WEBVTT 1 00:00:00.000 --> 00:00:17.000 Alright, welcome to Neuvottelijanava. We are visiting with the main host of the AI channel and practically the master database administrator Lasse Mikkonen and then regular guest Harri Juntunen. Hello. Hi. 2 00:00:17.000 --> 00:00:46.000 Hi. This was a bit of a failed joke here well, regarding information technology, or ATK. ATK is actually a great Finnish word for automatic data processing. And we have just discussed for an hour on your channel, where I was a guest, so we turned the tables. And today's topic is something like 'vibe coding' and in a way this world of coding and how right now in 2025, even ordinary people have discovered the coder within themselves, so 3 00:00:46.000 --> 00:01:14.000 Could you, Lasse and Harri, provide a bit of background on what kind of backgrounds you have? Yes. So Lasse Mikkonen is the name, and I've been for 25 years a software industry consultant. From day one when I graduated, I became a consultant, and so I haven't really done real work but have always been helping companies to create better products, better services, and then for the last maybe four 4 00:01:14.000 --> 00:01:42.000 years now I've focused on AI matters around this same theme and indeed I also seriously engage in this kind of investment activity and I have some angel investments myself and I also participate in some of these startups quite actively, and then there's involved there, uh, not in a full-time capacity of course, but in an advisor role, and 5 00:01:42.000 --> 00:01:55.000 we are also setting up a fund. So with this kind of combination, and you could say that the software development work culminated when Eficode merged with our own company, Contribyte. 6 00:01:55.000 --> 00:02:24.000 There I really got to properly dive into the DevOps world, and that was when AI also started coming into play. Meaning how to automate the entire delivery pipeline, everything that happens after hitting enter while coding—how that code is then safely delivered to production and the services kept running. So this is essentially the field where I've been active. Yeah, and at this point, So, uh, like this, in real life, we met there at the main investor's... 7 00:02:24.000 --> 00:02:29.000 annual meetings, where we founded the main investor's coding society. 8 00:02:29.000 --> 00:02:33.000 So all you listeners and viewers who are members of the private equity... 9 00:02:33.000 --> 00:02:37.000 association, please just get in touch with us if... 10 00:02:37.000 --> 00:02:42.000 you're interested in making small code snippets to make life easier. 11 00:02:42.000 --> 00:03:00.000 Yeah. And then we already managed to open source a platform like this, a platform that can be used to analyze investment documentation, so with that, the FVCA coding subcommittee can get off to a good start, and we can also share it in a locally executable form. So, well, it was installed on my machine so that it runs without... 12 00:03:00.000 --> 00:03:08.000 this kind of, you know, cloud cloud model, where it just sends an API call to that Open works well. Yeah. 13 00:03:08.000 --> 00:03:36.000 Alright, so how do we match this? Well, maybe Lasse and I have known each other for maybe 25 years—well, not quite, but 20 years. A product development professional, me, and on the other hand, I've worked as a management consultant for a long time, and that's where Lasse and I started with requirements management and creating test cases and things like that, so my maybe the big picture here in my own professional 14 00:03:36.000 --> 00:04:02.000 professional sense is, well, product lifecycle management, so what it's like then, how you get, like, vibe coding like, the life cycle could maybe be here, but like, a bit something even bigger, because products live on the market, services live on the market, they need to be maintained, even that breaks. About this, 15 00:04:02.000 --> 00:04:29.000 we've discussed with Lasse on the AI channel, how the hell it can break, that software, when it’s just ones and zeros in a row going through a process. But anyway, from this perspective now and then AI has been on my agenda for about ten years now, how AI appears in these different kinds of places, 16 00:04:29.000 --> 00:04:57.000 what kind of sensible uses there can be in different parts of this cycle, and then with Lasse I think Lasse did some really concrete work on this, really well, like demonstrations of how AI works, and that's how it somehow started, so yeah, subscribe to the AI channel. Specifically, there will be quite fresh test cases coming out, like when someone has released 17 00:04:57.000 --> 00:05:10.000 a new feature from one of these large LL models, like a Google Gemini extension, and you guys pretty quickly put it into 'production' so to speak, showing how it can be used. 18 00:05:10.000 --> 00:05:17.000 Yeah. And we try to explain it in plain language and show with a demo how it works, so anyone can then benefit from them. 19 00:05:17.000 --> 00:05:31.000 Yeah. But anyway, our discussion on your channel is definitely worth go ahead and smash that subscribe button there, so you can join the ranks of the big channels, and it's great that you're doing that. 20 00:05:31.000 --> 00:05:59.000 Well, let's start with this like vibe or wipe coding. So, indeed as some of my listeners and viewers already know, I spent a year at a technical university studying CS, and then I became a finance guy and actually forgot forgot that whole world of coding, and in Excel I used the macros that were there, but when they introduced Visual Basic, even though I know Basic quite well, I thought that I couldn't be bothered to grind out any of those kinds of macros anymore. But then this year, I somehow woke up 21 00:05:59.000 --> 00:06:27.000 again. Now this no-code low-code stuff has come so far that now, now it's worth bringing Sami's teen version back from the grave, and it's been really fun. So, right now I'm mostly using this Swedish wonder called Lovable, which really is like a no- code tool. So you can just prompt. In practice, I made for example a website in an hour, which I had just earlier this year spent 22 00:06:27.000 --> 00:06:56.000 about six months struggling with a couple of students using HTML and WordPress there and writing them and trying to manage with some subroutine libraries there, so in Lovable I wrote that here is a template, change the format a bit, make it a bit punchier but copy those and those functional parts, then tap tap tap tap tap, a little I didn't have to look into this Google API call to see how those YouTube playlists are called through Google's protocol 23 00:06:56.000 --> 00:07:15.000 via it, but Dania, this is like such an easy case, this kind of frontend homepage, but if you were to translate this into professional language, what was I operating in when I created this kind of functional homepage, or what style did you use to make it? 24 00:07:15.000 --> 00:07:32.000 Well, yeah, it goes like that, Lovable especially is a good example of this 'vibe coding' tool but you could generalize it as this kind of agentic coding aid; there are many others, like Cursor for example, I think. 25 00:07:32.000 --> 00:07:35.000 ...be also Swedish, as you were able to say, and then I've used one like... 26 00:07:35.000 --> 00:07:41.000 Cline, it's also very similar to Cursor, so they are again those kinds of... 27 00:07:41.000 --> 00:07:46.000 ...that I think can be found as web interfaces for both, but they are implemented differently... 28 00:07:46.000 --> 00:07:53.000 ...then in Visual Code and then, or rather in VS Code, and... 29 00:07:53.000 --> 00:08:00.000 ...then I personally use GitHub Copilot the most, which is also a Visual... 