E142 - 2024 in Review - Christmas Special

Join us for the grand finale of AIAW Podcast’s Season 9 with Episode 142, "2024 in Review", a reflective journey through a year of remarkable advancements in AI. This special episode revisits key discussions from Seasons 8 and 9, highlighting insights from 27 expert-driven episodes. Explore topics like AI in HR, recruitment, and banking, autonomous agents, ethical AI, and industry-transforming innovations in energy, healthcare, and software engineering. Featuring standout moments on AI's societal impact, marketing, and the Nordic welfare system, we’ll summarize lessons learned in 2024 while offering predictions for 2025's trends and challenges. Don’t miss this comprehensive celebration of AI’s evolution and its exciting future. 

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

Chapters

• 0:00 Data Innovation Summit Recap and Reflections
• 4:33 AI Podcast Achievements and Highlights
• 15:52 AI Regulation and Innovation Challenges
• 25:13 AI Innovation and Compliance Trends
• 34:01 AI Research and Enterprise Transformation
• 38:30 AI Impact on Music Industry
• 52:31 Experts Discuss AI Communication and Innovation
• 1:00:52 AI Nobel Prize and AI Commission
• 1:11:48 Engineering Gaps in AI Commission
• 1:19:03 Advancing AI Through Latent Space Engineering
• 1:30:28 AI Infrastructure Investment and Innovation
• 1:42:15 Future of Open Source AI Models
• 1:48:08 AI Future Trends and Outlook
• 1:57:02 AI Innovation and Engineering Challenges
• 2:05:39 AI Outlook for Swedish Ecosystem 2025

Transcript

Henrik Göthberg: 0:00

Ah, goran, you need to help me. This was Data Innovation Summit 19,. Maybe yes, so Data Innovation Summit 2019,. We had a Danish guy who was a really good guy working in a top job with data, of course, and analytics, and then we found out, sort of in the after party, that one of his main hobbies was to do stand-up comedy and one of the favorite things he had was the roasting style of stand-up comedy.

Goran Cvetanovski: 0:35

if you know this, I can't remember the name now I can't remember, but he roasted me so bad.

Henrik Göthberg: 0:43

Yeah, but it was two angles to the story. So, first of all, all he built this piece or like bits on roasting the data and AI industry that he has tried on stage like open mic stage and stuff like that. So he's been doing the stand up and he did some of the stuff on us and it was hilarious you know all the stuff. But then it took a little bit of a wrong turn because then and he did some of the stuff on us and it was hilarious you know all the stuff, you know but then it took a little bit of a wrong turn because then he started roasting us and and when he was roasting me and Goran, that was super hilarious because we knew the backstory and we knew what was going on. But he started roasting everyone, anyone walking past. So imagine beginning getting exposed to a roasting person that you don't know it's a roast, but I mean you know. So what are we talking about this? Well, you know what's the roast of 2024. I guess this is the end season finale. How would you frame it?

Goran Cvetanovski: 1:41

so, um, I think we didn't have so much Kardashian dramas this year, for sure, but we had one big drama, which was the strawberry drama.

Henrik Göthberg: 1:54

But even was it this year that someone got fired and got back geeky, and that happened this year too. The Kardashian drama is this year's.

Goran Cvetanovski: 2:02

Yeah, probably still here, but it was not so hype as it was uh as it was in 2023. I think that is when it became so yeah, but I think the big.

Henrik Göthberg: 2:13

What's the big thing last year?

Goran Cvetanovski: 2:14

I'm so confused yeah, but the times are running. Uh, that's very fast keep in mind. We have been doing this now for four and a half years, so next year we are starting the 10th season, 10th Season, 10.

Henrik Göthberg: 2:29

Season 10.

Goran Cvetanovski: 2:33

This is like a Seinfeld oh, season 10, blah, blah, blah. So it's quite. I'm extremely proud. Just basically talking about it, I think it's been quite a journey so far and plus I think it's been quite a journey so far and plus I think it's been a fantastic year. We had some weeks that we didn't, actually when we didn't stream this year, and we deeply apologize to all the listeners and etc. But it has been a little bit of a busy schedule for all three of us. However, we still managed to produce 33 episodes In 2024? Yeah, in 2024, which is not that small amount comparing to the other years. I think that maximum we have gotten 35, 36 maximum.

Henrik Göthberg: 3:17

So it's still fine. A couple of episodes fewer, yeah, still fine.

Goran Cvetanovski: 3:24

What can I say more from a production perspective? I think it was an amazing year a lot of topics to discuss, great guests we pulled in. We had the first basically online as well podcast, so that was a little bit new for us this year. That was Javier L new for us this year. Which one was that um, that was a javier lupescu from mosaic ai when the data bricks, which was very d bricks

Goran Cvetanovski: 3:53

launch. So that was very interesting. And then now what we tried this year is actually to go a little bit more into panels instead of doing just one or two men show. So we had like approximately three or four panels, and I think these are the most exciting ones.

Anders Arpteg: 4:10

For me it's the most fun actually.

Goran Cvetanovski: 4:12

I think that it's super, super nice and it creates this energy behind it and it's different perspectives on the same topic, so I think that we should actually try to do more and more of that in 2025. Panels on a theme, panels on a theme. I think that is all what it is, so let me just wrap up the year very quickly. First of all, I'm extremely proud that we are still remaining to be amongst the top most popular shows out of the podcast globally. Of course, there are quite a lot of them it's 3,500,000.

Henrik Göthberg: 4:46

3,500,000 podcasts at this point of different kinds.

Goran Cvetanovski: 4:50

So we are aligned within the technology and there is no technology and business classification, which could be actually a little bit out of turf, because it's a mix between technology and how this is applied in businesses and society. So I think that will be perfect. But in most of the cases we are ranked within the technology. So this year we will rank 15th on the list of best data science podcasts on a good post, good, good pods. So cheers and shout out to good pod for putting putting us on this list. I think it was great. Uh, we are still ranked around like 41 globally on the top uh ai podcast. That is basically by fit spot and and no other swedish podcast there in the list.

Anders Arpteg: 5:39

If I may say, no nordic.

Goran Cvetanovski: 5:40

Actually, this is a little bit surprising because it was the, it was some actually Finnish and Swedish podcasts that were on that list that fall out. This time, I guess it's as I mentioned at the beginning. I think it's the sustainability over time that beats. Keep in mind that Joe Rogan is right now what on almost 2000 episodes he has been doing this for 10 years getting close as well, yeah, but the pace that Joe Rogan has man come on.

Goran Cvetanovski: 6:11

He's doing like three, four podcasts per week. That guy is an animal, so shout out to Joe Rogan. So if you're coming to Stockholm, you know where to find us, that would be awesome.

Henrik Göthberg: 6:23

Yeah, that's the fanboys talking right here.

Goran Cvetanovski: 6:25

Yes, yeah, good. So let's look at what we did this year. So we had, basically, we discussed talent and AI in recruitment. We spoke about RAG models to autonomous AI agents. We spoke about AI-driven software testing. We talked about AI innovation in banking, spoke about AI-driven software testing. We talked about AI innovation in banking, the AI apartheid, which was surprisingly very popular on internet. It's a clickbait. It's a clickbait for sure. We discussed, multiple times actually, the Swedish AI ecosystem and implementation and applications of that within all sorts of the public sector. We talked about AI in lean software development. We talked about AI in lean software development. We talked about AI-powered product development. We talked about the power grid of tomorrow. We talked about DBRX with Haggai. That I mentioned before as well.

Anders Arpteg: 7:18

It was Databricks, right, Databricks yeah.

Goran Cvetanovski: 7:20

Databricks, but at that point of time it was still Mosaic ML it was right around the corner.

Henrik Göthberg: 7:26

No, hage started Mosaic and they got acquired. And then it was the but didn't we do this when it was the big DBRX launch?

Anders Arpteg: 7:37

It was the same week, I think maybe even the same day.

Goran Cvetanovski: 7:40

Then we talked about AI and the Christian democratic values with Lisa Maria Norlin, so that was a special occasion for the Swedish.

Henrik Göthberg: 7:47

Yeah, party secretary for the Swedish.

Goran Cvetanovski: 7:51

Kristdemokraterna. Yes, exactly, we had AI learning and assistance with Sophie. We talked about deploying AI products. We talked about generation AI the documentary is done by Alexander Noregna, so that was a fantastic as well episode. Ai and Marketing with Emma she was fantastic. Very great energy, love that. We talked about AI-enabled future of music industry. We talked about building self-hosted generative AI solutions. We talked about the latest trends, together with Luca, jesper and Salah Fransen, which was also the closure of season eight before the summer. So that was actually super great panel, I think. Great energy, a lot of laughter. So if you haven't actually listened to that one, please go back. Although, going back six months in AI, it looks like almost like a decade. The season nine that started in September. We spoke about AI in the Nordic welfare system. We talked about 70B, the strawberry and other rumors. This was the Kardashian moment as well, because everybody was like the AGI is coming. It's called strawberry.

Goran Cvetanovski: 9:01

Well, it turned out it was not. It was called-.

Henrik Göthberg: 9:03

Q-star strawberry.

Goran Cvetanovski: 9:03

well turn out it was not, it was q star. Yeah, it was called exactly q star, but it was not. Nothing happened from that. It's a 01 model 01 and now 01.

Henrik Göthberg: 9:14

They charge 200 q. Star and strawberry it's not so sexy as a strawberry, but I told you they will have a and they launched it and now they relaunched and you know They've done something again and the pricing is hiked. We can talk about that.

Goran Cvetanovski: 9:27

And I told you they will have a problem with trademarking that thing because Strawberry is a chain of hotels that we know very well in the Nordic. So somebody made money on that.

Henrik Göthberg: 9:37

We talked about HPC in Sweden and in European Union, and this is now becoming a reality because Sweden, just one week ago, got approval from the European Commission factory the Mimer yes, exactly, so they're in total seven, but this means seven in Europe, one in Sweden, but the podcast with Tors then a little bit of a scope, because we had him here talking about the application they were sending in and we were sort of this was not really news out around broader. And now you know the Mimer story. The AI factory story was pretty big this week, yes, and we were on the story. How many months ago?

Goran Cvetanovski: 10:18

Well, I think that there has been a lot of stories actually this year that we got caught up into or basically entered a rabbit holes. It turned out to be a very good prediction of where we are going, so I'm extremely proud of that thing. To continue, we spoke about Sweden's AI journey and a future possibility with Sverker. That was actually a very good podcast as well. Testing and evaluating AI. You need to hear and listen to this podcast.

Anders Arpteg: 10:49

It may not sound funny, but it's actually a very good podcast, I think.

Henrik Göthberg: 10:53

It was really, really good, and I have another additional scope on that. I'm not sure if I'm allowed to say that, but yeah. Either you say it, either you don't say it. No, keep it All right, yeah, either you say it either.

Goran Cvetanovski: 11:04

You don't say it? No, keep hanging. Okay, fine, all right. So we then went to high-performing enterprises in the AI era with Rainer Deutschmann, which, according to YouTube analytics, it's one of the most favorite this season. We were talking about revolutionizing healthcare in AI.

Henrik Göthberg: 11:23

Which was really a federated machine learning conversation.

Goran Cvetanovski: 11:26

Yes, exactly, we spoke about AI-based software engineering with Anton Osika from Lovable Shut.

Henrik Göthberg: 11:34

A GPT engineer. Gpt engineer or Lovable, or Lovable. And they're doing amazing stuff right now and the launch was it's already been the major launch now, I guess.

Goran Cvetanovski: 11:43

Yeah, it's been now for a month almost, and they are kicking ass at this point of time, so I'm extremely proud Doubling, doubling subscriptions per month.

Anders Arpteg: 11:53

Yes, it's going really well.

Henrik Göthberg: 11:56

It's an impressive piece of software. And the core fact when he said how many stars on Git Git, how many? There Like 50,000 stars at that time.

Goran Cvetanovski: 12:04

It's probably much more now on git github like 50 000 stars at that time. That is pretty impressive. I think that in eight days they got like a 1 million euro annual recurring revenue.

Henrik Göthberg: 12:15

So it's a quite, quite fast, lovable team. Yeah, anton, well done uh.

Goran Cvetanovski: 12:19

And then basically we finished with the ai, or applied ai research in practice with uh se, sepideh, me and Anders. Unfortunately we were not here that night, but Hendrik took the flag and represented.

Henrik Göthberg: 12:32

I really enjoyed Sepideh, the way she can sort of relate and evangelize around applied AI research and kind of position it. I find that that's a good episode if you want to understand the difference between research and engineering and what is applied ai research, and you can learn about technical readiness level six to you know four to six and stuff like this. Yeah, it's really good. He did a really really good job at this I am so bummed that we were not, yeah, you missed it you missed it.

Goran Cvetanovski: 13:02

There were two actually episodes that we missed this year me and uh andres combined. One was this and the other one was uh, with uh jeanette. Yeah, but we spoke with jeanette, so she's coming on the season 10, so it's going to be super cool yeah, and I butchered the season.

Henrik Göthberg: 13:17

You know, one of my worst episode as as a host was probably with jeanette, because I was so excited about all the stuff that she was talking about. And you know me, when I get excited, I fucking talk too much.

Goran Cvetanovski: 13:30

Yeah, but it was-.

Henrik Göthberg: 13:31

So I'm glad that she's coming back.

Goran Cvetanovski: 13:33

It's a preparation for the next episode that we're going to have, and there are quite a lot of things happening in Sweden and the Nordics when it comes to AI.

Henrik Göthberg: 13:40

So it's a very exciting thing, very exciting times.

