E182 - AI Vibe Marketing Is Here - Berner Setterwall & Tom Ström

0:11[music]

0:16small thing because you know, he's basically saying that today if you just combine combine all the the chip output that we are producing per year, it's basically 20 gigawatts. Yeah. And he says basically in in Tesla and Optimus Bot and in SpaceX and and in xAI and everything, he will need uh like 50

0:34times more. Yeah. Like one terawatt. So, that is the Terafab name. So, okay, he wants to use more compute than you know, 50 times more than it's being produced per year today. That that number in itself is just insane. And I mean, it's just the space and energy you need to run that. Yes. So, yeah, that's the second thing then. But also, you know, we have a big um bottleneck in terms of who is really

1:01creating the compute. Yeah. Of course, we have the European ASML company that's building the machines, the lithography machine, yeah. But then it's being you know, designed by Nvidia and then being manufactured by TSMC in Taiwan. Yeah. And then no one can really compete with them in terms of the most advanced chips. Now, and that creates of course a big bottleneck. So, it also makes it really hard to innovate because if you have a new design, a new mask that you have designed for the chip that Nvidia does, then they have to send it to TSMC and then manufacture something, they try it out, and they have this kind of really

1:34long loop to really do innovation. Months. Months, yeah. If not more. So, now what Terafab is planning to do is to take that on themselves. So, they they can basically do the design themselves. They can do the the manufacturing themselves, and they can do the the use, the application themselves. Meaning, they can really take that loop, the iteration loop for innovation, and put it down from yeah,

1:59probably years to to weeks or something. Yeah. And no one else is anywhere close to be able But but is that feasible or is it the Elon Musk claim? I mean I think he'd done it with rockets before. I mean when he was starting SpaceX he we went around to all the rocket manufacturers and asked I need a rocket I can do this and this and can launch and you know, relaunch and then so on and nobody said that they could build it so I had he had to build it

2:21himself. So I mean he'd done it before. He has done it before if you look at it that way. I agree with that. But then even the what you mentioned with energy, you know, the most crazy thing that TerraFab I think is they're going to build it on the moon. Yeah, that's a logical extension of the energy implica- If you want to build

2:38TerraFab, what's the energy implication? Hmm, let's put it on the moon. Yeah. And and then It's not part of the plan like because it's it's a logical consequence of saying TerraFab. No, but he's saying like within 36 months the cheapest place to deploy new AI will be in space. Mhm. And long term I fully I think he's correct. When if he if we have cracked how we put things in orbit and then he cracks how how he puts

3:03things cost-effectively on the moon. Yeah, it makes sense. using the Artemis, you know, moon uh uh project now undergoing with NASA. You know, he's going to use the Starships to actually put the people on the moon. It's not going to be the NASA rocket, it's going to be Starship. So he actually get NASA to pay for him

3:22building such Starship rocket. the list to So he can then use it to build the TerraFab on the moon. Yeah, they're going to like outcompete them on the moon maybe. They they create the thing that people and he got plan to put like a whole AI factory on it so I mean it's such a smart and strategic

3:39move he's doing. I mean but but at one time I mean he's always said that he's going to do this. Uh and it it's kind of now that you can see the clear view of it or the whole plan. But he's always said that this is the whole plan I mean, if they can build that much AI I think he said himself in the in the interview that there isn't going to be a single problem for a single human being

4:02on Earth that can't be solved. And then comes another question. Do we need that much AI or will we fundamentally come up with smarter ways of, you know, slimmer ways of doing things? I mean, like is is the scaling just going to continue or are we going to think about fundamentally different architectures where this is actually not

4:21the way we're going to scale it? What do you think? I think it depends what you want to do, right? I mean, I think EU was thinking in these terms for quite some time. Like we should draw down all of the energy use, like minimize, try to preserve Earth as it is. But Elon Musk's plan is to consume 3% of the moon and turn it into silicon in the Dyson swarm. And he's like frustrated that we're still like not a

4:47Kardashev one civilization yet. So let I need I need you to unpack the sci-fi reference here. Dyson swarm, what is it? Yeah, so so basically I think all of this goes back to this Kardashev who is like a Russian physicist during the Cold War. He comes up with this scale of civilization where like Kardashev scale one, if I'm not mistaken, means that basically you capture your whole planet's all energy that it gets,

5:12basically. And then scale two, you capture the whole sun. And then people start thinking, how do you accomplish this? And then basically means that you build like a solar panel around the whole sun. It's the most [laughter] efficient way because you can't really go there and boil or anything because it will just also turn into fission, so you can't extract

5:30the energy without building solar cells. So that's the Dyson sphere. And the swarm is just like starting to build that sphere because it's just like a Starlink style satellite swarm. that moves towards the sphere. Yeah, exactly. Like when we reach level two, I suppose we we take all the energy from the sun. And then, as you said, Anders, there is like the galaxy, the

5:52Yeah. I think it goes to the universe. Continuum. [laughter] For all the sci-fi nerds out there, right? Yeah. Or maybe it's not so nerdy anymore. I don't know. Exactly. It's going to be in a reality near you very soon. In what in what in in a in a mailbox near you so what what's the in a outlet

6:10near you? But but it's also very You can think about it in the reason for Elon Musk to do so. One is simply that it is he means what he's saying, meaning he wants to make human species multi-planetary. Or he wants to accelerate the transition to sustainable energy or abundance, as he now calls Tesla's mission, right? Or is actually something else. And and this is simply something he says to really get people excited about working with us because it's, you know, a sci-fi movie

6:38more more or less coming to reality. I mean, I think it's a super smart thing to really concentrate power now, what he's doing. Yeah, I mean, have you read about the IPO he's going to do? On two trillion dollars. I saw two trillion now, right? It's insane. So, I mean, he's going to I think like as an entrepreneur, like it's not only the money, it's it's also about doing something fun. I mean, if I had that those kind of influence that he has, maybe I would build them Yeah, he was he's always already the most like rich person on Earth. So, like, you

7:11can't compete with that anymore. So. But if you read the biography from Elon uh what's his name? The the famous guy who wrote it write it. Walter Isaacson. Walter Isaacson. That's the one. I mean, like, there is a pattern here that is actually not so dumb. If you put an extreme goal like we did in the '60s, let's go to the

7:32moon. When you put that extreme goal out there, it drives you to solve problems that you then figure out and commercialize on the way. So, if you think about it, by putting up the fundamental bold goal of going to Mars, he had to build SpaceX, and in order to fund SpaceX, he had to build Starlink. You know, so so all the time it's a little bit like if you put an an extreme goal up, it gets you focused, and then you figure out what is the commercial models along

8:01the way. And it's not I don't know if that's bad. I think it's pretty brilliant. Yeah, no, I agree. And then we talked about innovation before, and I think like during World War, for example, there's been a lot of innovation. During the Apollo program, there's been a lot of innovation. So, I mean, there's definitely room, I think, in humanity for people to set these goals. And I

8:20mean, maybe he fails. them, where is the innovation going to go and come from? Then you're going to have the, you know, I love the advertising, you know, what was the top European innovation? We put the bottle cap that it don't fall off, right? Then you you don't have enough strong goals if that's the main thing you can come up

8:39with. That's a joke, right? I'm already looking forward to to the end of this podcast. [laughter] We were to go even more philosophical and And by the way, try those beers upstairs. The homemade beers. Oh, yeah. I'd love to try those. But before that, I would love to welcome you here, friends, ex-colleagues, and I think extreme AI and engineering expert, Benner Satterval at Satterval and Tom

9:04Strum. Welcome here. Thank you. Thank you very much. And we we were together a while back in Campania. It was one of my most, I think, fun times. I learned so much as well. Yeah, when you did a project according to a hockey stick, I heard that one. Yeah, yeah, we were young. That was it. I still remember, you know, the days when we had this competition with an Netflix to win their the worldwide marketing, and we won that, you know, worldwide. That's such a extreme Netflix, Zalando, yeah, some pretty kick-ass

9:36Spotify, some kick-ass customers. Yeah. Avito, yeah. I would say that I think Netflix was a lot of fun because of course they were respected as an engineering company. Yeah. And their own engineers were kind of like, "Oh, why are you taking in this external smart engineers? You know, we can beat them." And then we kind of had to rally our engineers and like, "Let's

9:55go." Yeah, exactly. And they had respect inside Netflix. That's very cool. And then we had the the office right above um Yeah, a famous nightclub. [laughter]

10:14Yeah. It was a very good office. of karaoke nights and yeah, it was I mean Is that when your karaoke career started? No. It came about [laughter] Maybe he was a big I you know, part of that. Definitely. I brought the quality down. [laughter] I love it. Anyway, we are here to to hear more about your work and especially with Cogny and your current approaches to to use AI and agentic AI

10:41for marketing and so many more things. So we'd love to dig deeper into that, but before we go into what you're currently working with, perhaps we can have a a short background about each of you personally. Uh if we start with you, Berner, how would you describe yourself? What Who is really Berner? [laughter] Yeah, I mean [clears throat] that's a a

11:00complex question of course. Obviously like my professional self, I'm a self-taught coder. I went to to gymnasium. I was supposed to go study at Handels. It was my plan like I had the grades and everything, but then I was like, "Maybe I should just try like one company before I start studying." Dip my toe. Exactly. [laughter] And and I managed to land a job at one startup and then I went to the next and and then I I co-founded Kampanj and here

11:28I am. Yeah. Still uneducated. you. [laughter] And then in my spare time like I like to be out in nature. I drive my snowmobile. I have a bunch of dogs. I have a family, two kids and my We live in Jämtland. I moved there 2020. Where do you live in Jämtland? Uh in between Östersund and Åre in the

11:46countryside. I've been thinking deeply in these ideas. Yeah, in in in a small village called Alsén. Alsén? Yeah, it's a beautiful lake. It looks almost like Switzerland because you can see like Oreskutan in the background of Let's talk about this later, yeah. Yeah, yeah. What is Alsén? No, I'm an awesome man and I've learned so much from you. I I was always amazed by by your both AI skills but also engineering skills. So um very

12:09impressive I must say. Oh, thank you. That must be good to have a PhD saying that to you. [laughter] Let me ask you one follow-up question. What's the trick to become a great engineer self-taught? How did you go about that? I think that I was just like really really really interested in computers from an early age. Like it's it's not

12:30that I was like without help as well. Like I have my brother who's working at Ericsson. He's he's done his whole whole career there. So when when I was like maybe seventh, eighth grade like I got my first, you know, thrown out computer from their office. Installed Linux, started playing, you know, deleted the whole file system many times. You know, a lot of mistakes but I also spent like a tremendous amount of time in the

12:56terminal. Yeah, tremendous amount of of time in the terminal. TTT, right? Yeah, something like that. [laughter] So yeah, it's the way Cloud Code had it. I was like rejoicing like in deep in my heart. Like there's always been this chasm between like developers who spend a lot of time setting up their environment and those that do very little. I'm I'm on the little side like I rather be able to work just with VI on

13:21any server that I SSH into. Um so the whole like terminal format of cloud code was like really I love that. Yeah, I can see that. Tom, how would you describe yourself? Well, I'm Tom. I am you know, I'm not a self-taught anything. I would say I went to school and you know, graduated Royal Royal Institute of Technology. I was um focusing on kind of entrepreneurship and

13:48marketing. Mhm. Got you all. Got you all. Yeah. Mhm. And then I went to Schibsted where I basically you know, Aftonbladet was going through this very kind of big change you could say where a lot of advertisers was leaving. Everyone was becoming digital and was this approximately? 2010. Oh. Yeah, something like that. Yeah. Um Yeah, so I stayed for a while and and basically helping them to buy companies and to run growth inside [snorts] their portfolio

14:22of companies. And then I kind of I joined the uh How do you mean? Campanja team. How do you mean burner? At Campanja. Oh, okay. Yeah, or it was I guess through like our shared friend Marcus. Yeah, because you were working with him part-time on Konstanten, right? Yeah. Mhm. Um yeah. And then like just gravity pulled all of this together into Campanja kind of or

14:46Yeah, exactly. How big was the team when it was at the biggest? At Campanja? Yeah. Was it around 50 or so? Yeah, something like that. 50 or 60? I had offices in London and Palo Alto. Of course. Yeah, of course. I'm cool. And now you have a new company and you're working with that. Can you just describe it briefly how you got started with both I guess I guess Growth Hackers and also

15:11Cogney? Yeah, I mean I think after Campanja we sold it to an Indian-American company. So we founded then Growth Hackers because we wanted to go a little bit wider because Companiya was very niche. We were doing just bidding for SEM basically or search engine marketing or Google Ads primarily. Um So, we wanted to go kind of wider, work with more channels, also kind of follow the users into the web page and app. Uh and that's kind of what what uh

15:42Growth Hackers is basically. Mhm. Helping companies drive on key moments, right? I think like we worked with Netflix for example helping them roll out ads across uh like they had they were successful in the US, but they didn't have a global footprint. So, they came to us and like help us run ads everywhere pretty much. And we rolled out, you know, country after country and and we launched in Sweden. There was a launch party. They came here. We were all like but then it was also like getting the performance, getting the users cheap enough, getting the scale so that they could take this like streaming platform

16:13position that they have today. And then they had like [laughter] zero Swedish TV shows on the landing page. How's your life? the content that you have in like it's not even available here because of the content restrictions. So, you like maybe optimize this page and we can do a better job optimizing your ads. And they were like you know, you shouldn't care about that. Like let us handle You're an SEM guy, you know, keep keep

16:35on your corner. [laughter] That's when the whole idea is And that was kind of like uh Mhm. Out of frustration seeing But I mean I guess not only as well. Like we had this idea at one point in Companiya to try to build this like business cognition type of company where like we optimize the business metrics end to

16:54end with like advertising as the lever. Mhm. Yeah, you remember this that we built our own tracking and yeah. So, we were kind of going in. And perhaps we should explain a bit how that type of SEM marketing really works Companiya style so to speak and then if we expand [clears throat] it a bit more

17:11to what we did today. But if you were to explain a bit more for yeah, the type of second order price marketing and these kind of techniques. How does that work really? Yeah. Do you want to go for the goal? Uh yeah, go go ahead, burner. Yeah, okay. So, basically like Google has a site, a lot of people come there, and they Google things, and like they started out

