AIAW Podcast

The Artificial Intelligence After Work (AIAW) podcast is a weekly live streamed long format conversation aiming to demystify data innovation and AI, as well as their impact to future business and society by bringing the listeners close to the challenges that AI practitioners aim to solve today. The case-study, industry-by-industry, human-focused, and guest personal angle on the topic approach makes the podcast educational, emotional, engaging, and entertaining to all who are interested in learning more about AI, the future developments in the area, or simply getting exposed to variety of topics from practitioners and experts with first-hand industry experience and knowledge in the topic of the day. Hosts: Anders Arpteg & Henrik Göthberg. Program Manager: Goran Cvetanovski

All Episodes

AIAW Podcast

AIAW Podcast E174 - AGI in 2026 - Karim Nouira

January 29, 2026 • Hyperight • Season 12 • Episode 1

0:00 | 2:05:50

Season 12 of the AIAW Podcast kicks off with a high-stakes question: Can we reach Artificial General Intelligence (AGI) by 2026? In Episode 174, we’re joined by Karim Nouira, founder of sics.ai (Superintelligence Computing Systems), for a deep and provocative conversation on the technical frontiers of AGI. From LLM limitations and JEPA’s alternative path to robotics brains and latent space reasoning, we unpack what it would take to build truly autonomous systems. We also explore Sweden’s role in the global AGI race, the future of labor in an agentic world, and the societal implications of superintelligent machines. A powerful start to a season focused on what’s next in AI. Tune in.

Math, Fermat, And AI Proof-Checking

Anders Arpteg 0:00

I mean cool. So how how did you uh did you just prompt it or or what did you do?

SPEAKER_02 0:05

Yeah, I I uploaded it uh to Claude uh and uh I instructed Claude to check uh uh some part of the proofs and the theorem uh and went through the the proof stepwise uh and and got some very interesting feedback uh very early. Uh and this helped me to perfect the the part of the proof which is uh which is ready. Uh unfortunately um uh the the uh the the main theorem of the proof was flawed, so I still have some work to do.

Anders Arpteg 0:38

So I I don't recall the theorem exactly, but it's something about uh Pythagoras' uh theorem, right? Or can you explain the theorem's last theorem?

SPEAKER_02 0:47

Yeah, it's very easy to state the the problem actually. It's uh uh Pythagoras theorem a squared plus b squared equals c squared, but then to change the exponent to instead of two, we have three, four, five, whatever. So and the theorem, um the the margin note of um Pierre de Fermat was uh I have a wonderful proof that uh uh the case is that there are no um solutions in whole numbers uh when uh the exponent is three and above, and he said that the margin is too uh too small to allow me to write the proof.

Anders Arpteg 1:23

I wrote it in a margin or something, right? Yeah, yeah.

SPEAKER_02 1:25

He had a book about uh Greek mathematician Deophant Diophantus, uh, which is about whole number theory and whole number uh equations. And on the the chapter about the cubic equation, uh this note was uh he left this note in this book. And actually, when Fermat died, he had like 30 different types of problems that he stated that he had proven, uh also with this type of probably Morgan uh and this was this and and then the mathematicians they have struggled for like 350 years to prove this, and then there was one left, it is called Fermat's theorem. And it was proved by uh Andrew Wiles in 1994, and the proof of Andrew Wiles is like 800 pages. Uh and at the time there was maybe 10 mathematicians in the world who can understand what it actually done. And in order to prove this theorem, he proved something else called Tanyama Shimura's uh conjecture. And uh Fermat's theorem ended up as a corollary from this result. So uh since then, about 1995, I read about and read an article about this this problem, and I said, well, probably there might be a much easier solution to this problem. And I started to work with it. Uh, and I found some non-trivial results that uh resulting in that I was accepted as a PhD student in mathematics based on my results. If my results were unknown, I would probably be um doctor in mathematics. But unfortunately, uh they ended up as I found them in uh Paolo Ribens Ribbon's book, 13 lectures on From Asless Theorem. So, and I've been working uh with this problem ever since. Uh, and I have some had a flash of inspiration this uh this Christmas. So, uh during the melanagon och whatever that's in English Christmas days. Uh I found a probable a way of using number theory in different way to to simplify the problem. Uh again finding a mathematical structure that matches your problem, if you do that, you you have basically solved the problem. Because you can use all the results about this mathematical structure to find your result. And this was actually what Albert Einstein did when he had the he struggled with the um merging special relativity or relativity theory with uh with uh gravitation. And uh he was lucky because there was a mathematical structure already from Riemann of non-Euclidean geometry off the shelf accessible with the uh curved space and everything. So he just took his equation and uh within this mathematical framework, and he could he can show that uh gravitation actually bends space-time.

Anders Arpteg 4:25

So uh some existing theorem did exist, yeah.

SPEAKER_02 4:28

So it was actually it was plug and play for him.

Henrik Göthberg 4:31

So and now what made what amazed you? So now you basically tested your theorem or your angles and ideas in Claude, and specifically now what amazed you? Not you not just interesting, but this was cool.

SPEAKER_02 4:44

That it it it it uh appeared to know mathematics, but it it actually doesn't. It's it's a parroting of lots of lots of humans' uh uh articles, uh but it does it in an amazing way. So it was actually very useful. So instead of asking uh one of my mathematicians' friends to check the proof, I could I could Claude did the job and it did excellent work.

Anders Arpteg 5:06

Did you put the previous proof or did it more you know work with the the theories that you had?

SPEAKER_02 5:12

No, no, I have a much much simpler uh um uh approach. So basically elementary numbers, sorry, elementary algebra, uh some group theory, uh and that's it. So uh the the the proof which uh it turned out not to work was is about 10 pages. Uh and uh I think there might be a way to fix it, but uh I I I I I have to wait next Christmas.

Henrik Göthberg 5:38

So the shortcut here is like inspired by the um Einstein um anecdote is to find the uh uh the systemic architecture of math and all of a sudden you're applying that and and then actually you can prove two things fitting together.

Wiles’ Proof And A Simpler Path

SPEAKER_02 5:52

Exactly. And it's the same approach when we uh have crafted this artificial general intelligence architecture. So if you look at today's AI and the mathematics behind it, uh it kills the structuring information. If you do for for your analysis or uh uh synth synth programming, uh if I say um hello, uh how are you, I get a sound uh file, I can use uh 10,000 voices with uh different settings, volume, frequency, um uh phase, and I can have this system say, Hello, how are you? And if you look at the the different parameters, they are sine waves with virtually no information in them. So this is an analogy to how today's AI works. It's a lot, lot of parameters, it's a big black box. If you just have a big enough parameter space, you can fit any uh uh data in data to uh uh the the model can replicate base virtually any in data. But the structured information is lost. You have an implicit mathematical model. So ideally, what you would like to have is a different type of mathematical model, which a rich with a richer mathematical representation that's more structure-preserving than today's AI. And this is what we have in our model, and this is the reason why we believe that we will have AGI in 12 months.

Anders Arpteg 7:16

Wow, what is what a setup, huh? Yeah, I'm so tempted to go into the rabbit hole.

Henrik Göthberg 7:21

This is this is just uh let's let's introduce the guests and set up the theme.

Anders Arpteg 7:25

I think it's the perfect setup. Yeah, awesome. Uh so very welcome here, Karen. Thank you very much. No Ira.

SPEAKER_02 7:32

Is that the first one? Yeah, yeah, very good pronunciation. Thank you very much. Uh, I really appreciate to be here. I look forward to this since uh we spoke with last uh last year.

Anders Arpteg 7:40

So but you're a founder and CEO of um Super Intelligence Computing System, is that right? Yeah, yeah, yeah. Super cool, and uh congrats on uh on the progress you have seen there. I'm really looking forward to hearing more.

SPEAKER_02 7:52

Yeah, thank you very much that you're working with there.

Anders Arpteg 7:54

But you also you mentioned that you started to do a PhD as well in mathematics, so we already heard that you you have some interest there and a lot of knowledge there. So I hope we can really have a fun and interesting discussion here going deep into both AI and math in some way informed.

Henrik Göthberg 8:11

Yeah, and uh we we always try to find a theme for the podcast. And the uh the the simple the simple theme is the future of AGI. So it's gonna be one of those conversations which can be philosophical, but I would love to make it mathematical, not because I can follow, but I would like to hear that angle on it. And and the you know the architectural, the Jeppa paper and all this. So we have so much to paint on the canvas of the future of AGI.

Anders Arpteg 8:39

So yeah, let's not be too afraid about going a bit technical as well for mathematical sometimes.

Henrik Göthberg 8:44

No, it's it's I think it's about him protecting his IP. Yeah, right.

SPEAKER_02 8:49

We have to drag it out. We will drag it out. Oh, at least we'll see what happens if I happen to see something.

Henrik Göthberg 8:54

We've had people who are uh you know representing uh Microsoft with insights on on OpenAI's uh roadmap, and we were we've been trying to drag it out, and we've been trying to yeah. No one mentioned, no one forgotten, you know what I'm talking about.

Anders Arpteg 9:08

But I think you know when Anton was here before Lovable launching, he'd mentioned a lot there, so I think we dragged this out.

Henrik Göthberg 9:13

We got we we had Anton Osic uh here the month before they did the Lovable proper launch. Like he was a project before with a different name. Yeah, and we had him here and like and then it just took off. So cool. So, you know, so we were you know, have some more beer.

Anders Arpteg 9:30

No, but Kevin, uh drink water, thanks. Perhaps you can just very, very briefly describe a bit who you are, uh your background, and how you ended up at uh at the super intelligence computing system.

SPEAKER_02 9:41

Yeah, sure. So uh I'm a jazz musician. I play together with yeah, I play together with uh James Bradley Jr., who is one of the uh the best drummers in the world. I used to tour with Anita Baker and Lenny Kravitz and uh oh wait there's a side note here.

Henrik Göthberg 9:58

We have a non-statistical uh survey going on how many musicians there are that in AI. Yeah and it's actually it it is almost statistically valid by this point.

Anders Arpteg 10:08

Yeah, I'm the exception, I think.

Henrik Göthberg 10:10

Yeah, you're the exception because too not too late.

Anders Arpteg 10:12

No, because I don't have no talent in music.

Henrik Göthberg 10:15

So no, but it's like uh it's surprising when you have a conversation like that in a couple of beers afterwards and we start talking about life, and music comes up all the time.

SPEAKER_02 10:23

Yeah, yeah, music is a vital part of life, uh, and it's probably connected with uh with life in uh a way that we in a metaphysical way that we probably don't really understand.

Henrik Göthberg 10:34

Yeah, and what do you play then?

SPEAKER_02 10:36

Which instruments are your uh I play the f Fender Rolls uh keys uh piano.

Henrik Göthberg 10:42

Oh we had the keyboard here before I should have had it.

Anders Arpteg 10:47

Ah that's shit.

Henrik Göthberg 10:48

You need you need a property.

Anders Arpteg 10:50

Okay, but uh musical background then and are you still active somehow? Yeah, yeah.

SPEAKER_02 10:55

No, the company takes a lot of time, but uh after my last exit I played a lot in clubs, you uh uh Strandwagonet, uh Liedmar, Nefertiti in Gothenburg.

Anders Arpteg 11:07

Jesus, impressive. I'm getting very interested here. Damn it. Okay, but uh please explain, you you you started your career, I guess, in in uh in the studies as well, in math, right? Or how did you get started with a math interest?

SPEAKER_02 11:19

Yeah, I I studied at the Royal Institute of Technology, uh applied physics. Um I like physics a lot. Uh I planned to do research in physics, but uh I ended up uh doing research in mathematics. Um I still very interested in physics. Um and also artificial general intelligence is also a topic of interest. So um, and um when I studied uh lots of my friends they uh studied both at the Royal Institute of Technology and uh Handels School, whatever you say in Swedish School of Economics, yeah. Uh and I I thought about maybe doing that as well. Uh but instead of doing that, I started uh companies.

Anders Arpteg 12:09

So uh what was the first company you started then?

SPEAKER_02 12:11

The first company was a crappy one. We made uh um we had the enquête where market research. We m m we made market research with uh I don't know the English word for that as well. Talsvar.

Henrik Göthberg 12:25

Talsvar and phone, the phone when you yeah, the phone. You could it was you could you could use the buttons on the old phone and then is it a one, two, two, five, press the five now. Yeah, exactly.