30 00:08:00.000 --> 00:08:04.000 VS Code extension. Its special feature is that you can put it... 31 00:08:04.000 --> 00:08:08.000 into a sort of query mode or agent mode. If you put it... 32 00:08:08.000 --> 00:08:13.000 into agent mode, it becomes this kind of vibe coding tool. 33 00:08:13.000 --> 00:08:19.000 As an extra bonus, you can also change the model there. So sometimes or... 34 00:08:19.000 --> 00:08:23.000 I'm actually testing things all the time, maybe in the spirit of the AI channel. 35 00:08:23.000 --> 00:08:52.000 I'm working on a video theme to present exactly the differences between the different language models in the background, like Claude's Sonnet 4.5 versus Codex GPT 5 or Grok, as these three are, in my opinion, competing pretty much head-to-head. Sometimes it feels like Codex works the best, and then it suddenly gets stuck. Then you can continue from that same point by switching the model to Grok, for example, and then you can move 36 00:08:52.000 --> 00:09:20.000 forward. So it's kind of hard to even say which of them is the best for what, but it can be said that Loveable is a fairly safe bet when building a user interface, and it has its own database, but the components are sort of pre-selected, what what it then utilizes, but in turn it provides quite solid output and works quite reliably in their environment. Whereas 37 00:09:20.000 --> 00:09:49.000 then these other tools, and especially working through GitHub Copilot, then you can choose the language, databases, everything just by prompting and yeah, there, uh, Harri, when I then returned to this when that, uh, Chat GPT 3.5 came out, so I set it up with VS Code, so that I ran it there on Windows, that text editor and with it, like, the lines of code are there and 38 00:09:49.000 --> 00:09:57.000 Python is a fairly simple, like, interpreted language, which, well, was already developed back in the 80s. 39 00:09:57.000 --> 00:10:19.000 Yeah, so then I learned there that there's this API call, which you can like use to call the AI, but then it was too hard and there were too many errors, and since there weren't any 'training wheels' back then, well then I got fed up and only returned to it this year. But that's also kind of interesting, that it hasn't actually changed all that much. There's just been this kind of assistance added alongside it. 40 00:10:19.000 --> 00:10:48.000 Yeah. And, I mean, you typically have some reason why you want to build, say, a website or some software, so this gap from mind to matter has, in my opinion, radically shrunk now; as you described, a hundred-hour job might now be like an hour of work, and you're already like doing things, you're selling the product on some website and stuff like that, so I don't think that at this 41 00:10:48.000 --> 00:11:15.000 moment it's quite understood yet how much this, like in this prototyping phase, this radical product development, like time spent on product development versus market need, so how much it ultimately affects things. But it is like an incredible opportunity right now. And especially, well let's think about just some sales situation, where you would do a sales meeting and spend an hour making a PowerPoint, well now you 42 00:11:15.000 --> 00:11:43.000 spend an hour and make a working prototype. You go with that So, or at the sales event, you're already coding directly whether this is the kind of thing you want to buy, so to speak, so you get—there are no static presentations anymore; instead, they are more like, 'Hey, let's pull these features together for this,' so all that specification work and such can be present right there in the situation, and then once everyone agrees—especially the customer: 'this is what I want to buy'—then 43 00:11:43.000 --> 00:11:59.000 you can go straight to production. Of course, there are many other things in the background that need to be handled, like data security and such, but this whole prototyping aspect has, in my opinion, changed really radically now. And what I personally understood here... 44 00:11:59.000 --> 00:12:29.000 then there, when I was sort of using that 3.5 two years ago, running a couple of simple Python programs on my own machine, which also made some API calls, and then I realized why I went to Bubble, so that the database was right there, as I don't want to run on my own machine, I want to run in the cloud, but I don't want to start setting up any Azure or AWS instances, so Bubble solves that too. So it has the databases and the front end and back end and security, so they come from there. Okay. They are provided for a monthly fee, 45 00:12:29.000 --> 00:12:35.000 so my credit card is kind of stuck there as long as I'm running that system. But you don't have to do any of that. They take care of it. 46 00:12:35.000 --> 00:12:49.000 everything and then they force it to be quite simple. I would have actually wanted to do it with WordPress and PHP, the database like the, no wait, the homepage like the old one, and they said we don't do these, are these my tools okay, and then I said well let's go with that then. 47 00:12:49.000 --> 00:13:16.000 Yeah, exactly. And it's not expensive compared to if you ran it in Azure, the same thing, so you'd have to be quite a wizard to get it for that same monthly price, yeah running there. And then that demoing is interesting, that in a way I can now also make like a minimal viable product MVP for like certain use cases. For example So here we are now, headed to Slush together with Lasse, and I tried something for that too, because with all due love, Slush's 48 00:13:16.000 --> 00:13:42.000 UX, or user interface, is a bit, well, stiff, so I made something so I can analyze who I'm meeting there and others myself, and I ran a database of it there and built this with you a bit, like how we can enrich this data, like LinkedIn profiles and such, and it's again much easier than hammering through the Slush user interface to find if some guy is there, so I can click it myself, or see what the description of the guy and the company is like. Mm. 49 00:13:42.000 --> 00:14:09.000 And that took four days now, so it wasn't hours, but it wasn't like just incredible, like, well, this—or for me it's just like absolutely amazing, that you can even do this. So, so, so, so and eventually if we think about, like, the graphical user interface, so when it becomes this sort of generative user interface, then you don't need those menus or analyses, so you just have like sweet data, like, like 50 00:14:09.000 --> 00:14:35.000 you just have like data there and then the app is generated for your needs there for instance, and it enriches it enriches it and, and somehow that you don't need to like program the, like, interface anymore that is meaningful to a human, but instead you are directly somehow with that data, in which case your purpose, those questions, Why do you want to be there at Slush, who do you want to talk to? 51 00:14:35.000 --> 00:14:48.000 And why, so in a way it's no longer that kind of, like, clicking work, but the analytics come from there and you are able to, like, optimize the experience in that way. 52 00:14:48.000 --> 00:15:14.000 Yes. Yes, but that also required some vision, that through roughly that kind of HubSpot-style user interface it should come, and it maybe guides a bit too much, meaning what others have done before. But then I started thinking, since I'm a partner at Translink Corporate Finance, so for example, we produce this kind of SaaS company valuation report, so it's the kind of thing where younger colleagues do a lot of work for them to start gathering data into some Excel that creates 53 00:15:14.000 --> 00:15:41.000 graphs from it. So just for fun, I reached out to Standard Capital IQ and asked if I could get the API calls directly, because I wanted to see if I could sort of automate the creation of this report, and I was apparently the first person to suggest this globally, and they said it's an interesting idea, but that is also kind of like, taking this data source from here and that one from there, then creating your own analysis, and then the system sort of builds it, and then 54 00:15:41.000 --> 00:15:44.000 you can even build the user interface for it yourself, or like... 55 00:15:44.000 --> 00:15:46.000 Yeah, a really interesting world. 