Goran Cvetanovski: 13:42

Very exciting time for AI in the Nordics. So if you guys from US or everywhere in the world you're listening, join us here.

Henrik Göthberg: 13:50

Join us here. Shit is happening, it's really happening.

Goran Cvetanovski: 13:54

Good stuff. So let's talk about recurring futures or basically.

Henrik Göthberg: 13:58

Wait, wait, wait. You went through the theme so fast so you didn't even mention. Some of you know you didn't usually your name drop when you release like this. You didn't name drop, I did that name drop a little bit. So I mean like so. So you said rhino, you say it's happy there. We had janet nilsson yes, head of ai center in rice I mentioned all of that time.

Henrik Göthberg: 14:18

Yeah, but we don't know who they are. Who knows? Okay, it's so interesting people. Sorry I'm so. I just want to do some name dropping. You said that I should keep this short, sorry.

Goran Cvetanovski: 14:27

All right. So if we go in depth, like, just go and listen to the episodes, it's going to be fine. These are class people. It's basically one of the best.

Henrik Göthberg: 14:34

It's the top notch people in.

Goran Cvetanovski: 14:35

Sweden. Yes, looking forward to actually get more Nordic people next year. Yeah, we need to work on that. Yeah, so now people even from different countries are applying to to come to the show, so it's going to be interesting if we can make it happen. Yeah, so let's see, all right, good. So, um, this is what I wanted to uh highlight a little bit, because, uh, of course, this would not be an AI show if we don't get all the transcripts and uh them into AI. We use AI.

Goran Cvetanovski: 15:03

To summarize here, so I did this and I'm sorry to Gemini and OpenAI and everybody else, but I use perplexity to this Perplexity. You still rock when it comes to research, so I like that very much. The interesting thing was about where we are when it comes to summarization of the entire topics that we discussed, and those are AI and future of work, recurring themes yes, recurring themes. So those episodes are episode 115, 117, and 140. So you can actually go and check them out. I will put some information in the link below. We talked about ethical concerns in AI. We cannot actually deny that the AI Act came to action this year in May and it's been a very feisty dispute since then about how this is going to work in practice.

Anders Arpteg: 16:06

And more and more of the tech giants are actually canceling europe.

Henrik Göthberg: 16:08

Yes, yeah exactly we've seen cancelled. Cancel europe first months until you understand what you're doing. All the technology, more or less all of the tech, major tech launches have been delayed.

Goran Cvetanovski: 16:19

Yes, yeah, I mean we still cannot use llama 2 right llama 2, I think we still cannot use llama 2, right, llama 2, I think.

Henrik Göthberg: 16:29

But yeah, but they're all delayed apple, but they, they, they go out in in portions, in small portions, not the big launches that you see all of them, yes unfortunately, that is what it is, and it still looks a little bit shaky of what is going to be implemented and etc.

Goran Cvetanovski: 16:41

I'm not saying that is a bad regulation. I think it's a really good thought behind it. But is it an incentive for organizations in Europe to innovate or to, let's say, not?

Henrik Göthberg: 16:58

to innovate.

Goran Cvetanovski: 16:58

But I think we need some clarity on this.

Anders Arpteg: 17:00

I mean, it creates an insecurity about you know, will I get sued or not? And that causes companies to question if they should launch in Europe. Also European countries to not being willing to even use AI at some points, and that's similar to what happened in GDPR times and we need to fix that.

Henrik Göthberg: 17:19

And, by the way, we solved how we should be doing it here at the pod, where we say if the legislation is the what, now it's super important to ramp up on the how and putting things in place. That basically makes us smooth and safe innovation country of the world. So I think you know. So the bottom line here is a little bit like yes, yes, yes, you made, made the law. Now really step it up to make it easy to apply the law and guide us how to apply the law and then make it a competitive advantage of sweden that, oh, if I go, if I want to be complying with the act, is easier because I get the help to do it right. I think that's that we talked with several guests on this and it's actually quite clear it could be flipped into a competitive advantage for Sweden.

Goran Cvetanovski: 18:09

But the question is can we balance data sharing with privacy right? Keep in mind that now we are talking about in order for this to work, so we are making a European air act. How come we are not talking about how countries are going to share data? I mean, even if you're talking about in Sweden, it's a little bit like my fear that it's going to be similar, like with GDPR it's just going to scare people off to innovate, but everybody's going to be stuck in a limbo.

Henrik Göthberg: 18:39

The only thing to avoid the limbo is to have the people doing the regulation also taking care of the consulting and advisory and support to help us navigate it. Yes, and we don't want the consultants and the lawyers to make money on this. Yes, we want ourselves, or even our states, to make money on this in that case. So the law is great, but guidance is the key. I think it's super simple. So the law is great, but guidance is the key. I think it's super simple and we had several simple. We could prove that in many conversations that that's the way to go.

Goran Cvetanovski: 19:12

Yes, the scary thing about this me and Anders, we were on a conference where Jeffrey Hinton was basically presenting oh, that's cool, and what he said didn and instruct me the most. He said like um, and it's obvious, right. He says like you read all of that and then on the bottom it says none of this applies to governments and defense. Right, so we can actually use ai for defense, but we cannot use to innovate in the other sectors. So I think it's a little bit like again the same thing.

Goran Cvetanovski: 19:50

It was with the GDPR as well, because it didn't catch the let's see. So, and another thing that he mentioned, which was very important, if you remember, he said like the people that basically do not care about regulation, they were the ones that are innovating the most. So that means all the criminals and everybody else who wants to do harm. And if you're looking at the GDPR and what is happening right now, people can still go online, find data about 70 year old in Sweden, that is, you know, call their telephone number and do a lot of scams because of that right.

Henrik Göthberg: 20:26

But isn't that the biggest flaw of this conversation? That we want to regulate in order to avoid harm? But obviously, the one who has nefarious purposes, who wants to do something bad, you don't catch them, in this sense anyway.

Anders Arpteg: 20:42

I think GDPR, you know, missed the target and the goal to a large extent because the companies that have the easiest time to be compliant are the tech giants and the ones that don't care at all, the malevolent actors. They have no problem because they don't care. So the main target audience that is being hit by GDPR and the AI Act, I would say, are normal companies that are fearing about you know, but the ramifications of this can be and that's sitting innovation, unfortunately, and that's really sad.

Henrik Göthberg: 21:16

I remember how we even talked about the trifecta in order to master GDPR. You, you have the size, you have the lawyers, you have the industrial approach to data, you know real industrial approach to data factory, you know pipelines and all that, and then you have a superior knowledge of technology and how you build and engineer stuff and with that trifecta, you actually have you know. This is what's not a problem to solve that. And that trifecta is what the super data elite has, but what majority doesn't have.

Anders Arpteg: 21:57

And they have a Swedish AI commission. We can get back to that later, but they came up with a report saying that the, the larger companies in sweden, actually have an easy time to do this, and I think it. Nothing can be further from the truth. It's actually not correct at all, and even the biggest tech giants if they are failing to comply with the law, then who else could do it?

Anders Arpteg: 22:18

I mean this is an interesting one with the biggest you know legal teams you can think of and the best experience doing it, and they still are concerned.

Henrik Göthberg: 22:26

Of course, all the others will have super big concerns and this is interesting because in gdpr that wasn't stopping one beat in the tech giants in their approach to europe. They used, they used, fixed it, but all of a sudden the ai act, now even the tech giants are holding back their launches. So it means to the tech giants even this is more tricky. Anyway.

Goran Cvetanovski: 22:51

All right, let's continue. So we had AI and the future of work, recurring teams, ethical concerns in AI and, as you can see, it's a feisty debate. Still, we had AI democratization. So several episodes emphasize on the power imbalance between large organizations and open source initiatives. We had a discussion about AI in leadership and decision making and then what was the most interesting as well to put a topic on that was Top pick on that was we talked about quite a lot about cross-border collaboration in Nordics, but that we can go through, and I think that I spoke about it earlier. One thing that I miss here I'm just going through my notes I have the shift towards decentralized AI development and we talked about emerging technologies and applied AI.

Goran Cvetanovski: 23:48

So this is the the year in a nutshell the year in a nutshell, summarized with the support of perplexity yes, I mean there is quite a lot of things that are are unseen here, because the amount of content that we have put out and that we are putting out every year is quite yeah because, because on average there are two hour episodes and we're talking 33 episodes times two hours.

Henrik Göthberg: 24:17

So the way, when you want to go do analysis, you can either load the whole transcript or, how we did it this time, we loaded the key yes, key sections, the key sections.

Goran Cvetanovski: 24:26

Um, the interesting thing about this is that, uh, even if we put basically all the 66 hours of content and more 70 hours of content this year and we put all the transcripts uh, transcripts and everything else I don't think that chat gbt will have the token space for that, yeah but, uh, it's been a, uh, it's been a great year so far.

Goran Cvetanovski: 24:50

So I'm very thankful to you guys for um doing this as well together with me. Uh, we know that our thursdays for every, every thursday for the post, uh, past four and a half years has been uh here in this, but it's easy to do I mean, we love AI and we love to come here and have an after work and combining the two having an after work and speaking about AI.

Henrik Göthberg: 25:13

But to move on then, if this was the AI summary of the year, what is the host's summary of the year in terms of the guests we had and how would you reflect on it? And there's the human reflection from the hosts. You know highlights or what you, what would you sort of from the guests you mean from, if we're starting with the guests. Yes, if we start about the themes and the guests, sort of what, what, where would you sort of try? What would you highlight and put forward to our listeners?

Anders Arpteg: 25:45

I think you know, speaking about testing and also QA tech with William, where we try to use AI to do more of the let's call it MLOps kind of settings, where we really want to find value from AI and what it means to not just build a prototype but actually put something in production and being compliant with the regulation, all the sandboxing, everything connected to AI act. I think that is so important topic, so for me that was actually very interesting to hear.

Henrik Göthberg: 26:14

So rewind the tape. So with William Erheim he's a startup working with QA tech, quality assurance tech, and what they have been doing is to think about how would an ai driven or a ai assistant or agent driven way of you know, supporting, I guess, user acceptance testing, or how would you frame it any times of testing that you need to do in order to understand if something is ready to be shipped production?

Anders Arpteg: 26:43

trying to ensure that the quality of the code etc is actually sufficiently good and, um, you can use ai for that, and that's actually very powerful yeah, and used to elaborate for the people who didn't see the episode.

Henrik Göthberg: 26:53

I mean, like a lot of stuff we even need, we need to act as human, click our way through and test things, and we need to have many different test scenarios and this is a very tedious work to test every single stupid combination that can go wrong, which is a perfect, perfect storytelling for AI, of course.

Anders Arpteg: 27:12

And then moving to the sandboxing that we hopefully will get with AI Act at some point in the future, but we have really no good solution for yet.

Henrik Göthberg: 27:19

But maybe this is the way.

Anders Arpteg: 27:21

Then perhaps we can actually have some AI helping with this, which would be really good, and I think that Petra is driving this from rice.

Henrik Göthberg: 27:30

That's the test episode.

Anders Arpteg: 27:32

And hopefully that can be something that makes us actually stand out and it can actually be an opportunity for Sweden. If we actually do find a value in this and I do wish that we could actually see this as an opportunity instead and if we do this right and better than other countries, it should actually be something that makes Sweden take a lead.

Henrik Göthberg: 27:52

Yeah, and if I orientate the listeners from one startup, very, very interesting episode. We have another episode with Peter Dahlunde, who's heading up what is called a technical or test and evaluation facilities within RISE, which is an EU assignment. When we have the AI Act, we need to set up test centers, how to evaluate that we are compliant to this and then, basically, they need to figure out what is the methodology, how we evaluate that we have the models correct, that the data is correct, that the setup of the governance is correct, and they need to figure that out. So so what you did now you took oh, this is some interesting stuff, how the startup was thinking and I, we oh, maybe this is very similar to things that we need to look at in a teff.

Anders Arpteg: 28:41

I mean I think if companies and people understood, it's actually not that hard to be compliant if you do it properly and just know what to do. So with the proper examples in place. If you could just have some reference implementations of how to properly use AI, that can actually be something that makes people less fearful and actually much more compliant and simply more more responsible when it comes to use of data and ai but and I really like that notion if, if, if, if rice and the teff and the air factory or whatever we have takes the stance of we want to have safe and smooth innovation with ai.

Henrik Göthberg: 29:21

and actually, if you're professional and do things right by design, you avoid the risk and you don't take the governance or AI compliance as an afterthought, you really do it right.

Anders Arpteg: 29:34

But it's not easy. I mean, most companies don't know how to do it. Even larger companies don't know how to do it properly, so I think this is actually something we need to invest in.

Henrik Göthberg: 29:42

But then you can argue that this investment is equally important for AI innovation to adoption as it is for compliance. So they don't even know how to do it properly. So they are failing on value. So, if you flip it, whatever they're doing in the TEF is there not only to support compliance. If the mission is not compliance but safe value, it would be better. It's great, it's awesome. If the safe value, it would be better. It's great, it's awesome If the test is set up correctly.

Anders Arpteg: 30:08

Well, more innovation but for testing and compliance would be awesome.

Henrik Göthberg: 30:12

Yeah, that's perhaps a good way. Yeah, it's a good summary.

Anders Arpteg: 30:15

What else? Well, I know Anton Oseké from before and what he has done with GPT Engineer was something that I think is very seldom we see. It's one of the top open source projects from Sweden at all.

Henrik Göthberg: 30:32

Yeah, in all categories.

