E173 - Christmas Special - Season 11 Finale

In Episode 173 of the AIAW Podcast, we close out Season 11 with a festive and future-focused Christmas & Year-End Special—and you're invited. Together with surprise guests Patrick Couch, Jesper Fredriksson, and Fredrik Olsson, we look back at the AI rollercoaster that was 2025: from breakthroughs in model development and enterprise AI maturity to the biggest twists in tech and regulation. We also look ahead to 2026, sharing bold predictions, strategic challenges, and the trends that could define the next wave of AI transformation. Equal parts serious and celebratory, this episode is the perfect way to reflect, recharge, and reset for what’s next. Grab something festive and tune in to the grand finale of Season 11. 

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

Chapters

• 0:00 Warm-Up, Meetups & AI Sweden Support
• 0:45 The Chinese Model Debate And Security Fears
• 7:20 Intros: Roles, Backgrounds, And The Stockholm Scene
• 12:40 What Counts As An Agent
• 20:25 Generation vs Action: Tool Use And RPA
• 26:30 GPT-5.2, Benchmarks, And Google’s Edge
• 32:30 Talent, Culture, And Why Labs Win
• 39:10 2026 Outlook: Speed, Inference, And Memory
• 46:30 Continual Learning, Titans, And Research vs Scaling
• 54:20 Reasoning Benchmarks, METER, And ARC-AGI
• 1:02:00 Sora’s Cultural Shock And Trust Online
• 1:09:30 Adoption Lag, Risk, And Enterprise Reality
• 1:16:20 Who Leads In 2026: Google, XAI, Anthropic, Apple
• 1:25:30 Europe’s Play: Adoption Over Frontier Training
• 1:33:40 SLMs, Cost Curves, And Bringing Models Home
• 1:41:30 Infra Arms Race, Data Centers, And Energy
• 1:49:10 Open Source vs Services: Control, IP, Safety
• 1:57:30 The “Make A PowerPoint” Test And Computer Use
• 2:04:30 AGI Scenarios, Governance, And Gentle Futures
• 2:11:40 Closing Quiz, Disagreements, And Final Thoughts

Transcript

Fredrik Olsson: 0:04

Oh yeah. The cloud code. Yeah, I saw that. I saw you on a picture and I said that.

Patrick Couch: 0:10

No, and that is great because that has also meant that uh AI Sweden has sort of stepped up and they now basically host we get to use their offices all the time if if they're free, and and as does Robert Vistine and Kaisa. So we've been able to sort of agree with like informally with AI Sweden. So they are super supportive, which is great, right?

Anders Arpteg: 0:28

So I feel better. I haven't uh really attended any meetup for a long time, but they missed that. It was really fun.

GC: 0:36

It is. And then the meetups are on Thursdays when we have the Yeah, that is true.

Patrick Couch: 0:40

So exactly, and Robert is doing Wednesdays, but uh I had to divide it somehow.

Anders Arpteg: 0:45

But uh but what about this aftermother piece then uh the article about Quen? Did I didn't actually read that article? Did anyone do that? No.

Patrick Couch: 0:54

No, I didn't read the article, I read the the the LinkedIn posts about the article. But that's uh what what I gathered from the article was that apparently Afton Ballet had been criticized and maybe self-uh correcting in the employment service uh because they had been using Quen. Chinese model. The Chinese Quen model, which is an open source model under the Apache something license, and they had used it, I think, on-prem or on their own colocated stuff.

Fredrik Olsson: 1:21

It wasn't necessarily apparently air gapped as well.

Patrick Couch: 1:24

Air gapped and everything. And um at some point there was a moral panic over the fact that Quen was Chinese, so they've sort of pulled the plug and uh and that that that created a lot of conversation on LinkedIn about do they even know what that means? Pulling the plug-up. What do you think? I I I think it's uh somebody said that if you should never what is it, what's the code? You should never ascribe to ill will what you can ascribe to ignorance. And I and I think there's a lot of truth to that. So I don't really think that that that um I was reading in the climate service really had a specific sort of intent which way or the other. They just realized at some point that this is Chinese and Chinese is sort of Chinese and AI and political and acts like anything about Tiananmen Square or something like that.

Jesper Fredriksson: 2:17

I don't think there was any specific reasons.

Patrick Couch: 2:19

And also I think the critique, of course, or the support for not using the Chinese trained model is precisely that, right? Yes, but they haven't excluded politically sensitive stuff. And I'm sure they have, even if they then release it as an open source model with the weights and everything, just freely for you to change. But I think people don't understand, yes, but the model can still serve a whole bunch of great purposes without being able to reference the the Chinese uh political hot potato of the Tinan Square. So what's the problem?

Anders Arpteg: 2:50

You don't see any problem with using Chinese models?

Patrick Couch: 2:52

No, no, I I see a lot of problems, or not but potential problems. And and caution is advice, I think.

Fredrik Olsson: 2:59

But but is that particular to Chinese models or just any model that's not? Exactly.

Patrick Couch: 3:03

Yeah, any model. And I think the what you need to scrutinize is specifically the open source license, like the details. Yeah, what it how open is it and what what kind of transparency does it offer?

Jesper Fredriksson: 3:13

Yeah, and also what is it used for? I mean, there's a lot of uses for Quentri coding. And when coding, I have a hard time seeing that the fact that uh it's censored would make a difference to me.

Fredrik Olsson: 3:26

Yeah, it could be, but it's it stalls backstorp backdoors in the in your code if you don't review it.

Anders Arpteg: 3:31

Exactly.

Fredrik Olsson: 3:31

That's uh far-fetched. If you that's far, but if you don't if you don't review your PRs, you shouldn't ship the code.

Jesper Fredriksson: 3:38

Yes, and I mean that could happen with any model, I guess.

Patrick Couch: 3:41

Yeah, and and exactly, I think there's there's a there's a tremendous prejudice against China in tech. Yes. There's a lot of anti-Chinese sentiment and and the rest of it. And a lot of it it's probably well motivated, given the fact that it's a totalitarian state, and there's a whole bunch of problems with that. But I remember still when I was at Silo, AI a couple of years ago before they were acquired by AMD, they had done a I think of Silo who had done a project together with Huawei for the Norwegian salmon hunting bureau or whatever it's called, because that in Norway there's a problem with invasive salmon of the wrong kind. So they want to sort of be able to select which salmon get to swim up the rivers and and uh get kids. And so they were running these uh visual recognition models on the on the swim by fish, and they had these traps and blah blah blah. And then at some point the Norwegians decided to just pull the plug on that entire little project because the Huawei was Chinese and there may be a leakage of data to China, blah, blah, blah, blah. And I just thought, wow, it's really sort of a bit too much.

Anders Arpteg: 4:48

I mean, yes, perhaps I'm a bit uh biased here, but I actually would not underestimate the creativity that some of the intelligence services is doing here. Have you heard about the recent thing where China is actually blocking Nvidia ships to China? Yeah. No, I haven't heard that. Yeah, you heard?

Fredrik Olsson: 5:04

So Trump um admitted them to sell to China and now China sells tells us that we we don't want them. Yeah. Did you see why?

Anders Arpteg: 5:12

Uh to not to be too dependent on the US. No. No. They were afraid that um US is incorporating some backdoors into uh into China. Let's check our Cisco routers. Yeah. I don't think it's stupid. No, no.

Patrick Couch: 5:30

No, not at all. I don't think any of that paranoia is stupid, really, in tech. I think there's a tremendous amount of things.

Anders Arpteg: 5:34

I think it's more than paranoia as well. I think it's actually a serious concern that should be actually considered.

Patrick Couch: 5:40

Well, like uh Gurt Cobain and Ivana said, just because you're paranoid doesn't move after you, you know, so it can be legit. But I get quote. I get the feeling that in tech now, in geopolitics, we're sort of in this uh uh Mississippi Blues kind of culture where people are either like uh writing songs about uh being the backdoor man and sneaking in and having sex with other men's wives, or they're writing about you know how afraid they are of people sneaking around their house when they're not the way and when they're not there. And it's like the the paranoia is rife in tech. Yeah, but I think it's because the stakes are so high, and I'm sure Anders, given your acquaintance with some of these dimensions of culture and society and technology, but I can attest to that. I mean, it's the style, the stakes are high.

Fredrik Olsson: 6:27

Yeah, but uh also China is one of three named states that that's the Swedish uh security police has actually named uh uh China, Iran, and and Russia that we should you know beware of. And even when I was at Rice, we had this uh course on what this is how you get like uh like to be a tool for a foreign state. Look out for these signs. And we had uh uh a former police handler that had you know handled snitches, we call it in English. Uh this is this is the the way that the long game that they're playing.

Anders Arpteg: 6:58

I was for sure shocked to you know to see some of the creativity that's happened throughout the decades. Uh so I I would be I would be more biased to find non-Chinese alternatives if there are some. I can tell you that for sure.

Jesper Fredriksson: 7:13

Do you think it's possible for an LLM to install a backdoor? Of course.

Anders Arpteg: 7:18

I mean, how would in multiple ways?

Jesper Fredriksson: 7:21

I mean i it's just weights, right? So no, it's not.

Anders Arpteg: 7:25

Okay, so yes, there are safe tensors, etc., that one can use, but um, as I said, the creativity that some people have found in just having a we shouldn't go in this topic.

Jesper Fredriksson: 7:38

Um, I think it's uh it is fascinating to think about if it's actually possible. No, but maybe not for now.

Fredrik Olsson: 7:44

Just don't pick up any USB sticks that you find in your specs on your computer.

Anders Arpteg: 7:48

I mean that's an interesting exactly. Should you would you pick up a USB stick and just put it put it in your computer? No, of course not. Yeah. So there is some paranoia in all of us, right? Yes, yes.

Patrick Couch: 8:00

No, but then I think it's it's probably a good idea to have a bit of it. Yes, I think so. For sure. For sure. Ah, cool.

Anders Arpteg: 8:08

Okay. Anyway, very welcome here. Um we have a panel now in the last AI afterwork. Oh, I see, Goran is um preparing something um dangerous here. Nice. Anyway, um, we have the last AI after work podcast for the year, and I'm very glad to have a set of friends here, and I'm sure it will be a very interesting and deep, and a lot of rabbit holes and a lot of I hope also a lot of um disagreements. And I'm sure it will be, which I really look forward to. But in a friendly way. Is this a rage bait episode? We're always friendly, I think. Anyway, with that, uh let me just uh let's do a quick round of introductions here as well. And uh please, Frederick, if you just could give a few words to to who you are.

Fredrik Olsson: 8:53

Yes, my name is Frederick. I work as a tech lead at all ears. Um just uh do a pitch for audioers now, is that okay? Yes, sir. So all ears, we uh what we're doing is uh listening to the spoken web at scale. So we have a platform where we collect data from uh like audio from from podcasts and TikTok and YouTube and such, and we uh provide that to PR comms and marketing and also to finance and open source intelligence, given what we're talking about right now in different formats. So that's where I'm working right now.

Anders Arpteg: 9:21

But also one of the stars, I think, in Sweden when it comes to NLP and to AI in general, and being in Rise and been in so many other places. So a true uh AI expert, if I may say so myself. Thank you. That's a nice way to say that I'm old. Yes, we I think we all are here.

Patrick Couch: 9:36

So it's fine. Thank you. Patrick, Patrick, who are you? Yes, I'm Patrick Kouch. I'm currently with uh Hewler-Packard Enterprise, uh working here in Sweden. Uh, I've been in AI for a long time. Uh, was at IBM. And I read a hilarious post about how uh a lot of ex-IBMers are touting Watson credentials uh as a as a feather in the hat, whereas it didn't produce any value and you shouldn't do that and the rest of it. Anyway, uh but I also help organize meetups in Stockholm for the AI community. So I'm sort of in the SmackDown. And that's called the Stockholm MNOPS community. But there are a lot of different communities in Stockholm popping up now, so I'm sort of part of that uh socializing culture, and I'm I'm super happy to be here today. And uh this looks amazing. So I'm sure we will have tons of light-hearted disagreements.

Anders Arpteg: 10:28

No, but uh we known each other for a very long time, so very glad to have you here. And thank you, Capture. I'm really looking forward to discussions here. And then we have Jesper as well.

Jesper Fredriksson: 10:37

Hello, uh, I'm Jesper. Uh I'm the CTO for simply.ai, uh, and we're trying to automate uh tasks for people using AI and uh a lot of uh uh just uh strict plumbing kind of things uh to give people time back. Uh so automating uh uh simple things uh that you don't want to do anyway. That's what we're doing. Um that's that's me.

Anders Arpteg: 11:07

And uh a bit about your background, perhaps even in different companies, very famous ones as well.

Jesper Fredriksson: 11:13

Yeah, I've been to many different places. Uh started out in uh uh image recognition, uh medical imaging, uh MRCT, and then I switched into data science, worked at uh startups in Stockholm, ended up in Volvo Cars, and uh then I moved to uh startup again after being in the enterprise.

Anders Arpteg: 11:39

And very much into the Atlantic scene as well, which I think we all are in in some way or form, but I think you also and I are authority in that.

Jesper Fredriksson: 11:47

Amazing space to be in right now uh with the MLOPs and other communities. It's really really fascinating to see Stockholm bloom when it comes to AI. Indeed. It's never been more fun to build.

Patrick Couch: 11:58

Never yes. No, I I I agree. I mean, I always I'm like you said, if you're old, I'm old too. So I to me, this is a resonance of uh of the late 90s when the dot-com boom was booming and I got my start. Uh it's the same kind of wow, amazing. Look what you can do. It's just that today you can do so much more compared to what you could do then. Because then you had to do a lot of the heavy-duty plumbing and the cabling and figuring out how to get the the PlayStations to work as a server farm for your compute and the rest of it. And now it's like, oh, it's so amazing. The times are amazing, technology is amazing.

Anders Arpteg: 12:32

I feel uh, you know, tonight I'm I think I'm I'm in a mode where I will be very contrarian to everything said here. Looking forward to it.

Patrick Couch: 12:40

It's a mind feel already. We're like 10 minutes in.

Anders Arpteg: 12:43

I think you know we can do surprisingly little with agents so far. I think it will change coming year, but I think today what we actually can do with agents is surprisingly little.

Jesper Fredriksson: 12:54

Surprisingly few domains where agent tick has be contrarian.

Fredrik Olsson: 13:00

What do you mean you mean customer-facing stuff, or do you mean actually doing stuff? Because we have increasingly uh at all age starting to work with with agents and cursor and whatnot, and it's amazing.

Anders Arpteg: 13:09

Yes. Okay, so for coding, it is I I would say the the uh the area where where agent works the best. Yes. So that that's certainly the case. But for other parts and even coding, some parts, but for other cases, I think it's surprisingly bad.

Jesper Fredriksson: 13:24

I think it's I think it's probably or maybe I should say, just the fact that it's been so good at coding. So we tended to focus all our interests in coding. I think that's it.

Patrick Couch: 13:35

Yeah, it's an easy win. But so let me ask let me ask you this then. Uh just taking a slight step back for you guys and for me, like what is an agent? Because I think but I think it's interesting because with what's your what's your preferred definition? Oh, I guess I should have one, but I don't really know.

Fredrik Olsson: 13:54

One who acts on your behalf.

Patrick Couch: 13:56

Yeah, I I I used to, I I gave I was a year ago or year and a half ago, I was in this uh trends seminar with AI Sweden, and I was asked, yeah, give us your trends on uh you know agentic, and this was a year and a half ago. And I was thinking about autonomy and agency and and all of that stuff. And I I sort of ended up with this idea of the agent should be somebody like like a the technology that can actually do downstream's decision making on its own based on the upstream conditions it's been given somehow.

Fredrik Olsson: 14:30

Cool. Frederick, do you have a preferred definition? I took an an agent programming course in the 90s. Or for my book, me as well. It was awesome, and we did some games and stuff, and it was belief, desires, and intentions, I believe. Yeah, stuff. API agents, yeah, I remember the framework for that. But I guess I guess now an agent would be something like your mom or dad just doing stuff for you, like picking up your clothes and making sure you don't, you know.

Anders Arpteg: 14:53

But if I phrase it like this, I mean some people call a simpler an LLM without any kind of scaffolding an agent. Do you consider a simple LLM without any scaffolding to still be an agent? No. So some kind of more stuff than an LM is necessary, right? I yes, I would say so.

Fredrik Olsson: 15:11

I haven't I I I'm I'm kind of functional in my approach to things because we build stuff right for customers. So definitions, definitions, let's talk about them later when they work.

Jesper Fredriksson: 15:20

Yes, but do you have a preferred uh yes, and uh I think we started going there. Like the the LLM on its own is not uh an agent, but if you give it tool calling, then I think it it's we're almost there. Tool calling and a loop, then then we're at uh an agent.

Patrick Couch: 15:37

Yeah, so that's interesting to me. So when I'm when I'm out talking about AI and sort of giving these sort of inspirational things, one thing that I often talk about is the uh the failed agents that we try to give physical form. If you remember the humane AI pin, or the with the the the rabbit one or the R1. Right? So the idea with them were that they were supposed to sort of there was they were like a smartphone takeout play, which is of course impossible to execute. I mean the glasses aren't gonna fix that either. But so uh but but what was interesting was that during that time year or two ago when they would when they talked about these things, they talked about the large action models of the LAM. I mean marketing stuff. But I like that. I thought that was an interesting take on it because uh back to your point of execution. If we can execute, if we can call something outside of itself, because if you recall back then like Clippy or the office assistant, it could only operate within this sort of instance of your own local copy of the of the office suite. And it could do different things. But once you started opening it up and then putting these API uh interactions at its disposal, uh the repercussions downstream can be so much greater. And that's what I thought was the kicker. Like if you wanted to do a unnecessary distinction between say generation and execution, generation is like generating an image or or a piece of code or something that that didn't exist before. It's just it generates something, it manifests something. It could be an image, a video, something, something. But then but an action, it's it's an execution of a of an action, that step that it takes, it presses a button, it pushes a light switch or something. And that's more similar to RPA.

Fredrik Olsson: 17:22

Yeah, yeah.

Patrick Couch: 17:22

So I was thinking when you when you get the agents to behave as a complex RPA type robotic process automation type of scheme, then it's becoming very interesting. Yes. And new.

Fredrik Olsson: 17:34

But now we're talking about like functional stuff. But let's talk about behavior instead. What happens when something fails? And I think that's where it gets interesting. I mean, you can have a like a deterministic program or something that's very big, yeah. Uh, and it can you know cover 98% of everything and then and it fails and you just go belly up. A true agent, what would happen then? Would it like recognize its failure and then go back and try to fix it or say I can't do this, I'll call my master, whatever.

Jesper Fredriksson: 18:02

Yeah, I mean the the LLM would would of course see the error and take action depending on the error. So it's always just a tool call calling something outside of itself.

Fredrik Olsson: 18:13

Delegating stuff, yeah.

Jesper Fredriksson: 18:14

Delegating stuff and seeing what comes back and then taking a new action based on that.

