AIAW Podcast
AIAW Podcast
E175 - Proprietary Data as the Ultimate AI Moat - Anders Hammarbäck
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In Episode 175 of the AIAW Podcast, we’re joined by Anders Hammarbäck, Co-Founder and CEO of RedpineAI, to explore why proprietary data may become the ultimate competitive advantage in the AI era. As models increasingly commoditize, Anders argues that the real moat lies in owning and structuring high-quality, licensed data. We discuss RedpineAI’s vision of a “Knowledge Layer” designed to reduce hallucinations, power agentic AI systems, and unlock new applications in science and enterprise. From Europe’s role in data sovereignty to the future of AGI and the evolving labor market, this episode dives into how the next wave of AI innovation may be defined not by bigger models—but by better data.
Follow us on youtube: https://www.youtube.com/@aiawpodcast
Setting Up The Hackathon With OpenAI
Anders HammarbäckWe we set it up together with OpenAI and uh Lovable reached out and then Strawberry reached out to be part of it as well. Uh-huh. So we were Did you reach out to OpenAI? Yeah, we have an investor from OpenAI. An angel investor who works at OpenAI. So he's said I'm gonna be in Stockholm. Let's do something together.
Anders ArptegYeah, yeah, yeah. Perfect. That's a great setup. It's a great opening together.
Henrik GöthbergIs it official who that who that person is? Yeah, Colin Evans. Colin Evans.
Anders HammarbäckWhat's his role in OpenAI? He handles startups and VC ecosystems in Silicon Valley.
Anders ArptegBut he invested privately or is it? Yeah, okay. Privately. Cool. So then you got the message from him, and then you got in contact with Lovable and uh Strawberry. That's right.
Anders HammarbäckWe reached out to these partners to set up a fantastic hackathon to really draw on great models, great agents, and of course high quality data from Red Pine.
Anders ArptegRight, right. So how was the setup? Could the could they actually get like free access then to your APIs and then they could build the agent on Lovable or how Yeah?
Accessing Proprietary Data Via MCP
Anders HammarbäckSo we integrated at that hackathon with Lovable on the back end through MCP. Yeah. So that uh the hackers, the hacking teams could integrate and access uh proprietary data from our data partners to be able to build even better, even stronger agents to solve real tasks. It was amazing to see the creativity in just five hours. You know, so many solutions came out.
Henrik GöthbergYeah. And what type of uh companies joined hackathon or people?
Anders HammarbäckPeople, individuals, um early teams that were looking to build startups joined to try out test solutions that they had in mind. And uh yeah, so much creativity and and action and great solutions came out of that uh one-day hackathon. No, when was this? It was the end of last year, uh towards the end of November. So it's not that long ago. No, just a few weeks ago.
Henrik GöthbergYeah, and because it's it's almost important to ask that question because everything moves so fast.
Anders HammarbäckThat's right. So it feels like ages ago, but actually very recently. Yeah.
Anders ArptegAnd how did it build the agent? Did it use their own like code editors, or did they actually use Lovable to build the agent themselves?
Anders HammarbäckYeah, I think most of the teams used Lovable, some mix of OpenAI as well, a strawberry browser as well. So there were a lot of agent systems involved in that mix, uh, integrated with Redpine data. And tell us, what were the highlights? What did they actually build? Yeah, I think there were so many great solutions coming out of that. Uh again, in just a day. There were things like sentiment analytics.
Hackathon Highlights And Winning Ideas
Anders HammarbäckUh, we we partnered with a company called All Ears, uh, Swedish company, actually, the audio uh data space. Uh fantastic. And they had lots of great data that we managed to get into RedPine's API and into the agent. So one of the winning teams built a sentiment analysis for stock uh picking, stock purchasing, and stock evaluation using real-time alternative data, and that was pretty good. Another one built a reputational check for companies and brands and CEOs. So the ideas are endless. Endless. That's the fact that fantasy is the only limitation factor. Again, human creativity.
Henrik GöthbergAnd and I used got uh I was did a sneak peek at uh tech arena today. Right. And and as we are saying, there has never been a better time to start a company. The idea is to limit it, right? Yes, fully agree, fully agree.
Anders ArptegAll right, amazing.
Henrik GöthbergHow is tech arena?
Why Events Matter And Which Ones Work
Henrik GöthbergUh Tech Arena, I we can we can segue a little bit about that because for me, Tech Arena is very much uh uh to me a view for uh uh as a meeting ground. I think one one of the most important dimensions is the startups having a chance to matchmake with investors. And it and there and at that scale, uh I don't think it's that many. And and and and so we we get a lot of views on that. And then we have some interesting discussions and panels, but here I'm a little bit like I'm a little bit jaded because here it becomes always super high fly and more panels and and generalist discussions where I I want more meat. Uh but but uh I think it it serves a really, really distinct purpose. But I was gonna ask you what do you think about you know what what types of uh conferences and events do we need as startups? Is Techarina a good uh a good model, or do you do you want it smaller or bigger?
Anders HammarbäckI think techarina is great. I'm glad that someone has taken that relay stick in Stockholm and Sweden to do something. We have a fantastic ecosystem, of course, as we know. And uh we should have a great event as well. I think
Techarena vs Slush: Purpose And Crowd
Anders HammarbäckTecharina is doing a fantastic job. Then I love slush, you know, I've been there I think for six years.
Henrik GöthbergHow do you how do you contrast? Have you been to Techarina last year, maybe? Yeah, yeah, I've been there. So please contrast those two.
Anders HammarbäckOf course, the size um slush is much bigger, you know, spread out. Um perhaps not everyone knows what slush is, so it could just be. I think they've been around for probably 10 plus years, and and it's the the biggest tech conference in uh Scandinavia at least.
Henrik GöthbergBut northern Europe. But if you look at the purpose or the mission, would you say they are the same tech arena and slush, or would you say slush has another identity or another slightly different purpose?
Anders HammarbäckI think the purpose is probably the same: bring together the ecosystem, have people fly in during a very concentrated period of time uh for a lot of meetings and create serendipity uh on site, you know, face to face. So I think that purpose is similar. Then how it's set up, the atmosphere is very different as well.
Henrik GöthbergI uh uh is is it who is going to slush versus who is going to techarienum? Is that I I I'm I'm thinking it's even a little bit more slush has groomed their sort of customers or clientele or the community.
Anders HammarbäckYeah, no, I think slush uh is something special. I was an investor for many years, uh VC, and uh I think VCs uh tend to go to Slush. Also, of course, some startups to create those meetings and LPs. Um, I think those and large enterprises within tech. So those are the kind of main groups of of attendance to slush. Tech arena, I think is also at least when I've been there, it's it's also a mix, of course, startups, you get investors. You also have I think the speaker list is a bit more alternative and exotic. People like this year, I heard Boris Johnson and we've had Lotan and and so forth, not the traditional tech speakers, perhaps.
Henrik GöthbergSo maybe that's uh
Intimate Events And Access To Leaders
Henrik Göthbergyou know a different flavor to yeah, and you can either like that or not like that because it becomes a more generalist, it becomes more curious to go to tech arena, and so you get the I think a more mixed crowd, is my feeling.
Anders HammarbäckYeah, and I think also they're bringing in the uh the public ecosystem in a sense, you know, agencies and and sort of state-backed organizations as well, more so than I see slash do. But I like the you know, events I think is a great way to meet people uh randomly and by chance and so forth, and also plan. I went to the uh Italian tech event called Tech Week Italy in Torino last year. That was the best event I've been to for five years or so. And why was that good?
Henrik GöthbergWhy do you like that?
Anders HammarbäckI mean, not only the weather out, you know, apertivo as opposed to you know November slush weather. Um no, but I think it was the size, it was much smaller, more intimate. Uh was a wide mix of people in coming to Italy, also from the US. They've been instrumental bringing people like Jeff Bezos coming in and Sam Altman the year before. So really the top speakers and very intimate, very accessible.
Henrik GöthbergSo so it is I mean, I maybe you take I'm I'm just so curious on these topics. Yeah. I mean, like, so I I think one one one key question I always think about is like, what's the optimum size? Yeah, because to some degree, when you go into Friends Arena, you get one experience. Um and if you go to a little bit smaller event, it's slightly different, right? So so but uh but then it can be almost too small, or that then it's a completely different ball game. I uh you know, right? So how do you think how big was the one in Italy and how do you think about size of these events?
Anders HammarbäckIt was much smaller, much more intimate. It was fit into one smaller arena here in Stockholm. It's uh you know, the Friends Arena. Um, what's the new name? Strawberry Arena? Yeah, Strawberry Arena. Uh so much, much bigger, of course. So I think the one in Italy, maybe three, four times smaller than um tech arena in terms of the venue size and participation, probably the same. So I think that's good. That created you know a much uh closer proximity between guests and easier to access people who are there. Cool stuff.
Meet The Guest: Anders And RedPine
Anders ArptegBut with that, we love to welcome you here. Thank you, Anders. Anders uh the founder of uh and CEO of Red Pine. Yes. And we're going to dig much deeper into that and hear more about you know what that is and what you're up to these days. I know a lot of stuff is happening there, so really looking forward to that. But you also have um an extensive background in the VC arena, as we just heard. So uh perhaps we'll we'll uh dig a bit deeper into that as well. Happy to happens. But perhaps we can start a bit with you know, if you were to just give a brief introduction to yourself, who is really Anders Hammerback Hammerbeck? Sorry.
Anders HammarbäckYeah, no problem. So uh professionally, right now, uh together with the team at Redpine, uh building the knowledge layer for AI agents. That's our mission, to make agents uh smarter, more intelligent, and more capable to solve tasks, meaningful tasks. That's uh that's Redpine. Let's get into that in a while.
Anders ArptegYes.
Anders HammarbäckAnd and prior to that, I worked in many companies. I love scaling teams, businesses, um, and companies on a global level. So I had the uh fortune and privilege to really work many places around the world and and do so. So I started building my own company when I was 19, studying at high school, ET V Nosit, it's called IT High School, out in Schiesta in the first dot-com era that we all remember. Good old days. Uh that was great, everything's possible. And uh we built a team called Mimes Brun um that became an ed tech company here in Sweden. And it's been alive for 20 years, I think, up till uh last year, actually, sold twice since. So that was a good first foray into entrepreneurship. Everything was possible, and and then that led me into more of a business uh school actually. I went to Hans Hall School and Stockholm School of Economics in Stockholm, and uh also exciting times. Uh, got to work in
Global Career And Data-Driven Work
Anders Hammarbäckbanks in London and got to work in telecom in Nigeria during the summer, which is very exciting. Yeah, and Nigeria. Yeah, Lagos. I was uh building mobile strategy for a telecom operator in 2005 in Lagos, Nigeria. Internship, super exciting, and got to really launch data services, you know, at that time. So it was early days in the in the kind of mobile in the mobile data space, yeah, exactly. So we got to look at the Nordics, what had been launched, what will be launched there, price points, analyze the market of you know, such a boomy market like Lagos. So fantastic opportunity. Um and then I went to Asia for an exchange program and I stayed um for about eight years in total, um, all together in two different periods. Um mainly China, mainly China, and Beijing first Qinghua University, which is now famous for a lot of tech uh and AI uh superstars coming out of that school.
Anders ArptegRight.
Anders HammarbäckSo I went there in 2005 as well. To study or to yeah, to study MBA Exchange from Stockholm here, and uh decided to learn Mandarin. Uh so I'm I'm a fan of difficult things. I like challenges and deep parts of the pool.
Anders ArptegYou speak and write Mandarin?
Anders HammarbäckI speak, I can write on a phone and on a keyboard. Okay. I can write a few characters by hand, but I yeah, yeah. My wife is fluent, she's Italian on that part as well, as a sinologist, but I can only speak and listen. But it that's been exciting as a journey. So, and then I worked in McKinsey for a couple of years as a consultant, management consultant, doing all kinds of projects, always data-driven, of course, analytical projects, uh, getting large amounts of data, analyzing that, and drawing conclusions together with the team. Um, how many years at McKinsey? Three years. So came in three years as a first job after school, after graduation. Um, and again, data was I think the common uh bloodline through that period.
Henrik GöthbergUh well, you're overlapping with Martin Lundqvist at that time, because he's a good friend of ours. Good, he's also been on the pod.
Anders HammarbäckAngel investor in my company.