Anders Arpteg: 30:35

And that's amazing to see and people that haven't tried it. I really recommend to go there and try it out. You can get some free credits and you can see that you can actually build a product very, very quickly. I think they had a workshop recently where some of the super users of gpt engineer and now called lovable were sitting together and building some products and they were tasked basically to, in like 11 minutes, build products 11 minutes and it's kind of crazy. And they also have this community now around gpt engineer where some people are committing to building one product per day and when we say product, we're talking about something equivalent of some sort of web web application per day web with the front and back end and functionality and working.

Henrik Göthberg: 31:22

Yeah, so that's very different from.

Henrik Göthberg: 31:25

So we're talking about one of the last episode, anton Ossika, and where he basically started almost like a conversation with his co-founder in Depict AI, oliver Oliver, what's her name, I'll drop it. Yeah, but the core topic is like how can we improve the experience and how can we build an assistant for engineering or software engineering? And they did an open source approach very much almost like a prototype or an experiment, and it got popular. The way he's describing it himself is it's probably one of the most consistently reliable GPT engineers or assistants to build a functioning web application. So we saw a big launch from Devon, devon, devon, and I think this is much better. The Swedish version here is much better. Maybe they have slightly different target audiences in terms of is it for a very experienced software engineer versus you don't need to be a software engineer and you can work on it.

Anders Arpteg: 32:35

I think it simply changes a bit the way we will do software engineering in the future, and I think this is a small taste of it. And last week I was away from this, but I was on a workshop with Jan Bosch, who is a former guest, also on the podcast, but he's a professor in software engineering, and they had a workshop with a theme the end of software engineering as we know it, as we know it, touché. We speak about topics like this and it's not really software engineering as such, but the way we actually do engineering. And if you try out Lovable, you will see that this is probably the future we will see when it comes to building different pieces of software. It doesn't mean that we won't have things to do. It simply means that we can do so many more things we can do one web application flip and stop saying that it will replace humans.

Anders Arpteg: 33:31

Instead, say it will enable humans to do so many more things that we were unable to do before.

Henrik Göthberg: 33:36

At all we look at. It's not replacing the engineer, it's allowing the engineer to build one web application per day. You know, but you didn't. You need the web, the engineer, to build one web application per day. You know, but you didn't. You need the web application engineer to figure it out and prompt it to do the work. But the productivity is, you know, try to build one web application per day, my dear friends. Try what else?

Henrik Göthberg: 34:01

I let me take one in between, because you, you now took a couple of very you know, a little bit on the techie side, which was super cool and they actually they were not only techie in the terms of, but the entry point was startups, techie, really hardcore examples. One of my favorite shows was, of course, with Reiner Deutschmann. I was super happy to have the former group COO of Telia here and for me, who sort of also wants to see the business transformation of the traditional enterprise in this pod, what stood out to me is he has a fairly unique profile. He has a research background in physics and quantum technology, blah, blah, blah, and then he moved into working with digital transformation and product. So he has one for being on a business side and group COO, a very techie background and to then converse around.

Henrik Göthberg: 35:06

You know how do you drive change of a large enterprise like Telia, the journey he did for three years, or what is the enterprise of the future, optimized or high performance with AI? I find that conversation really refreshing, both in terms of of um, what we need to work out, the foundations, but also, equally, you know he has a world model. He has an idea we need, we need to work on. Do you remember the five, six p, so whatever we called it? Remember? So he had a, a quite mature view on the stuff you need to work on as a CEO and I think that was refreshing. It was really nice. So that was mine, one of my favorites.

Anders Arpteg: 35:50

I was impressed as well, you clearly have a lot of experience and knowledge, so that was really something.

Henrik Göthberg: 35:55

And then I must actually you weren't there, so I need to take this up you weren't there, so I need to take this up the presentation and the setup, the pod we have with Cepide on trying to frame what is applied AI research, and where Jesper Fredriksson was in co-hosting together with me. I think we had a really, really nice conversation and Jesper did a great job. Really nice conversation and Jesper did a great job, and I find Sepere is a very nice way to evangelize around this. She could be super nerdy, but she's not super nerdy. She can explain it in a way I was trying to steal your thunder what I learned from you, Anders. You know how do we differentiate the research from engineering. We could take that angle and we actually improved and added to the dimensions. Starting from that.

Henrik Göthberg: 36:58

There were many nice angles and then she had so many cool projects. Did you know that Seppi was working with the Olympic sailing team and so she has been contributing to medals in the Olympics, in the last Olympics? That's some cool storytelling. What's your benchmark? What's your KPI? What did your AI do for you? And this was, of course, about optimizing, building a coaching tool and helping the sailing coaches to strategize on how to draw out the optimum route depending on different wind patterns. Nice, you know. So there were so many angles that she, because she's working as a senior researcher, that is, you know, with the hands in the dirt in the different projects, and we could flip it from, you know, the project with Volvo over here, the Olympic Committee over here, and I like that. So we could sort of we could talk about applied AI research and we can exemplify it with really nice anecdotes and we could talk about the gold medal or actually silver medal. I mean, that was was nice episode for me. What else do we like?

Goran Cvetanovski: 38:08

um, I have to, um, I think that there are a couple of those that were very interesting for me in um any case. So I think this uh, I always like the um, those ones or those podcasts that are actually showing that there is a shift going on.

Goran Cvetanovski: 38:30

Disruptors of the past are being disrupted today, and I have to start with the AI-enabled future of music industry. I mean, if you look at the uh, we started the year when open ai was sued by new york times. I believe right at the beginning, you know, there was a lot of discussion. Can they use this um data for coming from this news? Um uh. Then there was a lot of uh. You remember, in hollywood there was a demonstration from the ai. Art uh was a demonstration from the actors and then from the musicians about banning AI.

Henrik Göthberg: 39:10

Screenwriters, screenwriters killed.

Goran Cvetanovski: 39:13

It almost feels like you know 100 years ago, when people were demonstrating for having like motor vehicles on the road because they were scaring the horses. Feels the same momentum, right. So I think the music industry is about to be disrupted quite a lot. Not in that way that basically artists will not make money. They will make more money than before, but now it's much easier to sample actually things and use it in your music without paying royalties forever to the original artists. So I think that AI is becoming this new sampler brush, whatever you want to call it, that people can utilize it in the future. Keep in mind that at the beginning, all of these companies were suing basically the AI companies and then on the end, when they got the money that they wanted, then it was okay for this, what is called content to be used, but it's like the business model needs to sort itself out.

Goran Cvetanovski: 40:15

Exactly so. Everybody doesn't mind, actually, if their IP is used, as long as they are paid. That was also the story of Spotify. At the beginning, everybody was against Spotify until they found a way how to monetize on that. And now, when everything is super distributed and et cetera, of course there will always be some people that are not fully what is called satisfied with it, but the business will find its way and it's very actually.

Anders Arpteg: 40:43

Do you think artists will get more money in the future?

Goran Cvetanovski: 40:46

Yes, of course they will get, because even today they are getting money from the concerts. Spotify right now is the biggest actually campaign ad marketeer for any type of a concert. You go there and you basically click on upcoming concerts and you buy. That is what it is. I bought it from there. Teddy swims is coming to town. I'm going to there. Uh, kent is coming. Uh, now it's uh. They, after the final tour, they're making a new tour you know, zombie, zombie, zombie tour zombie tour so I bought tickets from there.

Goran Cvetanovski: 41:22

I bought like a tickets for Ed Sheeran and I was just now to super great band. We were there last week. I should know the name. It will come. It's a couple of these beers that are not there, but I'm thinking about this. Then it was another podcast when we were talking about AI in marketing, and I love that because I'm very about ai in marketing and I love that because I'm very into marketing and sales and I think that then I did a good, great job yeah, she did a great job and I think that it's very important for people to start basically talking about how we can utilize ai to uh, effectivize our work.

Goran Cvetanovski: 42:00

So can we? We can be more creative because, if you look at the workload that we have these days, we don't have time to innovate, we are just tending to. You know, like there is something burning every single.

Henrik Göthberg: 42:15

I have this picture in my head where two guys are pushing up a wheelbarrow with stones, with the square wheels, and someone is trying to offer them round wheels. And we don't have time to innovate with round wheels, we are too busy pushing up.

Goran Cvetanovski: 42:30

I'm serious.

Goran Cvetanovski: 42:31

I'm looking at my work and what we are doing today and it's impossible these days to be as effective as it was 10 years ago, because it's just the workload is too much. So if I can find a way for me and my team actually to effectivize 40 of their workload and give 20 of that to just basically think and ask questions, that is actually the value for me in ai. Keep in mind, my perception is that we will not be even talking in two, three years from now about ai, because ai is going to be infused in everything that we are doing, and I think that it's also very, very it's not fair, because many people basically see AI right now it's a link, so AI equal to me not having work, which is not true, because it's people has been using AI for a long time. I mean, the first ML model from Google was released when 2001?

Anders Arpteg: 43:31

Everybody works AI has been around since 1950s.

Goran Cvetanovski: 43:35

Yes, but in application that has been all around, people have been using it Spotify. You go and run right and you have the music is changing.

Anders Arpteg: 43:44

You mean Google Search run right and you have, like the music is changing. At least you mean a google search. Yeah, google search have had machine learning. You know part of google search for a long time.

Goran Cvetanovski: 43:49

Yes, yes, uh, I was reading about this, so I think it was 2001 that they released the first. I mean for them, the worst thing for google was actually to release the transformer paper attention is all you need right yeah, exactly, but uh, it's going quite well.

Goran Cvetanovski: 44:10

Then we had, like, uh, another, um, I was looking for something which was, I must say, yeah, yeah, let me finish. Um, then the air in the welfare. I think there is a quite a lot of great opportunities for AI to be infused in the workload of all the doctors and all the assistants and et cetera. Right now, we are just basically, and all of these workers are overwhelmed with administrations.

Anders Arpteg: 44:38

But I think you know, I think here is an important point. We need to be better to communicate about the benefits of AI, and I don't think that you should say, oh, we can save 20% of administrative work, that's not going to sell it. No, I think you know what the person that we had on said, which is the alternative cost of not using AI. Saying if we don't use AI, 10 more people will be killed next year is so much more powerful yes, but that is harder to sell for me personally, because I think that in general.

Goran Cvetanovski: 45:11

So I cannot save 10 people today. Um, give me a second, just to think something. I cannot save 10 people today if, uh, if, I have too much to do. I mean, keep in mind that all the doctors after.

Anders Arpteg: 45:28

I think we mean the same thing. You know, because we can remove some of the administrative work from nurses, means that they can spend more time with patients, patient care, exactly that means that they can save potentially 10 more lives exactly it's just a way of communicating about the same thing.

Henrik Göthberg: 45:43

Yes.

Goran Cvetanovski: 45:44

I agree Okay.

Henrik Göthberg: 45:47

But listening to you both, I think there's two angles to this and I agree with both, because sometimes as a business leader in the smaller stuff, you need to be concrete. On what can we improve?

Goran Cvetanovski: 45:57

What is in for me?

Henrik Göthberg: 45:58

Yes, but from a financial budgeting strategic view of getting the politicians lined up and getting into a media debate that actually gets attention.

Anders Arpteg: 46:09

your ways were better, so I think in general we're really bad about communicating yeah, we are benefits about ai. Even we can agree on that.

Henrik Göthberg: 46:16

We can agree upon that, because I even, even I can highlight how we sort of, oh, we need to talk about value. We say, right, but what do I see in terms of talking about AI down in the corner, and then we then make a major leap to something very fluffy defined as value.

Anders Arpteg: 46:32

And we're not very good at connecting the dots, we easily get stuck with technology kind of super cool stuff, but it's not really going to sell it, no.

Henrik Göthberg: 46:39

So I think in general we are really bad, so we're talking about the technology cool stuff and then we are really bad. So we're talking about the technology cool stuff, and then we try to make a superficial leap to value instead of really trying to connect. What is the stuff that really contributes to value, which is not the technology, it's not the invention.

Anders Arpteg: 46:54

But it's not important though.

Henrik Göthberg: 46:56

No, the invention is important Schumpeter, invention, innovation, diffusion. This is a three-step rocket the techniques put into a commercial setting that actually innovates, put into an operational setting that is adopted at scale, you know.

Goran Cvetanovski: 47:15

But what else? And I have final one because I can talk about this.

Henrik Göthberg: 47:19

I know it's super good.

Goran Cvetanovski: 47:21

The final one was for me, for all European listeners. So this episode, which is basically, I think, episode 135, hpc insize companies understand that there is a European high computing clusters that you can tap into for free to innovate with, and that is like a very, very big incentive for the European startups actually to operate on. So we have the sandbag for us to innovate. But this message needs to come and keep in mind if you're in any country than Sweden in Europe, you can find a representative person in your country that has access to these supercomputers, and it's European Commission's what is called ambitions to get Europe to innovate fast. Coming from this, I want to conclude just that things are really happening. I can see it now. We have been talking for four and a half years, at least in ambition level, that there is some kind of a vision or a future that A Europe needs to wake up. We have some things that we can actually put in place right now to wake up. We have some things that we can actually put in place right now. Let's innovate. Of course, we have the GDPR and the AI Act and all these things on the side, but that doesn't stop us actually to be still progressive in the world and talking about Antonosika and Fevr right and many of the startup companies that we have, and we were on this workshop. We were in this workshop last week and for me it was a little bit strange how we don't realize what we have. Instead of that, we are chasing after the giants all the time. I'll look at where they are, look at where we are, instead of thinking like, okay, how can we actually compete with them? How can we actually make something right now without thinking too much? Just basically start doing it. Why do we need to chase somebody else? The biggest enemy that we have is ourselves. The biggest ambitions that we have is ourselves, and we need to be better in the nordics and in sweden, to be much better in just taking that pride and said like, hey, we are still the number two startup. Uh, what is called incubator in the world after silicon valley is here in stockholm. Why do we need to say all the time how bad we are?

