E148 - AI Update: Grok 3, DeepSeek, Embracing AI - Robert Luciani & Louise Vanerell

Episode 148 of the AIAW Podcast is here! This time, we welcome back Robert Luciani, AI wizard and CEO of Negatonic, alongside Louise Vanerell, CEO of Recursive Future, for an in-depth discussion on the latest AI breakthroughs. We explore DeepSeek, Grok, and other cutting-edge research, unpacking their significance for LLMs, AI development, and business applications. Key topics include how AI is powering national Swedish TV (SVT), the Grok 3 release and xAI’s rapid advancements, Europe's role in the AI race, the future of open-source models, and unlocking AI’s full potential today. We also dive into groundbreaking research on test-time compute scaling, the impact of JPEG and latent space reasoning, the evolving landscape of software engineering, and the future of AGI—will it lead us to a utopia or a dystopian reality? Finally, we discuss what companies should do today to stay ahead in AI’s fast-moving evolution. With Robert’s sharp insights and Louise’s strategic perspectives, this episode is a must-listen for anyone passionate about the future of AI and its real-world impact.

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

Chapters

• 0:00 Exploring AI for Non-Tech Crowds
• 9:12 Navigating the Future of AI
• 18:55 Self-Service AI Platform Development
• 24:52 Empowering Business Owners in AI
• 29:28 Advancements in Grok 3 AI Model
• 42:53 Innovation and Speed in AI Training
• 51:27 European AI Investment Strategy
• 58:35 The Future of Open Source AI
• 1:12:33 Future of AI Parameter Expansion
• 1:18:37 Europe's Strategic Plan for AI
• 1:22:29 Fostering European AI Competitiveness
• 1:25:50 Upskilling for AI Adoption and Leadership
• 1:33:50 Latent Reasoning and Recurrent Depth
• 1:41:25 Efficiency and Innovation in AI
• 1:56:02 Neuralink, Latent Space, and Future AI
• 2:00:15 Future of Engineering AI Platforms
• 2:15:00 Future of AI and Human Intelligence
• 2:24:03 Preparing for the AI Trajectory

Transcript

Anders Arpteg: 0:00

A bit what AI is. Is that what you're saying, or how you can use it, or what?

Robert Luciani: 0:05

Basically what AI is and then how you can use it if you're a developer. But it didn't sound like a developer crowd. No, it wasn't a developer crowd. I think there were technical people in the audience. One of the persons I spoke to afterwards was some kind of technical director for government function. In any case, you know, I started maybe by polling the crowd, trying to figure out what they knew and that kind of thing, and I asked how many people here have heard of Lama? And nobody raised their hands and I was shocked.

Henrik Göthberg: 0:39

But I can give some context because I was there. This is an exhibition which has nothing to do with the data and AI. It's literally for marketing, sales and for meetings and events. So it's like an expo of how you go and stay at Strawberry or go to this meeting event or use this CRM app.

Henrik Göthberg: 1:02

And in this context, then HP had a talk. They were one of the main sponsors and the main storyline is, of course, then you were there together with a product manager to try to explain to normal folks what is an AI PC and why do you need a more powerful machine now in the future. So the context was really about okay, even if we have everything in the cloud, there's also this fundamental topic that you're not going to be able to move all the data to the cloud. Maybe and you actually don't you want to do some work on the machine and then send it to the cloud and then send it to the cloud. So she was giving the overview and basically positioning. There is a new era of PCs coming. Be aware when you really want to work with AI models, even if you're a user.

Louise Vanerell: 1:53

Isn't it fascinating that we've come to a point where that is a central topic on such an event and there is room even though you maybe not got the response you were expecting from your Lama question but that we can have topics like that on a venue or event like this?

Henrik Göthberg: 2:13

I think it goes to hey, this is marketing folks and they're supposed now to do more and more stuff, potentially with AI, in their creative process, and they are stuck and sitting there with machines that is not there. They've been sold. Everything can be done in the cloud. And then, hey guys, if you really want to use this stuff, photo, whatever it is, at some point in time, your creative flow will be disrupted because you need to wait for too long.

Robert Luciani: 2:39

But there's a really nice parallel here with when smartphones started grabbing a hold, where it's both a matter of you getting the benefits of what smartphones offers compared to brick phones, but also that all of your customers are there. So if you're ever going to interface a customer with an app or a website or something like that like who here doesn't have a website that works on a mobile screen? It doesn't work anymore. You have to have one, and it's the same thing with AI. You both need to know how to use it. But also because that's where your customers are going to go, and we tested it by asking a question to this CEO woman that I met. You know, somebody wants to find a new office venue. They're going to ask a chatbot and the chatbot will answer and send some links and she's like why isn't my company listed there among the links that the chatbot supports? You know, google used to be the way to do it, and now my customers are trying to find things this way. Maybe I need to be savvy on this stuff.

Henrik Göthberg: 3:34

So we're talking about a search S-O-E going into A-I-O-E or something like this. I saw this.

Robert Luciani: 3:43

You just need to be aware of what's going on.

Anders Arpteg: 3:44

I think the future will be similar to smartphones, that everyone expects that a smartphone should work simply.

Robert Luciani: 3:50

And.

Anders Arpteg: 3:50

I remember I gave some kind of math advice to a friend or family and she was trying to struggle a bit to solve the math equation and I simply took a picture with the phone, sent it to ChatGPT and asked it to explain the problem and give a solution. And it did. And I said isn't this amazing? The AI can understand the text to solve the math. For me it's amazing. And she was what do you mean?

Robert Luciani: 4:19

That's obvious. There's nothing strange with that.

Anders Arpteg: 4:21

AI is supposed to do that. I use Snapchat every day. That's nothing strange For them. It's obvious that any kind of product that they use should have intelligence in it and it should be like chat, apt level. Otherwise they're very surprised.

Henrik Göthberg: 4:37

But for me it was interesting to be there. You know, chairman of Data Innovation Summit. We go to all these conferences and we can all nerd out right and here summit and we go to all these conferences and we can all nerd out right, and here you meet a normal crowd, normal people, normal people not, not not in our echo chamber, so to speak.

Henrik Göthberg: 4:51

Yeah, and it the the communications or knowledge gap is so big. So the stuff we are talking about here and we are throwing around llama and tropic blah, blah, blah, you know, they don't even know it exists. Yeah, that really shocked me a little bit. But at the same time, why would they, you know, why would they care? No one cares, I guess yeah, but I was um.

Louise Vanerell: 5:13

I was on a podcast a couple of weeks ago with um. You know completely different branch of people and knowledge, but they're into recruiting and executive search stuff. So of course they're more and more meeting these kind of questions you know, what do we need? What does it actually mean when we say that we want somebody with this kind of experience or knowledge or role or whatever it is? And I thought they started off really good because they asked me you know what is AI? And I gave this example.

Anders Arpteg: 5:45

I'll take your answer.

Henrik Göthberg: 5:45

That's a good one.

Louise Vanerell: 5:49

I realize I'm in the fighting pit there, but I tried to give as easy as possible answers when it comes to how to see what it actually can give us, and I gave them the example that if we see machines as something that can help us overcome our physical boundaries I mean, us around the table cannot lift like three tons of something, but so we built a robot or something that can do that for us.

Louise Vanerell: 6:17

So we build something that can actually make sure that we're better physically. So they do stuff for us and we're so used to the fact that we can get physical help so we don't see that as something that sort of interferes with our you know, our core of being a person or human. But nowadays, when it comes to AI and I that was the other part then that this is something that can help us or, you know, help us overcome our sort of what we don't do as good when it comes to our cognitive abilities, and that's where AI comes in and just the fact that we're not so used to the fact of thinking that something else an item can help us think.

Louise Vanerell: 7:01

It can sort of, you know, make us better cognitively or intelligently, or however you want to say it. Um, and I mean, it makes it fairly easier to understand and it's also maybe easier to then look at the thing that. Why do we see it, as you know, potentially a threat? Why do we think that this, this thing that thinks and maybe thinks better than we do, why is that something that we feel is, you know, oh, we're into human turf here.

Robert Luciani: 7:29

That's scary for a lot of people right, yeah, it is.

Louise Vanerell: 7:32

But I've been getting feedback, which is nice, that that was an easy way to see it, especially when you're not used to having, you know, the very depth of conversation which you are accustomed to, and that our little echo chamber.

Henrik Göthberg: 7:47

But it's also. It's also clear about the echo chamber that in reality sometimes we work in a large enterprise and we are trying to help them right, and we might be talking to the tech people and we might have a fairly good conversation and we understand what we're doing, but in reality that same conversation, where the marketing director doesn't know what LAMA is, will happen everywhere, from the large enterprises to the medium science companies. So, to excel in this communication and I thought you did a pretty good job I thought who was the woman from HP you share the stage with?

Robert Luciani: 8:22

I think she was the chief product.

Henrik Göthberg: 8:26

She was the chief product officer Chief product officer for this new line of products and I think she did a fairly good job in connecting with it. And then I like the idea that you were trying to also do the demo For them. What I think is going to be a leap is that in a couple of years, they will build their own agents. They haven't figured, they haven't really realized that. So now they're feeling okay, are we going to use chachi pt, we're going to use these products and like, actually, you might likely be able to string together an agentic workflow yourself. Yeah, and then you even more need this type of machine and this one is completely out of they don't. They are not there, but it's so obvious it's going there, I think.

Robert Luciani: 9:11

I mean, it's more than obvious. I saw a post on LinkedIn of somebody or perhaps it was Reddit and then reposted on LinkedIn, as is usually the case. Somebody was complaining and saying my Python project that has 30 files in it isn't working well in cursor and the AI isn't able to track all of the stuff across the files. That's pretty damn amazing that you can have a project with 30 files in it and the AI is still sort of keeping up and then sort of breaks down. Well, you know, we're a few years into the sort of LLMs that are able to do this kind of stuff. It's only reasonable to expect that it's going to grow in the next couple of years. What do you think? I mean, no, people don't work with projects that have hundreds of files in them. These are really really senior developers that do that kind of stuff. Most people don't do that, and it's cool that AI is almost there today.

Henrik Göthberg: 9:56

Yeah, and then it comes to this whole. I was thinking and I was talking a little bit with Albert at HP how this whole journey has gone. Everything is on-prem, and then the pendulum flips over or everything is in cloud. And you know, the last couple of years has been like oh, everything is about the cloud, the cloud, and people even stop forgetting about the physical edge now getting adapted to have AI flows in it, and maybe we're not even going to call it an AI PC in five years time because it's going to be so ubiquitous. I don't know. But it's quite clear that we're not going to send every single thing we do to the cloud. We're going to do something local and then send some stuff. I think that's a clear trend here. I don't know.

Robert Luciani: 10:56

It's a shame that there's a lot of AI fatigue out there. Everybody's hearing about AI and they're like I hear about it. I don't get it and it seems like it's a bunch of hype, and when I tested the model myself, it was no good Two years ago. I'm not sure exactly what to do about that, but sometimes it's nice to just stay in the bubble and enjoy it.

Anders Arpteg: 11:16

Yeah, tough discussions speaking about and handling and managing, I think, the AI hype that you know some people are. Of course, there is a true value and it's amazing the progress that they have had in recent years and over the last 50 years as well, but still people are so ignorant about what it can do and cannot do and believes it can do so many more things it actually can.

Henrik Göthberg: 11:37

And then we have people that are and we are truly, and that is really bad that we have this kind of people that are not really managing that properly.

Anders Arpteg: 11:47

Yeah, annoying.

Henrik Göthberg: 11:48

But I think we also need to have respect that we are actually in our own echo chamber. We really need to have respect that we are outnumbered here. The gap is not only a geopolitical gap AI divide it's a microcosmos divide between people who are working with it and people who haven't really touched it, and we need to respect that. Yeah definitely.

Louise Vanerell: 12:12

I mean, that's a huge. We've been talking before, I think, about the value, about science communicators, when you're able to make something that is either perceived as being very complex or that is complex in its own, and to be able to bring something that is either perceived as being very complex or that is complex in its own and to be able to bring that to people so they don't feel like it's super hard or, you know, not for me, I will never get that it's you know I can never understand those things.

Anders Arpteg: 12:37

I'm a great communicator that can actually speak about complicated things properly, Like Richard Feynman, you know, in the 1940s speaking about quantum mechanics you know, in a way that was so engaging and it's very few people that can actually do that.

Henrik Göthberg: 12:49

But I remember from this pod you said something. Maybe it was you or someone said it to you. So how do you define who's the best data scientist in the team? And then, basically, the bottom line is well, engineering wise, you know what.

Louise Vanerell: 13:04

We can argue back and forth, but the one that has biggest impact is the data scientists that can communicate with the rest, and that is communication again, and I mean maybe that then applies to the whole field, and if the three of you are, especially, then representatives for this field, of course there's a responsibility to make sure that you don't make it more difficult or harder to get close to and that you don't have a point of alienating people, but so are you, I would love to welcome you here as well.

Anders Arpteg: 13:39

Thank you, louise Van Rael, you are actually an expert in this. I would say Thank you. How to work with leadership, navigate the future of AI, and just how to keep up with the amazing progress that we're seeing and you have a company called Recursive Future. Right, yeah, I thought of this. What do you do there?

Louise Vanerell: 14:01

Well, Recursive Future works with precisely these things, which is, how do we make sure that the people working with these technologies and the people who are working closely to them actually have the knowledge and the capacity and the skills to be able to work with these kind of tasks? And I love to sort of bring it from not only talking about you know, what do people have to know, but rather what do we want them to be able to do? And as soon as we have that thing more figured out, then it's a lot easier to actually set up to make sure that people get that kind of knowledge. So that's what we worked at All the human factors regarding the technology that we are seeing the revolution of, so to say.

Anders Arpteg: 14:43

And dearly needed, given what we just spoke about as well before. So kudos for that, I think. Thank you. I would love to welcome also an old friend here, robert luciani. He is annoyingly a very multi-talented person that is both an expert programmer, julia lover of course an ai.

Robert Luciani: 15:04

Wizard this is the year of Julia this year is for real 2025 it's just around the corner.

Henrik Göthberg: 15:15

Mikkel Kinval always says that I'm working with. I can't believe Python beat Julia. Julia is so much better. What if the whole world went my way?

Robert Luciani: 15:24

let's pin that. We can get back to it. I have a feeling it's going to show up again.

Anders Arpteg: 15:29

But also, of course, an AI wizard. I think you call yourself actually one of the most knowledgeable people in AI that I know, so very happy to have you here as well. Awesome, as always, and it's going to be a bit of a special episode, I think we're we have even less prepared than usual, which is hard because we don't prepare, we prepare, we do.

Henrik Göthberg: 15:49

We have our themes, we have our questions and the guests have them beforehand.

Anders Arpteg: 15:53

Now we have topics yeah now, but we're basically planning to speak a bit about the news and have a proper after work sitting a couple of friends and speaking about the favorite topic that we have.

Henrik Göthberg: 16:06

Yeah, and in my way the news wheel of new things coming out is spinning so hard, so we usually have a news section that we do like 10, 15 minutes, and then we've had we did it last year with Jesper, like when we oh yes, yes, we used to take a full episode and dissecting some of these news, and I think we are right at that cusp again with it with you know, then last time there was the star, uh q star, and then all of a sudden we have a one coming out and now, yeah, after it was so annoying the day after you had it.

Henrik Göthberg: 16:39

You, you were predicting the day after it was released. And here now we have, you know, everything that has come out now with grok, with the deep, seek, deep research, and we have. We, we really have seen now the, the new breed of reasoning models, and we all had maybe some chance to play with it and all that. So it's a time to reflect on where we are right now again.

Anders Arpteg: 17:03

And more stuff coming up very shortly. You know Sam Oldman is announcing GPT 4.5 around the corner, so we can do the Q store, we can go back to the.

Henrik Göthberg: 17:12

Probably tomorrow they will announce 4.5. If we predict it today, they announce it tomorrow.