17:32without the business model pretty much. They were just like, "Oh, people like to find the structured information on the internet." Yeah. But then at some point like they had had added this AdWords feature. Like I think one of the first customers was like a lobster sales guy that was like, "Oh, like finally sell my lobsters." Lobster, you think of something else these days, but [laughter] Lobster, they

17:52Yeah. It's them and the Gremlins, I guess. Yeah. Yeah, but but then like they they took in like I think it was like an economics professor Hal Varian, like microeconomics professor, and and he was like, "You should really do this as a second price auction." Mhm. And what that means is that you can basically say, you know, these clicks are worth this much to me, like 10

18:16crowns or 100 crowns. I guess to give some more background. I mean, if you go to the normal Google page, you will see the organic results, like Yeah. the real web pages, but you also see ads, and the way then then you place your ads there is by setting some kind of price there. build a campaign in in a Google Ads

18:33interface. Yeah. You then add ad groups, and you then add your actual ads that will be displayed, the text, and also a set of keywords that will trigger like These days it's more complicated, but that's how it started pretty much. Yeah. And these keywords match different sets of queries. Mhm. Depending on settings, search terms,

18:52yeah. Mhm. And and then basically you also set what the price that you're willing to pay for this particular keyword. Your max price. Yeah, I mean, just like you would do on when when you do sell stuff on Tradera, basically. Mhm. But but the genius of the second price auction is then that you can say honestly what it's worth to you, but you pay the price of the next guy who hasn't

19:13the same value. Yeah. So so maybe the the the most valuable click is 10 crowns, but you are willing to pay 100 because you earn 100 per click. Then you're still just going to pay like 10.1 crowns per click. But if the other guy optimizes his business so he can pay 50, now you have to pay 50

19:31to be on top. And then on top of that, each landing page get graded in a quality score. So if you have a high quality score, you get cheaper clicks. If you have a low quality score, you will get screwed basically. Is it the same but much more sophisticated today? Because this is this started in the era of [clears throat] the keywords or search words. And now everything Yeah, it's fairly similar that I would say like more complicated. There are more like AI layers and optimization layers in between. Like that was one of the reasons that we sold the company

20:01eventually to 24/7.ai. That Google was basically removing some of the levers that we had used really successfully when building Kampanje and it felt like basically we were sitting in in a very dependent position versus Google. Yeah. And that window was kind of closing. So it felt like like we should probably figure out something else. Yeah, I mean yeah, Google is just removing levers to pull basically. And nowadays it's just basically all doing AI mass campaigns and then they do all the optimization for you. And of course they're in the business of selling you bad clicks and good clicks. But Kampanje was in the business of only finding the

20:39good clicks. Yeah. So yeah, they were the main optimizing the the way to set the bid. And I guess that's the hard thing with not knowing really what you will pay. So you could also try to game that if you do it properly and get a higher position basically but by doing so. Yeah, and that was I guess the core innovation of Kampanje that basically we realized that if we just re-evaluate this very very often, we can then pay the correct price at all times. And

21:05that's like high frequency bidding. And it turned out to be super profitable. I I think we beat Netflix own engineering team with like 30% better results multiple times, even though they went back with their own kind of build again, and Yeah. they couldn't win this time. They came back, and we beat them again. So, like it it was a super successful approach, like a very expensive to build. Like we had a big engineering org. We had like multiple people working pretty much

21:29full-time building APIs. But also very successful. And it was kind of hard to explain to customers. But you had to kind of go into this PhD Bayesian math, and they were like, "Oh, what the hell are you talking about?" But yeah. But I think that that was like not not in terms of technology maybe, but it was your most valuable contribution for company as a business, Anders, when you took this 3D chart of the value over

21:50time. Yeah, yeah. The love dot. Yeah, and then they finally got it. Tell us this war story that Anders did his creation. Anders' war story in a nutshell. I want to hear this. Yeah, but please. Yeah, but it was like this then if I should do it shortly. Yeah, do it. So, I had built this high-frequency bidding engine. But the problem was that like we were very naive when it came to what the value of clicks was. We could control the cost, but we didn't know the

22:13value. So, then we looked at the market, you know, what is the best talent to help us solve this. We found Anders. Uh an AI PhD already back then, which is I think it wasn't they weren't growing that fast. So, yeah. And and even when we found some more later, like not everybody were as good as you. So, so that was a fantastic find for us. And you helped us build uh basically AI models. What was called AI at the time. This is of course like always changing definition, but

22:43What was the underlying tech, Anders? Don't focus on [laughter] Yeah, but out of that model also came a graph that visualized the value over time. And that is something that the customers finally understood that, "Okay, by the same bid all the time, it's not following the change." And then And it's this beautiful, you know, it goes from like uh, red to green, and it's almost like a landscape. So, I could basically show Netflix this like, "Oh, when do you think your users are most likely to convert?" And then you could see this kind of 3D landscape where Sunday at 8:00 in the evening the the graph just

23:16jumped basically and they understood it. And there was a uh, I mean, a really innovative product, but it was also a very good way to explain what we do at the company. and the way how we communicate Exactly. Yeah, I know it's such a big leap. I mean, if you can't explain the fantastic product like you're not going to make

23:33much business anyway. My deck went from like, 1 hour to 1 slide basically. And [laughter] then it just said Netflix on the top and then You're exaggerating, but yeah. [laughter] I mean, good. But then, okay, you have expanded a bit. If you were to just perhaps explain the business model of Growth Hackers and and Cogni compared to Companiac, how would you

23:51differentiate it? Yes, I mean, if if Companiac was very deep in one channel, Google Ads, Growth Hackers was more taking another approach which is you need to do everything fairly right in each step of the entire user journey. So, from the channel into the site and all the way down to revenue in order to get a good

24:10funnel. And get kind of good growth going. So, [snorts] what Growth Hackers did was to help orchestrate this to get like the Google Ads campaigns or all the campaigns uh, trackable and testable and then get the user journey on the site, all that trackable and testable and then be able to run experiments in all these different parts of the user

24:32journey and do that as much as possible. So, the more experiments you run, the better growth you got basically. Mhm. Mhm. And we call that the growth process. And and we we talked about before at that point in time growth hacking, but this is now more or less established in around the people that know. So, we we call this growth engineering now. Or Yeah. Yeah. Or go-to-market people. It's a little bit different if you

24:58term I would say. it's kind of it's part of GTM. It's popular right now at least. Yeah. You know marketers, they go like It's like fashion. Passwords. But okay, so what's the difference between growth hackers and Cogni? Oh, but that was the fantastic thing like we had tried out AI in different forms throughout our years. Like I I I don't know if you remember when this like AI Dungeon game was launched on top of GPT-3. I mean that was the really first indication I think that that these new transformer models

25:30could deliver something special. When is this? This is before Chat GPT of course. Yeah, yeah, exactly. So that's like I think it's it's the summer of of '23 something like that. Or maybe it was Christmas '22 Christmas. Oh yeah, sorry sorry. Yeah, so like is it Christmas of '21 something like that? No, this is like deep fake, you know. [laughter]

25:51No. All right. So okay, so they shipped that one. I actually applied for the API, I tried it out a little bit, but like quality was too bad. Mhm. But yeah, no, sorry, that was of course 2021, yeah. And then as Chat GPT was launched November 2022, then it was like okay, now they managed to get it good enough and in a pretty

26:13short time window. So then we basically just went all in on Cogni and and I kind of we were in a pretty like hectic spot with Growth Hackers at the time, so we didn't have like time to work in on it on the day-to-day. But then as kind of Christmas break started, I I pretty much coded together

26:31an MVP and gave it to Tom for Christmas. I was very happy. I showed it to my whole family. Look at this fantastic Christmas present. Yeah, '22. '22. Oh, it was then? Yeah, Christmas '22, yeah. And then like early 2023 we shipped our first like SQL to like AI powered SQL generator copilot thing. I mean it was worthless because you more or less had to know SQL to be

26:56able to generate SQL. [laughter] Even though you write natural language, but it helped me at least. So, we had a user of one, basically. [laughter] Uh but then we kept building. Um and the idea was to basically take what we do by hand and do it with AI. Mhm. And I think it gradually kind of took over more and more of our work. I mean, in the end, we basically just got questions from clients. We fed it to Cogney, and Cogney created a nice

27:23analysis that we then presented. Mhm. And was mainly to generate the SQL itself, or did you do something more with visualization and stuff, or what's really the the core functionality of Cogney? Yeah, so it's evolved a lot over time, of course, with capabilities. Like we we have tried to trust AI a lot, and then we have been like burned. So, like the the first step was obviously like this AI co-pilot for your BigQuery data warehouse. Like you ask the question, and the AI delivers. And it's kind of still that in the core today, but it's

27:53obviously much more capable now. But then we tried to build like a if you when deep research was released, we tried to build it like a data analyst service based on the same principles for BigQuery. But like the analysis that came out weren't that good. It was hallucinating. I mean, if it's hallucinating a little bit when you're building a web page, maybe that's not you know, it won't destroy the the entire page. But if you do an analysis that shows, you know, you made 3 million instead of 300,000, that's a little bit

28:26bigger bigger problem. But but still like this was in the same era. Like we we have we know Anton from before. Like so, we we met him. You mean Anton in Lovable? Yeah, exactly. Anton was the guy, yeah. Yeah, so so we met him around that time. He was hacking on GPT engineer. Uh and Yeah, that's right. It's it's almost

28:42similar timeline. Yeah, and he was like, "You should try it for your growth business." And we like we generated some like dashboard pages and stuff but we're like nah, it's not good enough, you know. that. [laughter] Yeah, so so it was like at the time where models weren't super capable. Yeah. So then we kind of took a step back and instead built like more static reports where like the human could define the sequel queries like make sure there was

29:11no hallucination in the core data. And then do like AI analysis on top of that. So like what does the data actually tell me? Mhm. Um Because I mean if you present a dashboard to a client the next question is always going to be but like what does it mean? Mhm. And that's basically what you can use an LLM for to give you some

29:32you know, bite-size insights. Mhm. And then I don't I don't know how how far we should go. No, man. We we talked a little bit upstairs as well and I also a logical progression that okay, now we can produce all these reports but no one can read the reports. So now the next stage is on top as the But I think let's I mean pause it at a little bit because that was like a huge surprise to me. I mean I'm like think that everyone wants data, everyone wants insights. I don't care in which form it comes. But when we sat there and used Cognite to produce insights for us and

30:03we went there to present it everyone was like oh, fantastic. How did you find this? And then we were like okay, but you can get this every week for your specific role. Even everyone in your organization can get their own insights. You can go to the Monday morning meeting and everyone knows how the last week gone and you can get some recommendations. So we started sending emails but nobody opened the

30:24emails. Which is kind of interesting. Interesting, yeah. Because I mean insights should be insights in my world, but they're not. So you went basically to do some vibe analytics in some sense. Yeah, exactly. We actually And then like at the time that Lava Bell really started working I think and they had rebranded and everything, and that was, I think, the summer pretty much of '24, right? So, they they launched for the public Auto 24, but already in the summer they had like something that started to work much better, and that was obviously Sonnet

30:563.5. And then you had the hood, yeah. Yeah, exactly, exactly. Like, that that base capability then unlocked like a good enough user experience that could be used. We actually had Anton on this podcast right before he launched the LavaBell. So, he hinted a lot about what was coming. we we we we take the claim, if you want to have a really good rocket ship, go to

31:18us and see what happens 1 month later. I think it was only 1 month before he launched It was like 1 month before he did the Product Hunt launch and the rebrand. So, it was so fun. We take all credit for this. Congratulations. Oh, yeah. Yeah, yeah, yeah. After Sonnet was here, they were able to sell it, yeah.

31:36I see the trend. And he was much more open with how it actually works, you know. [laughter] See as well or No, okay. Anton outlier, yeah. You're in for a good ride now. Yeah, yeah, okay. That sounds good. [laughter] No, but but back to what was the point here with them you were making a point here at this point in time, Sonnet started working

31:54better. Yeah, and that kind of unlocked like real analysis again. Like, the the copilot we had built before all of the sudden started kind of working. So, then we went back, and and by the time we had kind of re-engineered the the analysis, even newer models were were out. And we were then like going away from this human-built templates to like it more of a agent spec. Like, you should do this,

32:22analyze this area. And and we were running that for some time, and then around, I think, Christmas we felt that, okay, this like agent-driven analysis, like the agents gets an area to analyze, but it doesn't get like a specific query to run or something like that. Mhm. Uh those started getting good enough quality that you can actually take

32:42action pretty much straight from those. And at the same time users weren't reading the reports anyway, even though like the quality got really, really good. So then we built this whole like uh ticket orchestration system with actual agent execution on top of the insights from the reports. What do you mean? So ticket means I mean basically if you go back to the example, like let's assume that everyone

33:03in the organization has insights. Mhm. What's the point of that? What are you going to do with it? Exactly. You're going to you know, create an experiment or act on it somehow. You want that recommendation to you to think about this type of AB testing for your your role in SEA depending on what role it is, it will look a little bit

33:22different what they should be doing. Yeah. So we then created this ticket view that we call it or and and that's another agent that then basically reads the insights that the insight agent created and then create actionable tasks. Like we recommend you should do this. We score it. We have the eye scoring where you basically take the most uh the highest impact uh ticket with the highest value and prioritize that first

33:45and present it in a ticket view. That goes then from That's cool. Yeah. And I think you mentioned also that I mean one approach would of course be to have one big agent or AI that does it all so to speak or you can split it up into different agents with different purposes. And I think you chose the latter, right? Or can you Yeah, I I

34:03Describe more. Yeah, it's So of course you can think of it in many ways, but like the tickets are really like uh a specification that the uh a ticket execution agent could then take and run. Mhm. So it's both kind of what we need to do, what we expect should be the outcome, and the why. And the why goes back to these uh reports that we had generated

34:25earlier. And it also has like how do we intend to do this? So it comes with a basically an ask for permission to you know, maybe publish a new page on the blog or maybe change budget on some campaigns or maybe create some new campaigns and send traffic to the new blog post. And if you approve that, the agent goes