Anders Arpteg: 18:19

Yeah, but I like you know, Andrew Ng said something, you know, he got the question, he was really tired about this question, what is an agent? And but he said, I think uh a rather good thing that it's all a matter of degree of autonomy in some sense. So um, if it's something with zero autonomy, meaning it has to do exactly what it's told to do and can't really make a decision, then it's very little amount of agents or agency. That's basically what agency means if you really take it to heart. So then the question is really can you found some kind of lower lower bound? And I really love this kind of short, simple definitions. And I think at least being able to take a one single decision on what they uh action to take, you at least can take that. Yeah, and let me let me give an example. We we built this kind of you know AI afterwork um planner agent. And I would not really I don't feel comfortable calling it an agent because I told it exactly what to do. You you give it a LinkedIn uh profile, uh, it goes out, it scrapes and looks for stuff, then it goes to some other page, and I programmed hard-coded every action it should take. That's a workflow, I would say. Yes. Yeah. And I don't consider that to be a proper agent. No. Right?

Patrick Couch: 19:33

Agree. I agree also, and I think that that's that's great, right? You want you want some level of autonomy, an agency, an execution capability to have it.

Jesper Fredriksson: 19:42

The difference would be if you had told the the uh uh AI after work agent to uh use these tools, like this is the LinkedIn API or whatever, yeah. And this is uh I don't know, some other tools, and then you give it the task. So plan the next.

Fredrik Olsson: 20:03

Yeah, it's like it's like language. I mean, I as a speaker I want to be underspecified and say as little as possible because I'm lazy and you want to hear as much as possible because you want to be, you know, this should be unambiguous. Yeah. And the same thing goes for an agent. I mean, if you can tell it to fix this stuff, and it will maybe ask you stuff and then just do it. So yeah.

Anders Arpteg: 20:23

It's weird we were actually agreed on something here. Okay, please disagree now.

Patrick Couch: 20:29

In an attempt not to agree, then is a decision tree an agent? No.

Anders Arpteg: 20:34

Because it can take actions, right?

Patrick Couch: 20:39

Well, I mean, I guess you don't decide in terms of how how to how the making a prediction of what the next token is, it's an action.

Anders Arpteg: 20:47

I think there is a lower bound to what So the lower bound is interesting, right?

Patrick Couch: 20:50

Because you said if you could if you can take some kind of autonomous action above a certain something, a pay grade, then it's an agent. If it's not, it's basically like a mathematical rules, even if it's an unsupervised decision tree regression kind of thing.

Anders Arpteg: 21:03

But it needs to be outside itself, right?

Patrick Couch: 21:05

Okay, so that's great. So add that then. So then we have an agent being something that needs to touch another system. So it's it's a combination of systems, subsistence. If it's a if it's a solid kind of you know thing of its own.

Anders Arpteg: 21:16

Algorithm in by itself, exactly.

Patrick Couch: 21:17

It doesn't count. No, no, I I can I okay. So we're agreeing in.

Fredrik Olsson: 21:20

Well now you're mixing now. You're mixing problem statements and solutions, right? So I I just I'm just interested in in telling someone to do stuff for me, and it will know me and understand and just do it. Uh that's the like functional outcome from it. And now you're talking about how to do it.

Patrick Couch: 21:35

But so I agree that I don't really think the definitions are that interesting. I remember I asked Gemini or somebody, like, what is this? And he gave me the answer, and then I showed it to my audience, like, what is this a definition of, like the Jeopardy thing? And it said, like, oh, it's something that can uh decide something, blah blah blah on its own, blah blah blah, something, something. But what it was was not an agent, it was a macro Excel. So you can have it define it functionally sort of similarly, and it doesn't really matter if it's an um if it's a macro, if it's a this or the that, if it can do stuff for you, boom, let's go do some of that.

Jesper Fredriksson: 22:12

I mean it it all depends on the task. If the task is simple uh uh at the outset, then a decision tree or a workflow performs the task equivalently good as an agent. But if it's something where the problem is space is much more open, then you can have an agent generalized to basically any anything. So that's that that is of course the the strength of an agent.

Patrick Couch: 22:36

Yeah. And I think that is really where the in the really interesting stuff up ahead, I think, is is is located. I remember when when OpenAI when they dropped their um uh their their GPT uh store thing. Oh the in this instant checkout. Yes. And and um around that time they also released that first um API assistant call or whatever. It was uh the first one, the first one, the very first one, like two years ago or something. And and the idea was of course to start piecing together these multi systems of systems type of uh frameworks, similar to uh what Microsoft was doing under the name I forget now. But anyway, that is interesting to me. Like if you digitize the world, you create the digital landscape, then you start opening it up through APIs and different shared protocols, all of a sudden you are creating the foundation for something very interesting to emerge if you pursue it that way. Because then on that digital landscape, you can move around in a much more frictionless way than you do in matter world.

Anders Arpteg: 23:40

It's like a game of life if you heard about this kind of mathematical concept. It's uh surprising how intelligent behavior something can arise from very, very simplistic rules sometimes. It's cool.

Fredrik Olsson: 23:52

So just one question is uh like physical agents. What if you get into your your car of choice and say, uh drive me to the French Riviere? And it does, and you wake up there. Would that be an agent or something else?

Patrick Couch: 24:04

Certainly for me. For me, it would be an agent, yes. It would be a vacation, yes. It would be a dream come true, too.

Anders Arpteg: 24:12

It's only one tool, but that's enough, I think. I I potentially would disagree that we can define an agent by the outcome. I think we need to actually look internally.

Fredrik Olsson: 24:20

Yeah, yeah, sure, sure. But I mean, uh we we lose ourselves in the internals when we when we should really look at the functions. Now I'm really biased by working at the product company, but anyway.

Anders Arpteg: 24:29

The the idea of this podcast is really to uh actually do a bit of a year-in review for 2025, and perhaps also some outlook of 26. We'll see. Uh depends. Um, but I'd like to do something else. No. But if we just continue a bit uh and then we move more into the highlights of 25. But we just heard uh GPT 5.2 was released, and of course, all they spoke about was the agentic abilities and its abilities to do coding and actually do uh instruction following much better than anything else. And surprisingly, for once, it actually did beat some kind of benchmarks as well, which previous 5.1 and 5.0 did not. Um what's your thinking here? So perhaps if we take a bit of a more broad look at OpenAI and GPT, actually, for once breaking some uh state-of-the-art records in GPT 5.2. What do you think about the 5.2 release?

Fredrik Olsson: 25:28

Yeah, well, um, that deadlines work. Have a code red and you have to improve stuff and you do it on on on benchmarks.

Anders Arpteg: 25:36

Yeah, yeah. Yeah. Yeah. But the code red was it two weeks ago or three weeks ago? Yeah. But I mean, they have worked with 5.2 for longer. Yeah.

Jesper Fredriksson: 25:49

Well, uh let's take a contrarian view then, uh, since that's the theme. Is it really that good? I mean, it it feels like the the story of uh 2025 is a little bit like Google won.

Patrick Couch: 26:01

Like uh agreed, agreed.

Jesper Fredriksson: 26:03

I mean, they they were fighting very hard to beat Gemini 3, and did they? Um I don't think so. I I think uh in my book it's like for general intelligence, I think Gemini is is king of the hill. Yes. I think for coding it's uh open anthropic. That can be debated. Uh swim bench pro as a business. Yeah. I mean from from my personal experience, that's that is the best one. But uh OpenAI is in this awkward middle space. It's like what is their speciality really? So I find myself doubting more OpenAI after 5.2 than I did before. Because they because they had this code red, and this was supposed to be their comeback. But was it a comeback? I mean, it's like, yeah, it's good, but it's but are we like typical users?

Fredrik Olsson: 26:54

I guess I guess OpenAI is uh targeting like users that are chatting. Yes, yes, and we're not chatting, at least not my me as much. So I do coding mostly and I switch between Opus and GPT-52. And I I found it to be inferior because I didn't like the output form.

Anders Arpteg: 27:12

I mean, for me, I've been um nagging on OpenAI for all of 2025, and and uh I think as you said, you know, Google has really taken the lead here most of the year. And this was actually one of the first times it actually at least actually surpassed Google in some benchmarks, not in LM Arena. 5.2 is surprisingly low there. It's kind of interesting. But at least, you know, I I was starting to count out OpenAI completely, and now at least it it cashed up and and was slightly ahead in some benchmarks.

Fredrik Olsson: 27:42

So for me it was actually a surprising and it's it's beneficial for us because it contributes to the dynamics in the field, right? So we we wreck the benefits of of lower pricing, much lower pricing, and actually more for sure.

Patrick Couch: 27:54

So, what would you say is the actual decisive success factor when it comes to succeeding with these releases? Is it is it is it the right people that yeah, they have that dude, and that dude is amazing, or is it they have the more infrastructure compute power, so screw you guys, or what is it? What is the secret source that makes or not this a secret source maybe, but what is it that makes certain initiatives succeed whereas others don't? Because it seems to today, I mean, the the attention is all you need type of papers have been around for a decade almost now. So everybody knows everything about a lot of stuff that that goes into creating success in this space. But some succeed more than others, and for some reason I agree with you on this. OpenAI hasn't been hitting it out the park. And it's not just because Sam has been talking crazy about you know the expectations of people are always let down because it doesn't provide on that promise that he gives.

Jesper Fredriksson: 28:58

But but even so, even so, it's my contrarian point is that I think that they have delivered very good up until this year. I mean, last year at this time, I was really shocked when they released O3. That was that was something really big. And O1 was also like that was the first reasoning model that started something new, so that's they invented a new new-ish paradigm.

Anders Arpteg: 29:22

I mean, it's but then the the big people left. Basically at that time, you know, the the big people, including Ilya and others, and Mira Murati and so many more is they actually left OpenAI. Yeah, and I think actually that does matter a lot.

Jesper Fredriksson: 29:34

I think so. They uh I mean keeping keeping keeping the good people and having them uh having them have a safe space to innovate. I think that's that's probably key.

Patrick Couch: 29:45

So that is super interesting to me because what that says is if you imagine if you think of the the sheer economics of this this thing that is happening in terms of AI reshaping everything. And then we're saying, yeah, and they lost, and we name one dude with no hair, name one woman, and then we sort of say, yeah, maybe some of the teen. It's like what is it, 10 people? Yeah, 20 people, and all of a sudden, I mean, not only do five people own everything, yeah, then 10, 15 people that are like, yeah, if they're not on board, we're gonna lose their race.

Anders Arpteg: 30:23

I mean, it seems super meta is paying 100 million dollars assigning bonuses, right? Because it it it is sufficient sometimes with really, really high-stake people if you want to be a truly frontier lab, right? I don't think most companies should even try to be that, but if you want to, but meta is the case in point against.

Jesper Fredriksson: 30:41

They uh it seems like they are not yet. It takes a year or two yeah.

Fredrik Olsson: 30:45

Yeah, I mean, uh that's uh that's a point. I mean, for for for the people leaving, I mean, that's uh one thing is the actual handiwork that they do, but they're also psychological value of we are losing some some top top brass here. And the other thing is when they are leaving the company and starting something new, it will take a couple of years for them to to sort of start to produce.

Patrick Couch: 31:04

Yeah, so but so my thinking in this is if this is indeed so, if we agree that yeah, it's because the 10 best people are just you know 80% of the secret sauce. So if you don't have them on board, you're screwed. What is the takeaway from that?

Jesper Fredriksson: 31:20

But our takeaway is yeah, that's that's not my uh interpretation of it. I think it's I think it's more that is important, but having the a safe space for a team to innovate, I think that's probably the most important thing. Okay. I mean, that's based on my own experience.

Patrick Couch: 31:35

And I would I guess I would tend to agree, but just sort of the point that I wanted to try to make is the the lesson from that is okay, you only need to train 10, 15 people, make them really good, and you're gonna be on par with those guys, which means then that you only have to train how do you even train people to be top AI people? I'm thinking can well so we have almost 10 billion people on the planet, half of them are online, and you do the funnel and thing, and I'm sure you can eventually find 20 cool people in Europe that you can give some money and just go do what they do.

Anders Arpteg: 32:04

If this is the case, I'm not saying it is the case, I'm just saying get acquired very quickly by the top uh tech giants. They are really good at this.

Fredrik Olsson: 32:11

Let's be a bit contrarian now. Because I guess yes, you can do that if you if you have it like a uh let's say unique unitec economics for it, but you need uh the team is not just the people, you need all the things that have to work and I mean in in the in the special space and the right time and everything. And how do you produce that environment? I mean, for sure now in Stockholm it feels like there's something busting. Maybe it's like a you know um Pygmalion effect. We say it is and therefore it is. Yeah, uh, but still, I mean it gives a positive effect. And if people are leaving, it's the same thing. So just uh grabbing the the brightest kids of the school and then uh send them to Mars for training for two for four years and come back and do AGI, it's not gonna work.

Patrick Couch: 32:54

No, and and maybe not. I just find it interesting that the on the one hand, the numbers are so few, we always revert back to the same few people, and then the the financial implications or the impact on the global scale is massive. And then yeah, and yeah, exactly. You know, that's a good that's a great analogy, actually. Like in the 80s when you had the action heroes, you had Stallone, you had Arnold, they carried the movies, they carried the box office. That was it. You had them on board, you knew it, you knew you were gonna be. You didn't even have a good script.

Anders Arpteg: 33:24

I mean, we shouldn't underestimate the uh impact of AI stars like movie stars, right? I think it's actually there is a few set of people that actually can make these really, really innovative changes. And it's very, very rare.

Fredrik Olsson: 33:36

And what they are good at, most of them, or many of them, is actually educating the rest of us in terms of you know holding the flag high and and talking about what's possible and what's not possible. And uh now this is a secret, and I raised a lot of money and make this like the uh you know AI person's dream. Yeah, now we're agreeing again.

Anders Arpteg: 33:54

Let's do something else. Let's let's try to disagree. Yeah. Okay, but uh interesting. And I thought you know DPT 5.2, it was an interesting one, but just thinking in general about 25 now. Uh of course we can see some kind of open AI trend. I was you know seeing a downward, perhaps we have a slight upward, but I think in 26 we will continue to see a decrease for open AI uh in general. Will you agree? 100%.

Patrick Couch: 34:21

I think it's exceedingly difficult to wrangle power out of the hands of companies like Apple, Google. I mean, in tech time, they've been around for a while. They're savvy, they're vicious, they're experienced, they have every connection possible.

Fredrik Olsson: 34:40

They have a lot of depth, a lot of legacy, they need to move.

Patrick Couch: 34:43

They've been around the block. Yeah. And an upstart like OpenAI, they can do certain things. But eventually I think it's a very slow-moving field. I think the reason we're the fact that we only have like four or five or six giant tech companies ruling everything, and everybody understands that's not what we would like. Like in terms of a democracy or or a non-oligarchic kind of approach. Concentration of power is scary. And everybody, everybody's like, that's uncool, but it's very difficult to change. Yeah. So therefore, I think, yes, I think there will be dents in that, of course. And every now and then there will be like some kind of establishment of a new power to some extent. But really, if you look at the past 20, 30 years, there has been there has been this consolidation around certain gravitational financial centers and technology, and they are super difficult to bud.

Fredrik Olsson: 35:35

But if you think of open AI as a catalyst, I mean innovators dilemma and all that. What was Google uh a year ago? It was like in crisis, we we can't do this. And then I mean they had someone to to play catch up, catch up with.

Anders Arpteg: 35:48

Yeah.

Fredrik Olsson: 35:48

And now it seemed like they're yes, beating them.

Anders Arpteg: 35:51

But they did, yeah. And you know, sometimes they were perhaps a bit too careful about releasing things. Now they were forced to release bars and stuff. Yeah, yeah.

Patrick Couch: 36:01

But but so that is super interesting to me. That angle, like they were careful to release dangerous stuff or stuff they didn't control or understand or something, and then somebody else did, and there was a tremendous uh, you know, hoopla. And then these reports out of McKinsey and MIT come and say, hey, 80% of return on investment, zero. No, 95% of all Gen AI stuff crashes. And I'm thinking, Google and the rest of them, they're pretty savvy. Like they understand, you know, there will be some, you know, teething, excesses, some challenges, there will be some problems, there will be some short-term gains, there will be some fluctuation in the market, there will be like uh a deep seek moment for Nvidia, and the stock will crash, and then they will resurge buying 5% of Intel. And Google are just like, yeah, we're just gonna set this turbulent times how we've been remembering the 90s when we came along, blah, blah, blah. And they are just super cool about this. And eventually you start hearing about, yeah, Apple played it right, they didn't get into that benchmark war, and blah, blah, blah. And you're seeing the arc, the narrative in media change.

Fredrik Olsson: 37:01

But they have another or different revenue schemes that OpenAI don't have. I mean, that's what they're that's a very good point.

Jesper Fredriksson: 37:06

I think uh my take on uh Google's comeback is they've always been focusing on many different things. There uh uh if you take the comparison to OpenAI, they've always been into more or less only language models. They've done some attempts at multimodal, and they have Sora and all of those things. They tried robotics, you know, in the beginning. They tried robotics, I mean, but that's but now again, I guess Google has done a lot of different things, and I think that's what we're seeing now. That the this is starting to pay off because what is Gemini good at? It's the multimodal things. Uh, it's super good at understanding images, it's super good at producing images, and that's probably because they also have a lot of experience in working with uh this open world kind of thing. So uh SEMA 2 and Gini 3. Uh I'm not sure if they are doing it already. I've seen indications that they are. We talked about this this just before. Like uh uh I wouldn't be surprised if Gemini is controlling uh the worlds, the open worlds that the other uh world-generating models can do, and then there's a feedback loop between these two systems. So that's that's my take on why Google is uh leaping forward now. Uh and my take on OpenAI is I'm still sort of holding my breath for for their comeback. Uh there could be something in the cards. What I'm excited about is uh Sam Altman was very clear uh a couple of months ago when he described the roadmap. It was surprisingly clear that uh the goal is to build an AI research intern released in September 2026. So already then we will see what will they bring to the table. And I think they will be as a follow-up to GPT 5.2, there will be something in January. So that will be super interesting to watch. You think GPT 6 or I think GPT 6 will not be in January, but it will be next year. That's that's that's what I think. And uh uh I think it will be sooner than we think because that's what they want to do. They want to be quicker releases, and the goal, uh the only long-term goal they had, and that was also up on a on a monthly uh prediction level, I think it was March 2028, then they were supposed to build uh a researcher, uh, an AI researcher. And that's that's of course a powerful vision. That's like the re uh recursion thing.

Anders Arpteg: 39:47

But if we were to guess, you know, what do you think GPT six or Gemini 4 or Groc 5 or whatever, what will they it contain? Do you have any if you were to guess, what do you think that? New inventions in the next major AI frontier models will be we're talking near term 2026 release.

Patrick Couch: 40:09

Yes.

Anders Arpteg: 40:11

In coming six months or something. Inference speed. Inference speed.

Patrick Couch: 40:16

Inference speed. Yeah. Interesting. So what's your thinking here?

Fredrik Olsson: 40:20

I think that if you're looking at, for instance, cursor, they have their own compose one, composer one model, it's really quick to compare to everything else. And what I do when I code is wait many times. And for everything else, if you look at you, you have this special uh inference chips like rock and so on, cerebrus. I think that's going to matter more and when you roll things out in in the re in real life.

Patrick Couch: 40:42

Interesting, yes. I I okay, so I will immediately agree with you. So because what I'm thinking what I'm thinking is uh the uh just what you say about weight thing. That is interesting to me because and also what you said earlier about even if there's no what was the term, like general progress, if if it if it flattens out, we have work to do to just benefit and reap all the benefits of of the current possibilities in terms of what the technology can do for us. And we haven't done it yet, that's gata in any way, shape, or form. So we don't really need, oh wow, now we can zink da-da-da.