SPEAKER_02 12:35

So that was quite new at that time. So we we uh uh we ended up with some projects, uh and then in 1998. Uh, uh, me and some um uh friends from Royal Institute. Uh we started uh uh a company more known at under its product named J Rocket. So this was a virtual machine for Java, which solved a big problem at the time because uh everyone wanted to use Java. Java was good for like small uh systems with small footprint, but there was a problem using Java uh on the server side. So uh because the Java virtual machine accessible at that time, uh, which was like some microsystems hotspot, right? Uh, didn't work properly. So uh we made an adaptive uh self-adaptive optimizing uh virtual machine, so sort of uh rule-based AI, we could call it today, if you like. Uh and uh everything's AI. Yeah, yeah, but at that time everything was IT, so it was this IT boom, so we had to call our company the IT company. Anyway, so so this was uh very successful, and we sold the company to BI Bia Systems in uh Silicon Valley, and BA Systems was that was then acquired by uh Oracle, and this product is still uh running on 20 million servers worldwide, and it's a de facto standard for Lidus. Uh and moving to today, uh uh we reassembled uh part of this team. So, me, you Joachim, Piet and Marcus from from the Rocket are in this new company. So I joke and said we we like the the the movie The Expendables, you know?

Structure In Math And Einstein’s Playbook

Henrik Göthberg 14:16

With uh That would be the coolest Alone and the Dolph and uh I needed to fix your web page but and and I first time I heard I saw we met at the one of the conferences, yeah, HPC conferences, right? Yeah, high performance computing conference and immediately.

SPEAKER_02 14:40

So I thought oh this must be a homage to the original Yeah, yeah, exactly. So that's so one part of the company is this uh expendables. The other part of the company is uh uh there are three uh well, at least three people from uh originating from six in the company. Uh so six uh Swedish Institute of Computer Science doesn't exist anymore. It's a part of Rice, so the name is not used anymore. And we had this as uh since we had this Niklas who is in the board of directors, uh Niklas Rudemou, he's from Rice, he was like uh deputy head of Rice, something like that, deputy director, and then we have um Don Yill Blonde, uh who is the head of AI, uh was the head of AI at uh at six.

Anders Arpteg 15:24

Is he part of the new company as well?

SPEAKER_02 15:25

Yeah, yeah, he's he's part of the company. Uh and then we had also uh we have also um uh Yussi Colgian, who's uh who is uh very skilled in uh in computer uh linguistics. So so he has helping. Yeah, so yeah, I'm sure you know him then.

Henrik Göthberg 15:42

Daniel has been here. Yeah, has has Yussi been here yet? We've been both Yussi and Dan. Dan, you you you're late to the party.

SPEAKER_02 15:52

I I'm the new kid on the blog. They use and Don, they've been around for like 25 years.

Henrik Göthberg 15:56

So and I've been doing the AI profiling, yeah, yeah. So super intelligence.

SPEAKER_02 16:00

Yeah, so so the name uh we we took the name as a working name for the company when we started it as a project, and then we got stuck with it. So we we registered we registered six.ai today. It's very difficult to find an A.ai domain, so uh so we stick with this. It's a short good name. So yeah, yeah.

Anders Arpteg 16:17

Superintelligence, it's the same as uh India SearchGiver's name, right? It's it's super intelligence super intelligence, it's his company.

SPEAKER_02 16:23

Yeah, yeah, he was after us, so let's go.

Henrik Göthberg 16:26

He was after, but I just love the homage to people working at six.

SPEAKER_02 16:31

Yeah, yeah, it's uh it's a very good institution. Uh there are a lot of interesting companies. Uh, for example, Databricks originates from Six. Yeah, yeah, yeah. You have a company valuation of like at least$38 billion last time I checked.

Henrik Göthberg 16:43

Did you have the um Gods? Was it God's? Yeah, the Gods, yeah.

SPEAKER_02 16:46

He's the founder of I know, I know, I know. And he was what years was he at six? I don't remember, but he worked together with Daniel for a while. Yeah, yeah. Wow, what a duck pond, yeah. It's a small world. Finally, I meet you guys, it really is really nice to meet you.

Henrik Göthberg 17:05

Now you're getting the connections.

Anders Arpteg 17:06

Yeah, but I worked a lot with six at that time as well. But still, it's yeah, it was uh a lot of fun.

Henrik Göthberg 17:12

Yeah, and we had we have uh lale, of course. I mean, like there's so many people great people coming out of the six era, yeah. So many.

SPEAKER_02 17:19

Yeah, yeah. It's um important, it was an important institution at that time.

Anders Arpteg 17:23

Uh yeah, yeah. But please tell us a bit about the origin of the new, not the old six, but the new superintelligence computing system. How how did that get started?

SPEAKER_02 17:32

Yeah, so I uh I I I I made an exit uh around 2015. I grew my hair, became jazz musician for some years, yeah. Uh, and then uh I had a discussion with some friends uh who are now in the company about well, what could be like next generation AI? Today's AI is very boring, so or or like real intelligence somewhere uh on the roadmap or where. Um then I spoke to uh uh Peri Kulson, who is one of the what he's co-founder of company. Uh he's also the uh uh co-founder of MySQL. Very brilliant guy, a genius, uh, with a uh capital letter.

Henrik Göthberg 18:16

And use MySQL in a nutshell for the people who don't know.

SPEAKER_02 18:20

Yeah, MySQL is a standard open source database uh used by like 150,000 companies worldwide, and today uh he's also like a J. Rocket part of uh of Oracle. Yeah, yeah.

Henrik Göthberg 18:35

And he was one of the founders of that open source project.

SPEAKER_02 18:39

Yeah. So so and Perek he's been working also with AI since the 80s, so he's really skilled. He's probably the most skilled AI engineer in Sweden. There's only as far as I know, uh Daniel who can who can match him in AI skills. So, anyway, so we spoke about this next generation AI, and then Perek said, because he at the time he lived on a farm. Uh I don't know why, but uh yeah, anyway, so he said, Well, I'm coming to Stockholm next week, I'll tell you. And then we had uh dinner and discussed AI. Uh and he showed me uh an American company, it's called Numenta. Numenta, there's a the the it's founded by uh Jeff Hawkins, who was one of the founders of um uh Pawn Pilots. Uh and they had some interesting ideas. So uh Numenta. Numenta, yeah. So if you if you want to have some sort of idea about where we started, you can have a look at um Numenta's homepage. But so so basically, what Jeff Hawkins did, he he wanted to make a computer model of the brain. That was his the end goal, uh a very ambitious one. Uh but he had some he had some interesting ideas about how the brain worked. Works uh which are not like I would say standard neuroscience um theories. So this was an was an uh amazing um inspiring uh moment.

Anders Arpteg 20:14

And this was in the year like 2000.

SPEAKER_02 20:16

2017, 2018, something like that. Yeah. And then we uh so we had discussions about AI, and then we brought in uh Don Gil Blonde, uh who's who used to he was the director of AI both at uh Chalmash uh Chalmers University and also this Swedish Institute of Computer Science. And then we spent like several hundreds of workshops um crafting a AI model or AI architecture for for for artificial general intelligence. Um so so basically from first principles, what is intelligence?

Anders Arpteg 20:56

What's the do you have a preferred definition?

Today’s AI As Black Box

SPEAKER_02 20:59

Yeah, sure. Yeah, it's an easy one. So uh intelligence is a is a feature of nature. Uh it's as a uh as a concept, it's not really well defined in science, but if you look at look at it as a feature of nature and uh try to understand what why it has uh appeared in nature and what it's used for. So intelligence, uh what what's the use of intelligence in nature? If I uh if I may ask you, what would you say?

Anders Arpteg 21:29

Well, I can go and speak about this for half an hour, but I have my preferred definition. Um I actually changed my mind about this. So uh the one I uh prefer today is coming from 2019 and the paper from um Francis Cholet. And he basically defines defined intelligence as the um uh skill acquisition efficiency. So the ability to actually acquire a new skill. It's not about having knowledge, it's really about the efficiency of acquiring a new skill, and not knowledge, but really skill, meaning the uh capacity to do something real, so to speak. So that's my preferred.

Henrik Göthberg 22:08

That's a very good uh definition. Let me let me let me give giving an other angle on this and then back to you, how you continue. I think that whole definition of intelligence is is almost very, very hard, as you say, it becomes uh philosophical. But you can you can look at what are uh intelligent behavior. So this is closer to when you say what is the feature of intelligence. And if you look at that, that I think holds both uh in in biology but would also hold an enterprise, it's ultimately uh let's see how we put it, but it's about doing the um most rational or relevant decisions and actions in mutually adaptive in a larger context. So it's about you know cybernetics feedback loops, it's about uh how you how you know that in order to have intelligent behavior, you need to have the right context uh to act intelligently, to behave intelligently. But but ultimately it's then something that is close to you know, based on your objective, based on your context, uh what is the most logical uh next uh step or decision? And that then sort of manifests intelligent behavior. It's a long one because I can't I couldn't remember my my my researcher's brilliant summary. I I I I was trying to frame it. But you understand what the essence of it.

SPEAKER_02 23:37

Yeah, yeah. So I would say in that concept, uh our definition of um intelligence or the the way to understand what intelligence is, it's much more naive, so and much more um basic in in a way. Um so uh in in a in a very uh down-to-earth way, the purpose of intelligence in nature is to survive. And what's the most basic skills in order to survive? Eat and avoid to be eating yourself. In order to find food, in order to run when there's a lion around, you need you there's one uh capability behind navigation.

Anders Arpteg 24:25

So like a shark, you know, a shark um is really good in surviving. Is that me does that that mean that shark is intelligent?

SPEAKER_02 24:33

Yeah. So so intelligence is closely related to navigation. And if you think about what navigation actually is, navigation is about planning, reasoning, simulation, prediction. And independent of what kind of um definition you have uh for intelligence, those are like basic intelligence skills. First principles, I mean first principles intelligent skills.

Henrik Göthberg 25:08

Yeah.

SPEAKER_02 25:09

So the conclusion is our conclusion is if you have solved properly the problem of navigation, you have basic intelligence.

Anders Arpteg 25:20

So is a dolphin more intelligent than a shark?

SPEAKER_02 25:26

Uh probably. Um uh it's uh how do you measure intelligence? So you can use an IQ scale, doesn't work for dogs. You can use some sort of uh you can have animals looking at themselves in the mirror and see if they recognize that uh they see themselves in the mirror or not. So this is one uh example. So how do you level diff how you measure different types of intelligence is actually also content-based. So uh there's one scientist I don't remember his name, but he he he argues that probably the liver in the body possesses some level of intelligence, and we are not intelligent enough to speak to the liver and understand what uh what the word is like for the liver, because there's a complexity in the environment, there's a problem to solve, uh serving the body with uh uh poison uh cleaning the cleaning the blunt.

Anders Arpteg 26:23

In your view, intelligence has nothing to do with adaptability or anything, it's just the capability of doing something, all right.

SPEAKER_02 26:33

It has uh there are uh I would say a relation to complexity.

Anders Arpteg 26:42

Okay.

SPEAKER_02 26:42

So uh if you can master a more complex environment, if you can master more complex tasks and solve more complex problems, then you're more intelligent than not. So for example, a rat is less intelligent than a cat, a cat is less intelligent than a human, and a human is less intelligent than our computers.

Anders Arpteg 27:04

But I'm trying to if we just compare it to the skill acquisition efficiency definition, if we take playing chess as a capability, for example, you could be a person that has trained throughout his whole life to play chess, and it took like 20 years to get semi-good, perhaps. I'm not speaking about Magnus Carlton because he's extraordinary, but just whatever person. And then another person learns to play chess in um in a week. They both potentially are on the same level, but one has had had to learn for 20 years and another one learned it in a week. Would you still consider them to have the same level of intelligence in that case?

SPEAKER_02 27:43

Chess is uh is a bad example. Uh any capability. Okay, but this chess is interesting because it's uh this is a quote from Jan Le Kun. It shows more like the that we humans are not so good at chess.

Anders Arpteg 27:54

So still so that's an example. But doesn't the efficiency of acquiring it play any role? Or is it is basically intelligence the same as knowledge or the ability to perform a skill?

SPEAKER_02 28:09

There's a definitely a relation to the the the speed of acquisition of the skill, I would say. Yeah.

Anders Arpteg 28:14

The speed of acquiring a skill, okay.

SPEAKER_02 28:16

Yeah, but but also uh there's another dimension. So it's easy to to work one dimensionally. You have but if you add like uh the volume of data you need to acquire the skill.

SPEAKER_03 28:29

Yeah.

SPEAKER_02 28:29

So you have like time, but you have volume. So um, and if this person who's learned chess in a week has consumed the same volume of data as this person who's done it for 20 years, so it's basically the same uh but something that requires less data is probably more intelligent if they can then reach the same level of quality.

Anders Arpteg 28:51

I would say that. So it's something about skill acquisition efficiency still, right?