Robert Luciani: 17:16

Yeah, we make it come out tomorrow.

Henrik Göthberg: 17:19

This podcast.

Anders Arpteg: 17:20

If we predict it today, it comes tomorrow right, but also before we move into some of these kind of most interesting news and perhaps some research papers and whatnot that we can discuss. I heard also, Robert, that you have a new adventure as well.

Robert Luciani: 17:33

Yeah.

Anders Arpteg: 17:34

What are some new exciting events happening for you?

Robert Luciani: 17:38

There's always side gigs going on in the background, but I have settled with one of my customers and taken a job there, and that is the Swedish National Television SVC. I'm very happy to be working there because everything they have is content which is great for AI people it's pictures, it's text and everything else. And it's a small team and, as you know, small teams can achieve a lot, and we have a very clear vision of wanting to be able to jump ahead really fast and it's possible, so I'm excited.

Henrik Göthberg: 18:09

I heard you had a cool job title.

Robert Luciani: 18:12

Yes, I'm the only person there called AI Wizard.

Henrik Göthberg: 18:15

So I'm very happy about that that was actually part of the discussions for getting the position. I'm not going to get on this job if I can't have AI Wizard on my list.

Robert Luciani: 18:25

Yeah, that's what I was. They're like all right, we'll make a concession this time awesome it's so cool you should wear that with pride.

Louise Vanerell: 18:35

So can you elaborate a?

Anders Arpteg: 18:35

bit more.

Robert Luciani: 18:36

What are you going to work with, or yeah, for sure I mean, they've been pretty advanced um for a long time with certain types of technologies, from streaming really high bandwidth, you know, because they have a mission that they have to work on all devices to everybody all the time and be super reliable. So things like translating between languages, archiving things and other things. They've been doing AI projects here and there and so they have very many teams that have a little bit of AI skills spread out across the company and even recently, you know one guy that just joined my team he was in another team earlier is a super expert at comfy UI and they've trained Laura's to do you know the things that they wanted to do, but with Swedish landscape or in the style of their own drawing artists and that kind of stuff. They've trained Laura's that make sort of not thumbnails, but you know kind of arts that they need for shows and stuff like that, and so now what they need is a platform that enables automation of this kind of stuff and scale out and self-service and a lot of kinds of things like that.

Robert Luciani: 19:48

And I have a bit of a grand vision myself, which is that you have one chat interface for everything and you should be able to ask it a question and it returns custom CSS where you can fill in your parameters and then it'll generate a picture for you. So you don't have to be an expert in Comfy UI. But the Comfy UI person can set up like a workflow. So we do need to enable agentic type flows of events.

Anders Arpteg: 20:15

Can you just elaborate what is Comfy UI? So?

Robert Luciani: 20:18

Comfy UI is a user interface for image mostly image and video models and it's quite complicated just because sort of the domain of diffusion models and all these things have a lot of parameters. So the UI looks really neat. It looks like you connect cables between all the inputs and outputs of these models and it can really look convoluted and you can download some pre-made configurations. But you know, this guy makes his own. But it is for generating images. It's for generating images and video. So if you have a picture or a video of a landscape and you're like I want it to look more like northern Sweden, you can run it through a pre-made workflow that he has made that will remake it into Northern Sweden style.

Louise Vanerell: 21:07

Stuff like that. So is this then to be used sort of instead of stock photos and stuff?

Robert Luciani: 21:11

Instead of stock photos. Also, for example, they have artists that work full time on drawing things, and they have drawn tens of thousands of pictures from courtroom proceedings.

Anders Arpteg: 21:22

Oh yeah, of course they don't want to draw more pictures like that.

Robert Luciani: 21:24

And that style is so uniform that creating a Laura which is like a custom little model.

Henrik Göthberg: 21:29

And what does Laura stand for?

Robert Luciani: 21:31

Low rank adaptation. I think Low rank adaptation that's just a way of sort of nudging a model in the direction that you like, and so this nudges the model to turn every photograph into a drawing in the style of these artists. So instead of having to ask the artist to spend an hour making a drawing that they hate to make, you can just use the Laura to take a photo and turn it into a drawing like that, and so we want all of this kind of stuff to be self-service instead of having to ask our comfy UI expert to do it by hand every time. So what we need to build is a system where people can help themselves, can manage their own prompts. To build is a system where people can help themselves, can manage their own prompts, can manage flows of prompts, and build agents and that kind of stuff.

Henrik Göthberg: 22:09

I love this vision and it resonates so well with with on a much deeper level that the way we really create autonomy for people using AI is to build products is to build, ultimately, self services.

Robert Luciani: 22:23

I just have to share one anecdote, which was basically on the first day that I was there.

Robert Luciani: 22:27

I got to meet a lot of people and I got to meet this prize-winning journalist woman who you know she's an older woman now because she's been around for such a long time won lots of prizes and is a good journalist, right, and first of all, everybody's very excited about AI at SVT, but this woman in particular was so creative with the way she wanted to use it and you know, you can imagine, let's say, you have a RAG set up with a database of articles and stuff and she's writing an article about something and she needs to know something in order to build an argument in the article she's writing.

Robert Luciani: 23:00

She can ask way better questions to dig up information than I can because of how good a journalist she is and she's like you know what. I bet this is relevant. I'm going to ask this question. So the way she imagined using the RAG system is way more sophisticated than what I would be able to imagine as an AI expert, and I think that's so cool that you see the domain experts being the ones that show you this is really how we want to use it. I thought that was pretty neat. So I find that it's always the people straight out of university or the oldest, most experienced people that have the best ideas when it comes to AI stuff.

Henrik Göthberg: 23:34

But it's also so clear with the self-service thinking, the co creation between the domain expert and the engineer. This is the power Anyway.

Anders Arpteg: 23:45

I mean also. I certainly subscribe to the idea that if we can make it self-service and truly enable non-technical people to make use of AI, we can scale up the value of AI so much more, because we usually have so many more people that are not super experts into AI. And I actually heard you mention something very interesting in Tess Vitti as well, which is something like and please correct me if I'm wrong, but something about the CMS for prompts or something.

Robert Luciani: 24:10

Yeah, this was actually an idea that they had before I even joined. You know, they used the chatbots to draft things and to double check, if you know, if they hadn't maybe sourced things correctly, and so they had sort of workflow ideas and so what they had done was a lot of work on prompts before the user asks their question. So it would be a long prompt and say I want you the language to be like this If the question is like this, then do that If and so on. So like long pre-prompts.

Henrik Göthberg: 24:41

I guess Like system prompts.

Robert Luciani: 24:42

Yeah, exactly Like system prompts and so like long pre-prompts, I guess Like system prompts System. Yeah, exactly Like system prompts. And they had generated so many of them that now they have a prompt content management system, a prompt CMS. So cool.

Robert Luciani: 24:52

And there's a business owner on the business side. That is not a technical person, that is like the AI product manager for the chatbot and his job is to manage the prompt CMS and talk to the business and say are you getting the responses that you want? Why is it good? Why is it not good that? Person is so mature that person is responsible for if the service is good or not. We have no idea. We can see if it's up or not or if it's many yeah, but he content wise, qualitative, yes.

Louise Vanerell: 25:20

Yeah, that's a better view on this like you and I have been discussing a couple of times like it's so important to see where the value comes from and having somebody in that position take that sort of responsibility, of course, sort of make sure that you get the value from doing this kind of works and having that sort of setup in place.

Henrik Göthberg: 25:40

But this is also interesting now, like there's a lot of people now they're putting on co-pilot. They this is also interesting now Like there's a lot of people now they're putting on Copilot, they're putting on ChatGPT, they're making an enterprise approach to it and they give you some basic teaching on how to do this and off company goes. And here, oh, come on, this is going to be Excel hell all over again. So here now I'm so glad to hear someone sort of having a framework idea, a product idea, and then thinking about how can we steer this in a responsible way? Super cool.

Robert Luciani: 26:12

One thing we probably think about as engineers are like evaluation metrics how can we check if our deployment is performing as expected? And we try to come up with things ourselves, and there are these pretty clever ways of doing it. And we try to come up with things ourselves, and there are these pretty clever ways of doing it. One of my favorite is my old colleague, liang Feng. He said you can ask a question from the model and then see if the model can infer the question back from the answer, and if it can do that, then it's a precise answer and not a vague one. You know, that's just one kind of metric that maybe is important or maybe not, it doesn't matter. And so engineers will try to come up with metrics for whether the models are good or not, but really we don't know. It's only the business people that know whether it's good or not, and so it makes more sense to have a business person managing and sort of being in charge of whether the service is good or not.

Anders Arpteg: 27:02

I really like that and this is what AI really enables to a higher extent. You don't need to have a super technical skill to make use of technology in some sense. Yeah, yeah.

Henrik Göthberg: 27:12

But there's a profound back depth to this, because we have anticipated that when you look at using AI and AI agent, it will fundamentally rebalance the workforce. And a simple way of saying it well, you will need people who are managing the AI, simply put, and then, oh, what the fuck does that mean? Here we have a very concrete example that we can't all be journalists. So within the journalist and within deep domain know-how, someone is now taking on the CMS or product owner role for the prompting. So is he an AI engineer? No, but is he way more advanced in AI than the journalist? Who's the user? So we are now getting a rebalancing of the domain teams versus the tech teams, and in the beginning maybe those guys need to be fairly techie, but the more we are better in the platform teams to build self-services, it will shift. And this is a major impact, because if you're not thinking about this rebalance, it's going to be troublesome when we think we're going to have exactly the same profile, balance of teams and not really fixing this.

Louise Vanerell: 28:17

But tying back to what you said earlier about who's the most valuable AI engineer on the team yeah, the one who can communicate, because obviously that's where you build the bridge to somebody. Who is that person, like in this case then, the domain owner or the product owner?

Robert Luciani: 28:32

This person has become technical. So, they weren't technical to begin with and they've been on every standup meeting on every Monday with the tech team, so they are a part of the tech VT.

Henrik Göthberg: 28:41

So if you look at the report from a reporter point of view he is the AI engineer in their team. Clearly, Then if you look at this to like you as a nerdy, you know hardcore, you know he doesn't do assembly coding.

Robert Luciani: 28:59

They're very far ahead of many other organizations in that respect. I thought that was pretty cool.

Anders Arpteg: 29:04

And perhaps they will start to do coding without julia in the future, because business people actually can code in natural language.

Anders Arpteg: 29:12

I think so how does that make you feel? Let's keep that question a bit and get back to the future of coding. Perhaps you know and see how we can enable potentially business people to not only use AI products but actually develop them. Yeah, interesting. Should we start digging into some of the obvious news questions that we have to at least speak about and then perhaps get more into some more interesting ones? But I think we have some obvious news items that we have to basically tackle and I guess we have to speak about Grok 3 then, to start with Anyone that wants to start off and give some thoughts about you know what's so special with Grok 3?

Robert Luciani: 29:53

I got Grok 3 immediately, because I've been using Grok instead of OpenAI since last year. And I can already say that Grok 2 was very good. There's something very weird in the, I would say, both the performance but also sort of the tone. I don't mean tone like it uses words that way.

Anders Arpteg: 30:16

Before we go too deep here, perhaps we just need to give a quick backdrop to what Grok 3 is Right.

Robert Luciani: 30:21

Grok 3 is a model like ChatGPT, but by the company X, formerly known as Twitter.

Anders Arpteg: 30:29

So X has its own machine XAI, rather XAI. Yeah, that's right.

Robert Luciani: 30:33

So a company with a mission of creating AI models, and it is served through the X platform. Right now there's going to be phone apps. There's one for iPhone just came out and one coming out for Android.

Anders Arpteg: 30:47

Pretty, soon and it's going to go into space. I've heard as well. It's going to go everywhere, yeah.

Robert Luciani: 30:53

And you know, everybody's competing to make the best model, obviously, and XAI is a very young company- Well funded though I guess I should say by strong people behind it yes, mr musk, deep pockets and um, but not just deep pockets. The xai team seems to be a real swat team of super, super experts.

Anders Arpteg: 31:18

I'd love to speak about the speed angle. So we can we can focus on that shortly, but I love that you actually have experimented with it yourself and have some concrete experience with it. So what if you just go back to like Grok 2 versus Grok 3, what was the difference in your view?

Robert Luciani: 31:34

I just want to mention the thing about tone. I always tell people to test models themselves to get sort of a feel for them. It's something about the way they understand what it is that you mean. You know, if you ask a model to give me a summary of this, and that some models will give you bullet point lists, some will give you succinct answers, others will be very wordy and soft Verbose. It doesn't have to be like politically correct or anything, it can just be the style of the way they like to converse with you. And you have to get a feel, for you know what you prefer and I just felt that Grok was to the point and always answered exactly the way I wanted. So I got a preference for it.

Robert Luciani: 32:16

When I switched from OpenAI Now Grok 3 that was over a year ago, right, yeah, like six months ago, something like that and Grok 2 was very good at that time. It didn't have as many features as OpenAI. Now Grok 3 came out and it has three big features the chat feature, the think for longer feature and the research feature. And I think we need to Deep brain as well. Right, maybe I haven't tested that. In that case, I think we need to talk about them all separately. So Grok3 has been secretly competing on Elem's chatbot arena for a long time and you know, as engineers I think we all have respect for that. Benchmarks are just benchmarks and they're interesting and they don't necessarily say more than what they say. But in any case, uh, chatbot arena for me is a nice benchmark because at the very least it says what do users prefer to talk to?

Anders Arpteg: 33:14

not how good it is or anything. I think it's special. Also it's not the normal kind of benchmark, I think in that sense. So just explain it a bit. I mean, it's basically where you have a blind test, people don't know which model has output at what, and they basically say I prefer this over the other.

Louise Vanerell: 33:28

So it's a user interface test User-driven is similar.

Anders Arpteg: 33:31

They have ELO scores like you know whoever? Is winning in chess get a certain ELO score and they can basically put ELO scores on models.

Henrik Göthberg: 33:41

But the real distinction is that this is a blind test that models, but the real distinction is that this is a blind test. That is really distinguishing.

Robert Luciani: 33:48

You don't know which model you're talking to and you have to say I like this one more than this one. And there's one more thing, which is that in the past year it looks like all models have been plateauing, so even though GPT-4 has been a clear leader, and you know, and ELO is logarithmic, so having a few more points means a lot, but in any case the models have sort of been converging a little bit, in the sense, at the very least, that small models now seem to be able to give answers that people think are pretty okay, despite being very small models. And what happened was Grok started competing and just demolished the others. It's the first model to have an ELO score of 1,400. And so it's clearly better than the other models.

Robert Luciani: 34:25

And I don't think it's because it's better, because I don't think users are using prompts that are testing its math capabilities and other things. I think people just like to chat with it. And it comes back to this thing about tone that I was saying does it get the way you want to talk to it? It doesn't patronize you and pretend like you're stupid and sort of over-explained things or under-explained things, and so that's the first cool thing. It's a pretty nice chatbot to chat with.

Anders Arpteg: 34:49

Does it still have the fun mode kind of thing?

Robert Luciani: 34:51

No, I don't know if it has a fun toggle anymore, but haven't, really haven't really used it. The second feature is the think for longer feature, and we got a sneak preview of that with some open AI models, but maybe more specifically with deep seek. That made it more popularized among people at home. But the thing for longer feature is basically that it's allowed to write messages to itself and then read them and use that as sort of like a scratch pad before giving its final message.

Anders Arpteg: 35:29

So basically we can pretend like it's… Is this the reasoning part.

Robert Luciani: 35:32

Yeah, I mean, models now need to spit out their answers before they've finished thinking, and so what they let it do is use a piece of paper that's hidden and it thinks on that piece of paper and then it finally gives you its official final answer. And it's very neat because it's very fast. I mean, this thing is spitting out over a hundred tokens per second really fast in the background and so I asked it a question yesterday where it thought for an absurd. Oh yeah, I asked it to guesstimate the performance of some computing stuff, and it just started talking to itself for a long time. But wait, actually I forgot to take this into consideration. So it was thinking for like two minutes and having this internal dialogue in its head before outputting the answer, but the final answer that came out was really well, sort of-.