Henrik GöthbergOh, I don't know.
Anders HammarbäckMartin, yes, shout out to Martin. I will, for sure. Very good. Also, Jan Sandal, who was a guest here a few weeks back. Yeah, yeah, that's right. Angel. Um, yeah, so McKinsey, and then worked in EF Education for a couple of years, a Swedish education um group doing uh work around the world, and we were building uh digital and physical education products in China for kids and teens.
Henrik GöthbergUm in Chinese for Chinese market. No, in English for Chinese markets. In English for so yes, EF Education Language School, English for Chinese market. That's correct.
Anders HammarbäckThat's correct for three-year-olds and uh so most of our customers were literally three. There were a lot of two-year-olds where the parents uh pushed them in one year earlier to get a head start in life.
Henrik GöthbergYeah, a head start in life too.
Anders HammarbäckSo that was
From McKinsey To Edtech And Payments
Anders Hammarbäckand as a new employee there, I got to teach some of those classes. It was a big challenge, maybe the biggest professional challenge I've ever had teaching two-year-olds to speak English. Um, but it was fun, it was good.
Henrik GöthbergOh I'm trying to look into that. I'm trying to picture that.
Anders HammarbäckYeah, yeah. I still uh remember that. It was very fun times, and many things happen. You know, China was opening up uh the doors, and and English was a vehicle and a key to make that happen. So fantastic booming market, and I think we were doing quite well as a player in that market and leading player. Digitization was going on across the stack from marketing to production of education, digital learning tools, everything. So came in there and just uh so much opportunity. And we grew very fast on all fronts across China. I traveled around to 55 cities as part of my my remit, my my responsibility, and opened up new learning centers across China. So super cool, uh great company. Enjoyed it, came back, then worked for a CNV company called Clearo, uh, which is a payments company um in Stockholm, competing in online e-commerce payments. And uh I was building that as a COO from from early period, um three years, uh, and then decided to go into investments into VC.
Henrik GöthbergSo uh, you know, data has always been the focus and uh exciting different angles that sort of build say portfolio of experiences that kind of adds up.
Anders HammarbäckYeah, I think so. I think so. Different different fields, different markets, and I think that's also how I look at business today. I think you know the parallels between industries and businesses and and markets for that sake create opportunity.
Anders ArptegBut then somehow Red Pine got started. How did that did that journey start off?
Anders HammarbäckYeah, Red Pine got started. So nearly now, a year and a half, almost two years ago, uh, my co-founder David Ostadal from a colleague of yours at Spotify. Yeah, we decided that this is the right time to build a company, to your point. And um, as a venture capital investor, I saw so many companies coming to us for investments uh in the app across the AI stack in application, uh, infrastructure companies, and even some model companies. And I think the common challenge I saw in so many of them was that they were using the same or very similar data. Maybe the UX was different, maybe the team was different, had experience and perspectives. But most of them were deploying a base model from a large lab. And that base model's data set uh was pretty much the same, scraped internet data from common crawl, find web, and so forth.
Anders ArptegBut then at that time for training purposes, right? So training AI models.
Anders HammarbäckYeah, the big LLMs had been using that data to train their base model, and the base model was used, of course, by the application layer companies. So as an application layer
The Spark For RedPine: Data Moats
Anders Hammarbäckcompany, I struggled to find a moat and differentiation between those companies.
Henrik GöthbergThat was exactly my question that you know and stole, because you saw that same different different variation for what they're planning to do, but they're relying completely on the underlying LLMs. Where's the moat?
Anders HammarbäckYeah. So people talk about wrapper companies and so forth. I think that's sometimes unfair, but I think it is true in the sense that most of them, when they start, have access to pretty much the same data. Sometimes they have a unique angle there. But anyway, and then the accuracy was often a challenge. You know, what's the accuracy of the prediction model or generative model? And um, that was also, you know, the data was the answer also to that. Um, so we started thinking about this and and digging, and at the same time, big lawsuits were starting to happen and appear on the market between media companies and AI labs, open AI versus New York Times, big lawsuit, as we all read about, and many others as well. So we started to see tectonic shifts in this field, you know, with original content and AI labs, copyright, fair use, all these questions that are still a little bit open and understand that something's big is happening here.
Anders ArptegAnd we need to it seems like the fair use case is swinging in some sense, right? That you are allowed to use copyrighted data in the US for training purposes.
Anders HammarbäckI don't think it's that black-white. I think we we saw a big settlement entropic that sell $1.5 billion for using 500,000 uh copyrighted books for training. So I think that's a good point.
Henrik GöthbergMaybe talks in a different direction. But but the whole mode here, I think we have we've said it in many different ways. But one one way we have said it on this point has also been like you want to figure out a use case or a business idea where whatever you are doing becomes better and better the better the underlying model is, and and not the opposite that you know when the underlying model gets uh good, it eats you. Yeah,
Lawsuits, Fair Use, And Model Training
Henrik Göthbergexactly. So if you can if you can figure that out, uh then you then you have a mode that where you become better by them when they mature, not letting you riding on their investments.
Anders HammarbäckThat's right. And I think data is part of the answer to that, finding those pockets of high-quality data. So I think starting from a different angle, you know, AI has, as we know, and and you know better than anyone, through this uh pod and your work, uh, so much potential and and has already explored a lot of that potential, but the most is yet to come within all fields, but including the the important fields that that we all have. Healthcare is a big one also for us. Education, where I spent a lot of time, there's so much opportunity in building uh personalized, uh smarter, more democratized education across the board.
Henrik GöthbergBut I feel almost we need to I mean I think we're gonna circle back to it, but I'm gonna I'm gonna start with a simple question. Yes. If I go to a more traditional enterprise and they think they want to go start doing AI, how many of them do you think they really understand where the gold lies and what the commodity is? Because without naming names, I think you end up thinking that oh, we need to do AI, and they go all in on that. When in reality it's their data and it's their knowledge that they haven't even started touching to clean up properly, they are missing the point. It's thinking they can do AI and jump, jump, jump to the LLM. So this is a very leading question. But but do you think people are getting it? What you're saying now is obvious when you set in the VC space. But when we get to normal businesses, how should we use AI? Sure. They have they have they haven't really put the laser on the right, the laser target on the right problem yet or the right opportunity. That's my leading question.
Anders HammarbäckWhat you lead to makes a lot of sense. And
Enterprise AI Maturity And Data Demand
Anders Hammarbäckand of course, from top down, a lot of enterprises well, uh, I think understand the potential and transformative potential and also disruptive effect of AI for that industry, almost look at any industry. And there's a top-down agenda to how AI will change both the company and the productivity, but also the services towards the market for that company.
Henrik GöthbergBut the the difference is then to understand what is it that we truly need to invest in, right? And what is it that we will get for free used by them becoming better commodities?
Anders HammarbäckYeah, yeah. And then the change management, even if all of that is decided to get that through the company, takes time. Yeah. The bigger the company, the more time it takes.
Anders ArptegSo you got you got all these kind of insights and you got an idea for Red Pine. Yeah. Right. And uh how how did you actually get started then and given this idea?
Anders HammarbäckWhat so David and I we started um looking at this potential, we started lifting some stones, and and as an investor, one of the takeaways I I had at least, you know, to validate the problem thoroughly uh before starting to build in a certain direction to understand what's the right direction to start. That can change later, but at least to start digging the right place.
Henrik GöthbergA solid thesis, an investment thesis or a startup thesis.
Anders HammarbäckYeah. So we we saw this eight-lane highway is massive, you know. No matter what data type you looked at, um, there's opportunity, right? We can start in almost any field, and there's opportunity. So we started talking to our networks, uh, 200, 250 people, I think, during a couple of months, very intensive validation research work to understand where we should start. And we we spoke to supply, i.e., where the data is, you know, where is their interest in unlocking some of that data from new revenue streams for AI use? And the demand side, which companies are interested in actually licensing or purchasing or acquiring data to have even better AI processes and models and applications. So I think we we put all that together and just realized the opportunities are immense almost anywhere we start. Let's just go full full on in this company.
Henrik GöthbergBut and how did you then narrow the startup thesis? Yeah. How did you go to that and where did you end up in at first?
Anders HammarbäckFirst, we decided to pinpoint AI maturity. I think back to your point about enterprises as a key factor on the demand side. The more mature on the AI drive and journey and demand, the better, of course, as a starting.
Henrik GöthbergThey understand what you are trying to figure out. They understand the problems. Oh, this is what I want. That's right.
Anders HammarbäckAnd and not surprisingly, the large AI labs uh have the biggest maturity because they are building the
From Training To Inference: RedPine’s Pivot
Anders Hammarbäcktechnology and therefore the appetite for data is and was um the highest across fields. So we kind of started there and got in touch with many of the US labs, Asian labs, and also European companies and started working there and looking at them as your first in in in innovator clients. Yeah, design partner clients, partners and so forth. And um also European.
Henrik GöthbergSo you gear towards them, I mean, like to get started somewhere to what is it that we're productive and you listen to what they need.
Anders HammarbäckYeah, exactly. And understanding what's you know, AI moves so incredibly quickly. So to understand what's happening three, six months from now. And then sometimes that wasn't clear and and the road turns very quickly.
Anders ArptegHow did you approach and find the AI labs? Did you find some contacts to speak to them? And or how how did you do that?
Anders HammarbäckYeah, partly through network, uh, through angel investors that we eventually brought on board uh through conferences. Um, again, uh I think that's a great way for entrepreneurs to actually find companies that maybe fly in for those conferences. So different sources, different pools, sometimes uh social networks like LinkedIn, tough, but it can work. Uh so I think different sources uh led us to many of the labs.
Henrik GöthbergBut it's a networking exercise from hell, you know, to really triangulate everything.
Anders HammarbäckEspecially, I think you know, Europe is amazing. This ecosystem is on fire now, but it's still a distance to other parts of the world.
Henrik GöthbergAnd you need to triangulate interactions to uh to get through.
Anders HammarbäckAnd we traveled, I mean we traveled intensely for the first years and still do all the time to Silicon Valley, to Asia, to you know, uh Central and
Building The Product: Data Science Layer
Anders HammarbäckSouthern Europe. Um we keep doing that.
Henrik GöthbergCan I ask you which which lab was the first that sort of gelled with the ideas and that you started working, engaged with more properly? So which one are the first ones?
Anders HammarbäckA lot of MDAs around what we do, a lot of confidentiality, which uh includes who they are. Yeah, yeah, yeah. So it's a sensitive feel because things are moving so extremely fast in this space. But I think the leading companies that we all know, everyone understands the importance of acquiring really good data. When we started, it was more for training, training models, pre-training, especially, mid-training to some degree, but especially pre-training. Then I think as our journey has evolved, uh the focus on data has also moved on towards uh post-training, fine-tuning, for example, uh, reinforcement learning, still very, very popular as a field. And of course, now, as we talked about before, inference is where RedPine, our company, is focusing most of our effort and energy in the product building as well.
Anders ArptegSo that's more than the training, both the pre and post-training. Now it's more the agent-based or inference mode that's that's right.
Henrik GöthbergSo we're focusing on inference data for agents right now, and this we see will carry us and will carry the market for the call that the pivot from your original uh thesis, or is it is it more like an evolution that this oh shit, it's it's an inference game we want to play?
Anders HammarbäckYeah, I think the more like an evolution. The mission is still the same to unlock proprietary data for uh the value of proprietary data for AI and how we do that. I think that has evolved. Our thinking about it has evolved in how we provide even more value to demand and supply from data.
Anders ArptegSo you had a good idea, I think, of the business model at least. And then you need to build a product, I guess, somehow as well. Or was that too simple almost? Or did you have how did you think about another product you need to build?