56 00:15:46.000 --> 00:15:50.000 Yeah, exactly. And so, what do they say over there on the standard side then? 57 00:15:50.000 --> 00:16:16.000 Well, they say that, well, now we mainly have these like coder types. Great that some, well, investment is closely connected, that we're still thinking about that. At the moment, our policy is that they go through like, via the Cloud, and then I said, I haven't yet opened the credit card for that Cloud, so I have to think if I can be bothered to go there. But then someone immediately started on this, but they were interested and didn't say absolutely no, and then they said, well 58 00:16:16.000 --> 00:16:23.000 ... that Excel first, and I'll look at it on this end, so that, well, that how it goes then. Great, great. 59 00:16:23.000 --> 00:16:51.000 Of course they have, well they naturally have, as they are a global data integrator, so they don't want me to, like, do too much, because they want the features I create to perhaps be taken through them, through their user interface and shared with third parties and Deutsche Bank and so they aren't just Translink's tinkering, so here of course the data owner probably, and for good reason, suspects that people might kind of steal their data. But it's perfectly fine to process it with Excel and 60 00:16:51.000 --> 00:17:15.000 make your own report. In my opinion, there's no difference to the fact, that now you're not just using Excel, but you're using Lavaboli, you're coding With JavaScript and HTML. I'll have to present this to them then, but but like in a way, I find this mindset interesting, because up until now I've been kind of a slave to the user interface, so if I don't know how to use some Slack features or these specific Hubspot database 61 00:17:15.000 --> 00:17:45.000 functionalities or, for example, the Statistics Finland data series user interface, then I just don't use it or then I'll learn from some videos how these work now and from where to find that and how to combine these, but in this new world, you can like, via AI or maybe even with API calls, these kinds of like data request calls, which are like AI-shaped. Isn't an API call also—I mean, if it were explained, 62 00:17:45.000 --> 00:18:13.000 so what is it actually? Well, it's maybe a general term for it. It's like Application Programming Interface, so there can be many types of those as well, many types, but the most commonly used nowadays is REST, so it is a web-protocol-style call. So you basically provide a short, human-readable message, usually in JSON format or XML 63 00:18:13.000 --> 00:18:39.000 format, and you call some endpoint or function there, let's say it's called from a CRM, for example, that the customer company's data is the call, and then it returns the address, name, phone number, and for example, a list of what has been sold to them. So it's like In its simplicity, one API call. It's a programmable interface, as in a... 64 00:18:39.000 --> 00:18:42.000 ...interface in the Finnish language. I might take this even further... 65 00:18:42.000 --> 00:18:47.000 ...by saying that I think software has always been... 66 00:18:47.000 --> 00:18:53.000 ...a giant brake and friction between the question, the idea, and the data. So... 67 00:18:53.000 --> 00:18:58.000 ...now this friction is rapidly melting away, so you can... 68 00:18:58.000 --> 00:19:03.000 ...actually be in direct interaction with that data. Yeah, and then I just noticed... 69 00:19:03.000 --> 00:19:06.000 ...there that you do have to understand a bit how... 70 00:19:06.000 --> 00:19:10.000 ...computer logic works. So, in this Lavapoli, I had this thing where... 71 00:19:10.000 --> 00:19:13.000 ...I turned security on, and then... 72 00:19:13.000 --> 00:19:17.000 ...well, Lavaboli thought it improved security that one... 73 00:19:17.000 --> 00:19:45.000 the database is like locked for me as a user so that it's then not written to, and it was really stupid because it was like just a basic database of those investors, so then I thought the program somehow went broken when I could no longer update that investor database. I started changing the code, like change this now and copy that and try this and then it was like accidentally revealed to me that it had just locked this with this security feature. So, but then in a way since I now had still a bit of this architectural view that 74 00:19:45.000 --> 00:19:52.000 even this kind of thing is possible, so I haven't like sent a fifth call anymore saying, fix it better now. 75 00:19:52.000 --> 00:19:56.000 So the cause was approached from somewhere else. Yeah. 76 00:19:56.000 --> 00:20:12.000 Yeah, yeah. I mean, of course, the basic pattern needs to be there, but semantics, ontologies, all these kinds of things are always linked to data, and above all the valuable questions you want to know, so we can get a handle on these much faster now. 77 00:20:12.000 --> 00:20:41.000 Yeah. And then, well, that's sort of this hallucination architecture, since in a way JSON is just this kind of text format. And often these are like really old protocols, it's just been decided that there's this kind of text format where text structures are sent, and that's how everything travels now, but but it's sort of like a free-text format can be for the AI, but sometimes it results in the kind of hallucination where it interprets inaccurately and can once return a slightly different thing than another. So, here we have this kind of 78 00:20:41.000 --> 00:20:45.000 tackling of this problem seems to be underway. 79 00:20:45.000 --> 00:21:10.000 Yes, and at least that ChatGPT or OpenAI API that I have used most in Ohmoni, you can turn on a setting there so it always responds in JSON format, or even follow that schema, but there's always the probability of hallucination where even the schema goes wrong, because of bad luck and the language model decides to choose that solution as the best. 80 00:21:10.000 --> 00:21:35.000 Yeah, yeah. But I mean, surely those the quality of the outputs can be measured in many ways, as it stands currently, so even though the model's internal operation might be a neural network such that we can't quite see everything in there, still the outputs can be verified, and then we can perform completely deterministic checks that are not done by the language model, but 81 00:21:35.000 --> 00:22:02.000 rather this kind of evaluation-type schema is already quite advanced. And if you combine that with asking for the answer again, it will likely work in your favor, and the fact that you don't have to get it in one shot right away, right, but you can do, like, multiple iterations and the fact-checking can be done from, say, any trusted database. 