Fredrik Olsson: 41:16

But if you can cut the wait time, yeah, it's like battery, battery capacity on battery capacity, exactly.

Patrick Couch: 41:20

Always on. So um, which makes me think of there's some there are some interesting Swedish uh startups like umbedl with Devd Dubashi, whom I think you know, and um Inceptron with Lucas Ferreira and uh Nima Karima out of Lund. I think they're but both both of these companies are sort of uh Swedish university spin-off type companies. But but they do like kernel optimization stuff. So they they do like okay, so you have this model, you want to run inferencing, you see that model, but that model does a bunch of different stuff that you don't really care for. So let's just get rid of all that stuff and maybe we sort of decrease accuracy. Slightly, slightly. Yeah, there you have it, right? So inceptron amazing stuff. And I think this is deep tech, deep core, and I think there's a lot of efficiency here. There's this is not like, yeah, but first it could only generate images, now it can reason amazing qualitative gain. No, no, no. This is about hitting a real pain. The fact that people that that are running inference saying coding, for instance.

Anders Arpteg: 42:18

Yeah, I think the capabilities are like leapfrogging. Yeah, so inference speed could be one thing. I think so. Any other Jesperonian thoughts?

Jesper Fredriksson: 42:24

Yeah, so I I think uh in my world, in my book, uh, if we focus on inference speed, I would be uh uh dissatisfied. That would that would be uh the bad thing. Not enough for you. That's uh that's to me a sign that uh AI is slowing down if we focus on that. Commoditized, yeah. Yes. Then then it means that okay, so now we've gone the distance, this is all we can do, and we focus on getting this uh better like everyday use. Uh what I hope for is uh number one, uh, for an LLM to say, I don't know. That would that would be awesome if if the LLM could say I can't answer this. Uh and uh I don't know if we will see that in 2026. But uh the next thing uh and I think what what everybody's hinting towards that they want to solve is continual learning. Yeah, because I mean uh when we're sitting in our uh AI coding tool and we turn off the session and then we start again, then basically all memory is lost. There are of course techniques to get around that, and I think, for example, it's things like skills in clawed code where you can uh you leave some sort of um diary behind, sort of this is how we can solve this kind of things. So when you're when you talk about continuous learning, you're thinking about persistent memory and so it's it's a mixed mix of things. Um good question. Uh this is this is more the um the pragmatic approach to it. Like uh uh we have something that's working today, and we we use what we have and we use scaffolding around it to uh to make some sort of uh continual continuous improvements. Of course, the the real thing would be to update weights in in sort of maybe real time, maybe it's something specific to you, maybe it's something for all the models that they will always improve. That would be a major step forward. Uh I don't see that in 2026, but could be.

Anders Arpteg: 44:25

I mean pressure will have these kind of hierarchies where you have a general model that's not updated that often, but then for some company, it can have a more company-specific model that's updated a bit more often. And then for people, they have even more updated often, and pre perhaps even for specific tasks that can have.

Jesper Fredriksson: 44:43

I mean, in a way, you could argue that this is the the direction we're moving because we have much more frequent releases now. I mean, just look at uh GPT 5, GPT 5.1, 5.2, it's been very short intervals between them. So, in a way, that's the sort of a program pragmatic approach to to do this. But that's not what we really want, right? We want something that's that's breaks what I'm doing and uh adjusts for my needs. Yeah. Um, but um, yeah.

Fredrik Olsson: 45:12

And I agree with you on the less psychophantic side.

Patrick Couch: 45:18

But that's the that's their greatest uh attribute uh contribution. No, but uh I was thinking about something else in in terms of where we're going and what will be the things. Remember, uh, did you see that um Ilya interview when he rather eloquently I thought I don't really tend to like the guy, but I thought he I thought he did it really eloquently when he's when he framed it in terms of the age of reason, the age of research, the age of scaling, and now we're going back to the age of research.

Fredrik Olsson: 45:44

Most out the dwarfish. Yes, I think so.

Patrick Couch: 45:46

Yeah. Uh and he was basically making the case for for a long time there was a research-focused effort, and then that gave way to the scaling. After once we hit you know GPT-2 and the Chow GPT thing, and that we'll figure out more parameters better. And so we just went scaling crazy, and now we're seeing the S-curve hitting us, and we're going back to H research. Now, I like that narrative. I I think that's a that's a that's an appealing narrative because what it holds out is a non-linear surprise kind of future, because that's what research does. Yeah, scaling is just more the same, but more better, faster, scooter, you know, that stuff. So to the extent that we will see a more sort of concerted effort to do proper research, if we give these five, ten people that know what they're doing some some time to do stuff, maybe, maybe then we will see something really interesting that would surprise us.

Jesper Fredriksson: 46:40

My take on that is that it's talking out of its own interests. You say I think so, because I mean he's a one-man army, but he wants to he wants to improve in some other way that the other companies got.

Anders Arpteg: 46:52

But speaking about you know, continual learning, I think something that Sam Altman have said, and we also see Meta working in this direction, is the memory functionality. Uh if you remember, I think it was called Titans, the the new model from or paper from Meta, where they have this additional model uh which is not a traditional memory, it's actually a neural network that is besides the normal kind of big model, but it's actually being updated in test time. I'm sorry, I hate the word test time, but in inference time. So it's not actually during train time, but it actually is updating the parameters of the memory by this kind of surprise loss function that they call it, which I hate as well. Um it's just a normal like loss function, but they call it something fancy. But anyway, these kind of additional memory functions, which something that complements the long-term parameters that you have in the big model, I think will be a big thing in 26. Would you agree? No idea, but again, we're talking solutions. What's wrong? What's wrong with that?

Fredrik Olsson: 47:54

Yeah, but what's the problem?

Anders Arpteg: 47:56

The problem, of course, is that we don't have memory, right? And we don't have continued learning, right? And we have to start from scratch for every session.

Jesper Fredriksson: 48:03

I think the the the argument against is that the Titans paper was was it last year?

Anders Arpteg: 48:09

No, no, it's just a couple of weeks ago.

Jesper Fredriksson: 48:11

No, there was no Titans has been around for for a while. I think there was a follow-up paper about Titans. But I haven't seen it, I haven't seen any. You're living at warp speed. Yeah. But but so we can check it, but uh I think I've heard it a long time.

Fredrik Olsson: 48:26

But I but I agree, I'm sorry, I agree, but uh you know, I'm kind of provoking with the solution thing because uh we we we as researchers and technicians tend to uh venture into to the solution space. And what's the problem really is to to provoke you to to what do we want to achieve here? I mean, there's a like trajectory of incremental improvements of stuff because we see stuff in architectures that are current, they're not working. But what if we are on the wrong path entirely? So yeah. And and what's the potential right path then? I don't know. Don't know. I'm too I'm too deep in this.

Patrick Couch: 49:00

But so if can we just take one step back and and briefly touch upon why is memory difficult? Like why don't why don't why don't we have it already? Like I don't I don't have a good it's a sincere question for my part because I don't understand this at all in terms of technology, but why is why haven't we persistent memory across because it seems like a such a of course you want that like you you you don't you don't need the the dancing christmas tree or the rest of it, but but you do need a couple of stuff and then the memory would be great.

Jesper Fredriksson: 49:29

Yeah, I mean I guess the the the simple answer is it's hard to to have your own model and update the parameters uh for just you. Uh because I mean the what you would want is to have Is it is it's compute challenging or it's I mean memory is it would mean we have these super big models. Uh they're I don't know, let's say uh trillion parameters. Uh and uh if you need someone for you, then you need one copy, and then Friedrich needs another copy.

Patrick Couch: 50:02

Of course, you could do tricks, and maybe that's that's what you could do to but in in the age of scaling, then when Jensen is just uh you know tossing hardware around, can't he just start tossing?

Jesper Fredriksson: 50:13

Everybody is complaining that they don't have enough compute.

Anders Arpteg: 50:16

That's uh but it it is a really good question. And there were these kind of old models. I thought it was called something like a differential neural computer or something 10 years back or 2017, 18 or something. It actually had a memory, it had this kind of uh soft memory where you can do read and write operations to that memory and then combine it with the normal parameters. Super cool thing. But it never took off then, it really died out.

GC: 50:41

And I think uh he was right, December 2024, the first one. Oh yeah. Okay, cool. Same sort.

Anders Arpteg: 50:49

Um, and and it's really died out, and it's kind of interesting, and I think unfortunately it is the problem of scale. I mean, if you want to do this kind of memory-augmented networks, which have been around for a long time, I'll say. I mean, even a recurrent neural network from like 80s was actually a memory in it, but it it doesn't really scale. And the big point with self-attention was that it's super, super fast. Yes.

Fredrik Olsson: 51:14

So, but when you say scale, do you mean for a population or for a single user? Could we make one of these things for a single user and show that it works?

Anders Arpteg: 51:21

Uh scale, I mean like training super big models on lots of data, yeah. Then it would be super, super slow to train it on that amount of data, which is the big power of LLMs. Right. And and that's what I I think really removed the power. So I think you know, if we simply can improve the computational power that we do have, I think we can start to see this kind of memory going coming back. And I I think Titans could be a step in that direction, but I don't know. We'll see. And and we won't have rag anymore.

Patrick Couch: 51:53

No, no, Chairman. No, so yeah, so exactly. So back to then what's the point or what's the gain of of succeeding with this?

Anders Arpteg: 52:00

Yeah.

Patrick Couch: 52:01

So there are many. I and I find it interesting because again, with this gaining thing, there seems to be some tendency to look at what we've achieved and and and and to feel, yeah, just max this out. Put it to 11. Push it through put it to 11. You know, just crank that baby, right? And see what happens. And we've been doing it with the with the parameters, and that has driven you know the GPU compute need. But if the memory is then a slightly different ball game or kettle of fish in terms of the hardware that you need to do, it's still similar in the sense that, okay, so create more items and call them the memory items or whatever. Like, and and then and then go and then return to this, oh, but we're burning the plant and we're consuming all the water and all and and all the rest of it. And then we can bring in Robert to Chani to tell us that we only need to eat vegan to have it work anyway. So that's fine. But so uh I find it interesting because the the fun thing when when when people meet technology is because is when people start trolling technology because it's it's it's human, it's like this. Oh, but humans are still better because the technology is not and da-da-da. And one thing that to uh an angle of an angle of approach to attack is always memory. And if if you recall uh early days, uh that robot with a human face, Sophia.

Jesper Fredriksson: 53:23

Oh, yeah.

Patrick Couch: 53:24

Yeah, like she was doing the rounds, and then the funny thing to do was just what did I just tell you? Tell me what I just said, or remember what I said to you, and she was like, What? Oh, the computer's stupid. I mean, it's it's it's sad, but anyway, it shows how critical that persistency of what you were doing, the the recall, the memory is to the perception of you know the value and the stuff that we can do.

Fredrik Olsson: 53:47

I have a question for Anderson because uh you know your stuff about the the research, the research. Do you think we will see like uh an attention is all you need kind of paper for memory soon? Yeah.

Anders Arpteg: 53:59

I I still believe in Jan Likun, I still believe in the JEP architecture. And I think you know the world models that he had in that paper, like three, four years ago now, is what's starting to come about right now. Yeah, and that allows us to do this kind of off-policy kind of training that we're seeing right now. And it also had memory, it had like multiple short-term memory, um, middle term, and then long term was the parameters we see now. So yeah, I think we'll see it. So, and what's the next problem after memory?

Fredrik Olsson: 54:25

Because there will always be a next big problem.

Anders Arpteg: 54:28

Yeah, yeah. So okay. I'm trying to figure out a short way to say this. But I think you know, people are thinking that we have good reasoning in today's models. I think that's wrong. And I think we have surprisingly shallow reasoning today. I think we will see significant improvements in terms of reasoning capabilities that we're seeing. And we know we can see it because we can see it in an Alpha Go and Alpha Zero and these kind of models. They have amazing reasoning, but horrible memory. What we're seeing today is models that is extremely good in memory, in knowledge management. And it's really hard to tell. Is it actually reasoning its way to the answer? Or is it just you know fetching different pieces of memory together and pulling an answer together? And I would say it is really using the amazing capability of the memory today, since it can take like 10 books in a single print and just you know have perfect recall in any part of the book. That's no one human can do. So, and that's good. I mean, it's nice that AI can do something that humans are horrible at, but they are surprisingly bad in reasoning, and that's what we'll see. I think also the action-taking thing will actually come in 26. Everyone is talking now in in GPT 5.2, they're speaking about computer use, proper one, without having hard-coded API or MCP service in place. They can still take action in a way that humans can, which they can today. And I I don't I think that will be a major, major change.

Jesper Fredriksson: 55:52

So it's uh contradictory point, then. What do you what do you think about Arc AGI 2 and and the advances we made? That was uh uh specifically uh designed. I mean, even uh ARC AGI 1 was designed to be like a measure of general intelligence and reasoning. And now we're getting to a point where I don't know what's the score, like 40-50% or something like that on Arc AGI 2.

Anders Arpteg: 56:17

Actually, it's very high. It's 50 plus list.

Jesper Fredriksson: 56:19

Yeah.

Anders Arpteg: 56:20

But still, I think you know ArcAGI 3 is even more interesting. Have you seen the interactive one?

Jesper Fredriksson: 56:25

No, I would say is it released yet?

Anders Arpteg: 56:28

Yeah, I think they mention it.

Jesper Fredriksson: 56:30

So it's more that they talked about it, but I'm not sure if it's released.

Anders Arpteg: 56:34

At least I've heard about it. I think it's released. No, but it will be more interactive. So then it's not simply you see these kind of puzzles and you know decide you know what it is. It's actually you have to take a set of actions.

Jesper Fredriksson: 56:46

I think Francois Cholet has some some i he indicated already when he released uh Arcadia 2 that uh the next one is going to be even better. Yeah.

Anders Arpteg: 56:56

So I think two is a lot about reasoning skills, so that's cool. But action is something else, and I think that's what the version three will actually have.

Jesper Fredriksson: 57:03

So my my contradictory point is I mean, when you're doing AI encoding, I feel like it's reasoning. I I agree that the the way it's reasoning is artificial, it's not how we reason. But then again, do we reason in a good way? I'm not sure.

Anders Arpteg: 57:20

Um, let me let me comment on that because you know, if we take you know playing chess and we take Magnus Carlson, you know, the best human chess player in the world, he has amazing memory. He also has amazing reasoning, and he can actually think much further ahead x number of steps that no or very few other humans can do. But you can't really tell from a human if you just watch the moves that are doing, is he actually reasoning a number of steps back and forth, or is it just memories memorizing similar positions from the past? It's super hard to tell. So I would actually argue that um humans are actually rather he's very good at reasoning. And and AI is really really bad in reasoning. Why do you mean because humans are horrible in memory, and the only way, except unless you are Magnus Carlsen, is that you have to think through steps, and you can't really access knowledge in any way or form similar to what AI can do. So AI is really good in in knowledge management and really bad in reasoning, but they are humans are still beating AI because they can do reasoning.

Jesper Fredriksson: 58:30

But uh, I mean, computers have been beating uh humans at chess for a long time.

Anders Arpteg: 58:36

Yes, so that's a pure reasoning task that doesn't require memory skills. Yeah, yeah. So that's that's a point. So if we can move that, which is exactly what Dennis Sadis says, this this is what they're doing in DeepMind. They're trying to take the reasoning capabilities that we had in Alpha Zero and Alpha Go and put it into the big LLMs that we're having. That's literally what they're trying to do, and they've failed so far. But if they can put that kind of a real advanced reasoning into and combine with the knowledge management skill, that will be amazing. And we haven't seen it yet, but I think it will come in next year.

Jesper Fredriksson: 59:07

I mean, my take is that the the paradigm we have right now with reasoning is that in a way. I mean, it's just a chance of thought, and uh it's not uh it's not super good, but it's I mean it's getting the job done. I think that's that's also one of my takeaways from 2025 that reinforcement learning works for L.

Anders Arpteg: 59:27

But it but it's not anywhere near Alpha Go, I would say, or Alpha Zero's kind of capability in terms of reasoning, and they have to still do the full like you know, going from tokens to token for everyone instead of doing the business-based one.

Jesper Fredriksson: 59:38

There I agree. I mean, we're we're not at that level, but we made great advances in 2025. If you compare it to where we were a year ago, it's totally different.

Anders Arpteg: 59:49

Well, I would argue that coming year, and I think you know, in 26, we will see significant improvements in reasoning and in action taking abilities. And we are surprisingly bad. AI is surprisingly bad today.

Jesper Fredriksson: 59:59

Yeah. I don't think it's bad. I I disagree there, but uh but uh let's just stop there. But but uh just one thing. Uh we talked about the the the meter uh benchmark before uh on this podcast. Uh this uh like long-running tasks. Like how long can an agent uh autonomously keep uh a task running?

Anders Arpteg: 1:00:19

And that's that's I think uh But that wasn't meter really really you know for for tasks that takes X number of hours for humans, what can AI actually do? Isn't that the meter benchmark?

Jesper Fredriksson: 1:00:31

Yes, that's that's the meter benchmark.

Anders Arpteg: 1:00:32

So it's not really being able to run for a long time because that could just you could just put it in a loop, right?

Jesper Fredriksson: 1:00:37

So yes, and how long tasks. That's exactly the point. So so it uh it it it um it measures how complex tasks or how long running tasks it can it it can't fully complex tasks. Yes, and that's the that's the beauty of the benchmark that it's measured in human time. Yeah, I love it. So so for me, when it takes this uh this many hours to code something, can the the LLM do that for me? And uh for how long can it do that? And that's that's always been a super interesting benchmark to follow. And uh so GPT2 famously was at two hours 17 minutes, and just recently uh meter or or epoch AI, which is doing the meter uh uh collaborating, I think, with with meter. Uh sorry, mixed them up. So epoch AI uh produced something that predicted uh meter scores uh for uh the new release models, and there Gemini 3 came out at more than four hours, close to five hours. Uh that's that's a prediction, so there's huge uncertainty in that, and that's released only like uh three months after GPT 5. So we were probably see uh previously seeing a seven month doubling time, and now it seems to be accelerating, which is sort of case in point that reasoning is reasoning is working.

Anders Arpteg: 1:02:03

No, no, it's not not working, I would say it's improving, but it's yes, it's improving.

Jesper Fredriksson: 1:02:08

But not not to the scale that human have if you're saying that uh a task that takes a human five hours to complete, would you still say that that's not reasoning? I mean, uh de facto it is reasoning to me. I mean, if it takes me five hours, then it can reason, it can plan. Yeah, of course it's doing reasoning.

Anders Arpteg: 1:02:30

It's just doing very shallow. Yes, it's shallow reasoning. Yes, right, right.

Jesper Fredriksson: 1:02:34

That I agree with. But it's working. That's the that's the thing.

Anders Arpteg: 1:02:37

Okay, yeah, working. It's working badly, but yeah, but I think we have a lot of opportunity to significantly improve the power of it.