Goran Cvetanovski: 50:01

Databricks one of the funders is from here. We have a lot of open source contributors from here. We have the best universities in the world here. We have great, intuitive, very smart people here. Why are we talking all the time about problems. Let's talk about opportunities. That is what makes basically the future brighter. Let's not talk about problems, let's talk about opportunities, and I think we have opportunities everywhere. Right now, it's raining money. The only person that actually gathers the money is the person that will start inventing like, okay, if I take the umbrella and tweak it other way around, I can actually collect the money that is raining. We are keeping the umbrellas as it should be, so they're bouncing off our umbrellas. Let's stop doing that and let's move on. All right, I'll leave it to you now, guys.

Henrik Göthberg: 50:57

Do you have?

Anders Arpteg: 50:57

any other favorite uh angles or episodes, perhaps um alexander norian still yeah, we need to mention alexander I think you know what he did with um, with the documentary series of six episodes of generation. Ai was really cool. Um, they had what was it? They had six episodes, something creative, I think.

Henrik Göthberg: 51:18

The teacher, the dystopian kind of future, the doom, the utopia, the paradise kind of situation and, yeah, different kind of settings what could happen in different scenarios, but to have a Swedish television like the Swedish BBC approach of a news business tech reporter going deep into our community and having an opportunity to, in my opinion, interview some of the greats in different ways and packaging it, communicating it well yeah, I mean, I think this is really a good example of good journalism, something I dearly miss sometimes and I can understand partly that.

Anders Arpteg: 52:04

you know journalism which has come gone down, I think, in the last decade partly because of they've been fighting against social media and they they have to chase clickbaits all the time and having this kind of more in-depth kind of half a year kind of work that he did, speaking to the top people in AI and then producing this kind of documentary series, I think is great to see and to have him here on the pod.

Henrik Göthberg: 52:31

There's a couple of things that stands out. First of all, the key, key insight how much better we need to be at communicating and what happens when you have a communication master at the in the pod and picking his brain on how did you think about this? Or how do you go about thinking, pitching this and all this. So this is number one. The other major standout for me was that someone that is not sort of entrenched in the industry did such an awesome job to navigate the space and also communicate it in an approachable way. You know this is one of my hardest topics on a daily basis. You know you ask me, Henrik, what do you do? And it's super hard, right. How do you extract out what is the key message that is worthwhile and that sort of connects, and to have him here is a masterclass in this skill.

Anders Arpteg: 53:33

That was amazing. Yeah, and I think that's communication at the best.

Henrik Göthberg: 53:41

It's one of the highest levels of communication in Sweden.

Goran Cvetanovski: 53:45

I have one more, and I will leave it this one to you. Actually, that was the episode where we had these three people from Ex-Potifiers Mattias Altin, niklas, Modig and Henrik. Knieberg, I think that that was.

Henrik Göthberg: 54:01

All right, let me set it up. So this is actually one of those episodes where, also, I'm a little bit starstruck, because if you think about Mattias Altin, niklas Modig and Henrik Knieberg, you need to pick apart who they are. So if we start with Henrik Nyberg, who's been at the forefront of Agile Before Agile was you know, maybe he's right at the cusp when the manifesto is written and communicating with the people in the US. Who's thinking about this? People in the US who's thinking about this? And what Henrik Nyberg did was he joined Spotify super early and basically came to almost like a research lab for Agile, where the whole culture was Agile at the core, not trying to learn Agile, starting out Agile.

Henrik Göthberg: 54:51

And there's this famous story about Henrik Nyberg working together with Oskar Stahl and how they are trying to. Holy fuck, how should we scale agile now when we are going from one team to 60 to 200 teams in a very short period of time? And out comes one of the most beautiful communications ever seen on YouTube on what is called the Spotify model and how we do agile at Spotify? So these drawings and stuff like that. And now, by the way, he trumped that in terms of views by doing the same story all over again, explaining AI and generative AI. So he's once again a communicative master, but also a worldwide renowned guy on Agile.

Anders Arpteg: 55:39

An awesome communicator.

Henrik Göthberg: 55:40

An awesome communicator.

Anders Arpteg: 55:43

That just signifies the importance of being able to communicate. It doesn't matter how knowledgeable you are if you cannot communicate about it.

Henrik Göthberg: 55:49

Yeah we get this. So that was Henrik. Then we have Niklas Modig. Niklas Modig is one of the most leading experts on lean and he basically when he was doing his research in PhD he was embedded in Toyota on the Toyota way, lean, kite Sea and everything we talk about and sort of went deep and wrote one of the strongest books ever written on lean which is actually not the biggest or thickest book but the most approachable book on lean and started a company and in my opinion in Sweden is top in terms of discussing lean. And then we have Mattias Altin, who came in sort of a little bit later in Spotify, worked in Spotify, came with a nice pedigree from FinTech and all that. And you know I was scared.

Henrik Göthberg: 56:48

The story is unfolding here in the podcast, which is beautiful. Unfolding here in the podcast, which is beautiful, where Mattias Altin is questioning some of the ideas in the Spotify model and basically he's a little bit scared of saying that because he has profound, you know, iterative and in the podcast it's so clear. Henrik Nieberg says we always understood the Spotify model as a snapshot point in time, how we were thinking about organization and people took that and did that as a hard copy. So when you came and you wanted to improve on that. I applauded that and we could see it in the pod a release, something that Mattias had thought about for 5-10 years. Oh my god, I was stepping on some toes here and Henrik said in the pod she was not stepping on toes, you got it beautiful, so good. Fantastic episode, fantastic episode and and and the level of how we come, we conversed and around lean, agile and organization of software and data and ai top notch, absolute top notch.

Goran Cvetanovski: 57:54

We cannot conclude this without that episode. I think it's one of the most epic ones, for sure.

Henrik Göthberg: 58:02

For someone to get a condensed view from the source. I like it. Really proud of that episode. Really proud of that episode.

Goran Cvetanovski: 58:12

All right, I also have one more. So we did some sentiment analysis of the data and it seems like uh, henrik has stopped talking about goosebumps and andres can stop talking so much about dylan musk, which is super good as well, right so but we haven't even mentioned that we did a weird podcast on stage with cassie kusorkov, you know to name drop yes, that was very good.

Goran Cvetanovski: 58:39

We did this during the data innovation summit, so it was super uh cool to have her uh on that uh and do and do a podcast on stage on stage with 3 500 people. So that was that's.

Henrik Göthberg: 58:50

That's a different podcast for me. Well, I mean, we're going to repeat it.

Goran Cvetanovski: 58:53

That's a different podcast for me. Well, I mean, we're going to repeat it again soon, so it's not going to be a big problem.

Henrik Göthberg: 58:59

Are you going to keep up? It's a good format.

Goran Cvetanovski: 59:02

Oh yeah, I think that we should. We had a lot of opportunities this year to catch up with some great people. We had, like Mikael from Meta this year on NDSML. We could have done a podcast there Something that haunts me this year but we will make it happen.

Henrik Göthberg: 59:19

Because usually our conferences are on Thursdays, we have some top names.

Goran Cvetanovski: 59:21

Yeah, and then we had some people in data as well that were on data 2030., so it was a good year. It was a good year. I'm very happy with this. But, anders, maybe you can guide us through what happened from a technical point of view, because it was completely. That's why I said, if we are looking at the episode six months ago, it almost feels like 10 years ago in development of AI. So, lead the way.

Anders Arpteg: 59:46

Well, as usual, the news and the technical project or progress in AI is going faster and faster, and it's certainly true for the year of 2024 as well. What is that saying, they usually say? Sam Altman says progress will never go as slow as it does today. Yeah, profound, it's kind of fun.

Henrik Göthberg: 1:00:06

It's a t-shirt.

Anders Arpteg: 1:00:07

Yeah, anyway, I guess we have to take the Nobel Prize. It's such a big thing let's start there I like it.

Goran Cvetanovski: 1:00:17

Let's start with the end. Let's start with the end. End, yeah, because they had a nobel price ceremony after that it was the week of air release, when everybody's yeah, okay, so in the midway let's start with the big one, I think it's.

Anders Arpteg: 1:00:30

It's one of the biggest ones and for one it's a tribute to AI, although there is no prize in the Nobel Prizes for computer science or AI or compute in general, but still AI won two prizes, both one in physics for some weird reason, and one in chemistry. I think physics is really weird. It was John Hopfield and Jeffrey Hinton then the right people, but the storytelling is. Yeah, jeffrey Hinton, absolutely, but he got the prize for Boltzmann machines, something that is not used at all today and hasn't impacted AI.

Henrik Göthberg: 1:01:07

And he himself was always embarrassed, but you told me the story about.

Anders Arpteg: 1:01:11

You told this story, yeah, well, during the Nobel press conference when he got the news. He basically got the question you know, what impact did your invention have of Boltzmann machine on AI? Nothing. But I did the other thing called in the back propagation, which has a huge impact and it's, you know, it's used in all kind of AI models of today and it must have felt really bad for the Nobel Committee to hear him say that on the.

Henrik Göthberg: 1:01:39

He didn't even want to play ball, because they're squeezing it into the category of course.

Anders Arpteg: 1:01:45

But he's a really cool person and you saw as well, goran, you know, when you heard him. He's a really funny and cool person.

Goran Cvetanovski: 1:01:52

It was a guy that you want to have a beer with. Probably he would take a beer with you he will take a beer and he will say he will speak his mind yeah, but he was speaking his mind all the time on the conference. He was not, he was unhinged at all.

Anders Arpteg: 1:02:06

Beautiful guy, man, beautiful person and just you know, without going too long into this, but he was one of the godfathers of deep learning, you know, together with jan lecun and and uh canada, canada crew yes, the canada mafia the deep learning mafia.

Henrik Göthberg: 1:02:22

Mafia, not the crew godfathers.

Anders Arpteg: 1:02:32

But anyway, you know what he done. He also was working at google for a long time, but he left google for a couple years ago and one of the reasons he said he left Google was because he's very concerned with AI. And the reason he got started with AI to start with was rather from a neuroscience point of view. He wanted to understand the human brain and he thinks the best way to understand something is to build that. So he wanted to build the human brain and he knew that backpropag propagation, which he was part of popularizing, wasn't the perfect way. He was sure that he knows that's not how the human brain works, but it's the best model that we have at least. But he knew that that's not really how the human brain works. But then in the last couple of years he basically he was speaking about the analog intelligence and digital intelligence, where the analog refers to the human brain, and he basically said okay, I know it's different, but maybe the digital intelligence is superior to the analog version.

Henrik Göthberg: 1:03:25

And that made him really just react and change his mind a lot. Then we don't really know what we're dealing with, because we're inventing something that is maybe superior to what we can think of ourselves but I think it's more like instead, instead of.

Anders Arpteg: 1:03:42

you know, I think the the comparison to the plane and the bird is is accurate here exactly. And then if you want to have something you know super powerful in terms of transportation, you know airplane is more powerful than a bird, and the same can be argued here. Backpropagation in the digital kind of intelligence that the deep learning networks have today is potentially significantly more effective than the human brain, and we know that from a knowledge point of view, there is nothing that beats Chateau-Petil and large language models today. They lack in reasoning, but in terms of knowledge recall, it's like no way humans can compete.

Anders Arpteg: 1:04:21

So now he's focusing a lot on the risks of AI, and that's what he wants to focus on, and he spoke a lot about the loss of control. We can speak forever about this. That was one, so let's go. Yeah, so the other Nobel Prize was then to AlphaFold and DeepMind and Demis Azabis and the other people that worked on this awesome deep neural network model that were able to predict the 3D structure of proteins. So AlphaFold.

Henrik Göthberg: 1:04:49

And who is the Swede on the team? That is sort of a little bit of a hidden famous person that we should have on the pod. Even, yeah, I've been large.

Anders Arpteg: 1:04:59

Yes, yeah but you know, I think it's an awesome example they actually open sourced all the inventions before. You know, being able to understand the 3d structure of proteins is really important because that really tells what the functionality of it is, and usually it took like four or five years for a single protein to have their 3d structure identified. Now we can do it in subseconds and they even took like 200 millions of all the known proteins, did the prediction of the 3d structure and just put it out. Put it out there for free to use for any science.

Henrik Göthberg: 1:05:35

So what magnitude of work are we talking about? We're not talking 10x, we're not talking 1000x, we're talking way more.

Anders Arpteg: 1:05:43

It's mind-blowing, amazing contribution to science, to the world of building knowledge. In terms of, is it chemistry? I would say it's more biology, but it's interesting is it chemistry or biology?

Henrik Göthberg: 1:05:56

You would argue biology, okay, but this also. Here we have a computational technique. It highlights the computational techniques in the actual field. That is then sort of the Nobel Prize. So this feels more accurate.

Goran Cvetanovski: 1:06:11

Yes, this feels more accurate. Yes, this feels more accurate. We also need to mention that there is a limitation in the Nobel Prize, right, because there are only five categories so far. Right, and that is basically how the world was perceived at that point of time, and it's true that they want to remain the legacy, but the world has changed, so maybe it's time to.

Henrik Göthberg: 1:06:32

Yeah, this was the conversation we had on the pod. We had this on the AI News and is it time for new categories? Yes, that reflects the research. It's been a hundred years, but yeah, maybe it will happen, who knows?