SPEAKER_02 28:55

Yeah, I would say, yeah. But that that turns and this boils down to okay, so if we uh if we uh in this context agree that intelligence is related to navigation. So what's what's navigation? How how is how is navigation related to to more general intelligence and higher cognition? So um moving something from A to B, that's a this is a that's a navigational problem to solve. And moving yourself is the same type of problem. Uh and if you consider all the type of objects that exist in the world, all type of materials and so forth, there's a lot, a lot of information, and there's a need to generalize in order to master the world. Um, and if you can master the most simple robot task of picking place, actually you have when you have that, you already have some uh um capability to do basic assembly, putting two things together. That's picking place, but maybe two arms and a more advanced destination. So you can go from picking place and a combination of pick and place uh actions, then you have a more uh complex behavior. And how how again, how is it this related to higher uh cognitive functions? Well, uh our view is that solving problems is still navigation and picking place in an abstract real. So when you're you're looking for uh, for example, when I'm looking for this uh Fermat proof, I I I I explore different paths in the head. I'm moving mathematical symbols around when I'm taking a shower, uh, and all of a sudden I can see a pattern, uh I can see some sort of destination. So uh so if you have navigation in in a sort of naive real world, uh you have the the basis for artificial general intelligence, and this is what we have built.

Henrik Göthberg 31:01

So this is so so it's somewhere to to crack the code of the fundamental lowest level first principle architecture for navigation, exactly. And then as you do that, it uh it's a matter of scaling that into more and more complex or abstract uh context.

SPEAKER_02 31:17

Correct. So you need to master pick in place first, and then you can do industrial assembly, maybe some construction sort of, uh, and then you can you can make your PowerPoint later, but you cannot start. It's like a baby, you can't put uh a baby in a car and expect it to learn how to drive. But the 17-year-old can learn how to drive in a few hours. So, and this goes back to the the other analogy or of the uh the principle of navigation. So, what do you need in order to navigate? A world model. Yes, a map of the world. You need a sort of map so you know where you are, where the objects are, the relationship between the objects. It must have some sort of uh model of basic physics, right? And this type of and this model is structured. The the keyword here is structure. Today's AI model don't have this type of structure.

Anders Arpteg 32:07

Hold you for a second because we're getting into a really interesting topic. But I just want to finish the previous topic, which was you know, what was the origin of the company, the superintelligence computing system? So if we just you know try to figure that out, and then we get back to this point, you know, what are you trying to do there?

SPEAKER_02 32:23

And yeah, that's so the realization was if you can describe superintelligence, if you have like a blueprint for superintelligence, you can actually build it. That was the realization.

Henrik Göthberg 32:34

So let's do it. And and then and the logic here is that how how can anyone have a blueprint for superintelligence when you start top-down, or if you start from the right angle, uh it becomes uh impossible. But what you're trying to do is to trying to sort out the principal pattern that you can now scale, or that you can now take into more complex so so so to crack the the inner pattern.

SPEAKER_02 32:57

Yeah, it's a very good summary. So so in our view, today's AI is a top-down approach. A lot of data so forth brute force. Uh, but we're doing it bottom-up instead.

Henrik Göthberg 33:08

Trying to crack the inner pattern.

SPEAKER_02 33:09

Yeah, and and it that if uh and and in order to to build a structured model, you need a mathematic uh mathematics that have uh which is good at capturing structure and which is structure-preserving. And when you have that, when you have this structure, you solve many problems with today's AI in one swoop. For example, with structure with structure, you you get you get stability, you get much more robustness. And also with structure, you have like you have better transparency but better understanding what the model is about. So it's not just a black box anymore. And also the structure makes the information available for many-step reasoning in a stable way.

Team History And Roots At SICS

Henrik Göthberg 33:54

There's so many tangents here because as soon as you now say like we we want to do deep learning, we don't want to do mathemat the mathematical fundamental principle that can scale and learn by itself, but you're also with the world model view, you're you're you're entering into this is structure somewhere also. And now you're getting into this whole neurosymbolic or whatever we want to talk about, how these worlds merge. Even these conversations happen anywhere else as well. So this is but but because you have some part here in your thinking which is structure and one that is learning.

SPEAKER_02 34:27

Yeah, exactly. Yeah, very good.

Anders Arpteg 34:30

Okay, so you came together a number of really famous expandable, no, not satisfables. Expandables? But okay, so you formed the company in at what year?

SPEAKER_02 34:45

Was it uh 2022. In december, we finally found the companion. Then we had the sufficient uh uh work done on the on the AI-model to to start using it in a real life setting.

Henrik Göthberg 34:58

So, workshop, workshop, workshop, or you know, gatogethers and whiteboards?

SPEAKER_03 35:02

Yeah, whiteboards.

Henrik Göthberg 35:04

Since between 17 and 22, Covid that was like that's special.

SPEAKER_02 35:10

Difficult to use whiteboards, but we managed somehow. Yeah.

Anders Arpteg 35:13

So the goal of the company, I guess it's evident from the name, but still, you know, what are you trying to do with the company?

SPEAKER_02 35:18

Yeah, the the end game is artificial superintelligence. Cool. Yeah, that's cool. It's pretty cool.

Henrik Göthberg 35:25

And but also know when you start with the uh with the pattern view, you can now start applying it and solving narrow tasks as you grow to more and more complex tasks and you can have usefulness, is I assume is your idea, without uh having reached superintelligence, whatever the definition that is.

SPEAKER_02 35:45

That's uh that's that's a a somewhat valid point, but actually, what we are so what we're doing, we are the development of the AI uh follows the development of a child. So the the first thing we need to do, like a child. When we're born, we have like 700 megabytes of information in our DNA. That's nothing. Uh a grown human has like 20 petabytes uh of uh information stored in the brain. All this information needs to be learned. And basic things about the world need to be learned by a baby. Between six and nine months, uh the baby spends a lot of time learning about gravity, how that works. Uh a baby that doesn't first understand the uh called the object permanence that when daddy leaves, he's gone forever in the baby's world. So these things, basic things have to be learned. So we need also to start at this baby level. So we have a baby AI right now doing picking place, like a baby playing with his toys. And this uh it's an it appears to be a narrow task, this pick in place. And when I spoke to um over license ring, he showed me an article from ACEA from 1984 where they had uh a rule-based system for doing pick and place automatically with a robot. So this is like very old news. And picking place has been solved in the specific case since that time. But the gen but the general case to hand to be able to handle all types of objects in all types of environments, all types of materials, that's an unsolved problem. And this problem needs to be addressed and solved first. When you have that, you have the foundation to grow further. So this is this general case is not a narrow task. It's a super super task that requires artificial general intelligence to be solved. And also the other way around, if you have solved it, you have artificial general intelligence or the base the basis of it.

Henrik Göthberg 37:50

And is this the reason why you quite fast moved into robotics as your or your physical intelligence, so to speak, uh physical movement as a key focus of uh of your AI? Yeah, it's great fun. Yeah, yeah.

SPEAKER_02 38:06

Yeah, but but but but the thing is, since we have an AI that has a different type of learning and uh which is interaction with the with the world and the environment, if you have a physical interaction with the world, it's much easier to understand if you're on the right track with your development. Yeah, it's something that is actually quite uh measurable or observable. Yeah, and and there's an abundance of training data. Just put the put some sensors up, cameras, tactils, or whatever. Uh you have an abundance of data to use for your training.

Anders Arpteg 38:36

And I know you don't want to go to too much into the details how it works, but still, if you were to, in an abstract way, still describe what is different from your design architecture of how your AI works compared to the AI of today.

Henrik Göthberg 38:52

And AI of today would be use the LLM baseline or that.

Anders Arpteg 38:56

AI in general, I say deep learning or some other traditional machine learning or even rule-based, if you want to.

SPEAKER_02 39:03

Yeah, that will require uh entering into the mathematical structure behind. So, but I can tell you what it's not, Randall. If I describe today's AI, uh, which is a brute force approach, you have trillions of parameters, uh, it's a huge black box. With a large parameter space, you can uh uh parrot any type of input data in your mathematical model. But uh the the the the parameters they are really uh it's a I uh uh this example with the synthesizer. So if I say if I say something, say say hello, I get a sound file, I can use 10,000 voices, uh uh adjusting the amplitude of the sine waves, uh, the phase and the frequency, and can I can make the system imitate and say hello? But there's no uh no understanding behind. So so this is the way that today's AI works. So in order to make uh The system is more intelligent. In order to build a foundation for this world model, you need to find a mathematics that's is richer in structure. I can give another analogy. This is from physics. There was a famous physicist, he became famous after uh an equation that he uh discovered the Dirac equation. So Dirac he struggled with uh merging physics, special relativity, and quantum physics. And he had this equation, uh, and he tried different uh variants, and then by some peculiar reason, nobody knows why, it's a mystery, but he replaced numbers pure uh or variables with a more sophisticated mathematical device called matrices. And when he did this, all of a sudden the equation worked. And with this more complex mathematical structure, he could infer theoretically from the equation that antimatter exists. So he dis he he he uh described antimatter through mathematics, through the structure, so the more richer structure mathematical um representation.

Henrik Göthberg 41:16

So the analogy here is that we need to replace with a more sophisticated math or more sophisticated architectural structure, it is still fundamental principles, but simply as he did go to matrices approaches, yeah, something then happens. Yeah. And this is this is now we are touching the the core IP, you know, like you're you're thinking about different mathematical structures at the right.

SPEAKER_02 41:39

And again, same thing with with with Einstein. So when he wanted to uh add gravity to uh to his special relativity, he needed another type of mathematics. And he used this Riemann mathematics uh with curved space, which was already existing pre pre-Einstein times, and also a lot of of the uh the results from physics about the uh length contraction and time-delation. This was also results known before Einstein. It's called Laurent Fitzgerald contraction. So Einstein he managed to put things together uh with again uh some fundamental and very simple principles actually. The speed of light is constant, uh uh and the the uh the laws of nature they behave uh in the same way in different of if you move uh different reference.

Henrik Göthberg 42:33

Would it be fair to say that when the team came together in 17 to 22, would you had this? Has this emerged as a strategy? What you're trying to do, replace some with more sophisticated math, or has was that the like the founding ideas that maybe one way to get get faster or closer to AGI is to simply do this finding the right uh more sophisticated math. Was that emergent, or did you think like that from the beginning?

Defining Intelligence And Navigation

SPEAKER_02 42:59

Well, the the development. So when you think about things and when it's solving a problem, you're going in that direction and then you go another direction. Uh so you you need to start to to uh to do this like uh homework about what intelligence is, what kind of feature you would like to capture, and how uh and that's a long discussion in itself. And that's a long discussion in itself. So this is an important feature. This maybe not. This might be uh could we borrow this from neuroscience? Is that a good principle? Yeah, it was a good principle for a while, then it was thrown out. Uh, so it's an it's an iterative development. So so we start in this this and then we ended up like over there, and then when then we're back, and then we reused some ideas from five years ago and so forth. So it's iterative, yes, yeah.

Anders Arpteg 43:44

But if you were to describe a bit how it works today, and the thing you actually had put in production, if I understand correctly, how would you describe it without giving away too much uh secret, so to speak?

SPEAKER_02 43:56

Yeah, the the the interesting most interesting feature is uh probably that it uh learns through feedback. So it's not the traditional batch learning uh AI. So it's more human-like in that sense. And it's turned out to be to it's turned out to be uh much more efficient. So the the the the training cost of the uh AI we have on site today is is virtually zero.

Anders Arpteg 44:20

So compared to how it normally works today, normally today you have this kind of you know batch or offline training where you where you have a huge data set and you train over and over again and an update parameter, and then you start to use it in inference mode. So you have this completely separate training and inference uh mode. But I guess if I understand correctly, you have more of an online mode, meaning you actually do both training and inference at the same time. Is that correct?

SPEAKER_02 44:46

That's a simplification, but uh that's uh that's I would say online learning is the is the end game. So the the thing we have uh in production right now is continual learning.

Henrik Göthberg 44:57

So so can you help me who is not as steep as you guys? What's the difference between continual and online?

SPEAKER_02 45:04

Continue continuous learning, online learning, then you learn all the time. Continual learning, you have chopped up it uh somehow. And you can uh for for this uh use case pick a place, uh you it's uh uh it's it's um you don't need to feed back the the model uh uh like like every minute or something like that. So you you have uh you can keep just long time intervals if you like, but you can make them shorter if if needed. So so depend I would say depending on the complexity of the um of the task, you you you need to go fast, yeah.

Anders Arpteg 45:48

So it's not like online in in like every kind of data point, but it's some kind of small batches that updates the model.

SPEAKER_02 45:55

Uh it's not really like that either. So but I can go on more into detail, sorry about that. Oh, you can't it's a really it's a different uh different setup altogether in that aspect.