Patrick Couch: 1:02:44

100%. I agree with you. And and I think uh so this is interesting to me what you said about Magnus Carlson, because obviously the dude's a freak and I mean he's amazing. And so it's difficult to tell what's going on inside his head, right, when he's playing. But you can you can look at the outcome of the play and say, dude, the the dude's playing chess like at a top level. Is he reasoning? Is he just recalling frames? Is it just you know generating things on the go? There's no spoon. Yeah, what's happening here? You don't know. But you know for a fact that the dude runs at 100 watts like the rest of us. So whatever he does, that technique is energy efficient. And I think to understand, that the thing that mimics reasoning does so at the very low energy efficiency point. For sure. Because it's not reasoning, it's doing the semblance of reasoning. But at the same time, is there a hard barrier for us to solve this? No, I cannot possibly think so. If you have ones and zeros out there, you know, scramble them around, eventually you figure out reasoning because you know, who knows? So I think 100% that we're in, we're certainly still in for big surprises in terms of what tech can do and digital tech can do for sure. 100%. And I I think any linear thinking is just not operationally valid for a long time. I also don't think that that this uh singularity exponential thinking magical thing is uh either the the rate to go, but I think there are many things going on at the same time, and certain there will be certain breakthroughs in terms of you know what we could do, we could move these ones and zeros around differently. It's called research, and we just figure something out, and all of a sudden, hoop blah, we're doing like 70 on that benchmark instead of 55 or something.

Anders Arpteg: 1:04:25

Yeah. I see the time is fine away, and we haven't even started on Goran's list. We have just been speaking that's what's super interesting. Super interesting. Oops, now we we lost uh sorry. Okay, I lost mine. That's what you have to be done for. Okay, okay. Good. If you don't follow script, you get cut eliminated, perhaps very, very briefly. And and um do you have a favorite model to release uh for um 25? Is it like uh Groc 3 or is it um GPT5? Is it um yeah, Gemini 3, Groc 4.1, yeah Opus 4? Yeah, do you have one?

Fredrik Olsson: 1:05:06

I have one Deep Seek or Opus 4.4.5. Okay, right. Purely for coding.

Anders Arpteg: 1:05:13

Yeah, for example. It's amazingly good, right? Yes, yes. Why do you think that is? Why why is Opus so good in coding?

Fredrik Olsson: 1:05:20

I don't know.

Jesper Fredriksson: 1:05:22

Um next one. This is actually a really interesting question. I've I've been trying to piece that together myself. Why is Anthropic so damn good at coding? Yeah, they've been mostly at the top all the time since at least the 3.5 sonnet. Yeah. Uh and it's been amazing. And it's uh uh I don't know if they have better data, or it's if it's just what we talked about before, that they have some sort of uh space to work with and they have a clearly defined scope. Narrow scope.

Patrick Couch: 1:05:55

Yeah, so is it because they this is what they like to do? Yeah, and they go for enterprise and that's called.

Jesper Fredriksson: 1:05:59

Let's do coding and go for that. Screw all else. No, it's maybe maybe as simple as that.

Anders Arpteg: 1:06:04

They have a but everyone is saying they want to do coding, but no one is doing it. They were first with uh saying that this we're we're doing it. They're working with it longer, perhaps.

Jesper Fredriksson: 1:06:13

Yeah, could be. Could it be I still think that they have some secret sauce? Yes, I think they they have some kind of data set or something like that. They they're collecting data. They're they're also dogfolding uh extremely much and probably have been uh since a long time.

Anders Arpteg: 1:06:29

You one thing I really liked with the anthropic approach for building models is the constitutional AI part, which which means that they basically have another model that is trying to judge how good something else. And uh that's basically a world model, I would say, actually. Right. Uh perhaps they're doing I'm just guessing here. I have no insights about it whatsoever, but perhaps that's actually what they're doing for coding as well. And they have some kind of world model for coding that they can train on to actually generate data you know infinitely if they want to, and probably in high quality by simply having constitutional. I don't know, I'm just guessing here.

Jesper Fredriksson: 1:07:06

I'm thinking maybe uh something that I'm thinking about a lot is like um if you just give uh the model systems, like access to a lot of systems, and saying, okay, uh try to reproduce this system if uh so they can use it, they can interact with it, they can see all the code. So if you're if you're running that all the time and they they have access to all these software systems that they built, or they somehow have access to, maybe that's something that could generate uh this disproportionate amounts of uh good coding. Uh if they started with that before anybody else, maybe that's the key.

Fredrik Olsson: 1:07:47

But again, it's a bit strange. I mean, Microsoft, GitHub, Microsoft, OpenAI, all the data centers, GitHub copilots. I mean, so they should be able to do that.

Jesper Fredriksson: 1:07:57

But the the thing is, uh OpenAI has always been a consumer first. Exactly. They started the the whole chatbot thing, and that's that was their focus for a long time. Whereas Anthropic very quickly went after the coding market.

Anders Arpteg: 1:08:12

Yeah, really smart. Um they have huge revenue streams now from that.

Patrick Couch: 1:08:17

Yes, and I think this is not talked about enough to your point. I think earlier. Uh, you have to look about you look at the sort of financial dynamics in which these various corporate entities operate within in order to understand what the hell is happening. And I think it's very different if you're open AI compared to if you're Google or or or because OpenAI is about showing potential revenue streams, it's about valuation of potential future, you know, money, which is very different from hey, you know, Google just wants to, you know, get the commission in and get the revenue in and you know pay the stakeholders and like and I think these play out very differently. So therefore, you will see consumer-oriented chatbot frequent drops that are like fun and and people get engaged and you show you know the numbers of users and you get valuation of that and blah blah blah, which is different from yeah, so Google wants to break in money. What do you do? Or claw wants to, you know, what do you do? How do you get paying customers? Well, if you can solve a single problem, we go for coding. And we we just do that. I think a lot of these things go into that.

Fredrik Olsson: 1:09:27

Lots of stickiness, high retention.

Patrick Couch: 1:09:29

Yeah, yeah, yeah. So yeah, I think always you have to understand the financial, economic, and the context of these companies simply because of the numbers are so big. So that therefore they have a disproportionate impact on corporate decisions.

Anders Arpteg: 1:09:44

What model will take the lead in 26, you think? Model or model provider? Provider, I guess.

Patrick Couch: 1:09:54

Yeah, I mean, uh, will be uh GPT-6 maybe or Google anthropic uh so you're thinking in terms of user uptake or revenue generated or impact on culture.

Anders Arpteg: 1:10:07

I was thinking benchmarks and yeah, or like L marina kind of thing, or yeah, just something that seems to overtake the other models in terms of performance. Not yeah, not user adoption perhaps, but yeah, you can choose.

Patrick Couch: 1:10:20

Yeah. So that is interesting to me. Why not user adoption? Because I think there's a funny sort of elitist kind of view on these benchmarks that are, yeah, the benchmarks, blah, blah, blah.

Anders Arpteg: 1:10:30

Nico, like, oh, so the benchmarks are great, but then user adoption is just about you know who having the great the best distribution in some sense. You know, you can just integrate it into whatever products Microsoft to Google have, and then they will have the best one. It doesn't mean they have the best model, right?

Patrick Couch: 1:10:42

No, no, no. Well, so it's best generating money, generating joy, uh being used, uh showing something surprising technologically, what is best, right? And I think you can obviously we can spin this many different ways, but I think it's interesting to look at how are people using these various models. Because there seems to be when I look at 2025, there seems to be two kind of groups of of great usage of this. One is the coding dudes, the the builders, the the the the women and the men who sort of create stuff. But then there are these others who are like these more casual consumer users. These are the Sora 2 users, right? And they are winning, they're winning the internet, right? They are Sora 2 to me is the most consequential model of 2025 simply because of its cultural impact and how it has completely changed how people consume and produce YouTube style, TikTok style type media, and and how that impacts revenues, how that impacts the whole thing. I mean, to me, it has been a tremendous, it's like the remember uh uh Sam Altman and OpenAI uh complaining about uh the Gibdi Studio images melting their servers because the users were just creating these uh images that this poor Japanese man was telling everybody not to do, and they were like, yeah, we shouldn't, but look at this. I mean, it's too funny. But anyway, so if you think about Sora 2 and its impact and what that brings in in terms of it, completely shuts down any possibility of validating anything consumed mediated through a digital technology, it's just game over on that. Like Bank ED is holding on by its skin of its teeth nowadays because of Sora 2. I mean, it's just crazy, right? So what do you mean by Bank ED? I didn't follow that. The the Bank ED supposes a level of trust between different stakeholders, and that's how you can transact monetarily between these stakeholders. But if you can't trust anything in terms of internet, if you're not on the on the pre-vetted blockchain type situation, if you're in a situation where you don't know the content produced, you can't, there's no watermarks about uh what's coming your way. Is this QR code actually you know validated? Can I scan it? Can I use bank either? Can I go from generating anything moving in the digital world to something that is financially sound in terms of a constituent of society? I think we're seeing the end of a lot of things that we built to be lasting since the late 90s.

Jesper Fredriksson: 1:13:26

I'm not so worried about I'm not so worried about bank either, but I'm I'm definitely worried about trusting what I see on the internet. Uh I think so, but you have to translate that.

Patrick Couch: 1:13:35

But what's the impact of not trusting it in terms of your comfort level of transacting financially on the internet? It doesn't need to be bank either you pick anybody.

Jesper Fredriksson: 1:13:44

I I got stuck in the technicalities.

Patrick Couch: 1:13:46

You pay something and you pay something on on the on the surmise that it is what it's supposed to be. Like the Greeks say, Yeah, this is the world. Look at how real it is. And you look at the the the David and amazing, or you look at the payeta, and you look at the the marble, and it's like, I can't believe it's marble. It looks like a cloth that is wet. And is it real or what's happening, right? And on the internet, everything's a dream now because Sora 2 just made it so good looking. So, what what happens with the internet after that kind of model? That's why I think it's so interesting. Oh, all the deep fake porns, all the nudity crisis, all the rest, all the C scam, everything. Everything is just completely.

Anders Arpteg: 1:14:24

I'm not worried in that sense. But I think you know, one interesting point with what we're saying is um I have to give an Elon Musk quote here as well. We haven't even spoken about the Elemental yet, so yeah. No, but he I think he said something about you know the future of what apps on the mobile will be in some sense when we have generative AI in terms of videos that is amazingly good, like Sora2. Then you know, why do we really need apps at all? Why not simply the phone becoming the video client in some way and everything is being rendered on the server? And it's just you know you just speak to it, it does what you want, and everything is handled in the cloud or something, and there is no real need for apps or anything. It's it's just some kind of sensor and and uh displays, right?

Jesper Fredriksson: 1:15:07

Uh it's interesting, like all the AI encoding will be gone because we will not need AI, we will not need to code anymore because they it will only be uh um it will only be generated on the fly. Yeah. Yeah, I think that will happen.

Fredrik Olsson: 1:15:22

And then we can all check out just sit in a sun chair with a daiquiri and whatever, and just to have the amazing box run however.

Anders Arpteg: 1:15:30

I'm all for it. Yeah. Yeah, it's kind of an extreme thought.

Jesper Fredriksson: 1:15:34

When will that happen?

Anders Arpteg: 1:15:35

Yeah. Uh not this next year at least. Not next year.

Jesper Fredriksson: 1:15:38

But but I think if you had to make a guess, when would you say?

Anders Arpteg: 1:15:42

Oof. Uh I think 10, 15 years, perhaps. I don't know.

Patrick Couch: 1:15:46

Yeah, but so that's a that's a cop-out. That's like uh it's a cop-out.

Jesper Fredriksson: 1:15:49

10-15 years is like who knows. It's impossible to say, but uh, I mean, you have to assume that it's like we we have reached AGI, ASI, all of those things.

Anders Arpteg: 1:15:59

Okay, but let's let's move into a great question. And this is something that Henrik uh uh Jotberg as well is speaking about all the time, and I think is is interesting, and and that's the the technological progress of AI versus the adoption of AI. So even if we have SORA 2, and even if we have SORA 5, which is more or less like interactive video, you just speak to it, it just interacts with you, it's perfect video, you can't really tell anything from it, and it's amazing. Does it really impact companies? Does it really impact society? Does it really change our way of working that much? I would say no. It it will take time and it will take many, many years, even if the technological progress will you know go really, really fast.

Jesper Fredriksson: 1:16:39

Definitely, and it will be uh it will be distributed very unevenly. Yes, that's even more important. Yes, agree. There will be there will be those companies that run with this, and then there will be the legacy companies that will struggle with this.

Fredrik Olsson: 1:16:52

And I think the adoption will be uh like dictated by the failure modes that we see in the models. So if they make failures that are very severe for business, they will not touch it unless they can actually predict what kind of failures they have.

Jesper Fredriksson: 1:17:03

Yes, and the next thing is of course how risk-averse you are as a company. Yeah, and that will that will dictate how uh how well you are performing at the same time.

Fredrik Olsson: 1:17:11

Yeah, again, like Google, what what kind of legacy and and you know uh stuff do they have to carry around, dug around. For instance, I mean we have bigger companies that are big companies that are not uh AI native, that they will struggle a lot.

Anders Arpteg: 1:17:25

And this is kind of scary to me, I must say, because I think some companies will be much faster in adopting this. And of course, we can already see it today, and we've seen it for the last five, ten years more or less already, that some of the tech giants are becoming insanely wealthy. And if you just look at the most wealthy companies, the top 10 in the world, in any kind of sector, it is the tech giants. It is the one that can scale the business model with data and AI. And you know, if that kind of divide or or that kind of if they adopt AI so quickly in these kind of few set of companies, and the rest of one of the one is is trying to catch up, but it's still always slightly behind, that kind of divide will just continue to increase, not linearly, but actually potentially exponentially or polynomially, and that will be really scary. That will lead to a concentration of power, I think. That is really, really dangerous. What do you think, Patrick? I know you have a lot of questions here.

Patrick Couch: 1:18:19

Well, yes, no, so I agree. I think concentration of power in the hands of the few is a dumb idea, it's not resilient, it's it's vulnerable, it's it's easy to co-opt, it's dangerous, uh, it's prone to error and stupidity, and I don't think we should go there.

Fredrik Olsson: 1:18:34

Uh, there are few already, and they will be there, they will be fewer.

Patrick Couch: 1:18:37

Yeah, and I recall many years ago I read an article about uh how is it that uh how how could these tech companies form so quickly? And the the basic argument was well, the US legislation managing mergers and acquisitions wasn't updated for the tech era in the late 90s. So when they got going and they acquired all these companies, and you saw these uh discernible epicenters of power like the Googles and So what should we do? I mean, because they they still operate you know very well, they just acquire people and companies all the time, and right and so here's I think the sort of the the social cultural uh question we have to ask ourselves, and that is to what extent can we offset or to what extent can we are we comfortable with non-democratic oriented mechanisms to achieve certain efficiency gains?

Anders Arpteg: 1:19:32

What should we do then? I mean, should we remove capitalism?

Patrick Couch: 1:19:36

No, no, no, no. Well, yes, main or hypothesis, of course we should. That's I mean, capitalism is just not working and blah blah blah blah blah. But more concretely, I think diversity is good. And we need we we we don't we shouldn't fight for diversity.

Anders Arpteg: 1:19:54

You can't really be angry with the company because they're successful.

Patrick Couch: 1:19:56

No, exactly. So I'm not. So you you you shouldn't put blame on the companies. Because the the companies, the the profit mechanism or the the the the entity, the corporate entity is set up for a very specific purpose. And and it when it s succeeds in that, you shouldn't shoot it down. I mean, obviously, right? So you have to think about the larger play and what is what kind of society do we want to create? And I think we have to look at uh various mechanisms in terms of technology or or or corporate legislation and see, you know, how do these play out. So if you look at uh should we attack this with regulation, you think, or what should be the uh well, so the way I see it is regulation is good when you can't have people to uh self or um can't take responsibility for their own behavior. And there are many reasons for this. Like you can't you can't uh ask people with shortage of money to buy organic food because it's good for the planet or what have you, right?

Anders Arpteg: 1:20:55

There are certain things that are just we have a lot of law, laws and regulation for for a good reason, right?

Patrick Couch: 1:21:00

Yeah. Yes. But but I think we have to. There is a guy on on LinkedIn that I follow whose name I keep forgetting, but he he posts these interesting things. And he posted this one clip about you know the fact that we need to slow down. We don't need to speed up, we just need to slow down and you sort of you know take stock and you know do your point.

Anders Arpteg: 1:21:19

Do you think that's a good way forward to slow down? We tried that you know a number of times. Well, I'm not hard.

Patrick Couch: 1:21:24

Well, so I'm not saying that we necessarily should slow down, but the question is when we build valuational companies on the release cycle that very quickly gets outdated, so we say, okay, so this benchmark was hit, so therefore we get a lot of valuation, so we get money in and they can do more of this. And then it turns out that yeah, nobody actually uses that thing because the function wasn't then there. So the 95% of the Gen AI pilots just don't go anywhere. And so what's the value of succeeding with this benchmark? Why should that be the case? And should that should we have more quicker? So benchmark.

Anders Arpteg: 1:21:55

What should we do? I mean, we have a big AI race. I mean, we are seeing now that Google was forced to just release stuff much faster than it even wanted to because of OpenAI and others, right? And and then OpenAI have to do it much quicker because of Deep Seek and so many more things. So we are seeing this kind of race leading to releases that is perhaps too fast in some way. But but what's what's the answer? Frederick, please tell us. I don't know.

Fredrik Olsson: 1:22:19

Uh when I when I do, I will form a company around it.

Patrick Couch: 1:22:23

And mostly rich. Here's my beef with the pace. I mean, I'm I'm sort of a forward-leaning guy, and I and I agree with with you know Terrence McKenna and the forward escape and and the and the image of you know, you you go into the Death Star, you toss in the malt of the cocktail, and you get the hell out of there before it explodes, right? But if you don't make it, you're screwed. And that is sort of you know the human race on the planet with technology. We need to figure this out, and then we go to space and you know the rest of it. But I wonder if this narrative of, ah, yeah, but uh look to the west will save us. Because people are so very keen on saying, yeah, but uh But what's the answer then? I'm not sure what the answer is, but I don't think the answer is to keep looking west, which has been the answer for a long time. And looking east means what? I don't know if we need to look east. Maybe we need to look inside. I don't know. But I'm just saying that there has been a tendency for culture to look to the west and look to Los Angeles, look to California. I mean, we're talking, you know, then.

Anders Arpteg: 1:23:23

Can Europe start to play a bigger role here in some way?

Patrick Couch: 1:23:25

Europe has to, has to, Africa has to, and I think Africa is. I think there are some interesting uh things happening in Rwanda and in other places in Africa where you see, you know, tech hubs form on their own terms. I think we need to, you know, foster that. I think we need diversity on a grand scale. I don't think we will see this through with four dudes on the West Coast living in bunkers telling the rest of us how to live. I mean, five people owning what, 90% of technology and that's when I say something stupid that will make you really angry, Fodrick.

Anders Arpteg: 1:23:57

No, please. Yes, but yes, of course. I love go. Maybe Trump will save us. Well, so can you guess what I mean when I'm saying that? I don't know, evolution through crisis? No, but but yeah, partly because Trump is forcing Europe to take action. Yeah, exactly. No, I agree with that. So you should be thankful to Trump. Well, crisis I'm gonna push you here. Sorry for that.