82 00:22:02.000 --> 00:22:27.000 Mm. Which is then brought into, for example, that response then. Yeah. And then there are these like, you know, this GAN model, where in a way another AI can be this kind of monitor, a quality controller. In fact, isn't it the case here—I mean, since I know Eficode, it's also because my study mate Juha Borsos owns this kind of, or is like a general partner in this Bootti, which was, like, a significant investor in Eficode 83 00:22:27.000 --> 00:22:54.000 before, so they have this kind of DevOps philosophy and self- As an investment banker, I was helping Nitore, a leading Finnish agile software development firm, find some new owners, so I think code development has changed a bit in this regard, but DevOps is probably also changing; explain a bit what DevOps is and what the philosophy is, briefly; I've 84 00:22:54.000 --> 00:23:24.000 explained it, or if one explains it briefly, it could actually be this: everything that happens after a coder has pressed enter or actually performed a git push, i.e., pushed the code into version control, then automation takes over and it goes through, possibly depending on security and other levels of ambition, it goes through looks at the code to see if, for example, good coding practices are followed, 85 00:23:24.000 --> 00:23:52.000 formatting, and whether there are security issues, and then runs automated tests, creates an installation package—it depends on the language; if it's a web application that is then deployed to another server, it might create a Docker container or an image out of it, and then proceeds to move it to the next development or test environment and finally to production. So it's automation around the 86 00:23:52.000 --> 00:24:05.000 production of that code, but in a broader scope, even before coding there is DevOps. For example, definition work, its documentation management, and perhaps standards there. 87 00:24:05.000 --> 00:24:32.000 even standards, and then especially the change management around it. So how we track and first analyze a bug report, for example, see everything it affects, what we need to change. Then perhaps people manage the workflow with some Kanban or Scrum process or any project management method. And then all of these are like DevOps on a broad scale, meaning the software—the 88 00:24:32.000 --> 00:24:41.000 support functions around software development, you could say, if the very core is what the coder or agent writes there. 89 00:24:41.000 --> 00:25:08.000 That is probably a pitfall for this 'vibe coding'. I mean, when I was fighting that one database update bug, well I had to, like, in terms of version control, I then went to a version that worked. I tried to bring code from there, but the copied code didn't work. So, because the error was at a different level, meaning in the write permissions, it didn't help, but I had to struggle with that. But then I, like, tried to look at those versions a bit. My brain wasn't enough to, like, 90 00:25:08.000 --> 00:25:33.000 analyze the code at that level, but this is maybe a problem with this vibe coding, that when it might come up with some solution, it could be any kind of messy code that is totally different from the previous version, so the development history can be very random. Yeah. And then, like, it happens really often to me too, that there's a bug or the code doesn't work the way you want it to. Then you give it some sample data, saying that this should now 91 00:25:33.000 --> 00:26:01.000 pass, but then the error is actually in your input and you ask time after time for the AI to change the code, and it does change the code, so that like 150 lines have been rewritten and the error finally reveals that your original request was wrong. And then if you try to ask to go back to the original, it writes even more code, supposedly returning to it. And yeah. But I, I, I'll maybe throw in. I found this maybe a month ago, something like, called Bllizle 92 00:26:01.000 --> 00:26:29.000 service. Yeah. Where we have sort of gotten rid of these human worries, that you can put into that Bllizle service and pay 100k a month or something ridiculous, but you can throw in there say, a bank's COBOL mainframe code, which has all that, and it comes back in a week, like, I've refactored this thing. So now in this AI era, we've noticed that these software 93 00:26:29.000 --> 00:26:50.000 development metrics and methods don't necessarily measure the way humans develop software, nor how we could develop software with AI, because because you don't have that, it's not found on GitHub The kind of task like refactoring the Goldman Sachs COBOL system. 94 00:26:50.000 --> 00:26:54.000 But for AI it's possible, and this is an absurd task for a human. 95 00:26:54.000 --> 00:27:20.000 Yeah. And then in a way, that inhuman machine like coding logic might emerge, so it's not just like some code snippet trial-and-error copying and merging, or it could be that the machine could even invent its own code. So in a way the machine could learn and crunch that code and optimize it like Yeah. But now we have decided to do it this way, so it runs there on Git. And actually this is like maybe good 96 00:27:20.000 --> 00:27:48.000 to explain, that for me even that is like developed by Torvalds So, like a repository. You have, in a sense, a standardized way to store code in the cloud. So couldn't you say that G— Let's look at it closely. So Linus Torvalds created a version control system called Git, and its original idea was that we can actually, in a distributed manner, develop and distribute code among developers located somewhat randomly anywhere. 97 00:27:48.000 --> 00:28:16.000 And it is actually based on the fact that you always have a local repository. The entire code is local to every developer, and then from there, those deltas, meaning changes, can be shared from one coder to another. And now, these GitHub, GitLab, and all other centralized or these kinds of central-server-based systems Git-based version control systems, which go slightly against that original distributed 98 00:28:16.000 --> 00:28:24.000 idea, but they are made so that we have one master repo, which is now there on GitHub's server, and it also has web interfaces. 99 00:28:24.000 --> 00:28:28.000 That's where the big help came in, because originally Git didn't have any interfaces. 100 00:28:28.000 --> 00:28:53.000 Being able to go look at code and download it via a git clone address— nothing like that existed. So these have brought usability to it and the idea that a main master repository exists somewhere in the first place. Yeah, and that was it for me when I started writing that Python code with VS Code, 101 00:28:53.000 --> 00:29:18.000 it really annoyed me that ChatGPT's codex said that I goddamn have to go there myself, to like, well, there to the command line to do some git stuff, that's where my patience ends, like okay. I have a bit of an opposite experience. It's like that, well, I'm a passionate vibe coder. I know how to do like Hello World in Java and then with Python, but nothing more than that. 102 00:29:18.000 --> 00:29:46.000 B- but but now that these like Claude's toughest features, they are damn good at it, so it starts suggesting, the AI from there, that do you want to install a python package here that handles version control for you, and then you just click ok. OK. And then when you have the architecture in your head, what this should be roughly it, and it's actually surpri- I mean I'm getting working software done without me knowing how to code, so it doesn't really yet 103 00:29:46.000 --> 00:30:14.000 fit right into Git. You were talking about a slightly different thing. Well, well, it could be, but I just mean that in that version control and everything else, the AI suggests to me to create this kind of GitHub and create this kind of version control and all that. So I could still return to that for a bit, to what it changes in the DevOps world. Now we were talking about how coding changes and design changes. when we remember that DevOps included all these, so there's now 104 00:30:14.000 --> 00:30:41.000 clearly one thing I've done myself quite concretely because these In DevOps tools, especially GitHub and GitLab, there's a lot of talk about how automation is built with these Actions or like pipelines. So when the code goes into a certain branch, then those Actions start to happen, and they are practically scripts that are then run in a Linux