Henrik Göthberg 46:04

So it's such a challenge sometimes when you want to describe your your your your thinking and philosophy here because we are we are now going down so deep for the last couple of years in some certain patterns and and we talk about you know it's it's pre-written. What's training, what's inference? So it's it's all we're already down in one rabbit hole, so to speak. And now you're saying something, oh those words I could use them, but they're not really my words. I need now the semantics to capture this.

SPEAKER_02 46:33

Exactly, exactly. And we have this IP protection because we we we're not a research product, we are funded by uh VC VCs, and about the uh difficult difficulties. Uh a VC asked me in this type of pitch session, what's the most difficult uh question you got from a VC? And it's like, how does it work? So yeah, it's a it's a big uh it's a big problem, and it's uh okay.

Anders Arpteg 47:01

Perhaps we can get back later too. If we can try to poke you in more later, but okay, I get you. We we can't really go into too much detail there, but perhaps you can speak about the application a bit because you you do have used it for real, so to speak. So if you could, could what is your current system used for today?

SPEAKER_02 47:18

Yeah. So if you look at uh industry and industry automation, there are like uh three core values. Uh first one is robustness, this is very important. A typical industrial robot can run like 100,000 hours mean time between failure if it's uh oil changed and so forth, we're properly served. It's like 20,000 hours to uh till the first oil change. So robustness is uh key. Uh and this is a problem with today's AI, with hallucinations and so forth. So, due to this problem with hallucinations, AI cannot be widely used in industry. Second uh is scalability. Same robot, same program. Anywhere in the world would do the same job perfectly. But with with AI, well, the AI deployed today in industry in picking place. Uh it needs a lot, a lot of data to run. Uh and the more complex environment you have, the uh the more progress you have.

Henrik Göthberg 48:16

Specific for the environment.

SPEAKER_02 48:18

Yeah, yeah. So so the the uh this uh training makes uh today's AI not enough scalable. So not efficient cost efficient enough. As you said before, that uh if you have uh the the the a big enough model and and uh uh enough data, then you can uh keep the um the outcome decent, but uh you don't really know when it will fail. Uh so the the the the um mean today you have in order to mitigate this is you don't have end-to-end AI. You have a mix, a hybrid between AI and program. So the program is needed for safety guardrails, it has like human heuristics, uh hand and special cases, so so forth, which uh degrades the uh uh uh the scalability. So this is the the problem today, to to summarize tradition robots robust, scalable, not so flexible. AI flexible, but not robust and not scalable. So what we have with our solution high level in industry, we manage to have flexibility without cannibalizing on robustness and scalability. So this is basically the problem we're solving. We're adding robustness, opening up for wide-scale deployment of AI in automation. And but you also do have yeah, and uh and uh yeah, and also if you if you look at industry today, if you look at uh uh warehouse logistics, you could potentially uh automate picking place. 95% of all warehouses have no sophisticated automation. Amazon's uh warehouse in Sweden, zero automation, 400 people picking goods. Uh uh Lieder's uh warehouse in Halstad, no automation. Because it's it's a complex environment with high variability, unpredictability, and unstructured. And if you look at industry uh and uh assembly, well, you think about industry manufacturing industry, lots of robots, yes, but between 70 and 90 percent of all assembly work is still manual. Because the the the environment is is too challenging for for today's robots.

World Models And Bottom‑Up AGI

Henrik Göthberg 50:35

So flexibility is not the the strong feed. So you know when Amazon is doing something super cool, it's because they are highly standardized a very specific warehouse and then then it works. Yeah, but but as and the same with uh in in car manufacturing, you can hand highly standardize certain moment tasks and then it works. But then all all the complexity around there, or you know, that that still takes too much time to retrain or find a new model, or uh each step needs its own model, etc. etc. So that's the flexibility that having something that you can redirect to other patterns, you know, to learn something new.

Anders Arpteg 51:14

But but I guess in short, it's basically that I think people overestimate what AI can do today. If you put it in real-world situation, it can't really do that much, right?

SPEAKER_02 51:23

Yeah, so solving real-world AI.

Henrik Göthberg 51:26

Yeah, that's the key to intelligence, to real intelligence. And this is I mean, many are going down this path that you know. Are we really solving real-world intelligence when we're going down this uh linguistic way? No. Uh this is a fast answer, right? And then and then we we've been joking about wh why is uh you know who is on top of this race? And we always said, uh, don't count Elon out because he's working on all the angles with with this Tesla car, with his robotics and etc.

Anders Arpteg 51:55

So that but it's something you're using the term real-world AI a lot. Yeah, yeah, yeah. But still, you know, we we love to just understand a bit, you know, what is the use case. I don't think people really have heard yet, you know, what are you using it for?

SPEAKER_02 52:07

Yeah. So back to warehouse. So uh in warehouse logistics, uh, the business of warehouse logistics is moving things from A to B. It could be products, it could be cars, it could be whatever. Uh and it's uh it's it's very important with uh precision. Uh, if you move the things to the wrong place, that could cost quite a lot. So so this is the the the core functionality you you must have is robustness, uh, and you have to master also a bit more uh complex environment than you have in in manufacturing industry where you're moving the same pieces uh all the time.

Anders Arpteg 52:46

Put words in your mouth, sorry for that. But in short, you have a robot that you control with your AI that can move parts on the production line in different places.

SPEAKER_02 52:56

Yeah, thanks. Exactly. So we do pick and place automation. Yes. So our AI is controlling a robot. You can say that it's a uh an AI robot brain that we're building.

Henrik Göthberg 53:06

Can you can you explain the you had a very nice example case before we started the pod, which is a quite quite interesting one with with a lot of different SKUs pieces and with a lot of variety every year. Yes. Can you explain? Because I think that sets the tone.

SPEAKER_02 53:21

Yeah, sure.

Henrik Göthberg 53:22

It's exemplifies for me really well.

SPEAKER_02 53:24

Yeah, so so again, these core values: robustness, flexibility, scalability, flexibility a way of of uh um having flexibility the need for flexibility is if you have high variability. So in this case we have uh uh 20,000 different types of products, and you have 2,000 new products entering every month. So you have this problem with viability, and we can handle that with with our AI, and for a uh very um uh I would say reasonable effort. And then we have the the robustness, the the pick in place accuracy, and we we are picking with nearly 100% accuracy for these um products, and how are they presented to to the AI robot? Well they they they enter uh on the conveyor sort of uh not really, but they enter um in a in uh a bin, unordered, uh in a pile. So it's it's what you call 3D picking. If you have uh the the items neatly on the table, it that's 2D picking, but this is really 3D picking.

Anders Arpteg 54:38

So okay, so with picking and placing the different parts in a proper place, and it is adaptable, I guess, because it can handle new parts. So you had you said something about 2,000 new objects coming in every month or something, yeah, and replacing others. How how does it know what to do properly then? Is is some human try showing it how to do it, or is it uh learn adapting or learning in another way?

SPEAKER_02 55:01

Yeah, so so we are engineers, so we're taking all the shortcuts we have uh available to bootstrap the system. So before it's put in production, it's it's bootstrapped in different ways. If we have ready-made metadata, we can use that, and then the uh as that at that starting point, it improves from from that on. And then we come to the third uh dimension, which is scalability. So this uh AI model is crafted in such a way that it's one AI model that controls all the robots. Uh how is this possible? Uh well, all the robots that live in the same world. So if you have a good generalization in your world model, it can be used by different types of robots.

Anders Arpteg 55:45

But what happened if a new object comes in there and never seen that before? How does it know what to do with it then?

SPEAKER_02 55:51

How would you do?

Anders Arpteg 55:52

I asked a friend or a colleague to show me what to do.

SPEAKER_02 55:55

That's one way of doing it.

Anders Arpteg 55:57

And how does your system do it?

SPEAKER_02 55:59

Or the other ways to to learn?

Anders Arpteg 56:01

I don't know. So you tell me.

SPEAKER_02 56:02

No, if you if you if you experience this this can or and you can you can you give me the can and what and you don't know how to lift it? What do you do? You try out, yeah.

Henrik Göthberg 56:12

You try out, you read the label, you see what it says. Yeah, you you you take it up and you try to figure out what information can I uh gather from this. Yeah.

SPEAKER_02 56:21

And then yeah, you have you have your memory, have your word model, this resembles something that you have experienced before, maybe, maybe not. Uh and if it's completely new, then you need to be like curious.

Anders Arpteg 56:34

So no human feedback at all?

Robotics Focus And Pick‑And‑Place

SPEAKER_02 56:36

No. No, no, uh I would not say no. We can use human feedback if if we like, and to bootstrap the system. So uh and also uh I haven't spoken about language, but again, if you have structure, you have the basis for language. So you can have uh a language which is semantically grounded in the real world, and you can if you have that, you can actually teach this the system using language, which is bottom-up grounded and not this type of uh large language model type of language. So this can also be used to to teach the system. And the seats the system can also communicate back with uh and ask questions, and those are real questions, like um so I spoke with Yussi about that. So what what's the first what's the first time type what's the easiest type of of language that you can can have in this type of system? Is it like is it like yes or no? And Yussi said like no. Why is that? Yeah, well, no is a very complicated linguistic uh term, whatever. How is that? Well, no is heavily context dependent. You can't understand what no is unless you have you know you know a lot of things about your environment. So this is a very complicated thing to to uh to use. Okay, so what what's what what can we use instead? What where do we start? Well, uh the probably the most uh basic uh linguistic start for the system is you uh it can com communicate uh normal conditions, and then when you have something that deviates from the normal condition, the system can can uh alarm you like a baby.

Anders Arpteg 58:27

So so so does the system understand text and uh speech today?

SPEAKER_02 58:32

Uh if we if we would like it to do so, it it uh I would say we don't have the linguistic uh capabilities uh implemented in the system today. Yeah, but there's nothing that in principle prevents it from having that type of capability. So okay.

Henrik Göthberg 58:49

But we are uh I think this is super interesting, and I think we could do the whole pod here, but we also have a bigger theme. Do we uh you do we want to I'm I'm you I could stay here, but also I'm I'm curious to stay on the bigger topics of AGI because I even you know before the camera was on, was you can we can come into a real discussion and debate on what is AGI or uh and all and and and down into the theme. So what do you think, Anders?

Anders Arpteg 59:13

Yeah, I'm gonna think it's gone one hour. The question is if we should do a new section or not, or should we if we want to?

Henrik Göthberg 59:19

I we it's since this is the first pod for the year and there's been some stuff going on. Yeah, uh, I think we should have a little small use section. We we have you usually break in the middle and you know talk about, you know, did you hear about this uh or have you tried that? Yeah, well let's do it. Let's do it. Let's do it. It's time for AI News, brought to you by AI AW Podcast.

Anders Arpteg 59:46

Cool. Yeah, as I said, Henrik, we usually take a small break. We we usually fail in making it small, but we try to make it small and the other thing. But it's fun to, you know, to speak about some uh favorite news that happened in recent weeks, or it gets longer in this case potentially. Anyone that wants to start to have some favorite news? Have you something, uh Kream, that you heard about recently? Some news article? Something you tried, a new model or whatever.

SPEAKER_02 1:00:13

I like these human or robots doing karate kicks.

Anders Arpteg 1:00:17

I don't know if that's the news, but this is a unitary thing, or which one? Yeah, yeah, that one, yeah, with a watermelon. Yeah, yeah. That was really cool.

Henrik Göthberg 1:00:25

I mean, like the one topic that sort of no one has um missed is the whole the clo uh what's it called?

Anders Arpteg 1:00:34

Short bot, but it's changed name to multibot now.

Henrik Göthberg 1:00:37

But yeah, yeah, yeah, yeah. The whole thing and that where people start setting up uh their Mac minis at home and with with no clarity and no authority and you know letting it loose, and it's amazingly powerful, but you know. Should we say something about that? What is your take on that?

Anders Arpteg 1:00:56

No is really cool, and just give some background about this. It's based, you know, we had Claude Code, which is like a command line uh tool that is really powerful, and you can do a lot of coding with it, and a lot of other stuff besides coding as well, and you control it through the terminal. And then recently, Anthropic also released Claude Cowork, which is uh a bigger framework where you can actually operate. Yeah, we should have mentioned that first. Yeah, so you can basically operate um on your machine, um, and it can have access to files and it can do much more than you otherwise could.

Henrik Göthberg 1:01:29

So it really and the value propose to get it out of the hands of the pure coder and and make normal people find the value of it.

Anders Arpteg 1:01:39

Yes, so it's it's more of a hygentic assistant or what we uh should call it, um, but it really enables it to take action to an extent that wasn't possible before and in a very much easier way of using it as well. But then um there came this kind of open source test a point, right? From a single person that um has created what they call the Claude bot.

