteg:That's hardware that's hardware, that's UX. Let's try to capture at least one or two topics before we close, if that's okay.
Magnus Engström:And also I should flag that I'm sort of running out of time. Yes, I know If you have to take the hard topics from 10 minutes from now, then Hans will be happily answering those.
Anders Arpteg:Let's try to end in 10 minutes if that's okay, so you will have to be able to catch your train and whatnot, and perhaps actually moving to you a bit, magnus, I've taken up a lot of airtime in this podcast, but you kind of provoked me into it also.
Magnus Engström:I must say Fine blame it on me.
Anders Arpteg:Okay, but just thinking about you, thinking about actually value capture from AI in general and thinking a bit about a lot of companies in general are really struggling to find value from AI. They get stuck in the build a lot of prototypes, the scientific see can it work or not, but don't really make a product that actually do provide value. I would like to hear a bit, perhaps, how Bonnier is approaching this. Do you have some kind of idea of really trying to see how AI can provide value in as quick way as possible?
Magnus Engström:Yeah, it's a short answer that will lead to a longer discussion, but the answer is that Bonnier News doesn't do AI projects discussion. But but the answer is that Bonnie news doesn't do AI products projects. We don't say that let's sit down and figure out a new AI product. What we do is that we sit down and like saying that, right, we have this process, uh, and like we have, we have different ways of identifying. We have different ways of identifying when things turn into something that we should probably try to approach with AI and that is the whole idea of cross-competence teams and the idea of domains and areas and trying to make sure that everything kind of flows together is just making sure that whenever someone have an innovative ID and the only way to kind of get that off the ground is by using AI, then we need to be there and identify that that's the case, right, and you can also I mean, you could talk about being innovative and like finding new applications for AI, saying that we have those ideas that we couldn't do five years ago because we didn't have the resources.
Magnus Engström:We figured that this would be a great thing to have. This is a product feature that we would love to give to our users, but we don't have the resources and it's going to take two people working full time to provide the content for this functionality or whatnot. But it's not a new idea. It's just that it was very resource intensive and so we kind of scrapped it. And now, five years later, we can kind of revisit that and say, hey, we don't need two people working with this anymore, we have this technology right. So I really try to be clear with that. We don't do AI projects.
Henrik Göthberg:You do innovation projects. Not always innovation projects, yeah, not always innovation projects either.
Anders Arpteg:Let's see if I can summarize what you said and if I understood what you mean. Either you can say that we have a set of business processes today and potentially we can see can we optimize them in different ways and find AI where it's most useful in existing processes, not for new products. The other thing that you're saying, if I understand you correctly, is that you had thought about old ideas before that was not really feasible to do, but now potentially are. But that means building a new process or product.
Magnus Engström:Yeah, sometimes it depends on the case right. For example, it would be great to have more relevant content exposed or shown in connection to the article the user is reading. But going back like five years that meant that someone has to kind of sit and make those connections manually, or putting metadata and some granular metadata often, and build the algorithms on.
Anders Arpteg:I'm trying to rush you here. Yeah, sorry, because we have a final question that we'd like to capture. That's right and we'll continue a bit with actually Hans later.
Henrik Göthberg:Yeah, that sounds like a great idea.
Anders Arpteg:But I think this sounds awesome. It's actually resonates a lot with me as well, if you really want to quickly find value from AI, it's better to use the existing processes.
Magnus Engström:Otherwise, it takes a super long time to change the organization.
Anders Arpteg:Yeah, because it has inbuilt processes that are hard to change and if you can start to optimize them, that means you can quicker find value.
Magnus Engström:Yeah, yeah, so I figured that you meant I was thinking that you were going the opposite way. Haven't you kind of rebuilt the organization to take AI into account? No, we have taken AI into account. That's why we haven't reorganized Exactly.
Anders Arpteg:That's, I think, how everyone should think. So it's a really good approach, I think. But then you know, the big thing I think happens when you start to transform the processes as well, actually starts to have new process, but that takes longer time Because it means changing the way you're working. It means you have to have new processes, but that takes longer time because it means changing the way you're working. It means you have to have new things in production. It means you have new processes in place and that is hard, but it potentially has a bigger impact, but it's more long-term. So the best way to find value in the short term is to optimize existing one. I think that makes a lot of sense. Awesome, okay, quick one, and then we take the AGI. I was thinking, henrik, you can let me know. Great.