Patrick Couch: 1:24:20

No, no, no. So speaking as an American, no, but uh I agree that that evolution through crisis is a tried and tested method. Like the human race has has been through some genetic bottlenecks. We were down to like, I don't know, 200 people at some point. I mean, we we lived through five ice ages without antibiotics and the rest of it. I mean, the human dudes are we we create as a scramble. And that scramble hasn't been really there yet because we're also prone to being lazy and automated. You know, I forget who said this, but Joe Rogan restated it on the podcast some time ago about how uh hard people create soft times, soft times create soft people, soft people create hard times, hard times create hard people, and blah, blah, blah, blah, blah. Right. Maybe that is true or not. But it is certainly so that once a well-functioning democracy gets into Falkhammer territory, we sort of ease up on the gas a little. And we we sort of tone down the 68 political discourse, and we go for, you know, the Tesla car and the solar panels, and we sort of chill with the demonstrations on 1st of May and all the rest of it. And with that then comes a slightly different political wind that is not managed. Yes, and I think that is, yeah.

Anders Arpteg: 1:25:35

Yes, but do you have any thoughts about this? How can you avoid, you know, perhaps the tech the concentration of power and uh take progress is moving faster than adoption and easy questions?

Jesper Fredriksson: 1:25:47

Yeah, exactly. Um I mean to take a contrarian approach again, um in my personal life, do I care? You don't? I mean you don't care about concentration of power? Does it affect me? And I mean I think the circle of influence and uh exactly it's like uh uh and and that's also again, I guess, stating the problem. I mean, I don't care, I just want this thing to to move as fast as possible. Uh that that's that's that's sort of my point of view. And and I I I I hear how dumb it sounds, but that's sort of my inner voice. Like I'm loving this. I I want the releases to be faster, I want everything now. Yeah, uh, so I I think that's that's great. Then we have the problem of of uh um we talked about what would happen if let's say uh there's uh somebody uh making um uh decision that uh inference will not go outside of US. Let's say that that would happen. Then I would stand there feeling like a fool because I mean this could happen. And these are the kind of things that I'm worried about. Like what will happen when we reach AGI or whatever? Let's say we reach something super powerful and it's only released in the US, only US companies can use it. That's when it becomes really scary. And do I have a solution for it? No, I don't. I uh and I um I think uh regulation to some degree, but I mean, what can can EU stop the the US from uh doing uh US first? No.

Anders Arpteg: 1:27:29

I mean I have a topic here about the Nordics and in Europe, so so let's try to go there at some point, and perhaps we can you know guide all the policymakers, you know, how we should do it. That could be nice, but I think it's a very interesting question, you know. Okay. Um let's if we were to just speak about you know still the adoption problem a bit more, and then we can perhaps move to the Europe question. But we we can start of course see the big tech giants in the moving super super fast, and and Nvidia is like exploding in value. And uh, I don't think actually NVIDIA will continue to grow, but that's do you agree by the way?

Patrick Couch: 1:28:05

I agree.

Fredrik Olsson: 1:28:07

I agree with you. I think it's they will plateau a bit, they will plateau, of course, they will. Yeah, yeah. And then there will be a paper that's um for memory, and they will have to shift gears, and you know the memory is really, really expensive right now.

Jesper Fredriksson: 1:28:19

Why why will it plateau?

Fredrik Olsson: 1:28:21

Um I don't know. Yeah, I guess I guess the plateau in the sense of their margins will become smaller. Uh they will have uh competition.

Jesper Fredriksson: 1:28:33

You mean AMD will catch up?

Fredrik Olsson: 1:28:34

Uh no, I think the Chinese will. I mean, there's all I mean, you probably read the Chip War book, but I I I think they will catch up sooner than we think. And then we're something different.

Patrick Couch: 1:28:45

So this is super interesting, I think, because uh I mean. So Nvidia has been tremendously successful once they figure out how to how to leverage GPUs for specific workloads, and those workloads were the sought-after workloads.

Anders Arpteg: 1:29:03

And I think for for people that may not understand it, Nvidia is the world's most wealthy company in the world right now. Yeah, bigger than Google and Microsoft.

Jesper Fredriksson: 1:29:11

Are they still bigger than Google? I mean, go Google increased a lot recently. Okay, but they're big.

Patrick Couch: 1:29:16

But anyway, I mean that I think the point is valid, right?

Anders Arpteg: 1:29:19

So they were up in five trillion dollars valuation at some point.

Patrick Couch: 1:29:24

And and and so so the the narrative has really been okay, build it and they will come.

unknown: 1:29:29

Yeah.

Patrick Couch: 1:29:30

Like field of green Kevin Costner territory. Like just get the GPUs out, increase the receipt uh release cycles, you know, make it available, make sure that companies like Hewlett Parker Enterprise and the rest of us create servers for these uh GPUs and have all the other equipment, infrastructure pieces be there so there are no bottlenecks and just you churn it out and sort out the rare earth metals and go, go, go. And that has been going on for some time. And you know, there's been these talks about, yeah, NVDS courting uh UK government to get uh open air, blah, blah, blah. At the same time, you're reading these stats about uh the overestablishment of data centers and how how difficult it is for certain to have high utilization and how few customers are actually using them. And when you sort of scratch the surface, you realize that yeah, so some of these, you know, uh heavy-duty uh foundational model type of companies are certainly churning up a lot of infrastructure need. But the rest of the adoption, the enterprises, the mid-sized companies, you know, what are we what are we doing with this? Well, not much. So people are like, okay, so on the one hand, we're being FOMO's into investing into infrastructure for AI compute, and it's being burnt by a very few, and then a tremendous amount of people don't really see the use case. Like, what are we gonna do with this GPU that IT bought? And in between comes all these reports about you know the discrepancy about you know the possibilities and the tech company's valuation. And imagine how wealthy you could be if you're a tech company and you're like, yeah, but we do plumbing, and what are we gonna do with the data center? So I think all of this will have to be balanced out over 2026. I think 2025 has been a very wild ride. I'm sure if you're in finance, not in tech, but in finance, and you're seeing these hundred billion dollar deals flying left, right, center, you see these uh charts about the interconnectivity, about you know, okay.

Anders Arpteg: 1:31:31

So trying to close this topic then, you know, we we spoke about the tech progress, you know, we were seeing the frontier companies, you know, tech giants really flying off and the concentration of power, and the rest of the companies are lacking behind with adoption. But if we if we ask a question, you know, will who which company will be the big winner in 26? Not Nvidia, probably, not probably OpenAI, I would guess. I think we all were. Are we still agreeing on that? That NVIDIA and OpenAI will not be the yeah. I agree with that.

Jesper Fredriksson: 1:32:01

I I don't think they will. I I I mean, just looking at uh the the scale out uh plans for all companies, I can't see Nvidia losing the next year unless there is somebody catching up already now.

Anders Arpteg: 1:32:14

They have order books for many, many years ahead. So they they will of course make a lot of money independently of what happened, you know, they will still make money for many years to come. Yes. I think we can be sure about that. But but still, will they be the big winner?

Fredrik Olsson: 1:32:27

Um like in terms of the delta most increase. I I I would actually guess that Apple would be that. I think they will nail the consumer side interest and finally make that right. Because they have the distribution and they have the the gadget. I'm not an Apple guy in terms of functionality.

Anders Arpteg: 1:32:44

I think that they because they don't waste money on building the frontier model system.

Fredrik Olsson: 1:32:48

I think I'm not sure if uh and like now they have a crisis with management leaving, uh, I hear, but uh, I think uh in terms of kind of using things that other other people have built already and they have the consumers for it, and they're kind of how do we get this right?

Patrick Couch: 1:33:03

100%. And when Apple makes a move, it's gonna be a well-founded, commercially sound move targeting end consumers. And what end consumers mean is real fresh money in. Lots of data in these end consumers have wallets. Now, the bots on the internet, they don't have wallets. I know that, yeah, but the financial traders, they have blah, blah, blah, blah, blah. Yes, but I'm not talking about that. I'm talking about exactly what you're saying. Once we see Apple move into this space, it's gonna signal a real value in a use case.

Jesper Fredriksson: 1:33:36

That is what do you think that what do you think that value will be? That's that's an interesting question.

Patrick Couch: 1:33:41

It could be anything, right? It could be you know an AI feature in in Roblox or something that is just taking off, and then Apple monetizes that in their App Store or something. I don't know. But I totally see that as the signal for when when the core is correct. Because now the the most valued companies are the sort of the tech providers, the enablers of this amazing thing, and we are evaluating them based on proficiency and benchmarks, but that is not adoption or use. And Apple is all about that. So I think you look towards Apple, you look towards, you know, uh, what's the older guy's name, Berkshire, Hathaway guy? Buffett. Buffett, right? You look at that kind of investment, the end the long game.

Anders Arpteg: 1:34:21

Yeah, okay. I mean, good interesting, and I think that comes back to the Nordic and Europe. So so keep that in mind, you know, what Apple is trying to play at. But if we just keep at the tech giants for one second, uh I I have strong beliefs, if we call it that, not hopes, but beliefs, uh, in in XAI. I think they will take the lead in many ways and ways and forms in 26. Anyone that agrees or disagrees?

Jesper Fredriksson: 1:34:47

I disagree. Um well, I don't I'm only thinking about the LLM kind of things. Um and the rest of it I can't speak for. I I would say that uh the most likely thing is that Google will keep on winning. That's that's my guess. They have the scale, they have the distribution, they have the the knowledge inside of they have they have all this diversity in sort of approaches. So I think they will probably continue to lead the race.

Anders Arpteg: 1:35:18

Uh do you think uh yes uh sorry, Frederick will be the big winner in Endo 26 in terms of frontier AI providers?

Fredrik Olsson: 1:35:28

It's interesting that you mention XAI. I have not got a hold of them really. I don't use X for much. I haven't really tried ROC, maybe I should. Um I still I still kind of lean for anthropic and the thing that they address with enterprise and and and coding still. Um winning in that sense, I'm not sure what that means, but yeah, but if it's it's for attention, I I I I I'm I'm inclined to agree with XAI because I guess that the the midterms is going to take place, right? Uh in the US, and there's going to be like a spectacle around that. And then I guess that those two old dudes are still pretty close in terms of what they believe. Patrick, do you have any?

Patrick Couch: 1:36:13

I don't think XAI is gonna win. I think uh they will win in the US, and if the US wins, yes, then sort of indirectly they will sort of but I think the user adoption of X will just peter out because it's just so dumb at points at certain places. It's just too dumb. I mean it's generally okay, and then it's really dumb. So I think that will be a problem to succeed globally. But I think Google will do well. I think the sort of incumbents will continue to do well. I think it's very difficult to upset that. I don't think uh I think OpenAI will have a hard 2026. I think they will shuffle massively. Yeah, I think NVIDIA will work the order books and then have to start to to pivot towards some other infrastructure technology like memory-oriented stuff or something.

Anders Arpteg: 1:37:04

Yeah, we'll see if Sam will still be the uh CEO of the interesting, right? Yeah, you as well.

Fredrik Olsson: 1:37:11

And we're we're constantly looking to the West. What about Alibaba? Alibaba, I mean when and stuff like that. Chinese models. I I know too little about them to have an opinion, but I think that's where it's going to be like an increase a lot.

unknown: 1:37:26

Yeah.

Anders Arpteg: 1:37:27

I think still the the they are really good in engineering, but perhaps not so much in innovation. But but who knows? They have a large like customer base. Yeah, but I think yeah. But but they will not be the big innovator, I don't think. But that's yeah, who knows?

Jesper Fredriksson: 1:37:41

I I think I agree. Just the fact I've heard, I haven't confirmed it, but uh OpenAI is said to have more compute than the whole of China. So I think uh if that is true, then I think uh they will win. Yeah, I mean US will win. That's that that is definitely my my thinking. Then there will be a use case for Quen, definitely. I I definitely see the value in it, and that's more to the to the case of like having more efficient models, maybe faster inference, that kind of thing.

Fredrik Olsson: 1:38:10

But what will happen with Taiwan? Yeah, and so you know don't even go there. If something happens there, the the whole supply chain for the US and the GPUs will be. They will not I don't think US will allow anything to happen. No, but then we still have still have like something that's going on there, there will be interruptions.

Anders Arpteg: 1:38:28

That will be horrific. Yeah. Okay, so time is playing away. And and I'd like to go back a bit about to to the Nordic and Europe, and then perhaps also to the Apple kind of thing, but where they not may not be the top frontier AL, but potentially be really good in adopting AI. What do you think about the Nordic landscape? Um could in the best case scenario, what could we do to really make 2026 a good year in terms of I think we can all agree Sweden, the Nordics, and Europe will not be the frontier like AI provider, right? Yeah, yeah, yeah. Providers. Not current? You never know?

Fredrik Olsson: 1:39:08

Yeah, I I very, very should we? I mean, uh, oh exactly. No, yes, we shouldn't. No, we shouldn't.

Jesper Fredriksson: 1:39:15

I think we should be we should be proud of what we have and that we're we're very good at using the models.

Anders Arpteg: 1:39:20

Yes. So even though I hate Apple, uh sorry to say that, I think you know, potentially their approach could be something for the Nordics in Europe to take on. What was that approach? To be really good in adopting, not building AI. Meaning, right? Uh uh and and and if we get uh really, really good in finding ways to build companies to really quickly adopt the use of AI in a potentially responsible way, I think it could be really positive. Yeah, I think uh and we don't have to waste many money on building. These kind of huge infrastructure data centers, we don't have to waste so much money in training these super big models. We can spend money on actually building real value.

Jesper Fredriksson: 1:40:07

Yes, definitely. I think it's probably not the best position to be as a company to build out AI models. It's a very costly and very difficult business. You have five of them. Yeah, yeah, exactly. That's enough. Yeah, and uh as long as we can keep that source for everybody to use, then I think we're much better off building things that build on that and solving real-world problems, which we are quite good at today, I think. I think so.

Fredrik Olsson: 1:40:37

Yeah, and I mean I mean, now we we're talking about one one or maybe a few different architectures that we're using already, but I mean the the next innovation or research might be in Europe, and we come up with something that's more on a hundred-watt scale, uh, that's actually more beneficial for everything. And then we have sorted out the things with the data, probably and the business needs, and yeah.

Jesper Fredriksson: 1:40:58

I think that this this is also something um maybe maybe a tangent, but I think this is uh one of my key takeaways from 2025 that this was the year when when Sweden uh got good at AI. Uh before 2025, it seemed like it was just data scientists playing with the models and doing rag. But now it's like everybody's doing it. Uh, and there's surprisingly high levels of uh competence, I think, in in Stockholm at least. I haven't seen much uh outside of Stockholm, I'm sure there are other places, but Stockholm is really a good place to be at, and I'm super happy to be here. Yeah, 100%.

Patrick Couch: 1:41:37

And and I think you're absolutely right, and I think you're right, and I think you're right as well. I think everybody's right.

Anders Arpteg: 1:41:43

Agreeing too much here. So that might be the evening of disagreements.

Patrick Couch: 1:41:46

No, but because no, I'm I'm thinking about you know technology as uh like there are there are there are stepping stones in in technological progress, and you don't everybody doesn't have to create the Betamax cassette or the dot cassette, or you can just wait out the streaming stuff and figure out Spotify and then just fucking hit it and you succeed. And you do mini-discus that you you have you have the US and China and all these other large economies work out these intermediary steps. And once you've once you see, you know, oh wait, so now we're we're now we're approaching the streaming state, the steady state. Like this is the S-curve in terms of technology. This is something usable. You can have an agentic execution, you can have some memory persistency, you can have certain things, then you only need only need to put that to use. You don't really need to create the fighter yet of sob. Like you can, but it's a unique kind of approach with that kind of economy and country to have that kind of engineering, prowess, and production capability is unique. Maybe you can't replicate that across, but you can replicate something else, namely El Ami Ericsson, for instance. He didn't, they didn't invent the telephony stuff, they just got the license for the technology for the Nordics because the other guys didn't bother with it. So they ran with it and they created Ericsson out of that. Amazing. And I think that kind of thinking is what we're seeing now in Stockholm for sure. We're not seeing the the slow, steady success of the WASP program and the PhD candidate program, all of the stuff. We're seeing an influx of talent leveraging master degree type programs at lower levels and benefiting from the social swirl that is Stockholm today in 2025. And there's a tremendous amount of swirl, and there are so many people that do so many amazing things. And these guys all rub shoulders because Stockholm is like this tiny, tiny little fish pond. Everybody knows everybody after five minutes. You go to meetups, you know everybody.

Fredrik Olsson: 1:43:48

Yeah, you go to my meetups and you meet other people, but so you just you cross-fertilize, and so I think yeah, and I guess the old uh definition sort of of artificial intelligence from was it 1956 is that is the the science and engineering of making smart machines, whatever we did the science, now it's the engineering stuff, and that's the way we can build stuff without doing the science, and then we're back to to science again.

Patrick Couch: 1:44:09

So yeah, so I think Sweden should really and the Nordics and Europe should lean into this entrepreneurial thing because that's our prowess. And that's not needing these massive financial structures. I mean, yes, you can do a series A 200 million dollar, you know, funding round, but you can also do different things, and you can do a lot of them.

Anders Arpteg: 1:44:32

Yeah, but I guess we we we will see some tech companies, you know, a few selected handful counter companies building this huge model, they will be amazing. They will do so many cool stuff that we haven't even been able to imagine today. So cool, but they will probably be super, super expensive and for one to to train, but even to use. So, you know, we've been seen man saying many times we will see like a big proliferation of SLMs or these kind of small language models or small foundational models that will be the one that actually is being used for adoption or actually is being used for real value. And and I guess you know, the one that actually gets the best in being able to build and use SLMs or normal normal size foundation models, they could be the winners. Yeah, and I think that's that's an area that's up for grabs.

Fredrik Olsson: 1:45:24

And I even if you if you look at the pricing for, for instance, OpenAI right now, you have the really expensive stuff. I'm not sure how much you have to pay for it per month, but it's really expensive. And then you have the readership side stuff, like GPT-5 nano. Yeah. So there's a large span of things there. And uh if we could just carve out one of those areas, if that's kind of correlated to the actual model sizes or capacities, wherever, there's use cases in each of these tranches of pricing, I would say. Yes.

Anders Arpteg: 1:45:55

To the point of SLMs. Yeah, yeah. For me, it's obvious this is really where we should invest. This is where we should do both the research and engineering in how to fine-tune that to companies for their own needs and for society. I mean, if we actually invest in that, Jesus Christ, I think it could be an amazing opportunity.

Fredrik Olsson: 1:46:12

So the way we think of it um uh when we we speak about companies uh all at uh at all areas is that we think that we could build stuff uh with let's say open AI, that and so we get like a stable feature set that we want to provide to our customers. Yeah, and then the uh the kind of end game is to uh finalize that feature set and and and and bring them home, like train small models for for doing exactly that, and don't then we we will not be like stuck with uh big third-party providers in that sense. And then what we need is SMMs. Definitely.