Henrik Göthberg 1:02:02

Claude bot open source.

Anders Arpteg 1:02:04

But it wasn't spelled like Claude does, it's basically spelled like the crawfish kind of thing with a claw. But then uh Anthropic Susan said you can't use this even though it's spelled differently. It sounds like Claude, so you can't do it. But it's basically based on the idea of Claude cowork. It's just adding a lot of integrations to it and a lot of functionality and makes it much more general than what uh cowork was. So it's super easy to install. You can run on Linux, Windows, Mac, or whatever, and um, then you can connect you know all the different types of accounts to you. You just give your password away for to all the what's give your life away and do telegram signal accounts or whatever, and it can read it, it can write there, and it can start to interact with people in a way that um that is very, very nice, but also a bit scary.

Henrik Göthberg 1:02:52

Yeah, but this is so many ongoing, this is so many tangent stories here. So like one story is about people put it straight on the port with no security, and no, so it's completely open. Uh, it was a side story that all of a sudden now all Mac minis got sold out because people went, you know, people realized oh this this the the the the architecture in a Mac mini is quite uh useful for this. Uh, you know. So so there was this tangent story, uh GPU and Nvidia is dead. Now Apple will take over with the world with Mac minis. You know, so there were many side stories that made this whole very interesting couple of weeks.

Anders Arpteg 1:03:29

But it's also a big shift in how traditional chatbot works compared to this one, because the traditional chatbot is very uh reactive, meaning it doesn't do anything until you tell it to do something. This one is much more proactive, so it can basically go and check you know every five minutes or whatever you want for something that happened, and then it contacts you. And it also has this kind of uh I don't know what they call it, uh but self-learning kind of abilities. So it it tries to do things, and when it finds out that suddenly the way I usually answer an email or whatever doesn't work anymore, then it actually self-corrects and it does so without the user telling it to do so. So it's it's trying constantly to improve its skills, so it has a set of skills, you know, similar to anthropic skills, and and it's really cool. So so becoming these kind of more proactive agent or assistant or butler or whatever you like to call it, uh that I take actions without you telling it to do, uh, and then having access to all your accounts and whatnot is a step change, I would say.

Henrik Göthberg 1:04:36

Yeah, but and I'm interested in the bigger picture here. I mean, like so one of the key bigger pictures is that I think I mean, like you you you you are you are discarding the whole LLM story altogether.

Continual Learning And Efficiency

SPEAKER_02 1:04:46

No, no, no, it's an amazing engineering um uh uh feat and very, very useful. Uh and uh Perirek is trying to convince me to use this uh this clawed uh claw bot or whatever. Uh and I've been really skeptic. But when when I tried claw with this mathematical proof, uh I now I uh uh uh it's on my to-do list to install this uh but but but the whole McMini story to me really shows that if people think AI will only grow up around one axis, the fundamental logic of large LLMs and more compute on that.

Henrik Göthberg 1:05:22

I think that I think that in all its patterns, you know, it's should sho it's showing that the way we reach AGI is you know more access that will need to be better. And here we have the example of well, oh cloud you know, the Mac Mini will not take over, but it clearly separates where where what you want to do locally versus what you want to do centrally. Yes, and and and that's a shift to the for us this is obvious, right? But for a lot of people, I don't think we they don't see that.

Anders Arpteg 1:05:53

And I have to say, you know, you don't have to use a Mac Mini for this. I hate Apple driver, so I should say that. You can use MLIDA has awesome DDX Spark machines and whatnot that is perfect for this. So you don't have to use a Mac Mini otherwise. Anyway, uh but I think you know one uh interesting part is really the adaptation it is having. So, you know, we have of course the LLM underneath, and it is hard coded or it is fixed, it's it's offline crane, right? So it doesn't update the actual model it uses, it actually uses the API. So you just go to Gemini or Cloud or OpenAI, whatever. Um, but it does adapt other things. So it does basically have a set of markdown files, simply text files that it saves on the file system, and then it updates those all the time. So it has a long-term memory that it does, it has skills, and it has I think it called it a soul personality. So it try to adapt to the personality that you like, and it constantly updates that by itself. So by adding this kind of scaffolding or extra parts of the system to it, it does start to have a memory that the normal kind of LM does not have. It starts to actually have a more of a behavior adaptation that a normal LM does not have. It does actually come a bit closer to continuous learning, meaning it doesn't really change the parameters of the model, but it changes the skill files and the other parts autonomously.

Henrik Göthberg 1:07:18

Yeah, and and this to me, I want to try it with you. It also pushes to me a much clearer view that you know what, if you want to do AI properly, you need to have a systemic thinking about it and you need to build and you know, with today's technology to make that work and to you know minimize hallucinations or make it you need to build AI compound systems. So you're sorting out, you know, you might have a core LLM somewhere, but to do it with today's technology, you wouldn't you need the other parts. And what you see, what this open source project is, is several moving parts as that together makes us closer to continual learning, makes us closer to this. I think we should go there because I know you also, as me, has um a big interest in the Jan Le Kun kind of JEPA architecture, which does have multiple modalities or we have talked about this for on this pod for several years now, that it it's not about the problem that with the LLM, it's the problem that people think that's the only thing.

SPEAKER_03 1:08:26

Yeah.

Henrik Göthberg 1:08:27

And and actually quite sophisticated big project projects or enterprises talk they are they are they are they don't understand the problem with learning and navigation, and that and that okay, you can solve that with you know and yeah, yeah.

SPEAKER_02 1:08:41

And then if if you if you in some way, magical way, could make the the LMs have like a uh dynamic long-term memory, then the problem is solved. And this maybe what some people are hoping. Uh in the brain, you have like uh the the the neocortex where you have this long-term memory, and then they have a sort of scratch more short-term memory in the hippocampus. So with this new type of uh claw thing, whatever you have some sort of hippocampus layer which works. Uh huh.

Anders Arpteg 1:09:12

But but it's not really weird more things than memory that's required to achieve AGI buttons there. Yeah, yeah.

Henrik Göthberg 1:09:20

But maybe do we have some more. This was the one main news. I I I couldn't not help also have you followed the whole new hype word of context graphs that also, I mean, like if you if something we didn't talk about before Christmas that really exploded in January, context graph this and context graph that. Someone wrote a paper and someone tried to push that whole logic. Nothing that you really knowledge graph is like uh old stuff now. Not knowledge graphs, context graph.

Anders Arpteg 1:09:58

Okay, let's not go there. Perhaps the small story could still be that they started to put ads in in Chat GPT though. But I guess they have to do that, you know.

Henrik Göthberg 1:10:06

That's the that's the monetization model problem. Yeah. Anything else that was sort of there there was some new I we can stop here and we can take it next week. I think I think this one was the fun one to talk about.

Anders Arpteg 1:10:21

Yeah. Okay. So getting back to that, and I'm not sure how to, but should we go perhaps to the Jeppa? That's a good segue. Um just continue there. What's your view of how would you describe what Jeppa is?

SPEAKER_02 1:10:34

Regarding Jan Le Kun, yeah. We share some uh high-level ideas with Jan Lee Kun, but not the Jeppa, I would say. Okay. And there's something that that works with, but um there's some there's some fundamental flaws uh in it which uh which prevent it from really working. And I can't I can't really talk about it. I would I would love to do it. But it's yeah. I can just give a hint.

Henrik Göthberg 1:11:02

It's some can get if you what's the pro what's the what's what what is good with Jeppa and what's problematic with Jeppa?

SPEAKER_02 1:11:11

Uh if you want to build AGI. And I can't really tell you why. Yes, you can.

Anders Arpteg 1:11:16

Come on. Okay. So much fun. But uh okay, how would you describe Jeppa then? If we if you start there.

SPEAKER_02 1:11:25

It's um okay, high high level. Oh, can you can maybe explain it this this way? Jan Le Kun is not into robotics. So uh if you if you challenge with the the the real world and the interaction with the real world, you you you you have uh another another set of uh problems to solve than then is on the table for Jan Lee Kuhn.

Industrial Robustness, Flexibility, Scale

Anders Arpteg 1:11:51

So that's one way of But if I give a quick like context then, or my view of Jeppa perhaps, and then we can continue from that point. And um, you know, he's been famous for saying that he doesn't like the auto-regressive nature of LMS, meaning that the error uh compounds as you know predict next token after the other, and it can't really backtrack you know once you predicted one of the tokens, so it just will have an exponential increase in error as you auto-regressively continue to predict token for a token. So he wants to find an alternative to this. And then he has this kind of joint embedding uh predictive architecture where he wants to do the prediction in some kind of latent space or energy space, as you call it. So instead of having to predict one token all the time and going you know through the whole model with a full feeward forward pass for every single token, and then you know taking the previous token that you generated, which is syntactic, you know, what token should be, going back into some kind of middle layer out of these hundreds of layers, a semantic understanding of what the token means, and then back to the syntactic representation in token space. It works semi-good for text because text is rather high-level abstraction, but it's impossible to do for pixels, for example, for videos and images. So even today's image image generators do do not do like this. So they're not auto-aggressive in this way. So even today's image generators work in a non-autoaggressive way, I would say similar to the path of Jan LeCoon's predictions, meaning it has an auto-encoder around it, it takes the pixels and um and find an uh a latent representation, then it does the diffusion transformer part of it. So it actually instead of doing the auto-aggressive, uh, do a diffusion from you know a very noisy uh version to the clear picture, as you see when you uh generate image today. And and so so it's already moved, I would say, in his direction. He doesn't say do not use transformers. He was basically the one that came up with the self um super super uh vision uh techniques. So which you know auto-aggressive models to use. I mean, he's very much in favor of transformers or uh using that type of technique. It's just the the auto-aggressive nature and being able to do prediction in kind of a sensor space or text space or a pixel space that he appliced to. But instead of doing it in in some kind of relatent space and then having these kind of world models and memory and different components that he has a beautiful picture of. And when they have that compound together and can do prediction, multiple step prediction, which I heard you speak about as well, in this kind of semantic space, just as we do as human when we reason and do not speak until we have thought. Some people do that anyway.

SPEAKER_02 1:14:40

Uh but anyway, that's that's basically lots of system one going on in speaking sometimes. Yes.

Anders Arpteg 1:14:46

Yeah, but I think that that's a good, you know, then it's that's kind of hierarchical Jeppa where you can do you know reasoning in different levels of uh abstraction, etc. But I I would say would that be a fair kind of description uh of what uh what Jeppa is gonna say?

SPEAKER_02 1:15:01

That's it's probably a good description.

Anders Arpteg 1:15:04

Okay. But you also spoke a bit about, and then I really hmm it's hard when you don't you know want to divulge. But let me try it like this.

Henrik Göthberg 1:15:13

There's some if if we don't go now in the techniques, in the detailed techniques, but but talk on principle level. So world model is something that uh you believe in. Yeah, that's central. And that's central, and that I think is central also for Janikun's reasoning. Yeah, yeah.

SPEAKER_02 1:15:29

We share that idea.

Henrik Göthberg 1:15:31

That's a fundamental idea as a principle. I think you're sharing uh the fundamental idea on online learning that we we need to go there. Yeah, exactly. And so so basically it's not enough until you have those components. And are those components the same as as the core underlying LLM or whatever? Or is it you know, an is this an architecture? So so on so I can see how you share principle ideas, yeah. Then how to tackle them might be what where the magic source is. Yeah, that's different. So let's let's argue on that one. What on principle level do you share ideas with Jan Lee Kuhn on these topics? The one we said now, you know, and what and are there are there some fundamental principle you don't share with him?

SPEAKER_02 1:16:18

You you you uh summon it in a good way. So a word model, structured word model, paired with online learning. That's the N sense. That's the the the core uh navigation. Navigation is part of a word model. If you a word model is a necessary condition for navigation. So navigation is like operations on the word model or based on the word model. Paired also with uh, of course, the sensory input. So you have the member in the word model and you have sensory input. Uh, and and this uh what a learning principle, which um is uh uh say based on uh uh interaction feedback.

Anders Arpteg 1:17:01

So you you still have some kind of backward propagation kind of learning, right?

SPEAKER_02 1:17:07

I'm not going to details. So I really I'm really sorry about that.

Henrik Göthberg 1:17:12

Okay, respect that. Okay, but but let me that is technically stupid, so I can't ask ask the uh the tricky question. So when we say online learning, have we solved the short-term memory versus long-term memory, or is this something completely different?

SPEAKER_02 1:17:29

That's part of it. You need the you need a long-term memory, uh, and then you need a short-term memory. And they work uh differently and they work together, and there's uh there's a stabilization in the learning with these two types of memories. So online learning is key to get the stabilization right. Yeah, the online learning is necessary to learn.