34:45ahead and executes. Because I mean, we started with this ticket view. We were like, "Oh, this is cool." But then we started to assign ourselves. Okay. And that's not so cool. Yeah, exactly. [laughter] Because I mean, I'm not going to have an AI telling me what to do. And then the AI basically tracked what I was supposed to do. So, then we were like, [laughter] "Okay, let's just have the AI execute this as well." And then the whole MCP trend came. So, we started playing

35:08around with that. So, then so then because now you get into thinking about recommendation, but then actuation. Now we're getting into actuation space, right? But but keep to the topic here. I think it's important you know, thinking here from a more architectural point of view, should you have one agent to do it all or should you have separate more uh specific agents that have specific purposes or

35:29Can you go a bit more in depth? Like I I think it like we always have the context window to to think about of course. Like it's not infinite. It's 1 million now, but like that's easy to burn. So, if you if you specify like a sub agent's task really well, sub agent could execute it maybe even with a weaker, cheaper, faster model than than what you need to actually ideate it and and like make sure that

35:53this makes sense to do. I mean, this is essentially how human orgs are kind of supposed to work as well. load ultimately. Yeah, as well. And also of course, insight judgment, context like I mean, the CEO knows the direction that it's going and could take like, "We need to do this." Delegates it to somebody that has like an idea of of kind of how to do it within their specific area. And in the end, it it maybe lands in an expert's hands that only thinks about how to make that really intelligence in an enterprise is

36:22distributed. That's the point, right? Then no one can have the can understand from direction to coding decision. No. And like maybe AI could do that, but you also want parallel execution. You want like item potent task that could like You want to impose judgments on different layers as a example, right? So how do you if you have no way of going in and What do we? I mean that's I think it's a

36:45question, right? It's not obvious, I think. But I think for us it was more like as we build out the product, we needed to do it in steps and you to understand that it was doing it. It makes [laughter] sense. Yeah, exactly. I think so. And I think also today when we sit in demos and explain it to marketers or CEOs, I mean they they understand it directly because it mimics

37:05a human process, basically. Can you go a bit into the text stack or using like LangGraph or something or No, no, no, nothing like that. We're like we build in TypeScript. Mono repo pretty much. Okay. And and yeah, no, then it's our own agent harness that we that we Yeah, we haven't talked about I mean like that's that's the new fancy word, harness engineering. You built your own harness. What are you

37:31What are you talking about? Yeah. I think the harness is the product pretty much, but like what what people talk about is of course that like the the underlying models themselves have like great capabilities. On top of that, obviously like Anthropic built Claude code. Claude code is a set of prompts. It's a set of tools. And they have the added benefit that like all of their RL is probably done with the same set of tools and prompts. So like it becomes really, really strong as a companion to

37:59the model. And [snorts] that that tool is is made to, you know, fulfill coding tasks pretty much for maybe a single terminal session at a time. Now obviously they have way better all of that, but but that's roughly what it is, right? A set of primitives for the agent to use like from from the token generation like now I should emit the tool call, it

38:21looks like this. I expect this to come back, etc. I mean I I don't think people, you know, recognize sometimes the value or the USP that the harness or the scaffolding or the logic around the model actually do have. I mean I think Sam Altman once said something like um if your startup or your company doesn't get better as AI improves, then you're

38:45on the wrong track in some sense. Yeah. Yeah, but I I I really want Bernard to go deep on harness because I think this is one of the most biggest misconceptions by people who don't understand AI engineering that we haven't we have a we have one type of evaluation criteria for a very large broad language model. It should be general, right? And now we want to use that generality in a specific context, you know, for whatever you want to do

39:08optimize growth or whatever. And all of a sudden now if we are not putting a harness in relation to the eval criteria we have right now, it will hallucinate by our new standards, right? Uh and and I think this is so so so this is becoming super clear, I think, in Silicon Valley or, you know, we had we had the guy from Google here, Magnus,

39:30right? You know, who who's who was running the eval criteria for Gemini, right? As thinking about how the we put eval criteria or something like that is general. Yeah. Impossible, you know? So I think this fundamental is one thing also. I mean, one thing is to put guardrails on it and another is to put the logic on it, right?

39:47Yeah. What what one do you actually put so on? I mean, it is Yeah, I mean, we have our own set of tools, obviously like we have a loop that goes around the call that goes to, for example, Anthropic that costs us money, right? That that's what the harness is at the very core.

40:03Do you do you do Yeah. But I like you The loop is one thing then, and you have to have some kind of logic in you know, when you should stop or when you should do some other calls and H Yeah, of course. Like what prompts do we have? What tools do we have? What type of like, descriptions comes with the tools, are skills auto loading or not, like, what happens when the task is done, all of this is kind of orchestrated into the whole that is like our cognitive cloud offering. We're like as the task gets delivered, if it's a coding agent task,

40:34it becomes a pull request. If it's a, for example, advertising task, it becomes like analysis ticket in the end where like another AI takes over, looks at what the first AI did and tries to do an analysis. If this analysis comes to the conclusion that it's good or bad, we log that. If it comes to the conclusion that this needs to be analyzed in some time because like SEO takes time, campaign evaluation takes time because

41:02it wants statistical confidence, etc. Like all of this is the harness. But me me now who really I'm really trying to understand and learn this because I'm not an engineer myself. So are are we distinguishing between guardrails, harness, or scaffolding? Are we using them interchangeably? Because I think you you were highlighting, right? There there is more there is something to think that harness is about guardrails, but I think in this sense harness is more of the scaffolding, the

41:26logic. I mean I think about guardrails about it not up, basically. [laughter] Yeah, I mean the harness thing is way more. And I mean [clears throat] yeah, I would say so. I mean that's more about creating the value or distilling more value from a general LLM, but the guardrails is like you're not allowed to change bits and pieces. one thing, but a harness should be very tightly connected to the eval criteria for the objective function you're trying to solve, right? I mean like so if you So So when you say harness, it's like you should be doing this, but we also want to evaluate. Is it Is it becoming

41:59better? I don't know. Is that part of the harness as well? I mean we don't train our own models, so we don't have a like the objective function loop in our own LLM. We could do that, but it would probably be kind of expensive, and the question is like, can we keep up with the Frontier Labs? And like, I think today

42:16the answer is probably no. Like, at the same time, we're building quite interesting data set with like, these are marketing experiments that have resulted in growth or not growth, for example. like, where where do you put the memory? Stuff like this all Is that part That's our harness, yeah. That's your harness, right? like, we we build like a organizational

42:32memory for the the customer's account. We try to figure out like, what is driving growth for them, what is not, what should they do, what should they not do, like, what's the positioning in the market they should like, lean into or lean out of. Like, all of this is orchestrated within the harness. This is so much more I mean, like, so I like I love the definition AI compound system that Berkeley AI research lab, I don't know, they put in paper a couple of years back because it's it a highlights this there are different things that needs to work

43:01in in orchestration here. And it may be the harness way of thinking about this is one way to get more people to understand what this is all about. I don't know. What do you think? But one way to speak about this is also to use terms like Yeah, going from software engineering to agent engineering, meaning it's more about the engineering you're doing is really trying to orchestrate the agents properly, and then perhaps seeing how can you get them to collaborate properly to do what

43:28you want them to do in some way. Yeah, no, for sure. We're leaning into that like like mad men, actually. 100% yeah. [laughter] Yeah, perhaps that could be a term to use potentially. And and what we launched recently is something we called Yolo mode. Yolo mode. Yeah, and it looks like Yeah, because I mean Exactly. It was another like feeling when you sit there and you kind of just approve ticket, approve ticket, approve ticket, approve ticket. It's like, why do I need to sit here and just, you

43:54know, smash smash smash smash smash? Maybe it should just auto approve. So then it kind of Scary. Yeah, I saw this podcast recently of Andrej Karpathy, and he, you know, he doesn't like the term vibe coding even though he really invented that or coined that term. As you know, he said you know, in the beginning you know, since he's a programmer since very since very young age, he loved to actually program and he wants to see the code and he wants to

44:20review what the AI is doing. But it came a point very recently like December last year. Exactly, yeah. Where you know, he said you know, the AI does it right every time. There's no point for me to even look at the code. So I become what I don't want to a vibe coder. [laughter] Yeah, I mean I had that exact same moment with Cogni actually because I mean we worked with you know, 10 years training growth hackers and becoming these kind of uh innovative business people that know technology and marketing and they are supposed to come up with growth hacks or ideas how to grow a business and of

44:55course I've uh been pitched a lot of these ideas uh through the years and just in December actually. [laughter] Cogni started By chance. It's a which underlying model? Uh four or five Opus four five. Yeah, Opus four five moment. Yeah, exactly. Yeah. So we were sitting like looking at in the ticket board and we saw all these experiments that Cogni suggested to us and it was like, oh this is good. And that's that's good. Wait, wait, is that hallucination stuff? This is actually better than my growth hackers I used to train. So I mean that was our holy

45:28moment basically. Yeah. That's so cool. Nice, it's really cool. You mentioned by the way a bit about MCP as well and then I mean I think for the people that are a bit tech interested here it would be fun to see you know, how are you making use of MCP today? Yeah, MCP is is kind of like USB-C, right? It's just a standard to connect something and for for the longest time it it like did have some momentum but now I think it's really gathering momentum like at an

45:57insane level. Uh so we basically leaned into that. We build all our own connections to other platforms as MCPs. We federate those MCPs into our own MCP. Mhm. And what is MCP stand for and what and do we have and do we have alternatives? there are a few that are like semi-alternatives, I would say, like agent-to-agent protocol would be one, for example, proposed by Google, but like I would say pretty much no

46:23traction. So, MCP stands for? Model context protocol. Model context protocol. It's now owned by the Linux Foundation, but originally from Anthropic. So, Anthropic Linux Foundation, it becomes then the standard language connection interconnection. yeah. It's a standard protocol for connecting agent work. Yeah, the way that I explain it to business people is just that it's an API

46:45that an agent can talk to. Yeah, it's an API and a protocol specifically for agents. Yes. Yeah. And on top of that, it has like authentication primitives, for example, so like I could come to Stripe. I could say I'm a Stripe user and now I want to use your thing in my agent. Then I connect to their MCP server and I can then authenticate with Stripe using the MCP protocol. Like one thing that is notably missing is, of course,

47:15payments. Yeah. There is apparently like a request. Yeah, it's in the works, but it's not on the prioritized list of things that they will focus on this year. But and if I understand this like so I have never done MCP engineering, but I know API engineering really well. I've been using APIs all my life. Is it

47:34different? It is because the MCP is kind of a little bit more alive. It's more agentic from API point of view, if you will. So, the a typical API I would like expect you to give me maybe your Swagger spec or an OpenAPI definition or something like that. I'm going to go off as a happy developer and hand code it like now we're in the 2010s era. have have your documentation. I have all this, so now I I do my trust that you do exactly what you say, and then I can do exactly the same, and

48:02like we're all happy. Yeah. But model context protocol can like change endpoints depending on context, can provide example prompts. So they can be more dynamic with each other. Yeah, they can. Which I guess also less less demand on humans, so to speak. In an API, you normally have to have some kind of coding being done to use the

48:23API. Yeah, this is just plug and play, like the agent figures out in real time how to use it. You can just take a URL to to an MCP and add it to another agent, and it starts to make use of it without you having to do any specific coding to Yeah, exactly. And now there's even MCP apps. So like the MCP can then like inject UI that optionally

48:43could be rendered by the MCP client. So that's really cool, like I I mean that means that for example in Cogni, we have this concept of core metrics, like we we ask the user to, obviously with the help of AI, define what they really care about. And then we're going to optimize that, of course. Uh but it's much easier if we know what

49:02they measure. And then this can become like your core KPI card, and we can actually design a little UI component, so if you connect your cloud co-work to the Cogni MCP, and then you ask like, "Ah, how is my KPIs going?" We can even send like a user-friendly version of that rather than just like the JSON numbers or

49:22whatever. So when you start thinking layers on layers on layer, imagine what it what you could use for it. It's so much. Yeah. I mean, if you think about like I think that most people are in this kind of copy and paste mode with their AI, they're like going into this system, copy-pasting that, and then pasting it into the AI sitting, and they're like back and forth and back and forth. And I mean, the MCP just completely destroys that, and you don't have to do anything anymore. You can just you know, ask the AI to go and fetch your data in that system, crunch it the way you want it,

49:51and then send it off to that system. So, I like a concrete example. Like so yesterday there was this Y Combinator event. Obviously, like that's kind of our ICP. A lot of startup founders who want growth. We provide growth engineering as an AI. So, then I asked Cogni to launch an X campaign, X ads campaign targeting Stockholm, targeting people like Y

50:13Combinator profiles followers. And it then went to um basically Nano Banana using MCP, generated ads like we were ideating about like what why we should go with etc. And then it launched the campaigns on X. And it was like, you know, I was cooking food for the kids in the meantime. So, But we we we are getting to a very different world now with people who are getting it and using these tools versus the rest. This is the the gap is

50:42widening right now. Fast. And I mean, we can notice that when we sit in in demo meetings. I mean, we can like in 5 minutes in, we know what stage they're in basically. Yeah, and and how how do you how do you differentiate your selling or your communication with them when you realize someone is less mature and someone gets it because this is now getting

51:01This is Mars and Venus stuff. Yeah, exactly. I think it has to do a lot with trust because as this December moment, that's when you start to trust AI basically for coding. And I think every human needs to go through that phase basically. So, if we notice that they're not leading in that much to AI, then we focus more on the kind of analytical AI part and then they're oh, you probably run a couple of campaigns here and there and it's probably a hassle for you to you know, again, paste and copy all the data and create a report so that Cogni can help

51:31you out with that. really. Yeah, exactly. And it's like you can start with reporting stuff. Uh your team will get more insights and then when you're, you know, happy with that, you know, we can connect a couple of MCPs and then you can start executing basically. And how does it sound when someone is

51:46switched on? Like someone is on it. They just want to know more and more and more. They just connect every MCP that they have and like let's go. We show the reports and okay, so this I can really use this to execute. Yeah, wait. Let us come to the next point in the presentation. They hook up their GitHub account and

52:04yeah. Actually call this I mean if we have agentic development or agentic coding, is this basically agentic marketing then or what would be a good term? code it as coding as well. So I mean yeah, I would call it like data-driven marketing and coding. In this The same thing is it's coding as well, right? Because as you are coming up with an execution, you build just in time