Anders HammarbäckYeah, great question. So I think the concept we have is is very simple. Great data will power better AI. And then how we deliver that. We have quite a more sophisticated view on how to build that. We, of course, the data itself is valuable, the licensing of that data, and the rights to use that data. And then we're building a data science layer on top of that, which has evolved over the past
Distribution Via MCP And Agent Tooling
Anders Hammarbäckyear. And and our founding uh data scientist, Leonora Vestebacca, she's been here on the pod, I think, yeah, as well. Um, so she's building that layer to really make sense of the data. Uh, and we sometimes look at companies like when Google won the search engine race 20 years ago now, building a page rank algorithm, yes, which was smarter and and better than like us, uh, Alta Vista, excite, and Yahoo and the other companies that we remember. Yes. But at least uh old people like us, not everyone else. Exactly. Not all theone else. So anyway, I think what they did was uh quite innovative, right? Yeah, uh to say the least. And and we think about the same, how to build that data science layer to really retrieve the right data sets. We built a um a sort of subproduct within our product called Red Mirage, uh, which is is also evaluating data. We built another product called Red Anchor, which looks at the retrieval. Are we pulling out the right data sets at the right time? And and are those data sets accurate to the ground truth and so forth. So the data science layer is uh you know super exciting. Uh super exciting. And then the distribution layer on top of that is an MCP right now. It's an API uh access point where companies can integrate with RedPine to access proprietary.
Henrik GöthbergSo as long as they have a strong agentic uh comp AI compound system approach with good tool use, that's right. They simply have that as one part of the tool. You you're looking at your product as distribution through tool use, and therefore you're having your main user persona is the AI engineer, so to speak, in within the company that sort of figures that needs to sort this out.
Anders HammarbäckYeah, or the agent uh itself, herself, you know. The agent as a customer is sort of very exciting to think about, right? It's yeah, marketing to agents. How do you market to agents? What are the kind of purchasing behavior and is it automatic like that?
Anders ArptegSo it more or less finds your proper API and data sets automatically, or is people involved still?
Anders HammarbäckI think it depends on when people listen to this podcast. The answer is yes or no. But right now, in January or February 2026, uh we're getting there fast. Uh, if you listen a little bit later, I think the answer would be yes.
Anders ArptegSo but now people can choose that they need to
Healthcare As First Vertical
Anders Arpteghave some data. And can you do discovery somehow through your API? Or can you just give a concrete example of how a customer would use Red Pine?
Anders HammarbäckYeah, sure. So the first vertical, the first uh industry we're working with is health. Yeah, it's a huge sector, it's very accuracy-driven, compliance-driven, and you know, great impact to improve uh the healthcare sector from drug discovery and to the optimization of healthcare. And and so, therefore, that's a very concrete case. Um, so there's many healthcare agent companies that are you know well-funded, and I'm very happy to see that. Startup scale-ups, mainly. Okay. Scale-ups, you know, so relatively well-funded that need better data or need more data and to be able to do even more things with their agents. So integrating with them, they need data at point of inference, and we have licensed data from uh data partners who hold various forms of research data within fields like oncology, immunology, uh, respiratory diseases, and so forth.
Anders ArptegSo those scale ups, for example, have some kind of product where they want to help patients or the health uh organization somehow. Yeah. And then or pharmaceutical companies. Okay. Oh, okay. All right. And then that application that they have have some kind of agent integration underneath, and that can speak to RedPine then in some way. Exactly. Okay, cool. Um, then
How Agents Discover And Buy Data
Anders Arptegpotentially can interact back and say, do you want to use this data? Or can you just go through a little bit how does it work in practice then?
Anders HammarbäckYeah, so it's it's all uh MCP driven right now, and we'll expand that in the future. The the agent will then query or prompt uh our API and say we're looking for data with these characteristics. You know, this is the also the context where we are in base. I think context is it's a very hot word right now, but it's also important. You know, so the context and the prompt, we take that, interpret that, and build that into a query into our data stack.
Anders ArptegOkay, so you see more the context they're in, and you try to define what data potentially is proper. Yes. And then you give a set of suggestions back somehow, or it's sort of like that.
Anders HammarbäckWe make a recommendation and the purchasing decision is with the agent, so to speak, or the agentic program. Sometimes human in the loop as well. Sometimes it's an autonomous decision by the agent with a budget set that can then procure an extra money right to the payment.
Anders ArptegSo it's per each transaction, so to speak, a separate payment. Exactly.
Henrik GöthbergI I think this is so important that to just reflect on what you said, is is it human in the loop here or not? And and right now, in the next couple of months, you will see many, you will set the frame of the budget for the exploratory work of the agent, and then it's and then it's not per
Autonomous Spend And Agent Commerce
Henrik Göthbergse a human in the loop exactly that happens. And and that is happening right now, experimentations and interesting it is it is happening right now.
Anders HammarbäckAnd I think from the personal use, separate from Red Pine, we start seeing all this use cases, shopping agents, and you give uh the agent maybe uh $100 or $500 to buy an outfit that um matches your clothes that you like and so forth. And maybe they get it right, maybe they get it wrong. And if you get it right, maybe increase the budget, travel booking, so forth. And we've also seen examples of with that not going so well, and people booking their own tickets and flying everywhere.
Henrik GöthbergBut anyway, I think uh that's an evolution and but the behavior is in the cards that it's it's going in this direction, totally, totally, but but simply by the sheer size and speed that that this this unlocks.
Anders HammarbäckIt makes a lot of sense, right? It's uh the purchasing whenever we as people buy things. That's kind of annoying. The kind of thing.
Henrik GöthbergBut you can argue it makes a hell of a lot of sense, or you can argue it makes no sense, depending on how risky verse you're on these topics. But if you look at technology-wise, that we will solve this safely, it will be done.
Anders HammarbäckIt will be done. All the big uh payment companies and finance companies and banks are launching their own um payment solutions, agentic payment solutions, ACP from Stripe, and everyone has their own agent payment protocol now to integrate with to make uh agent commerce automatically. So that's the ChatTTT has their own instant checkouts.
Henrik GöthbergAnd
State Of RedPine In Early 2026
Henrik Göthbergyou can imagine how strong driving forces these for for the financing companies are different. I mean, like I'm thinking about anything that increases RPU for telcos, anything that increases spend, of course they're gonna build those protocols and make sure they work.
Anders ArptegTotally. Cool. Okay, so you have a working system right now, I guess, or what's the current state? How would you describe where red pine are today?
Anders HammarbäckSo today, again, timestamping, uh February 2026. Uh, we have a working system, which uh I'm very proud of and it works well, working with early customers, early clients, especially healthcare clients, but quite broadly also looking at other uh sectors where you think this makes sense. Um and yeah, is it's getting good feedback, getting you know good um appreciation from the market. And we're scaling now as quick as we can in this early but fast-moving field. Cool.
Anders ArptegIf we move to another topic, I think you you mentioned something before that I found interesting saying something about uh proprietary
Proprietary Data As The Next Frontier
Anders Arptegdata potentially will become like the the ultimate moat in some way. Yeah. What do you mean by that?
Anders HammarbäckI mean, I think it's still an undiscovered space in a way, proprietary data. And the way we look at it, of course, there's internal company data. That's a massive opportunity in itself for a company and enterprise to scale up to understand their own data and set that up in in ways that make sense so that they can use that data in a good way. That's that's a great opportunity. We don't focus on that right now, that's not our focus. But instead, the external data, which is offline from the internet, non-public internet data, and just uh orders of magnitude. So LM's trained to about 20 uh to 30 trillion tokens as a kind of pre-training set. And that's a lot of data that's massive. Outside the public internet, there's up to 100 times more data, which is not public internet data, right? It's almost like we're now on planet Earth only, and there's the entire Milky Way, Vintagotan, a galaxy with inhabitable uh you know place. And then the same with data, like there's so much data out there which is not used for AI, which is great. I mean, great potential. And that's the one we're exploring and uh digging up and making available for data and for AI. So that's why I say proprietary data is the next frontier for so many companies to understand that you know there is a different way.
Anders ArptegBut I'm speaking about the moat concept, you know, and sometimes you know it's uh interesting to think about you know what the big moat for Google or or even a startup is sometimes, and how easy it is to copy in some way or form. But I guess the point with proprietary data is that you can't really copy that because unless you do have access to it, you can't really copy that, right?
Anders HammarbäckYeah, and proprietary, of course, can have different meanings to different people. The way we look at proprietary data is non-public data in a broader sense. Proprietary can also mean data that only I have within my domains that no one else can use. That's one sense of proprietary, but we use it in a broader sense, okay. Basically, data which is not public that I have access to. And even if another company somewhere else has that data, it still adds value to my AI, my algorithms are more accurate, my sources are more trustworthy, and so forth. So it still helps me. So that's where we I think it's the next layer mode.
Internal Knowledge Vs External Supply
Anders HammarbäckIf that makes sense.
Anders ArptegYes, but I think even you know, internally in some companies, if you have an algorithm that can discover and recommend you know what data source to use for a specific query, I can imagine that even you know working internally in companies. Absolutely.
Henrik GöthbergI mean, absolutely. I I do the example where I spend a lot of time with companies like Scania. And I think this is this is they have an in they have a really great moat in their internal data alone. Yes. And it's massive, and it's also a very, very untapped potential because right now it's simply not that easy to make the data move around. They think they have it. Then we come to the next interesting topic here. How much is in the end data, and how much is in the end the whole knowledge layer? Something like so I take Scott as a good as a good example. All the engineering data or IP data is very compressed. It's used to data points. But what you really need is the how did we arrive at that decision that made that data point make sense at that point in time. And this is also, in my opinion, a completely untapped potential of creating data which is simply decision traces and like knowledge. So we have we're having a whole plane of data not even created yet in the old companies. So all this is smote for me.
Anders HammarbäckYeah, I agree. There's so much potential. I mean, you have information, you have data, knowledge, um, the context, the skills, and so what this whole knowledge here is if you can tap into that and make that uh encoded and pro and that's the potential, right? I think for us it starts with the data and then you add um the intelligence to make it into knowledge, and knowledge that it's is is more rich in terms of context and the linkages between different fields and different areas within that domain, but also outside that domain.
Henrik GöthbergBut and and for me, it's obvious that you start with the data you have and you're trying to bootstrap knowledge as far as you can, yeah. But then it's
Defining The Knowledge Layer
Henrik Göthbergsimply how you set up and encode your decision traces that is a never expanding view of data that continues to grow.
Anders HammarbäckI agree, I agree. And to build that into reinforcement learning mechanism as well on the tracing and being able to this is the point.
Anders ArptegBut perhaps you can actually speak a bit about Red Pine's thinking about the knowledge layer. What do you mean with that in and at the RedPine?
Anders HammarbäckSo I I think it means a few things. Being able to also be scientific about how strong uh the knowledge is for the agent's use. That's kind of a problem that we're cracking and and have come quite a long way in doing. Not just sharing data, but actually um distributing knowledge to the agents to be able to do things. You know, I think it helps sometimes to look at agents like people, you know. People, if we're dropped into new situations, we need to understand the context first of all. Who are we? Why are we here? What are we supposed to do? Those kind of parameters set the context of the of the task. Someone's parachuted into a new situation. And that context helps us to understand what data do we need in order to do something that and and what's the value of that. And then I think the pre prior experience of that person or the agent for that sake helps us to decode uh information and turn it into knowledge. If we take a field like healthcare, there's um a rich set of data which is turned into knowledge when we uh connect it to other forms of data. For example, we take patients' history, for example, and you know, important data, research uh data for say cancer research, you know, publicized in leading journals and good academic journals. Then you also have the payment history, you have insurance data sets as well, you have guidelines from public institutions. All of those are in some ways different types of data, but connected together, they form a much smarter
Chunking To Tokens And Ontologies
Anders Hammarbäckknowledge layer about how to go about setting up the future of medicine in AI when there's linkages and good. We talked about before, ontological linkages between those data points. Even you know, case law, looking at the legal field, connecting that to medicine, for example. Then you're going broader, but that makes it more knowledge, knowledge dense.
Anders ArptegSo drawing the insights to build in information about you know what really the data says, and then perhaps even the knowledge about you know how to do use it properly. But do you do something specific to do that at uh Red Pine? I mean, even if you have access to the core data, so to speak, it could be a bit difficult to actually build up the knowledge in the end. Uh or yeah, it could be difficult. You showed me, you know, you showed me a bit of the demo before, so I'm trying to hint a bit about that.
Anders HammarbäckAbsolutely. So we're breaking down the data, you know, to subcomponents and to almost atomic units of data, which can then be put together into new combinations as well, as it makes sense. So it's not just documents here, PDF here, it's really breaking down chunks, breaking down tokens, even the atomic use unit of AI data we see as tokens.