Henrik Göthberg 1:17:49

Yeah. It's necessary for the learn in the batch old way, but for for real-world usefulness.

SPEAKER_02 1:17:58

Yeah, yeah, yeah. The different timescales and different applications, so so forth. So everything I say is uh like simplification, uh the first approximation, and so so forth. So yeah.

Anders Arpteg 1:18:09

Okay, but moving perhaps a bit to AGI kind of questions, and um thinking you have made some statements about you think can come rather soon or very soon, even.

SPEAKER_03 1:18:18

Yeah.

Anders Arpteg 1:18:19

But perhaps starting, you know, what's your preferred way of describing what AGI is?

SPEAKER_02 1:18:25

Yeah, so um if we said today's again, simplifications, but today's AI is good at facts, not so good at reasoning. It's good at pattern matching, statistical pirating things, but there's no true reasoning behind. Yes. So uh you can say that is some somehow it's system one, but not system two.

SPEAKER_03 1:18:48

Yes.

SPEAKER_02 1:18:49

So AGI is system two. AGI is the ability to do many-step complex reasoning or planning, planning and reasoning is basically the same thing in a stable, robust way. So if you have that mechanism, you have the foundations of AGI. The rest is training the uh the AGI to perform different types of tasks. So it's the ability to be able to learn any type of task. And like a child, you can't start with a complex task. You need to start with a uh a more um mundane task.

Anders Arpteg 1:19:26

Basic task, yeah. Yeah, and then learning. So you start with something similar.

SPEAKER_02 1:19:30

Exactly, exactly. So you have curriculum learning and you move towards more and more complex complex um environments, tasks, and so so forth. So uh an important com component uh besides the architecture is something Donnie Yillblot called uh machine didactics. You have to put the AI uh to school and have a a well-crafted like uh curricula. Yes. Okay, so. So so so the the the the the the the problem with with answering the question will we have AGI in twenty-four months, sixteen months? Uh it's it's a gradual development, but it's based on one specific architecture that enables this gradual development.

Henrik Göthberg 1:20:16

Yeah, and and because then we can come up to how do we define the cutoff point when we reach AGI and we can use um you know, if we use uh Sam Altman's uh I think this is one of it's it's a simple one to understand. What when the when you literally the computer can be doing what an average person can do, or which means you're new at the job and you need to learn the job, but then you can do it. And then we can talk about doing that in the in the digital space versus doing that in the physical space, and then we can talk about AGI doing this at scale, and then we can talk about AGI regulated, you know, from the lab and and i into in into into everyday lives, right?

Warehouses, 3D Picking, New SKUs

SPEAKER_02 1:20:55

But if you look at intelligence, we had the discussion, how do you measure intelligence? Is a dolphin more uh intelligent than a shark? That's a very interesting question. Uh and I said it has probably something to do with complexity. Uh so uh, and you have a again uh a rat is more uh intelligent than uh a bacteria, a cat more intelligent than a rat, probably, a human hopefully more intelligent than a cat. And then what levels are there beyond? Well, superhuman level one, and then you have superhuman level two, and then you have a skyscraper of intelligent levels. There's no limit to uh how intelligent the system or ever can be. Uh, and if our hypothesis is that once you have this AI arkitektur properly um form and you you teach you teach it so uh the first year is the most challenging corresponding uh knowledge. The second year is is probably easier to if you have properly managed if you can manage the environment and understand physics and obvious so forth, it's easy easier to add like uh more complex behavior, it's just high complexity, more of the same. So so our hypothesis is that the adolescence years of the uh four, three, four, five years uh uh knowledge knowledge of the child, uh those are the most challenging. When you have that in place, the development of the AI or the the uh capability of the AI will accelerate. So we'll you will fast uh reach uh um uh adult level and then even faster superhuman level. So this is our hypothesis.

Henrik Göthberg 1:22:45

Um but but it you know there is a debate here when you want to put uh put the number on something in terms of year, when when will it happen and all that? But uh what I'm sensing now, I want to try it if I if I actually follow you right. So it could be like this, because this is of course a gradual thing. Yeah. And and and you say a gradual thing that is potentially exponentially. Yeah, bottom-up and gradual exponential. Yeah, so it means then that we will not understand, you know, if from a surface observation level, as you said, right, you are solving a task that has been solved since the 80s, yeah, we but in a completely different way, which is narrow and limited, and now in an in in an adolescent way, on its way to superintelligence. So what we so so maybe then a way to uh you know manage a debate on it cannot happen in one year is is like when we look back in 10 years, 20 years, when when did we get on the path of superintelligence? The way I understand your definition is basically well, if the if the architecture is there, yeah, that's the year when that's that's the that's the birth of superintelligence. When will we recognize it as general intelligence? When will we recognize it as, you know, well, when we see it out in the world and we see it performing in the real world, which is maybe then you know, with regulation and everything, that could be five years, ten years in in practical AGI, so to speak. But what you are referring to is your ground zero for when did that AGI, when was that born? And you're saying then that But but you're putting a lot of words in. Yeah, yeah, I'm doing it.

SPEAKER_02 1:24:22

It's great.

Henrik Göthberg 1:24:23

No, but I'm I'm I'm I'm uh reflecting. Yeah, this is my words, and please shoot me down. Uh exactly.

SPEAKER_02 1:24:30

No, no, it's it's a very good description. It's a good way of putting it so uh and and also if the AI learns uh a skill, uh it very fast become uh superhuman.

Anders Arpteg 1:24:42

So so when do you think we will have a practical AGI then potentially?

SPEAKER_02 1:24:48

We will have uh something that can manage a more complex environment, uh well practical. We will have the ability to reason and do complex reasoning uh in 12 months if we get properly funded and we can be amazing if that's it. So you have a good feeling about this now, but it's it's but it's uh the proof is in the pudding of the floor. Yeah, and then I got this question when can it write the PowerPoint? Yeah. The answer to that question is uh I I can't really answer, I can't really set the date. It will be uh on my birthday, 20th of October. And it also if it it's also it's like this you can always showcase if I really wanted to make a PowerPoint, I can I can do the development in that direction and we can have it by tomorrow tomorrow. I'm exaggerating.

Henrik Göthberg 1:25:33

But it's your curricula. Would you put PowerPoint first in your curricula or not? I think a lot of people would love it.

SPEAKER_02 1:25:40

Yeah, actually putting PowerPoint early in the curricula is good. No, the the the the question is more how how how much can you accelerate this development from from a baby AI to an adult AI? Uh and and and we think we think it will go it would be pretty fast, but it also depends on uh other resources like uh data and so so forth. But um what's the bottleneck? Is it compute or is it a curricula? There's uh with the bottleneck, the scarce resource. Um I would say we we wouldn't we don't need that much compute as our as our peers in the business. So compared to them, uh that's not that's not the bottleneck. Um about curriculum, yeah we need training data, but since the system can train by itself, uh it's not it's not could be a bottleneck, but not the same type of bottleneck again as uh as those um building human or robots.

Anders Arpteg 1:26:42

So it is some kind of self-supervised learning, at least. Um if I'm trying to interpret what you just said.

SPEAKER_02 1:26:48

Well you can you can interpret if you want, but I won't help you labeling uh it will.

Anders Arpteg 1:26:53

I love this, it's so fun. I'm trying to okay. I should stop uh interrogating you here, I think. But but we're very curious, that's why we're asking us a question.

Henrik Göthberg 1:27:02

I I appreciate that. Uh yeah, yeah. But you had so many more great questions here, Andesh.

SPEAKER_02 1:27:06

Yeah, okay. Um, but let's Okay. Can I can I ask you, how would you define uh AGI or supertelius?

Anders Arpteg 1:27:15

Well, you know, I think actually the Sam Alton one is rather good, you know, saying simply when we have an AI system that can operate uh at a level that an average human coworker can, then we have AGI. ASI is something else, but AGI at least. So even just you know being able to take person whatever X in your company and actually putting in AI that does what he did. We are not we are not there today for sure. But uh but when we can, um for you know for a large variety of tasks, some tasks we actually can perhaps today. But but for most tasks we cannot. And when we actually can that for most tasks, then we have AGI potential.

Henrik Göthberg 1:27:58

Then it goes from you know doing it in the lab to do it in the real world, right? So because Sam Altman then maybe it's a is is Sam Altman's definition a physical one or a digital one? Or he has layered it even, right? I think it works in both ways, but you know, yeah, but but the complexity potentially to do it in the physical world could be higher than to do it in the digital world.

SPEAKER_02 1:28:20

You need you need to stay in the in the real world, I think. So so so again, we uh we we don't work in a lab. Okay, we made some early work, but we we we are uh in 2000, we we started the company in 2022. In 2023, we had our first pilot customer. So the development has taken place in the messy real world without system integrations and uh stupid layout of the infrastructure which is not optimal.

Henrik Göthberg 1:28:50

Uh so uh and um you said before you took in you you you said before, maybe before the pod that you really approached it with the first principles, and you said engineering mindset. What do you mean with that?

SPEAKER_02 1:29:05

It's this physical real mind. So I stole a quote from Elon Musk. So we're we're more uh what I say, rocket engineers than rocket scientists. Yeah. So it means like we have a problem. Uh we we uh we try also to work with end-to-end systems and build the if you want to build a rocket, you don't start with building a moon rocket, you start with a little firecracker or whatever, and then you build increasingly uh larger um equipment and then you can reach the reaching moon. So we work in that way. So the easiest um robot task picking place, narrow it down, narrow it down, and solve that first, and then more and more complexity uh along the way.

Henrik Göthberg 1:29:50

We have had this conversation for several years. What's the difference between science and engineering? And can you do the research? Is it only science that is research, etc. etc.? And you are clearly in the uh Elon Musk camp here, like engineering, problem-solving rocket engineer, not rocket scientists.

Language Grounding And Teaching Machines

SPEAKER_02 1:30:10

Yeah, so so in in AI it's interesting because the the the uh what the frontier between engineering and and science is is blurred. And a lot of what's uh going on in in research institutions is actually engineering, and also what's going on in companies is is research. Um and I would say that we said that we we don't do research, we don't publish research papers. Uh but what we're doing is probably at the frontier what's considered research. Uh on the other hand, on the other hand, uh there are uh a lot of research going on, which is more probably research than than engineering. Uh, and and ultimately we're using research results from neuroscience, which is not uh and physics, which is not uh not related to computer science in a traditional way. So so so the research part is very important, and we rely on it on other people's research, but we don't do it uh ourselves in that way. So so uh uh because this is a debate if if research is dead or not, I would say definitely not. Research is very, very important. And for Mass Last Theorem, that's that's research, it's it's not engineering, it's just you know fun mathematics.

Anders Arpteg 1:31:23

Do we have a preferred definition of research and engineering by the way? Yeah? Yeah, do you have I have one, yeah. Yeah, please. Um so of course they're overlapping, as you say, but I think the big distinction is really the purpose. So the purpose of engineering is to build some product or create some real value in some sense. You know, building a Spotify app or a robot that actually does work, like a product is the key. So the purpose of doing engineering is to build a product in some way that that works and create value. But the product of science or research is to build knowledge, then I would say. So that means like publishing a paper or or whatever, but it doesn't have to be published. The point is really that you do what's necessary to come up with an answer or some kind of new truth that you want to see. So then if you may have to do engineering to come up with uh new knowledge, meaning you have to perhaps build a prototype or something to see if it does work or not, because then you can answer a question does it work or not? Then you create new knowledge. But if the purpose is really to build a product, then it's more engineering.

SPEAKER_02 1:32:26

That's interesting because uh you you used the description you might need engineering to do research.

Anders Arpteg 1:32:33

Yes, and you and you need research to do engineering, yeah, yeah, it's blurred.

SPEAKER_02 1:32:36

Yeah, yeah, it's blurred definitely.

Anders Arpteg 1:32:37

But but it's it's very clear in the purpose. And I think what Elon is doing, you know, and a lot of tech companies are doing is very, very clearly engineering, meaning they want to build a product, yeah, sure. And then they do what's necessary to make that product work, right? Yeah, yeah. So then it's clearly engineering first mindset, right? Yeah, yeah, yeah. And unfortunately in Europe, and I think a lot of research um uh investments is clearly not, and and uh that's a bit sad.

SPEAKER_02 1:33:00

I think it it is it's actually both because there's a lot of good research coming out of Europe that's that's very valuable and useful for engineers, so uh yeah.

Anders Arpteg 1:33:10

Yes, but if you don't have the engineering, it doesn't really add value, right? So that's a different story, yeah. So if you just have the research and do not have the engineering, then it doesn't really change society in that way, right? No, we need both. And I think you need at least 10x more investments in engineering than you do in the research to actually do come up with.