Magnus Engström:Let's leave the AGI questions to the society in the future Should we?
Anders Arpteg:I have the creativity and music thing, but we can take that with Hans later. And if you have to leave at seven. But perhaps just thinking in general, coming 10 years in AI in a media landscape in general, the progress in AI is moving in an exponential phase, as usual, and a lot of things will happen. How do you see Bonnier News in 10 years?
Magnus Engström:Super simple question, yeah, yeah so, to start off answering that, um, I think paper, yes, yeah, yes.
Magnus Engström:So I think that to a large extent, I think that actually this is also something that we talked about before me and Hans Like, in the future, we're going to see more and more agents, kind of sharing the same context. I think I have multiple agents and they're going to a large extent, I think that assistance and agents on the user side will act with more, with incentives. Today, if you look at the operator, you give very strict instructions of exactly how to act, right. I think that if you give it 10 years, I think that the agents will work towards objectives, yes, and I think that that is something that will take into account and building news products as well. So I think that the agent representing the end user is probably going to have more agency, so to speak, and more be more object drivendriven and have stronger incentives to act one way or another, because it's a very much like a better connection between the user and the agent in terms of what the user actually prefers and wants.
Magnus Engström:So you're combining agent evolution and personalization evolution, yeah, but you're still going to have information need. You're always probably going to have information need and that is kind of where news products come in. So we need to kind of map the. You have the user and you have the much, much, much more sophisticated agents in 10 years, and then you have products that satisfy information need and then we have to understand kind of how will those Matching. Yeah.
Henrik Göthberg:So we will get in all the noise. We will get the more surgical personalization in terms of with an agent with a clear objective, helping us with the news we really want. And so it means that on the production side, that it will evolve in order to have the news bits in such a way that it's completely modularized and can be fitted into a different packaging for each user. And on the other side, you have the agent assistant. That is objective, driven to support the personalization for consumers.
Magnus Engström:Driven to support the personalization for consumers. Yeah, so today, for example, if you look at an app like Spotify, you have like Spotify, I guess, is building like this model of the user.
Magnus Engström:And they use that model to understand what music is recommended, what times and so on. I don't think that that model will not be at spotify. That model will be on the agent side, right. So agent will go to the spotify who provides music and the agent will know exactly what to search for because it will act upon whatever state it's an interesting thought.
Henrik Göthberg:It's an interesting thought that you will have our personal agent that we will load with music objectives, news objectives, etc. Etc. Yeah, and with that with that, on that cliff, on that bombshell.
Magnus Engström:I always wanted to I always wanted to kind of stand up and leave, so this is my chance but that was top gear and on that bombshell, I know you have to run, magnus Engström, but thank you so much for coming here.
Anders Arpteg:We'll stay and continue to speak a bit with Hans, a bit more.
Magnus Engström:Yeah, really nice to be here.
Henrik Göthberg:We loved it so fun.
Magnus Engström:And also I love the arguments.
Anders Arpteg:Same here. Because, and I love your comment saying and I know this is very controversial, so let's go there.
Magnus Engström:Yeah, I mean, I mean either we can't talk about it or we can right, so if we can, we might as well. Yes, that's cool.
Henrik Göthberg:Thank you so much. Thank you so much and yeah, don't miss your train.
Anders Arpteg:Cool With that. Do you have anything to add? Before I'd like to talk a bit about music with you as well. Yeah, since a lot of people here is very interested in music and how AI is impacting that, and I know you are as well, hans. But do you have anything to add? What do you think about the next decade in the media landscape and how AI will impact it? Any thoughts about that I.
Hans Hjelm:I think this is um. I really like this idea that magnus is presenting. I'm not exactly sure. I mean, how will, uh, this agent that magnus is? Projecting yeah how will it, this agent that Magnus is projecting, how will it talk to Bonnie News Like, how will that business model?