environment. Whether it's 105 00:30:41.000 --> 00:31:11.000 virtual or local and so on, creating these scripts used to be quite laborious and required specialized expertise, because you need to understand so many different layers: those individual Linux/Unix level scripts and then their automation in GitHub's language or GitLab's language and then possibly Docker containers or Kubernetes environments, like infrastructure as code, these pieces need to be made to work 106 00:31:11.000 --> 00:31:39.000 together, and now creating those seems to be, for AI, really child's play, because they are practically scripts of a few dozen lines at most. They might be just two or three line scripts, but from small pieces, a pipeline description like that is assembled, so these can be made in a way that you can automate it and then you can also generate the test cases. Whereas before it was said that testing is always 107 00:31:39.000 --> 00:31:53.000 a compromise, that we can only test 20% of features because no one has the time to write test cases. Well, that's no longer true. Now we get 100% test cases. Of course, those are written by AI. 108 00:31:53.000 --> 00:32:14.000 So if a person spent their time going through the ready-made test cases and looking at which ones are current and which are not, then we'd reach a completely different level in quality assurance and also in automation. You no longer need to wait for the DevOps team to come and write those automation scripts. They are obtained by requesting the code's continuation. 109 00:32:14.000 --> 00:32:23.000 Yes, this is exactly what I was trying to say, that since I don't know how to code, then AI quickly coded everything for me that was in the catalog. 110 00:32:23.000 --> 00:32:52.000 Yeah. Maybe then regarding this tech stack, sort of like a full-stack coder has always been the god of the coding world who knows everything and runs all sorts of things. So back when I was coding, everything kind of boiled down to the hardware you have. So, I myself started programming on the Commodore 64 with BASIC and even some machine language, so everything is sort of hardware-dependent; everything in the end is eventually compiled with something like C or assembly, at the latest into 111 00:32:52.000 --> 00:33:18.000 the machine language that the processor understands, but that container thinking has sort of brought something new, that everything kind of runs everywhere and you just bring the stuff needed to run it in some environment, so this has also been, in my opinion, liberating, that everything sort of runs everywhere and the machine's raw performance is no longer the main thing, everything just, everything just runs. 112 00:33:18.000 --> 00:33:21.000 somewhere and then AI handles the platform it runs on. 113 00:33:21.000 --> 00:33:46.000 Yeah, well in practice, to put it simply, it's virtualized— virtual, even a container is like a virtual machine, a virtual computer, which usually runs the operating system on Linux, so it's just a standard way to do virtualization that can be run anywhere. But, but really, it's not necessarily like a walk in the park, as I've also noticed when encountering VIP, 114 00:33:46.000 --> 00:34:13.000 that I've also become a VIP debugger and a VIP nerve-loser, when it didn't quite work after all and so on, because when you don't understand it at a really deep level, there can be like microscopic errors and others, so that you sort of at least get that prototype stuff done quite well, and test cases and all that, but if we start talking really seriously 115 00:34:13.000 --> 00:34:41.000 about, say, moving money or something else, then yes, there still need to be those quality gates in between, so that yeah, or if there's an existing service, I wouldn't right away start building on it. Well yeah. Neither would I just throw, like, a bank's mainframe onto some rapid coding platform, but just this prototyping nature and this kind of thing, that you get so much further with the functionality and everything, but if you're going to make like 116 00:34:41.000 --> 00:35:07.000 serious business, so there is still like work to be done. And then in some cases, performance is the deciding factor, so you shouldn't start writing code for a quantum computer with something like Python, it has to be done as fast as possible, and you have to prioritize performance, and then again in things like card games and such, things like pixel-level detail and others are so important that 117 00:35:07.000 --> 00:35:13.000 you won't get by with a tech stack that's just 'close enough'. 118 00:35:13.000 --> 00:35:41.000 No, but what I've discussed with other coders and what I do myself, I mentioned that you can put GitHub Copilot, for example, into 'ask' mode or 'agent' mode, and that agent mode is where it agentically modifies the code, completes it, and even runs the code and checks the logs to see if it works. So it can also test. But then this Ask mode is sort of a limited version of it, where you 119 00:35:41.000 --> 00:36:09.000 can provide the exact same prompt, but it just tells you how it would be done and shows you the code. That this code should be changed by this much and this kind of snippet should be added to this code. So I use it quite a lot myself, to see that, oh, that's how it thought it should be done, and then I might go and code it myself, sort of correctly inspired by that. Or sometimes if it looks like that's such a 120 00:36:09.000 --> 00:36:14.000 simple snippet, then I copy it, read it through, and then sort of accept it. 121 00:36:14.000 --> 00:36:39.000 into that entire codebase. So you could kind of think of it as, a bit like what Visual Studio or VS Code does in general, helping you to— put brackets and commas in place more or less. So this is like an extension, that in a similar way helps you write entire functions and review the logic of if-statements, for instance, better or faster than a human would. 122 00:36:39.000 --> 00:36:56.000 Yeah, yeah. And I mean, good point, that binary thinking often isn't very smart, so there is of course a spectrum of doing things, how much autonomy you give, for example, to the code—I mean, agent or AI, versus how much control you want to maintain. 123 00:36:56.000 --> 00:37:06.000 personally, I think you should think about these a bit like a centaur, where machine and human are joined together. 124 00:37:06.000 --> 00:37:25.000 Yeah. I've had Mikko Alasaarela as a guest a couple of times and he has really properly dived into this agentic activity and even founded a company for it called Agentics Finland, and during Slush week we have a third gathering there. So, what is this agent activity exactly? 125 00:37:25.000 --> 00:37:53.000 Agent coding? Well, there is some kind of explanation for this as well, but in short, I think it's... this is how I would describe it, that it is like, if a language model and a bot made from it is like a first-level robot that you always have to give a question to and it gives you an answer and with that bot it can be just a raw language model like ChatGPT. You ask it for example to write a poem or 126 00:37:53.000 --> 00:38:22.000 then it can be with these kinds of documents or something else, a trained bot that then knows how to, for example, the company's contract matters or customer information to answer. But then an agent is like a step forward from that. So it's this kind of bot that can still do some things. It knows how to do things, that it knows how to, for example, send an email. So you can ask it to send Sami an email to invite him to lunch next Tuesday. M 127 00:38:22.000 --> 00:38:44.000 and go check his calendar too. So then it would be able to, like, [the tasks] given to it tools to use for assistance. If I have authorized it for my calendar and my email, it can go and complete that whole task from start to finish. And now this coding agent, well it is authorized to change the code at various different levels. 