Anders Arpteg: 1:06:46

Anyway, it's weird that we didn't get a prize in economy or in the Peace Prize or literature. Perhaps you know with ChatEBT in economy or in the peace prize or literature. Perhaps you know with chat ebt in the future the next year's nobel prize in literature will go to chat ebt, who knows? Maybe not. Well, I don't think so that is so provocative.

Goran Cvetanovski: 1:07:07

It can be that, uh, you know the the brush that we were talking about earlier all right, but other topics that were sort of big.

Henrik Göthberg: 1:07:15

I mean like so this was the Nobel Prize. Okay, If we try to do the year in review and pick out a couple of things to spot down on, yeah.

Anders Arpteg: 1:07:26

I mean I think also. I mean it's a couple of things now that's happening late in the year, but still, I think the Swedish AI Commission was a big thing as well. It's been working around the clock, so to speak, for the whole year of 2024. I've been trying to help them as much as I can and giving some recommendations, but they came out with a report earlier than they should have.

Henrik Göthberg: 1:07:47

Which is a nice gesture.

Anders Arpteg: 1:07:49

Yes, showing To actually do something ahead of time. Urgency, yeah, and they basically did because of urgency, as you say, since they came to the conclusion that Sweden is not only, I mean, they are lagging behind and we need to take action as soon as we can, and therefore they tried to publish this report, a lengthy and very well-packaged report, 130 pages or something like that, did you?

Henrik Göthberg: 1:08:14

36.

Anders Arpteg: 1:08:14

Yeah, something like that. Did you read it? Yes, I actually did.

Henrik Göthberg: 1:08:18

I actually did not finish it, but I started in a strong pace.

Goran Cvetanovski: 1:08:23

You were dedicated.

Henrik Göthberg: 1:08:24

I was dedicated for half of it. You know what is funny.

Goran Cvetanovski: 1:08:29

This is basically Henrik also. In a nutshell, he's like Goran, you need to read Plato. It's like okay, and I went and bought the book. So I bought the collected works of Plato Shut up.

Henrik Göthberg: 1:08:40

Don't sell me out like this. Don't sell me out like this. I'm not selling.

Goran Cvetanovski: 1:08:43

And then I'm getting this book and I'm starting basically reading before I go to bed and I cannot finish more than two pages of that.

Henrik Göthberg: 1:08:51

It's heavy shit.

Goran Cvetanovski: 1:08:52

It's some heavy shit and I cannot actually move to another page because I cannot. If I haven't understood the first page, I cannot move to the.

Henrik Göthberg: 1:09:00

I was like strong commitment to reading, but it was some great stuff, yeah, yeah but so here we have the AI commission, and the backdrop is, of course, that it was commissioned from the government in Sweden and it's in relation to creating a more clear AI strategy in Sweden and maybe even a clear AI budget and everything like this. That would be awesome if they did. What was the standouts in the report? For me, there was a couple of things that I thought was quite cool. Stabsleg was fun.

Anders Arpteg: 1:09:35

Yeah, the task force thing.

Henrik Göthberg: 1:09:39

Yeah, this will never happen. But to show the urgency in a political report and recommend that the government goes to stabsläge, like go to high alert, defcon 4 level on a certain topic so you can put a taskbar right under the prime minister to get moving and then stabilize, I think it's the right, I think it's a very refreshing. I don't think it will happen, it won't happen, but it has signals value.

Anders Arpteg: 1:10:09

Yeah, and I think it's great that they actually have a very. I mean they came out with like 75 concrete suggestions and a budget. Yeah, they actually have a very. I mean they came out with like 75 concrete suggestions and a budget, yeah, and actually did the work, you know, to try to estimate the cost for each of these actions, and that was really really well done. You know they did their homework for it. So that's, you know, very concrete, a lot of very substantial, actionable, you know suggestion that they came out with. Then I don't agree with everything they said. I mean, I think it's great that they recognize that the tech giants of the world, the AI gap that we have been speaking about so many years here- Four years, to be precise.

Anders Arpteg: 1:10:49

Yeah, they recognize that and of course they are much further ahead in the US and the few selected tech giants and also the Chinese ones. But then the actions they suggested, saying we should build our own infrastructure, not building on top of their top cloud providers it's not a practical and useful suggestion in my view. So I think that's the really kind of naive and arrogant to even think that you can build something that they spent billions and billions of dollars into bringing up.

Henrik Göthberg: 1:11:23

But let's go here, because when we had this on the AI news, when did that happen? I don't think you were here. I can't remember Were you here.

Anders Arpteg: 1:11:32

I'm not sure.

Henrik Göthberg: 1:11:33

No, here, I can't remember you. Were you here? No, but my take, if I'm a little bit more, um, I was not referring it to that what they were doing was wrong. I, I was trying to pick it, you know, okay. So this is all great stuff. We can tick a lot of boxes here.

Henrik Göthberg: 1:11:48

What are the blind spots in in the commission setup and the blind spots in what was not mentioned? And we had a very nice conversation here on the pod when we discussed this and I stand by it. If you look carefully at who is in the commission the main characters in the commission you have academia and you have CEOs and you have politicians. You don't have one single proper engineer. You don't have one single proper engineer. You don't have one single proper CTO of a company and you don't have one single representation from, like, a CTO of Spotify who could tell it like it is.

Henrik Göthberg: 1:12:26

And that is actually reflected in the way the report is structured and it mentions all the important stuff that Sweden needs to do from an academic point of view and it reflects how the politicians and how the CEOs think about this. But when it comes to building product engineering or building the fundaments correctly, like if you had a CTO with deep knowledge how it works, they would say we need an AI factory, but they would have been much more nuanced. We should not build our own infrastructure. We should build our Swedish layer on top of whatever Stuff like this.

Henrik Göthberg: 1:13:06

I think, that's the core gap, this engineering gap, and I think they're missing a point.

Anders Arpteg: 1:13:12

I mean they're trying to say that, okay, we have the tech giants that are leading to a large extent, but then they say we need 600 new doctors in.

Henrik Göthberg: 1:13:21

AI To do what?

Anders Arpteg: 1:13:22

If we compare it to the tech giants and say how much investment do they do in engineering versus research, it's probably at least a factor of 10 times more in engineering compared to research, and then we have, if you look at the report, zero.

Henrik Göthberg: 1:13:36

I would say in terms of engineering. So we have a very unbalanced view on this. And then, even to prove this point not with ourselves, but with the most knowledgeable people in Sweden we had Sverker Jansson in here. So Sverker Jansson sits heading of the AI Center in RISE, but what is more important is his pedigree dating back to setting up SIX. What's SIX?

Anders Arpteg: 1:14:02

RISE is the Research Institute of Sweden.

Henrik Göthberg: 1:14:04

Yeah, research Institute of Sweden. But in the early days it was the SIX. What does SIX stand?

Anders Arpteg: 1:14:11

for Swedish Institute for Computer Science.

Henrik Göthberg: 1:14:14

Yeah, what does 6 stand for? Swedish Institute for Computer Science? Yeah, so this is where data and AI as an institute grew up before it became RISE. And when he was on the pod he said something profound I think he was talking about. Well, the real difference that is so hard to copy is how the tech giants and the startup community in Silicon Valley has been churning out data, ai, elite data engineers, ai engineers. So it's not about the research they've done in Stanford, it's about how many are exposed to what elite AI and data engineering looks like, and that is the real money. This is the real game. And so then, that is an engineering problem more than a research problem, and this is I don't that nuance. So nothing is wrong in the report, but there is a spot on engineering and and the profoundness of what swariki could use. He said it like in. He said like, yeah, it's so obvious, but it's not obvious if you're not an engineer yourself. In my opinion anyway, sorry, rant, it's a rant, yeah, but we had guests who? Who makes this point? So well, what? So? Okay, that was the commission.

Henrik Göthberg: 1:15:38

Okay, we need to talk something about all the tech releases. Which ones do you want to highlight here we had the final 12 days of OpenAI in the lately in competition with Google's releases simultaneously. And then, of course, if you look at the whole year, we have the whole strawberry debacle. We had the whole. You know what we anticipated, what we actually got in the 01 release. We had Antropic with the computer. We can now run a. You know we can release the agency. Computer use, computer use. What else else you know? So I'm used. Which one do you want to pick out of which?

Anders Arpteg: 1:16:21

in a year pet thing I've been speaking about, I think, for a number of years. This is kind of latent space reasoning part, and this is literally what 01 is starting to do. So I for me, this is a very important and big trend that I think will happen going forward. So, in short, 01 no one really knows exactly how it looks like because they haven't released a technical report that that gives details about it, but it is doing something else than just the large language model approach for it. So then, if that's a star thing, the self-taught reasoner or if it's just test time compute, I would say it's more the star approach. But in any case, we need to have something that can take multi-step reasoning approach for not being as limited in terms of reasoning as the LLMs have been. So, going from the type one type of reasoning that we have today in LLMs to a type two type of deliberative reasoning, kind of thinking fast and slow.

Henrik Göthberg: 1:17:31

Yes, type one, type two Kind of thinking fast and slow.

Anders Arpteg: 1:17:34

Yes. So this, I think, is a super, super big trend, and we have just started to see the beginning of this. This is something that we'll continue to see for 2025, where we will see improved levels of reasoning, which will also reduce the demand for these kind of super big models that we have today, with super big data sets being trained on. So this is a super exciting part, which I think all models now, even Gemini 2.0, which was just released last week also have this kind of additional reasoning mode. They call it advanced reasoning or something, but all of them are taking this approach.

Henrik Göthberg: 1:18:13

But help me out here, because we you talked a lot about this and I even read an article I was trying to find it right before how now there was a paper of I can't remember it was google, who was, who was how they, how they're now thinking about how you switch between the data space and the latent space. So you need to switch back and you do some part of the prompt, the way we have done it, and then you switch to a higher reasoning latent. Could you elaborate what do you mean with the latent space reasoning?

Anders Arpteg: 1:18:45

I like the Jan LeCun kind of JEPA approach for it and that says it rather clearly. And I think if we want to move and bring the multimodality into a really functioning foundation model in the future, we need to move to a space where text, audio and images are the same. And we know, in the human brain we have some kind of abstract concepts and we know in the human brain we have a prefrontal cortex where you can have this kind of deliberative, conscious type of thinking and that's what's lacking a bit in the models of today. So that's what we're trying to to get more into. And that means that you know, instead of we have what we've done in the past is simply to predict the next token in token space. Yes, token space versus latent space. Yes, we could move into doing the next token, if you call it that, but in latent space instead, and this is basically what, yep, I is saying. So they they're speaking about energy space or they have this kind of energy latent space.

Henrik Göthberg: 1:19:44

For me, who doesn't really get the technicalities, can I understand it as different abstraction layers, levels, so you're working on the concrete token space, on the concrete task abstraction, and then you're trying to find, you know, like multimodality You're trying to find a similar, you find the semantics that you can.

Anders Arpteg: 1:20:04

I think it's easy to understand if we speak about images and pixels, for text actually does work rather well to work in token space, because tokens in text space are rather high abstraction level, but pixels are not.

Anders Arpteg: 1:20:16

Pixels are super low in abstraction level and super high dimensional. So if we were to predict just the next piece of pixels all the time, it would be super slow and super inefficient. So we need to move back into some kind of compressed space and that's why we're seeing all this kind of image generator that we have today the text-to-image generator and the video generator like Sora, or we have the image generator like Dali or Imagine or Midjourney. They all basically have an auto-encode around it that takes the images and compress it down to a latent space and then they do the diffusion, the transformer diffusion in that latent space. So they already operate in a latent space and this is where all the other models need to move to, because it's much more efficient. And to me it's obvious that you need to move to this space and do the reasoning, if you call it that, the planning, at this kind of latent space, and this is what O1 is starting to do as well.

Henrik Göthberg: 1:21:18

And so, therefore, you're moving and doing some of the work in the compressed space, the latent space that then directs what you do in the token space.

Anders Arpteg: 1:21:29

I mean, just think about the human mind when you think about how I should act or respond to a question in like a discussion that we're having now. We simply do not use reactive kind of thinking. We actually do consider a couple steps ahead, at least at least a couple of them, even though the human mind is very limited. Still, we do consider in the prefrontal cortex some steps ahead before we speak. At least some people do, maybe not all people.

Henrik Göthberg: 1:21:58

I don't I think you do?

Anders Arpteg: 1:22:00

Of course you do, but anyway, it's a joke. So I think this is where we need to move to and it's kind of obvious. And then if you take the JEPA space, that also have the hierarchical kind of JEPA.

Henrik Göthberg: 1:22:11

So JEPA was then the summary. This is the Jan Likund paper position paper on where we need to go yes, and JEPA stands for Joint Embedding, predictive Architecture.

Anders Arpteg: 1:22:21

Yeah so joint embedding meaning you move from the sensory space into some embedding space, a latent space, and then you do the predictive architecture in the embedding space and we saw some of the.

Henrik Göthberg: 1:22:34

This is already a year ago when we saw some of these talks we are trying to look at. It's not one big model, it's more like different functions or different features working together, which resembles our brain. Yes, it does Work.

Anders Arpteg: 1:22:50

So O1 is moving in this direction and we will continue to see progress, I think in 2025 for this, and that will be exciting times, but I think O1 was a super big part of it. It has been similar work in the past with self-taught reasoning, et cetera. Now everyone else is following, gemini is following, claude is already there as well with additional reasoning, so obviously this is a trend that everyone is following as well with additional reasoning.