Magnus Engström:work, will you?
Hans Hjelm:still buy a subscription to Bonnie News to have your agent talk to.
Anders Arpteg:Yeah, the whole business model idea that's a big question. How will that?
Hans Hjelm:work in 10 years, and I mean you could imagine that you pay a premium to someone like Bonnie News to get this nicely packaged news information that you can trust, and then maybe you have open source that it's free. Just to elaborate a bit about this and I think it's super fun.
Anders Arpteg:We have spoken a bit about this, you and me, henrik. But let's imagine that we have a future where we have a few, a super big frontier, but a large number of small, open-sourced or otherwise available models that all can generate whatever news they can in different ways. I think and hope that there will be greater need for finding sources of truth and trust in the future. I think and I hope that media institutions can be that kind of source of trust, so to speak, because it will be so many different news angles and it will be so easy for anyone to produce their kind of opinionated kind of view of whatever news, opinionated kind of view of whatever news. But people, hopefully, will be very drawn to a place where someone is providing like objective journalism in some sense it's even.
Henrik Göthberg:Like Magnus said, you don't need to have fake news, you just need to have angled news. We're doing it all the time. We're cutting out snippets and taking it out of context. That has been done for hundreds of years. Like we're doing it all the time, we are cutting out snippets and taking it out of context. That has been done for hundreds of years. So we don't need fake news in order to have the wrong news, so to speak, or two different truths. This is proven already. So that also leads me down the path and I was talking a little bit about like that.
Henrik Göthberg:I was going in that mindset before when we were talking, but it actually fits better here. You know, almost like the brainstorming around product development, what are the core products? And I think you are onto something. I think one of the core products is potentially trust, and how do you package trust? This is what I was arguing that we have these new types of news site and this is new products where the core product is trust, and that they are giving you a 360 view on what has been written, and they are also clear with articulating that this angle I don't say it's wrong or right, I'm saying this is the blue angle and this is the red angle.
Henrik Göthberg:This isn't the blue media, this isn't the red media, this isn't the blue community groups and this is in the red community group and this, for me, gives me as a reader a more nuanced view of the world.
Hans Hjelm:It's not the place of Borny News to present the blue and the red. I think Borny News has to make some kind kind of decision. So this is how we see this event happening in the world.
Goran Cvetanovski:This is the difference you are one of the news out there and this is an, this is essentially an aggregator I'm discussing an aggregator product, so good.
Henrik Göthberg:so you are a news, so you are standing for a point of view and this is clear. But you can still have product features that are trust-centric or balance-centric. You can do different things, Even if you have your angle. I think it brings more credibility, Even if I know you have a blue angle. But if you present other topics or related topics or other articles and clearly showing, or related topics or other articles and clearly showing, I think there are different ways where you make Bonnier the number one choice where I go in and look first and that is also about this I guess I would summarize it as the trust product or the trust features. I think this is going to be a big deal.
Anders Arpteg:Just before we go to the next topic. If we flip the argument and just think the most horrific way forward, it could be the case that no one can argue against the big frontier models that are so knowledgeable, that are so trustworthy. Potentially I don't think they will have it, but there could be Today, when I get unsure about something, I go and ask Chateau, petit or Gemini, not Perplexity, yeah, that's as well, I like it. I know I like it.
Henrik Göthberg:When it comes to questions that I want to clarify, I actually go to Perplexity first. I usually like their way of their UX.
Anders Arpteg:No, but it could be the case. I mean, I'm already going there as my most trusted source in some senses.
Hans Hjelm:Yeah.
Anders Arpteg:Do you think that could continue? But is that news? Yeah, that's my question as well. Is that informational news? You know, the first thing I did when I used Operator, or actually this new kind of task functionality that you can schedule tasks in OpenAI is to simply say give me once a week the top AI news, and then it gives me the top.
Hans Hjelm:AI news each week.
Anders Arpteg:And it's rather nice.
Hans Hjelm:I've set up something similar, yeah.
Anders Arpteg:Wouldn't it be hard to beat that kind of service, you think?
Hans Hjelm:But it gets the news from news outlets right.