Henrik Göthberg: 36:19

Well-structured.

Robert Luciani: 36:20

Well-structured well thought through and everything like that. So it is a pretty cool feature. And then the final feature that grok3 has, which and these are things we've all seen, but, uh, they, they do really well with grok3 the final one is, uh, I think they call it deep search instead of deep research, which openai calls their version, and this is really, um, a great example of agentic behavior, meaning we know what the workflow for doing a bit of basic research is. You find some sources, you compile what, you read across them, you draw some conclusion and then you structure the output and you research a number of times as well, and you research a number of times, and so we can describe this quote unquote workflow in a very well-defined way. And so that is an agent. It's a very simple agent, but it's still an agent, and it's an agent that everybody has benefit from. And people will ask questions like give me a breakdown of the performance of this company financially, blah, blah, blah, and it'll generate a really really nice report for you.

Henrik Göthberg: 37:20

So the traditional market research. Can you do a report on a market entry strategy to UK for Vattenfall? It?

Robert Luciani: 37:28

could even be. Could you compile me a research report on what I need to think about when buying running shoes? Exactly, Teach everything about foot arches and stuff like that. It'll make research on anything you want.

Henrik Göthberg: 37:41

And the core topic then is that it thinks it's through and then it goes out and can more or less search for those things and they it. It can produce the report. Reference to what? Is what the what is, its sources and how it's yeah, so this is very powerful stuff for a lot of different disciplines yes, in normal day life for sure, would you say Grok 3 is an incremental improvement or more of a step change.

Anders Arpteg: 38:08

Improvement compared to OpenAI and Gemini.

Robert Luciani: 38:13

I think they're still playing the catch-up game on some fundamental stuff, but there's so many things to appreciate and applied science is just super important. It doesn't matter that somebody has some secret weapon in their lab down in the basement somewhere and I'm sure OpenAI has lots of that but the fact of the matter is that Grok 3 is out and it's just so good and people can use it and they know how to use it. So I would say in terms of usability, it's a huge leap. It's just so good for the user.

Henrik Göthberg: 38:44

But do you think it's better in terms of you define usability from your perspective?

Robert Luciani: 38:48

then I mean usability in the sense that it feels like it knows what I'm asking.

Henrik Göthberg: 38:52

Okay, the way you connect with, okay, what you had in mind for you, the way your brain is working, it nails that, yeah, I mean ask open ai.

Robert Luciani: 39:05

Look, I'm starting to learn how to cook, uh, and I also have my system prompt, you know. So open ai knows that I'm a 40 year old engineer, I'm I'm not stupid, and I'm trying to ask it how to make lasagna and it's like great question. Let me do this for you, and I just find it so patronizing it.

Louise Vanerell: 39:23

And it's just a lot. Yeah, I don't like it.

Robert Luciani: 39:25

Grok just is like ah, here's some cool ingredients. I think this tastes good. It's a reference from this recipe. See if it works for you.

Anders Arpteg: 39:32

I'm like, actually I don't know how to fry, Can you?

Robert Luciani: 39:33

tell me that Sure It'll give me the answer. It might just be personal preference.

Henrik Göthberg: 39:40

But I think this is very important, and I think it was even Elon who said it. I mean like we will need to find the LLM that suits our tastes. I think that is quite telling, because I mean like you have a very acquired taste, both in music and you know and how you are, and also in how you converse.

Anders Arpteg: 39:57

Yeah.

Henrik Göthberg: 40:01

So that is obvious, that what is better or worse then becomes a personal preference somewhere here, when it's like on the micro percentage benchmark, better or worse.

Robert Luciani: 40:09

I think we're creeping in on sort of design preferences topics that designers are really passionate about, so we probably shouldn't step on their toes too much. But there is an aspect of both learning a tool and having it adapt to the way we want to work. And, um, I think rock three hits the sweet spot, but um, have you played around with it or?

Louise Vanerell: 40:30

not with the, with this one, but I just came to think about the fact when you spoke about tone, just uh, not only in this one specific, but, like you said, uh, you need to find the one that sort of speaks to you. And I remember my first sort of conversations with um chat gpt when they came out and I was sort of thrown off by it because I thought it was so chatty and it was so, you know, trying to um, I don't know, please, my ego and I was like I'm not here to be pleased and it really. I mean, it took me a while to find the ones that weren't as chatting.

Henrik Göthberg: 41:06

And I remember we had Luca on here and, like the whole discussion, which one is the better? Which one do you prefer, coding with Claude or ChatGPT? And then he's like, oh, I guess ChatGPT objectively might be better, but it's too verbose for me. That was the end game, right? So he preferred Claude.

Robert Luciani: 41:24

Here's a question, though have the models actually regressed, meaning you know, remember when, first, what was Microsoft's version of GPT-4 called Sydney, sydney, yeah, version of GPT-4 called Sydney? I feel like the first versions of GPT-3, 3.5 and 4 and Sydney and all these were really really good, and then, quickly, they were locked in and then they weren't good anymore.

Louise Vanerell: 41:49

What do you mean with good then? I just meant. Like they regret they cut off their token space.

Henrik Göthberg: 41:56

What do you think happened?

Robert Luciani: 41:57

I don't know. I just feel like they got more canned answers.

Anders Arpteg: 42:01

I mean, I remember the initial failures of Sydney was kind of amazing. It actually became angry with me. You're stupid.

Louise Vanerell: 42:12

You'd never see that today and that kind of non-PC kind of behavior is really really not okay today and you don't see it anymore. But maybe they aren't really really. You know how do you call it. Uh, when you take antidepressants so you just sort of flatline your, your mood swings and that's what they've been doing to them.

Robert Luciani: 42:30

Oh my god, there's too many analogies to be made here, because I think antidepressants affect your performance in many other respects than just that Good point you know we were talking about how there are some models now that have been uncensored and they have improved in benchmarks because the censorship was removed and it's not obvious why they would be better at math when censorship is removed.

Henrik Göthberg: 42:53

But have you tried Rock.

Anders Arpteg: 42:56

No, not Rock 3, no, so I'm not a paying user yet, but I probably will be. Yes.

Henrik Göthberg: 43:00

Yeah, I'm intrigued now you sold it.

Robert Luciani: 43:04

It has image generation too I forgot to mention. And you can upload images. I think I feel so bad about doing this, but you know how old people will take a picture of their screen and then paste it into a Word file and then zip it and then email it to you just cutting the text.

Robert Luciani: 43:19

I did something similar. I got a. What was it? Some kind of kubernetes error? No, a rock m error for my amd gpu and I just took a picture of the screen and sent it to grok on the phone and it did ocr and and debugged the message for me. I felt really sinful for doing that, but OCR is better in these models than real OCR models.

Henrik Göthberg: 43:42

That's crazy, but there's so many more things to talk about here, because one big topic here is like I guess right now the biggest supercluster is the one in Minnesota, memphis, memphis I mean Sorry, memphis we are talking 200,000 CPUs. Should we go to that topic?

Anders Arpteg: 44:02

I think the speed issue here is a really interesting one, and it's connected to the data center that they built specifically for Grok 3 and what it's been trained on, and should we start with the data center issue then? But I think the whole thing connects to speed in some sense. I can start to give some of my thoughts here, and I think this is actually what makes me really, really impressed with how they were able to develop Proc and what will set it apart in the future. So, thinking about this, elon was late in the game of LLMs, early in the game of self-driving cars, but not in LLMs. And actually XAI was started like one and a half year ago, summer of 23 or something, and they built the first version, grok 1, grok 2. And one and a half years after, they're beating all the top AI labs in the world, including OpenAI, google and Anthropic and so many more, which is incredibly impressive.

Anders Arpteg: 45:00

Then, if you take a specific thing like the data center that they built, so they recognize that they need a shitload of GPUs to do this. And yeah, you mentioned that as well, henrik, but they went to the normal kind of infrastructure providers and said we wanted a center to hold 100,000 GPUs. How long would it take? And they said, ah, about 18 to 24 months. I said, no fucking way, that's too slow for us. We will be dead if we wait that long. We're going to do it ourselves. So what they did then is they tried to find an abandoned factory somewhere.

Anders Arpteg: 45:31

They needed a building A building, so they found an old Electrolux.

Henrik Göthberg: 45:35

Yes, that was a Swedish connection, it's an old. Electrolux manufacturing plant, so it's more or less Swedish. Basically Sweden, we can take the credit for this we as Swedes can take some credit here.

Anders Arpteg: 45:51

Okay, so they found this building and then they basically in 122 days, went from the building to a working data center where they can start training, which is unheard of. But the other thing that is, you know, besides the speed of being able to do this, which normally would take two years, they did in like 122 days is the scale. So no one else has actually built 100,000 GPUs that operate in a coherent mode, and if you want to do inference, it's kind of easy to scale up. You can have a large number of data centers that don't need to be connected to each other. It can be distributed, that's no issue. But for training they need to be connected and they need to communicate to each other.

Louise Vanerell: 46:32

When we say connected, it means that they have to work together. They cannot just be stuff being distributed.

Anders Arpteg: 46:38

They have to actually think and use and for every kind of training batch they have. They have to send the gradients to all other GPUs in kind of all reduced ways.

Robert Luciani: 46:47

So they have to communicate a lot, so extreme bandwidth demands and let's go a bit technical- there's one more thing here that I think is super remarkable about these it's not just XAI, but it's all sort of companies that try to do anything over a billion parameters Is that you almost always only have one chance to get it right. You have like one giant training run and you do one pass through the data and that kind of stuff. It's not like they can train, they can distill smaller models from the big model, but they they can interrupt it during the training though.

Anders Arpteg: 47:19

No, you're right, You're right.

Robert Luciani: 47:21

But it's not like they start from scratch. No, no, but they do tweak during the training.

Anders Arpteg: 47:24

quite a lot, for sure.

Henrik Göthberg: 47:26

But I heard the long format with Lex Friedman and they had the word for this. The do or die run. What do they call it Like you're really betting a lot of money?

Robert Luciani: 47:36

Imagine they it's not like they go wild with the architecture, right? I mean there's some commonalities and stuff but you know they try to do some stuff there and if they make a bad bet and halfway through the training they're like you know what this was maybe not a good idea. It's not like you can start from beginning again.

Henrik Göthberg: 47:51

You lose months and the story they did on Lex Freeman was like they were calculating and say you know what, sam Altman at one point bet the whole company on. Okay, now we have tweaked. We tweaked the tweaks. Okay, are we ready? Do we feel like, are we going to press the button now and now? If we do this now, this is billions burning or millions burning at least. Right, and they had a word for it, like the run of LOL. I can't remember they had a Silicon Valley saying like the death run or something like that. It was funny, very Silicon Valley-ish.

Anders Arpteg: 48:24

Just to go through the whole kind of training. I think it's cool with the data center. So there's a lot of innovation that happened during these 122 days. It's not just that they actually did it so quick, they did it in a very innovative way. So the whole communication. Normally when you have this kind of GPUs, they use something called InfiniBand to really communicate super quickly between the GPUs, but they didn't do that, so they collaborated with NVIDIA and used something like normal Ethernet. It's a Spectrum X Ethernet kind of thing which enabled them to communicate so much faster than they otherwise could. So that's one big innovation.

Anders Arpteg: 48:57

The other thing you know was that they actually use water cooling for the GPUs, and that's normally not used, and you have normal air cooling, except TPUs and Google. But still that was a new innovation that they had to fix. The whole thing is cooled, water cooled and then how do you do that? I mean, they need to have this kind of big factory building and then need to put a shitload of containers beside it to basically fix the water cooling to make that work. But not only that.

Anders Arpteg: 49:22

Then if you actually have 100,000 GPUs using the power grid of the state in Memphis, that's going to disrupt the whole power grid in the state. So that didn't work either. So they have to come up with a solution for that. And then they took all these kind of power walls that they have anyway for Tesla cars and SolarCity and just put a shitload of those kind of containers on the side to basically balance out all these fluctuations that you are seeing during the training. So that's another innovation, and all of this happened and we're able to make them build this kind of the biggest training data center ever.

Henrik Göthberg: 49:56

And you can find a three-minute clip where Elon is sort of sequencing the problem From the building to when you get to the building, we needed to have all the containers of power next to it, and then we needed to have all the containers of cooling next to the building, and he actually mentioned a stat they apparently used a quarter of the total fleet of mobile cooling in the US, a quarter of what you can find and then the Powerwall, and then they had to recode the Powerwall for this problem. So there are many things they used to say. Oh, it's like you know, you solve one problem at a time.

Anders Arpteg: 50:37

The scale of this is so insane and I don't think people understand the scale of things happening here. And what's even weirder is that after a soon time they said ah, this is not enough, we're going to double it to 200 000 gpus and they did that in 92 days yeah, but okay.

Henrik Göthberg: 50:52

so one stat here I heard it and you can explain this better is is this facility the first one where they actually run 100,000 CPUs concurrent? So even if you go to OpenAI and some of the other ones, they tap out at 50 or 60,000.

Anders Arpteg: 51:08

20 to 30,000.

Henrik Göthberg: 51:09

20 to 30,000. So it's a magnitude of three times more complex networking going on at scale, and so this is an engineering feat in its own right, for sure. Nice.

Anders Arpteg: 51:24

Yeah, and they're aiming for one million GPUs now. So that's the next step. And I mean then, when you think about the other big thing a couple of weeks ago was the Invest in AI initiative from European Commission and they were saying like, oh, we're going to invest 200 billion euros, which is actually good. It's a bigger investment than we've seen ever before. Before it was like a few billions, now it's 200 billion. It's kind of nice. But they said you know, we're going to build 100,000 GPUs and you know, for EU to do that it's going to take years probably.

Henrik Göthberg: 52:03

And in a couple of years, where will Colossus, as this data center is called, be? Yeah, but if you want to contrast that with EU I saw someone make a comment on it how can you make this engineering feat as a committee of 50 different partnerships and contrasting that to a team of super swat team led by elon musk, going bananas? You cannot do this as a committee. You cannot do this type of engineering as a committee.

Robert Luciani: 52:26

I don't think that's one of the main problems we have I think you can do it for maybe some topics where there is no rush. You know, look even there, like in fundamental physics and that kind of stuff, we've seen a lot of reasons why these things sort of hiccup. But we've seen precedent for projects that have run for three or four years and have become outdated during that period and it's not like it's slowing down now. So building a data center now and then hoping that something comes up in three years from now is not tenable.

Henrik Göthberg: 52:57

No, but the engineering feats we're talking about. If we follow the history, you know how was the nuclear bomb invented, the Manhattan Project. You take the brilliant minds and lock them in a desert and you know, and they can't move.

Louise Vanerell: 53:12

Yeah, but I think so.

Henrik Göthberg: 53:14

I don't think you can what I'm trying to say here. This type of engineering feat requires an extreme focus from the right people and cannot be run as a committee.

Louise Vanerell: 53:24

Yeah, definitely. I think there's a lot of point to that, but another thing that could be interesting for me to ask you, then is I mean what you're talking about? You know the Manhattan Project or similar things?

Anders Arpteg: 53:36

The CERN. They're comparing this to the CERN project. Yeah, which is interesting.

Louise Vanerell: 53:40

But I think it's the difference between being you know, really the front runners, the one who pushes the boundaries of what is possible, who achieved these things that nobody said was possible to do in such a short amount of time, and you know just the feat of sort of gathering up a quarter of all the water coolers that's even you know exist, um. But there's also the second part of you know, sort of the, the second gang on the ball which can do, can learn from the forerunners, and they're going to compete with being. You know the ones who pushes the boundaries, but they can be the second runners up who do things smarter. There's two different discussions. Runners, then, they're going to compete with being.