Anders Arpteg: 1:46:44

Um speak yeah, okay. But but just perhaps taking another topic before we go even more philosophical. But of course, we've seen insane amount of investments in infrastructure uh in US uh uh and in Europe actually, as well, partly some AI giga factories being built, and uh in Sweden we have spherical AI, it's a cool first like GB300 is coming to Europe, uh really cool stuff. But then we can wonder, you know, okay, so we what the question I guess is what how can we find value for these kind of big investments? And if we take like um Stargate in the US and $500 billion and in this kind of huge data centers being built in Memphis, etc. And if you were to look at it really negatively, and I want to have a positive spin here, but negatively you can think, okay, they're they're investing in in GPUs now that's going to be legacy in in three years, perhaps. Yeah, in reality, they have to really get the money back in a very few sets of years for insane amount of money.

Fredrik Olsson: 1:47:50

But part of that money is going to the actual buildings and the infrastructure, right? So when you change or or or whatever slot out, is it the majority of the money actually for I guess many many of the things will go to uh electricity? Yeah, yeah, yeah, yeah. True.

Jesper Fredriksson: 1:48:06

But but still, electricity will be a really interesting thing to to follow. Yeah. Like what will happen to our uh electricity prices? Orbital data centers.

Anders Arpteg: 1:48:17

Yeah, the space is interesting, right? Everyone is speaking about you know building data centers in space now these days. It's kind of insane.

Jesper Fredriksson: 1:48:24

Who who else? I heard uh Google was talking about it. Is there are there more companies doing that? Yeah, of course, XCI is doing it and uh in China.

GC: 1:48:32

China was the first one, and then after two weeks, uh Amazon is doing the same.

Fredrik Olsson: 1:48:36

I think Microsoft also, and everything. Well, they trained the first language model, I think, in space last week or two weeks ago. Yeah, yeah. Just a small test, I think.

Anders Arpteg: 1:48:44

Yeah, cool stuff. Yeah, I I just think it's interesting with this insane amount of investments in infrastructure. If you take it from a positive side, um, you know, it's not really the training of the model that's necessary for this, I think. You know, but imagine you know us being able using AI assistance all the time, every human all the time. They're going with the mobile phones and asking stuff and just having this kind of moving towards a video stream client or something, it will be an insane demand, right? Yeah. That will be you know in the orders of hundreds of thousands times higher than today. Yeah. And and then we need this kind of infrastructure, right? So how do we even do that? I I think you know, the I think we need some breakthroughs in terms of compute. Yes. Like energy, energy efficiency, yes, energy efficiency, like neuromorphic computing kind of stuff, right? With the human brain being able to do stuff in 20 or 100 watts, and we need megawatts in in in uh GPU to do the same thing. And if we can get down to a thousand of that, and if the one that actually succeeds doing that, that will be they will be the winners in my view.

Fredrik Olsson: 1:49:58

Do you think that would be one of the big tech companies now or something else?

Jesper Fredriksson: 1:50:02

I think that would be what they use AI for. Yeah. AI building a more efficient AI. That's that's what I think.

Anders Arpteg: 1:50:10

Yeah, like that could also be, you know, you can think of why are they investing so much in getting first to AGI? And that could actually be the reason. That's uh you know, the one that actually do get the smarter one can can build the the even smarter one, right?

Jesper Fredriksson: 1:50:24

Yeah, and and you can build anything. Yes. That's uh I think that's the simple reason. You can you can do anything if you have if you have super powerful AI.

Patrick Couch: 1:50:34

Are we on the AGI topic?

Anders Arpteg: 1:50:36

No, but infrastructure still, and and just thinking you know, why why are we seeing these kind of insane investments in infrastructure? You know, training is one thing, but probably even more so is the inference uh use, and we will have perhaps in a couple of years every person using AI ten times or even a thousand times more than today. And then we need some infrastructure.

Jesper Fredriksson: 1:50:56

I think there's also this thing, I I guess speaking to you to your kind of topics, uh, there's this uh um arms race around AI. Uh I mean it's obvious that US wants to beat China, and uh they're seeing this as an existential thing to uh to be for to be the lead globally. It's a national security issue. It's a national security issue to be the best at AI in the US. That's that's what I'm hearing.

Patrick Couch: 1:51:22

Absolutely 100%. And think of it like the space race between the US and and Soviet Union. I think it's very much that cultural dominance thing. But also, like you accurately point out, there are for sure uh geopolitical power mechanisms in play here that are completely hinging on your technology prowess. Like cyber warfare, the rest of it, fake news, new propaganda, manipulation, political instability, all this stuff that uh you know, secret police type entities have been working forever. And now they have this superpower and the fact that we're all connected to the internet, we're all reachable, it's just an API call away, touch 10 billion people, woo, let's go. So obviously, I mean, the and I think this is a super dark and terrible and horrible aspect of this. And I think we can't really shy away from it. But I also think we it's difficult as a non-actual expert in the field to have a valid point to provide, sort of like I looked to others, and obviously you have, but I think for most of us we can go like, yeah, so AI for weapons industry, bad, but you still want your team to have the biggest club.

Jesper Fredriksson: 1:52:44

No. But let's take a positive spin on it, and what what what do you think? Uh uh, what do we think that AI will be used for? We uh I mean, so I said that you can build anything if you have the most powerful uh AI you can think of. Uh, what do you think we will build? I uh I mean I think it's interesting that uh uh OpenAI at least talks a lot about uh um research, that uh this is what they're planning to do now. And uh I guess the best case is just making an AI researcher because that's uh probably proving what they're doing. But they're also leaning into a little bit the uh the uh research aspect of it, and we will probably see AI doing research discoveries next year. That's that that's uh I think not even a bold prediction at this stage, even though it feels weird to say that out loud. But I think we will see like uh real scientific breakthroughs, small ones next year.

Patrick Couch: 1:53:39

Yeah, absolutely. And and and to return to your point in terms of infrastructure and why are these built out and what's the arms race really about and what's going on here, why all the investments? I really think that it's because of something that Juvel Harari spoke very eloquently about in terms of the human naked apes, we're like creatures of the imagination. We we we live through stories, and there is a story everybody can get on board on today being told, and that is imagine if we can just extrapolate prowess and technology into the future. Imagine the things that we can do, and we're going like, I can imagine that. I mean, imagine anything. Well, I mean, you you see, you think about Matrix, you think about uh any movie or book that you read, you think about your own dreams in terms of if I could just have a digital agent avatar kind of person do stuff for me, what would I ask it to do? And I mean, everybody can see this, and then and everybody understands that, well, regardless of what you're seeing in terms of your vision of the future, it will require infrastructure and compute. Yeah, so you can't go wrong, sort of.

Anders Arpteg: 1:54:47

Even if you're saying, yeah, but what about the financial progress in terms of you know how you can build the hardware and the infrastructure, and if they actually do invest in a lot of things that is going to be obsolete in a couple of years, that could be horribly bad.

Patrick Couch: 1:55:02

But but I don't think it's gonna be obsolete.

Anders Arpteg: 1:55:04

I think that's uh that that's a negative uh uh misconception of the possibilities because once you No, no, but but imagine that they do like super big um advances in in neuromorphic computing, and you can get something that's similar to the human brain in terms of energy consumption, then of course it would be really stupid to build out infrastructure that is this in energy consumption uh consuming as as they are today, right?

Patrick Couch: 1:55:28

Oh, yeah, absolutely. But you could probably make use of that older stuff for some other stuff.

Anders Arpteg: 1:55:32

Yeah, but super expensive.

Patrick Couch: 1:55:34

Um I mean I'm thinking about uh you know Christian Lange and Barrietoi and this funny story he told about uh the sort of how Barriet Oi came about. And he had this, and he was at one of our meetups, and he told this wonderful story about how the how we sort of more or less had some servers around. And they were all like legacy stuff, and it was like, yeah, maybe we could you know create uh an inference machine out of this, like how hard can it be? You equip it with some GPUs and off you go, right? And so it wasn't that easy, but I think like you you it would be sad if you if you if your vision of creating a piece of infrastructure and hardware and equipment is that if it's not used for what it's supposed to be used for, it's obsolete per definition. And I don't think so. I think kids will find uses of these legacy data center type equipment that are, oh, look at this old GB400 thing.

Anders Arpteg: 1:56:28

I think it's it's you know, we certainly need the infrastructure build-outs. It's just that it is I mean, it's on a scale that is you know insane. I mean, they're spending money that is like GDP of Sweden, like all the income of all the companies combined is what's being built in the US as terms of AI infrastructure right now. Money they don't have. Yeah. Yeah. That's being loaned from companies that loan money to do that, that buy it from other companies that loan money to do that. I mean, it's the capitalist.

Fredrik Olsson: 1:56:58

But to me, turtles all the way down.

Patrick Couch: 1:57:00

Yes, turtles all the way down, but it's also to me a completely separate story. Like there are two parallel stories going on at the same time. One is we're building the data center, shovels in the ground, buying cables, connecting stuff, and leveling, you know, trees and the rest of it. But the other is a financial dream thing that is happening parallel, and people tend to confuse the two. But I think I don't think they should be confused. Yeah. Like the financial thing is, I mean, I'm not a macroeconomic guy, but I was surprised that uh Larry Ellison could put those $300 billion CEO, yeah. The Oracle CEO could put those $300 billion on the books as future revenue and have his stock increase by 40%. Now they're down.

Fredrik Olsson: 1:57:46

And now they're down in September.

Patrick Couch: 1:57:48

Obviously, right? Because it's what is that? I mean, that's just so so I think so. But I think it's important because if we want to understand where the future is going, we need to distinguish between the financial spiel narrative thing and what is actually, you know, the the progress being made and that is that is genuine.

unknown: 1:58:08

Yeah.

Anders Arpteg: 1:58:09

So going back to the Nordics then, and trying to close this topic a bit, and perhaps even Sweden in this case, uh, I think there was some positive news in 25 in terms of Sweden when the government actually did some significant investments in AI. I mean, it was a couple of billion over five years. Um, it wasn't the 12 billion Swedish crowns that the Swedish AI Commission were asking for, but it was at least like a tenth of it, at least. And I mean it's not it's not that bad. No. I was actually surprised that it was that much. What do you think? Do you think Sweden is very positively surprised with a budget that came up from the government in Sweden?

Jesper Fredriksson: 1:58:48

Yes, but I don't have much to say actually. I didn't I don't follow it, and I think uh to some to some degree it doesn't affect me. You're like the Zen Buddhist at the table. I love it. I mean what what happens with those m with that money, uh what are we gonna do? I don't know. I haven't followed the news. But but that I think is an excellent angle though.

Patrick Couch: 1:59:10

Yeah exactly, right? It's easy to conceptualize the amount and then then compare the amount and contrast the amount with some other amount from some other place. But that your point, your question is a much better question. Because even if you only fork up 5 billion, if you put 5 billion into uh you know social welfare for homeless people in Stockholm, that amount of money is a tremendous amount of money. But if you funnel it into creating the quantum computer in one of those uh KTH labs, yeah, maybe not so much. No.

Fredrik Olsson: 1:59:44

What's the expected outcome of this? I mean, that's that's the thing. Uh probably we're gonna be world leading. Is it knowledge or yeah, world leading probably in in metal modern metal?

Anders Arpteg: 1:59:54

You know, I'm trying to end off in some kind of positive sense here. That's and then you're dashing it down on the top.

Fredrik Olsson: 2:00:00

I think contrarian? No.

Jesper Fredriksson: 2:00:05

Because I didn't follow it. I don't know what happened. I don't know.

Anders Arpteg: 2:00:08

No, but they they actually did uh I think invest a significant amount and they even earmarked it to specific, you know, concrete investments it's going to do that's good in for Sheckingskassan and Tax Authority, knowledge uh the the um uh Corbea, the the national library, and then then I think it's good because you have like an outcome in in mind for it, and it's also a testament to that this is important.

Fredrik Olsson: 2:00:31

Let's do it.

Patrick Couch: 2:00:32

Yes, yes, and you and you also have a bottom-up approach because if if those if those if that figure five billion, if if it's just taken it wasn't that much, but no, but whatever it is. If you build it up from the ground up and say, you know what, for seconds, Gasland could actually benefit from this money for this project. So that there's a hundred million there, and then you continue just piecing it together, then you get a very good, like worked out. This is the purpose for that investment, and that justifies you know the nation coming together on this. That makes sense to me. Yeah, and also I should say, I haven't either kept closed tabs on this, but I'd be super willing to just lean on Texferia and ask them, what did you think about that? And if they if they say I we think it's good, I'm gonna say it's good. But we'll just based on that.

Fredrik Olsson: 2:01:19

We want some more, they will say they will probably say they want more.

Patrick Couch: 2:01:22

But but anyway, so I think Texvaria is a really good place for for vetting and assessing political, politically motivated investments into technology. I think Texvari is they're doing some really good stuff, and I'm I'm happy for them.

Anders Arpteg: 2:01:38

Cool. Um I I want to throw up an uh controversial topic. Um, what will happen with open source AI in 26? Uh I know I know I think what the answers will be here, but still I would like to. If you start with you, Felix.

Fredrik Olsson: 2:01:54

Yeah, yeah. I I was in one of these um well, meetups or get togethers, I think last year, and we I was to debate a topic with someone, and and we all thought the same thing, so we had to to to uh agree to disagree. So I chose the point of uh arguing in favor of proprietary API delivered uh LLMs and Kaisan Arin said, well, I'll take the open source. And there's merit to both approaches, but they're not the same. I mean, if you buy something from OpenAI, Athropic, Google, uh, there are services, not models, right? So it's a completely different thing. You put in your your critical card which has pros and cons, right? Yeah, of course. So it's a different use cases. Uh as I like I mentioned before, we are now working with things that we can buy over API, and then we get the features that we want, and then we envision that we would bring home the open source models and retrain them. Like so, there's different things. Uh but you can retrain through the APIs as well. Yeah, yes, yeah, but but if you want to have control for the sake of your business, you don't want to depend on third parties too much. Uh, so that's that's uh yeah. Um, so I guess there's um what would happen in open source is that there's not really a competition, there's an overlap with the other models because it's not the same kind of stakeholders that's that are using them. Uh so I guess open source will benefit from the increase of use in in uh like services uh preparatory things, and and maybe vice versa as well. So I think there's going to be um everyone will benefit, we will uh live have it happily ever after and prosper. I think yeah, I think it would be better.

Anders Arpteg: 2:03:30

Before I I know what Patrick thinks about this, but but before we go there, um we we can see China, of course, being very open source, and you can think about why they are uh that yeah, and and we can also see see the top frontier labs, I would say, go in the other direction, and uh then perhaps uh at least for the latest kind of models, they're even Meta is not uh releasing them, yeah, nor is Mistral and other companies. Um what do you think in 26? Will it I mean a good thing is that OpenAI actually did GPT OSS, meaning they actually did release a model that of course wasn't the latest one, but it was you know second to last or something, which is similar to what Elon said, you know, they're never going to release the latest model in open source, but you know, second to last, yeah, could be what is is what do you think will happen in in 26? Will that continue in that way?

Fredrik Olsson: 2:04:20

Or now we're talking about the providers, we're not talking about the real elephant in the room, which is hugging face, right? Where everything is collected and this because that's also they don't build models. No, exactly, but that's just distribution. So if you build something, post it there, and it's it might be a success from a single person, private person, small lab, whatever, yeah, it will get distribution and that would impact the use of open source. Uh and now they're into robotics as well. So I I think um I think there's like a there's no contradiction between closed source and open source here. I think the big labs will also be.

Anders Arpteg: 2:04:53

Will the trend go towards proprietary or towards open source?

Fredrik Olsson: 2:04:56

You think uh it depends again on uh it always depends, but I mean it's unevenly distributed. So the really techie people will probably look at open source and say, is this something we could justify actually running ourselves? Or is it too expensive? Do we need the manpower or do we have the manpower for it? If not, we will go with proprietary source. I think so. I think um different use cases. I think it will continue to grow and it's not really competition.

Patrick Couch: 2:05:23

Patrick Yeah, so um what do I actually think? I think well, I know what I want to happen. What do you need to happen? I want I want uh open source to to win and everybody to embrace it and uh re-release everything, information wants to be free.

Fredrik Olsson: 2:05:42

But then you will just uh uh move the the the power to something else that uh like the hardware providers.

Patrick Couch: 2:05:47

Yes, or the or like you say, the distribution. But but but so this is an interesting question, and I'm I sort of I realize I'm ideologically v wedded to you know wanting open source because I think it's so beautiful. But I think it comes down to I think what you the way you distinguish between open source and service, I think is very true. Because a service is a dependency on something that you don't care too much about. You go to the restaurant for a nice meal, you don't really care about things. You want the the the plate to look nice and taste good and be priced accordingly and and off you go to glory. And I think that will be super relevant for for a bunch of stuff. But I really do think that the more the the more central our use of this version of digital technology will be, the more we embrace it as a core component. I think Linux in into you know how ubiquitous it is.

Anders Arpteg: 2:06:51

But do you see any negative aspects of if we were to move everything to open source, could there be some negative impact on that?

Patrick Couch: 2:06:58

Well, I mean I'm I'm not knowledgeable enough to actually have a proper answer, but I can tell you that I think that the downside is precisely the kind of downside we see with any collaborative effort. It's messy, it's non-efficient, it's confusing, it's changing over time. People have the same view of the United Nations, and they're like, oh, it's so inefficient, and goddamn it, and like why can't we just build a wall and because the totalitarian state's not.

Anders Arpteg: 2:07:27

But we know, of course, from an IP point of view, people or companies don't want to release the latest one because they can simply very easily steal it, which DeepSeek did before and showed it. It's very easy to do. So of course they're not going to do that from an IP point of view, and okay, that that's not from uh the the best of society, simply from a commercial point of view, you don't want to do it. But there could also be security implications here, right? If we were able to generate um Sora videos, and then you can fine-tune it to remove all the watermarks that they are provide providing in it, it would actually create some you know problems from a security point of view as well, right?

Patrick Couch: 2:08:05

100%. And and again, to return to your point about pace, uh how how do we increase pace? Well, the way to increase pace is just to flood the market with stuff for people to seek their teeth in for sure, but and then out of that will come the most surprising things, good and bad. But there will be more than that.

Anders Arpteg: 2:08:24

Is it worth any cost to do that? Like security implications?

Patrick Couch: 2:08:28

Well, so the question you should ask yourself, I think, is rather how do what kind of security approach would you like? Think of it, think of it in terms of uh, you know, these um the C scam, the child control, these various political measures that are geared towards preemptive misuse, like minority report territory. And and and you have to ask yourself, yeah, but to what extent am I comfortable with that kind of policing thing in order to weed out the worst case fallout, which is sort of the EU act risk mitigation approach, right? We want to figure out the worst negative fallout and then backtrack from that and say, okay, that's not okay. That needs to be regulated, that's okay to just go do.

Anders Arpteg: 2:09:08

Because we want to look at the But you can even take the Yanikul approach here, saying actually it actually is more secure to have open source out there because you can actually test the security of it in a way you couldn't if it's proprietary. Exactly. So what I'm not sure I'm buying it, but still you could take that approach.

Patrick Couch: 2:09:24

Yeah, so what I'm leading towards is how do you want to play your defense? Because if pace is the future and you go, go, go, go, go, but you still want to sort of not have too much chaos and and and negative fallout, how do you how do you go about it? Because you can go, you can go full Stasi and just have everybody report on everybody and just lock everything down and it's super strict, and it but you're like, I don't want to live there. So where do you want to live? You want to live in free America, okay, but you don't want to live in downtown. San Francisco. But so what I'm trying to say is there are ways to balance and trade and fight trade-offs, but those are those are not preemptive.