Henrik Göthberg: 1:23:16

So obviously this is um a trend that everyone is following, and we had some of the last um talks that sort of that was trending on youtube, used, I guess, the last week I can't say his name uh, exactly, if we were sort of you know, predicting, you know the way we have understood pre-training, it will be something else. The trajectory that we've been on that took us where we are is not the same trajectory.

Anders Arpteg: 1:23:43

from 2025 onwards, scaling will not continue to work by simply adding more parameters or more data. It will need a change in the algorithm or the architecture as well.

Henrik Göthberg: 1:23:51

Yes, and he was highlighting the example of, okay, the compute laws and how we can be more and more efficient. We are there, but we are getting to sort of the finite view of the data. He uses terms like the data is the fossil fuel that is finite. So he used a funny rhetorical metaphor that I liked.

Anders Arpteg: 1:24:16

that makes sense, yeah but I think the best way I would explain it is to say that we have large language models that have a lot of knowledge today but poor reasoning. Then we have actually models that do reasoning really well but have really poor knowledge, which is the AlphaGo of the world. This is your story. Then we know that in in alpha go we can play chess or go, uh, in in a very, you know, advanced way, much better than any human, literally much better than any human, doing moves that no human could even think, think of, like move 32 when they play lisa gold in the in alpha go, like Mu32 when they play Lisa Gold in AlphaGo Chess Master Championship. So we know it can be superhuman performance for that type of reasoning tasks. When it comes to very specific, specialized tasks like chess and Go.

Anders Arpteg: 1:25:10

Now what you should do is you want to have that type of reasoning combined with knowledge, and this is literally what people are doing now, and we know that for AlphaGo and AlphaZero, it was about the reinforcement learning kind of techniques, and this is actually what O1 is doing. I mean, if people are not seeing this, then you know. This is obvious. This is what you should combine and this is what we're seeing.

Henrik Göthberg: 1:25:30

But now let me not challenge you, but let's bring in the very refreshing angle from Anton Osika on this, because now we are talking about how can we continuously improve the way we do the model. And then when we were trying to press him and discuss with him and actually I got more of his point when I was listening myself to the podcast afterwards in my car and his main argument for 2025 was you know what? And now let's go away from the frontier and the Sam Maltmans and this and go back to us. His basic argument was that you know what, in order for us to get value and improve the efficiency and productivity of AI has very little to do with stressing the models further, because we are so low in our capacity to build compound AI systems. So he was actually trying to convey to me and you and we missed that a little bit in the conversation it stood out more that he was trying to tell us that the way I look at how I improve GPT engineer, lovable, you know what.

Henrik Göthberg: 1:26:48

Model here, model there, yeah, this is quite important, but the engineering pieces, that's where the real blind spots and these sort of performance gaps are right now. That's the real problems to work on normal value. So that is also a little bit like backing out of this conversation. You know what? That is fantastic, and this is where the frontier should go. What should we do? It's an engineering problem still that we haven't at all mastered. What do you think about that comment?

Henrik Göthberg: 1:27:18

yeah, of course I agree that engineering is is what we're lacking ultimately, if you go back to the, what should sweden be better at? It's engineering man, yeah, interesting, okay, um, where are we in time? You know how do we want to wrap this up, do? There's so many other topics, I don't know if we should pick a couple of more it's a new year, so we need to add some with some predictions, I guess yeah yeah, but are we, are we done?

Henrik Göthberg: 1:27:46

because there's so many more key topics for the year, but we could stop there.

Goran Cvetanovski: 1:27:50

We could go one more there was definitely a lot of other topics, technology wise.

Goran Cvetanovski: 1:27:56

So you got like microsoft and co-pilot releases and agents within co-pilot yeah, we haven't talked about agents yet we didn't talk about how google is trying to beat the race, but they are already late, because they released the the paper, but now they're late, and etc. So there is a quite a lot of things, but in conclusion, I think that there is uh, it's been things, but in conclusion, I think that there is it's been a much more interesting year in development than it was in 2023.

Henrik Göthberg: 1:28:23

Is that true?

Goran Cvetanovski: 1:28:24

For sure. I was just looking at the timelines of the releases and et cetera Just this December. I think it was crazy. It was a one week within everybody released new models yeah so you can see that right now it's like a really competitive edge for everybody that are working in this field there's a fun anecdote here.

Henrik Göthberg: 1:28:47

do you remember how we were planning? Let's do an episode to speculate on what is q star, speculate on what is Q-star and what is strawberry. And you were there prepared with Jesper, and it was literally the day after we did that speculative podcast, they released O1. And okay. So now I can ask you did O1 live up to the expectations of the speculation or was the speculation on point, because this was speculation like one or two days before?

Anders Arpteg: 1:29:20

I'm gonna, I think, the technique behind it, even though we don't know the internals of one. It's basically the star approach. It's basically what everyone thought. You know the q star would be the star kind of thing. So but in that sense then the question is more did it improve in in skills that much? Did it improve in benchmark results? Did it improve in reasoning capabilities? What's been interesting, I think, when they released the 01, the full 01, compared to the 01 preview and the 01 mini that they released earlier. The 01, the full version did not actually improve that much when it comes to compared to 01 Preview, and that's kind of interesting. But it did improve in speed engineering kind of KPIs rather than research KPIs, and that's kind of interesting.

Henrik Göthberg: 1:30:12

Yeah, and then you said it I think we talked about this with Anton that O1 is super over-performing or fantastic in very narrow type, really proper reasoning topics. And once again Anton was like when do you really need it for the practical? So it's hard to judge, you know, because if you go in and try, if you try O1 with the wrong problem, you don't really see it. But if you try it with the right type of problem it's fantastic.

Anders Arpteg: 1:30:46

But the full release of O1 tries to make that classification a beginning. So for simple problems. If it's just a factoid that you need to retrieve from the knowledge, then it does it very quickly and then the speed is up again.

Henrik Göthberg: 1:30:57

But if it's more of a reasoning task that requires you know thinking to a much larger extent, and then it will do so well, and how did you reflect on the price hike for this pro model, like 200 dollars or something ridiculous like that, trying to show that it's super something else, what is your take on that?

Anders Arpteg: 1:31:19

well, if you get more compute time, potentially that's what you get with the pro, with the 200 per month kind of thing. Okay, there's a fundamental cost topic here yeah, I mean it is very costly to have this kind of extra test time compute or the inference compute that you get. So, unless you have some kind of other business model around it, I can understand OpenAI's view on this.

Henrik Göthberg: 1:31:43

It runs a lot more compute potentially.

Anders Arpteg: 1:31:46

It's very costly for them to operate a one. So if you want to have this huge amount of reasoning and do the really advanced kind of questions, then they need to have another price model of reasoning and do the really advanced kind of questions, then they need to have another price model than selling everything for $20 a month which you get. You know you get the full, you get the O1 in the ChatDB, t+, but with potentially some limited time for the more advanced reasoning tasks, and then you get, you know, more time in the pro. I can sort of in some way get it, but I think, yeah, it's very expensive of course it's also because it almost becomes a marketing stunt.

Anders Arpteg: 1:32:24

Oh, 200 dollars, wow, it must be super good I think it's fun if we just speak about, you know, the open ai 12 days kind of releases, the Shipmas thing.

Henrik Göthberg: 1:32:35

So they have now until Shipmas, shipmas.

Anders Arpteg: 1:32:37

Yeah, ship one every day until Christmas, and they first released the O1, the proper big one and the pro subscription as you spoke about, but then they had an awesome new fine fine tuning approach. It's a second release, which was really fun, I think, and that is really cool. So you can, with reinforcement learning, do fine tuning with just like 12 exams or something, and and you can take the mini version, number one, which is 10 times more efficient and cheaper, and make it work better than the full version by fine-tuning it for your own needs, and I think that for many companies that would be super valuable.

Henrik Göthberg: 1:33:20

So here we are coming to fine-tuning on how to make this useful in production.

Anders Arpteg: 1:33:25

From an engineering point of view, from an engineering point of view exactly. And then really Sora as well. So this is available for anyone to do video generation, also with improved engineering capabilities.

Henrik Göthberg: 1:33:36

You get Zora functionality now with your normal chat, gpt Plus yeah.

Anders Arpteg: 1:33:40

That's pretty cool.

Henrik Göthberg: 1:33:42

Not in Europe, but Not in Europe yet, but yes. And what about? You know, pick apart a little bit like because someone I was reading sort of oh, who had the better Christmas release party. You know, google did a lot of cool stuff as well and someone was sort of rooting that, oh wow, open ai and someone's rooting, oh no, no, google's release was more profound. So which one do you think stands out? What happened in the google space? I?

Anders Arpteg: 1:34:11

think open eye trumps time. Yeah, and they seem to time their releases so that Google's releases seems like nothing. Every time they seem to have some insider and a great marketing department.

Henrik Göthberg: 1:34:23

But I saw several arguments that this time Google did good, but you haven't changed your mind.

Anders Arpteg: 1:34:34

Not so far. I mean, it's a lot of things the Astra thing, the Mariner thing, the Jules thing. Astra, Mariner was the ones I was thinking about and also, yeah, Jules was some kind of you want to summarize what they were.

Henrik Göthberg: 1:34:48

What was the Astra? What was the Mariner?

Anders Arpteg: 1:34:51

I mean Astra was more. Think how you can take your phone and just walk around and it can actually understand from a multimodal point of view what you're seeing. It's like what OpenAI released with the advanced voice and video mode, so similar in that sense. Then, Mariner let me see if I remember it correctly, but I think it was more the browser experience, so you can actually control your browser and let it actually take actions with your browser, similar to computer use.

Henrik Göthberg: 1:35:22

Computer using traffic.

Anders Arpteg: 1:35:23

Yeah, which is cool stuff. So moving more agentic there as well, so that's really cool. So it's taking a lot of steps there as well, but a lot of engineering things been moving into. The Flash models is similar to O1 Mini but Flash has I mean, it is still copying OpenAI and getting it slightly worse.

Henrik Göthberg: 1:35:47

You are right now. You've always been a Google fanboy and right now you're an OpenAI fanboy a little bit more, I don't like OpenAI.

Anders Arpteg: 1:35:54

I actually, I'm just trying to.

Goran Cvetanovski: 1:35:56

It is called closed AI. It's not OpenAI.

Anders Arpteg: 1:35:59

I actually think OpenAI you know, given you know, if it's something that's happened to OpenAI in 2024, it's surprisingly high level of churn of people.

Henrik Göthberg: 1:36:10

Very big churn.

Anders Arpteg: 1:36:11

Ilja Sutskever left, then, you know know, mira murati left, the cto, basically the whole safety team left, more and more people. Some other person, high roller, left, now today as well. I think it's going south for open ai coming here it's tricky.

Goran Cvetanovski: 1:36:28

It's tricky situation yeah, so we already opened the predictions for 2025.

Henrik Göthberg: 1:36:35

No, no, but before we leave, I have one more way to reflect on 2024, and then we can go to predictions. Or did you have something else?

Goran Cvetanovski: 1:36:43

yeah, because I I think we are missing like uh um the chip battle yeah, the cheap battle.

Henrik Göthberg: 1:36:50

Oh, good points, good point, thank you. Uh, to do a couple in mind, that Amazon now is in play.

Goran Cvetanovski: 1:36:55

Yeah.

Henrik Göthberg: 1:36:56

I heard that, oh, the chip battle. Yes, the chip battle.

Goran Cvetanovski: 1:36:59

So it's quite a lot.

Henrik Göthberg: 1:37:00

The whole infrastructure battle the whole infrastructure AMD let's go, let's go, let's go. Chip.

Anders Arpteg: 1:37:06

I think you know, something that's been insane this year is simply the investments in infrastructure and Sam Altman started infrastructure and some old man started out saying, you know, in the beginning of the year.

Henrik Göthberg: 1:37:18

We need seven trillion dollars in coming years infrastructure for AI and and at some point Nvidia top tops, I mean.

Anders Arpteg: 1:37:22

Then Microsoft came out and said we going to invest 100 billion dollars for the Stargate thing in 2028, and it's already starting to roll out these hundreds of millions of data centers. Meta came out saying also, they're going to invest hundreds of millions of dollars sorry, billions of dollars. It's a big difference and going to get hundreds of thousands of GPUs in these clusters. Elon, you know and I actually like to mention this a bit more because I think that's the upcoming star now when it comes to AI in coming year Elon, you know, released this colossal colossus and it's their new data center with 100 000 h100 gpus.

Anders Arpteg: 1:38:05

And what they did is something even more astonishing, which is that normally, if you want to have a large training session for a super big model, you can't get more than 20 or 30,000 GPUs in sync. So when training, you need to basically have them all in sync so they all know what parameters to update etc. And no one has really figured out how to do that to an extent larger than 20 to 30,000 GPUs. But now Elon suddenly made it work for 100,000 GPUs, so more than 3X more than anyone else thought it was even possible to do.

Henrik Göthberg: 1:38:40

So there was this thinking thing. That was amazing, and then I saw the CEO of NVIDIA summarizing how the fuck they put up a huge data center in less than 21 days.

Anders Arpteg: 1:38:53

Yeah, in 19 days 19 days for something that the rack was put up until the first training, but then they had the whole data center built in 122 days as well.

Henrik Göthberg: 1:39:02

But both two metrics are insane to anything in mankind history. Computer wise, it's such feats that we can't.

Anders Arpteg: 1:39:11

So I think right now and it's just a few weeks back since they launched the Colossus thing.