Henrik Göthberg: 54:17

You know the ones who pushes the boundaries, but they can be the second runners up who do things smarter there's two different discussions here, because I truly believe that you can have a vastly, you can succeed if you find your niche and your approach. And therefore there is I'm not talking about that we are second on the ball and therefore we can lose. So, so that we can find our niche for a european data center, for sure, and whatever that is. But then you come to how you go about. Yeah, of course, getting there and this is and this is when I say it cannot be run by committees- no I'm not saying that we cannot find a good no, my

Anders Arpteg: 54:54

question is a good value proposition.

Henrik Göthberg: 54:56

It's about we need to find a format, how this works, to be able to do this work?

Louise Vanerell: 55:01

Yeah, of course, and my question was not so much if it's possible or doable, you know, with a committee in place, because I think that's always a hard time unless you're doing things very slowly, but more on, like, what could the value be of actually being, you know, the second runner up? We have the innovator's dilemma, being, you know, the second runner up. We have the innovator's dilemma here, which?

Anders Arpteg: 55:19

is actually partly applicable, I think, and I think it's stupid for Europe to do the frontier model developments and then try to compete.

Anders Arpteg: 55:27

In that we could be actually winning using them finding the application for that, and I think that would be amazing if we actually did invest in that and then potentially you know they call it if we compare the big Stargate thing with $500 billion in US to invest in AI, you know Stargate was basically three main people or companies that are going to drive the whole work, which are commercial, private industry companies the Oracle, the OpenAI, the SoftBanks. That's going to drive it. They have a proven track record, they know they can build it and they have a small set of companies that's going to drive it and they're probably going to succeed in some way. If you have this kind of committee that you speak about, henrik, and then saying we want to have a more inclusive approach.

Henrik Göthberg: 56:15

We're going to have 50 actors and no one has done it.

Anders Arpteg: 56:18

No one with a proven track record, no one that really is accountable either. It's super hard to orchestrate. So if they are going to go into this kind of idiotic race and trying to compete in building this kind of frontier models, how can they compete in terms of speed when it comes to Grok and what we're seeing there? But what if they instead compete of using it? Yeah, I mean that would be an amazing thing.

Henrik Göthberg: 56:45

So there's one thing how you raise the money and now you need to raise the project. You need to raise. How will this work? And then you need to raise. You know what is our competitive edge? Where are we going to put this money towards?

Louise Vanerell: 56:58

But I mean, it's always so much easier to bash on things and to say, like you know, we can never do this or we can never this is not possible.

Henrik Göthberg: 57:03

My argument is we can.

Louise Vanerell: 57:05

But then you know, turning the question towards you, who come from, you know you represent different fields in this area. What would be the smartest way, then, for europe to go about? How do we make the most of these kind of uh money?

Henrik Göthberg: 57:22

that's being poured in? How? How do we utilize the thing that is our? Can we stay here a little bit? I love this question. It's a good question, thank you. Do you want to start or should I start or should you start?

Robert Luciani: 57:29

yeah, I mean, I got a sense from the money answer and I I, if I and if I understood it correctly I agree that there is just such an untapped potential in the way these technologies are applied Exactly that we could be making all the coolest applications, programs, services and everything. And then the fact that the weights are pre-trained somewhere else is sort of immaterial in the grand scheme of things.

Robert Luciani: 58:00

And I think that the engineering track record required for doing that kind of stuff already exists here Not building data centers, but building cool tech companies. I mean we have that, we have all the good universities and everything. So it would be neat to see I don't know if there is something that needs to happen with the funding, but obviously the money exists and then sort of deciding how to put that into the right companies is sort of another matter, but I think the focus should probably be more on applications. And what is your take?

Anders Arpteg: 58:37

I mean, I think you know, as we said a bit before, if you take a tech company the one that actually have found a lot of value from AI and see how much do they invest in research versus engineering? That's one question. And I would say they basically invest 10 times more in engineering work than in research work. If you then see how Europe has done it traditionally, they basically only fund the research and very little investment in terms of engineering, and that's a really bad idea. And if you just take the previous kind of Euro HPC investments that has been poured billions and billions of euros into, they couldn't give the compute away for free. They literally said you get compute in this kind of AI clusters for free for a company that want to use it, and they didn't want to use it.

Louise Vanerell: 59:27

And why is that?

Anders Arpteg: 59:28

Because they were supposed to use it for innovation purposes, not for production purposes. So, okay, let's say you actually are using it. You build a model and then you say, okay, I wanted to put it in production. No, no, no, no, no, no. Then you have to go somewhere else. Then you have to switch the whole environment, then you have to buy your own or go to the cloud or something. And if you look at the cloud offering and the functionality they have there to do the serving, but also the training, it's insanely much better than what the Euro HPC is doing. So I think this is a twofold. One is simply the miss out on investing in engineering efforts, which is really the hard part. It's probably less than 10%. It's more close to 1% of the effort is the research part, and so much more is needed to be spent on the engineering part.

Henrik Göthberg: 1:00:19

And then you can actually lead on to that. You need to have a product mindset here. You need to have something that is usable, useful, adoptable. So if you do something that is research-oriented, well your offering is very niched to researchers who know what they're doing. But it's not easy to flip that product. That is not a product ready for a small, medium business or public sector. So if you take those things, you get to a couple of things Okay, we need to spend this hundreds of millions on engineering doing something cool with a very clear product and a very clear use application.

Henrik Göthberg: 1:01:01

We need to figure that out. You said it yourself it's an application oriented. We joked about this in the Christmas episode and we say you know what? Europe should use the frontier model to distill out baby models for. Europe should use the frontier model to distill out baby models for Europe. You know, we should use those fun, we should applaud these crazy investments, use them to our advantage and then basically flip it into with our data for whatever you know domain that we want to be good at Automotive, hint, hint.

Robert Luciani: 1:01:35

Transport, hint, hint, you know and then we build the world's best uh, you know baby model for for very clear industry domains that is suitable for the um profile of european industry yeah, I think um one thing that's worth reminding the audience of is that the innovation OpenAI has done tons of innovation, but the thing that was very specific about chat GPT was the interface. It was a product innovation it wasn't a technological innovation.

Henrik Göthberg: 1:02:08

It was not a technology, it was not a research innovation at all that could have, in principle, have happened here in Europe.

Robert Luciani: 1:02:13

There's no reason why somebody couldn't have launched a really sweet chat service from sweden it was easy to, to, to use to, and we, we joked about this and we had the joke like there's no research in here and that, no, the the research was in transformers.

Henrik Göthberg: 1:02:29

And then so, no, no, the research was not in transformers. The research was what the transformer techniques are based on. You know, you can go on forever. So what is research? What is product and what is then even ux? We had? We had the guy from bonior media and we had a very good argument and he was just thinking it's all about ux you remember that?

Louise Vanerell: 1:02:48

but it's some truth to it, right and and and then you know if you take that it's very interesting yeah, but I mean you can, you can tie that into both what you're talking about, about the engineering part, which for me and I'll say this from my perspective but you have the research part, which is you know what's the ideal, and then you have the engineering part, which is how do we make this work in the real life, in the real world of things? And that's sort of similar to what we talk about in the user face of ChatGPT or something similar, because something is only bringing value or making a shift of something when it's actually usable. Yes, and you can use the ability not only to the, you know, to everyday people, but also the usability between, like, a researcher and an engineer, or a researcher or engineer to a product owner or whatever.

Anders Arpteg: 1:03:38

Having this kind of engineering first mindset, I think is so important. And actually that is what Elon is doing all the time. He's thinking you know, what should I use it for? Then he goes back and see what do I need to research to be able to make it work, if it's a self-driving car, or if it's Rock 3 for Twitter, or whatnot but, it's always putting it in production super, super fast.

Robert Luciani: 1:03:59

Coincidentally the approach Microsoft took to designing their latest quantum processor. Can we go there? I'm just kidding, let's not go into that one.

Henrik Göthberg: 1:04:08

That's a whole example of the opposite, which is kind of interesting, yeah, but we went down the rabbit hole, starting from Grok and we took Grok into the engineering feat and from the engineering feat we sort of reflected on how should we spend our money in Europe, and I think it's a quite nice turnaround. I mean, do you have anything more to say about Grok? If we flip it all the way back, oh yeah, good question. Anything else we want to sort of highlight in this whole?

Robert Luciani: 1:04:37

We don't know elon has said that he's going to open source grok too right, so for every version that he releases.

Robert Luciani: 1:04:42

He open sources. The previous one. Good, um, we know that when he open sourced grok one, it was basically impossible to run because it was designed for their big, huge machines. And uh, if I were to suspect, they didn't design the models with charitability in mind not that they don't mean to be charitable, but uh, they're not meant to run on your laptop so well, in any case, uh, they'll release that and uh, what else? No, but I think it's a good point.

Anders Arpteg: 1:05:07

I think we should spend a couple minutes here, like on the question of what is the future of open source ai models. And if we start with what you said there about elon, because he's been promoting open source from the start, he actually did start OpenAI on that specific purpose, to put the latest type of AI in the hands of everyone and be open about it. And then he left and then you know, sam turned it into a private for-profit company. But still we can see also now actually elon um which you know in some sense is a double standard here, saying grok 3 will not be open source. Now what he basically said is the the second to first model will be open source, but not the first one, and also he's not.

Anders Arpteg: 1:05:53

You know, if you go to the thinking mode that it has, it shows a bit how it's thinking. But he said basically we are deliberately hiding a bit how it works to avoid being copied, because you know DeepSeek and others will do the same again and just copy from it, and we know DeepSeek distilled from OpenAI and he wants to prevent that. So there is a commercial reason to not open source here and I think there is a regulatory purpose as well. But still, we made a prediction, I think, before saying frontier models will not be open sourced, but other models will, and I think this is actually what we're seeing here. So even Elon, that is a super promoter for open source, is not open sourcing. Now the frontier models.

Henrik Göthberg: 1:06:39

And the prediction goes, we will have, in the end, three or four real frontier models and based out of that, the rest of us will build product distilled using open source approaches.

Robert Luciani: 1:06:52

It's very hard to say what's going to happen five to 10 years in the future. Yeah, that's true, because imagine we have models that are good enough that can run on personal devices, then maybe frontier that distinction won't exist as much.

Anders Arpteg: 1:07:05

There was a little movement, but don't you think there will be some, a few really, really huge models that will never run on a?

Robert Luciani: 1:07:11

phone. Maybe they will be the God models. These will be the God models.

Henrik Göthberg: 1:07:14

They will exist, but for practical purposes.

Robert Luciani: 1:07:18

Well, this would be the God models. They will exist, but for practical purposes. But well, I just want to say there was a little run on Reddit recently. You know, sam Altman went out and did a poll on X and they said do you want us to open source a cool model or a telephone model? And everybody was like, oh, I want a telephone model, but we can distill a telephone model from a cool model if we want to. And people didn't get that. So they started voting for the telephone model from a cool model if we want to. And people didn't get that. So they started voting for the telephone model and everybody on reddit's like please go in and vote. And so you know, with the reddit flood they managed to turn the tide of that. But it's not like open ai is trying to be charitable. I mean, he even admitted we sort of messed up on the openness part and now he's trying to what is controlled the problem.

Anders Arpteg: 1:07:56

But that was a big change in their approach and they certainly went from open to closed. But now suddenly Sam is saying we may have gone too closed.

Henrik Göthberg: 1:08:05

Right, but let's move into that as a new section. That is a segue, then, but this is the open source segue. Okay, yeah, this is the open source we are there, I'm getting it Sorry.

Anders Arpteg: 1:08:16

So what do you think Was?

Robert Luciani: 1:08:22

it the deep seek that forced Sam to change direction here a bit, do you think? Yeah, it was probably the straw that broke the camel's back. I think he's realizing that, people, you know, you can take bad publicity for quite a long time and there are some politicians that are masters of that, but OpenAI is not a master of that and bad publicity now has sort of caught up with them. With the openness thing. And when deep sea came out and everybody got nervous and said, well, look, open source is just as good, I think. I think he just wanted to sort of fix their bad, the damage that had been done to their brand. I don't think he's being. This is not for the greater good, this is a PR thing.

Henrik Göthberg: 1:09:03

And this is also a marketing strategy thing and we can talk about it. If you're the front runner, you want to build a moat in one way, and if you're the second in the game, you want to kill that moat in any way you can. So I mean like, so, of course, an open source is a way to take away their competitive advantage in. In the end, my prediction is that I mean, like, let's call them god models, to really show that this is something very, very different. We will have a few of them and maybe we'll tap into them at some point, but I think, for practical purposes, I think the whole edge model or the small model is more feasible in many, many, many cases because of the cost of running it. What?

Robert Luciani: 1:09:44

is the distinction? Again, I forget. What it's called Open source is when the project is open and people are able to commit and partake in the community of it. And what's it called when the source code is available but it's not open source?

Louise Vanerell: 1:09:56

So when it's more transparent.

Henrik Göthberg: 1:09:58

I would talk about open weight is one.

Louise Vanerell: 1:10:00

Well, open weight is one.

Robert Luciani: 1:10:01

I'm thinking more like you know how, chrome is not really an open source project just because the source code is available. It's like GPL.

Robert Luciani: 1:10:08

No, it's more like only Google can really commit to Chrome and then everybody else can clone and do their own fork of it if they want to. But it's not really an open source project. It's Google's project and it's the same thing here where, I mean, these companies are not running it like open source projects. It's not like Debian or FreeBSD where anybody can be a part of the project, and so Lama and all these other things. They're not open source projects. They are things that are put out into the wild to undermine the competition as part of some kind of strategy that we're not entirely privy to. And, you know, in the minds of the people that are doing it, they have an idea of what they hope to achieve. And we I mean, we don't have open source AI projects really anywhere.

Anders Arpteg: 1:10:53

Meta is still trying to go that line, but I think they will change soon. I think during this year Meta will also stop producing their frontier models.

Henrik Göthberg: 1:11:02

But I get your point. Open source, yes, but can anyone commit to the source code?

Robert Luciani: 1:11:08

Contribute or change.

Henrik Göthberg: 1:11:09

This is the real line. It's hard to run.

Robert Luciani: 1:11:11

AI projects like that, I think, like the.

Henrik Göthberg: 1:11:13

Apache project.

Anders Arpteg: 1:11:15

Yeah something like that.

Henrik Göthberg: 1:11:16

It's not, yeah, it's even.

Robert Luciani: 1:11:20

Lama, that I'm not sure. Well, maybe we can talk about that at the end, cause one of the papers I want to talk about is about a model that didn't require as many GPUs as these super giga models require and what maybe little companies can and should be. Yeah.

Henrik Göthberg: 1:11:32

All right. So so on the open source topic then, what's the future? What's the future? What's our predictions? I think I stick with my predictions that we will have some few frontier models, but for the practicalities, we should be focusing on distilling out our own models and then, most likely, they will have the open source core, and this is how we build cost effective there's one more weird thing happening when it comes to open source no-transcript.

Robert Luciani: 1:12:22

I mean, it's not the only source, but it's really a pressure.

Henrik Göthberg: 1:12:25

Yeah, and it becomes a pressure of memory usage, storage uses, bandwidth use, all this.

Anders Arpteg: 1:12:30

But just thinking future, do you think we will see? Let's assume that we have 2 trillion parameter in the latest or in GPT 4.5 being released soon. That's obviously super expensive to run anything on and probably if you build a product that had 1 million users and they all make use of this model, it will be far too expensive to make use of such a big model, right? Would you think that upcoming models coming years will have even bigger number of parameters, a larger number of parameters than this, or would you think the future will be there?

Anders Arpteg: 1:13:13

I think they'll go up and then come back down again, because it's easier to scale up certain things and then, once we learn the tricks, we can bring it back down again, but potentially there will be models that, even if you use reasoning tricks which I really excited to speak more about soon there will be some things that is super expensive to train and use for anyone and elon, when they have a million gpus in their next uh, colossus version, they will train something that is really, really awesome but far too expensive for most companies to build a product on. Yeah, I don't know.