Henrik Göthberg:Yeah maybe, maybe. So that means also who's your customer? Yeah, I mean like so this, this is the value chain shifts you know. You know when are we the commodity and who who do we have. You know who's the customer? Yeah, in this industry, shifts you know vastly common and if we look back in history, 100 years, these there have been major shifts like this, this in different industries. So I think that if we look 10 years ahead in the media landscape, I don't think you can assume that there's the same fundamental actors of value chain. No, I don't know, not in 10 years.
Hans Hjelm:No Predicting anything 10 years forward 10 years is a long time.
Anders Arpteg:I think it's hard enough to do two years.
Henrik Göthberg:But in essence, the interesting point is that your argument fits with Magnus' argument because essentially this is a news outlet and you are using ChatGPT as your personal agent. You are giving your personal agent an objective function. I want AI news weekly and it fetches it for you from whatever news outlet it can receive. So essentially, you are playing out the vision of Magnus right now.
Anders Arpteg:Yes, it's no different, it's true, it's true, it's true, it's interesting, I think he's right I think so.
Henrik Göthberg:That is interesting then what? What is? Who is going to build the agent? Even Magnus said so.
Hans Hjelm:Yeah.
Anders Arpteg:Cool, let's jump a bit in the topics here, since you have a music interest as well. I have as well, even though I'm not at all to the level that Goran and Henrik, and I'm sure you as well, hans, but if you think about that, you know there's so many. We need to play some guitars after AI is progressing so much in generated music as well and I've used Sonos a lot and Udio and these are really cool stuff that actually produce music that I think is enjoyable to listen to. Have you tried these?
Hans Hjelm:Yeah, is it called Suno.
Anders Arpteg:Suno is one of them, yeah.
Hans Hjelm:I've tried that one and I found it fascinating, I must say. I write mainly instrumental music and what I tried is so you can upload your own music into the service and then if what I tried was just write lyrics that fit the melody and then you could ask it to either continue the song or make a cover version, so that was a way of me hearing my own music with like a professional singer singing the melody Crazy, yeah, I thought it was fantastic.
Henrik Göthberg:And it's not that much work even.
Hans Hjelm:No, no, it doesn't take a lot of work and it's free, unless you want to do too much. Yeah, I think if you would use it more.
Henrik Göthberg:It's like a I don't know 10, 10 dollar subscription but you, you who are in a position, um, where you're trying to I mean, I just you said it yourself I started a career as a musician and then I figured out oh, it's a difficult thing to make a living. How will this change with ai in the mix? Can more people actually make a living, or even fewer make a living?
Hans Hjelm:That's a good question, I don't know where this is going.
Anders Arpteg:I can create music now.
Hans Hjelm:But I don't think I can make money from it. No, but I mean people. You have to have people who are actually interested in the artists somehow. I think. Otherwise maybe your music will become popular as background music when you go in the department store. But if you want talking about fans, then you have to create some level of interest in the artists behind it, and the artists can use AI if it's like and people can use.
Anders Arpteg:If I want to listen to Taylor Swift or something, I can simply have AI generate music in her style. Yeah Right, and what would that mean for Taylor Swift? She has more money than she needs anyway. But you see, but someone has.
Henrik Göthberg:We have already had this discussion like it's also shifting that we, the artists, that we that have fans, it also becomes the whole idea to see them physically and go to the shows, and so so you, you can also see also how this is merging. You know how you're selling tickets to the shows inside the spotify app, you know.
Anders Arpteg:So, yeah, I mean if we compare it, to stress, I think it's an interesting comparison. You know, we know we have no chance against ai in chess, but people and humans are still playing chess against each other yeah so they do appreciate, like the, the human experience of playing chess. I think what you're saying we've already seen that you know what artists make most money from these days unless you're a super big artist is actually the physical kind of gigs that you do, right.
Hans Hjelm:Yeah.
Anders Arpteg:And perhaps that will simply increase and you do find value for seeing a human being doing the musical act in some way. Yeah, and that perhaps will continue simply.
Henrik Göthberg:I think both spectrums will exist.
Hans Hjelm:Yeah.