52:26what you need to build. Yeah. Yeah, right? obviously subject to human approval. We don't have an auto PR merger, but And yet. [laughter] I mean for sure in many cases like the so the solution could be partially changing the code and also partially doing actuation in different marketing systems or sending a newsletter or like But because this is also a little bit I asked you before, are you for product teams or are you for marketing teams or you right on the intersect? I think you I mean like you're right on the

52:58intersect, right? Yeah, yeah. Yeah, and I think like one of the core IC pieces really CEOs. Like so somebody said that yeah, like I just hired a AI transformation guy blah blah and then I think it was Garry Tan on on Y Combinator it was like the CEO should really be the AI transformation guy like And I think it's true like it's in the end like it all boils down to that. Like it's it's where the the main agency of the company comes from or should come from and So you saw I understand. So when you say this is an e-CEO play is like you really

53:31need to decide where you want to put autonomy agency and where you want to put people, hire guys and where you want to put in agents now. Yeah, but it's also like That's what you mean, right? No, but it's also like Cognite does all these roles for you. So now you don't necessarily need specialists. You can just, you know, That's becomes Yeah, it become it becomes the it becomes the step to the

53:52one-man unicorn. And that's why it's a CEO question. And it could be like I mean should the marketing department decide your pricing strategy? But maybe your pricing strategy decides your marketing success. And this kind of converges usually on the CEO or that's sort of the Yeah. Yeah, exactly. The interface. I'm thinking if we should move into I mean you're both experts in search engine optimization as well and in search engine marketing, but I guess nowadays, you know, you have AI agents answering questions more than search

54:26engines. Yeah. Do you have any thoughts about that? Is this changing how you do the work if you want to get noticed these days in terms of Definitely. I mean it's It's all I think about, you know, how can I present [laughter] myself in the right context. How can we be found in the agent space? Exactly and like who will use this data

54:48that I enter here how? Yeah. Exactly. So what What's the answer? What's the answer? Yeah, but I mean [laughter] Yeah, of course. I mean you can go all all all different directions with this, but I think where we are right now in the market, you kind of just have to make your content readable for LLMs and structure it in a way so that So let's break that down. Step one, getting out of the search engine into the agentic mode. So ChatGPT or anyone can find you, you need to be readable, you need to be structured, you need to be code coded in a certain way. Could you

55:19just elaborate on that? Yeah, I mean I mean they like bullets, they like tables, they like things like that. So it's it's not And they they also like all of that to be server rendered. There's There's a couple of things I learned about this like server rendered. Sorry, I'm getting excited. It's exciting. Go on. Go on. [laughter]

55:41Sorry. Yeah. This is fun. Yeah. Sorry. Yeah, no, but I mean I the whole internet is being rebuilt as we speak. Yeah. Yeah. Yeah. Yeah. And and at some point like somebody thought it was a good idea that the client computer should be big and should do a lot of computation. That's why like

55:55my Chrome takes 60 GB of RAM. All right. Yeah. And that means the browser is a like a fully built super complex application where you can find like a lot of security holes etc. Yeah. And and it also means that to consume this content you need to execute JavaScript, you need to have a lot of

56:13RAM as well. But most agents are not full Chrome browsers. It's not a very like computer program friendly interface. That's why it's better to have like server-side rendered so that you take back this computer that you have offloaded to your customers and you render the page from day one, right? And they can just download it as text and then parse out the information they need. On top of that, there are ways to structure the information with for example structured data standards like this LD JSON format that makes it easier for agents

56:47to consume the data on the page. So we're getting back to how you structure your web page and you back to good old HTML rather than all these scripts and everything else. It could be that, it could be like markdown versions being served. That's another track that people are taking like obviously with an HTTP request you could specify which format you prefer to get the response in and like you have the option to then if they say they want it in markdown actually give them markdown rather than all of these extra worthless HTML tokens that will make it look pretty for a human but it will make it

57:20like look messy for an agent. But I mean these tactics aren't bad for SEO either. So I mean you you you could do both at the same time because LLM googles as well. So I mean Yeah. it's not bad to SEO you You say Google but you mean search engine and and if you do use an yeah, chat chat GPT or whatnot, I mean, they do still use the normal kind of search engine

57:43even you don't always see it. Yeah. But then is there I mean that there are similarities then as you say that so it's probably good to to optimize for both. It wouldn't hurt you to do that but I'm not I don't know the answer but but my is there a way you know, besides the format that you put on the page how should you be recognized for when an agent do search, when an agent do web

58:06search? Yeah, they Google in a very different way or they search in a very different way. They use more words. They they formulate themselves differently like they use they do this query find out is what they call it. So like you the user have one question but they will do like 10 to 100 different web searches

58:23and and visit a bunch of pages. Different it's like a different prompts or questions they ask. Yeah, exactly. Small aspects of this what you're trying to find the best apple tree to plant this time of year. They do five. They do 100 versions of that question literally. Yeah, slight variations and if you can start figuring out what these variations are like commonly, you can then start

58:44optimizing your content for that. Mhm. And there's a quite interesting metric like if we want to talk hard numbers here because most of the search providers aren't sharing data around this but Microsoft Bing is doing that. They have a tool called Bing Webmaster Tools. You prove that you own the webpage and then basically you will get impression statistics, click statistics from their traditional search and then they also launched an AI

59:11citation statistics page. Oh. Yeah. First That's the that's the only official because all the other providers you see out there, they basically come up with probable prompts and then they run that prompt through all the LLMs and then they read the results. Yeah, and then they also pollute their own measurement. [laughter] Oh, you will appear more and more in the

59:32searches related to what you monitor. But this is so interesting. How many How do you look at your the industry that you're coming from the with the CEO people? Because a lot of the marketing people and CEO people were not really so deep coders. Are they Are they keeping track Can they Can they follow into this Will we learn this or is this a different type of agency you

59:53need to have get help from? Yeah, I mean I I think that has historically been what why you go to growth hackers. Because you need all that technology to help you, and it's not necessarily you going to see that on billboards or anything like that, but you get solid data you can trust, and you get like an infrastructure that you can run experiments in and things like that. And that's actually the the problem we're trying to solve with

1:00:15Cogni. Exactly. We remove all that scaffolding basically that you need You You need to have GA4 accounts, you need to have GTM set up, you need to have a good structure of your homepage, you need to have, you know, all this. Now you can just get recommendations from Cogni, and Cogni will take care of all that

1:00:32scaffolding for you basically. Cool. And um Yes. It's time for AI news, brought to you by AI 8-W podcast. So we usually take a small break in the middle of the podcast and just talk about the recent most exciting AI news that we can think of. Could we perhaps ask Tom and Bernard, do you have any news stories that you read recently that you'd like to share

1:01:03something about? I mean, what I what I read this morning when I woke up, it's always something, right? Yeah. [laughter] Yeah, it is always something. But But it was that Stripe had launched a link CLI. Right. And And obviously like CLIs are this like third leg of MCP alternatives, not A2A, but like the other way to do it is just that the ship is CLI that can do everything that you want the API to And you're very technical, but I think you know, CLI it's a terminal tool

1:01:30basically. Yeah, command line interface. [laughter]

1:01:43[laughter] And and why is Link CLI then so cool? Yeah. It basically is a command line tool where we now can create payments that the user then can approve. And Link is like Stripe's pay with one-click solution. So basically now I can give my agent the possibility to spend money. But this is sounds super dangerous, right? If you give open claw then access to Link CLI, then they can make payments

1:02:11for you, right? Yeah, I think there's still a human approval element, but the whole like commerce layer of it is unlocked and that's commercial infrastructure is there now. Yeah, I mean this is This is an attempt to let it started. So I think that now I can like programmatically tell an agent that like it's a payment you need to make, you need to input this into your CLI, and then it will prompt a human to approve

1:02:34it. So like I think the example from the CEO of Stripe was that like I asked you know, if it was open claw to buy something for yourself, and then he had like a screenshot of the payment approval prompt that he got. And I think that like this whole agentic commerce area is of course super interesting both from like just a shopping perspective with like B2C products, but also like from a SaaS perspective. Like how can we enable

1:03:03agents to buy our services? But the it it it fundamentally we we talked about it, the behavior around commerce is going to get to into into a new arena eventually. Definitely. Different world. And especially in marketing because we are usually have to, you know, buy advertising spaces here and there and if you can get that authentic then that would open up a lot of media space for a lot smaller

1:03:27Yes. If we we take Peter, you know, who created Open Clo. He actually prefers CLIs to MCP, by the way. Yeah, he did that MCP now though. So, he partially reversed that take, I think. But yeah, I know. It's it's very common like we we think about building a cognitive CLI as well. It's actually already being built. So, So, what's the difference? I don't

1:03:47understand. Yeah, MCP based or CLI based? What is the distinction? So, MCP is basically an HTTP protocol at the base. Like it's HTTP RCP, remote procedure calls. Like it's a very defined way to build an HTTP API Mhm. that then the agents can follow. But the CLI, on the other hand, it could be anything. It's just a command that you can run and then it will print text

1:04:17and hopefully do something as well. So, it's even more open, more flexible in that sense. I mean, at the same time you need like a command line interface like your own computer to run it basically. Install the command, etc. But for Open Clo, it's easier because it's it has a terminal. It can easily just execute the CLI. Yeah, it's a bit

1:04:37more flexible, I guess, in that sense. But Yeah, and a lot of services don't have MCPs. Mhm. Um and you can like if you have a coding agent, you can produce a CLI maybe more easily than an MCP. Mhm. So, that's why it's very good for Open Clo that it can basically tell the agent come And your your line of thinking is to

1:04:56have both. Yeah, so we're going to build a CLI that then calls the MCP. [laughter] But then we're serving both niches. Yeah, yeah. I good story. Good one. Do you have anything specific? Well, I do share this kind of geopolitical interest that you have on the show. I mean, I like the story about

1:05:15the the the Meta Monus deal. Mhm. Yes. And how China just basically reversed that. Yeah. Which I think is kind of hard to do because those people at Monus are now working in Meta. Uh and And some more background for Yeah, what is Monus? What is What's this the background here? Yes, I mean Monus is a

1:05:37Chinese model or they started in China. They moved to Hong Kong and then to Singapore to kind of wash their Chinese almost like a pre-open core version of It's more of a harness than a model, I would say. You could plug in your multiple models into to Monus, I think. But it's it's for sure like it I guess it means hands. So it's like a closed hands. It's a similar idea. But more cloud hosted than than you run it on

1:05:59your own computer. And what was the fundamental pushback? Because China wants more AI power, basically. I guess that they I mean, Meta, of course, desperately needed some type of AI hit. He Zuckerberg has tried a lot and He's just buying stuff. Exactly. So he wanted to buy it and got pushed back that not He bought it. It went through. They signed everything. You know, the people went in. They even integrated Monus into the Meta advertising platform. And apparently, you know, the the revenue for Meta jumped like 30% for a quarter when they did that because the Meta interface for us marketers is extremely complicated. So having a

1:06:37conversational agent to talk to is, of course, fantastic. And then China kind of went in and said, "Hey, this can't be happening." it's interesting how more and more of the trade or the collaboration is being blocked, you know, over the geopolitics. And normally, it's US that block access for China to get access to Nvidia chips

1:06:57and and other technologies. Like the TikTok. Yes. Yeah. The TikTok. Now it's reversed. Yeah. And now China is blocking US instead. So, um yeah. So it's it's getting I guess more and more polarized in in the US. Yeah. And back to what we talked about before, like this is being such a big change for humanity that it's not

1:07:17something you can really just leave. They have to pick the fight basically. Yeah. I think there's a there's there's another interesting twist here in this like who is leading the AI race story, who is blocking who. Now it seems like this is one other indication from that angle that they're doing great stuff in China, they're doing great stuff in US, so so it's not like one is always the top dog blocking the other ones. No, it's much

1:07:44more similar. Yeah. Now and if you factor in robotics of course, that's where we are. Now it gets interesting. Yeah. Conversations are happening. And of course power. I mean China has a lot more power than US, so I mean Yeah. electric power Yeah, exactly. Anyway, super interesting future for sure. Yeah, thanks for that one. geopolitics plays out in coming years and we'll certainly be in

1:08:08What about Yiannis? Well, this is not geopolitics but is more US politics perhaps but I think the acquisition or offer to acquire Cursor which SpaceX did was interesting. Yeah. So you can you know SpaceX they recently acquired XAI. It's part of one yeah a set of these Elon companies and and XAI had Grok as you know one

1:08:32frontier model. It's not been going super well for Grok and this I guess not having had the progress they were hoping and he fired a lot of people from XAI recently and could be because of the acquisition in SpaceX. But still he said that it needs to be rebuilt from the ground up or something like that. That is a strong indicator that

1:08:51something is not right. So so one of Elon's companies is buying one of his other companies. Yeah, no, no, but that was before, so SpaceX already acquired XAI. Uh a month ago or something. Yeah. But now in AI time. [laughter] But now then a week ago then SpaceX put an offer up for Cursor and Cursor is you know very famous coding environment and very popular in the enterprise business similar to

1:09:23Claude Claude code. And but I mean they they are the ones that define this tap tap tap paradigm pretty much like I guess it wasn't in GitHub co-pilot before but they made it like widely popular and has grown to a bigger size than lovable and like Yeah. Yeah. And I think we had this a lot of coding

1:09:39users. On the show what's your favorite combo has been the question a couple of times. Oh, it's a cursor. Yeah. Yeah, I know cursor Cursor and that that is the answer for many people as the main combo, right? And you can think about why he's doing this. One of course could be that Grok is not progressing at the scale that he wants to and Cursor actually are building their own models as well composer and And then some are rumoring that Anthropic is you know fighting Elon and they cut Elon or Tesla and SpaceX off from Claude and potentially they have to have their own coding agent and then

1:10:17Grok is not you know progressing in coding as as as I hoped. So, this could be a way for that them to simply accelerate a lot. And it's also of course a treasure trove of data around coding successful coding outcomes etc. that is like he he can't build that by himself because nobody uses Grok.