Henrik GöthbergBut if I understand it right then, because if I look at it from uh from the eyes of a more immature customer or client, I almost sense that if you have a very guided approach where you're pushing them to chunk it and to draw the ontology around it, you're actually still guiding them into how to think knowledge, even if they haven't got the whole thing sorted when they start. Is it's that part of sort of how we look at the the the system should help you not only have data points randomly but start sorting them ontology. Absolutely, absolutely, with proper sources.
Anders HammarbäckWell, because people need to trust AI to actually make uh decisions
RedPine’s Three USPs
Anders Hammarbäckon AI. So exactly like you said, uh, with a knowledge layer or knowledge graph on top of the actual data that we have, uh to draw the right linkages between data points and data sets that we have access to, to create something which didn't exist before, and then draw on the right such tokens and chunks for the agents use case at the point of interest.
Henrik GöthbergSo could we could you argue? I'm trying to put uh my words on it. Could you could you argue that the way you are guiding them, the tool should help them to get to emerge and building a scheme or a knowledge graph out of what they are starting to chunk up?
Anders HammarbäckDepends on if we talk uh you know supply or demand in our model at least. On the supply side, I I think that's part of what we do different than many other companies. Um so as a supply company that has amazing data and see the potential of generating new revenue on the data that they already have. In that case, I think it does help to know that we we take the data seriously and we turn it into knowledge for agents. Um, on the demand side, definitely it helps to know what we're doing in the data science layer before providing it to agents, before the agents need to use it.
Anders ArptegIs that even perhaps one of the top uh like USBs, uh the unique selling points that RedPine has to not only be a raw data provider, yeah, but actually to uh refine the data into knowledge channel or opening. Absolutely. What is the main USBs of Redpine?
Anders HammarbäckI think it it definitely is. I think the you know, number one, licensing data sets is you know very important to have access to the right data. That's the fundamental of this, you know, the data itself. And um, sometimes we look at companies who've done the similar journey before, as Swedish Spotify, a fantastic example year old company, licensing music to be able to stream that to uh
Market Dynamics And Quality Signals
Anders HammarbäckAudio listeners and video listeners. So the licensing part of data, number one. Number two, exactly like you said, turning data into knowledge through data science. Number two. Exactly. And number three, a very good integration interface through an API. We looked at Stripe as a good role model in this regard. Stripe had just a few lines of code to integrate with many companies, where other bank solutions had six months of integration timelines, KYC, AML. It's six months later, you still haven't integrated. Whereas Stripe says, you know, here's your seven lines of code and off you go. So of course you win the developer community very easily with that kind of solution.
Henrik GöthbergSo it becomes so when you're building your USP, it's it's it's uh several vectors here.
Anders HammarbäckIt is, it is. And I think that's you know, can be or should be common for many AI companies. It's probably not one individual mode. It's tough today, I think, to have one moat.
Henrik GöthbergYou know, it looks like one moat, but then you decompose it into what makes it lovable.
Anders HammarbäckYeah, I think you need a set of modes that together form uh your company's you know unique edge.
Henrik GöthbergI like that.
Anders ArptegCouldn't you I'm just thinking out loud here, but couldn't you actually turn this into a marketplace situation where they have to bid basically on saying I want to provide my data and sell that to this uh consumer somehow?
Anders HammarbäckYeah, I mean marketplace is is a concept that can mean a couple of different things. There's a market dynamic dynamic in what we do. So is there a marketplace? We we believe in uh selling it to or licensing it to more demand side companies, yeah, as opposed to you know that's the great thing about data, right? It can be duplicated. Yes, uh but yeah, but you can if you have two suppliers of the health data about patients for cancer treatment or whatever, and then you have to choose you shouldn't use both, perhaps, then you could come to a situation where you have to choose and then you can make bids or something on which one to use, or the way we see it is quality should lead the way, you know, quality and the context of the exactly what data does the agent need, and then look at those two and say, okay, this is a little bit better. Um, so that's
EU Data Spaces, EHDS, And Data Act
Anders Hammarbäckat least you know the guiding star for us right now.
Anders ArptegYou might take like a pay train kind of algorithm situation here where you define the quality by the pay track, exactly.
Anders HammarbäckQuality could talk fit. Exactly. But uh, you know, there's uh you're right in the sense that there's market dynamics, the supply and demand, yeah, uh, which is very exciting.
Henrik GöthbergUm I I have a I have a tangent question here. Sure. Because when I listen to this now and I look at this, I can't hold I I get a nagging side question in my back of my head when we're talking about uh in Europe, it's almost jumping to the geopolitical topics. We can park this question. I just want to put it on the table right now. We talk about oh, we need to have open data spaces. What I hear so far is such a naive conversation around that. On on you know, on raw data, uh no understanding that we need meaning, hardly even metadata, and then very raw APIs on that. Yeah. So for me, I think this is the route European. If you because this is also building the European mouth uh moat in a sense, but I think it needs to go in this direction. What would you I mean like we can park this, but I but it's just a tangent here that if you want, I can jump in and or should we do it now? Yeah, sure. It's a good I think it's a very interesting topic on the internet.
Anders HammarbäckI think Europe, you know, uh first of all, like Europe, America, Asia, and so forth. I personally see myself as a global person, and first and foremost, and then you know, I'm happy when our Swedish ecosystem, European ecosystem thrives. But I see it as a global market. And um red pine, you think for red pine and in general. I think if there's a great solution in healthcare education in one market, which isn't Europe, that we can use here and the people can benefit from that, that's amazing. Um, an open source model from anywhere can sort of increase the competition and make those products better across the world. That's better for everyone. But to your question, I think Europe has an unfair advantage in some way by having the market that we do have opening up some of its data. And we see that happening. PSD2 was a you know positive reform in the fintech financial space that allowed for some use of uh data.
Henrik GöthbergUh because we were coming back to what the hell should Europe be good at, you know. And I think now if if we are thinking this is the commodity and this is the remote, then you if you flip that whole story, oh my god, how far away. You know, we need to heavily increase our maturity around these topics and put that back into the open spaces. Open. I mean, I because uh to me, it's the wrong people with the wrong knowledge working with these topics right now because they're not even close to this conversation.
Anders HammarbäckI mean, a lot can be done, but I think there's also positive signs out there. There's a thing called EHDS in the healthcare space, European healthcare data. Um they're going in the right direction. Absolutely. It's it's already enacted by many member states. I think all of them should enact it. It means you know, healthcare systems need to open up some data for
Licensing, Rights, And Revenue Sharing
Anders Hammarbäckprimary and secondary use. Primary is by the patient. You know, you can go to your healthcare provider, get your data. That's almost a given. But the secondary use is you know, as a healthcare company, you can go to a healthcare system and say, I want to actually analyze this data in a positive sense. I want to build a solution on top of this, anonymize it and give it to me. This is news. Not so many people know about this. It's similar to PSD2, but it's for healthcare. Same in industrial data, which is a massive field for Europe. I mean, look at the European production machinery, European Data Act. That's also a new thing.
Anders ArptegThat also is very important, I think, for your U actually has something called data spaces as well, or open data spaces, right? So they're trying to promote that, even for health sector.
Henrik GöthbergMy argument is a little bit like, oh, we are these don't they are thinking right. Philosophically, they're thinking right. Yeah. Practically, they are they need to be way deeper into how it's done. Right. The how I think that's what I'm trying to say.
Anders ArptegIf we really should have uh a data space which works, I mean we should be able to make some money. Otherwise, you know, who will let their very valuable proprietary data go away for free? But if they actually can start to make money from their own data, that could open up for some serious use and value creation.
Henrik GöthbergSo this is a key point. Data spaces, yes. How and what do we mean with open? Open API-wise, open sharing data, but but I think this is really interesting.
Anders HammarbäckYeah, absolutely. I think data is the new goal or oil or whatever kind of resource you want to put there. And and of course, Europe as a region should build great companies on top of the data that we have.
Anders ArptegPerhaps you can get some. I'm not saying you should, because EU funding is horrible to work with, but uh you you should be a good candidate, in my view. In my view, it's the same.
Henrik GöthbergI'm like, I'm like I I I want to throw you into a couple of conversations, like, dude, please start talking to Andesh and the other guys that know what they're doing. Thanks.
Anders HammarbäckI appreciate that.
Anders ArptegYeah, love to take those. Yeah. Uh then uh really cool. Um I think you mentioned, I'm not sure what they are, but you mentioned some partnership with something
Case Study: Aceda Sciences Partnership
Anders Arptegcalled uh, let's see, ACEDA Sciences. Is that the proper name?
Anders HammarbäckIt's a new uh it's a client that we uh are working with, a very exciting company, by the way. Um they they're from Switzerland, but they're also global as a company, building in the biochemical space, including drug discovery and toxicology. Um, and they have uh many different data sources, and then RedPine is helping them with additional data sources through our MPA, uh MCP into their system. Uh so they're doing fantastic work. We just launched that partnership publicly recently. I think it's a good example of the type of clients that we like to work with.
Anders ArptegI think that could really explain what RedPine can do. Yes. And if you just were to describe briefly what they actually do, it's something about like visualization, right? Or correct.
Anders HammarbäckSo data sciences, they uh you know, again aggregated many sources that they have um built up, many of them themselves, through their wet lab work and their domain expertise in toxicology through many years. And the team there has worked for some of the leading pharmaceutical companies and medical companies. And the visualization is the way that they enhance that data by again bringing it from data to knowledge to make people understand how to make decisions, create new compounds, substances uh from that work, working with research institutions and pharmaceutical companies as well. So great company, really love them. And and we're providing data which they don't have access to to create.
Anders ArptegWhat kind of data can you help them with?
Anders HammarbäckIn in that case, one of the sources is published clinical studies uh from various diseases that require proper licensing to be able to be used for AI use in inside their system.
Anders ArptegRight. Okay, but how how do you do that then? Because if they need to have proper licensing in place, have you already written some kind of license agreement with them then that they
Regulation: AI Act And Training Transparency
Anders Arptegor yeah? So that's part of how we access the data.
Anders HammarbäckYou know, we we secure the AI rights for the data so that this data can be used in a certain way for AI. Uh, it could be for inference, it could be for training, fine-tuning, and so forth. So you get an agreement and you pay for it in advance, or or how does it work? Business model uh can vary a little bit. I'm not gonna go into exact details, but I think the key point is that the data owner gets a fair and good and increasing compensation for the use of their data. Um and that's the key point, and that's part of the unlock of proprietary data that we work with. So the more it's used, the more revenue that comes back to data owners.
Anders ArptegI'm hearing some kind of kickback going to Red Pine or something, but who knows?
Anders HammarbäckYeah, I mean, we have a platform fee, and that's no secret. So our business model is you know, the more data that is used by our partners, uh, the more data, the more money and compensation goes back to data partners, and also we're retained part of that fee.
Henrik GöthbergOne thing I didn't catch with the ACI case. Yep. Are you only are they only working with you and RedPine in the sense when it's proprietary data from the external sides? Or when you said you're talking about their own proprietary data from the wet labs and so forth, are they also bringing that in the same way? Or or or is it a demarcation between how you work with internal data, or could you potentially also help them?
Anders HammarbäckSo they are up and running and have their own existing data and have been building visualization on it. Fantastic product. Um, we're providing external data which they haven't yet.
Henrik GöthbergOkay, so actually becomes the interface, it becomes the way how they can enrich what the that's a good way to put it.
Anders HammarbäckYes. Yeah, exactly.
Anders ArptegCool. Um, I'm thinking perhaps one more topic before we go to the news uh session here, but continuing the EU discussion a bit here, we have a lot of discussions about the legal systems as well. Um, and you can think a lot about back and forth, but we have the new like uh Compass Agreement, etc., that trying to actually deregulate a bit in Europe as well, uh which could be interesting. What do you think about like the European strength, so to speak, when it comes to privacy and regulation and the legislation that we do have, can it somehow be beneficial, or is it mainly a negative outside from a European point of view?
Anders HammarbäckYeah, I like the angle of the question to see
Privacy, Secondary Use, And Pragmatism
Anders Hammarbäckthe positives in the legislation. And uh to give a few examples, I think uh part of the EU AI Act, um, which is I think put a bit on hold, but it's probably coming in some shape or form uh later the year. Part of it is I think uh opening up the training of models to understand what's the training set, what's the data that goes into the AI services that you know millions of people use every day? I think that's very important to make sure there's no misinformation, there's no uh you know overly biased information that goes into that data set to make sure that you know the citizens get fair and trustworthy information. So I think that's a good type of regulation that actually says, as a lab generating an AI model, what is the training, like the you know, declaration of content inside that model? It also makes it very clear. I've used this data, you know, is it fairly sourced? Is it fairly acquired, which I think is also a key component.