Henrik Göthberg:And I can understand how I can enjoy an AI DJ not only AI DJ, but an AI DJ channel on Spotify when I go on the radio and then I want to listen to Bob Dylan, or, you know, I want to listen to my and I want to go to a concert and see my favorite band.
Hans Hjelm:I heard an interview with a woman who works as a digital artist, but she has physical challenges, so she can actually paint herself, but she can generate art using image generators, and so her point was that now it's not a question of talent or technical skill, it's about taste. So she says she has fantastic taste. Everyone likes the art she generates, but she doesn't have she. She doesn't have the physical skill or the physical ability even to generate the art herself.
Henrik Göthberg:Maybe you could see something similar happening with music but that that can even be, that can be explained as truth in music. I mean, if you have producers like alan rubin or someone like this weird guy in a beard who always were able to hit the taste of the time. You know max martin, they, they have tape. I mean like they know what. What's their brilliance? I mean like they can produce music, yeah, yeah, yeah, many can do that, but he can actually nail the taste that sort of resonates with society. Maybe that's it, yeah.
Anders Arpteg:But in general will it go up? Will it be easier to make money from music?
Henrik Göthberg:in the future or not? Yes, the answer is yes. Why dollars? Yes, why okay? Why because?
Goran Cvetanovski:we are basically making too big point of this, and we have discussed this before this is a brush in the round like 2000, maybe a little bit earlier than a brush it's a brush right.
Goran Cvetanovski:So because I'm right now very much into edm electronic dance music right. So now you have this AI tools where you basically you can write midi notes, right, and then you choose a voice and then you have a. You write the lyrics and then you get basically a singer. You choose between different voices, right, and you have now vocal for a track, as you're saying, right, the problem is that you are limited to four, five, six voices.
Goran Cvetanovski:In the 2000s there was a program called EJ EDJ a beautiful program for EDM, but you had the samples and everybody basically could produce music. They had the same samples, right. Music is about patterns and making new things that sound a little bit stranger, new voices and stuff like that. So I see this as a empowerment of a music, not actually as something that is negative. It's another brush. So now I have my air assistant vocalist that I can actually create new music, I think new genres will be born and then you will see basically a mixture between those things. Now we have synthetic data data uh, synthetic music. For a long time we have screen lakes. We had, like a rap.
Anders Arpteg:Yeah, yeah, but but uh, the background in the eighties it was sampling.
Goran Cvetanovski:So what are we talking about?
Henrik Göthberg:It's just basically, and we have already seen, uh, the the rise of our Swedish House Mafia, avicii, kugel, who they're almost growing up as DJs, turning producers and from producers all of a sudden. No, this is good music and we want to go and see them as artists. I even went to a Swedish House Mafia and Kugel concert and of course it's production.
Goran Cvetanovski:You go and pay for MasterChef. You go here in Francena you eat. He cooks the same food as we are cooking at home. But he has a taste in how it's combined. His taste is very.
Henrik Göthberg:So your argument is actually we are opening up the creativity for more people that have the great taste. I like that way of framing it.
Hans Hjelm:I mean, the flip side of the coin is that maybe you will flood the market.
Henrik Göthberg:Yeah, and then foreign yes. That could actually be worse, it will be a lot of garbage In a short term, for sure. The problem is also there's going to be a flood of garbage.
Hans Hjelm:Yep AI slash, what's it called?
Henrik Göthberg:We talk about this on YouTube, we talk about this on Twitter, we talk about this everywhere.
Anders Arpteg:Just before we move to the final topic here. I mean, we can also think about the question about you know who is really in control. We're all assuming that the human will be in control when it comes to producing music, and then you use AI as a brush for a tool. But it could be the opposite side, saying ai is actually in control and then using humans to produce music in some way. Um, and that would be the ai producer, the ai people.
Henrik Göthberg:Yeah, yeah, record company, oh, why not? They are they. Are they having the perfect algorithm to know what's going to sell? They have the best taste. They have a scientifically optimized taste.
Goran Cvetanovski:Yeah yeah, yeah, but then that means that the AI needs to be in a position to decide what is important agent.