1:10:34Grok coding business. Yeah, you could try to simulate it but it won't be the same as like coding in the wild and This this makes sense. He knows the benefit of data in his cars. You know, driving data in his cars. So, now I want driving data in the coding I mean getting access to all the enterprise users that they have but also the data they have provided throughout the years of course will help them to build and then on the other side XAI have the big Colossus data center, which is the biggest one in the world more or

1:11:01less. Um so then Cursor can use that to continue to build their models even faster potentially. But Cursor was separate, right? Who was the founders of Cursor? Wasn't it even a Swedish guy in there? Yeah, one of the four was Swedish. But he had actually left Cursor now, so he's not part of it anymore. But he was But have they been able to stay independent the whole way through until now, Cursor? Or have who who may who is

1:11:25it buying Cursor from? I don't know this. Yeah, I mean they they had I think three or four founding rounds or something. They are still VC funded. Yeah, yeah. But I think the the interesting part here is also that SpaceX, you know, is rumored of course to do the IPO very soon for the small amount of like $2

1:11:40trillion or something. Which is insane. I mean it's it's the IPO wars now with OpenAI and Anthropic and and and SpaceX. And then they're saying that you don't want to be number two or three because the one that goes first is basically going to take all the money that's exist in the world basically.

1:11:57I was I was what are you talking about? What is that? ever. So it it it's insane how big it can be and then given that it's going to happen rather shortly and that I'm sure they will doing all the filing preparation now for the IPO, then it's super hard for them to actually do this acquisition. So they are actually not acquiring Cursor right now, but they are giving an offer. I think they actually got the right to buy Cursor,

1:12:20but they're not going to do it now. Mhm. Yeah, exactly. And I mean I think he is kind Elon uniquely positioned to get value from buying it. So like for most companies it would be pretty expensive to buy Cursor today for 60 billion and like they are under massive threat to both from obviously Anthropic in closed code and then like more

1:12:40recently like OpenAI GPT. He's a logical buyer. Yeah, but like he is has the scale on the data center side or is planning to build it. So like for him it's super valuable. But does it you said something uh um on the IPO. I just want to make sure because I thought that was pretty cool. You mean the one is not announcing it, but the the one executing it. You want to be first execute on this super

1:13:06massive IPO. Is that what you meant? Yeah. Because then the money dries up. Yeah, basically. I mean, I listen to these financial analysts and they're like, you know, so many people wants to go into these type of IPOs. Uh and if you do them like it's like a movie premiere. If you have Avatar going up with another big movie at the same weekend, then people can't go to two

1:13:28movies at the same time. But it's also even worse because of obviously all of this cash is not sitting as cash. It's sitting as assets in other companies and like they're going to have to pull it from somewhere. And they can't pull infinite to put everywhere. So But imagine the capital they will have now from this IPO and then they're going to build Tera

1:13:46Fab as we spoke about in the beginning. Yeah. Yeah, I think that's the way the reason why he's doing it. I mean, Elon has never had problem raising cash, but it takes time. He has to go around, meet the investors and and now he can just basically use the the exchange for that. So I think it's with his plan, he needs this hose of

1:14:05cash. Yeah, and I think potentially the IPO could go up because of this because it will create some kind of because Grok is not progressing perhaps at the same pace he's hoping for and but with this acquisition locked in so to speak, the price in the IPO could potentially go up. Yeah, and I think also like 90% of the

1:14:24value of SpaceX is actually Starlink. Mhm. Right. So I mean, that's what people are investing in and I think Starlink is a hugely fantastic business model. So Yeah. Yeah. Awesome. Good news. I didn't think we're going to have any [laughter] I don't have a mic. But what do you think should we talk about the nerfing of Claude?

1:14:48Uh no, please. Tell me. Yeah, so I guess the the vibe or the the rumor or what people are talking about is of course that this new like Claude Opus 4.7 is way worse than 4.6. Way worse than Yeah, it's it's like they and if you look at like the the benchmarks that they published, like one of the things where they did some progress was in

1:15:11safety. Mhm. And and this is of course related to the meat those thing as well. I think that they started to think more about like the potential cyber security offensive capabilities of the models. But when when they've done like maybe that safety improvement or if it's just like running it more quantized on like

1:15:32yeah, cheaper, faster way to run it. Basically, you decrease the number of bits for each parameter. Mhm. Then it has resulted in the model not being as good as it was a month ago. Mhm. So I think that the vibe like you see a lot of people saying they're moving to Codex and GPT 5.5.

1:15:49Mhm. I mean the word on the street is if you if you look at the data center investment they last. They have this deal with AWS, but that's going to take Mhm. And I don't know a year or two. So this is the problem like they've gotten too successful for their own compute. Mhm.

1:16:03Oof. Yeah. Hard. Yeah, and [laughter] I mean I I think there was this really interesting interview with Dario like a few months back where he was describing this dilemma where like if you overestimate your amount of customers, you have to basically invest so much that you're going to go bankrupt. [laughter] And if you underestimate like uh then you won't have any compute to train

1:16:28the next model. Mhm. And now they're trying to thread that needle. So like giving less to inference, they still need to keep training the next model, but users are noticing because there's just so many of them now. Wow. It's a fight for the compute, it seems like. What was it? The um It was Who was it? Jensen Huang, I think, in Nvidia CEO said the five layer

1:16:49cake. It was the energy, chips, compute infrastructure, model, and application, I think. Right? And if you lack or have a bottleneck in any of them, you will be, you know, or yeah, screwed more or less. Yeah. It seems a lot are bottlenecked on the compute layer right now. Yeah, and will continue to be, I think. I mean, Elon Musk has said that he thinks that in 36 months the cheapest

1:17:15place to build compute will be space. Yes. And I mean, he's Space is in the center. Yeah. A lot of A lot of like linear curves on log log plots point to that, right? But it was energy then. But then, being in space, you basically have infinite energy energy from the sun as well, so

1:17:32And you don't have to cool it. Yeah. Well, yes, I think you do, by the way. Do you? Do you have space? No, it's it's super hard to cool I think in space. Is it? I thought that you know, space is cool. No, you you can't transfer the heat out. It's kind of tough, actually.

1:17:46Okay. That's interesting. has some interesting techniques from all the Starlink satellites he already has. So, he knows how to Yeah, exactly. But it's not that easy, apparently. Okay. But I heard somewhere that he needs to also send like a rocket every 5 minutes to get them [laughter] up, which is insane. But sometimes I just wonder, okay, IPO this terror of that, is it just or is is it any sense in this or is it just blown out of proportion and we're going

1:18:16to see something different in reality? But what do you think is driving him if not like To me, it looks like his decisions are are mainly based in this like almost existential dread that like we're going to somehow it up Yeah, I think maybe he as a person is driven by those kind of

1:18:36motivations more than normal people. He's taking on the mission to make life multi-planetary. Like if his main motivation was to be able to drink like pina coladas every day, but he could have stopped after the paper. Okay, so let me rephrase the question. I have no doubt he is actually completely sincere in the intention and what he wants to do, but is there a is there a feasibility in it or is it something you can aim for and have that as your hardcore goal, but shoot for the stars, but but and you will reach the treetops. What's the

1:19:09reality here or is this the reality? That's my question more. So, I'm not I'm not questioning his in being insincere. I'm just questioning where will it land? I mean, these tech CEOs are probably I don't know, not the best people on the planet. [laughter] But I they do see farther into the future than we do. I definitely think that and I think that the whole Claude Nerfing is a tell that people are using AI a lot. I think that I read another analysis that the like the revenue coming from Anthropic, they can't just be coming from like IT budgets. They're like eating into HR

1:19:46budgets now. Um if you you know, you can't grow that fast basically. So, I think we're it's moving faster than we think and I you know, we all lived through the IT area or the internet area and everyone was like, oh, it's going to flying cars and everything and then you know, that didn't really happen. So, maybe it's kind of that situation or they see

1:20:06something that we don't. I mean, it's a good old saying like you always underestimate the short-term impact, but underestimate the long-term. Yeah. Yeah, maybe that simple, right? I was just thinking exactly like this like we were over overestimating how fast the internet would grow, but in reality 15-20 years later everything and more happened,

1:20:26right? Exactly. Except if you're Ray Kurzweil and then it's like his own timetable. Yeah, exactly. He's nailing it by the way. [laughter] Cool. 29, right? What? 29. 29, yeah, I suppose yeah. AGI. I think you mentioned the banner and something something about, you know, code bases may be replaced by CMS platforms or something in in the future. Can you elaborate what you meant with

1:20:53that? Yeah, I I think that's the other way around. So basically like CMS content management systems was basically built as a way to to enable non-coders to build out content on websites, to manage websites. And it helped humans back then, but now it's actually in the way because they add so much extra scaffolding and limitations where like basically AI natively speaks code, maybe even more

1:21:21than language even. And if you take like the the SEO and geo perspective on that as well. When Google is supposed to read a site and it's a WordPress site, it has a lot of other code that's not maybe relevant for the specific site. If you go like pure code, then there is no code kind of unnecessary there. So it becomes faster, it becomes just easier to read, you rank better, it's just basically So what you're saying in in layman's terms, we used CMS systems like WordPress or others in order to basically improve our speed of

1:21:55building websites. And when we Or yeah, I would say democratize it. We democratized building websites to designers or whatever you want to call it. And now in reality, we can democratize in another way through through coding assistance and we build real websites in code that way instead. And democratize it even more because like all CMSs have like limits in terms

1:22:16of like what you can do. But what is then the ultimate if now what if you let's let's follow that trajectory. I love that idea. Where would you go as a designer con code now then? Would you continue with WordPress or would you go into Laravel or how would you go about or you cursor and

1:22:34on it? I think code design is pretty good like you could you could do that. Or you can continue working with Figma which is you know and and you can just have a MCP with cloud code and you know. So the logic is Figma maybe and then with Figma into how? Yeah and you kind of cloud code connect Figma and then you can uh you know publish your sites because I mean when you ask cloud to generate a site

1:22:57you get one of their their templates. Yeah yeah. Uh and and you kind of have to push it rather hard to do something unique. So that's what Figma is still strong at. Exactly yeah. But that's at the moment because like Figma is still also kind of like a CMS for design right? It has a framework for like what you should put in set uh schemas and like they have built this whole product with with uh some ideas in mind how it should work. And obviously code is more expressive than that and like with the the coding capabilities just scaling like cloud or some other model will win with

1:23:33code in the end. Yes but this is vastly interesting how this plays out now because I I I've been in onto this conversation even for my own company right? Yeah. Where we are not growing we are very careful of how and what type of competences we are growing and where we are building stuff right now. And the the core logic that we sort of installed was we want we want to follow

1:23:56the trajectory of the core coding tools. So okay. So where is our content management right now? Are we are thinking about GitHub right? Yeah. You see what I mean right? So we we are building our and okay everything we write let's write it in markdown. Yeah. Immediately. So I'm writing in my daily blah blah blah in markdown and then I'm taking it different things and this is what I'm

1:24:19this is exactly the point. Should we go WordPress should we go HubSpot, or no, no, no, we should go somewhere else? Yeah. Very, very interesting for the very big portion of the market. Yeah, I think I mean I read the statistics that WordPress is like 43% of all the sites globally. Yeah, this is the statistic I'm thinking

1:24:37about, right? Yeah. But that also makes makes it a target. I mean, we had a customer that got hacked. Yeah. And their main page got basically redirected to a gambling site. Ah. Which meant that they got thrown out of SEO and Google and you know, so being on a platform like that is also a risk.

1:24:57Mhm. Whereas if you have your own code, you can do. But what is your bet? I mean, like you're going to advise me. I'm going to buy your consultancy for my for Datax, you know, how are we going to think about this? So we're going to go document, you know, the same way as we do code, we do text. No problem. MD,

1:25:12blah, blah, blah. And then we're going to go uh no, HubSpot was the first one that went to. I want to go more open source. I want to do something which is fundamentally open source coding frameworks and build my own code. I mean, I React, you know, what what am I doing? Yeah, I think Next is a pretty strong candidate, which is of course like a

1:25:29React framework. Yeah. You could also do like Astro is another that has like excellent track record in terms of page speed, so you could do that. But this is this is going the code route rather than the WordPress route. of course, 100% This is the Mono repo is also the way to go, of course, because if you have multiple repositories, it's much harder to coordinate your coding

1:25:50agents across them. Like, it's possible. I do it for like one particular repo, but we have only two pretty much. Uh so, otherwise, you want all the code and all the context in the same place. Like, where your context makes the most sense. This is the whole point. This is why everything is in GitHub. Yeah, and we talked to, you know, VOs or customers that basically said like, "Oh, I was on a rock concert on the way home from the train. I just sat and live

1:26:12coded a replica of my site. Put it on GitHub, and now it's performing way better than my CMS. But if we go a bit more philosophical here and and say that okay, AI is allowing us to remove some of the abstraction levels we had in the past. Like a CMS system is something that non-coders can use to then build stuff

1:26:32and Figma is as well. And then you can move down the abstraction level and say we can work on the coding level for non-coders directly. But then if we think even further here and going a bit extreme, but I'm quoting Elon and I'm actually quoting Andrej Andrej Karpathy as well. Saying he tried to build this um some kind of menu app or just for for for sake of experimentation, but then you can think okay, before you you you had some kind of menu you wanted to have, you could do that in a CMS. Then you can switch to code, you can do it very easily. But

1:27:07potentially you can move down and say, I just want to use Nano Banana or an image generator to to do the actual menu and I don't need the code in between to do so. Because the code in itself is just you know, rendering some kind of image anyway. So why not cut the code out as

1:27:22well? Yeah. And then Elon has said you know, sometimes you know, he believes you know, the future thing with apps and having a mobile phone and that you may not need apps at all and the only thing you have really is an AI and some kind of visual interface. Where the pixels are being generated directly by the AI. And you don't have these kind of steps in between with some machine code or some Python code or

1:27:46TypeScript. Or some kind of CMS system. You just have the AI generating the pixels like Nano Banana style. Mhm. And that's the way interfaces will work in the future. What what do you think? Is that too extreme or? I mean, it's probably an interface for humans. Yeah. But then you have of course agent to agent interface which if you go in that far into the future then I mean, I see a homepage today like uh you know, a rendered page basically, but in the future you would probably have some part that is code, which is like open APIs that the agents can use, and some part like