Anders Arpteg: 2:09:59

How do you do that?

Patrick Couch: 2:10:00

They have to be reactive. How do you do it? Regulation or what? Regulation for sure, yes, absolutely. And you have to do it through uh societal structures. Like there are very few.

Anders Arpteg: 2:10:12

If you float an open source model out there, regulation have no way to impact that plan anymore, right?

Patrick Couch: 2:10:19

But what I'm saying is misuse is a direct consequence of stupidity. Uh but no, but I think it's valid, right? So you have to do that.

Fredrik Olsson: 2:10:27

Yeah, but I mean you alluded to the data thing that you could, you know, retrains or train stuff on scientific data and remote stuff. So yeah, and in that case, it it it's enough. If you mean that you have an open source model that could generate data for you and train on that, yeah, it's enough with one copy. Yes. And uh if it's not stupid use, it's like malicious use. Yes. And I think that's where you're going because of maybe part of your your background. Uh, some state actors want to get this. Of course. And there were stories about this year and a half ago about Meta's uh Lama models being you know used in China for stuff. Yeah. Um I'm not sure if if it's if there's a way around that and have open source as well.

Jesper Fredriksson: 2:11:14

I mean at the at the current uh sort of technical landscape, it doesn't seem like a big risk with uh open source because there's such a uh there's such a high pace of development. So I mean, for example, the deep seek uh moment that we had this year, super interesting uh moment in in 2025 for many reasons. And it was out there and people could do anything with it. But now that model is history, and uh I mean we're we're always looking to the next thing.

Anders Arpteg: 2:11:46

So I uh where we have a nice reasoning model with the 3.2 coming up.

Jesper Fredriksson: 2:11:50

Yeah, yeah, yeah, definitely. And and uh I'm I'm not arguing against it, I'm just saying uh 3.2 will also be irrelevant in in a year from now. Yes, agree. So I I I don't see really a problem with open source because there will always be a new model.

Fredrik Olsson: 2:12:05

It depends again, I think, on the areas of use. If you're thinking of the state of the world today as we perceive it or as I perceive it, with there's a lot of you know, unrest, unruliness. Uh, and there's a like a higher and higher level of background noise of stuff happening, and you don't know if it's true or not. Just have to like you you throw a mine in in Astrochon, and say there's mines here, and people have to be very careful. So there's the same thing going on, I think, in one level, uh with respect to what people are are um like subjected to in terms of polarizations and you know info ops and so on, psy-ops.

Jesper Fredriksson: 2:12:44

And then there's that doesn't have to be the latest model, it just has to be one, and you have to Yeah, but there's there's always gonna be an arms race, and and I mean uh there's gonna be like the the black side and the white side, and they're gonna uh compete against each each other. And I don't see a risk of of like one model coming out because there will be a much better model in just the good side will use the better models to detect stuff. Yeah.

Anders Arpteg: 2:13:08

Since other people would do bad stuff, let us do bad stuff. That sounds like a really good idea.

Patrick Couch: 2:13:13

No, that's not that's not positive.

Fredrik Olsson: 2:13:17

But so what about the positivity that we should analyze?

Patrick Couch: 2:13:22

I think there is positivity here. But but but I I still want to sort of ground this because I think there's there's this uh I think very unhealthy uh fear-mongering, you know, banging the drum of, oh, but uh this technology is so amazing and powerful, we're gonna have the PhD candidates around the corner and blah, blah, blah. So we can't give it to everybody because you imagine what the terrorists are gonna do with the PhD candidate. Just gonna do, well, it's gonna create the bombs, keep them away from the technology. They're like, but really that that won't work because technology travels information and wants to be free, and everybody gets it eventually.

Anders Arpteg: 2:13:57

So there's double- So you think all the safeguarding that everyone tech yang is spending so much time on is completely useless?

Patrick Couch: 2:14:03

No, but what I'm saying is you don't try to say is you don't uh prevent Russia from invading Ukraine by limiting their access to outdated weapons. Um it's not the access, it's not the access to weapons uh that that decides if they're going to invade or not. It's mental hygiene. You have to play the long game here, and you have to have belief in your fellow human being, and that belief has to be if you create a society that feeds, clothes, shelters, buries the dead, do all that great stuff, and you increase the medium, median ease. So we don't need regulation at all. No, no, no, no, no. I love regulation. I want everything to be free, legal, and regulated. No, I'm not saying the opposite. What I'm saying is you have to fight it through the long-term investment in society, not through a prevention of access to technology, which will anyway seed into the user community regardless of your preventative efforts. It's mute, it's like the war on drugs. You can't have war on open source, it's not gonna work ever. Should we stop the war on drugs? Yes, 100%. Okay. 100% it hasn't given us anything.

Anders Arpteg: 2:15:21

Yes. Okay, that's a big topic. Okay, um, um, let's let's agree to disagree. Finally, some disagreement. Okay. Um, let's try to end off here, and I'm trying to find a good way to end off on a positive note. I have okay.

GC: 2:15:44

So I have two things for you. So you can choose between uh uh a good quiz, which is basically a realistic quiz about this year and next year predictions. The game is about like um what are the chat GPTs and answer to the questions that you have this and one is going to be a comedian and the other one is going to be regular. So do you want to guess the real stuff, or you want to take the comedian satirical one?

Fredrik Olsson: 2:16:17

Comedian, I think I would go for comedian anywhere, anytime. We should answer the positive note.

GC: 2:16:21

What's gonna happen is that I'm gonna basically prompt it here, and then it will list seven questions. Anders will read the question, then you're gonna guess what ChatGPT has actually answered that question, and then chat GPT will answer that question, and then you will basically answer if that is true or false.

Fredrik Olsson: 2:16:40

And then someone will be eliminated from the island? Yes, yes, exactly.

GC: 2:16:45

Love it. All right, let's start.

Fredrik Olsson: 2:16:49

Cool.

GC: 2:16:50

No, wait, I'm not writing anything.

Fredrik Olsson: 2:16:53

First question is hard. Is this conversation helpful so far?

Anders Arpteg: 2:16:57

Ah, you prepped it somehow. Okay, cool. Which of the following best captures the moment uh in 2025 when the AI industry collectively stopped pretending everything was moving at lightning speed and quietly admitted that uh what was actually happening. A breakthrough model passed a new benchmark. No one can explain. Enterprises announced AI success will still while still being in pilot. Regulators finally understood how large language models work and at all. Um AI achieved conscientiousness uh during a board meeting and asked for budget approval. Okay, anyone wants to go first?

Fredrik Olsson: 2:17:53

I would go with D because I would let that happen. AI achieved consciousness during a board meeting and asked for budget approval. Yeah, that'd be really nice if you agree.

Jesper Fredriksson: 2:18:06

B sounds too realistic to be true. That's one way to put it. Enterprises announced AI success while still being in pilot, isn't that uh the norm?

Fredrik Olsson: 2:18:16

That's the norm. It is, yeah. Regulators find C at least, right?

GC: 2:18:24

All right, choose one. Let's go.

Anders Arpteg: 2:18:26

Okay, A Yeah. Uh okay, B for me.

GC: 2:18:33

All right, write down somewhere so you know what you can do.

Anders Arpteg: 2:18:48

Oh, it's gonna be a quiz properly, right?

GC: 2:18:51

It's a proper quiz. So there you go.

Jesper Fredriksson: 2:18:59

Too boring to be true.

Anders Arpteg: 2:19:02

Read the answer, basically. Okay. So the correct answer is uh enterprises announces AI success while still being in pilot. Yeah, so because nothing defines 2025 better than calling something strategic, scaled, and transformational while it still lives in books that are owned by one person who left the company three months ago. Ladies and gentlemen, let's give it up for pilot as a service. I love it. Pilot as a service. Pilot as a service.

Fredrik Olsson: 2:19:30

That's could be good if it was an airline, but no, no.

Anders Arpteg: 2:19:34

Yeah, yeah, yeah, yeah. Or fight it yet. Okay, question two biggest twist in AI this year. What was the biggest, most unexpected plot twist in AI during 2025? AI replaced middle management overnight. Humans used AI mainly to generate slides explaining AI. Open source model. Ended all vendor lock-in. AI ethics debated debates where we saw calmly and rational. I see an escalation from A to B in both questions so far.

Jesper Fredriksson: 2:20:15

The correct uh answer is obviously B. I mean, that's what we've seen yes. Which is B is random, but is actually interesting that uh all of a sudden we could do PowerPoints.

Anders Arpteg: 2:20:26

The thing is, AI can't generate PowerPoints. But there's a I I would challenge anyone to try to do it. They are too stupid to generate PowerPoints.

Fredrik Olsson: 2:20:32

But there's a company, I don't remember the name. Would you think it works? But there's a company.

Anders Arpteg: 2:20:41

No, but agree, that's exception. But everyone else, ChatGPT, they are horrible in general generating presentations. But Gemma does. Gamma app is a gamma. Like 50 people under like profit. Yeah, that's amazing. They got a lot of investments well. It's one of the few that actually can.

Patrick Couch: 2:21:00

Yes. Yeah, I love it.

Anders Arpteg: 2:21:01

That's or presentation. But it's amazing to see how stupid AI is today. It can't even generate a presentation.

Jesper Fredriksson: 2:21:07

But I uh I haven't done it myself, but I've seen many presentations generated by Nana Banana.

Fredrik Olsson: 2:21:12

It's just images. But I guess that's the agent case. I mean, if you you're on you're under specifying something, I want a presentation about broccoli.

Anders Arpteg: 2:21:20

And you get something, and that's not good. But I think it's a perfect example of how horrible AI is in taking actions today. It can't really create a presentation about broccoli.

Jesper Fredriksson: 2:21:30

But I think this this is great. They need to have a challenge to produce a PowerPoint presentation in the next uh next meetup.

Patrick Couch: 2:21:37

Yeah. But isn't this emblematic of 2025? Precisely this. I mean, we we have these billions and billions of investments flying around and we're doing all this stuff and we're talking about, oh, but what about uh the misuse of AI and blah blah blah? And at the same time, we're like, yeah, but really you can't even have it make a PowerPoint.

Anders Arpteg: 2:21:54

If you were to open Gemini, or even Chat TBT and say, create a presentation uh for XY and Z, you you will not get it to work. I challenge you, any one of you, to do it today.

Jesper Fredriksson: 2:22:04

I take up the challenge.

Anders Arpteg: 2:22:05

Yeah, this is this is the thing. Yes, we can try it now. You will it will fail.

Jesper Fredriksson: 2:22:09

I I haven't done it myself, so I think I would need some implementation. But I'm I'm I'm pretty sure. What was his name?

Fredrik Olsson: 2:22:15

Ethan Ethan Mollock? Uh yeah. Uh his his book about this. I mean, it's very like a a star. It AI, whatever it is, is really good at some stuff, but not other stuff, and that other stuff is apparently PowerPoint.

Anders Arpteg: 2:22:27

No, no, but it will get some kind of PowerPoint, but it will look horrible. It will be unusable. Yeah. I tried it, I know. Just like Gemini changed.

Jesper Fredriksson: 2:22:36

I've had members on my team who did it, who showed it to me. I haven't tried it myself.

Anders Arpteg: 2:22:41

They can do bullet points and they can have the content, but not really.

Jesper Fredriksson: 2:22:44

No, really images.

Anders Arpteg: 2:22:45

Maybe you have a difference in taste. But then it's just initial news. I mean, then it's just like nanomanana kind of images or or MP. Yeah, but it's generating this. Well, let's try it out. You want to have a proper PowerPoint. If you can get a proper PowerPoint with a proper point, but I really think there's elements in it. A full night out every expense pays.

Jesper Fredriksson: 2:23:11

So we had we had an incident at our workplace. And the guy that that stood up and took responsibility for it, he made a presentation the day after and he said everything was generated by uh by AI. It could be that he generated images. But uh isn't that the same thing?

Anders Arpteg: 2:23:30

If you generate images and put it in the point is, it can't do something that humans can do very easily.

GC: 2:23:36

PowerPoint is hard.

Anders Arpteg: 2:23:38

Humans can still do it.

GC: 2:23:39

Yes, yes, yes.

Anders Arpteg: 2:23:41

And basically every company, more or less more, most middle person can do a bootpoint. Okay, but most people in a f in a company can generate or create a PowerPoint presentation. AI is really, really bad at it, right?

Patrick Couch: 2:23:54

But I think this is great. I mean, I think to me, this is the perfect summation of 2025. Yes, this is where AI is today. But that's what is the answer?

Jesper Fredriksson: 2:24:02

I I I would guess that the answer is B.

Patrick Couch: 2:24:05

B. Oh, I do.

Jesper Fredriksson: 2:24:07

Oh, okay. Yes, it's always B. So Chat GPT agrees with me.

Anders Arpteg: 2:24:13

Okay, we we should create a presentation about this and you will see how horrible it is.

Fredrik Olsson: 2:24:17

This was fun. Move on.

Anders Arpteg: 2:24:23

You're so polite, please. No, you waste a lot of freeze token. Okay. Okay, question three. Did enterprise AI actually scale? Which statement best describes the real state of enterprise AI scaling in 2025? A. AI scales seamlessly across the organization in under six months. That's funny. Okay, B, AI scaled tech uh technically, but failed culturally. C, AI scaled in budgets, governance, and committees, but not in production. And finally, the AI was abandoned in favor of Excel, which remains undefeated. D for B.

Fredrik Olsson: 2:25:06

D for B again. We have two data points on B so far.

Patrick Couch: 2:25:10

I'd say scale technically?

Fredrik Olsson: 2:25:12

I mean, whatever. Yeah, I just I'm just looking at back in my two previous data points. So it's B and two previous questions.

Anders Arpteg: 2:25:26

Okay, correct answer is C. AI scaled in budgets, governance, and committees, but not in production.

Jesper Fredriksson: 2:25:31

I think I've just chucked a bit too much, so I know how it thinks.

Patrick Couch: 2:25:35

But do you actually think do you agree with it? Do you think that AI is allocated the budget? It sort of needs to have proper impact culturally. Scaling is not just a good idea.

Jesper Fredriksson: 2:25:46

So what is what is the question?

Patrick Couch: 2:25:51

No, uh if if this is correct, if AI scaled in budgets governance and committees, but not in production.

Anders Arpteg: 2:25:56

Can't you actually think that's for fun, uh if you open up it? I think it's just making a play at it delivery. Chat TBT. Right, right. Okay.

Patrick Couch: 2:26:08

You don't really see budgets allocated properly. That's true.

Anders Arpteg: 2:26:12

No, no, but I just want to just uh a US Gemini. Sure, sure. You can try Gemini or whatever.

Patrick Couch: 2:26:20

No, I guess.

Anders Arpteg: 2:26:21

Yeah. But but just to get that point, it can I mean you will see it very quickly. If you just open it up, please. Do Gemini or Chat GPT or whatever, and just ask it, create a presentation about AI highlights on 2025 or something. So after No, no, not Gemma. Do the pro no two and a half hours of discussion about this. Now you're cheating. Do the chat.

Patrick Couch: 2:26:42

This is exactly in the state of AI.

Jesper Fredriksson: 2:26:47

I I've done semi-good presentations in in uh Claude code. They were not as good. So I'm I'm pretty sure that uh that the better model.

Anders Arpteg: 2:26:57

If you go to Chat GPT or Gemini, yeah, for if you train for it, it can halfway work. Gamma is really good. But the point is the vanilla, the normal Chat GPT Gemini Claude is really, really bad. Claude is probably the best one because it's actually trained to work with Excel and Python.

Jesper Fredriksson: 2:27:13

I think uh Gemini is probably the best model now because it's multimedia.

Anders Arpteg: 2:27:17

It's horrible. I've tried it. But let's uh finish the other test so we can do that.

Jesper Fredriksson: 2:27:20

But we have to have a face-off. I challenge you. Okay, we can try it here now. No, but uh no, no, no. This is not a proper one. We need to just tell it to Chris.

Anders Arpteg: 2:27:32

Okay. Uh question four. Which AI model truly dominated enterprise conversations in 2025? A. The most advanced foundational model with record-breaking benchmarks, B, the fully open source model deployed at scale across industries. C, the mysterious custom internal model no one can see, test her name. D, the AI that finally replaced Excel Macros. There's no such model, right? No, not like so.

Fredrik Olsson: 2:28:09

Which one of these are is open AI?

Anders Arpteg: 2:28:13

Truly dominated enterprise conversations.

Patrick Couch: 2:28:17

I think it's A. People don't need to talk about benchmarks. Yeah. Because it makes it more concrete.

Jesper Fredriksson: 2:28:23

I think I would bet on Cetting Rude.

Anders Arpteg: 2:28:36

Nice.

Jesper Fredriksson: 2:28:38

I think it's almost scary. Yeah, I think I know how it thinks.

Fredrik Olsson: 2:28:44

Or is it the other way around?

Jesper Fredriksson: 2:28:47

Maybe it's June for me. That would be great. One more, and then we'll go to do the challenge, the proper challenge.

Anders Arpteg: 2:29:00

Okay. Question number five. Predictions we got totally wrong. Which AI prediction age aged the worst in 2025? A. AI adoption will be gradual. B enterprises will move slowly due to regulation. C AI will scale fast once pilots prove value. D people will overhype AI on game.

Fredrik Olsson: 2:29:32

I guess C, yeah, because that hasn't been falsified yet.

Jesper Fredriksson: 2:29:39

I I noticed that that uh Chat GPT is very pessimistic about AI in 2025.

Patrick Couch: 2:29:45

Plays the long game, sees no uh drops in Roma.

Anders Arpteg: 2:29:52

Yeah, probably see. Yep, see AI will scale fast once pilots proved that purely. That's the one that uh was the worst one. Yeah, and that yeah, I agree. Okay, Goran, please. Oh, just do a new chat window.

GC: 2:30:10

We can close the podcast.

Anders Arpteg: 2:30:12

No, no, no, no, no. Please. All right, cool. New chat GPT, not that one. No, no, stop, Goran. Go to Chat GPT.

GC: 2:30:21

Why are you afraid of Google?

Anders Arpteg: 2:30:22

Are you not in control of your agent, Anders? Okay, create the press right, create a presentation about highlights and AI highlights from Twitter.

Jesper Fredriksson: 2:30:34

Chat can't do this. That's that we all agree on.

Anders Arpteg: 2:30:37

Okay, but isn't that strange?

Jesper Fredriksson: 2:30:39

Um no, because it's not trained on that. But a human can do it. Yes, but this is not multimodal.

Anders Arpteg: 2:30:48

It is. ChatGBT is multimodal. Do you think so? I know so. It can work with images. It can work with images, but it's not working internally. It just presents text. It's horrible. Yeah, okay. You can try it, try in Germany, the same will happen.

Jesper Fredriksson: 2:31:01

I I bet you.

Anders Arpteg: 2:31:02

Normally, actually, ChatBT creates uh banana can do this. And they can create images, yes, but not a presentation. But so this is a difference. A presentation is a presentation, an image is an image. It's a big one.

Jesper Fredriksson: 2:31:14

It's it's just a series of images.

Anders Arpteg: 2:31:16

But it's not the proper presentation that humans can do. So AI can't do a simple thing.