Anders Arpteg:They can make decisions.
Goran Cvetanovski:I want you to objective function to sell the most records families, husbands we're not listening how are you going to? Listen to AI.
Anders Arpteg:I listen to AI today. If I want to learn something, I use AI to do it. You could be doing the same for music.
Goran Cvetanovski:As long as it's not saying can you go out to a block to buy some things for me? Then I will see how. You go out to a block to buy some things for me, then I will see how you're going to respond to that one.
Anders Arpteg:Cool, it's going to be an exciting future, right, toms?
Hans Hjelm:Yeah.
Anders Arpteg:We can be sure about that. It's going to influence the market for music and any kind of market, I guess in the future to such a great extent?
Henrik Göthberg:And do you still play a bit for fun at home? Yeah, I do. Yeah, do you have hobbies live and kicking? Absolutely Good to hear.
Anders Arpteg:Hans, if you were, do you believe that AGI will come? Yes, any idea about timeframe Two years, five years, 15 years? Without the definition, without the definition.
Hans Hjelm:Okay, let me give a definition.
Anders Arpteg:You can also say what you want. But if you take a definition similar to Sam Maltman, he basically says that AGI will happen when you have an AI system that can do tasks equally well as an average coworker. People today, I think, don't really understand how hard that is, because AI today is basically about perception. It can't really do reasoning good, it can't take action as well. So to really being able to use operator, but in a much better way I mean any human is better than the operator today, for sure. So it's easy to see that we are not at that level today, and far from it, I would say. But it will be a point where it would be. Instead of asking Henrik to find a consultant that can do that, I simply will do something else and ask an AI to do it, and they can do all the tasks, including taking actions, making decisions and we're talking about computer use, not like physical work.
Anders Arpteg:For me, I've had three levels of AGI Cognitive one without actions is one thing, I think digital, but taking action is another one. And then you can take physical one as well, and potentially a fourth with societal disruption, but it doesn't matter really, let's take a physical one, then that's probably much further away, really.
Anders Arpteg:Let's take a physical one, then that's probably much further away but if we take a physical one, where you have humanoid robots that can do what humans can and can literally do what an average co-worker can physically, not only digitally, that will come at some point as well, I guess, right, so let's take that we have that point where we can have an ai physical, ai humanoid robot that can do tasks to the level of an average co-worker. What would that mean? You can think of it in two extremes. One is the dystopian future. We literally have the terminators of the world and the matrix that will be using us for bad gains in a really bad way, that will be using us for bad gains in a really bad way. Or a utopian one, where we actually have AI that helps to cure cancer and fix the climate crisis and energy problems and everything becomes a world of abundance where we can live for free, more or less, and pursue our creativity and passion. Where do you think we will end up in this kind of scale and spectrum?
Hans Hjelm:I have been very optimistic, but I am slightly worried about the development in the US currently, I must say. And if we get a very powerful government that goes down a totalitarian path and they also get access to AGI, then I think that is not the state we want to end up in, because that will not be fun, I think.
Henrik Göthberg:We are talking about an AI divide where we have total inequalities, where there is a very, very few who then control something that is so much more potent, or productive. Yeah, I mean. And the argument would then stay very segregated or the power would be very concentrated to a very, very few set.
Hans Hjelm:Right. I mean reaching, that point first, I guess will be very important, because that will enable you to do a lot of things that your competitors can't.
Anders Arpteg:I think that's an interesting question, perhaps if we could spend some time on that. A lot of people are saying it's the one you know, the one that reaches first will win everything. I have a bit hard time seeing that actually. I think even if you have one company that develops AGI first, I think the other ones will be not that far behind and they can still find a lot of value from a second to best model.
Hans Hjelm:But combining that AGI with the powers of a nation state. That's.
Anders Arpteg:Yeah, of course the abuse is horrible. So that's my current worry so you're more to the negative. If you were to say like 50% in the middle, 100% positive, 0% negative, are you more on the zero?
Hans Hjelm:negative. Still in positive, I would say okay, over 50% at least. Yeah, good, yeah, we had some. How about? It's still in positive? I would say Okay over 50% at least.