1:28:20you're describing images or um, you know, custom instructions. could talk in latent space, right? Or they they don't need pixels, right? I guess. Exactly. That's it just it would just burn too many tokens, I guess. Well, why would you need exactly, right? Yeah. But on the other hand, I mean, I in I think in this example, which is

1:28:37interesting, right? The fact that the model is supposed to take this menu and then visualize like his app was supposed to do is probably also some prompting or memory, right? Because he would need to ask for it the first time or if it happens without asking and he says like, "I don't do that again. I want to actually read like with a through my AR glasses or whatever device he's he's using, right?" Then it

1:29:02needs to remember that and don't do it. And then this like instruction do or don't is also kind of like code or a prompt, right? I think we should at least I mean, it's the human kind of sensors. Yeah, exactly. So, the question is like like do you want it to be like but you're going to store it, right? As some sort of context in this like whole memory for this multimodal LLM or something like that. But so then the question is like do you want it to be

1:29:26probabilistic or deterministic? If you want it to be deterministic, then you probably want code because it can be repeatedly, verifiably done again, hopefully. But if you want it to be kind of more probabilistic, more intention level rather than like exact specifics, then you can do it more as a prompt. Can you give an example when you want it

1:29:47to be deterministic? Yes, absolutely, absolutely. So, I have this project at home. Okay. I'm rebuilding a Volvo up. I know it's going to be a bit difficult menu. So, okay, okay, but please. Yeah, okay. So, I have a Volvo up that I'm rebuilding to an electric car. So, I'm going to swap the petrol engine to electric. And to do this I need a coupler between like the gearbox of the old car and the splines

1:30:09of the shaft of the new electric engine. Right. And this is a fairly custom part like there are one guy that I know that has done this before on the whole internet pretty much. [laughter] With Volvo valve specifically. Volvo valve specifically. So basically then I need like uh somebody to cut me these splines for the gearbox because the the motor choice is

1:30:30kind of common. Okay. And to to test this then I basically had Claude like I took some pictures of the splines, gave them to Claude, and had Claude produce code that will then 3D print to these uh splines. And then I can test it. But now that I tested it, I want the exact same outcome again when I go to the machine shop and cut this out in the metal. I don't want a slight variation like the first attempt actually like Claude made the the splines inverted

1:30:59shape. Mhm. So like they were a circle in the middle, they were wider at the wrong end. So after you tested it, you want it to be the same all the time. So I love that this is uh deterministic. But this has to be with a model. Yeah, but if if I just like told the same thing again, it might make a a different mistake the next time. But now I want it reproducible because I tried it in plastic and next time just to memorize something rather than actually generate something? I mean But then it would memorize the the code you like this is like it's a Python code

1:31:30that outputs a 3D model. Yeah. So should it memorize that specific code? But then it stores code in the memory so it's like Maybe you could just memorize the exact like 3D model that you have produced from it, right? So you don't need a code at all. Yeah, yeah, potentially. But then like that's also like uh I suppose uh I actually had these kind of similar discussions in the past and you know, it it's actually I think it's super hard to really motivate really really strong

1:31:58determinism here. But I could be wrong. But I I it's a deeper question here because you are speaking on this on such a high level or nerdy level and when and when you hear the fundamental media or someone who's not into the details also ah how can we have stochastic systems ever do reproducible tasks? And I say we well you are not calling LLM the system I hope. I hope you have a harness. I hope you have a compound system, don't you? What do you mean? I like oh my god. I mean like so so we're literally trying to use math that is not done for reproducible

1:32:34things. So it's so what what I'm hearing is like how we look at the whole to where where do we want the how do we want to use it and then when then then this is actually memory or from memory. So then you worked around the deterministic problem, right? So this is what your angle is that yes, we need reproducibility and exactness, but we can get it to a different in a different way or what is your angle on

1:32:56this? Well, some people argue from a security point of view super important to have deterministic solutions that you can also have transparency in for example and also that reproduce the same thing. And verifiability in my case, right? Yes. Uh but in a verifiability I think you can solve by by having good memory instead. Yeah. Yeah, potentially if you memorize it in the same form, but then maybe they like the most dense way to save this is actually the

1:33:26program that generates the 3D model. Mhm. Mhm, could be. Could be. But yeah, maybe maybe not. But anyway I I think it's an interesting question. I think so. [laughter] No, but it's it's unsolved. I think you know so many people are stuck in the idea that you need to understand code to trust anything and that is the problem

1:33:45I'm having. So if you simply believe that just because you have the Python code available somewhere in some repo, then it's trustworthy. I don't buy that argument. Mhm. Because code you can't really trust if you have a a million lines of code, how can you trust that? You can't as a human have a full

1:34:03overview of everything happening there. And especially not Python code since there's like [laughter] infinite layers below pretty much. I think people have an over belief in the trustworthiness of code compared to models. And that's what I'm arguing a bit. Okay. Do you write code by the way? Yeah, I'm reluctantly. Reluctantly, exactly. I can hear that. [laughter] But I'm getting But do you read everything that that it No, but I can see many cases where AI do not solve the task and I have to help it. I mean, if we take a simple example, like some kind of bug that is a bit more

1:34:38difficult than a standard bug. Mhm. As a human, you can actually put in a breakpoint and actually do troubleshooting in a very interactive way, which an AI actually cannot do today. Yeah. How can it? I mean, depends if you use a CLI tool to do it or not. Well, I can put like print statements and stuff and and run it and and see what happens, but we can as humans do a much more, I would say, efficient interaction in the debugging phase than AI at least today can until they have, you know, that type of

1:35:06functionality. Yeah, but like if if you take like a C debugger or something like that, like there are interactive interfaces, but if they would break down that into like a CLI with sessions. Yes, but they don't do that today, right? No, of course, but like you should you in the future, absolutely. I think it's so strange that we don't have that today, right? I think it's so strange that we don't have that today, right? I think it's so strange that we don't have that today,

1:35:27right? Yeah, some people probably do, but like I mean, it's very unevenly distributed for sure. It would It would It will come for sure because I think it's so powerful and I can see myself when I fix some fix some bugs that uh I try to use the agent, I always do, but then it fails and then I go in and I

1:35:43can then so much easier fix it. So, there are many cases just By reading the code. Yes. Okay. But there's like I mean, there are examples like this. For example, like this whole Chrome debugging protocol where that basically gives you this Chrome inspector that web developers use to fix pixel errors like We we will come to that. can use that already like through an

1:36:03MCP. And I think super super powerful. The only problem is that it's also super super slow because it entered this like debug mode of Chrome. So, I don't use it all the time, but if there is like a pixel error that I can't solve by just prompting and prompting again, then I'm like, "Okay,

1:36:18go in, check all the CSS, all the HTML." Like, "I can't bother to do it. Like, you solve it." And it solves it. So, And and I'm sure that you know, Claude and and Cursor and others will have debugging proper, you know, functionality properly soon as well. And they will unless they will have computer use and simply use keyboard and mouse, unfortunately. But of course, that's where where most models are really trying to to achieve, you know, good results in. And GPT-5.5 that was just released, I mean, they're really

1:36:43performing well in computer use. A lot is happening there. But then we can keep this CMS because there's more than that. [laughter] Have you had the agent using WordPress? Great. Back to [laughter] square one. Which way to go? Yeah, that's actually funny. Like, we had this data set like back back in early early 2025. This was like the first computer use beta was released by

1:37:05Anthropic I think autumn 24. And we're like, "Oh this is amazing. Now we don't have to build MCPs. We can just like let it loose." And we we tried we built our own like a copy of their sample project for like spinning up your own VM, like opening the browser, letting loose the computer use. I was going to do a demo to Tome and like, "Look here when it posts in WordPress on our own site." And it just missed the fact that it was still in the title field and just typed out the whole blog post in the title. I was But I think people, you know, don't

1:37:36understand that AI in some ways are actually very stupid. And if I just can elaborate a bit on that, I mean, I usually use this kind of pyramid from Open AI to explain my view on on AI and it's extremely good in in memory. Meaning you can actually put 10 books into the prompt and it can have more or less perfect recall in all these 10

1:37:54books. Uh in a single prompt. Like and that no human have any way any chance of do anything similar to that. But then if you take something like Arc H I 3, the recent kind of benchmark that came out, then you need to be able to interact with some kind of interface and you need to reason about it and take

1:38:11action. And it performs horribly. For humans, you get 100% accuracy. Any human put in that test can solve these kind of tasks. And that's the state of the art AI get less than 1% accuracy. And and they they don't really have the ability to interact in a in an interactive way. Um and and then reason about that in the

1:38:33way that humans can. Yeah, they have no training in that modality pretty much. Well, they are partly. I mean computer uses that partly, isn't it? Yeah, in a sense, but it's still probably like screenshot based and like pretty slow in the sense where like I mean Gemini can watch videos now as well. So it's getting a lot better. But You can you can say that it it's not trained that way, but in reality it

1:38:56doesn't have that capability. Yeah, because it's not trained that way. But once they do, like it will it will ace that. Why don't everyone is trying to beat Arc H I 3? No one has done it. Yeah, but it's been out for what? Less than a year. But everyone is trying to solve computer use for ages as well. I mean it's it's

1:39:13super hard. Uh I will Wait wait until 2028. I think you know I mean it's harder than people think. I I would really argue that. And uh But why do we need computer use so bad? Uh there's one reason. It's it's called like banks. [laughter] Don't have empty space and like still need to go in and find boring receipts and stuff. So like you just use computer

1:39:36use for that. It's amazing. We can go into the whole Jeff and the Jan LeCun argument here, but that's the reason I think you know, we need to do some changes to actually get this type of intelligence. I don't think we're there yet. No. And I think there are a number of ways that we need to step uh we'll take to to

1:39:51actually achieve this. And you think that goes like beyond the transformer models capability? still based on transformer, but you do latent you know, latent space reasoning then in a different way. And we actually seen a number of papers going that route. So, it's starting to move in that direction. And of course, all all the image generator is not reasoning in pixel space. They are first having autoencoder around the the image and then doing the diffusion with

1:40:16transformers, you know, in latent space. So, we already have any kind of you know, model that's working without text. It is need needs to be in latent space. So, I think we are seeing clear signs in moving into latent space reasoning, which JEPA is is primarily about. And we need to take that to to I think the next level to have this type of reasoning which is not auto-regressive. Um uh to to make this happen. And I think

1:40:41it will come. And is that basically like the innovation of JEPA then that that you do reasoning in latent space rather than just like tacked on after the fact pretty much with tokens? It's a big I mean, the joint embedding is really about that. So, joint embedding predictive architecture is really you know, taking both X and Y and put it in the embedding space, which is the latent space. And then you do the prediction there not the auto-regressive one, but but once you have that and the world models that everyone is speaking about right now where it can actually

1:41:08judge if a position is good or bad. Uh if that's for physical AI where you try to understand if if the gravity is working in a certain way or whatever it is, uh that can that can be something. But if you have these kind of you know, they have like four or five different modules that needs to work together to achieve this. But if you have that together and and then do the reasoning in the embedding space or latent space or as Elon call it the vector space, then I think we're

1:41:35we'll get somewhere. But it's interesting now because I mean obviously Jan LeCun was like head of AI at Meta for quite some time and like it should have had a lot of resources. So like is it that much harder to do that like they tried and tried and tried and then they did like Lama on the side and and it failed or like well because I I kind of disregarded him because he he just sounded like a grumpy old man pretty much like complaining on all the

1:41:58other labs. Yeah. Uh No, but I think you know he has failed and that's a bit sad. We'll see if they succeed with a new AI my lab that he's having. But I think you know with all the papers and even all the the the other models that moves a bit beyond text because text is really latent in some way. The level of abstraction that text has is rather high compared to

1:42:17pixels and audio. Right. It audio and and and images is super high in dimensionality so then it's super hard to reason in that space. So to be auto regressive in pixel or audio space would be idiotic. And no image model today is doing that. So they already have moved to that space. So it's already taken that kind of step. But you can also I mean like we had Karim here the founder of what's his super intelligent six six super you know so so startup in Sweden that is more working towards um you know they they are saying uh we we believe in a different route towards AGI

1:42:58which is more or less more carefully understanding well going language and text as way to AGI. Hmm, will that really work? Do we have the real real real understanding of how we build stuff? So they are rather going the physical route. How do we build a physical world model for a or how do we learn to simply grip something and move something in a warehouse and that is sort of the the baby step of something that basically builds a world model that becomes more and more effective where basically there is some research in Sweden which basically says hmm maybe not the text way is the right way." And and and I

1:43:34think but but here we put the emphasis even stronger on the world model, the memory, and the recursiveness. But just to summarize it, I think the important point here is to still understand that humans are better than AI in some tasks. Very clearly demonstrated by Arc AGI 3. I mean Yeah, for sure. It's so easy for humans to do this task and AI independently of why it still can't do it today, right? So, there are things that AI cannot do today and humans do much easier than AI, but certainly the opposite as well. So, if you have that understanding, I think you have a much bigger chance of

1:44:07I would like but the and the core question is also do we simply need to do more of of the same we're doing, or are there fundamental architectural tweaks missing? And I think that's I think something is missing. This is my view. Who knows? Who knows? Agreed. Could be, but I think you told me this, Tom, from from the Y Combinator event yesterday that like actually AI doesn't have a problem. So, there's still like a space left for humans to have like the problem that AI can

1:44:35fulfill. And that's where like if AI is going to take all jobs, like humans going to have problems and that will kind of generate new work in itself. So, there there is this like human agency aspect that that is like uh I think at least for the time being like uniquely human and and obviously like what Arc AGI 3 is kind of edging towards is maybe like is part of the same

1:45:01Mhm. capability set. I mean, just learning so quickly as humans do from a few examples and then extrapolating from that is something that today's models have a really hard time doing. Yeah. And uh but it will come there, of course. It will come, but I there there are some breakthroughs left. Yeah. Some small ones, maybe.