Anders ArptegI think that's very important and very I mean, I think no one really disagrees with the intention of the EU AI Act. Then the question about perhaps the execution or implementation of it, but still the intention is is really good, right?
Anders HammarbäckYeah, I think it is. I mean, um, and then you can argue, you know, where where do you draw the line, what's reasonable, what's not reasonable. But I think those kind of things, it makes a lot of sense, right? AI is affecting all of our lives, and it makes sense to have some sort of regulation on it. I think that's good.
Anders ArptegAbsolutely. I think no, yeah, very few uh uh I think would say we don't should have or shouldn't have any regulation. That would be really, really dangerous.
Henrik GöthbergI think but let us test one angle we have talked quite a bit about on this pod when when we had these conversations, is that I don't know where you're going.
Anders ArptegThat's yeah.
Henrik GöthbergI'm not sure. We I don't know where you're going. One of the theses has been I want to test it on you. It's actually not the regulation per se or the ideas that are the problem, it's the legal uncertainty or the lack of harmonized standards. So people, engineers,
AI News: New Models And Agentic Tools
Henrik Göthbergdoesn't don't really know what it means to be compliant or what is the right view of doing this. Would you agree with that? Because I I think even putting uh this on hold and pause for me is a little bit like yeah, you you released regulation, but you had you also had no there was legal uncertainty because you hadn't solved it for the engineers to know what the fuck to do. Yeah. Well, what do you think about that sort of thing?
Anders HammarbäckI would agree that certainty is always better than uncertainty when it comes to legal. Uh so absolutely. And and I think now we're in some sort of limbo. I think in the meantime, I mean, builders build and and companies keep innovating in the meantime with this limbo, and then you know, hopefully we get some more clarity over time. Um so I think in some fields like credit decisions and and so forth, it is probably a problem because you don't want to invest too much before you know what's going to come out on the other side in terms of regulation. So I think it is probably a problem. I like I like the the signaling from Brussels now about the EU Inc. kind of standardization, harmonization.
Henrik GöthbergThis is cool.
Anders HammarbäckIt is cool, you know, to have one legal form and maybe one IPO standard and quicker time to build.
Henrik GöthbergTaking away the the downsides of being a multifaceted market and making that something. Exactly. Exactly.
Anders HammarbäckAnd then I think the the upside of it is of course the cultural diversity that we have. So having the values but taking away the frictions.
Henrik GöthbergExactly. Exactly. I think that's that's good.
Anders ArptegAnd another angle that we've spoken a lot about when it comes to the EU AI Act is um, you know, should you regulate the use case or should you regulate the technology? But in this case, perhaps we could add a third angle, which is should you regulate the data or the use case? So if we take healthcare here as an example, you have data that you do provide, we could choose to say that no data about uh cancer treatments should be allowed. Then you regulate the data itself. Or you regulate the use case. You know, if you use the cancer data for this, then that should be regulated to do that properly. How do you lie on that kind of scale? What should we really regulate? Should it be the data, the tech, or the use case?
Anders HammarbäckI think the privacy protection is important, first of all. The individual's privacy protection uh is extremely important. We're early in this AI AI field, and having that privacy protection, I think, is really key. Uh so for us, any data we use is anonymized and it's it's not
Claude 4.6, Codex 5.3, And Co-Work
Anders Hammarbäckthere's no sensitive GDPR data, so to speak. Uh so I think that's good. Um, no matter if it's about cancer or if it's about other forms of data and legal data or so forth, it should definitely not be privacy sensitive. I think it's super important. So I wouldn't say that it's just the use case itself, this type of research and that kind of research. I think just having some sort of privacy filtering, then I think it's probably gone too far in some ways with the privacy. So you have to, it's an unreasonable amount of work that needs to be done to clean up like any small signal or indication that could potentially indicate some sort of linkage, even for information that is already public, so to speak. So yeah, I think we can pull it back a few notches, but we still I think need to keep some sort of privacy filter on it. That's key. So like GDPR, you know, maybe we can bring it down to 70%, 60% of where we are right now. We'll probably be good enough to kind of protect uh the kind of core privacy aspect.
Henrik GöthbergSo be very vigilant with the core, but then understand what is uh work practic pragmatics. Yeah, exactly. I think so. I think so too.
SpeakerCool. It's time for AI News, brought to you by AIW Podcast.
Anders ArptegCool. So let's do a quick uh or hopefully quick news break in the middle of the podcast, and then we get back to the discussion here with uh Anders as well. Um so we tried to discuss a few uh interesting highlights from the recent week's news. Uh, each one of us. Uh do you have something, Anders, that you'd like to bring up, or should we You get started. Yeah, and I will uh add. Henrik, or should I uh you can start this time. Well, I think the the obvious one this can't uh this time is uh yeah, once again a new model comes up and it's released, and it has new amazing functionality and and uh performance. But still, I I think it's interesting with the codex 5.3 that was released from OpenAI. One hour after uh anthropic claude uh opus 4.6, right?
Henrik GöthbergSo so we are actually have we talked about 4.6 even. No, no, it's so there's then there are two
Computer Use And The PowerPoint Test
Henrik Göthbergnews then.
Anders ArptegYeah, but I think the easy combined, you know, they were released more or less exactly the same. Okay, of course, from a very strategic from a marketing point of view, uh they have exact inside information when each of them will release, and then they do it very, you know.
Henrik GöthbergSo this my question was for the ones who hasn't been following every single so we have actually anthropic and open AI releasing versions in an hour.
Anders ArptegYeah, yeah. But we've seen in the past, you know, when Google released, you know, OpenAI went out like also like a day after or something and completely debunked and made the whole news from But this is almost choreographed. I mean it is insider information. They of course have detailed information about you know whenever they will release something and they purposely time the releases to to uh you know to hurt the other competitors, yeah. Is it through a proper like AI race, of course, here so but cool. I think both were surprisingly interesting. And um, if we go a bit through the two uh models here, uh you know we've been saying for a long time that 2025 was the the year when OpenAI lost their throne, so to speak, in being the Frontier AI lab. And I think they still are. But 5.2 was surprisingly good, and and even this Codex 5.3, which is not, I don't think, public yet, but they released some results from it, is really cool. So of course, everything is moving from like standard like code generation, it's a lot of focus on coding as usual here for these models, but moving from just you know normal code generation to, of course, agentic uh tool use and being able to have this kind of orchestration of agents built into the model itself. So both models is really going that route, uh, so it's very clear path forward. I think for Codex 5.3, they actually used Codex 5.3 to build Codex 5.3, right? So it was truly this kind of um, they used their own model to build itself, which is kind of cool. So um more so than in the past. But even how they trained a model, how to manage the GPU clusters and how to evaluate the results and how to steer the training for that model was done by the model itself or the other versions of it, right?
Anders HammarbäckIt's similar to Claude, at least you know, publicly announced that to build Claude 4.6 and the co-work module as well. You know, Claude built it with human oversight and support, but you know, that's the new uh norm. And of course, we will never know exactly the percentages that is built by those models, but pretty clear that's that's the trajectory is there. For sure. For sure. And we saw in Davos, you know, Dario said you know,
Startups Operating AI-First
Anders Hammarbäckwithin six to twelve months it will build all the software.
Anders ArptegYeah.
Anders HammarbäckSo whether true or not, but anyway, that's the trend.
Anders ArptegBut it's a clear trend for sure. Very much. And uh, you know, it it never will be trained as little as it's done today by AI. And in future, it will be majority trained by itself, right? So it's that's really cool.
Anders HammarbäckI thought, you know, if I can add, yes, I thought the for the Anthropic launch, the two things that were non-coding were almost the most uh you know um eye-catching. Um the the um legal module and the financial model. I think that's super interesting because even coding has been leading the kind of research for the labs until now, but all of a sudden we're going into new verticals of automation.
Henrik GöthbergYeah, please go here because I think this is also one of one of some of these like LinkedIn blew up. Like uh, did you you see what happened? Sauce is dead, and like 40% down on the on the SaaS uh uh software you know app in those agents, and what and we now see a trend where they where you get all the way into a vertical. Yes, and that is so that is reshift shaping the whole sauce ball game completely, and and for companies, enterprises as well.
Anders HammarbäckYou know, everything from how to staff teams and and what to invest in, what software to buy, everything is shifting very, very quickly. And I think the you know the legal module of Claude looks super. Interesting, right? The plugin they call it, but it's just the start, the financial module as well, bringing cloud into existing software like Excel in PowerPoint.
Anders ArptegAnd integration, right, to other kind of finance systems and the same, right? If I understand it correctly.
Anders HammarbäckThat one I haven't seen, but at least in in the uh kind of spreadsheet softwares that you know hundreds of millions of people are using every day, a lot of that work can and will be done by AI models and AI agents.
Henrik GöthbergSo I think that was a big but I I this is so mind-boggling. Let's take an example. Let's say I was used to Tech Arian and talked to a head of strategy of a large Swedish company who is sort of in a in a hyperscaler mode. Yeah. And
Think First, Then AI: New Team Norms
Henrik Göthbergthey are thinking carefully about how to do how to now in the same time get their ERP systems in place or get their the stack in place. How to think about this topic now? I mean, like because I don't think it's like we can replace the core transactional systems, but you need to be uh very careful now to modularize everything so you have replaceable components. This is this is you know if you're in a in an ERP shift now, and unfortunately, when you go to the enterprise space, it's not the same people that knows this stuff that is working on those types of uh projects. But right now, how to look at that becomes I don't even know where to start, how to unpack it.
Anders HammarbäckI mean, uh SAS companies, software companies are quickly becoming agent companies. Salesforce is agent force as part of their work. Microsoft is an agent first company, I think they said in in one of the announcements. Or you know, like an agent. I mean, I I think in general the SaaS industry is moving from just a seat-based price per month model into agentic pay-for-use, pay for value kind of model. It must be like that. You know, instead of having a software suite sitting on the local hard drive, get an agent that solves the task, period, no matter if you have the software or not. That's what you're going to pay for. And I think these launches, back to your news section, is just a precursor of a much bigger shift.
Anders ArptegAnd I think the the classic like uh slogan of you know, we won't have AI replacing people or companies, but companies and people that use AI will replace the companies that don't. Yeah. So if you don't move into the space, totally.
Henrik GöthbergThis is the point, right? Because in the last enterprise in the enterprises with the proud analog history, they are these are fundamentally different teams that are working on oh, what should we have in SAP or you know, those decisions versus who builds the modern data stack? Exactly. And this is blurring, and also the whole point with the they don't back to what you're doing, they're not differentiating with this is compressed data in the transactional system. This is not knowledge. You cannot get knowledge from here. Exactly. You even if you wanted to, you can't squeeze knowledge out of that. It doesn't sit there. On what the data is that is not what we're talking about. We're talking about the knowledge traces. And and and so you need to converge in understanding on these topics now. Otherwise, you will, you know, what is an AI strategy and an IT strategy all of a sudden got blurred.
Anders HammarbäckTotally, totally. It's moving so fast. And I mean, I think the open claw, maybe you talked about that in a previous uh news session, could be like that's also massive, right? Yeah. And how that's just an early indicator of what's to come on the agent side, on the autonomous agent side.
Human Strengths: Communication And Synthesis
Anders ArptegBut another I think interesting aspect about uh Codex 5.3 is it's actually not just for agentic coding uh or code generation, it's actually for computer use point. And the reason you know, I've been a bit outspoken a bit about how bad agents are and how bad they are at taking actions. I agree. Yeah, yeah. Because the use case I've been saying is you know, an agent can't even make a PowerPoint today. Um you can't really replace a human because you know every human more or less unfortunately needs to have PowerPoint at some point. No, horrible example. I mean, I agree, it's a big part of it, a big part of the work for me. But it should be a simple thing, right? Yes. I mean, a PowerPoint is not a hard thing to make, but AI is really bad at it. Yes. And that's kind of surprising. Now it's actually moving in that space and specifically into the computer use space. So actually, uh Codex is finally getting some good scores here. So they achieve like 65% accuracy on this kind of OS world benchmark or something like that. Yeah. And I think humans are like 72. So it's getting close, almost close to human performance now. And what computer use means is basically using keyword and mouse, right? So if you don't have like direct API or MCP access to do something, you need to use the human interface, which is computer and mouse, right? And and for humans, they they of course can do that and build a PowerPoint. But once an AI can do that, then suddenly the possibilities for an agent will significantly improve.