Henrik Göthberg: 1:39:16

I think right now XAI Elon's XAI have the most efficient a data center for training ai yeah, so, and you're basing that on the topic that he's syncing 100 000 cpus, so he's three gpus so so he's syncing 3x more than anyone else with a new technology, together with nvidia as well, and even and even, jack ma, like he was, like he was, like he looked like a school boy. How do you do something like that in 19 days? What?

Anders Arpteg: 1:39:53

that's engineering, it's engineering.

Henrik Göthberg: 1:39:55

Right, it's not research, it's engineering.

Anders Arpteg: 1:39:59

Ok, alright so we're going to see a lot of infrastructure going to happen and everyone is building their own ship. Openai is building their own ship as well, yeah and this whole geopolitical discussion on.

Henrik Göthberg: 1:40:10

You know what's the cycles of getting off the um uh addiction to tm tms, tm tmc, tms tsmc semiconductor manufacturing company so. So basically, they had had a world monopoly for tens of years and now everybody is thinking about how do we get out of this situation. Monopoly and it's insane money in terms of your political power struggle that's going on and in, in in this this, in this bloodbath. At some point, I think, nvidia also topped the highest valued companies in the world from nowhere, boom spikes yeah, they, they've grown.

Anders Arpteg: 1:40:55

I'm not sure how much this year, but 300 percent or something it's not 300 percent. This year I think it was around 150 yeah, but yeah anyway they went to the top and surpassed even microsoft and apple for a while market cap okay.

Henrik Göthberg: 1:41:13

So that that was the semiconductor space. I want to do one last sort of stop down. Where are we right now in the conversation end of 2024, open source versus proprietary models? We need to say something about that. Where did we end up? Because we had predictions, we had conversations, we talked about ai, apartheid. We had like, say, like we need to make sure we go, we go open source, and then we have now hint on saying it's the most crazy thing we can ever do. It's too scary, you know so. So where do we stand?

Goran Cvetanovski: 1:41:50

it's 300 percent no, it's 161 percent 160.

Anders Arpteg: 1:41:56

okay, okay, yes, okay, but what you know how?

Henrik Göthberg: 1:41:59

would we make me so happy. But how do we wrap? How do we summarize the open source or open weight? Anton, also we need to talk about open weight.

Anders Arpteg: 1:42:09

Yeah, well, let's do that and also I think we should. If you put up the Tesla stock market after the Trump election, that would be a super fun graph to see which one Tesla stock after the Trump election yes, I'm glad actually.

Goran Cvetanovski: 1:42:28

Anyway, okay, let's go. It didn't spike that much.

Anders Arpteg: 1:42:31

Go back like three months or something.

Goran Cvetanovski: 1:42:33

Yeah, we'll do a year, so let's do that.

Henrik Göthberg: 1:42:36

That's a big spike.

Goran Cvetanovski: 1:42:37

That's a big spike, it's not a big spike, man, it's 82%. That is not a big spike, it just looks big.

Anders Arpteg: 1:42:44

I think it is In normal. Normal terms. 82% is good. I made a lot of money from it, so I'm happy yeah, you should buy a Tesla car.

Goran Cvetanovski: 1:42:53

No, that's throwing money into whatever okay, shut up, we both drive a Tesla what about your bike?

Henrik Göthberg: 1:43:02

how many times it got stolen? Fuck off, oh yeah, okay, but to go to open source.

Anders Arpteg: 1:43:05

You know my thinking.

Henrik Göthberg: 1:43:06

How many times has that got stolen? Fuck off, okay, but going to open source.

Anders Arpteg: 1:43:10

my thinking there and please disagree if you want but I think we're going to move to the frontier models being less open source, but then we're going to have a spectrum of models that are more specialized, that are getting open source, but for the frontier models, I think even Meta that today is launching L getting open source. But for the frontier models, I think even Meta that today is launching Lama open source, not in Europe, but at least in the other parts of the world. But I think they will at some point soon stop releasing open source as well. Why? What is?

Henrik Göthberg: 1:43:43

the driving force for this prediction.

Anders Arpteg: 1:43:46

Because of the abuse of AI.

Henrik Göthberg: 1:43:48

It's too scary, it's too powerful Because of Jeffrey Hinton, it's too powerful.

Anders Arpteg: 1:43:55

It's too easy to abuse AI if you start to release this model completely open source, and OpenAI, of course, stopped that for a long while back. Google didn't and will not do it, and the only one really left is Meta that have these kind of frontier models.

Henrik Göthberg: 1:44:12

Jan Lekun is really pushing this story still.

Anders Arpteg: 1:44:14

He's still saying that Elon is also trying to. They released a lot of open source as well. I think Elon will not either.

Henrik Göthberg: 1:44:23

But Jan Lekun is still philosophically inclined to open source, like we had Eric Hugo, who has a very ideological view on this.

Anders Arpteg: 1:44:32

Yeah, so I may be wrong, but I think in 2025, we will see a model that is released from a meta that is not open source.

Henrik Göthberg: 1:44:39

Okay, I agree. Let me put another angle on this hypothesis, and that is for the normal companies who really want to benefit from AI practically in healthcare or whatever. I'm not sure we always need to look at the frontier model. We are much more better off to cost effectively look for open source approaches or different things.

Anders Arpteg: 1:45:05

I think you can use the top frontier models to do model distillation to your model you have. So, you use them as a teacher and that's how you're going to use a lot of them, because it's not practical to use this super huge model.

Henrik Göthberg: 1:45:20

So break that down slowly what you said now, because I think that is exactly what we're going to go.

Anders Arpteg: 1:45:26

So we're going to have frontier models to teach the baby models yes, Because it's super expensive, you know, to run this kind of, let's say, one, two, three, five, ten trillion parameters.

Henrik Göthberg: 1:45:38

It's not even funny to argue with you.

Anders Arpteg: 1:45:40

We think the same, but it's super expensive and you don't really need it for more specialized problems.

Henrik Göthberg: 1:45:46

Yeah, but the, the profound thinking which is so obvious, like, like when you say, oh, we need to combine the thinking and the reasoning, the data, you know. The thinking is it like this is too fucking expensive. And then we have the normal bipolar argument oh, we should stop it Fucking. Meet in the middle, optimize. What do you use it for? In the outlier case of training the baby models, it's obvious. But the media and the way that we communicate is always bipolar. The same goes with the argument where we should combine the two Bipolar. We are fighting these stances instead of combining them, and I think it's so when you think about that.

Anders Arpteg: 1:46:30

Of course the frontier models have a place, but of course they are too expensive to run in order to run your fucking small appliance at home and if you have a million users or even a thousand users sometimes, and they have to do a lot of reasoning or a large number of, you know, inference steps, it will be too expensive to run it on the big one yeah, so that.

Henrik Göthberg: 1:46:55

So then we get to a new one's conversation, and I want to bring it a little bit back to the apartheid idea, ideology we talked about with eric hugo, who actually point out in a very nice way when we talk about the AI divide, it's easy to think about the AI divide from the tech giants to Europe. Then, of course, there's another spectrum, from Europe to Africa, or the farmer who is forced to have 5G in order to work with ICA in Sweden. It's technology apartheid and in that sense, open source is so detrimental and so important in the way Africa or you know, we can get a more balanced distribution of intellectual power in relation to AI. So then it makes sense, right, we're going to have that, but it's not going to be the way the normal people, the normal businesses, scale.

Henrik Göthberg: 1:47:49

This is I don't. So there are going to be huge opportunities for the companies, fine-tuning models on certain applications or industries, and they're probably going to do it. You know, training them, and maybe here is going to you know how we're going to use supercomputers or hpc and all that, and then we're going to package and we're gonna do something else. I, I can really see this. I don't think it's that hard to put those ideas together yeah, I'm not.

Anders Arpteg: 1:48:18

I don't think hpc is a solution if we're trying to say that that's a good thing, but um, no, another deeper conversation.

Henrik Göthberg: 1:48:26

We'll take that afterwards.

Anders Arpteg: 1:48:28

But still that we need to have specialized models.

Anders Arpteg: 1:48:36

And I think what OpenAI just released with the new reinforcement learning, fine-tuning approach where you can train the mini-model, which is like 10 times less, and you can make the performance of that be superior to the big one by having 12 example data points, Nothing. It's cheaper to just training that. You can put it through the API. They train it for you. It's like AI as a service, if you call it that, and then you get the model back. It's very cheap to run and it's better and faster.

Henrik Göthberg: 1:49:08

This is obvious. It's so obvious. We will figure this out. Is that the end note of sort of summarizing the year, or did we miss something major? Otherwise, I'm quite happy to find.

Anders Arpteg: 1:49:20

Perhaps just one point. I mean, 2024 has been a super election year as well.

Henrik Göthberg: 1:49:24

Yeah, Was it a fake? We talked about, we were predicting. You know this was going to be the AI election year, Was it or not?

Anders Arpteg: 1:49:38

I think it was a lot of AI being used, generative AI being used, but not too bad sense.

Henrik Göthberg: 1:49:45

Not as we were fearing. We were fearing more.

Anders Arpteg: 1:49:48

I think we were having bigger concerns with potential manipulation and disinformation being happening in a conscious way. But in reality, of course, elon's use of putting up a video of Kamala Harris saying that she knows nothing about how to run the country but it doesn't matter when you're just a puppet or something. I mean everyone understood it's a parody, everyone understood it's not her saying it. And a lot of people used generative ai, more like a parody kind of thing, more as a way of communicating.

Henrik Göthberg: 1:50:29

We used it as a traditional political satire approach, as a way to make our points, communicate our points.

Anders Arpteg: 1:50:39

I think it's an example of how people and society in general are surprisingly adaptive and society in general are surprisingly adaptive. It was actually the same in the election in Argentina back in October last year, where they have a huge use of generative AI. But everyone understood that this is fake, you know, and it's just for fun, and they had the two people, massey and Millais, who were the two final candidates. Millais won later, but both of them were putting the competitor, so to speak, with a zombie or something. No one really believed it was true. They didn't really try to fake it.

Henrik Göthberg: 1:51:23

It's satire, it's political satire, and political satire works in this context, but it's not the same as deep fakes or fake news.

Anders Arpteg: 1:51:31

So, in general, the election. Of course AI was used a lot, but not to the extent in a bad sense that people feared, which is good.

Henrik Göthberg: 1:51:39

But now we can talk about bad versus tasteless. Did you see when they were trying to prove how grok, how un-sensual it was? And you saw this they had videos of pictures of Donald Trump together with Kamala Harris and how their baby would look like, and it was all over. It was all over. It was spinning for a couple of weeks and it was all over. It was spinning for a couple of weeks and it was funny. And you know it's it's not even politics, it's used hilarious and it was like, and it was more like you know what you can do, whatever the crazy thing you want with grok. That was, of course, elon free speech lobbying yeah, what happens?

Henrik Göthberg: 1:52:23

yeah, okay yeah, that was. That was another geopolitical yeah, there's other. Yeah, no, I think we stopped there. Yeah, I think we should go. We haven't. Let's, let's bring in some kind of future outlook. How do? How do you want to frame that? My, okay, I I try to frame it, but you can flip it. I was a little bit thinking about okay, where are we going in 2025? And, not to be sort of too bold and go too crazy, I was kind of looking at what are the key themes and guests and topics that are this if we want to be on point, we should be talking about this stuff next year. So what should we have on the pod to be on point the next six to 12 months?

Anders Arpteg: 1:53:17

I like actually the opening eyes levels of AGI. I think it's actually more. I think it's rather educational and pedagogical in a sense. You know it starts level one, I believe, is conversational ai chatbots that we had have today. Next step is reasoning. That's basically what we're working with right now. Third level is autonomous, meaning it can take action, becoming a bit more agentic. But autonomous is is the way they use, which I think is good, and then I think they go into level four in innovation or invention, but innovation. And then level five is organization. So I guess in coming year we will see a lot of for one, reasoning, of course, but also starting to come more with autonomous and the agentic approach. That's obviously, I guess, where we will see a lot of progress going forward. But also then, in terms of engineering, I think, at least for the tech giants, they know and they do focus a lot on engineering aspects of this and it's not scalable to have these huge frontier models being used for real.

Henrik Göthberg: 1:54:36

So that's another part that I think we will see a lot of work going forward. If I pick up the thread on the outlook for next year, we spent a lot of time trying to wrap our head around the word agentic because it became such a trending word and it was all over the place and we understood. So what we are working on there is a lot of times not only about the techniques, but how do you organize the teams and the people and operate and the governance, and we realized that we we wanted to try to reclaim the word agent and again, tic is, of course, very old in both of software engineering, building agent-based systems, and then you can talk about this terminology in business and organization. What does principal agent problem mean in a completely different setting? So we were trying to basically do a leading with a hype, starting with the agentic workflows and talking about, you know, agentic workflows in the technology, and then move that into what does it mean to have agency between people and you know how you organize autonomy of teams versus autonomy of systems. And this leads us down the path and now I'm connecting back to you that another way of talking about these levels, if I'm trying to communicate that to someone who's not in AI, where you can kind of see it.

Henrik Göthberg: 1:56:07

We saw it in 2024, was the year of the Gen AI and the year of we are starting to talk about. We are prompting for solutions and if you go the next level of autonomy in this, we're going to start prompting on the problem level, where the system then is more autonomous to sort out the steps, the sequencing and the reasoning in order to solve the tasks. So we are going on a trajectory where we're going from, you know, getting support with our workflows, then we're getting, then we can prompt for solutions and getting automation around solutions. We are prompting and we're going to now prompt in more and more abstract levels and therefore giving agency and autonomy to our systems to do things. And this has to do with the elevation of reasoning and all those skills. And if you connect it back to old school approaches we talked about RPA a couple of years back, robotic process automation. What was that all about? We were trying to build autonomous or automating very dumb processes, low knowledge intensive processes, and we are now trying to really do autonomy and automation augmentation on rpa.