1:45:21Might be. [laughter] Who knows? But uh it won't be that long, Adam. I don't No, that's another question. Moving If we move truly to the how long before 3 will be resolved. I guess one year. I don't know. Yeah, yeah, probably. So that the crazy thing we've had a couple of benchmarks. Oh, it's impossible and then 3 months later it's oh we're on the we're

1:45:42on the move now. Yeah, yeah, so we've seen that one. Cool. Let's see I was skipping some topics here. Time is flying by here. Um Yeah, so Okay, so we have more complicated potentially marketing tactics happening. Um What do you think you know, what will the impact of AI agents you know, being put in in a more agentic way be on you know, on on the tactics of

1:46:12marketing being the future. Do you see what I mean? It's a bit complicated. Yeah, I think I mean when we say marketing marketing is a very broad word. I mean we have lots of like branding and creativity that I think is still very human domains. Um Where in our space which is more kind of data-driven marketing, that's where an agentic kind of setup works really well because it's all about analyzing data, you know, adjusting what's going on, running experiments and then create that loop. But I think like Legora did with

1:46:45Jude Law. Did you see that? So they basically have Jude Law as their brand ambassador which was like holy I mean Jude Law is one of my favorite actors and I never thought he would go out and basically sell himself especially like you know, Hollywood with AI that's like it's not the best combination. [laughter] Um but in the end he decided to do it and I think that's a you know, truly Klarna

1:47:12did it with Snoop Dogg as well. You know, that type of marketing I don't think any AI would be like oh, spend uh know how much to law tech, but he's probably not cheap. I can't argue with that. Yeah, like spend a couple of million dollars on new law to get attention. But but I think the flip argument to that is like I mean if you listen to bio labs and stuff like that, they say like we try millions or billions of molecules

1:47:37now with AI. And I mean Meta recently published a prediction model that will predict the emotional response in humans based on the video. They've done like fMRI scans. That's crazy. So you just do the same setup, right? You generate the billions of potential ads, check the emotions of the potential simulated users that you want to reach, and you're going to find new law or something better, I think, with AI doing this

1:48:01computing and the agents to work. I mean at least if you were to listen to the latest years of like influencer marketing, everything should be authentic, you know, even big companies like Vattenfall or whatever. They shouldn't spend millions on commercials. They should have a handheld mobile because it's authentic, and that's what people relates to and things like that. So I mean, of course you will probably be able to and Meta also have have this vibe social network where they generate all the content with AI. So if you kind of combine that with the brain scan models, then you have a pretty scary uh setup where you can basically decide

1:48:38what reaction [snorts] you want from your ICT. puts this as like AI is going to one shot the human limbic system, and I think he's right. Like I wouldn't bet against against Elon. Yeah, that's something to do. But but speaking in general, let's take the the Swedish election now coming up here, and you can imagine a lot of marketing being done here for for that,

1:48:57of course. We could even see I think um the election in Was it Venezuela? No, Argentina. Argentina in uh uh 2 years ago or something. I think call it the first like AI election or something, and they used a lot of generative AI to put up like bad pictures of the They were but they were Do you think it's fun because they were roast roasting each other? So, it was people it was obvious it was fake. But, it was still communication-wise it was quite

1:49:23powerful. And of course we're seeing the same now with Trump and one of the other kind of weird AI-generated images being actually not trying to make it look real. They're just using it as a communication tool, I guess. Yeah, and in the Iran war like they have been pretty aggressive, haven't they? With like the Lego Trump and like the lots [laughter] of Trump waiting for the

1:49:42Iranian delegation. Yeah. Like fake obviously fake videos and and like they still use that to try to get their message out. but this point is that it's completely AI-generated, but it's not trying to fool us. It's trying to use it as a powerful communication persuader. Yeah, but I think it's very hard to draw that line, isn't it? I mean, I I've seen some on LinkedIn where they even write on the uh image itself like AI-generated, but people are still like blowing up in the comments and be like, "Oh, this is a Palantir CEO with uh you know, our state head of state, you know?" And so,

1:50:17I I think that that line is hard to Yeah. Draw basically. But, if we just try to be really uh visionary here in terms of what marketing will look like in the future. One one of course is is simply to use AI and energetic to uh to optimize the the marketing that you have.

1:50:34Yeah. But, I guess you can also think of the product that you you market to be more AI-driven. Meaning, just as Spotify could have Discovery Weekly, which is personalized to hundreds of millions of user and then adapted to this is actually the music you'd like. Yeah. Perhaps we can have ads in the future that is so personalized and actually having this kind of they understand you as a human so much so they actually have different

1:50:59you know, videos that is optimized. Like one example that we've seen recently like with our our growth agency Growth Hackers is that we started running Cogni. Yeah. Cogni does a lot of analysis of the data. Yeah. And all of a sudden it's like, you should really like build a some SEO page for marknadsföringsbyrå,

1:51:17like marketing agency. Mhm. We're like, we're not a marketing agency, we're a growth agency, you know? And and this hadn't crossed our mind really, but we were buying it on search, of course, because like people looking for our services think of us as a marketing agency. Right. So like from a brand positioning perspective, like it didn't make sense as our identity, but it made sense like how people tried to find us anyway, because they didn't know about us maybe, but then if we popped up, it was a

1:51:45interesting proposition. So that's kind of working our way into more of the idea space, the feeling space, the like what what people yeah, how they reason about this. So in that sense like AI can already help kind of take away our prejudice about who we are and try to position us better in the market.

1:52:08Mhm. I think also like historically, you've had a page or a landing page and and you you try to run AB test on that in order to improve conversion rate and things like that. But then and if you're even a little bit more advanced, you start to do personalization, which means parts of the sites adapt depending on who you are, like um And I think AI in a sense completely will rewrite that because why would you need one single user journey? You could have like Spotify for hard rock players and then the the that user journey can be specifically for people that like to listen to hard rock. Or you can have

1:52:42someone for blues. You can have like all the subcategories, you know? AI can generate millions of user journeys that are specifically for But I think that's what you're also aiming at then, because then marketing and product is converging in some ways. Yeah, definitely. I mean, I would say that it depends like in in our industry in the startup space, I think I mean, if you are at Spotify and do marketing or if you are at Lovable, you know, what are you doing? I mean, it's a great product. It sells itself, right? You You

1:53:13just try to amplify that. Uh and I think that's a big difference than if you're at a I don't know, an energy company that Yeah, but I think it's also from the from the loop, right? So, what is a product, right? And ultimately, if you have fundamentally personalized perceptions based or or journeys. Yeah. Then the meta product is fundamentally

1:53:32different products. Yeah. I mean So, in this sense, then you have a marketing intelligence product loop Yeah. I mean, it's the same value that you give. It's just adapted to who you are. Because I I mean, it's it's Spotify is still a music service. It serve any music lover, but if I feel that I am at home as a hard rock lover, and the user journey is adapted to that, then I'm going to feel more connection to the brand. And that basically makes the brand Yeah, but it makes total sense that because if if we are talking about product as something that you employ to do something that solves

1:54:06friction or something you want done. So, I want to employ a product that is perfect for my use. Mhm. And all of a sudden then, what what where is the line between what is product, what is marketing, what is UX? It's It's used It converges. It converges. into your personal AI agent and like

1:54:25Yeah, exactly. product companies left to doing more like uh agent-based payments, APIs, information that's really clear is then your meta product or the the way you experience the product is part of the meta product in that It's about how you explain it to different people. I mean, the value in the product in itself is what solves the

1:54:44problem, right? Yeah, so the core the core functionality might be the same, but the but and and maybe that's an inner core, and but I'm I'm not I'm not layering and layering layer until like I mean, that's a That's That's an interesting I think it converges. common interests and things like that. But at the basics, it needs to be a strong relationship. I mean, it's kind of same that with the product. You have to love the problem or you you have to have a problem with the problem that it solves, basically. But, then it helps if it's closer to you, it looks kind of the same to you, and it it talks the way

1:55:17that you talk. Then, you come a little bit closer, which means that you can you will pay more for that than uh an exact replica of that service that just looks But, I I just wanted to close this thing going further. If I put the thesis down, marketing will never be further away from product than it is right now. It will use conversion, get closer and closer and closer in terms of how this You know, we always talk about product, marketing, and then we have the distance between marketing and sales. I mean, like so then we are talking about really that distance goes away, and I'm saying,

1:55:52"Well, it's not going to go away. It's just going to go closer and closer." So, it's the context convergence. Yeah, isn't it? Yeah, it is. I think I think so, too. I mean, as ads expert as well that you are in from the past, you know, if we do consider, of course, Google search going down in popularity, it's just agents, you know, looking at Google

1:56:13search in the future potentially. And then, people are looking more and more at the yeah, text output or video output or yeah, image output output. I mean, OpenAI wanted to move ads into their output, so to speak. Mhm. What do you think? What this will be the proper way to just, you know, get the message out there, so to speak, when you

1:56:35can't show ads in the traditional way? Yeah, so there are multiple ways, right, to show ads as well. I mean, Google has had a super strong position because they have a way to show ads where people has already have they already have the intent. They are at the bottom of the funnel. They figure out, "I need now to book a concert ticket to blah blah blah." And they Google that, and they find it, and they're happy. But what happened in the moment before like we call this the upper funnel or awareness

1:57:01stage for example. Uh that is of course still left even if the agents do like figuring out how to help me solve my problem but but creating the problem, that's still marketing as well. Upper funnel. Upper, middle, Yeah. Yeah, I like that. We're associating ourselves with the problem area. Exactly. And that might become even more like a creative like exploded by AI basically like I mean this Jew Law example is is a good good example I think that they're basically like they're borrowing a little bit of his

1:57:37brand to also amplify their own. But like maybe there's a crowdsourced version of this like I figure out that I need everybody to eat my ice cream so I like give it to certain people. I have AI help me find this and then like by them eating this ice cream it now becomes valuable in the eyes of others and now I done marketing and it's like agents then get hooked into find ice

1:57:59cream and buy it, right? I liked I liked the way you painted the funnel now, upper, middle, lower and how we can understand how agents are helping us in different ways depending on you know, what do you have a lead generation problem or you have a lead nurturing problem or you have a closing problem. Uh makes total sense to understand then that we need to categorize different types of marketing workflows in relation to creating the demand versus closing

1:58:24it. Yeah, but I think the the like where you start with an LLM, that's not fully understood like do you start in the awareness stage to be like oh list all the telecom exactly or are you just asking like what's the best energy provider in Stockholm? I mean like I I I I if you like watches, it starts with you nerding around watches, going on YouTube, doing different things and all of a sudden oh I like pilot watches oh la la la learning about that and then all that's my nurturing story already there

1:58:50right? So upper of to bottom. Yeah, but I think in historically what we've done is we of course mapped all the clicks that happens to a page and then we can see like oh, this user had you know within the 90 days that we know that it takes to consider a new telecom provider for example, we can see that oh, they started searching here, they started looking at which type of open

1:59:11mind or like subscription they want. And then they like 90 days later they ended up at what they want and they can kind of choose that. And then we can map out all those clicks and understand like these clicks are top of funnel and then we can do what's called attribution, we can attribute the value all the way back. But I mean the LLMs kind of

1:59:29destroy this. That's a different story. The whole that whole exploration is done in the LLM basically and there's not a single like call that goes out to the page. But isn't this where the power shifts then? Because somewhere in my chat repository chat GPT knows exactly which toys I'm dreaming of. You know what I mean? Yeah, but I guess like if

1:59:53how can you monetize that? But the but the difference is like if you present I mean I heard someone compare like Google Ads with magazines. Because basically if you look through a expensive magazine, the ads are as good as the content. Yeah. And that's true for a Google search as well. I mean that the ads are as good as the organic content. But if if you ask an LLM a simple question it can't give you some like oh, this is kind of good

2:00:18things, but there's also this. Mhm. I mean they have to give you an answer, probably a pretty deterministic answer. They can't give you Otherwise someone's going to happy with his search. kind of hard to fold in ads into the format in one sense. Yeah, so that's an interesting one. Where where will that go? Or will it will it mean that the like the normal search and ad space is not going away? What what what is the

2:00:41prediction here? You're you're pointing at something. Hold on. Yeah, exactly. I mean Meta has this model, of course, that's trained on the 3 billion users that they have. So, they're supposed to know you. And I think that if you have a model like that, then that model can interact with different MCPs or clearly different interface, whatever it is. And then you can kind of sort between those and then

2:01:05get a couple of responses back. So, so basically, it's a data exchange between the advertisers and the LLM. But you you can't present that in in the Will will ChatGPT ever sell the top right corner to a Yeah, exactly. Will I have back-to-back banners in ChatGPT, right? Yeah, but maybe I mean, a gentle commerce is an obvious way to to try to to bring that monetization in there in

2:01:32the like a genuine way, right? Do you You know, like prompting You know You know when you're prompting and then now I could do this this this this even even in the prompting you know, structure. He could easily slip in, "Do you want me to search where I can find this?" I That even happened right now. It already happened right now. Do you want me to find which

2:01:50which places you can buy this? [laughter] We are doing it already in some ways. Yeah, it is happening. I mean, if you ask you're going to get an answer if you ask an LLM which telecom provider you should have or, But of course, also like discounts could be a way to do it. Like you you ask for five Mexican restaurants, you get five, but then it's like, "Do you make the Tex-Mex restaurant like with a little

2:02:10bit of Texan flavor with a 5% discount?" And like, "Yeah, maybe. That's fine. Like I can do nachos with steak or Where will this go? Yeah, but in the end, I guess you're nudging the trust of the LLM if you start showing ads. And I think that's the whole point of Anthropic's whole commercial series that we're never going to do ads because it jeopardizes the

2:02:32whole trust of the platform, basically. Is it? If you have to pay for it in the future, perhaps people will still accept it, you know, if they get it for free. Yeah, yeah, I know. I mean it's the whole also like B2B, B2B, B2C plays that like Anthropic heavily enterprise focused while like OpenAI has gone like really broad, a much more adopted free plan and like maybe ads is the only solution that they can provide this in the future with. I mean they tried this like scan your eye for crypto approaches

2:03:05well, but I guess that backfired then. Yeah, a lot will change for sure and uh yeah, I mean marketing is is still very important tool to have, right? I mean otherwise we can't grow businesses and we need to figure out a way to do this in the in the gigantic future in some way or form. Yeah, and especially if like for example SaaS is becoming like completely commoditized or you can even like just spin up your customer stuff like I mean I think we're looking at the new super cycle of sales in a sense like we have all the incumbent companies, they sell their products, they have their stock