Anders Arpteg:Good yeah.
Henrik Göthberg:How about? One of the key arguments has resonated with me how to answer such a speculative question has been several guests who went away. We will have both. So we will have both abundance and we will both have. You know, like we we have today, we have some people who are rich and we have some people who are not. So the problem then is how huge are these inequalities, is my kind of second thought. So the answer then has been a little bit like okay, it's really hard to say we're going to be on 50. It's more like we're going to be on 70 and 40 or so, I mean like it's going to be quite shitty and it's going to be quite good. And the problem then is if that quite good, is that a very, very small percentile? And that is my take on it.
Anders Arpteg:I think we're going to have both and we want to minimize those inequalities that divide um I have, if you take you might start an answer to this, is basically I'm more afraid about the short-term use of stupid ai that we have today than the long-term ati.
Anders Arpteg:I'm I'm more optimistic saying if we have truly reached ati, we can have AI systems that control other bad AI systems, so that, I think, could be a good situation to come to. But today and literally coming few years or months, we could have people, bad actors, that actually abuse AI to a large extreme and that could be really detrimental and dangerous for our society. So I'm more afraid about short-term abuse of normal, before AGI kind of situations than I am of actually having AGI. Another really scary for me that we are seeing, I think already today, is the concentration of power that AI leads to. We have already seen this kind of hyperscalers that are richer than most countries are. They have more power than most countries have. Yeah, now we have a single person that had more power I would say that many countries have. Um, that is very scary, yeah, and I think that will continue, and that kind of concentration of power is not a good thing. No, and that's something I'm very afraid of as well.
Henrik Göthberg:And there's a follow-on question here we should regulate that power for that person, where I think the real, the optimistic way to solve the AI divide is to focus on that. More and more people start using it and start moving their way up, a little bit, like Magnus was talking about how important it is that the companies that can do the research in this they should and they should do it open in order to get the snowballing effect around us. So I think this whole mission is to wake up the sleepwalkers. I think we are sleepwalking. I think a large part of society and leaders are, to me, sleepwalking on this topic and if they meant business, they wouldn't do prototyping and ending up in a prototype graveyard with an innovation theater of AI projects, but they would simply innovate process as hell. And if we innovate process as hell, we would reach a productivity frontier that we need to use a lot of agents in order to further it.
Anders Arpteg:So I think we have not really started 90% of us and people don't think understand how fast the hybrid skillets are moving. They are moving much faster than most companies and then if you take Elon Musk companies, they are moving much faster than the hyperscalers and that type of increase in divide that we're seeing it's very scary to me.
Henrik Göthberg:So, for me, this is why we started the problem to talk about the AI divide, to demystify it, to inspire, in order to, rather than pointing fingers at the one at the top. This is scary shit. Get your fucking act together, Not, you know? Don't blame them for being smarter. Blame yourself for being slow. To be a little bit cynical, why aren't we getting into it? You know for real.
Anders Arpteg:Well, let's try to end on a positive note.
Magnus Engström:Now we went into a rabbit hole here, very negative about the future, I get frustrated.
Henrik Göthberg:It could also be the case.
Anders Arpteg:You know that we actually can use ai for very positive purposes.
Henrik Göthberg:We've seen like alpha fold really being a super positive absolutely valuable purpose we can find, you know, gains in curing healthcare there's the other angle we we will not solve the climate crisis, we will not solve the energy crisis. We will not solve these things without technical innovation. We need to innovate ourselves out of it yeah, yeah, I mean I'm so sure.
Anders Arpteg:I'm so eager and hoping to get to AGI as soon as possible, because I think it will be a very beautiful future, as long as we avoid these kind of problems.
Hans Hjelm:Yeah, yeah, that's the bottom line, I agree.
Henrik Göthberg:Yeah.
Hans Hjelm:Thank you for putting it up.
Anders Arpteg:Hans-Jelm, it's been a true pleasure to have you here. I hope you can stay on for a bit more discussions in the off-camera after-after work, discussion and try the guitar. And perhaps play some music. Yeah, that would be super cool. Thank you so.