128 00:38:44.000 --> 00:39:12.000 Yeah. I think the Finnish language is playing a bit of a trick on us here, in that we have 'agent' and with an 'i' at the end, it's 'agentti'. But 'agency' in English of course means agency, that you have like agency, that you have like some kind of autonomous agency. So whether it is then your agency with that AI, so do you have the agency to use your credit card to book hotel rooms or trips or things like that, so there is 129 00:39:12.000 --> 00:39:40.000 exactly this functionality. where some software acts on your behalf without you defining it in advance. And I think Lassi put it well, that the actor or the agent is also able to plan its own task. Yes. We actually had a big commotion a few years ago when Osuuspankki moved from this kind of managerial prerogative-based organization to a self-organizing one, and 130 00:39:40.000 --> 00:39:58.000 there are a few legends about how well it worked in practice, but it's a bit the same thing, meaning giving autonomy or reducing that strict managerial prerogative and giving people responsibility, and with AI, we can then regulate these rights to act. 131 00:39:58.000 --> 00:40:27.000 decisions and actions. And I have this, sort of, way you can basically test quite easily, that in this web browser there have now come quite a few of these agentic or delegation rights for it, so one of them is this comet.perplexity.ai perplexity.ai, just like a new web browser, so it gets this AI window, and there you can write commands for that browser, and I tested it when 132 00:40:27.000 --> 00:40:34.000 it was brand new, so I did exactly that. So I put it in HubSpot, I put it in an email, and then I said, like, copy a message from a colleague. 133 00:40:34.000 --> 00:41:02.000 Luckily, it had this feature where it still asked that I don't like hit send, because it sent like a document that I really wouldn't have wanted to be sent to this guy, and then another one was I sent, then I went to meet a lawyer over at Fisken på Disken, so I tried if it could get it, like, the restaurant booked. Well, it didn't succeed in that, but it sent the invitation, but it sent it so that the address was completely wrong in it, and in that it didn't ask anything, so it went like directly into that lawyer's calendar, 134 00:41:02.000 --> 00:41:30.000 so here is this delegation, that whether there is still some kind of, like, human check, so it's really important that you don't at least right away delegate it by pressing. And, and so now we are like at the edge of this kind of AI distribution channel, these browser wars and others that are currently ongoing, we're essentially in the early stages, the first light bulb has been lit, but then in exactly a year's time, that kind of delegation will be quite routine, so that 135 00:41:30.000 --> 00:41:54.000 yeah, but even then you have to be careful if the other person says, 'what's your...' I don't know anything about it'... well, and the agents have apparently talked to each other, so we might live in this sort of deep fake world a bit there in communication... but then again, I don't think it's necessarily a good use of human time if you haven't familiarized yourself with what we would be, for example, negotiating about. Well, exactly. 136 00:41:54.000 --> 00:42:16.000 It is, but I mean, if it pressed 'send' a bit too unexpectedly, unsolicited, too early, so like yeah, yeah, yeah, but we'll have to learn culturally and otherwise, and then probably various things, we have to learn to trust programs like people, how much slack can you give to Tampio there, so well, it's like a smart summer trainee or a dumb summer trainee. 137 00:42:16.000 --> 00:42:32.000 The only thing is, it might change overnight to be a bit smarter or less smart. That's how this vibecoding is, you don't know what kind of Miettinen has pulled a completely new code there, like you trusted for a year that it was fine, and now some Miettinen vibecoded it. 138 00:42:32.000 --> 00:42:35.000 ...like over again, and what came out of there was total garbage. 139 00:42:35.000 --> 00:42:39.000 There's actually a pretty funny story about this from another perspective, where... 140 00:42:39.000 --> 00:42:44.000 these models really do get updated, and an organization had just learned... 141 00:42:44.000 --> 00:42:48.000 to trust these answers, whatever comes out of there, and had done things, and then... 142 00:42:48.000 --> 00:42:53.000 when it was updated, OpenAI did this kind of 'out of the blue' type of... 143 00:42:53.000 --> 00:42:56.000 update and didn't provide any continuity for it, so then the... 144 00:42:56.000 --> 00:43:01.000 organization was completely at a loss because these answers were total crap, like... 145 00:43:01.000 --> 00:43:08.000 like come on, how, because it's interesting, it's semantic in a way, that... 146 00:43:08.000 --> 00:43:11.000 thing, so the model update actually led to the fact that the... 147 00:43:11.000 --> 00:43:15.000 organization received different information than what they... 148 00:43:15.000 --> 00:43:31.000 was used to. Yeah, and it was classic how people said that the five is blunter than the previous one, that it left those pleasantries aside more. It's different, so quite a vibe-based feel, could I put it that way. 149 00:43:31.000 --> 00:44:01.000 Yeah, but those are like amazing tools and can increase work productivity, or then it can lead to people making demos of things that don't even solve any problem. Just like it's nice to make some new user interfaces. But actually, this UX hacking, in a way, this uh user interface, I think it's perhaps too little discussed how revolutionary it can be, 150 00:44:01.000 --> 00:44:20.000 that the world is full of really bad user interfaces, both physical interfaces and things made by coders, so in a way, bringing AI into that, alongside your screen, keyboard, and mouse, as an equal component and a source of intelligence to it, well, at least for me, it increases my productivity quite a lot. 151 00:44:20.000 --> 00:44:48.000 Right, well, there is also talk about whether speech will eventually become the user interface, so you no longer have a mouse and keyboard, and or from Elon Musk's Neuralink, it will then come directly with a slightly higher bit rate. If there is a queue, I can certainly sign up for it, so there would be this kind of brain-to-computer directly. Yeah, but often software, the so-called frontend, is like built a bit, well, around human senses and the mouse and, well, 152 00:44:48.000 --> 00:45:07.000 keyboard and maybe even the microphone now, but AI doesn't need that, so that surely brings a like, a new, new opportunity, where we don't even design interfaces for humans to use, but rather for other AI or other software systems to use. Mm-hmm, exactly. 153 00:45:07.000 --> 00:45:15.000 Right. Exactly. So in a way, we're bringing in a bit of that machine-language communication there. 154 00:45:15.000 --> 00:45:23.000 Yes, exactly. And humans are then served or given, like, confirmation views or visualizations of a certain situation. 155 00:45:23.000 --> 00:45:28.000 Right. So, this Model Context Protocol-type interface that... 156 00:45:28.000 --> 00:45:34.000 Yeah. What about that MCP, as even something like n8n is almost rushing into it... 157 00:45:34.000 --> 00:45:37.000 I was looking into it, but then I didn't bother. So, what is it trying to... 158 00:45:37.000 --> 00:45:41.000 achieve? Well, maybe first what it is—we just... 159 00:45:41.000 --> 00:45:45.000 went through the API, meaning the interface, which is how you can programmatically... 