Robert Luciani: 1:13:46

I mean we can look at the price of hard drives over 40 years and maybe the price of compute will take 80 years to get there. I don't know, it's just very speculative. But with regards to open source on its own, I think there's so much cool stuff happening in open source that people will keep themselves busy. So if we tie it back to, like the SVT stuff we were talking about before, the little fun loras that we're making, the stuff that we're doing with image, the stuff that we're doing with text doesn't need all those frontier models and so I feel like there's space for open source to thrive and grow.

Anders Arpteg: 1:14:21

Yeah, I mean, that's a second line kind of models that you spoke about, right, and that will be what we really should spend time on, right.

Henrik Göthberg: 1:14:28

And this ties back to we had Anton Osyka, founder of Lovable, and we discussed with him you know what to care about and how to build a great AI product, and he was really putting it you know what at some point in time. Yes, yes, yes, the model is important, but I need to really focus on the engineering parts and making the whole thing work and the UX work, and that is way more important now to get to useful performance than the actual percentile of the frontier model versus a mini model, and I think this is what we are almost. You know, the headlights is rising to us and we get the shiny lights, so the real problem is in the blind spot, which is build a awesome product and then in that sort of that whole brick wall, you know, the model is just one brick in the brick wall, and I think this is what we need to figure out in Europe that you know what we need to be really fucking good at building the brick wall and not only staring at the frontier model.

Anders Arpteg: 1:15:31

It's one brick. I see Goran here is playing around with GROK3 and doing some deep search.

Robert Luciani: 1:15:38

Is he making a research report for us? Yes, I think, he just pretended that he's thinking.

Louise Vanerell: 1:15:46

He's thinking about something.

Robert Luciani: 1:15:48

I guess we'll find out soon what he's thinking about.

Anders Arpteg: 1:15:49

There's confusion in the numbers. Robert, let's go back to the report when it's done and see what it came up with what's your question to it.

Robert Luciani: 1:15:59

So if Grok would have, a mission to make Europe leading AI by 2030,. How would he do it?

Henrik Göthberg: 1:16:11

Oh, that's a good report.

Robert Luciani: 1:16:13

Yeah, send it to the European Commission.

Louise Vanerell: 1:16:15

Sold it. Charge it as a centering.

Anders Arpteg: 1:16:21

It's kind of. I've used this kind of deep research things from OpenAI and Gemini and now deep search here as well. I haven't used deep search, but still it is kind of impressive how you in 20 minutes can do something which normally take months to do otherwise.

Louise Vanerell: 1:16:37

I read this small study from BBC. I think it was where they had done you know very old-fashioned deep dive into how different AI-supported news outlets, how they compile their news and if they're correct or not, and I don't remember the precise numbers, but it was quite a large number where they were, you know, making stuff up when they were sort of summarizing the news, and I think it was maybe even half where there were you know errors or where they didn't, they changed sort of the tone of the news or they changed sort of what the angle was Changing the angle.

Henrik Göthberg: 1:17:25

No journalist is changing the angle, ever. No, no no, never.

Louise Vanerell: 1:17:29

But I mean it could be troublesome if you read sort of what you think is a summary of a BBC News article and you expect certain things and then you get something else. Yeah, that's true. I thought it was just kind of interesting, especially today when we get so much sort of compiled information in that way, and of course that maybe ties into how perplexity or other similar services are being used today, which maybe ties into how I think we should read out what Goran just did.

Anders Arpteg: 1:18:00

He doesn't have a mic here so I can read for him, potentially, but you basically asked some kind of question about how Europe can win in 2030, or something.

Henrik Göthberg: 1:18:09

How do you care about the leading AI race? This was sort of the question that we answered or asked, and the hypothesis is if Grok will do it and if he has the mission to win, oh okay, go on with leading questions.

Anders Arpteg: 1:18:28

Okay, so I'll go back to the first part and then I get just going to read out Okay, key points. Europe needs to boost AI funding. Okay, spotted on that Attract talent. I think actually we do have a lot of talent, I think so too. Streamline regulations Certainly agree. Lead globally to win the AI. Okay, to win the race by 2030. It's very important. You need to have a mission to win, yeah, but I think it's to have a mission to win?

Anders Arpteg: 1:18:58

Yeah, but I don't. I think it's actually a bad mission to have, but that's another thing.

Robert Luciani: 1:19:02

It doesn't matter, it's just my saying. Must invest in infrastructure and foster collaboration I suppose collaboration within the EU. Okay, a unique advantage is Europe's focus on ethical AI. That's a unique advantage in becoming the world leader. Okay. Which could set global standards and attract talent?

Anders Arpteg: 1:19:22

It could attract talent for sure it could actually be a good point.

Robert Luciani: 1:19:26

Can you scroll down Steps to win? Boost funding Doubling the public AI funding to 2 billion yearly by 2027. Create 50 billion European AI investment fund to track 200 billion in private investments. Matching US levels Makes sense to match US levels. Launch a European AI fellowship with Stipends to build custom AI curricula to train 1 million professionals. This I want Luis to talk about for a sec, because she has spoken on other podcasts about the idea of creating AI courses in university and whether that's a great idea to sort of speed up AI professionals.

Louise Vanerell: 1:20:06

Sorry, maybe I'm getting ahead you want me to do it now?

Robert Luciani: 1:20:10

continue reading this one, then let's uh, let's pin the and remember to get back. What is the best way to um get people good at ai right now? Okay, good, I write it down. Streamline regulations. Oh man, that's like beating a dead horse right now. Build infrastructure. Invest 10 billion specific, very specific number in computing centers for three exascale computers by 2030 and develop a european ai cloud cloud platform for secure resources. There it came, there it came. We need it within our own cloud. It's for inference, that's unspecified. Maybe later for inference, we don't know yet. Foster collaboration ai hubs in major cities. Cross border projects to unify efforts across 27 member states. Set global ai standards okay, still pretty high level so far. Comprehensive analysis let's just check maybe the first one and see sort of current AI strategies, focus, excellence and trust. Recent initiatives include invest in AI initiative launched in February 2025, mobilized 200 billion Okay, found the latest one including 24 gigafactories.

Robert Luciani: 1:21:17

However, it's found the latest one Including 24 gigafactories.

Anders Arpteg: 1:21:19

However, here's the detailed plan I think it missed the point of engineering. I think so too.

Henrik Göthberg: 1:21:27

I think you need to be more explicit on that fundamental topic.

Robert Luciani: 1:21:30

The only one that sticks out here is the exascale supercomputers that it thinks.

Anders Arpteg: 1:21:36

Only for innovation purposes.

Robert Luciani: 1:21:38

Right.

Anders Arpteg: 1:21:40

Okay.

Robert Luciani: 1:21:41

Citations. You prompted it wrong, Goran. That's what happened.

Henrik Göthberg: 1:21:49

Anyway, should we move to another topic?

Anders Arpteg: 1:21:52

Yes, okay, which one should we take? Should we take your topic, dan, and let me see if I can phrase it properly, but something about what's the best way to learn how to use AI today, or how would you phrase it, louise?

Louise Vanerell: 1:22:05

Yeah, well, I think it's double-sided questions, because we're both talking about how do we bring forth or do we bring forth people who are sort of on the edge or on the front when we talk about AI development, and obviously there we have the research field and obviously there's a lot of stuff happening there. But if we then take the other side of the coin, like how do we make the most value out of it? How do we make Europe something that's competing on those levels where we see that there's Embracing it in our enterprises.

Louise Vanerell: 1:22:43

Yeah, exactly when the action happens. How do we make money out of it, which, of course, is the sort of the end line here.

Anders Arpteg: 1:22:51

Sometimes societal gain as well, yes, of course, of course, of course.

Louise Vanerell: 1:22:56

But usually, you know, those two goes hand in hand in one way or another. Yeah, either you, you know, you work smarter, you make better systems, you do all these things and then you.

Anders Arpteg: 1:23:05

But for public sector it could be other things than monetary gains. Yeah, of course, but still, I see you for it.

Louise Vanerell: 1:23:10

I agree with you. And then I think if we're talking about the regular sort of educational systems with the university and stuff, then you know, at the least you have a seven, eight year sort of frontline to when people from. If we were to decide today, yeah, we need stuff, then it's, you know, seven, eight years at the least to get those people out into the industry again. And then we're talking about junior people. They're, you know, they're not even dry behind their ears yet. So then how do we make sure that we get people the right competence and the right stuff so they actually learn? And then of course, like you're talking about, then we need to make sure that people actually doing the work today get to know and get to know how to handle the different tools and the advancements and what's actually possible. And I was actually thinking when you said that the team that you work with today at SVT is a really small one, and I find that so interesting because the smaller you make the team, of course then you have to make sure the people, all of the ones in the team, knows what they're doing. You can't have these, you know, humongous teams with 25 data scientists somewhere doing some sort of work, and I think that's a super interesting sort of field. If we're talking about how to make sure that more companies today in Europe get to use this kind of technology, how to make sure that more companies today in Europe get to use this kind of technology, we need to bring in people who sort of spread out the knowledge and get people to do stuff.

Louise Vanerell: 1:24:43

I don't know who said it, but you know, when you talk about engineering like putting it out in a production part, that is and I thought about that when you said these huge models that you were talking about from the start and you sort of at some point have to push the button do we know enough?

Louise Vanerell: 1:24:59

And the same thing is on a very, very small scale when you talk about companies and industries. You have to at some point be like now we just have to try it out. We may not know everything, we may not have everything in place, but we just have to try it out. We have to start somewhere, and I think that's where most of the learning actually happens, just as it is when you talk about you know huge projects with scaling or learning or developing these kind of models. It's not until you actually see what goes wrong, that you can be able to adjust and you do something new, and then you learn, and then the other ones coming after they're like, ah well, this is the trouble that they were, you know, running into. How can we make it smarter? How can we be better engineers so we make sure that the production, the actual results of something, gets better?

Henrik Göthberg: 1:25:49

So a very long, maybe very broad answer. The bottom line then the university will fix a small small part of the problem. The university will fix a small small part of the problem, but you're alluding to that. We need another type of adult learning on the job, experience learning.

Robert Luciani: 1:26:05

It's so obvious when I think about it, I had just never considered the sliding time window of universities. We have like an AI course, yeah. Then they have to go the course for two, three years. So we're still like four years away from getting those people. And then, like she said, you know, they're junior developers too, so we need to use the people that we already have as well.

Anders Arpteg: 1:26:26

Isn't this an example of you know trying AI versus like building AI? Sure, we need to have education that brings out like super, you know technical people in what AI is, but the big value isn't from those people. I would say the big value is when people that are domain experts in whatever kind of field there is, start to use AI. And that is when I think what you're doing as well, louise, is really upskilling people to have the level of knowledge they understand. You know what to use it for and what not to use it for, and how to do that properly. And if we can get people to just be able to use it properly without having to build it I mean it's similar to the europe versus us kind of thing as well then that that is really the valuable thing. And then you don't need to go like a university course for four years. You can take a smaller like upskilling course rather quickly.

Henrik Göthberg: 1:27:17

But but there's a deeper problem here, because, on the one hand side, I fully agree with you that we then need to focus on adopting data and AI in our core operations, so that is not the same as having another AI researcher out there. The key problem that I have seen that has worked extensively with enterprise is that a lot of times, we have a fundamental way of organizing and processes that were designed in an analog context, so that education is not only about using AI, but the education is okay. Maybe I need to have a cross-disciplinary team, I need to have this prompter expert, I need to have a product owner, so there is an upskilling that is not related used to using AI, but it's the context of organizing in an agentic world, maybe. So I think this is all about user adoption, but it's actually about leadership and it's about organizing things slightly different. Otherwise, we are putting AI onto analog processes and I don't think that is really the productivity frontier either.

Louise Vanerell: 1:28:27

That's not really a new problem either. Hasn't that been sort of around for quite?

Henrik Göthberg: 1:28:32

some time, yeah, but we never solved it.

Louise Vanerell: 1:28:34

No, we haven't solved it. And now it's maybe. You know it's on steroids, so even though we've been able to, We've been discussing that exact problem.

Henrik Göthberg: 1:28:43

I've been trying to sell to that problem since Data Innovation Summit, but ultimately we have this divide that people that are setting the organization and designing the processes are AI illiterate, and I'm not saying that they need to be engineers, but they need to understand. What does it mean to have a cross-disciplinary approach? What does it mean to have a product-oriented approach?

Louise Vanerell: 1:29:08

Yeah, but I mean, we had a discussion and I think I asked you about it which was you know if you want to have, if you want to get started, what's the smallest possible team you can have?

Louise Vanerell: 1:29:23

And what does that consist of? And we spoke about, which again ties into the fact that you need the engineering part, because of course you need the full stack engineer, who preferably knows, you know, all the ins and outs of how your system setup looks today, about what do you need to think about security-wise or governance-wise or whatever it is, all those bits and pieces about how does our setup look, what makes it work, what makes it fail, what makes it blow up. And then you combine that with somebody who actually knows not necessarily a data scientist, or rather not, but rather somebody who's a Gen I developer, who knows how to build programs on these new technologies, and when you combine those two, and then preferably with some sort of AI product owner we said right, then you have like the smallest task force that you can have, but still have the capability of doing stuff that works.

Anders Arpteg: 1:30:25

Yes.

Louise Vanerell: 1:30:26

Would you agree? Yeah.

Anders Arpteg: 1:30:28

Yes, but it's a deep topic, I can speak about this forever Just a small comment here you can think about should you optimize the existing business processes that you do have or should you invent new business processes? And of course, the big value comes when you start to transform the way you work. But I wouldn't say it's a bad thing to optimize existing one either. That's actually a good thing to start with. Yeah, because it's so slow, it's so hard to change an organization and the way they're working.

Henrik Göthberg: 1:30:55

So don't be afraid to start the easy way I always do, and then you have a strong point there, because it's also about if we think about this as learning by doing is the only thing that will work. So maybe then, okay, I up, let's now try to improve our analog process with the wrong organization as a project, and then we realize by learning that, hmm, we maybe need some other team members here, et cetera, et cetera. So I see your point that actually don't get hung up on that. Don't make the whole problem too big.

Anders Arpteg: 1:31:29

It's not the first thing you would start with. I would say no exactly.

Louise Vanerell: 1:31:33

I have a very, very tangible sort of example from that and that is when I go in and I try to set up.

Louise Vanerell: 1:31:41

You know the smartest way of making sure that people learn stuff in their work environment and people are always looking for that sort of you know magic bullet of just give me the tool or whatever that's going to, you know, suddenly make everybody love learning.

Louise Vanerell: 1:31:59

And then you're like, nah, you know, I try to be, as you know, as how do you say, reluctant to bring in any new tools or sort of disruptive ways. You know, if you work in the Microsoft universe, I will set something up in your Microsoft universe. So people know the tools. They don't have to log into a new place. They don't have to, you know, mess around with getting a new account somewhere which they will forget in two weeks anyway. You know, just do it the path of least resistant when it comes to sort of bringing on something new and eventually they will be like, oh wow, I know these things now. So if you're setting up some sort of readiness program where you want people to start using these kind of tools, you will make sure to set that up then in the tools that you want them to use, so you make it not too hard.

Anders Arpteg: 1:32:51

Yeah, awesome, and time is flying away here a bit and I'd love to come to the paper that I haven't actually read, but I get very excited about when you speak about. So if we could, let's go there.

Louise Vanerell: 1:33:05

This is today's learning session.

Robert Luciani: 1:33:09

I think we might be jumping over a. You know we've been talking about competing with the United States, because it's a competition. There's another big competitor out there as well, apparently.

Henrik Göthberg: 1:33:26

China.

Robert Luciani: 1:33:27

China, china, china, and they released a pretty famous model now.

Henrik Göthberg: 1:33:33

We haven't said anything about DeepSeek. Do we need to go there a little bit first?

Anders Arpteg: 1:33:39

It's another new model that came out.