Henrik GöthbergYeah, because if if you use then so the inst the interesting thing here is that you're you're you're getting to these network effects where you simply get to a critical tipping point on one feature and all of a sudden that unlocks a new level. Yes. Which which then goes really fast. Super fast. Super fast.
Anders ArptegI don't think I don't think people understand. I mean, some people say, well, why shouldn't AI use a human interface? Why shouldn't they have their own? Of course they should have their own, and that will happen in many years. But if we want to quickly unlock value and be able to take a lot of actions, they should use what exists today, and that is the human interface. Just as robots need to operate in a physical world with uh the human kind of size. That's why we have humanoids being built everywhere, right? It's for that simple reason.
Henrik GöthbergBut how fast will this happen now? Because it's one, it's always back to the it can be done in the lab. But but some of these things, if if if you take the PowerPoint example, when that kind of works, it's up to me when I do my PowerPoints. So it's not really it's out of the lab very fast. Or or I mean I think look at uh the Google.
Anders HammarbäckWe as a startup you work a lot in Google and the Google Suite. And uh I think it's just in one year gone from,
Cross-Discipline Work And T-Shaped Skills
Anders Hammarbäckas you said, not very nice looking presentations as a former consultant, you know, high bar on uh PowerPoints. I gave this life for the best. Still have it as part of your toolkit. But anyway, uh I think now it's much better. Still not, I think, to the human level at least, but moving very fast. Trajectory is clear. So and with the you know, open claw regime of agents, yeah, it's just gonna much, much better.
Anders ArptegIt can flip very quickly here. So that's that's kind of interesting.
Henrik GöthbergVery interesting. But I think the the point is this it flips quickly, but then you're gonna have the fundamental divide between uh which is which boils down between uh the technology invention, exponentiality, and the very linear adoption, invention adoption capacity curve or whatever you want. So you will you so this what happens now? The the the competitive threat area, the surface area for competitive threats. We're gonna go into this innovation pressure tipping point where it it can go faster than we think now.
Anders ArptegAbsolutely. Cool. Still, if I were to rank the two, uh it would be fun to hear what you think, Anders and Enric, as well. But um, if you go to El Marina and look at you know who is in the top there, uh of course Opus is now in the top uh compared to everything. I think it's even number one and number two spots. So they have this kind of high thinking mode, and then even the second one, which is not high thinking, is still 4.6 in the top, and then Gemini and then Grocki, I think. Um they I don't think uh Codex 5.3 is in there yet because it's not the public. We don't know the numbers yet, yeah. Yeah, so it's not public really yet. So so it could be, but still it's not there, but still Opus is amazingly good, but it's very expensive, and uh and uh yeah, there are other purposes or negative points, I would say, but still for coding purposes, I think it's hard to beat. I would say opus. What do you think?
Anders HammarbäckYeah, I mean I speak to my my developers on the team all the time, and I think uh you know Claude is is a big favorite um still, and also now with the cowork module, it's actually spreading to other parts of the team. Yeah, so I think that's the that's a big shift.
Anders ArptegYou know, it's been a you know developer-centric uh type of you know, with the cowork, people don't perhaps realize it, but it's actually has its other app or other way of working
Reasoning, Innovation, And Organization
Anders Arptegand integrating with a desktop. And actually, OpenI did the same. So I forgot to mention that, but they created something, uh, what was the name of it? Uh some app that runs only on a Macintosh show. Yeah, I won't try it because I don't like I don't like Apple. But still, they have an app um that can basically show how it's thinking and it can access the desktop and uh even it can do like interactive debugging properly. So if you are doing coding, before it couldn't really control like the IDE, you know, it just has to produce code and run something, but it couldn't really do proper debugging, but apparently this can. And I think it opened up so many new avenues of how an agent can take actions when you have these kind of extra applications.
Anders HammarbäckWasn't that the OpenAI agent? I think they call it. Is that the one you mean? There's a part of OpenAI's uh ChatGPT suite called just OpenAI agent or chat GPT agent.
Anders ArptegNo, Codex app is called it's like a command center for managing multiple agents.
Henrik GöthbergBut but how how are you starting how are you yourself thinking about this as a startup? Yeah, how how deeply have you gone in in AI-enabled coding right now and and and or whole operation enabled? Like because one of the benefits of being a startup is that you don't have the legacy, you can you can try to think and be uh surf the wave now. So are you doing that or is it still you need to focus on doing the traditional doing startup work?
Anders HammarbäckWe have to be AI first and we want to be AI first, and it's a fantastic opportunity to start with a canvas, blank canvas, and how do we design the teams in programming in in uh computer science and and software engineering?
Henrik GöthbergSo you're setting up teams from this perspective from the beginning, thinking we need to be we should be better the better the AI agents get. We shouldn't be in need to reinvent our process.
Anders HammarbäckDefinitely like that. So that's part of our core value. Um, to really be we say AI first, but I think one caveat that we have built in also is to do human thinking first and then AI second. Because it's so easy, and I see that all the time, not just in programming, but across AI slop and people just hardly think, just put something in the AI, something comes back, hardly look at it. Maybe it's awesome or it's terrible, but you know, I think it's still a value of doing the thinking first. Like what do we want? How does the prompt look?
Henrik GöthbergNo matter if it's so has this been sort of almost like a heuristic or design principle, or have you thought carefully when we say AI first? We want thinking first and AI first.
Anders HammarbäckThink first, AI second. And I think that goes a long way still. It might flip in five years. Where but right now I still think we have so much context, intelligence, knowledge
Agent Swarms And The Hive-Mind Idea
Anders Hammarbäckto put into this.
Henrik GöthbergBut it this resonates so well when we talk about which coders gets the most benefit out of the AIs, and it's it's the best communicators, it's the best so you the leadership skills of communicating are delegate, be clear on your goal, and which is the thinking, what do I really want before I get the shit out.
Anders HammarbäckThat's right.
Henrik GöthbergThat is that is a real important thing.
Anders HammarbäckI think it is, and no matter if it's about writing or you know, visual development or coding or it says something on the new leadership skills, and it says something about what we should train as humans for sure. It really does. I'd love to go deeper after your new section, and we hijack the new section. But now we can go deeper in what it means to be a professional and a person in the HVAI.
Anders ArptegBut let's table that for another write it down here. Yeah, um before we go there. Uh Goran, do you have anything you want to add or anything?
Goran CvetanovskiUm I had three, but uh I will make them very short because we don't have much time and I can need to leave at seven. So yeah, yeah, right. So you have the the C dance uh 2.0, the new video uh released model, and I think it's uh fantastic. Uh then there was a new what is called like a report about AI um ethics and governance. Uh over 200 actually scientists, uh including Joshua Bangio, who was leading actually the panel, created a new report about the dangers of AIs, which is super cool. Uh, there are many things in the past two weeks that actually has happened, right? It was also the ancients talking to each other, yeah, but we missed the bots and but we can take that in another time just to enjoy a little bit uh more their story.
Anders ArptegSounds okay, or do you want to add something?
Henrik GöthbergNo, I I I'm actually quite keen to segue on the on the on the angle you you went with. I think this is a nice angle and and start zooming out a little bit. What do you think, Alan?
Anders ArptegYeah, no, let's go there. Perhaps you can frame it, Rick.
Henrik GöthbergNo, I I I give it back. You you were doing the framing perfectly. So take take you had your you had a train of thought. Take us back to the start of that train of thought.
Anders HammarbäckAnd I I think it's
Toward Scientific Discovery In Three Years
Anders Hammarbäckprobably something that all of us who are you know working with or in AI are thinking about. You know, on one hand, there's immense excitement about the opportunities of coding, of writing, of creative and development and so forth. On the other hand, there's that, you know, hmm, what's gonna happen? How is this transforming industries, jobs, professions, our kids' future? That question my identity, my strengths. What is my strength? Exactly. What what are the jobs of the future? What's gonna happen, job displacement or not? All these questions I think are are all around us all the time. Seeing you know, back to your news about these new modules. And I spent some time thinking about this as well. And I can give my humble two cents on this topic, which I think I'm an optimist at heart, and I think we have to be. And I overall I'm I'm extremely excited about what AI would do to so many important fields: climate change, health, education, right, material, and so forth. So that's sort of putting that aside from an I do think that what it means to be human and the value that we can bring into the professional space, especially, I think it's going to be um very important to follow. And and and I do see you mentioned before that the best programmers are the ones who communicate and so forth. That's in a way cross-disciplinary skill. You have the communication skills, you have the programming skills, and then you can leverage AI to be better, to be more productive. And I think if you take vertical by vertical, like industry by industry or job by job, I think it's the same. So, legal, for example, the best lawyers that I've worked with throughout you know many years, and probably the same, are not the ones who know the law, the book of law, the best now communicate and pedagogically get you to into the topic. Communication, but also strategy and business. They understand okay. A legal field is also about risk balancing. There's the finance, there's the industry domain expertise. I can pull insight here. I've seen that example from the past. I can pull data into knowledge, and I can use that to give a recommendation, which is different from just citing from the book of law. And I think that's the the very best in each profession. Have that sort of cross-disciplinary human touch, which is still very difficult, still very difficult. In the future, let's see what happens.
Anders ArptegBut in the next human touch, I mean, I would argue that, you know, I think, of course, communication is the if you see
Online Learning And Real-Time Inference
Anders Arptegsomeone being really good about communicating, about coding or legal, or marketing, or whatever, that's a very positive sign. For one, that they can communicate. Secondly, they actually need to understand it to be able to communicate it. You can be the best communicator, but if you don't understand the topic properly, you it doesn't matter if you can communicate or not. So if you can say something profound about something complicated, that's a very good sign. Yeah, that's a good step, right? So even Einstein said something famous, like, what was his quote? Um, you don't master a topic until you can teach it or something. And I I think it's a lot of sense in that kind of statement, right?
Anders HammarbäckSo that's true, but also drawing things from different disciplines and make sense out of it, not from the most obvious angle. Investment, like take investment, I think the best investors are not the obvious investments because everyone else makes it, the logical ones everyone will jump on, but the contrarian ones, the ones who see things differently from a different vantage point. And it's not sure that just because he increased the temperature of an LLM, they will do contrarian investments, right? So I think seeing things, I've seen that before in that industry. I take that into this context, I place a bit. I think those are the investors.
Henrik GöthbergLet's really, really try to put the word on that because we have the communication skills. Yeah. And then you say we need to have communication in relation to a field, so it's actually knowing it to proper intelligence to understand it as well. Yeah, this is step two. Now you said the third one. And let's really try to put a good naming on that because sometimes we say, Oh, you need to be curious. No, no, no. What you're looking into are people who are deliberately seeking out uh intersectional innovation or meshing things. So it's this fundamental ability to mesh different concepts. Yes. So not disruptive innovation, intersectional innovation. So it's something about then is about seeing how these different things pull together. So this is is it synthesizing? No, it's it's it's it's taking in and then synthesizing, taking in different perspectives.
Anders HammarbäckI think so. It's almost like a chemical reaction. You bring in different compounds, and what happens when you mix them? It's something new that that has not been foreseen. I think it's the same in creative industries as well. You know, writing, music creation, you know, art and so forth. It's not the most obvious things, the most naturalistic paintings that are the most sought after, but maybe sometimes the most provoking and so forth. And I think that's a human element that humans have.
Henrik GöthbergBut but but if you follow that line of thought about taking different
Beyond Transformers: Architectures Ahead
Henrik Göthbergthings in, it leads to another very different professional skill set where you can see a difference in the divide, who has it or not. And what we are talking about now is the fundamental ability to work well in a cross-disciplinary setting. And to me, there are it it goes back to the fundamental understanding of being T-shaped. Yeah. So we are creating a lingo and a joint purpose that are bigger than ourselves, that are bigger than, you know, I'm I'm not my identity, it's not my task. It's our purpose, and within that, I'm confident in my discipline, but I also have a language shared with completely other disciplines that makes the team brain index bigger than the one. So that goes together with we want to then I meet people who sort of they have a hard time outside of their comfort zone of whatever uh topic they are, and they they it's so obvious they're growing up in a division of labor type organization. And when I put them now, we're gonna be a product team. The most fundamental problem is that they don't have a shared language. I don't I don't have a shared respect even for the other person's competences.