160 00:45:45.000 --> 00:45:51.000 query information from an application, so in a way, built on top of those is... 161 00:45:51.000 --> 00:45:56.000 a kind of language model or AI interface, where... 162 00:45:56.000 --> 00:46:00.000 those APIs have been connected to it. If it's a CRM, then you could... 163 00:46:00.000 --> 00:46:05.000 ask for everything, like what our leads and opportunities and others are, and and... 164 00:46:05.000 --> 00:46:35.000 of course, the company information and all the data, in a way, that is available from the API is packaged together, and it's been made into an interface for a language model to use. It's an interface too, but it's, like, uh, simplified in the sense that the language model can then, with a very simple call—most of which are still quite human-understandable calls—request information and then use 165 00:46:35.000 --> 00:47:04.000 it sort of, or specifically, those agentic, uh, AI functions use them. So the language model notices that it now needs information from the CRM, so it uses its agentic function to ask for that data from the MCP server—the CRM in this case. Then it continues the execution of that task, and with n8n, for example, you can, like, visually program, or define this logic, and then you can 166 00:47:04.000 --> 00:47:32.000 connect several of these MCPs or memories and other other building blocks. So, in a way, this kind of AI application that has agents, like an agent team, can be built, in a sense, with this n8n. Maybe about that division of labor, because programming back in my youth was linear, meaning you entered commands and then they were executed. Maybe there were some subroutine calls or loops or this kind of minor 167 00:47:32.000 --> 00:47:49.000 recursion, but mainly it just proceeded in a sequence. Then here now this kind of parallel computing and expertise has been built, and Of course, the dream was that the quantum computer would be like a most ultimate kind of serial operator, in a way. 168 00:47:49.000 --> 00:47:58.000 So doesn't that agentic thinking guide us toward doing things not in a sequence, but rather in parallel? 169 00:47:58.000 --> 00:48:09.000 Yeah, exactly, so it can be done just to make the execution times shorter. It's like a larger team being able to do different things at the same time. 170 00:48:09.000 --> 00:48:35.000 Just as more gets done in a human team during the same week, we can think of getting more done during the same minute, and there's also the benefit that a single agent only needs to focus on one thing, so it's really good for reviewing contracts, for example. It has using a vector database as an aid and possibly some other sources, and then a neighboring agent 171 00:48:35.000 --> 00:48:55.000 might write Python code for us, which then produces that report. So, yes, if we return to that MCP again, it's always good to consider what those words mean: it has Model and Context and then Protocol, so you want to use MCP to provide context for the model. 172 00:48:55.000 --> 00:49:23.000 for example, with CRM data, so the model can function much better in your context, and then the protocol, which is shared, so that it works this way for everyone, precisely creating that context for the model is what it's about the most valuable thing one can do, and there the company's own data repository is perhaps the unique thing that the MCP has available. Yeah. But that also comes to many firms 173 00:49:23.000 --> 00:49:48.000 as quite a surprise; I've even been watching the actions of Atlassian, the provider of Jira and Confluence, as they built their own internal AI tool, investing probably tens if not a hundred million in it, called Robo. And now, in a way, you can basically do the same thing when you add an Atlassian MCP server alongside your own Jira and Confluence and 174 00:49:48.000 --> 00:50:17.000 query ChatGPT to connect them. In principle, you'd get it, if only Atlassian allowed all the data to flow without any reservations from the MCP through the interface, so you could perform the exact same functions. So here we come to the question of whether the data in the system is owned by the customer, or whether the operator who hosts your data there [does]. What is the context there? 175 00:50:17.000 --> 00:50:29.000 Right, so in a way, how easily will companies go along with allowing all sorts of data processing through the MCP interface? 176 00:50:29.000 --> 00:50:45.000 In principle, they should, because the company using the service owns that data, as I understand it. And this just limits the route through which it can be used. Whether it can only be used through their own UI and AI or? Personally, I would keep all my own things in my own basement. 177 00:50:45.000 --> 00:51:13.000 my data, and you can use like MCP there, but very deliberately would I put it into any hyperscaler cloud anymore. And there maybe we'll return to exactly that, that Miettinen thought about that SaaS, uh custom-coded box that would pull data from the great and mighty S&P Capital IQ data, so they definitely don't want me to start reselling their data, for instance, so there are some of these 178 00:51:13.000 --> 00:51:16.000 kinds of things too. That's a really good example of exactly what I was getting at there. 179 00:51:16.000 --> 00:51:43.000 Yeah. Because the original ideas, some good questions, and that vision, there's a lot of huge value in that, which AI generically doesn't Oh. And and now Nvidia entered the consumer market, so they released a box that has, like, a huge amount of processing power. That box probably costs about 5 grand. But the, but the 180 00:51:43.000 --> 00:51:47.000 idea is that then all the stuff is there physically with you. 181 00:51:47.000 --> 00:52:17.000 Yeah. And one of the capacity limitations of that Microsoft Windows 11 update seemed to be that Microsoft is probably tired of running those Copilot things in the cloud, so some of them will be offloaded to the hardware. So I think this is kind of interesting that while during my long absence everything went to the cloud, now we're kind of returning to hardware. This, this is like in technology there's never an extinction, they always just go a bit like this 182 00:52:17.000 --> 00:52:45.000 yeah, and I saw a story like that too, I watched a story on YouTube, that these crypto coin or miners have moved so that with the same rigs with which they used to mine Bitcoin with graphics cards, they are now offering to the network computing services for AI models. So it might be more profitable, but this leads to the fact that computing capacity is needed at all kinds of very different levels, but in massive amounts. 183 00:52:45.000 --> 00:53:14.000 Yeah. And I mean, all the computing that can be built in the world, it will be used. So this is like quite a clear thing, that like data centers or whether it's rigs or those personal computing devices, but wherever computing is possible, it is computed and it has a price. Yeah. About that, I'd like to get you to envision that since at the chip level, NVIDIA is already like designing hardware for people's home computers over there 184 00:53:14.000 --> 00:53:41.000 in an AI sense. Then we sort of mentioned Atlassian, a listed company, so for some reason, this private company's Jiras and Trellos and suites are used in this management. This Microsoft's GitHub and sort of the Copilot programming environment is one way of doing it, so how is that going to play out, will people start vibe coding these components as well, or are there just a few that are so 185 00:53:41.000 --> 00:53:46.000 giant, like, mammoths, so they'll be running that infrastructure? 