Henrik Göthberg 1:33:28

But but but on a macro geopolitical scale, this is a very interesting uh topic. Let's take the example of the AI Commission. Very, very brilliant people were highlighting what do we need to make Sweden better with AI. And we pinpointed oh, we need to have very, very strong research. Yes, of course, on that one. But there was one uh quite big missing gap in in relation to that commission, and it it didn't speak about engineering. It spoke about what we needed in terms of education, it's spoken about more money. I mean, like so you could really see who was sitting on the board writing the commission report with academics and senior leaders of big corporations, but the core fundamental engineering angle into that, and here we had what was the guy heading up uh AI science, AI Center in Rice, Sverker. Sverke said it so well when he was on the pod. You maybe know Sverke as well. So Sverke said it beautifully. You know, one of the key magics of Silicon Valley is that they have been schooling the best of the best engineers in computer science through the silicon, through the software industry, and they continue and do new startups and new startups. So they build a real engineering ecosystem. So and if you think of if you if you if you search for that particular angle specifically in the AI Commission, it's it's actually a blind spot.

SPEAKER_02 1:34:54

Yeah, I was uh I had the privilege to be an advisor to a very, very small extent. You see what I mean? Yeah. Well I'm I was invited to a discussion there. Uh and then the the show of the PowerPoint with these are the stakeholders you have in big companies, and then you have uh research and then you have whatever. And uh startups w w w weren't mentioned.

Henrik Göthberg 1:35:17

You saw you you saw you saw the same uh observation.

SPEAKER_02 1:35:20

I I comment on it. That's uh a lot of a lot of the the development of AI is actually done in startups, more than in traditional companies. And this was exactly as I say, the blind spot. Another blind spot in in the AI Commission uh report is that it doesn't mention AGI. It was it maybe in one sentence on one page, uh there was something mentioned. Yeah, it's uh and and I I got the explanation of one of the people involved in writing the report that they would uh the their mission was to provide the uh the establishment with like low-hanging fruits and not to complicate things too much and and find out like a realistic roadmap for um short term, something like that.

Anders Arpteg 1:36:00

So just to not be too negative, yeah. I think the work of the AI Commission was actually surprisingly.

SPEAKER_02 1:36:08

It's a good step forward.

Anders Arpteg 1:36:09

It's a uh yeah, so and also that the government actually later acted on it, not to the extent they wanted with 12 billion, but they actually did put up billions to AI. So I think that's really yeah, a big uh compliment.

Henrik Göthberg 1:36:24

Yeah, so it's not misunderstood. It is a really good step forward, but it was just an interesting segue to how we discuss here engineering as the core.

SPEAKER_02 1:36:32

And at least I was uh they they asked me the question and they asked for my advice.

Anders Arpteg 1:36:36

So so um if we could, since you since I don't want to feel bad about trying to interrogate you more about let's talk about some other systems then and and and connected to the physics and robots and real-world AI and Yellon, and if we take um autonomous cars, autonomous driving, and you know, now Tesla is doing uh the the test beds with the with the robot taxis and everything in Austin and trying to build that out. And he also speaks a bit about you know the difference we we have the Optimus bot, you know, that that is supposed to be like properly autonomous, meaning it's not teleoperated but actually make decisions uh properly. I think you mentioned something about you know please correct, I don't I don't recall exactly, but you said I think you said something about it is so much more difficult to do proper autonomous robots than just having teleoperated like human mimicry um with a robot, right? I mean just having what Boston Dynamics did, which was basically the body of a robot but not the brain, now actually in in um in the CES they actually did come add Gemini to Boston Dynamics and actually did get your brain as well. But but still I hope you see the point here. It is so easy to have a robot that you teleoperate, but that's not really you know what we want and and need. We we need proper autonomous decision making, right?

SPEAKER_03 1:38:08

Yeah, sure, yeah.

Anders Arpteg 1:38:09

And that's much harder, right? Would you agree?

SPEAKER_03 1:38:13

Yeah, sure, yeah.

Anders Arpteg 1:38:14

Yeah, it's kind of obvious perhaps. Yeah, it's obvious.

Henrik Göthberg 1:38:16

But maybe maybe picking it apart, because for you it's obvious. For you it's obvious. Uh you know is it that we I mean like it's like we've solved problems before, but in a in a certain way, and that doesn't scale, or actually hasn't doesn't have the three things together. Yeah, but doing it with proper autonomous decision making, all that is actually about fundamentally solving the three problems smartly. Is that a key?

SPEAKER_02 1:38:41

If you if you have if you have this is the requirements of industry. So if you have that, uh then the industry will trust you and will implement uh the humanoid robots or whatever you like.

Anders Arpteg 1:38:51

But what do you think about the current robo taxi kind of or and autonomous driving approaches? And they all try to you know scale that, and Waymo has their way to build it. China is really progressing super fast here as well. Of course, Elon and Tesla is doing that as well. Do you think their approach will work or do you think they should? I guess you think that they should use your approach, but still, do you think that can scale later on?

AI News: Agents, Co‑Workers, Local Runtimes

SPEAKER_02 1:39:14

I think they they're good doing good work. Uh and if you compare it with pick in place, yeah, it's an easier problem because you're you're just moving the uh moving the car, you're not moving uh it's one object all the time. Um to be fair, it's a much more complex environment than in a factory floor. So but but uh the proof is in the padding, it it's it's uh it appears to to scale good enough.

Anders Arpteg 1:39:39

So you think um Elon in this year will have uh robo taxis in most in most cities, or or do you think that will scale in that way?

SPEAKER_02 1:39:50

Maybe uh you they talk about uh level five autonomous driving. I don't have the exact definition, but uh that that that could be Hard right. But in a fairly structured environment like a city, or where you also can can you can cheat, so to speak, or or have uh engineering solutions on uh other type of data and um GPS data and so so forth.

Anders Arpteg 1:40:14

So but if you take level five, and one definition is simply the no-wheel definition, meaning there is no steering wheel, you can't even take control as a human, right? And and they do have you know, even built these kind of yellow robo taxi things where you can't even, if you wanted to, control it, but you can scale it remotely, right? So I mean, even Waymo that has you know several cities that's operated, they have remote control. So of course it's driving normally uh autonomously, but as soon as some kind of problem occurs, they have uh like uh a call center or whatever you should call it, yeah, that can take over at any time and and you know and drive the car. And I guess that's that's good to have that kind of fallback, right?

SPEAKER_02 1:40:58

Yeah, the system deployed today is with a human in the loop with a type of uh uh what do you say, uh lifeline.

Anders Arpteg 1:41:08

Yeah, yeah. Yeah, okay. But uh it's interesting that uh that you know Elon used the same term as you when it comes to real-world AI, and it says basically to have full self-driving, you basically need uh AGI more or less.

SPEAKER_02 1:41:22

Yeah, to to really solve the problem you need AGI. So you're back to navigation and picking place and uh that uh that thing. So yeah.

Anders Arpteg 1:41:30

What do you think about the benchmark for AGI then? We we have this kind of art arc AGI, you know, one and now uh arc AGI two and and three is coming up with an interaction part of it. Do you are familiar with these benchmarks?

SPEAKER_02 1:41:43

Or have you uh not not by name, but what are the physical uh benchmarks or whatever benchmarks?

Anders Arpteg 1:41:48

Uh so ARC AGI is basically like a visual problem. So you see some kind of uh boxes on like um like a play board and and you should find a pattern and then given three other boards, you should um decide what the fourth one should be and and things like that. So, like a visual kind of intelligence test.

SPEAKER_02 1:42:08

Yeah, so we we're in general not that fond of that type of uh benchmark because uh again, an ICU test, if you put an ICU test in front of a two-year-old, uh he or she will fail. So so uh you you need some other type of uh it's more more related to actual skills if you can solve skills with a certain type of complexity.

Henrik Göthberg 1:42:32

So but this is a deeper, much more interesting question. What is a good way of benchmarking ADI or towards AGI?

SPEAKER_02 1:42:40

Well, it's so as Sam Altman uh if you if you have a if you can do uh a wide range of tasks and you can acquire them with uh uh a reasonable volume of uh training or time or whatever. So uh and how how to measure it? I don't know. There's a it's like how do you know if a person is rich? How do you know if uh um how do you measure measure wealth? Well, you you measure it in in money. Uh but but uh if you look at uh money doesn't capture complexity, it's just a just number. It's just number. And if you look if you look at Earth from space and you compare Earth with Mars or the Moon or whatever, then you can you can say, well, the in some aspect the the the Earth is richer than than Mars. Why? Because there's a level of complexity in the environment that you don't have at Mars. And there are there might be some some sort of uh physical way of measuring the there there are you can find measures for complexity. Yeah, so so so a measurement of complexity in some way could indicate that you have a actual a more intelligent system or something. And these benchmarks they have simply too low complexity. Uh, and uh you you can always if you have a benchmark, you can always find a way to beat it by cheating. Optimize for the benchmark. Yeah, exactly. It's like you you can do in school if you're uh uh adapting to the system. Uh so just what what would the hard question?

Henrik Göthberg 1:44:23

It's a hard question to find a yardstick.

SPEAKER_02 1:44:26

Yeah. So so um and and I think this this uh low-dimensional yard sticks, they don't really apply. You you need to have some sort of complexity measure.

Henrik Göthberg 1:44:36

Uh but it's interesting how that would be so you know, to I think part of uh stabilizing the whole industry and stabilizing things like that is to find yard sticks that we can agree upon is quite useful.

SPEAKER_02 1:44:48

Yeah, it's useful. Uh but it can you it's like uh Siemens, the the the famous uh German physicist says that I think no, no, maybe not. It's for someone else, maybe anyway. Uh you you get what you measure.

Henrik Göthberg 1:45:01

Yeah, exactly. I was gonna go there.

SPEAKER_02 1:45:03

Yeah, yeah. No, Siemens said medicine is the vision. So that's that's another quote.

Anders Arpteg 1:45:09

Uh I'd like to just hear what you think. We're going a bit more philosophical now and thinking about society and what will happen when AGI potentially do exist, and and we will see changes happening throughout society. And and something we're seeing already is that we have, of course, the big hyperscalers. They are the most valuable companies in the world and they use AI and data to scale their business model and are very successful. Now potentially, you know, as AI increases faster and faster, um, we can also see the adoption of AI in normal companies not really following along at all. So even if we have exponential improvements in technology when it comes to AI, the adoption is linear in some sense. Do you see what I mean? Or yeah, yeah, yeah. So I mean, even if you have amazing clawbot or multibot kind of things that you can use, it doesn't really change our society, it doesn't really change our companies that much, or even the people that much. And and the reason for that is basically that the adoption doesn't happen. A few companies are very successful, of course. The hyperscalers are successful in adopting it, but most even potentially according to MIT study, there are 95% is not. So basically they're failing and and and don't get uh return on investment on any AI investment that they're doing. What do you think about this? Do you think that kind of you know divide between the the few that have succeeded and the rest of the world and the rest of the companies will continue to accelerate? Or what do you think will happen as AGI is?

SPEAKER_02 1:46:43

Yeah, sure. If you look at uh the adoption of uh internet, it was like uh things for universities, uh military, and now it's everywhere. It's generally adopted by my mother who uses internet every day. Um cell phones. Everybody has a smartphone, not a phone, but it's something else, it's a terminal smartphone. So there is actually um quite the heavy adoption of uh technology and also this infrastructure uh for new technology, which the iPhone is. And and the the the paradox with with AI uh is like well everybody I know uses ChatGPT, that's AI. And the thing with AI is when the AI gets more and more complicated and more and more uh sophisticated, there's another parallel development that the interface towards the AI becomes more and more intuitive. For example, and this is the the success of Loveball, for example, that you can that you can you can vibe code, you can code without having this uh long uh education as a developer. So so with this type of of development with uh an interface that gets more and more uh easy to use, this will get um the the the deployment of AI or the use of AI with with the will explode.

Anders Arpteg 1:48:10

But still, what we see, if you take Chat TBT as an example, of course we see huge success then with OpenAI, and uh lovable is another example of something that's gone really, really well, of course. But most companies is not on that way. So even if we have a few selected uh frontier AI labs that is extremely successful, but we can count it on one hand, more or less, most other companies fail, at least to gain the value. So the concentration of power in some sense as AI grow in capabilities seems to be just increasing. And I guess that's that's a bad thing. Or would you agree with that? That you know we see the big frontier labs, but it's very few sets of companies, and the rest is just trying to keep it up, but they can't.