Anders HammarbäckYeah, no, good point. Good example. And I I do think that's gonna be super key, and also to bridge the divide between agents, AI, and people in the future, you know, going back to cross-disciplinary. We have you know the agents doing things and the people and making that all work together requires a new type of communication skills and collaboration skills as well. Um, so yeah, I think there's um that we're jumping a bit around in the topics here.
Henrik GöthbergGood idea.
Anders ArptegOne is the more like cross-disciplinary kind of um understanding or innovation that we can have, and then we can argue if that's a human skill or not, or if it's something else, and then we get back to what you originally said, which is the excitement or the concern potentially about you know the progress of AI and impact on human market or lot job market, etc. But just before we go to the job market, I just want to see you know, is this a human skill episode or not? So just to try to elaborate a bit more there, um I would argue today you you are right, that it is a human skill to be able to find these kind of more abstract kind of connections, to see that ah well, actually from the health sector, we can actually see that the data science algorithm there works in this use
Neuro-Symbolic Via On-Demand Software
Anders Arptegcase, and they can find these kind of high-level abstraction connections, which is really hard to find, and that's the true tell sign, so to speak, of a very senior and knowledgeable person, right? And I think AI today can't do it, they are extremely good in memorizing, so you can put like 10 bucks, of course, into Gemini or even into Opus now, which has one million tokens window as well. Um but it can mainly you know find uh data from that in a shallow sense. But it does that insanely better than humans. That's right, right? So so just being able to recall from a large piece of data is something that AI is significantly better and actually a very valuable thing to do because humans are horrible at it, right? Totally. But the reasoning ability, I think, today, to be able to do this kind of high-level abstraction connections is something that humans do better. So so I do give you the point, it is a human ability today.
Henrik GöthbergAnd but you're putting a timestamp on it.
Anders HammarbäckYou know, who can predict in five years when we sit here? Maybe I'm invited back, maybe not. But at least when you do the AI news in 2031, you know, about human skills at that point. Yeah. Then I think we'll be talking more about robots and and dexterity as a human skill, you know, a human, unique skill where AI and physical AI are not yet. You know, maybe at at your companies you will have, but to be able to lift up the glass without crushing it, and those things are. Then I think the the big focus.
Henrik GöthbergBut if we t if you if we go back, we had uh the first part this year was with Karim at six. I don't know if you know him or you know better. I don't know him, but I heard the pod. Yeah.
Anders HammarbäckBig fan of the partnership.
Henrik GöthbergYeah, yeah, and and then then ultimately then we we are talking about I mean his view on AGI and reasoning and navigation and all that. Where he basically thinks it's it's it's a core pattern, and then and then what we're doing is we're actually moving in complexity of the fundamental pattern. But um if I'm if I'm taking that in, that would would that sort of be would that lead to the point where the reasoning and this sort of pulling things together in the end is done better by the AI? Is that what we you know if if you take that trajectory and towards the AGI view on this and and we so it's like the knowledge layer, and now we go up in the in OpenAI's model to reasoning.
Anders ArptegYes.
Henrik GöthbergAnd when we hit that mark, we're probably not gonna be better
Closing Reflections And Near‑Term Outlook
Henrik Göthbergthan AI there either.
Anders ArptegNo, but I think still the top layers in this pyramid. So we're speaking about this open AI pyramid. So they like it two years ago or something published um an ATI path. Like then we said it had the first the layer which is conversational or knowledge. Layer five. And then had reasoning in level two, and then autonomous, basically agents in some sense in level three, and then we had uh innovation, they call it, and then organizational.
Henrik GöthbergThose two are a little bit more blurry for me, yeah.
Anders ArptegBut they mean the top layers. But yeah, but yeah. We can go there if we want to, but I think you know, it I think it's a rather good, it's not a perfect model, it's kind of old, it's many years old.
Henrik GöthbergYeah, but it allows us to reason about this and it allows us to distinguish which human capabilities are important. I agree. In l in in within the timestamp of where the AI maturity is. Correct.
Anders ArptegBut just to go there a bit, I mean I I think it's very clear that you know for autonomous level three, we are of course seeing a lot of progress in AI being able to take actions, but it's primarily through APIs and MCP uh interfaces that we have pre-built as humans for it. If we didn't do these kind of pre-built interfaces, it would not be good at taking actions. Now we are seeing computer use through uh Codex 35.3, it's starting to approach human skills, but it's not there yet. But it will, of course, improve a lot. Yeah, but it's not there yet. No. So we will see autonomous layer number three increase improving a lot coming years.
Henrik GöthbergBut we don't think people realize the big shift that will no, you know, and and and and what I my my thesis is that this will be completely blurred because from wherever we are right now, which is sort of the knowledge layer, we will uh someone's someone I can't remember who it was on Silicon Valley and said the the the AI edge is jagged. So what I mean with that is that like we we're not jumping cleanly up like this. We are we are we are we are spiking up here, and all of a sudden we are spiking on the agentic frameworks, which is so so all of a sudden now it's gonna be a mismatch of progression in the reasoning layer and in the agentic layer. And and some some of it right now seems like the agentic layer might confer before the reasoning layer. I don't know. So, but you see what I mean? Like so that's that's that that is then that's why I think it's gonna be a pot a really catch-up effect. It's gonna be what it's gonna be to the point where you have a tipping point where you level up. I agree.
Anders HammarbäckIf I can push the email up a bit on that point that you made, I think the uh you know the collaboration skill, coming back to what what can humans do, what can AIs do, uh and robots in the future. I think collaboration is still in a uh caveat, but I say is still a human distinguishing piece that I think you know also make us distinguish you know in the you know human race versus animals. Like why are we where we are today's race?
Anders ArptegThat's actually level five, the organization. Yes, because that's my view a bit about what it means. It's really organizing together to form like a company, yes, or a country, or a society in some way, and that's something that humans actually do very well.
Henrik GöthbergBut this was great because now we now we now we could uh comportmentalize uh the conversation because the reasoning in terms of seeing, getting signals, this is one dimension of cross-disciplinary or taking that in. That is that that belongs to the reasoning level. But collaboration most likely then be belongs in the organization innovation.
Anders HammarbäckI think it does, but I mean what we saw two weeks ago when OpenClaw you know spun up millions of agents in just 48 hours that started organizing, you know, at least they reportedly they created religion, they created companies, they created social media, they created uh uh you know their own cryptocurrency. Starting to see that, right? So that is and that to me tells me a couple of things. Like, first of all, you know, that's an organization layer maybe happening before number four. Uh, second, uh, what if that is really gonna have the biggest impact? You know, we've been looking at soda scores, soda scores, state-of-the-art scores on one reasoning model by after another. What if the real impact in AGI is to have a total organization and community of agents? Yes. Look at how people and the human uh species have evolved. You know, yes, we invented fire, we invented uh tools, tool use for agriculture. So the tools were important, but the ability to spread those you know, created a whole new society.
Anders ArptegBut the capability as a society that we have, you know, no single human could ever create an iPhone. Or that bad example, a Samsung or a Google Pixel or whatever. Uh anyway. No, but you see your point. So no single human have a skill that a set of people do. So an organization like Google or OpenAI or Red Pine, that organization have so much more skills than any single people do. Exactly. So that ability to form and collaborate and group together to do something together is actually much bigger than people realize it. I think so.
Anders HammarbäckYou know, taking that as a corollary to the future and looking at what if you know millions of agents can become billions of agents and they organize and everyone has you know good intelligence. Is it you know autonomous AGI for each of those agents? Maybe not, but it's still pretty damn good. And that put together as a kind of value layer, you know, can really build meaningful things we don't have.
Henrik GöthbergBut we we have you also talked about this before then, because how how do we envision this? Do we do we envision it as a huge brain central intelligence, or what you what I I would refer to this as a hive mind? Yeah. So this is what we what we are talking about here is also then how do we define when we reach AGI and ASI? Well, if I look at the one component of the hive mind, it it is on level on one level. Yeah, but when you look at something that is brilliant at organizing itself, it it it it the brain index, so to speak, is one plus one becomes five. That's right. Is that what we that's that's the sort of conversation we're having right now?
Anders HammarbäckYeah, I think it's an open plane.
Henrik GöthbergBut it's an interesting train of thought. Super interesting, super interesting.
Anders ArptegAnd people think sometimes it's super easy to organize, for example. They can see, oh well, I can make uh that agent speak to that one. This communication and collaboration is easy, but it's not. No. And if you take the innovation, the level four here, uh, there are actually some interesting experiments being done here. They're saying basically, let's remove all the data up to uh 1905 and then see if an AI or a set of agents can recreate the special theory of relativity that Einstein did. Now that would require a significant like ability to reason and innovate at a higher level of abstraction than AI ever could do today. So even though it can actually do a lot of great memorization, which is very useful, the reasoning and the innovation and being able to think outside the box here that Einstein did is something that is much harder than people realize. Yeah. But it will, of course, come there at some point. We will. I think in the next three years.
Anders HammarbäckThree years, I think. Scientific discovery of high level, I think three years.
Henrik GöthbergYeah. Do you think do there will be a tipping point? You have do you have this mindset in terms of it this will be vastly different in three years?
Anders HammarbäckI think so. I think AGI is uh you know kind of a blurred line when do we hit it, when do we don't, and different definitions. Is fiscal AI part of AGI when we can do you know lab experiments and so forth? Yes or no? But I think in three years we will be able to do scientific discoveries and and hopefully we can have a dent in that curve through Redvine's work as well, providing better data to companies that are pushing the boundaries in that field.
Henrik GöthbergBut okay, so this is super interesting because if we look now, we have a if we if you if we extrapolate on this as an example, sure. Within three years, we uh the thesis within three years we're gonna have massive uh uh scientific discoveries. Yeah. What are the core things missing or the core things? Like you're saying something we are here now, in three years we're here. What are the key things that happened or that succeeded for that to materialize?
Anders HammarbäckI mean, I I decide to dedicate these years to the data layer. So I think that you know a lot is to be um won by excavating and unlocking the best possible data that already exists to be able to make that available cross-disciplinary for the benefit of having AI's super skills play at its advantage. And I think asking the right questions is part of that as well. How can we ask the right questions, use the right data to come up with alternatives, iterations, alterations, and test those quicker in the computer, so to speak, and then doing it in reality, like in the labs. So um I think data plus questions and and it goes back into combining the best of humans with the best of AI's algorithms and the best of data in the near term before AI can help jump up a few notches.
Anders ArptegDo you have a preferred definition of AGI, by the way? I usually I can I can share mine to start with. Yeah, and then I can share mine again, but we go first. Okay. So I actually like a Sam Altman definition. Um, he changed it a number of times, but still I like the old one, which is basically we will have AGI when AI can do the level of work that an average human coworker can do. And that's you know, average human coworker can do a PowerPoint, so it should be able to do that, but it can certainly not do that today. But when that happens, and I actually agree with you, three years is actually what I think as well. It's similar to what Ray Kirchfeld said.
Henrik GöthbergYeah, we're we're back on that number. 2029, you know.
Anders ArptegSo it is actually three years from now. So that's uh what I've been saying also for for a long time. But we'll I think it still holds. I think that's we will see a lot of stuff happening in the coming three years. Well, what do you think? What's a good definition?
Anders HammarbäckYeah, I think you know, one of our team members, Fahad, um, I spoke to him um about this as well. I think when when it's similar to Sam Altman's definition, when um the computer systems really can resemble the work of humans across the different fields, then I think we're close. Then uh if you lift in you know, Dumis and a few other people, or I think Andre Capaty also talked about the physical aspect of it, you know, the physical AI. Yes, you know, even if the knowledge work piece is similar in three years, yeah, how good are the robots? I I hope they're really good because I look forward to to that paradigm as well. But I do think it's gonna take a few more years than three to have them as performant and and capable as people are today, maybe five, but uh not much more.
Henrik GöthbergBut I just want to circle back so we you the the data aspect will be a key part. In order to reach this, I'm not saying we have AGI, I'm we were you saying that we're gonna get to this uh major breakthroughs or like something bigger will happen. Will we need to have had um online learning to get there? Do we need online learning as a fundamental step in these three years? Online learning in what sense do you mean? I I I mean I mean in the sense that we we we are not doing patch learning before or training before, but but the model continuously recursive recursively learns all the time. Is that what you mean?