Robert Luciani: 1:33:44

So let's jump straight to the nerdy bits. I wanted to challenge you to show me a cool paper, because I found a really cool paper and I already know that you like this kind of stuff. Tom goldstein is a guy that I know that runs a cool team in the us I forget which university it is, I should probably be able to find it here it's on archive, uh but his, the paper they released uh, and they released the source for their work is titled Scaling Up Test Time Compute with Latent Reasoning, a Recurrent Depth Approach. Okay, I get less interested now, but okay, I know you have your favorite approach to the latent space.

Henrik Göthberg: 1:34:30

He missed the word space and then he's seen that.

Anders Arpteg: 1:34:32

No, it was the test time thing that I reacted to, but still, tom.

Robert Luciani: 1:34:36

Goldstein has done a lot of stuff a long time ago with sort of visualizing the internals of neural networks using 3D and that kind of thing. So I always thought it was pretty cool.

Henrik Göthberg: 1:34:46

Was that? You showed me ages ago some really cool videos.

Robert Luciani: 1:34:50

I made my own Julia implementations that were much faster than the Python ones that they made, but they invented that sort of visualizing the loss.

Henrik Göthberg: 1:34:58

The really beautiful one was made in Julia.

Robert Luciani: 1:35:01

So what they did, was they introduced? What was the title? Again, please if you could repeat it. Yes, the title is Scaling Up Test Time Compute, so inference time with latent reasoning. So, rather than the latent part, I love.

Robert Luciani: 1:35:16

Yes, and it's because there are a couple of. They chose a specific architecture. They trained on 4,000 AMD GPUs for 12 hours, for 21 days, and they built a I'm pretty sure it was 3.6 billion parameter model Pretty big, but not at all big compared to today's models. And the whole point is that they chose a non-mixture of experts approach because that was too complicated. Oh, there's many cool things here. They added recurrent layers in each layer so that, instead of it doing chain of thought reasoning per token afterwards, it was allowed to determine how many recurrence operations to do inside each layer.

Robert Luciani: 1:36:11

Now there are a couple of cool things that happened here. The first thing is, when they slid it manually, they could do tests on whether it was their architecture that gave good performance or the recurrence that gave good performance. By limiting the recurrence to one pass ie no recurrence they got a certain score and then, by bringing it up, they got another score. So the score that they got with no recurrence was pretty competitive for a 3B, 3.6b model. When they turned on recurrence, they could pretend it was more parameters, ie as if they had had that many more parameters depth-wise. So they can turn it up and say I'm going to do oh man. Yeah, but like do four cycles, which would be the equivalent of an 8.3 billion parameter model otherwise. But it's still 3.6. It's just, it runs many times through and now the performance suddenly goes up and it seems to sort of plateau out at around 30 to 50 B, hitting peaks at around 80 B. So it it as a 3.6 B model it's competitive with 80 B models.

Anders Arpteg: 1:37:17

And then it doesn't seem to cycles 32.

Robert Luciani: 1:37:20

Oh, that's a lot.

Robert Luciani: 1:37:22

They tested 48 and 68, 64, and it didn't really improve. But they tested on arc challenge, open book QA and a couple of other benchmarks and it's very cool. It really does get that much better. Now there are other things that one might want to do use mixture of experts and that kind of stuff. And what's really cool is you know now you choose one or many experts and then run through them. But with this model, for every recurrence you might want to choose a different expert. So the downside with experts is you still have to have them in memory, so it takes up space and you only get an advantage because fewer parameters actually have to be calculated, because you're only activating a few of the experts. But here you could even have both advantages you don't need as many parameters because you're doing it recurrent and you can have lots of experts, so you get really fast forward passes.

Robert Luciani: 1:38:20

Now what else is cool? Well, the way they use sort of skip layers in here to the novelty of why they got this to work was specifically that they use skip layers and other stuff to sort of get the model to train, because what's hard about these big models is just getting them to train at all. So the architecture is a little bit neat. But there were some sort of cool things here, like if you want to do some performance improvements, like having a shared key value file across layers, with a traditional transform architecture you have to train with that in place. You can't change it after the fact. But in this one, for some reason, it seemed like the recurrent layers were able to adapt to having a common key value, like a fixed one. Yeah, so they can share it. Fixed one, or I mean yeah, so they can share it, and so you can iterate over the recurrent uh layer and have one key value, uh matrix, meaning you don't have to have uh, all that extra computation. So, again, increase in the forward, uh, in performance in the forward pass. And then finally, because I said, uh, this was one of the coolest things because this team was so good at visualization, I suspect they sort of tracked the journey of the activation through latent space and they found because it's happening in the same recurrent layer all the time, they can track, sort of where things are headed in latent space and they found that certain types of questions resulted in.

Robert Luciani: 1:39:55

I don't know if you can bring it up there, because it's a very nice visualization, let's see here, here we go. Let's see. There was no title for it, but when you visualize, using PCA, sort of the trajectory of a token through this recursive layer, the type of question generates different types of orbits or movements throughout latent space. So a mathematical question and this is really a geometric representation of the compute graph inside latent space. So when it's thinking about math questions it makes sort of elliptical orbits and when it's thinking about other things like right and wrong, it'll have more. I don't know what you should call it, maybe polytope orbits or something like that, but it's so cool that you can see a geometric shape inside of latent space, how it thinks that you can see a geometric shape inside of latent space, how it thinks yeah, basically like a geometric shape, that is the representation of the computer graph to solve a problem.

Henrik Göthberg: 1:41:00

This is new thinking. It has this shape Exactly.

Robert Luciani: 1:41:02

Solving logic has this shape Computation. That's a little bit cool.

Anders Arpteg: 1:41:07

So they call it test time, but I guess the recurrence happens in training time as well, right, yeah, yeah, of course. Why do you think they call it test time?

Robert Luciani: 1:41:15

but I guess the recurrence happens in training time as well, right, yeah, yeah, of course. Why do you think they call it test time? Well, because you can choose how much recurrence to give it, to sort of scale up the performance. But it still happened in training time. I'm not sure. I think they only trained with one. I'm not sure actually how they did. I probably should.

Anders Arpteg: 1:41:31

The reason is I'm very annoyed with the term test time compute, so that's why I'm getting a bit interested here.

Louise Vanerell: 1:41:36

What's the annoying part of it? Because?

Anders Arpteg: 1:41:39

if you take O1 and R1 and whatever model. If you think they're only doing test time compute, you're wrong. I mean they are doing the reasoning and the multi-step reasoning during training time as well. And people still still keep calling it test time, and it's not just during testing, it's actually during training as well. It's just simply called it multi-step reasoning or something.

Robert Luciani: 1:42:01

It's not only in test time architecture but for this one I'm not sure.

Anders Arpteg: 1:42:04

I think they are doing recurrence during training time as well.

Robert Luciani: 1:42:08

I would be surprised if they didn't. The thing thing is, I wouldn't keep it against them because I know Tom Goldstein has made other papers where he has trained optimizers to do one step and then do multiple steps after that, and so I know it's difficult to do it on one step and then just do a test time multi-step predictions. So I think that's probably what they would want, but given that it's so difficult, it wouldn't surprise me if they trained with multi-step for convergence and for everything else. But it is cool that you can have such a small model and that you can get it to be big just by giving compute. And I guess I just want to do a small shout out here.

Robert Luciani: 1:42:52

I know a lot of hard this is tying it into hardware nerd stuff. But a lot of laptops that have gotten attention now are the MacBooks because you can load big models on them and they have super high memory bandwidth, which is nice when you're generating tokens per second. But for computing stuff and for pre-processing stuff and for evaluation and all that, that all compute bound, and this is all compute bound as well. So you know the gpu in a macbook is pretty sissy and you still want a powerful gpu and hopefully.

Robert Luciani: 1:43:20

That means, that future models you don't want a macbook no, you don't no, no, you want in their right mind, yeah.

Louise Vanerell: 1:43:27

So anyway, I just thought it's cool that this is very compute heavy rather than memory bandwidth heavy but, uh, coming here then, uh, sort of representing the um, the, the village idiot, um, when it comes to these things, how would this tie into the, the common thing that we had earlier? How do we? Is this something that can be utilized if we want to make sure that we're not losing in this game of making sure to bring value from these kind of things?

Robert Luciani: 1:43:53

can I just make a little attempt at explaining exactly what?

Louise Vanerell: 1:43:55

this sort of entails? I think yeah that's yeah.

Robert Luciani: 1:43:59

um, so we've learned that if we make really big models, they do better. And what this one is trying to do is the following one of the reasons big models do better is the wider they are, the more things they can remember, and the deeper they are, the more steps they can use to think through things. And what this one is doing is, instead of having many steps built into the actual model file, it's making the file smaller and then sort of calling the same layer many times, so it's pretending to be deeper than what it is by recalling the middle layer many times, so it's pretending to be deeper than what it is by recalling the middle layer many times. And then there are other techniques mixture of experts which, instead of calling all of the width, only calls part of the width. And all of this is to get the advantages of the big model and run it with as few resources as we can.

Louise Vanerell: 1:44:51

When you say expert and then you say certain things, does it know which part that it needs to call it? Does I guess, one thing that's worth?

Robert Luciani: 1:45:02

clarifying is that, let's say, you have a model that has 10 experts. It's not like one is an expert on medicine and one's an expert on biology. It's more like this. One knows a bit more about this question, so we'll let it answer. Let me try to also explain it for the purpose of this is so fun to try to distill it.

Anders Arpteg: 1:45:25

But you know, we are saying that some kind of reasoning, even though it's a shallow type of reasoning some kind of reasoning is happening in these big transformer models, including ChatTPT etc.

Anders Arpteg: 1:45:36

And people have said. You know, if you increase the number of layers in a model, the amount of reasoning, the number of steps you can think, increase in some way. Also the context length the more tokens you can look at in a single window, the more reasoning you can do. So the reasoning powers in some sense is bounded by how many layers, which increase the number of parameters, and the window length, which also increases the number of parameters, making these models that do reason well super big. What this one is potentially doing is saying we can have many layers without having more parameters in some sense.

Anders Arpteg: 1:46:16

So they are going through the same layer many times without adding more parameters, meaning you can have more reasoning without adding new parameters to it, and that's really cool. Potentially and that's why you know potentially they are having recurrence in some kind of middle layer, which is a latent space where they do more reasoning than it's not the type of latent reasoning that I would hope to be, but it's a small, simple one.

Robert Luciani: 1:46:42

And it's worth comparing to what Grok does today. So, basically, grok does this even though it doesn't have this middle layer that spins.

Henrik Göthberg: 1:46:49

Instead of spinning in the middle like this, it actually spits the text out and then sends it back to the beginning and then reads it again, which is very inefficient yeah, so, and what we, what we can look at here is then, basically, when we're talking about innovation, this, there's so much research and innovation going on not only around making it more intelligent, but making it way more efficient, and this is cost efficiency, compute efficiency, bandwidth efficiency.

Louise Vanerell: 1:47:18

Which of course makes it more usable.

Henrik Göthberg: 1:47:20

Which is ultimately the whole discussion on total cost of ownership. I saw a number. You know, if you have a standard question and you wanted a certain token with a certain quality of answer, there was some number In the beginning that was like a $60 question. The amount of compute that goes into this question equates to $60. So if you now compare, if you do the same type of question in the, in the, in the latest models, we are talking about the cost per question going down to 60 cents. So it's, it's, it's. So. This is what we're talking about, right, and this is part of making this, you know. So we we're talking about, you know, uh, this is going to go crazy. But layering, you know, when we are compressing and doing all that, it we can fit in the phone, you can fit it on a normal computer and it you know it's part of making this useful is the cost in the end ultimately.

Louise Vanerell: 1:48:20

But I also think this is, uh, this is super interesting because it's I mean, we love numbers, we love big numbers, we love, you know, the people who, uh, just throws everything in there and does huge things and just like you were talking about you know what you have in your sort of field of sight. Is that the correct one? Yeah, but there's so much stuff happening on the sidelines where people are just trying to make things smarter or more efficient or make it work better or those kinds of things. But it's not as sexy because it's as sexy, because it's not the world champion.

Henrik Göthberg: 1:48:56

And tying it back to what should Europe be doing? Should we really be there? To build the biggest models, or should we be there to be the most smartest, most?

Henrik Göthberg: 1:49:05

efficient useful model for normal application, because there's one thing to crack AGI, but we can have 90% of the immense value of the work we do today as people. We don't really need AGI to automate that, we just need it to work efficiently, cost efficiently especially. So I think you know, the more you think about this, what is our competitive edge. You know where should we go with this. Then you can go and do amazing engineering and do something that's very, very useful and very, very much cutting edge and has nothing to do with building the biggest colossus cluster it's the opposite.

Anders Arpteg: 1:49:49

Okay, anything more about the cool paper, by the way? Yes, I just wanted to answer your question.

Robert Luciani: 1:49:57

It says here, to enable synchronization, we sample a single depth R for each micro-batch of training. It sounds like R is variable, so you were right. But I wanted to say one last, just super cool thing, because you could potentially do one forward pass and sample the output before you keep going. You get like a fake kind of speculative decoding from this.

Anders Arpteg: 1:50:21

It's sort of cool that you just get that for free. Yeah, yeah, I agree, I still want to go to the JEPA paper because I think that's really the end goal. It's the best one. I mean, that's still, I think, the best way to go in the end.

Henrik Göthberg: 1:50:34

Because you sort of latent space has been a topic that we've been circling back to. Maybe we should do it a little bit here, like you've been talking to us about this before it is latent space even.

Anders Arpteg: 1:50:45

Yeah, okay, okay. So okay, we're moving more to the futuristic kind of you know what could the future of AGI be? And speaking about different paths towards that, and just connecting to this paper then, and speaking to Jepa potentially but we are. Let's do it like this If we take image generators of today like DALL-E and Stability, AI and Flux and everything they actually do, reasoning in latent space Recurrently.

Robert Luciani: 1:51:19

Yes, with noise and stuff.

Anders Arpteg: 1:51:20

No, but they basically have an auto-encoder around the pixel. So if you take pixel space, if you were to do reasoning in pixel space, it would be super high dimensional and very low level of abstraction and it would be super inefficient.

Anders Arpteg: 1:51:32

So what all of them do is basically to first compress the pixel space into some kind of latent space, and then they do the diffusion to generate the image in that kind of latent space. So they already moved in this direction that I'm really excited about and I hope that we will come to, which is what Jeppa is speaking about having these kind of joint embeddings, embedding into a latent space and then doing the reasoning there. So we already see reasoning like for image generators happening in latent space. We're not seeing it for text, though, yet, which I think is a bit surprising.

Louise Vanerell: 1:52:05

Why is that?

Anders Arpteg: 1:52:06

So text actually is in some way actually rather high level of abstraction. It's far less high dimensional than pixels are and they are actually rather high level of abstraction. It's far less high dimensional than pixels are and they are actually rather high level or semantical in that sense. So in that sense it's not so horrible to have to encode and decode to every token through every forward pass or every auto-aggressive pass that you take, even though it's super idiotic.

Robert Luciani: 1:52:31

It makes more sense in character-based languages than in English, for example.

Anders Arpteg: 1:52:36

Like.

Robert Luciani: 1:52:36

Chinese, where one character can mean an entire word.

Anders Arpteg: 1:52:40

But still I can follow Jan-Le Kun's point of autoregressive models is stupid, Because if you want to have this kind of reasoning, why should you need to decode to a human understandable language every time and instead do the reasoning every step in the latent space, which is actually what image models are already doing?

Anders Arpteg: 1:52:59

Why couldn't you put an autoencoder around a piece of text and then do the generation happening in that space? I think this is actually where we potentially will be going at some point. This is just me speculating, and then you can think also, if you can join this kind of latent space for different modalities meaning text and images and audio potentially and say they actually understand each other in this latent space, which is like think clip, if you know that kind of model, that would be amazing. This would be a possibility for us to do a reasoning in a super efficient way, because we're working in this compressed latent space now and then we can avoid having to encode, decode all the time. We can understand across modalities, and it would be so much more efficient. I think it's for me obvious that this is a direction we need to go?