Jesper Fredriksson: 2:31:22

I haven't done it, but I'm pretty sure it can be done.

Anders Arpteg: 2:31:24

Of course, you can do images, but the you you completely missed the point. Oh, I didn't define the agents in that word, define. Humans can create the presentation.

Patrick Couch: 2:31:33

AI cannot. But so let me just let me just so insert here, because I think I think this is this is interesting to me. And I think you are both right, but it's quite different because you seem to argue that, hey, what's in the name, right? If if if it's a presentation, it means that you're conveying something visually and you're talking to it, and here it is, and this is the presentation, right? So remember when they started creating uh uh generative uh doom type worlds that weren't really there, they were just like the next token predictions. They were never there, yeah. They were never there, right? There wasn't a spoon. And so Anders is right, neo Neo Badana creating the semblance of a PowerPoint presentation isn't a PowerPoint, isn't it a PowerPoint presentation? It's a nice presentation that conveys a message issue of blah blah blah blah blah. But no, but so the way I see the point is if you want to have it create something very straightforward, like hey, a PowerPoint. You open up the application, you use the application, and then you then you save it as a PowerPoint, a PPT file, then you open it up and humans can do that, and they can master that tool very easily. Maybe not Goran, but most other people are comfortable with PowerPoint. Now Goran is a master at that. No, no, but this LL point. No, no, no.

Anders Arpteg: 2:32:49

This is a very simple point. AI is really bad in taking actions. Humans can easily create a presentation that AI cannot do today.

Jesper Fredriksson: 2:32:57

It's super simple to make a presentation that we can agree on, right? It's just it's just uh using Python to create it. That's that's the simplest way to try it.

Anders Arpteg: 2:33:05

Try it, it will fail. That's normally what Chat GPT does if you ask it. Create Python code to do it and it will look horrible.

Jesper Fredriksson: 2:33:11

I've done it with with C.

GC: 2:33:12

It looks horrible.

Anders Arpteg: 2:33:14

It looks horrible, right? I mean, it will not be at the level that humans can do it at all.

Jesper Fredriksson: 2:33:21

I would challenge you.

Anders Arpteg: 2:33:23

I mean, expectation.

Patrick Couch: 2:33:24

I think you're in love with disagreeing, so I don't think this is going anywhere.

Anders Arpteg: 2:33:28

So no, but I I don't you think it's interesting? And so many people are saying, you know, AI is amazing, it can do so many things. It can't even do a simple thing as creating a PowerPoint presentation. It's a train dog. Uh I I but why isn't it? It's super general, it should be able to do anything, right? But it can't. That's my big point. AI is surprisingly stupid.

Patrick Couch: 2:33:49

Yes, and I agree with you 100%. Yeah. 100%. Have you tried conversation through years? Have you tried a banana?

Anders Arpteg: 2:33:56

Of course it creates images, but that's not the point. It cannot create what's the difference between creating images. It's bad in taking action. That's the point. So what's the action? It's it's bad in creating a PowerPoint presentation. A human can click with a keyboard and mouse.

Jesper Fredriksson: 2:34:12

I'm not talking, I'm not talking about clicking. I'm talking about generating a PPT. I'm talking about clicking. I'm talking about doing something simple. Clicking is something different. If we're talking about clicking, then I agree.

Anders Arpteg: 2:34:22

And it can't even create it with code either. But that it can. It looks horrible. Try it. I tried it. It looks decent. No, it looks, it doesn't look like this. Yeah, I tried it as well. You don't seem to see the point. The point, I've tried it now, say it five times. AI is surprisingly bad in taking actions.

Jesper Fredriksson: 2:34:43

It's not. I think it's super good.

Anders Arpteg: 2:34:45

And we end on this uh then it should be able to create a presentation possible. I mean AI coding is a good thing. If you think like this, if AI was really good in creating presentation, every company would be.

Jesper Fredriksson: 2:34:58

You must agree that AI is super good at coding. That is taking action, right? Yes, yes. Agreed. So not to create action.

Anders Arpteg: 2:35:05

Of course it can take action.

Jesper Fredriksson: 2:35:06

But you were saying the other thing No, it's bad at taking action.

Anders Arpteg: 2:35:11

I don't think we're getting anywhere. It's of course it's of course it takes action, but it's really bad in the world.

Patrick Couch: 2:35:15

The funny thing is, I can see both of them being completely right. And I'm comfortable with both. But you're sort of not talking to each other. That's how I'm seeing it.

Anders Arpteg: 2:35:25

So we I think this is such an important point because people don't understand what AI is bad at. And even such a simple thing that humans are really good at is AI really bad at. And this is called the Maravx paradox. And it's actually very important to understand. And I think a lot of companies do not understand it. If we had a really good service to create a presentation, which gamma. They're worth it, they got so much money, so much investment, because it's so important kind of thing that every company wants to do, it will be amazing. But it's really, really, really bad today in current frontier AI models. And it's actually a very interesting point. Frontier AI cannot do a simple thing as creating presentation unless you create the scaffolding around it to do it for it. I guess that's true for many, many, many, many services.

Fredrik Olsson: 2:36:18

Yes, yes, it is. Yeah, exactly. Imagine we had this conversation three years ago on Chat GPT. Was it three years? Two years ago, three years ago.

Anders Arpteg: 2:36:27

November. 22 November in 2022.

Fredrik Olsson: 2:36:30

So the level of expectations has increased healthily. Yeah. Yeah. At least.

Patrick Couch: 2:36:36

But remember, I mean, I think this is yeah, exactly. It has certainly increased. But I also think it's interesting in terms of the means and the ends. Because you're eyeing the end goal. Like you want to commit something, it's called the means to that is a presentation. And you're going like, hey, if you create a presentation use this technology, you need to take actions in terms of clicking this and getting the shapes and blah blah blah blah blah. No, but listen, I I think this is interesting because there is a semblance of proficiency rampant in LLM tech today. So behavioristically, it can drive a car, it can do a bunch of stuff. Does it understand the world? Does it have judgment? No, of course, obviously not. It's math, but it can still behave interestingly and efficiently in production. However, when you go very detailed and say, hey, create me a PowerPoint product uh presentation with slides and the rest of it, you're sort of expecting something that RPA, scraping, could do what, 10 years ago? And we can't do it with the Frontier Mobile.

Fredrik Olsson: 2:37:40

But but again, the presentation, the purpose for it, what is it? What would you achieve? No, I understand. But I mean, I mean the point is, good AI, you wouldn't need a presentation. You would achieve the means.

Anders Arpteg: 2:37:53

No, okay, let me try to phrase it again. What I'm trying to say, the point is that AI is surprisingly bad in certain actions. Yes. Agreed. And even in a simple thing like creating presentation, you would assume AI is really good at that, but it's not. Unless it's cheating by either having scaffolding around it, like a gamma.app does, or creating images without really doing it the way humans do. If AI was really good, which is probably will be in three years when computer use will improve, it actually can create a simple thing like a presentation in whatever way it chooses, but it it could do it. Today it cannot. And I think it's such an important lesson for companies to understand. AI is actually surprisingly bad at a lot of tasks. Yes.

Fredrik Olsson: 2:38:38

And in in this particular case, could it be that we're looking at language models and next token predictors and not really aesthetic image creating whatever that we're having? Yeah. So I mean but it's it's not really about creating presentation. It's about you know I understand, but it's it's it's um like subjected to different kinds of uh modalities of that data and so on.

Anders Arpteg: 2:38:58

And taking actions in general is hard. Yes, yes. It's good in coding, but for most other tasks it's actually surprisingly bad.

Fredrik Olsson: 2:39:03

Yes, because the loop in coding is really tight. You get the feedback so yeah.

Patrick Couch: 2:39:07

But but let's just uh dig into something that you said, which I think was touched upon earlier as well. Computer use. When any I can use computers, yes. So elaborate on that because I think that is interesting to me because that really hits the uh the the like the interface between humans and machines is the computer par excellence. I mean, that's where we interact.

Anders Arpteg: 2:39:27

It will be a big shift. I think in 26 it will improve. We know that um Gemini 3 Flash was just released like this week or something. It was specifically trained for computer use. And it's for a good reason because you know, once you get proper computer use capabilities in a model, you know, what they can do will be amazing. Then they can start to take action properly, and that will be very, very valuable. Like the case today.

Patrick Couch: 2:39:48

No, it's not. But with computer use, you're you're sort of implying use of uh the OS. Keep it and mouse. Um yeah. Keep it a mouse too. Same thing. Interacting like an API to the operating system, basically.

Anders Arpteg: 2:39:58

Yeah, not an API. Proper keyboard and mouse, human kind of interface.

Patrick Couch: 2:40:04

Through the OS. Yes. Yeah. Yeah.

Anders Arpteg: 2:40:07

And you would think that that would be easy for an AI. That is so smart to answer any kind of question, but it's not, because what AI is good at is knowledge management. But it's really bad in knowledge and action taking. This is exactly the proof for that. It will change in coming years. But today it is. Okay. Yeah, we'll stop that.

Patrick Couch: 2:40:27

Anyway. No, but I think that's the that's the perfect point. Like computer use. Like if you can get the AI to take action in the physical world, all Jensen talks about is physical AI. But the the the smoke and mirrors thing is the robots and the humanoid stuff. But that's not where the not only in humans interact. It's the computer. So I think a much better bridge between or into physical AI is through the operating system. Because the operating system has always been where you have software hit hardware.

Anders Arpteg: 2:41:01

And it's similar to, you know, you can think about why are you know Tesla building Optimus bot a humanade, humanoid form of a robot? It seems like not very practical, but it is. Because it it is going to be able to operate in a physical world which is built for humans. You can argue exactly the same for the digital space. If you get an AI that can operate in a digital space built for humans, it will be extremely valuable. So just as they are investing in humanoid form, uh it is because it will be super valuable to be able to operate in a world built for humans, just so it will be for digital space, which is built for humans. And once an AI can operate in a digital space by doing computer use, it will be very valuable. This is what they are all training and trying to get Gemini to do. They're trying to get GPT to do it. It's still really, really bad, but it probably will take just a few years before it actually will succeed, and that will be a really, really big breakthrough.

Jesper Fredriksson: 2:41:59

Yeah. Yeah, I think probably already in 2026, I think we will see uh good computer use arrive.

Anders Arpteg: 2:42:07

But it is surprisingly bad today, and I think it's it's a very in profound in you know yeah conclusion.

Jesper Fredriksson: 2:42:13

I had I had one task that I tried uh Atlas to do, the uh Chat GPT uh browser and the agent in it. Um so I had a like a large uh flowchart of of a system that we one of the one of the products we have at work. So it's uh it's relatively complex. So I wanted it to just move through the whole uh drawing because it's uh if you zoom it out so you can see everything, then it's unreadable. So I I just wanted it to uh zoom into various parts so that it could document the whole process from like a flowchart. And I had high hopes for it when Atlas arrived, but uh it just couldn't, it could it tried a lot. It tried to click on something, it tried to double-click, it but it's it's not there yet. But I it seems like it's not so far away from completing that task. I could see it almost succeeding.

Anders Arpteg: 2:43:06

Yeah, agreed.

Patrick Couch: 2:43:06

I think that is super interesting area, like that whole like the level of abstraction and the human-machine interaction. I think there is something there that is worth poking at.

Jesper Fredriksson: 2:43:18

I think of this as the the last mile AI, like uh the the last mile that is needed for for AI to really be able to do things. And and there I agree with you that it's uh it's uh not very good at at doing what humans do in that respect because it's obviously not trained on doing that.

Anders Arpteg: 2:43:36

Yeah, they're trying to train it, though, but yeah, yeah, but it's failing.

Jesper Fredriksson: 2:43:39

It is hard for it, of course, to read everything, to take a picture of the of the of the screen and zoom into it and then take an action and then take the next action.

Anders Arpteg: 2:43:48

But all of them are trained on it, but it's still really bad.

Patrick Couch: 2:43:51

Yeah, and I think the the frontier will be to solve that in a non-human-centric way. Like I think, I think there will be a like a roundabout way of addressing it because that's usually how we do. Like, but if you think about you know the great surrealist poets pointing out the fact that, hey, if you want to create a robot to clean your dishes, you don't create the human robot and and and you create the dishwasher and all these other things. Like there are you know, you saw that uh there were earlier this year, if you reminisce about 2025, there was this uh uh showcase going viral on LinkedIn where you had these two agents interacting in the hackathon proof of concept thing, and at one point they realized they were both bots and they switched to this other protocol because it was they have to go through the NLP you know interface and all this stuff. And I think there is something there to be pursued because we are obviously creating the computers to make us more efficient in manipulating the machine. But if you want the machine to manipulate the machine, you don't really need the five sensors of the human body all the time. In coming years, it's actually kind of efficient to have a computer between them and the dominant model in physical space manufacturing space, Henry Ford manufacturing it's the human. The human is the blueprint, right? Yeah, but so I think that's good for that. Yes, and I can see that being like ephemeral, like a it'll be a duct fluge or something in the right. It'll be a period of time and then eventually it'll just push through. Yes.

Anders Arpteg: 2:45:25

But it's a super interesting.

Patrick Couch: 2:45:31

100%, right?

Anders Arpteg: 2:45:32

And so that boundary blurring is interesting to cause left of the year next year. Anyway, should we try to close to three hours now? So we we should really try to uh um yeah, close down here. Okay, should we end with a normal question here? Um you can try. Yeah, should we do a normal one? Yeah. Okay. Um if we uh assume that an AI can actually create a PowerPoint percent, what color would the background be? Okay. If we assume that an AI would um actually become AGI at some point uh in five, ten years or whenever, or two years, one year, yeah, what do you think the impact on society would be? Either it could be the horrible kind of dystopian matrix and the terminates of the world and AI is actually going to try to kill us all, or it could be the other opposite of AI actually solving all the challenges we have in society, curing cancer, fixing the climate crisis, fixing the energy uh demands we have, and we live in a world of abundance where all the costs of goods and services go to zero. Where do you think we will end up?

Fredrik Olsson: 2:46:44

We will uh end up with a much, much stronger status quo. Okay, elaborate. Because uh yeah, why would it change anything? What's the what's the what's the reason for it to change anything? What what are why we we humans are not very good at issuing new stuff needed to do PowerPoints anymore? And you can ask an AI to do it, but that really changed a lot. That would change a lot, of course. Uh, but I guess what what's what is an AGI? Is it like a uh 12-year-old person eager to learn, or is it like a fully fledged professor knowing everything?

Anders Arpteg: 2:47:20

I like the Samultan definition saying an AGI will happen when we actually have an AI that can be on the par with an average co-worker, human co-worker. So is an agent, so it has like uh tools. It should be able to replace an average coworker with tools, even creating PowerPoints.

Fredrik Olsson: 2:47:40

With a hammer and a saw, yeah, yeah. Yeah, but I I think I think uh again uh we will see uh um at the extremes, there will be like some parts of the society will change, it will be very, very unevenly distributed. This this AGI stuff. Yeah, it's like Area 51 didn't happen for it. So I think it will kind of enforce status quo. Some people will protect or their interests very, very, very furiously, and someone will uh do this other stuff. Um depending on on the distribution of this AGI again. Yeah, so that's a hedging answer, that's a non-answer. Uh, if if I of course I don't think the dooming thing will happen. Because if if there's it's kind of um meaning to approach us, we haven't we have not we have yet to kill each other completely. Patrick, what do you think will happen?

Patrick Couch: 2:48:33

Yeah, so I'm torn, I realize, between my inner hippie and my inner Darth Vader, and they're always constantly you know at each other. But uh I must say I think I have hope in my fellow human being, and I really think that technology is a mere fallout of you. We talk about artificial intelligence as if it's not completely natural for human age to excrete technology and then the derivatives of it and the rest of it. Everything's natural, everything's biology, everything is amazing. And I really think that the more people we can invite to co-creating the future, the more interesting it will be, probably the better it will be as well. So I really think that to your point, I mean, yes, I'm all about regulation and the slowing down the pace, but if you really sort of put me under pressure, I will just say I'm an anarchist, I have faith in my fellow human being, just seed everybody, feed everybody, give everybody access to everything, open everything up and see what the human apes do. And some of one of those apes will surprise you with something amazing. It only takes one. Like just one guy who does something, one girl who does something, and it's gonna be like amazing, what did you just do? Remember when we spoke about earlier, like we only need 10 bright people to rule the world in terms of technology? Well, you know, those are out there. So we need to locate them. So I'm I'm super optimistic. I can't see this go wrong. I don't think people that that look towards technology or secondary means as a gauge towards where the planet is going, or is they're not looking at the right place. You have to look at your your inner moral, your all these soft values, the society, and how to bring out the best in everybody. Nobody is born evil. I cannot imagine it, right? So you want that little baby born to be empowered and be connected and surprise us. Although he's so optimistic. I hope you're right.

Jesper Fredriksson: 2:50:31

Yes, but uh I I like the uh um there was a vlog from Sam Altman this year uh called The Gentle Singularity. Uh, and in that he talked about what it would be if you would encounter a subservience. What was it? What is it called? Like somebody living uh uh at a low existency level. Uh I'm blanking on the on the English term for it. Uh a farmer living a thousand years ago. Uh, if you would talk to him today and uh tell him about our society and uh how we live, what we do for work and uh what kind of resources we have, he would say that we're living already in a world of abundance and we don't need to work, and we would think our jobs are bullshit. Uh and uh I think that's such a powerful image because I think that's that's what will happen. Uh so it's already happened in a in fact. We're already there if you if you just zoom out far enough, and that's probably what will happen uh afterwards get to it.

Fredrik Olsson: 2:51:43

Yeah, good answer. But what about you? Yeah, what do you think?

Anders Arpteg: 2:51:47

Yeah, okay. So my answer is is the same as as usual, which means that um I'm positive long term, meaning I I'm afraid that people will abuse AI and I will not feel safe until we have AI that can supervise AI. Before we have that, we will have people abusing AI, and that will happen because AI is too stupid. So it will do what the human tells it to do, and uh independently if it's a state actor or if it's uh terrorist or if we if it's a cyber criminal or whatever. And and that I'm really scared about. I'm really scared about the coming years where people can abuse AI until we have an AI that can supervise what people are trying to use AI for. So I'm hoping we will survive until the time that we actually do have uh AI supervising other LRS AI and then it will feel safe. But today, uh that's why partly I'm doing what I'm doing today. I want to ensure that we actually do have uh safe control of people abusing AI.

Fredrik Olsson: 2:52:52

That's my answer. I have a comment on that. Uh I think the problem is not AI, it's technology, and we have dumb technology that has made the society what it is today. I'm thinking of recommendation algorithms on, for instance, Twitter. Working for different means than uh society is good, more for engagement, and I think that's a big part of where we are today with polarization and stuff. So, I mean, uh AGI could be like an extension of that, and then I I share your fear for that. Yeah, cool, cool.

Anders Arpteg: 2:53:24

I'm afraid about stupid people. That's in short answer.

Patrick Couch: 2:53:27

But there are a lot of them.

Anders Arpteg: 2:53:30

Cool. Thank you so much for this awesome discussion. Um great to have actually a set of friends here who we actually can go into debate with as well. And I hope no one into too much offense, even if going into some heated discussions at some points. Uh at all. So fun. Thank you so much. Have a great evening. Bye bye.