2:03:38valuations and all of that, but like AI factories is coming, humanoid robots is coming, AI capabilities is coming so like everybody that's not a like pure play AI company is going to get squeezed pretty much. And then they need to do more selling and more ads. Yeah. But yeah, okay. I'm just trying to think if we go even further, do these kind of extreme use case thinking like Elon often does, thinking that okay, in the future we won't have ads, we won't have apps, we just have a graphical interface or a video being streamed to us all the time by an AI then understanding what we're saying and

2:04:17and doing all the time. We still want to have, you know, companies that want to influence us and inform us of of their products or their services somehow. Still want the illusion of choice. Yes. [laughter] Yes. Yeah, I'm unless you're talking about like the post-capitalism world. I mean that World world of abundance and Yeah, exactly. That might eventually happen as well. And then, of course, the the rules of the game are different, I

2:04:43think. Speaking of the world of abundance, uh if we were to move to an AGI future, we mentioned some years here, perhaps 2000 29 or something or or do you have you think, Anders? Well, I've been saying Ray Kurzweil's year for a long time. him. Okay. I can [clears throat] I can concur. He copied Ray. We on the first episode, I think, we said Yeah, it feels like the safest bet,

2:05:08right? Yeah, but it's interesting how it's sort of Did we I don't know. Yeah, it's the safest bet. Yeah, but I think this actually makes a lot of sense as well. I mean, a lot of people are saying 1 year or something Some people are saying we have it here, and I certainly don't I mean, it's so easy to show. Just it can't do our AGI 3. I mean, obviously, we don't have it

2:05:27here today. But But, of course, like from a commercial intents and purposes, like I don't earn money by solving our KGI, but like but there are things AI can't do. So, if you define AGI as something that can do on the par of an average human co-worker, we know it's a lot of things it can't do today, Yeah, yeah. And at the same time, it can unlock like tremendous value that like human can't. So, like it depends like which lens you're looking at this from. Like, I mean, from a from the SAS world or whatever, like

2:06:00maybe it's already here. But, obviously, from the like I want to help with the finding out the meaning of life world, like maybe it's not. But then and then you have, of course, okay, we can do now AGI uh desktops. And then, what about physical AGI? And what about then the regulatory landscape as well? We have AGI at scale, right? So, when we say 2029, which one are you thinking about then? The first level or the physical level or the abundant

2:06:32level. I I think 29, maybe I can agree with that, but I'm thinking the first level of sort of Yeah, I'm thinking digital space. digital space AGI will probably take a bit longer. And and I'm not saying physical space at scale. I don't I because I don't think the regulatory will be there. But it could actually be I think the physical space is moving very very

2:06:51quickly as well. Yeah, yeah. Yeah, for sure. Like faster than humans obviously. My bad. Yeah, I will see. So, it's not impossible both can happen. But are you are you on that sort of trajectory of 20 years or 50 years? I I think my timelines are like really compressed. Like I I I usually maybe my timelines are too compressed that I

2:07:10expect more to happen than does. Yeah, we have this problem that we usually are too far ahead and we sit and talk to people and they're like, "Ah, we don't understand what you're talking about." and then 8 months later, there's another company doing the exact same thing and they're having success. So, I mean uh market window

2:07:27Yeah, exactly. But but I think I mean I'm I'm worried that like uh do we really discuss what a post like AGI level one if you like then world looks like in the this upcoming election in Sweden for example? That's a good point. And I think we should like because there should be like a democratic mandate for the plan. Like is it just are we just going to all the leave off the like per year findighet like this big rare earth deposit that they found outside of

2:07:59Kiruna? Because I mean a lot of the Swedish budget that goes to the state is based on labor Mhm. taxes, right? But I think I think it's also important to distinguish between the you know, technical progress of AI which will go probably very very fast and that is what I mean with 29. Meaning that we have the technological possibility and potential of AI. But then that the actual you know, companies will adapt, the society will adapt that will take much

2:08:25much longer than that. agree. I think that when we talk to clients now, it's more about them adapting to AI. It's not about the latest models or like it's it's about them, basically. And it's an interesting because some people, of course, lean more into it and others are more hesitant. And I'm really scared about and I wish politicians spoke about that more is the concentration of power and the AI divide now being created because that will go

2:08:49just faster and faster. Yeah. Some companies that really do make the proper use of AI, they will just accelerate insanely fast. Yeah. Yeah. That's very scary to me. Yeah. Yeah, and it's it's I mean, it's putting a lot of companies or like livelihoods or like revenue or BMP under exposure. I mean, for example, all of these like dark factories that they talk about in China like I mean, they will be able to produce a bigger and bigger share of what we consume. And if they do it at like AI and robot prices, which like cuz they almost do that already like with with suppressed wages at etc. but like

2:09:26And I Consumers like cheap things. And it's going to be harder and harder if you haven't adopted your processes to AI. And and obviously like Europe has this pretty interesting situation with like no real big super successful foundational lab and also like having given away a lot of the robotics to China, I think.

2:09:44Yeah. But I think like back to the income gap because that's already a problem. I mean, I met what was that after parties to the offer met some politicians. Um shouldn't say which which party, but and they were talking about this, you know, what's going on in New York and you know, we can already see this like grassroots happening without AI. And of course, AI is going to amplify this

2:10:11Right. a lot. So, I mean, I think the politicians should start some type of universal basic income. I know Finland has done some testing around it. I mean, why not just start to test so that when or if it happens We need data. We need to have an understanding for what will happen. How can we really understand that if we

2:10:31haven't even experimented? Exactly. Or at least have that debate because maybe people want to vote on like let's turn everything into UBI and like skip all the basic services or the other way around like zero income you just go and like you get like a multiple or like some some kind of stamp card where you consume using that and no money like But back back to I I haven't thought about that but if AGI technically comes within the next election period

2:10:58Yeah. which we are saying then if we are following Ray What are the things we should be discussing then in this election? I think the UBI definitely. The UBI? I mean we need to get some type of program like that going anyway. We need to debate what's the options right? We will start to eat I mean I I believe in experimentation right? So I I think we should start testing. Uh just figure out I I'm saying question I mean it's so far away but at the same time No, you said

2:11:29it this is within this election period. Uh it's the supersonic tsunami coming here but But it's also adoption is later I mean the real effects you know what we are seeing society will probably take a longer time but but you can't know that. No, we can't know that. You can think that you think that based on past changes etc. but like eh if we think about the slow timelines like it was around like pre-World War II like there were very many horses and not

2:11:55so many cars. And then that war happened and like in the '50s there were almost no horses. And of course this was not with like online 24/7 intelligence like a country of geniuses in the data center type level execution. But if you bundle that with like self-replicating robots and stuff like it could go really really really really

2:12:16really fast. And and one point we can be rather sure about is that some companies will move much faster than others. Definitely. They try to build, you know, data centers in space and put people on the moon. Like somebody will do that. Imagine Imagine you have a company, just hypothetically, that has income that is on 70% of the

2:12:36world. Yeah, or like 100%. Like Elon Musk thinks that the total global economy going to 1,000x or something like that. Yeah. And most of that growth will happen in space. And if that's equal SpaceX, I mean, it will close to 99 or worse. But let's say 50% just don't make it too extreme. Even 50%. If one company owns 50% of the like turnover that we're seeing in the world,

2:12:58what would that mean? I mean, they will be much more powerful than any kind of country that we have. I think we're already there at some point. You're right. We are. But but And now comes why this debate is important regardless of exactly how far we are. It goes back to the William Gibson quote, "The future is already

2:13:17here. It's just unevenly distributed." And if you follow that trajectory, if we think we going to go closer to AGI, technically, it means some companies will from now to here reap a lot more benefits. So this this this gap, this distribution of the future will be there. Therefore, how will we as a society cope with the people on the different sides of this? So the whole UBI income discussion is also about how do we redistribute uh balances. Yeah. That comes a a

2:13:49question we can have now. Which which balances are left as well? Like I mean, imagine that that we had zero tax income in year 2030. Like what do we do then? Maybe we tax like we we're currently giving away mines, for example, the Swedish like um um politics for like if you want to establish a new uranium mine in in Jämtland. It's a big local discussion, all [clears throat]

2:14:12Where does those profit go? And currently like English companies or whoever comes here and builds the mines, they can reap a lot of those profits. But in like a world where raw materials and electricity determines outcomes pretty much, which is like I guess what the robot factories and everything is, then we need that to look a lot different, I think, if we want to have

2:14:35prosperity in Sweden. So these are very valid conversations to be had now, I think. Yeah, but I think like because we seem to politicians to make these decisions, but they haven't made decisions for quite some time now. I mean, if we look at the the data center build-out of 930 billion dollars, who and and compare that to the Apollo project or the railway project,

2:14:59which dwarfs all of them. Who founded that? Did the government fund that? No, it was private companies. So we're already there, I think. And and this has been so for quite some time. So But then it then then the world will not get balanced, cuz then it's then it's will always be that's the richest is capital Yeah, I think we're all oligarchs in one way, you know, runs the world. And and unfortunately, I mean, in in we do see some uh maybe hopeful sparks that politicians can start to take back some of that power, because of course that needs to happen. And if I had the

2:15:37100 richest people in the room, I would say like, you know, you should do this, otherwise, you know, it's going to be French Revolution on your ass. And you know, you don't want that. So I think everyone gains from a more equal society. Yeah. And that's a the discussion needs to be had, basically. What is it the Lex Friedman always says, um power corrupts, but absolute power

2:15:59corrupts absolutely. Exactly. [laughter] If some person or company gets absolute power, it will corrupt them somehow. And it also destroys innovation, right? I mean, if they don't have anything to compete with, then they probably just Okay, coming to this last question then, at some point we probably have AGI, perhaps very fast and perhaps a bit later and perhaps technology is moving very fast and adoption a bit slower, but still at some

2:16:25point it will happen. We can imagine them two extremes here, either either it can be the the horrible future and we have the Terminators and the Matrix and machines trying to kill us all, which, you know, given the geopolitical situation is not unfeasible either. They trade on us on our data. [laughter] And then we can take the other extreme. And that's the utopian future and we will have AI that solves the climate crisis and finds cure for cancer and fix the energy problem and basically creates the world of abundance that makes price and services and goods go to zero and we don't need to have a capitalized society

2:17:07anymore. Where where do you think we will end up? If we start with you Tom here, if we do you think what's the chance of us moving in one of these two extremes? Yeah, I mean, there there good arguments for both sides, I guess. I I heard this like when we uh happen to step on an ant, we don't feel that much you know, hurt or we're not we don't care that much about the ant. Why is that? And science has that apparently it's because our intelligence gap is so

2:17:39big. Right. So, if you argue that we would create an AGI that is many times smarter than us, that AGI would probably not feel so bad if they need to remove a couple of people because we're too many, we're consuming too much and so on. So, that's of course an an uh a rather dystopian um

2:17:59scenario. Um but then, you know, I think we created in this liberal capitalism system that we have across all countries today. We have this system that kind of just drives innovation and there's no way to stop it. I mean we can't go back and say like oh we're going to do plan economy now and you know even if that has its own ideas you know it's never going to you know that's never going to happen. So I think we just have to jump on the train and I think being positive and like the abundance movement for example I think that's probably the the way to go and hopefully you know we create this

2:18:35super intelligent being that can solve our all our problems and everyone can do what they like to do most which is probably being with friends and family and enjoying your hobbies. And brew beer perhaps. Yeah yeah yeah lots of beers. [laughter] More beer. Yeah. So we have the beer brewtopia. Yeah brewtopia definitely. Daniel what do you think? Yeah no I actually used to be in in like the camp of that like yeah no we need to like really put the brakes on this and like you know embrace the climate movement and like draw down consumption and like do all of this really seriously like otherwise we're going to up the

2:19:14earth. But now I think the irony is that like I think people don't like that Elon Musk is on his way to space and is going to spend so much money on spaceships etc. but I think that it might actually save the earth's climate because of it. Because if you move all of these fantastic AI activity out into space because it's more efficient you also don't have to like paperclip the whole earth. It can stay as a like little museum of the origins and and like AI can go out and do stuff in space because it's just more efficient better scale and and

2:19:49yeah way more inspiring to the AI. We make AI multi-planetary. Exactly. And we stay here. [laughter] And that's that's biggest scam. Look how much it's over there. We use ants. You want to go in that direction. It depends, you know, it singularity can happen as well. We can follow on the hard drive or something. I haven't thought about that. That's a really profound idea, you know. We actually create an outlet for the AI

2:20:12somewhere. Yeah, and then like the the whole market economy or like the consumption part that goes with it kind of just get 1% of the profit from the AI. That's written into the source code. [laughter] And we have abundance on earth and like I mean, he's even argued that this Dyson swarm that he he plans to send out like people say that jokingly, I think because he's really planning not to circle it around the sun, but have it somewhere

2:20:38close to earth primarily. And he can even use that to shadow some of the sun rays coming in and thereby dampen the whole like greenhouse effect or the effects of the greenhouse effect, if you will. So, I think there might actually be like a climate solution in the the whole like background here, you actually solved the climate we went to the abundance. We solved the climate crisis in a

2:21:02completely different way. Yeah, because if we build out AI like as much as we will want to in the next 50 years only on earth, I mean, then we living we fixed it. We fixed it up. You can forget about taking a stroll in nature after that. It's going to be like pure industrial area pretty much, but but yeah, going to space is a great

2:21:20solution, I think. Interesting. Very inspiring uh Yeah, I like this. That was a new angle. We we've tried this question like 100 times or something and uh yeah. We're we're building data here. [laughter] Going to train an AI on this and What what what what's the optimal AGI solution here? We we we all we almost

2:21:39have the answer now. I mean, this is already streaming to the internet and we're already [laughter] contributing to the contributing Yeah. Yeah. It's being rocked. Okay, thank you so much Banner, Setteval, and Tomstrom for coming here. I loved that you could came here come here and and speak about so many amazing things and I hope you can stay on and continue to speak after after work and the camera turns off. So,

2:22:04thank you so much for coming in. Thank you so much. Thank you for having us. Yeah, thank you. The pleasure it was really ours. It was great to be here. Nice atmosphere, good drinks, and good company. So, thank you. Thank you. Yeah, definitely.

2:22:19[music]

2:25:44Mhm.

2:25:53Mhm.