Henrik Göthberg: 1:53:50

Have we seen any paper? I mean like JEPA paper. It's already one year old now, or more.

Anders Arpteg: 1:53:54

Yeah, but that's a position paper.

Henrik Göthberg: 1:53:56

That's a position paper. Have we seen any real work?

Anders Arpteg: 1:53:58

So we have iJEPA and we have vJEPA, so they're starting to do a bit of it.

Robert Luciani: 1:54:05

We see all these kind of image models already going in that direction, not so much momentum behind the existing models and that there's sort of clear paths to improving them, as is that nobody dares to start from scratch with it.

Anders Arpteg: 1:54:20

Yeah.

Robert Luciani: 1:54:21

And that is what you know. This kind of work is a little bit, but I think there's two avenues to research in this field. One is probably trying to model things after the way we think humans do things, and there's good reason to do so, because we believe that the optimizer of natural selection has done a very good job at making our brains, at the very least, energy efficient. So it's a good way to sort of process data cheaply. But there is probably another avenue of research, which is what is the best and coolest way to do it, and so I don't think they're mutually exclusive, necessarily. But I think one of the cool things that Clip showed us, and even before Clip, was that early language models improved between two languages when you taught them other languages that you weren't actually translating.

Anders Arpteg: 1:55:14

You learn across languages Exactly.

Robert Luciani: 1:55:16

And then with Clip. It's the same thing that understanding images made them better at language, presumably and this is where we speculate, and I just think some people are better at speculating than others that's just the way it is. But presumably there's some kind of semantic representation. That's why I really liked Jan LeCun's representation learning, like the coining of that word, that there's some kind of semantic representation of something in our minds that is allowed to occupy the space that gets activated for both the word beer, the sound of beer, the taste of beer and everything else. And there is a lot of scientific evidence to back that kind of stuff too. It's a shame that the fields aren't really converging on stuff and that people like Hinton are getting so much flack for trying to sort of speak across fields. Have you guys heard of the? What was the woman that played Storm in the X-Men called?

Henrik Göthberg: 1:56:12

Is this Halle Berry, halle Berryurton? Yeah, there's a paper called wow, that is the reference.

Robert Luciani: 1:56:17

There's a paper called the halliburton neuron and it was a guy that had electrodes connected to his brain where they were trying to figure out what parts of his brain were activated for certain kinds of stimulus.

Robert Luciani: 1:56:31

And they found out in the end that he had one single neuron that they could find that always activated when they heard, when he heard her voice, when he saw her picture when he heard movies about her and everything else and they ended up calling it the Halle Berry neuron and presumably that is sort of the, the one neuron that activates maximally when every other neuron sort of the hearing part of his brain and other parts of his brain, and I just feel like it's so analogous to the way we do things and these multimodal models in latent space. Of course they should have a single representation for multiple modes of input, so it feels intuitive to me at least, that that would be the energy efficient and good way to do it.

Henrik Göthberg: 1:57:10

And maybe this is also because we are talking about something that now becomes very abstract because we have our different senses, so it's really hard to think about it. Can the brain have another lingo that ties them all together? Why not? But it becomes fairly abstract.

Robert Luciani: 1:57:27

But here's what we can't do as humans. Oh my God, this is super. Now we're going sci-fi. Remember those papers that said that when robots were able to talk to each other I think it was meta they came up with a language of their own that was more efficient, that seems natural. Okay. Well, we can't speak to each other in latent space. That's why we use words, and it's a very what do you say Crude? No. What do you say Crude? No, it's very high compressed serialized High level abstraction.

Anders Arpteg: 1:57:55

Yeah.

Robert Luciani: 1:57:56

Serialized unidimensional format of transferring. And I remember there are these sci-fi books where aliens can talk to each other by sort of exchanging DNA via osmosis and sort of talking really fast that way In any case. Ai systems could potentially speak in latent space with each other and we wouldn't understand what they're talking to each other about. That's what Neuralink is about right.

Anders Arpteg: 1:58:15

Yes. That is exactly it, exactly so. The reason behind Neuralink is that the communication bandwidth between humans is horribly slow, like bits per second.

Louise Vanerell: 1:58:25

Sorry, that was my sort of jump into what you were saying with being crude, because of course our use of language is very crude.

Anders Arpteg: 1:58:33

That's why we put so much effort into very good, very low yeah and people are good at communicating so if I had a neural link ship in my head and another in henryx, we can actually communicate in gigahertz instead.

Robert Luciani: 1:58:44

Maybe would it require that you two have the same weights. That's a good question that you two had the same weights.

Henrik Göthberg: 1:58:52

That's a good question. We have to synchronize latent space.

Robert Luciani: 1:58:54

Synchronize latent space. Yeah, yeah, which is a good point.

Anders Arpteg: 1:58:56

It's probably very hard to do.

Robert Luciani: 1:58:57

But it is my dream to be like in Ghost in the Shell and speak telepathically. That we should be able to do with Neuralink.

Anders Arpteg: 1:59:03

Awesome. Before we end, I'd like to hear speak a bit more about Julia, perhaps, and just the future of programming, and the role of a software developer, etc.

Anders Arpteg: 1:59:15

so we we spoke a bit in the beginning about you know what will happen as we are getting more and more domain experts that may not be technical experts but still using chatbots efficiently, because you can simply prompt them in different ways. If you can prompt it properly, you can actually build stuff and do stuff. What do you both think, if we start actually because I know a bit what you were going to say, but I want to hear Luis first, perhaps what do you think a software engineer or someone that builds an AI system? It may not be, it would be wrong to call it a software engineer but what do you think the future will be? When we build AI systems in five years, how do you think that will look like?

Louise Vanerell: 2:00:01

Interesting. What will it look like? I think that there's still going to be, I mean, the people who say that you don't have to have any sort of field knowledge. If we then talk about building stuff, engineering, I have a hard time thinking that we're seeing that that's going to be obsolete in the near future, both tying back to the example that you gave about the woman at SVT, the old journalist who was so much better at sort of using the tools, and I mean there's something I don't have perfectly formulated.

Anders Arpteg: 2:00:40

But is she using the tool or is she actually building applications by prompting it?

Louise Vanerell: 2:00:45

Exactly, and that's probably the question. Then the follow-up question about what do we mean when we say building stuff, because as soon as we start going into this whole agentic thing which obviously is not my area of expertise, but what you are doing is giving people, uh, sort of the possibility of of building things by being able to give instructions, which, of course, can tie into yeah, how do we build? How does Elon Musk build this? You know Colossus? Yeah, he's really good at understanding what's needed and giving instructions to people or to something else. So, and we see him as somebody who builds stuff. So, of course, if you bring that sort of into what building means on a smaller scale, that is the same thing.

Henrik Göthberg: 2:01:33

Let me go next. Before we get even nerder like this, I think, when we are talking about engineering, I think, on the one hand side, we will get to more and more better self-services platforms that allows everybody, without the backend know, back-end coding skill or really infrastructure skills that cognitive problem is gone and now they can focus on building whatever problem they need to solve. So here we now get one type of engineering which is very close to what we do and which is then basically based on very efficient self-service architectures. So this is engineering in the one hand sense that actually we might have many, many, many more engineering people. So we are democratizing engineering close to the workflow flow. However, to do that the intricacies of building that self-service platform and to run that platform and to operationalize that and to have the quality controls and the data and et cetera, to build the agentic AI platform and to build the agentic AI monitoring and governance, to build the agentic or the data underneath the data mesh serving the agentic plane you know this is a shitload of engineering that goes into building these self-service products. So I think there are two distinctions of engineering that we need to think about in the future.

Henrik Göthberg: 2:03:05

And then the third final topic. So even when we are now going into the platform, the really sophisticated stuff, the engineering field, the toolbox that these engineers will use, will abstract away how we code. Also, we will work on a different level. But you need to be a very, very you know building robust systems. You need to be very knowledgeable of what makes the system robust. But the way you get there is not in assembly and it's not in c and it's not in java, it is in something else, it's in different types of editors. That was my take.

Louise Vanerell: 2:03:43

So three angles to that question yeah, I just want to add on that I don't think the the role of the sort of the expert, the person who knows all the ins and out of stuff, that's not going to disappear. It's just going the sort of the expert, the person who knows all the ins and outs of stuff, that's not going to disappear, it's just going to sort of change.

Henrik Göthberg: 2:03:57

If you build a platform, even if you're telling someone to do something stuff for you, you're instructing it. You kind of need to know the ins and outs and the blind spots of building robust systems.

Louise Vanerell: 2:04:08

In order for you to give good instructions, you need to know what it is that you're asking for Exactly, and that's not going to sort of suddenly disappear and I think that's super interesting and probably not for this conversation. But you know, how do we see the future of grunt work? How do you actually get to the point when you have good intuition? How do you get to the point where you are the most able to use all the different tools that we're given? And I don't think that's super clear and I don't think definitely not. It's clear. You know out in the industries and in the fields.

Henrik Göthberg: 2:04:43

There is a thought experiment here. If I will never any more work in the process, I will always work on the process because I will then set up an agentic flow in the process Then we all are engineers on some level. For our data work, for our workflows. We are engineering our workflows up here. So engineering thinking, or system thinking, becomes super important to be a guy that can work on the process rather than in the process. So then we're all engineers. That's an interesting thought, right.

Robert Luciani: 2:05:23

We're all AI experts, but it's like, oh, we don't need an engineer.

Henrik Göthberg: 2:05:28

So I just flipped the script. That's why I needed to coin this.

Robert Luciani: 2:05:31

AI wizard term so that there's another tier of it I think um, uh, there's probably two kinds of ai developers in the future one that uses a lot of pre-made stuff.

Robert Luciani: 2:05:43

And I know like if you base things on azure and microsoft stuff, like you know, microsoft will say, this is how agent stuff works, deal with it, and and you're like, yeah, I'm cool with that. Well then everything's great, because then you just press next, next, next, next, next and you can just build these incredible things. But as soon as you want to first of all transfer your knowledge from Microsoft to something else but also, you know, just edit it a little bit you basically need to know everything, and I'm sort of catching myself noticing how unreasonable it is to have to know all this stuff. You have to know infrastructure stuff. I mean, just consider this as part of the stuff I'm doing at SVT. We're using Redis.

Robert Luciani: 2:06:23

You know a NoSQL database as a state machine. Not everybody has studied state machines and know about race conditions and stuff like that. There's just so much you have to know more than what you had to know as a data scientist before and it's just fewer and fewer far between people that will be able to know this stuff. So I don't think it's going to get easier in the future. It's only like really enthusiast people that are going to be able to keep up. I feel like I can barely keep up with this stuff. But to sort of address the question of programming in specific, I was just a little excited. The other day Julia released a library called GPU kernelsjl, which is a way of writing GPU kernels that is a runtime agnostic. So if you want to run on metal, you want to run on Intel's platform or Cuda or RockM. You can just write once, run anywhere. Trademark.

Anders Arpteg: 2:07:18

But Will you code Julia in five years?

Robert Luciani: 2:07:23

I don't even code Julia today, which annoys me a lot. But no, I'm actually going to try and code some Julia stuff. Don't ask for permission approach.

Henrik Göthberg: 2:07:34

You had to regress to Python. Yes, unfortunately.

Robert Luciani: 2:07:40

It's impossible to rebuild all the AI stuff from scratch in Julia.

Henrik Göthberg: 2:07:45

It just takes too long. The hybrid is there, come on, yeah Well, nothing is written in Python.

Robert Luciani: 2:07:50

Everything's scripted in Python. Don't get me started. Don't get me started. Nothing is written in Python. Everything's scripted in Python. Don't get me started. Don't get me started.

Henrik Göthberg: 2:07:55

There's nothing written in Python. Everything is scripted.

Robert Luciani: 2:07:59

But no, actually the guys at SVT are experts in TypeScript, so they are super good at typing.

Louise Vanerell: 2:08:07

Why.

Robert Luciani: 2:08:08

You know, TypeScript, compared to Python, is even a good language. It's strongly typed. It has lots of support from tools that are really, you know, made for making big enterprise software.

Henrik Göthberg: 2:08:20

How did they get into TypeScript? What was the angle there?

Robert Luciani: 2:08:23

They started with front-end stuff and then Node and then TypeScript, because JavaScript, is you know, for building serious applications is not good enough, and so it became TypeScript. And, surprisingly, typescript is pretty good, cool stuff what?

Henrik Göthberg: 2:08:36

what are you coding right now?

Anders Arpteg: 2:08:41

not julia it's not in english probably no, no, no, it's still python, of course, but um, but also a lot of other. Like I love type script as well. I think it's actually very beautiful but you know who coded into something else. If you take DeepSeek, do you know what they coded in to build their models?

Henrik Göthberg: 2:08:57

Have you heard? No, I haven't heard that one PTX.

Robert Luciani: 2:09:02

Have you heard about it. Isn't that stuff straight for? No, it's not for Google's processors. They didn't use that. Did they use MTT stuff? Or did they use? They used NVIDIA right Old NVIDIA cards H800.

Louise Vanerell: 2:09:20

So what is it? What is?

Anders Arpteg: 2:09:20

PDX, so they have different levels. Like PyTorch is a certain Python level, you can go to CUDA level, which is basically C-level language. Then if you go to assembly language underneathDA it's actually PTX. Yeah, I heard now I get it which then compiles to the machine related language underneath. So they went super low in the other direction here.

Robert Luciani: 2:09:42

Actually but that's because they had to rewrite the NCCL libraries and stuff like that.

Henrik Göthberg: 2:09:46

That's exactly right. I have you heard the Lex Friedman, you know, five hour marathon on the AI. You need to. But they have a couple of cool guys and they're picking this apart and they were explaining the whole the geopolitics. They're talking about geopolitics down to coding, so very impressive guys and they were explaining this exact point that you know what? Because they had these limitations, they you know, and what they wanted to do with mixture of experts, the cuda library couldn't keep up, yeah, so they need to do it from scratch.

Anders Arpteg: 2:10:16

How many experts they had is like over a thousand, right, yeah, so anyone that says a deep sea guys just a cool kid. So five people that were on the side time and I did some stuff, that's horribly wrong, I think that's a great description.

Robert Luciani: 2:10:30

I think people underestimate I mean when I mean cool kids like you're not cool if you're that, that stuff is cool.

Henrik Göthberg: 2:10:36

I mean, these are cool hackers. No, but this is not bullshit stuff?

Robert Luciani: 2:10:41

This is hardcore engineering stuff. Yeah, this is hardcore stuff.

Anders Arpteg: 2:10:44

Yeah, really hardcore stuff. So these are really advanced people with a lot of skills.

Henrik Göthberg: 2:10:48

Yeah, and when someone, some people frame this as something else. Why would people frame it as something?

Anders Arpteg: 2:10:54

because it wasn't research, it was stolen, it was blah blah blah anyone could have done this thing because I mean the data is made it sound like it was super simple to do and all the value that the media had was useless in some sense, right? So all the time that open the eyes betting on this is not really worth it. And it's super simple for five. Kids that have no skills can do a lot of stuff, but that's not the story.

Henrik Göthberg: 2:11:15

It's so wrong and the story was literally I mean, everything just plummeted when in reality there's been some amazing engineering down on the PTC level.

Robert Luciani: 2:11:24

This is an ultra ninja team.

Anders Arpteg: 2:11:26

This is ninja team.

Henrik Göthberg: 2:11:27

This is proper ninja team.

Anders Arpteg: 2:11:29

Okay, anyway, I think you know one other good anecdote, or metaphor rather, to have about the future of software engineering, which is also from Sam Alton, and he said you know, they actually made a bet, I think, to Elon or something. No, not Elon, someone else, they're not friends. Anyway, they made a bet saying there will be a point in a couple of years where we'll have the first one-person unicorn company, a company that is worth a billion dollars yeah, a billion dollars and it's going to be started and operated by a single person. Actually, instagram was not that far off. It was basically 13 people when they sold it for 1 billion to Meta.