World Models vs. JEPA Principles

SPEAKER_02 1:48:54

I I think there will be some sort of democratization of the of the technology, uh, and it will happen when when AGI uh is established because that's a different way of using data, and you don't need that uh uh that heavy uh compute or whatever. You you probably what will happen is that uh everybody will have like this Claude again. You you have your own data, you have something that's that's happening on your laptop.

Anders Arpteg 1:49:26

Uh um and uh so but but an argument against this is some people believe there is a first mover advantage here. So, you know, the insane investments we're seeing now with hundreds of billions of dollars and even trillions, you know, even that's great.

SPEAKER_02 1:49:40

That's great.

Anders Arpteg 1:49:41

I mean, OpenAI has invested$1.4 trillion, yeah, even though having uh revenue of like 20 billion or something. I mean, it's an insane amount of investment.

Henrik Göthberg 1:49:51

Um that's entangled. I think it I think there are two fundamental uh tangents here. On the one hand side, we have um the growing divide of uh power or wealth, uh, and we have then uh the the AI divide between the ones who are who have a much higher invention adoption absorption capacity. Yeah, yeah, there's two two sides, two discussions. Uh so so we have we have on the one hand side that fundamentals, and in is that good or bad, or what happens when the few really cracks the code and actually then not only crack the code, but they have the wits to absorb it into their. So this is one conversation. I think the other uh conversation which you started on, where we have the exponential invention but still have quite linear absorption capacity, adoption capacity, has to do with that we are still not nailing down, understanding what is important to train and shift and and govern and set up. So it's not so much about the tech itself. If I look at the internet, it's the same way, right? So, okay, intuitive UX uh helps us to go to abundance faster. Then it became it became seamless faster. That was the sort of the pivotal moment in the year 2000. Yeah, but then it took us 10, 15, 20 more years into architecturally and structurally and get it into our operations of our big companies, right? And now it's e-commerce way bigger than we ever thought about it. So, isn't that another part of the problem now? That if we want to do be AI ready, so to speak, we need to continuously work on the fundamental patterns, how we organize companies that are basically that are, you know, AI and digital at the core, that moves at a different speed, that looks at uh mandates and decision making and uh agency. I mean, like we haven't solved agency for humans yet. And that will slow us down for you know realized AGI. Yeah, yeah.

SPEAKER_02 1:51:51

This is not a discussion about large companies. And uh a company, there's a big difference between a large uh company's been around for a while and a startup. Yeah. And you can actually look at the company as an organism uh that has a life. With the the startup company, with the baby company is just uh consuming resources without doing any value. And then with the the baby companies grown to uh scale up, then you have a shift in organization and shift in priorities, and you need to add lots of structure uh in the company to make it work and uh make it stabilize the company. And this is like a the company is like a teenager resisting this change. And then you can have an adult company that works well, that's it's cash cow and uh has a stable relation to its market, but then the company starts to to get old, and you there's legacy and there's bureaucracy and there's a hierarchical level. This has happened to Google lately before it's there's been um uh shit. Yeah, so so so um, and it's a challenge with this kind of this old rigid company to to make it more agile uh again. So uh and this is this is this is preventing all types of change. It could be AI, it could be whatever. So this is this is another fundamental problem.

Anders Arpteg 1:53:09

So if we get back to the question of the first mover advantage, and I let me try to phrase it properly here, um, because you said something similar uh or very positive, and I I wish it it really will become true, and that is that AGI will be democratized in some way. But I think a lot of companies do not believe that, and I think a lot of countries do not believe that, and I think they that's the reason we're seeing these extreme investments that we're seeing, like hundreds of billions of dollars being spent, like the size of the Swedish GDP is now being spent by single companies, even so you know what do you think here? You know, because you know one motivation for the first mover advantage is that once a company do have an AI that is smarter than the average human, they can simply scale it up easily, and then they will have uh first or they will have uh access to intelligence and capabilities that no other company or country will have, hence they will have a first mover advantage. And then the rest will just you know continuously be f further and further behind. So you could think like that, right? Or do you think it's wrong? Or what's wrong with that kind of thinking? I hope it's wrong.

SPEAKER_02 1:54:20

Yeah, yeah, it's that's that's one uh that's one scenario. Uh and if uh if you if you can't predict the future, uh and if you want to predict the future, yeah, but that's a classical quote. So and that's what that's uh uh one of the main missions with the the company. So um if we are doing making AGI and making it work, then we can uh uh but we can also um and since our company is supported by the European Union, the the European Investment Bank is considering investing in the company, so there will be like a um uh a joint there will be Europe will be a shareholder of the company. And I think for us that's a good thing.

Anders Arpteg 1:55:04

Um so um and uh but are you afraid about the current kind of investment foreseeing? You know, I think there will be a backlash on this. I don't think they will get money back from$1.4 trillion investments in in a couple of years of depreciation that they only have. So I'm almost hoping for a backlash, but but still some people, you know, they they probably have super smart people that thought a lot about this and do believe that there is a big first mover advantage here when they get AGI.

SPEAKER_02 1:55:35

If you look at these uh uh big five tech companies in the United States, Facebook or whatever, they they have uh I think Facebook had like a trillion dollar uh evaluation, Nvidia also, but Nvidia is different because they have actually uh hardware. But but but Facebook, who could see that coming? And it's it's been that that level of uh valuation for for years. So if Facebook can can keep up with uh with and deliver uh at par with that type of valuation, then there's nothing that says that open air won't do it. Uh and with some if they add like uh ads uh in the in the feed, maybe that's a good idea.

Anders Arpteg 1:56:15

But Jan Le Kun has left Meta, so now it's now with a go to.

System Two Reasoning And AGI Timeline

Henrik Göthberg 1:56:20

But I think the the there is another question here on on that we that you put up, uh Anders, that I think is a nice sort of segue or bigger picture question, and that's sort of then w what do we what do we speculate or the trajectory of the AI ecosystem? I mean, like so because if we think about uh this uh I mean like several superpowers or geopolitical superpowers has defined this first move advantage as as a strategic uh topic, and we are now seeing uh USA, China, and we are now seeing even you know Europe waking up. We have to be first. We're working on it. We want to be first, we have to be first, we need to understand this, and we need to have our own the sovereignty topic or the cultural so many dimensions, right? So, but what is your prediction? Is this gonna grow as one AI ecosystem, or are we actually already fragmenting it in two or three? Because I the latest is almost like it looks like three paths US path, European past, Chinese paths.

SPEAKER_02 1:57:23

I'll quote my grandfather uh in Arabic: it's uh Kulumen Flujudi Yatta Busida Mohtalifeti. And it means that everyone is um uh chasing a prey, but the means of hunting are different. So there are probably many ways to to achieve uh AGI-like technology. And there was this it would not be like one winner takes it all. It would be like a fragmented set of solutions.

Henrik Göthberg 1:57:53

But it is important that Europe has its own view and stake and push in this.

SPEAKER_02 1:57:58

I think it's it's utterly important. So I was invited to to Brussels to the committee. Uh uh, and it was me and it was uh Mistral and the the founder of Legora, and one of the founders of uh uh Silo Silo AI, Peter. Yeah. And uh we got less a few minutes to to give our view about the development in AI. And and my my point is because uh the Europe is competing with the United States and China and so on. I said this stop copying the United States. Stop copying the United States, make it disrupt in AI and take the lead that way.

Henrik Göthberg 1:58:40

Um and this is we have talked about this on this pod. Uh sorry for jumping in. Because also, what is it for us to be competitive and what is it for us to succeed? Yeah, why why why would we ever look at USA or someone else because they have a fundamentally one mon monolithic market, we have something completely different. So all those things uh highlight what's what's our strength? What are we good at? What are we gonna use it for? All these things, if you don't answer those fundamental questions, then what are you chasing?

SPEAKER_02 1:59:10

Yeah, and if you don't innovate, so so my my friends, colleagues, uh they they're just doing more of the same large language models, reinforcement, whatever. And you can still do amazing companies with it, but it's not it's not the disrupt that will uh put Europe on the map and and uh put us in poor position. So that was my message to the to the European.

Henrik Göthberg 1:59:31

And then, of course, what is the disruption and on what uh abstraction level do you need to build it? This is another deep question, like because maybe we can be disruptive, but we don't need to do everything ourselves. We can use other ones' technologies. I mean, like if you have the core technology disrupted at the core, the fundamental architecture, that would be fantastic. But we can also be disruptive without doing every we don't we don't need to copy from the lowest pattern either.

SPEAKER_02 1:59:56

Yeah, yeah, of course, yeah. But but most of all, Europe needs. uh brave VCs. And there are a lot a lot a lot a lot of of uh money available for for investment. There's like 100 trillion euros in in pension funds. If you take like a small small fraction of this and instead of just uh uh feeding a lot of money into these uh um equity funds uh uh take a part of it and make a more diverse in investment in startups and in technology and you will have a fantastic uh growth in in Europe both in terms of money in terms of company in terms of technology and are we seeing this are we seeing this sorry sorry is this the trajectory we're seeing or we're seeing something else are you hopeful that we're going in the right direction here what you're describing as I I fully agree with you by the way it is the thing is if you want it to run in the right direction you have to put an effort into it yourself uh uh and if you don't if you don't do it then you uh you have to set take some responsibility in that yeah and and and even if you do a little little little little part uh you you you uh you you're fine but uh if you're just watching and and complaining uh you you can always find the uh you can always have the the some filter and looking at the the development and think that this is uh this is the wrong direction or we whatever but uh yeah I love that you walk the walk so to speak and just don't do the talk but actually try to be the disruptor so so it's so awesome and I wish I could um dig deeper in what you're doing but yeah hopefully we'll be able to do that later I I look forward to that so yeah yeah I mean I can we uh we you you know that you need to have some beers with some of some of his friends and it's fine at the back door maybe joking but Kareem um to end off here and uh go even more philosophical um we've spoken about AGI a lot and uh you believe it can happen really really quickly potentially which uh would be fun I guess and that would be great fun yeah yeah okay so but then will it be fun or not?

Anders Arpteg 2:02:03

So one way to to think about this is to look at like two extremes. You know one in in the worst case it could be the dystopian kind of night where nightmare where where AI will kill us all and uh we'll have the matrix and the terminator kind of future where where machines try to kill us all. Or it could be the opposite the super super positive utopian future where we have AI that can solve cancer and fix climate crisis and energy and and basically make the cost of goods and services go to Syria and we live in a world of abundance.

SPEAKER_02 2:02:36

Yeah that's the right way to do it.

Anders Arpteg 2:02:38

Where do you think we'll end up we probably won't end up like in in the extremes but where do you think we will end up potentially well if you look at in at the word in general intelligence is not the problem stupidity is the problem.

SPEAKER_02 2:02:52

So that's a short answer to that's a short answer.

Henrik Göthberg 2:02:55

The other the another answer we should be scared of stupidity and we should be scared of what stupid people are doing before we even reach AGI.

SPEAKER_02 2:03:04

Yeah yeah and and the other another way of putting it uh Nobel Prize winners aren't usually the scare killers and the the people um the violence dangerous people in the world so I think there's uh some sort of correlation between intelligence and good ethics.

Henrik Göthberg 2:03:21

Interesting one we've been on to that uh why you know why would uh the AI be what are you benevolent or Benevolent yeah yeah what what what what was why do we think that a more high intelligence being would be evil?

Curriculum, Memory, And Online Learning

SPEAKER_02 2:03:36

Yeah well but then you can argue with the wrong objective functions but for me that's a pass this is that those kind of problems is AI gone wrong that potentially it's not even proper uh ASI or AGI probably not and if we if we circle back to the the this the discussion about how do you measure intelligence uh and I I had um some some sort of idea that it's related to complexity so I I would believe that uh a highly intelligent being or machine or whatever will value complexity in a different way that we do. So for example the jungle the most complex complex system in the world is much more valuable than the the petroleum oil for example in that context uh and uh that will put different types of priorities uh for example the concern of nature will be probably more important for a superintelligence uh than uh the hunt for for for energy or oil or the the the if if energy is important then uh the the superintelligence we will make an effort to to solve a fusion energy well but the only problem is then when when when the superintelligence looks at the complex system and we are used to ants and the ants is the problem yeah I I would say that we significantly that that context add to the complexity of of of this uh Earth's ecosystem so I I don't think we uh the the superintelligence will uh will draw that conclusion actually that we're ants no anyway yeah I I love that you actually do try to be the contributor to the positive future and be the destructor be the one that actually drives perhaps Swedish and European um AI forward and have such a positive view uh that's that's really really inspiring I think so thank you so much Karim for for coming here and uh talking about us super exciting yeah thank you very much I really enjoyed it uh I look forward to this uh uh Paul uh since we spoke again so uh yeah really really interesting discussion and uh really nice to meet you also uh it was really fun so much fun thank you thanks