186 00:53:46.000 --> 00:54:13.000 Well, good question. I've personally used something called AnythingLLM a lot, and I ended up with it because it's basically this component that you can run on your own machine. It doesn't really consume much power, even though it performs AI, uh, functions, but you can then authorize or, like, connect it to use the OpenAI API or clouds or any cloud- 187 00:54:13.000 --> 00:54:40.000 based AI. Or you can run local models and connect them back behind this AnythingLLM. And AnythingLLM provides a single interface where you can then first of all put documents in. It makes them into this kind of RAG, meaning that vector database, and talk with your own documents either so that you have a local model or a powerful model in the cloud with tokens. And all of these can be controlled there, like with mouse 188 00:54:40.000 --> 00:55:06.000 clicks locally, and it provides a normal chat interface. It also has a REST API as well, so you can program against it with Python. But this has been, in my opinion, super handy for when you want to compare how different models work for the same task or if I could manage with a local model, if I only have ten medium-sized Word documents in the Finnish language. It starts working quite well with something like Gemma 3 nowadays. 189 00:55:06.000 --> 00:55:31.000 But that kind of, as I see it, there is also clearly one such trend, that you can keep the data locally you can run it completely on a normal PC, you don't need any- -thing. Yeah, and well, since everything went to the cloud, is it now then like, people start looking into, 'well, let's buy one of those five thousand [euro] units and make our own server'—are we going back to the roots here? 190 00:55:31.000 --> 00:55:58.000 Yeah, yeah, yeah. But I mean, this is a bit of an 'I want to believe' type of thing, because right now there seems to be this kind of land grab going on, where OpenAI is building some stargates here and there and they're talking about billions—meaning American trillions—and like but but like in a way, then on the other hand, if we look at machine learning long history, so it's pretty much open science, open protocols, 191 00:55:58.000 --> 00:56:25.000 open everything, like the hacking-phase-type communities we have in existence here, and these Web3 technologies and everything else, whose core idea is specifically decentralization of power, as we want a situation that is a fair, good situation, where there is a fair playing field for competition, so I at least don't want, and wouldn't want, and will oppose it with all my heart and soul, so that it wouldn't somehow be a situation of a few 192 00:56:25.000 --> 00:56:34.000 gatekeepers, like the situation we currently have, for example, in search, or rather search functions, or social media or something else; in my opinion, now... 193 00:56:34.000 --> 00:57:00.000 during the era of AI, we should just get this kind of distributed, fair, fair market, and where the key enablers are are open, managed in a Linux-like way somehow, and then on top of that, you build whatever visionary thing, but but but so that the, like, from the, like, the... 194 00:57:00.000 --> 00:57:26.000 infrastructure, the AI infrastructure issue, it shouldn't become something that only a few control. Yeah. And then locally you can do—I started, actually, after Lasse's point, I felt like installing a home server running AnythingLLM and doing some agentic stuff at night, that I code it to do, and it probably costs nothing since the API calls don't even go online. Yeah yeah yeah yeah. But I mean, this is what we 195 00:57:26.000 --> 00:57:53.000 want, I mean the innovation, everything you can do there, like at home with that own server, because we want a free market already to go to the moon, but like all sorts of things, well, yeah and and and then when you have that like trillion-dollar idea, you might even be able to pay a bit for it to some hyperscaler, like here's 100 million for you, I get to run this like trillion-dollar business here, so it's fair, but but like in a way that there wouldn't be like these kinds of control points here, and we 196 00:57:53.000 --> 00:58:01.000 have quite good chances here. And it does look good, if they are like vision-oriented, or let's say, it looks better than it could be. 197 00:58:01.000 --> 00:58:26.000 So there are these players. OpenAI is the biggest now and a bit ahead, but there are indeed Grok, Perplexity and Claude, Claude, so many others. Yeah, and then the Chinese are coming with DeepSeek and others, so in a way the model—now there is a lot of talk about these models and such, when the applications are somehow in my opinion, yeah, but those competitors Google is also there, of course, so there is genuine competition and they 198 00:58:26.000 --> 00:58:55.000 have a real interest in creating slightly different solutions and everyone is investing really heavily, and a large part of it indeed pays off specifically because of the market-driven competition there Yeah, and then there are also these Hugging Face-type communities and all the others who can produce, like, open models and such, so that what is currently the role of top-tier models today will, in a year's time, be miniaturized and so on, so that such performance 199 00:58:55.000 --> 00:59:21.000 from that model actually runs on a phone; in my opinion, this should be nurtured and promoted by all means, so that the world remains open. Yeah. And I think this is, like, a great time, and I would almost say to Finnish companies that if you have an IT department, well, have them demo, like, vibe coding trends to the rest of the team, or invite a consultant 200 00:59:21.000 --> 00:59:50.000 doing a bit of a pipe coding day, or then like organically someone just gets excited, like let's grab this glove and make a couple of demos and you try them out, it's a great time, and I think here you learn, learn a lot even in a short time, and that few tens per month fee for these services is, well, quite little compared to say the salary of some subcontractor or employee, so that here even for a small amount of money you get like a lot of bandwidth to do things and maybe this 201 00:59:50.000 --> 01:00:17.000 returning to this, how with Lasse we founded this Pääasioittaja Pipe Coding Society, so like these kinds of communities, that like 'uncle LANs' are popular again, like we also sat down there at Trans in the room, you learn so much when you compare, like laptops, like 'show me what you're doing and how you'd do that' and so on, and specifically, I mean learning and sharing are really, really important here, so 202 01:00:17.000 --> 01:00:45.000 nothing goes right on the first try; instead, you have to do something a hundred times and then you might find it, but learning is the most central thing in my opinion, and then the sharing, like 'hey' yeah, and then that kind of quick testing, so in that sense, the AI channel is worth checking out and subscribing to, because you can find fairly quick overviews of such tools and practical to solutions, what you're discussing and figuring out there. Thank you. I've 203 01:00:45.000 --> 01:00:55.000 had Lasse Mikkonen and Harri Juntunen as guests, and since you've watched or listened this far, please subscribe to the channel and also watch the AI channel's counterpart episode. 204 01:00:55.000 --> 01:01:22.000 I suggest we go to the insider side and move this now to that secret chamber there. So, I started mulling over that, um, sort of this Web 3.0 and this kind of, in a certain way, specifically own data, maybe a bit of the crypto side too, what could be done there, that if we really want to keep that data and run things on our own machines 205 01:01:22.000 --> 01:01:32.000 instead of letting everything run in those closed systems, then what do we could we do? So I suggest we talk more about that world a bit more over on the insider side. Thank you. 206 01:01:32.000 --> 01:01:36.000 Okay. All right. Good. Thank you.