Louise Vanerell: 2:12:13

That's quite many years ago, it is.

Anders Arpteg: 2:12:16

But we still haven't seen a single company being worth $1 billion. But that's potentially going to happen, and if you think that's true, it means that humans, as a single person, needs to be able to handle a huge variety of tasks becoming increasingly general.

Robert Luciani: 2:12:33

Not generally bad, generally good. I think some people mistake that we need to be increasingly good at very many things.

Henrik Göthberg: 2:12:40

Yeah, but you don't need to do anything. You can ask your chat GPT for your market strategy. You can do this and you can do that. So I think it's the thought experiment you get into then. What would be the university degree or what would you study, or what is the competencies to build the first? Well, you know, if I'm going to be the first one man unicorn, what is my skills trait, what is my profile to succeed with that? What I need to have? I need to have some. I need to have vision, creativity, I need to have, you know, from zero to one thinking, peter Thiel. I need to have system thinking as well, because I need to figure this shit out, even if I'm prompting it, or blah, blah, blah. But it's an interesting thought experiment. What would that be in terms of competences?

Robert Luciani: 2:13:30

It is. I want to say one last thing. It is to the expert programmer. That's exactly what has happened with recording at home studios. People are like, oh, it's better to record at Ocean Way Studio. Yeah, it costs millions of dollars, and now this person can get 95% of the way there at his home studio, or maybe just as good. Anything that democratizes exclusive technical or physical skills or anything like that, I think, always lifts the boats for everybody. And just as a fun anecdote, this morning I listened to the Donkey Kong soundtrack, which was translated to sound like flamenco guitars by a sort of translation service, and it sounds perfect. My guitarist said it sounded like a real person playing guitars and that is awesome that an amateur can make something like that.

Anders Arpteg: 2:14:38

As a musician, are you happy with all the music generation services coming out now?

Robert Luciani: 2:14:43

All the musicians that I speak to that do it for the love of writing meaning they're poor think this is awesome. It's just more cool stuff to make art with, so we're all pretty excited about it. I think the record labels are probably mad.

Anders Arpteg: 2:14:58

I'm sure, I'm sure.

Henrik Göthberg: 2:15:00

What was your take on the software engineering based on the?

Anders Arpteg: 2:15:03

one man unicorn thought process I mean you can think about the skills. I think that's a super interesting question. What should you tell your kids to actually, you know, learn and teach in school? And should I quote Elon? People get so angry when you mention the name Elon these days, but I think still he has a good point there.

Anders Arpteg: 2:15:25

His thoughts is basically physics and computer science. And you can think why is that? Well, physics, I think, is rather easy to understand. You need to understand how the world works. In some sense, I think that's a super useful thing to learn and rather complicated Computer science. Well, you know, the world is digital. You need to understand how that works. Even though you may not be an expert programmer, you need to understand the basics of how computers work to be able to tell it how to build stuff. So I think that's actually a good point. If you compare it to other things like okay, do I need to know how economics works? Do I need to know how marketing works? Not, really.

Anders Arpteg: 2:16:04

Do I need to know how to build a factory? Or physics, okay, I think you know. Potentially going closer to the STEM fields mathematics and physics, and engineering in terms of computer science is probably, I think, is a good idea.

Henrik Göthberg: 2:16:19

So if you start with those two and then we have already spoken about the third one, communication skills, Good point. So I think you can never. If you want to build a $1 billion company, you damn sure need to be able to sell and communicate your ideas.

Robert Luciani: 2:16:34

So communication Interpersonal skills.

Henrik Göthberg: 2:16:36

Interpersonal communication skills, building relationships together. And the other angle. I'm reading Adam Grant there's a book called Think Again who's basically exploring the challenge of unlearning to relearn, so to speak, and he's arguing around how we can get more methodic doubt into our ways of thinking in order to see our flaws and fix stuff. So he argues then also, you know, another angle why the stem fills is that there is a difference between our mental models. Are you thinking like a scientist putting up a hypothesis, having a methodic doubt and then trying to reject it? Or are you thinking like a politician trying to convince people? Are you thinking like a prosecutor always trying to make your thing the right way, executor, always trying to make your thing the right way?

Henrik Göthberg: 2:17:34

And and he's sort of having an understanding that, whether we like it or not, we fall into these mental models, that we in the arguments and and the most useful thing, when we really want to understand our blind spots, our flaws, and even to the point like, if you want, if you're a startup, they were doing a survey and this is psychology, and people were training in um, in scientific thinking, and then ending up pivoting on their idea, and you know how many that came to success versus the one that sort of had. They were stuck in their idea and they were prosecuting everybody or a politician for their first idea. That was actually wrong. It didn't have market fit. So this STEM thinking also goes into fundamental scientific thinking and we are not all trained in that.

Louise Vanerell: 2:18:18

No, definitely not. I mean it's the difference between you know, when we're talking about communication or something in whatever, but the difference between arguing which is sort of making the other person admit or something, or just you know yeah exactly and having a discussion where you actually, you know, end up understanding what is the, what is the problem we're having, what is our view on it?

Louise Vanerell: 2:18:46

do I understand you, do you understand me? And then you can actually get to a point where you get to a proper solution of something and, of course, that is the scientific way this is the scientific way actually to have methodic doubt that I might be wrong, instead of me prosecuting you because you're not thinking like me anyway, like Trump right but okay, so we have the Elon approach with STEM.

Henrik Göthberg: 2:19:10

We add communication and interpersonal.

Anders Arpteg: 2:19:12

I kind of like it cool.

Henrik Göthberg: 2:19:14

Should we go with the?

Anders Arpteg: 2:19:15

Elon approaches STEM. We add communication and interpersonal. I like it. I kind of like it Cool.

Henrik Göthberg: 2:19:22

Should we go with the last obligatory kind of question as well? Perhaps? I don't think it fits today. Okay, I think we need to have another question.

Louise Vanerell: 2:19:28

We need to think it up on the spot.

Henrik Göthberg: 2:19:30

No, no, no.

Louise Vanerell: 2:19:32

I'm putting myself on the spot. What's?

Anders Arpteg: 2:19:35

the usual, we haven't heard Louise's points of this.

Henrik Göthberg: 2:19:39

Okay, good point.

Louise Vanerell: 2:19:40

Are you going to put me on the spot now? Let's put Louise on the spot.

Henrik Göthberg: 2:19:43

Let's put Louise on the spot. While you do that, I'm thinking about a final question that will blow your mind.

Robert Luciani: 2:19:51

Okay, you can take our question.

Anders Arpteg: 2:19:56

So, Luis and Robert assuming that there will be a point, perhaps a number of years in the future, where we'll have AGI, where we'll have systems that is at the level of an average co-worker, not only for perception but also reasoning and agentic kind of taking action abilities. What would that mean? You can think about two extremes. One extreme is basically we end up in a utopia, what Nick Bostrom wrote about in his book Deep Utopia, meaning we've solved, we've cured cancer, we've fixed the climate crisis, we have fusion energy and we have basically a world of abundance where goods and services are available for free, Like Star Trek.

Louise Vanerell: 2:20:40

I like Star Trek.

Anders Arpteg: 2:20:41

Yes, Then we can have the other one, which is the dystopian world, where we have the matrix of the world and we have the terminators, where machines are being built to kill us all and we're living in a horrible society. Where do you think we will end up? Will it be more towards the utopian, more towards the dystopian? What do you think?

Robert Luciani: 2:21:05

Super easy question yeah.

Louise Vanerell: 2:21:07

No, I'm definitely gonna go for the utopian one. I'm I'm a positive spirit but I think um, I don't know why, to how to reason to it, maybe because it's so boring?

Anders Arpteg: 2:21:19

are you afraid? Are you afraid that an ai system will exist? That is because I um.

Louise Vanerell: 2:21:41

I would say it would be a hard time to sort of judge what that would look like, because I don't think there's a such a sort of binary um divide between what is or what isn't. So that makes me fairly confident to say that no, I'm not worried. And I do think you know it would be so boring to think that all the good stuff that's happening would somehow, you know, lead us towards complete destruction. You know, lead us towards complete destruction. And as a historian I could probably say that you know when we've faced big shifts and it's hard to wager, you know which has been the bigger shift over history of time.

Louise Vanerell: 2:22:25

But people have always, always painted the picture of like this is bad, and of course, because we always predict the future, because we disrupt systems and ways and rules of how things have been established and worked, and that is, of course, really, really frightening for the people who are on top today. So I mean, of course people are going to be very negative about it or paint the picture of like this is you know, hell knocking on our back door?

Anders Arpteg: 2:22:56

And I don't buy that Awesome.

Robert Luciani: 2:22:57

Robert, do you have any? Are you capping the intelligence at human level intelligence.

Anders Arpteg: 2:23:02

No, no, it's ASI, yeah, yeah.

Robert Luciani: 2:23:04

There's no, there's no reason that I see where that we would be like the pinnacle of intelligence.

Anders Arpteg: 2:23:10

Absolutely not, and so I'm so limited myself, I feel pretty confident in saying that I am superior in every way to a five-year-old, any five-year-old I I would just beat them at everything, everything you can imagine, except maybe humor or being silly or something asking questions.

Robert Luciani: 2:23:31

Asking questions I don't know, man, yeah, I think so. So sort of. By by that way, I I can envision an ai that is similarly superior to us in every single way, and that doesn't necessarily mean it has to be the matrix dystopia. But I don't see how we are the ones in control of anything after that happens. No, we wouldn't let kids run our governments today. Why would the ai? But it doesn't necessarily have to mean that we go into dystopia.

Robert Luciani: 2:23:50

So if you go, control of anything after that happens, no, we wouldn't let kids run our governments today why, would the ai but it doesn't necessarily have to mean that we go into dystopia.

Louise Vanerell: 2:23:57

So if you go, that route?

Henrik Göthberg: 2:23:59

what? Where do you end up? Let's say that the ai will consider you a five-year-old.

Robert Luciani: 2:24:03

I've left breadcrumbs on the internet, so hopefully they remember me as one of the good guys I don't know, maybe they keep us in a zoo or something.

Louise Vanerell: 2:24:11

Yeah, but I just want to add on to that it's very dystopian, yeah. The zoo. Well, you know.

Robert Luciani: 2:24:17

I think dogs have it very good. Yes, exactly.

Louise Vanerell: 2:24:22

No, but I think my response to your question about if I was worried is that I don't see ourself as the pinnacle of you know, the height, but I'm not worried about it.

Robert Luciani: 2:24:35

No, but then children shouldn't be worried about their parents helping them out either. I mean, I would be so proud if an AI that I was part of making would be good enough to be able to help us to do better things. Why wouldn't one want that? Agreed.

Anders Arpteg: 2:24:50

Henrik, have you thought about some other questions? I have a question.

Henrik Göthberg: 2:24:52

I have a question. I have a question, let's see if I can. It needs to be this whole setup. It's a whole setup, okay. So we start with the setup.

Henrik Göthberg: 2:25:02

You know, way back Ray Kurzweil wrote about the law of accelerating returns and maybe we can today understand it as an accelerating productivity frontier. Sam Altman said you know somewhere I think it was in 2024, you need to think about the trajectory, not where we are. And then he did the example that if chat GPT 3.5 was good for I had quality for five second tasks, then uh 4.0 was good for five minute tasks, then the next one in line, if it's chat gpt five now is good for five hour tasks, and then you multiply that you get to a month right. So if the core question then becomes how should companies think here and now to better organize for a trajectory and not where we are? So what would be our advice or what would be your sort of number one?

Henrik Göthberg: 2:26:08

Okay, if you really want to think about this, not where we are. So what would be our advice or what would be your sort of number one? Okay, if you really want to think about this, not where we are now, but to be ready and to be surfing the wave and not surfing the AI tsunami and not drowning in it. What is the key ideas on how to organize for the trajectory? Where would you go with that answer if someone asked you?

Robert Luciani: 2:26:35

I, I think, I think a lot of companies should be asking themselves if they have the guts to make the changes necessary and who are the people in the company that will sort of spearhead that charge do we have people in the company that know the right stuff to make that kind of a charge.

Robert Luciani: 2:26:50

I've always felt that ai is always for the past few years, that AI is very much about entrepreneurship. You know you have to like, dare to reinvent the company, and I don't know if everybody has the entrepreneurial spirit or willpower or mindset or any of that kind of stuff. So I feel like it's linked to that very much and I know I'm sort of leaving out where things are headed or what you can do about it, but it's more like if you're fundamental, yeah, we're ready.

Robert Luciani: 2:27:16

I mean I'll be preparing to go, to go here look, if you're dropped into the jungle, uh, you can't say, oh, I'm going to prepare by doing this and this. No, you have your knife, you have your matches, you have your compass, and then you're. You can only prepare so much, but you can't prepare for exactly what's going to happen. You can only prepare by being ready.

Louise Vanerell: 2:27:34

Being ready but I mean that then, if I may uh ties into my question, which is practically what you just said, but phrased differently know your problems or know your business? I mean if, if, if the equivalent of knowing your business is, you end up on you're being dropped on an island and you know that I need to fix water, I need to make sure that I, you know, have some place safe to sleep. That is knowing your business. Without knowing that parts, there's no way for you to even start navigating about what you should aim your efforts for.

Henrik Göthberg: 2:28:07

So you need to be clear on your goal, know your goal, know your value, know what you're trying to do. I think this is one part of the story. I go back to this fundamentally understanding that this is if you're not ready or prepared to take on the trajectory, if you are not prepared to walk out of the jungle, because you're stuck in the fucking jungle, if you're not going to walk out, are you going to live in the jungle or are you going to die in the jungle? So if you're not going to walk out, are you going to live in the jungle or you're going to die in the jungle? So if you're not mentally prepared that I'm going to try to walk out, it will never fly so this is, this is something else.

Henrik Göthberg: 2:28:39

Right, this is knowing your business. I know a lot of people who knows their business, but they are simply not prepared to start walking. I think what you're saying is quite interesting about courage here one last thing.

Robert Luciani: 2:28:50

um, one of the things is that preparing for being dropped in the jungle takes a while. You got to build up some muscle, you got to do some stuff. A lot of people think we'll handle the AI stuff. When it happens, it's going to be too late.

Henrik Göthberg: 2:29:01

You need to learn now. You need to build your muscle memory now to be dropped in the jungle. Okay, good Thank you.

Robert Luciani: 2:29:07

We should take jungle analogy, really far in the next few episodes Cool, thank you.

Louise Vanerell: 2:29:13

Good answer, thank you.

Henrik Göthberg: 2:29:13

It's good.

Anders Arpteg: 2:29:14

Awesome, yeah Time. Oh, it's two and a half hours. It's too far Okay.

Robert Luciani: 2:29:20

We didn't even talk for so long about deep secrets.

Henrik Göthberg: 2:29:23

We even skipped topics. Anders, this is nothing, man.

Anders Arpteg: 2:29:26

Right, it's good. Well, let's stop it there, and I hope you can stay on for some after work discussion as well, off camera. I'd love to continue to discuss about quantum computers as well.

Henrik Göthberg: 2:29:39

It's one of my favorite topics.

Robert Luciani: 2:29:40

We don't call that Microsoft released a new.

Anders Arpteg: 2:29:43

Majorana chip.

Robert Luciani: 2:29:45

It was so superlative in all of the press releases and everything.

Anders Arpteg: 2:29:50

I have so many thoughts about it, are you?

Henrik Göthberg: 2:29:51

opening up another topic now. No, no, that's for the off-camera. Sorry to everyone else. Sorry to everyone else. Now the real cool stuff happens.

Anders Arpteg: 2:30:00

Anyway, thank you so much, luis Bernal and Robert Luciani. It's been amazing to discuss this and I'm looking forward to the coming discussion as well.

Henrik Göthberg: 2:30:09

Thanks guys, Thank you.