Anders HammarbäckYeah, I think recursive aspects of the models are a key feature which is not uh yet solved to the ability we need to reach Age.
Anders ArptegIf we just elaborate a bit more on that, I mean today we do the pre-training and post-training offline first and then we use an inference mode, but we could blend these two together potentially at some point so it has proper online learning, which is that you all the time update the weights of the model, which we don't. So today we have the parameters being updated once and then perhaps retrain it uh half a year later or something. Exactly. But then we can take OpenClaw, for example, this kind of cool new open source project built by a single person. He actually stores things online in markdown files inside this container that he's running in. And that way it actually is continuously learning.
Henrik GöthbergYeah, so you can solve it technically, architecturally, in different maturity levels. Yes. So you don't need to stare yourself blind on the most on the parameters. On the parameter or the most sharp, hardest, hardcore definitions. It's similar to the human brain. You can look at it pragmatically.
Anders ArptegYou can look at the human brain. We know the human brain has a working memory which is very short, and we can just remember a few seconds or minutes there, a couple of things, but then we have mid-term and long-term memory. And we can think about, of course, the uh the AI system uh where we have a foundational model which is long-term memory, but and it's very rare that that updates. And then we have a working memory, which is the markdown file.
Henrik GöthbergSo, in this sense, then in its embryonic form, he's solving online learning. He's not trying to solve it uh on the weights in the core large model, he's just technically solving it solving it as an AI compound system. That's right.
Anders HammarbäckAnd I mean, Radby, in the way we look at inference, the way we think about inference is also data will have a more you know fast-moving nature, a real-time type data all the time. So impossible to get that into the training set of big models, even now and in the future. So better to provide a smart infrastructure system to provide it when the agent needs it, you know, short-term memory in your analogy and to be able to, okay, I need this reference information, I need that. What happened in this last second? Okay, I get this data, I pay whatever it's worth, I solve the task next time. I get this data. I think that's a more dynamic way for agents to be pragmatic, right? Yeah. Um as opposed to training the whole world data, like real-world data, even and uh you know, real world models, I think it's a huge trend as well. So yeah.
Henrik GöthbergAlso, if we are if the the core question was what are some major stepping stones? I I just want to test uh and provoke by by by trying to uh iterate uh reiterate on Karim. Like so he he he thinks the whole context view of uh of orientation navigation as he calls it, yes uh is a is a major stepping stone here that you uh you know can we can call it world model, we can call it called edit. It learned it like learns like a child. Like, what do you think about that aspect?
Anders HammarbäckWhen is that also something that you would I mean I haven't thought I haven't thought about navigation as deep as as he has. Uh I think it makes a lot of sense to train um navigation as in you know the physical world and and use data again that exists. I I really enjoy how we can use things like computer games, and Demis talks a lot about computer games, right, as a way of training models, but also for robots to use uh you know physical first-player shooter games or even use uh FIFA games, football games to train you know AI models on on real football and so forth. So I think that's part of navigation, at least the way I see it, to use the immense data set that exists you know everywhere and use that to train the next frontier of the models for navigation and for orientation and context.
Anders ArptegDo you think the current current type of architecture that we have with the Transformers and GPT kind of style uh models is sufficient? Or do you think we need to change that to achieve AGI?
Anders HammarbäckI don't think it's efficient uh enough. And I mean I think others can probably answer this better, but the way I see it, it's I mean it's great for text, it's fantastic for text and predicting next word and so forth. Then when we look at tabular data and numerical data, there are probably other architectural um structures that are smarter. Same for you know, let's see what John LeCoon and his lab comes up with. Yeah, right. I you know, it's doing the transformer.
Henrik GöthbergWhat's his like we know his YPA paper?
Anders ArptegYeah, so ECE is pursuing Yippa. I haven't heard anything more from their new AMI, right?
Henrik GöthbergAvance AMI labs, or is it advanced intelligence, machines, machine intelligence, I think.
Anders HammarbäckSo I think they're focusing on real world models, right? Uh next paradigm and bringing AI into the physical world as part of the scope as well. And I think that makes sense. And if transformers is right, architecture for that to be decided, can it be modified? Not sure. But probably no other thing.
Anders ArptegI think it's important to just distinguish here because I get this question so often. So I just want to say it clear. I mean, transformers can still be a part of it, but if you take any kind of image generated today, it isn't using auto-regressive models. It is actually having an auto-encoder around the pixels, so the pixels get translated into a compressed space, a latent space, which Jannik is talking about all the time. He calls an energy space, but still, in that space, you don't do next token prediction anymore, you do a diffusion model, but it is a transformer diffusion model that is used to generate a new image. So we already moved away from the next token kind of prediction auto-regressive models that he hates so much, but but he's still using a transformer. So but I agree with you, they are extremely inefficient in terms of compute and energy. And I think we will need to, for many reasons, to find more efficient uh solutions to this. Yeah. And I think there are ways to do that.
Anders HammarbäckAnd I think also for text, I think some diffusion model papers and experiments we've seen look very interesting. So, you know, I think also on the architecture side, um, let's see in a few years, there might be other ways of combining models, like mixture of uh models, you know, like mixture of experts to kind of combine into smarter model selectors.
Henrik GöthbergWell, what about this whole uh argument around neurosymbolic that we need to sort of combine things and we need to steer something's software as something we can do?
Anders ArptegBut I think it's flips around here before before they said they should move symbolic into neuro uh neuro neural networks in some way, but I think today we're saying we're seeing models generating programs. So instead of moving programs to the AI, AI is generating the program which is the symbolic script on the side. I mean that way it's become neurosymbolic, but it is by the LLM actually generating the code and running it. I haven't thought about like that. So it's flipping around with neurosymbolic means yeah. So in that sense, I think it will. So we we know that these kind of models that we have today is um it's not good at certain tasks, it's not good at just doing multiplications, actually. This kind of simple thing, but it to write a program about it, that's that's very easy for it.
Henrik GöthbergYeah, so the LLM will reason about how do I solve this in the best way, and then it will come up that I need to generate a program so I can so I can maneuver more correctly in this space. And then the the neurons have solved the symbolic. Yes, exactly. So it's indirect symbolic in some way. That's an interesting mind fuck.
Anders ArptegIt's like you know what some open also called on-demand software. I think that's an interesting concept. You know, today we have this kind of weird thing where we build software like for years sometimes, and then we run it, and then we try to use it in that kind of and slowly adapt it. But we can come to a point where software is actually written for a single use case and that's then that's thrown away, right?
Anders HammarbäckYeah, it's like the fast fashion of software in a way. Like I need this piece of clothing, so I order it, then I wear it, and what happens is the same with the software.
Henrik GöthbergAnd and logically, it's very simple because the AI reasons to what is the best way to solve this now, here now. Ah, I write a small program.
Anders ArptegYes. I write a small Photoshop and I do it and then I throw Photoshop away.
Anders HammarbäckI think back to sort of Claude in your news session, that's in a way new uh Claude in Excel, how it sort of operates. It doesn't write the formulas in the kind of Excel that I've been taught and trained, you know, over years and shortcuts and so forth. It writes a software piece that sort of operates in the schema, the kind of spreadsheet environment, but it doesn't write the formula, so it creates a piece of software that overlays the software as a service.
Anders ArptegYeah. Anders, I know you have to leave and um pick up your kid uh soon. So I thought we'll try to end us a bit earlier. Um thanks. He's an aspiring football star signing to so how old is your kid?
Anders HammarbäckUm, Alex. And where where does he play? Play in a soccer team called FC Stockholm. FC Stockholm. Stockholm Internationale. Yeah.
Anders ArptegSo ending up here, we've been a lot about the excitement but also the concerns about the future impact AI will have in different ways. And of course, we can we can think about that in different ways. And one way to think about it is to take the kind of extreme cases here. One extreme case would be the dystopian kind of future where you know the matrix and the Terminator gets right, and all the machines is trying to kill all humans. That would be kind of sad. Um then the other extreme is the utopian, and you were into that as well and spoke about you know, imagine what AI can do for climate change or you know, curing cancer or fixing the energy crisis, and and so many more things that would be you know the big challenges that we have in our society, and AI potentially could help with it. And we start to live in what Elon calls the world of abundance, exactly where the cost of goods and services goes to zero, and we don't have to work unless we want to. Yes. So these are two extremes, and we probably End up somewhere in between. Or what do you think? Where do you think we will end up here in uh five years? I'd say five years.
Anders HammarbäckIn five years, you know, despite doing this 24-7 and working with this, thinking about it, reading, listening. I just still think in five years, our lives would be quite similar still. I agree. Good point. Yeah. You know, and I think that's Sam says that all the time as well. Like three years ago, we didn't have generative AI in the same way. But you know, our lives by and large, you know, we're people with feelings and things and families and emotions and all everything. You know, and that that core component is still the same. And then a lot of exciting model releases happen. So I think in five years, hopefully, we have solved a few you know diseases and the cures for those diseases. I would be excited about that way.
Anders ArptegCan we agree that we should separate like the technological progress? We could in five years truly have AGI, right? Or an AI that can do what most humans can do. I think on the software side, I would agree.
unknownYeah.
Anders ArptegBut it's a different thing to use it. Well, the adoption is still so bad, right? So adoption is usually linear, exactly. The technological progress is exponential. So even if we have that technology, we haven't really used it to solve cancer.
Henrik GöthbergBecause then you get to the deeper definition of AGI. Have we solved it in a lab? Yeah. Have we solved it even so it's doable at scale? Exactly. But how is the uptake? How is the regulation? And what and and so and when do we have pervasive AGI in society? You know, so so that's an interesting discussion. Like I believe in the 2029 that we actually have a AGI, and it's not just in the lab, it's in production. But I but pervasive is a completely different adoption hasn't happened, right? No, it can't, it can't.
Anders HammarbäckNo, provoking even more. I think by that time, 2029, I don't think we will talk about AI even. I I think, and of course, this is speculative, but the same way, who talks about broadband today or cloud? You know, we have cloud, we have broadband, we have you know 5.6G, we have internet, but it's like in the background, it's ambient, it's just part of the infrastructure. And I think AI will, you know, I we're getting some funding for a company. I can say that. Part of the investment DD form was fill in which AI systems you use. And that's not like, yeah, yeah, you can do the obvious ones, but then you go into software as a service that have AI components. Are there systems that don't have any aspect of machine learning or AI? No. And I think in five years, AI as a kind of transformative technology will be already ambient everywhere in almost everything we do. So I don't think it's a subject matter as it is right now, but it still will sort of support many things that it doesn't right now. Uh so I'm very optimistic.
Anders ArptegYeah. So I agree it will happen everywhere, but still imagine if we come to the AGI or even ASI, artificial superintelligence, and we can actually speak to a machine and say, analyze what will happen in a geopolitical situation now, and it can do that with the power of a million people. That's still kind of scary, yeah, and powerful. So, in that sense, a ASI could be kind of a scary topic in five years as well.
Anders HammarbäckLet's see. I'm have to come back and revisit the topic in five years or three.
Henrik GöthbergWarren, have we had five year anniversary yet?
Goran CvetanovskiNo, we haven't. This is the 13th season. This is six and a half years.
Henrik GöthbergSix and a half years. But we haven't had a celebration. We haven't had a celebration, so that's why I didn't feel it.
Anders ArptegMaybe it's the same. Maybe that's the same.
Henrik GöthbergSo in the 10th years, this is you know, we have less than 50% left to answer the question. That's true.
Anders ArptegLet's go on in uh maybe in June. Yeah, okay, good. So I was looking forward to that. Is it six and a half? Come on.
Goran CvetanovskiIt's 13th season, we're just starting in 13th season.
Anders ArptegYeah, it was a surprise. Time flies when we have fun, and it's certainly been fun. And we learned so much by this, and we learned so much from having you here. My pleasure, and thanks. Was it 20? Was it 21? Yeah, it's a pandemic. And it was a true pleasure to have you here. I wish there was mine here for the after after work, but you have to run away and pick up your page. Take a rain check on that. Great, I'll come back. Thank you so much. Thank you. My pleasure. Great pod, keep up the good work.
Anders HammarbäckThanks so much. Cheers.