Henrik Göthberg: 1:57:29

Could finally have some ai in it so rpa could finally have some ai in it and we could maybe move into more knowledge intensive processes and and and. All this is on this trajectory and the technology is unlocking it, the path of what you're talking about, the latent space. The reasoning is on this trajectory of we are going, we are giving more and more increased agency to the systems and therefore more and more autonomy on more complex tasks. That is a given. So this is for sure a topic that we will have.

Anders Arpteg: 1:58:08

Yeah, I wouldn't phrase it as strongly, but yeah, I'm sure I agree it's going to move in that direction.

Henrik Göthberg: 1:58:14

Yeah, but then now I come to the next pet peeve that I think we need to add to that which we are not talking about enough, and that is the fundamentals of we need to stop talking about the model and we need to start talking more about the engineering around AI compound systems. And I can refer back to when we had the podcast with Cassie Kusorkov and we talked about this on stage and she did a couple of things. She, she did. She did a reflection of when you build an ai system that really work enterprise grade the model is just one brick in the brick wall. And what she was really doing she was giving a picture of the paper from google technical depth of machine learning from what is that?

Henrik Göthberg: 1:59:04

that 2015. 2015, right, which basically says you know what? In order to get this into engineering production? This is just a very small part and every single guess we've had. If we talk with fever on how we build machine learning in algorithms or anything else, everybody who knows anything about anything downplays the model and upplays the engineering, and so there's a huge gap in conversation on compound AI systems. This is what I want to see more of this pod that we can demystify, not only AI, but what it takes to go to production in AI.

Anders Arpteg: 1:59:43

Find true value. Find true value.

Henrik Göthberg: 1:59:46

Because that would be awesome, right? Because then we need, you know, we have the LALIS of the world. We have many different angles here, but I think if we can help bring that picture of engineering around the model, that would be awesome.

Anders Arpteg: 1:59:59

I think. Another thing, and this is harder to be sure of, but I think we will start to see the demise of open AI coming here. You think so? Yes, why? I think it's falling apart a bit internally.

Henrik Göthberg: 2:00:17

You think so?

Anders Arpteg: 2:00:18

With all the people that have left already this year, I think, when Sam Altman and his ideas going forward. But on the other hand, he got in a deal with Apple that in the 12 days of releases that they just did, they also announced the Apple intelligence using ChatGPT. So it's not like Apple will be dead using chatty PT. So I mean it's not like PayPal will be dead, but I think there will be an increase in concerns with how it's being driven.

Henrik Göthberg: 2:00:55

Okay, so let me be very provocative. Why do we need to care about that? Do we need to care about the demise of OpenAI, or is that just a?

Anders Arpteg: 2:01:12

natural cycle of Silicon Valley, or is it something bigger here?

Henrik Göthberg: 2:01:14

at stake. They've done something amazing.

Anders Arpteg: 2:01:16

Yeah, I agree, Given that. So they've been a big part of driving AGI, I think to a much faster pace than we otherwise would have done unless they've actually been there.

Henrik Göthberg: 2:01:28

so they've done a lot of great work in in that direction they, they sped up, they sped up the whole trajectory of this for sure and then the question is how much will the safety of ai concerns play in?

Anders Arpteg: 2:01:48

because that's where they're lacking more and more. I think you know, yeah, and the way you do it, and and if, if the ai race between the tech giants just keep accelerating, then it won't matter, because then it's simply, you know, the one that makes the most money that will win. But but I do think actually it will matter. I do think that they are racing ahead without thinking, potentially, about the humankind. The mission of OpenAI, what is it? It's something like they want to find AGI for the good of humanity or something.

Anders Arpteg: 2:02:25

I'm paraphrasing here a bit, but it's something about the good of humanity or something I'm paraphrasing here a bit, but it's, it's something about the good of humanity, that that should be part of it good for humans and it seems that you're not really emphasizing that part that much. So I, in that sense, and given also what xai is going to do now in coming year, I think that will be more than people think.

Henrik Göthberg: 2:02:51

Should we put more attention on the frontier models and what's happening in Silicon Valley? Or, if I flip it, how big impact in what we should be talking about is about this and how much, if in Europe, should be about implementation of the AI Act? I really want to flip it to smooth and safe AI innovation, to adoption. If we can flip the script on this as a control function of something that is stupid to actually helping us engineer better systems, then I'm all for it. Actually.

Anders Arpteg: 2:03:25

Yeah, and I wish Europe and Sweden would invest in exactly that.

Henrik Göthberg: 2:03:29

Yeah, but that's not what we're seeing, unfortunately well, how can no good, if that's what we want to see, who should we have as guests? What should we talk about? To take a compliance topic into smooth and safe Innovation topic at this port, how can?

Henrik Göthberg: 2:03:44

we help, that we need engineering, okay, even better, smooth, oh, you know it. No innovation in terms of the commercial value, value, the value, the engineering, smooth, effective innovation, adoption, and you know, put engineering in there for sure I agree, because I agree very much with what you're usually speaking about the invention, innovation and then diffusion and crossing the chasm in different ways, but when you look at what the HPC clusters that we're investing in and what the AI fabric in Sweden is going to work with is, they're using the term innovation.

Anders Arpteg: 2:04:24

I think using it in a wrong way is invention, because this is invention. It's just, you know, you're just allowed to use it to do some kind of prototype, and that's really bad.

Henrik Göthberg: 2:04:35

But then okay when I flip it if we have the opportunity to influence the Pietras Alundes and the Taurus of the world and flip them into not only thinking about the factory as a technique topic, if they were to help you with the important engineering aspects, about being compliant, about finding the affordable kind of sustainable models to run and doing fine-tuning from their frontier models that actually helps companies.

Anders Arpteg: 2:04:59

That would be amazing.

Henrik Göthberg: 2:05:00

So if you flip the AI factory, if you flip the TEF approach into engineering innovation, to adoption type topics, proper innovation, proper innovation, engineering, product development and I put it to you, it's as simple as understanding properly a software development lifecycle and ultimately understanding what questions should be dealt with in each part of the life cycle, including safety issues, including engineering issues, then it would be amazing. It could be. It could be amazing. I'm going to cut this out and send this to Petra, yeah, but I would really really Okay. So let's take another future outlook and this is very Swedish, yeah, but I would really really Okay. So let's take another future outlook, and this is very Swedish. Yeah, where is the Swedish ecosystem going in 2025? And let me give you a little bit of a backdrop. We had the AI Commission and we had a lot of interesting people working on the AI commission and my personal take was too few engineers on that, but that's another thing. Then you have had now, for one or two years or many years, we had Vinnova working into AI Sweden and that set up.

Henrik Göthberg: 2:06:22

I would argue that in this podcast, we actually made a fairly large push. It wasn't intended, but we had several guests from Rice this 2024, simply because they were doing fucking cool stuff. They're doing good stuff, man, and one of the topics where we talked about oh, you should have been here with Cepeda is why is rice so important in the AI ecosystem? And she said it beautifully and she exemplified it with the Olympics engagement. So when they put the team together to maximize the opportunity and win a gold or silver medal, there was the AI piece, then it was the manufacturing of composite materials piece, so it was the whole applied AI, where you need a real, real muscle of institution from different disciplines, because AI is cross-sectional across all this, like Sverker Jansson tries to work across all of this, which makes it really, really hard if you're very narrow and only focus on one thing or you don't have the body behind it. So I got one or two aha moments on where the Swedish ecosystem needs to go and I think rice should play a bigger role in my opinion.

Anders Arpteg: 2:07:49

Sounds good to me.

Goran Cvetanovski: 2:07:51

Well, I have to cut it here. We have been on now two hours ten minutes. I think it's enough, and plus, we have a dinner table waiting for us, so I think it's time to say goodbye From my side. I think that we are expecting a very good 2025. The world is going to as we know it. Don't end like that. Come on off. I had to be a little bit like uh, like that, but uh, my point is actually the different. Let's just uh like that, but uh, my point is actually the different. Let's just uh, do good with the eye. That is what is most important. We are living in a troublesome times and we need innovation and we need we need innovation, but it needs to be for good yes, exactly.

Goran Cvetanovski: 2:08:36

So let's do good, and that would be my sentence for uh this year. Thank you, everybody.

Henrik Göthberg: 2:08:43

No, no, no, no. I need to ask the producer where do you want to take the show next year?

Goran Cvetanovski: 2:08:47

What is your top top top one, two, three, so we are looking forward to coming back on 16th of January.

Henrik Göthberg: 2:08:55

That's when we kick off. Who are we going to have on the first?

Goran Cvetanovski: 2:08:58

episode, the first. We're going to have Pietro here from Silo AI. We're going to talk about deep learning. So we're going to go a little bit deeper in some of the topics because in my aspect, I think that we're going to go a little bit too narrow ai rather than into large language models and practical use practical use cases a little bit more towards agentic ai and and how I understand, agentic AI is just basically a very specialized AI model that will help you out with specific things.

Goran Cvetanovski: 2:09:34

So I think that that is actually going to be a big trend next year, because most of the organizations are still struggling and have challenges of getting a value of AI. That has been a topic of all the conferences that we did this year across the globe. I think that we will see quite a lot of geopolitical I would say investments in AI.

Anders Arpteg: 2:10:01

It's going to be a Trump year as well. That's going to be really interesting.

Goran Cvetanovski: 2:10:05

But you can also see that uh, if, if, if we ask all the governments why they're putting all of these uh investments in ai so much. If you're talking about usa, we are talking about silicon valley usually, but we are not talking about the country per se, right? Um, although they have done quite a lot of big investments and they're still investing quite a lot. But we can see that Europe is actually moving. We can see China, of course, moving a long time ago, and I can see Australia. I can see the Middle East investing quite a lot heavily in this. The question is just basically why, and for me, the why is very simple because AI is here to stay. Ai is going to be infused in most of the software that we're using today. People will be using AI to optimize their work, to be more effective. I think that we need to drop this. That AI is equal to taking people's job rather than it's going to create jobs. We need to think about that.

Henrik Göthberg: 2:11:03

It will augment it will be another tool.

Goran Cvetanovski: 2:11:05

It's just another tool. I mean mean, nobody lost job because internet came and the computers came. He just changed job from basically going in a factory to going and working on a computer. So it's going to be the same could be a tough transition period it will be a tough transition, but no transition is easy.

Goran Cvetanovski: 2:11:22

Yeah, that is what it is okay and keep in mind like that, this is only gonna. Right now it's affecting the, the white collar people and etc. People that are still working with construction, etc. They will not be utilizing ai for many years to come. Maybe our government, the reality when it comes, uh, it's going to be much better because then they can see everything um before they um maintain stuff. But I think it's very important to see the macro perspectives of this that AI is here to stay. Companies are investing. These people are investing. People are already using AI, right Since chat GPT came. Everybody use it in universities, in schools, in everyday jobs and et cetera.

Anders Arpteg: 2:12:05

And AI has been used for 30 years.

Goran Cvetanovski: 2:12:07

Yes, exactly so it's not something that is actually right now, it is not new anymore.

Henrik Göthberg: 2:12:13

That's the thinking that is wrong. It's just more powerful.

Goran Cvetanovski: 2:12:17

Exactly it's not new, it's more powerful. So, keep in mind, we were talking about this four and a half years ago, yeah Right, and I can see that people are fed up with ai, but they're not fed up with the transition and the value that they get from it, and I would like to dedicate the next season and year in exploring actually the bound, the frontiers of how we can implement this technology in a different set, because we need to have innovation and application on the same top. You remember, in the 90s, there was this TV show which was like you know how the future will look like. I don't know what was the name right now. It was like a documentary series when the future frontier was called right, and they had all of this. In future, you will see satellites. You will have, like people using mobile phones and everything else.

Goran Cvetanovski: 2:13:06

We need to start believing us, continue believing in such things. We need to explore the ai field and how we're advancing in this in a more deep well way. But also, I think it's very important for us to stop for a moment. Do not think about ai. Replace the ai term with something else which is going to be the value for the person that is actually using it. If that is an artist, how you can actually create art in a different way. If he's a doctor, how you can actually utilize this tool to be better in what you're doing. So it's quite a lot of things. It's enough talking about just AI in a platform way or a the technique.

Henrik Göthberg: 2:13:45

The technique Going from the invention to the real innovation and diffusion.

Goran Cvetanovski: 2:13:49

Exactly so. I would like to see a mixture between innovation and application, and that is where we are going. So I'm looking forward to meet all of our listeners once again in 2025. It's going to be a great year, of course, a lot of changes and etc. But you will find us here every Thursdayursday five o'clock having a beer and being open to you guys, and the chat is always open.

Henrik Göthberg: 2:14:12

so communicate, speak with us and uh yeah, you promised more nordic than swedish guests. Yes, can you, can, you can, we can actually even do better than that, uh.

Goran Cvetanovski: 2:14:25

So we will try to actually do quite a bigger thing. So, from me, thank you for this year. It's been fantastic to have you as listeners and viewers. It's been fantastic to have all of these podcast guests here and to have you here as well. You two guys. You have done a marvelous job for four and a half years and I'm looking forward to continue and see where this is going to bring us to. That is what it is.

Anders Arpteg: 2:14:49

Thank you so much, goran, thank you Everyone else.

Goran Cvetanovski: 2:14:52

Cheers, cheers. Happy Christmas, happy holidays Everybody, and a happy new year. That is what it is. Cheers to everybody.