บทถอดความฉบับเต็มของ Agentic Software Engineering - Dr. Ralf S. Engelschall

primary focus, actually, well, I would say 90% of my CTO job is located in the AI field currently. 00:02:11.280 --> 00:02:20.960 And uh I guess with like customers having to uh wanting to, needing to, and uh and wanting to have AI in their things and not knowing how to. 00:02:21.039 --> 00:02:32.000 So I think it's it's easy for a startup to embrace the coolest, newest stuff, but for a company of like 600 people with like a software package that's like 23 years old, not so much. 00:02:32.240 --> 00:02:39.039 But do you still get a lot of requests from customers to actually say, like, we need to do this because there's a FOMO kind of feeling? 00:02:39.280 --> 00:02:40.000 Uh definitely. 00:02:40.240 --> 00:02:57.039 Uh all uh enterprise customers have uh AI transformation programs and other things running where they on one side try to uh transform their own uh delivery and their own employees to use AI and uh gain uh product uh productivity here. 00:02:57.120 --> 00:03:02.800 And the other one is uh to do in their business, put more AI for their end customers in into their products. 00:03:02.879 --> 00:03:24.560 And this is where uh companies like MSG help out through consulting, but also to through software, because uh, well, today you cannot ship any piece of software without at least an AI label, and uh, this AI label underneath uh has to show AI-based features and functionalities, more or less to show uh new features and new uh possibilities to AI. 00:03:24.719 --> 00:03:32.240 So all enterprises evaluate AI things and uh actively and aggressively put those AI things into their own offerings. 00:03:32.400 --> 00:03:41.360 And uh companies like MSG, which are in the second row and delivering those solutions uh this way has a lot of stress to deliver those AI solutions. 00:03:41.439 --> 00:03:55.360 But even on our internal delivery, we also uh have an AI transformation program where we have to enable more and more our developers to gain even more productive through the use of AI for ourselves in the software engineering itself. 00:03:55.599 --> 00:04:03.520 The other problem that I found is that we're we seem to be in a world where there's like the the uh uh the first level people that get it, and then Europe gets it. 00:04:03.759 --> 00:04:18.879 When I worked for Microsoft, and especially with when I talked to Microsoft Germany, because I work for Microsoft US, they were all like, we're still trying to get German customers to embrace the cloud rather than like have the one server farm that they bought in 1983 and have back in the basement of their of their housing. 00:04:19.120 --> 00:04:26.879 Do you think the the it's now the speed of what we have right now forces people to actually embrace new technology more? 00:04:27.120 --> 00:04:32.879 And if they do, oftentimes it basically is like here's a cool new model that you can only use if you're in America. 00:04:33.279 --> 00:04:34.160 How do we deal with that? 00:04:34.480 --> 00:04:37.920 Well, at the end it it boils down to the sovereignty discussion. 00:04:38.000 --> 00:04:48.879 Uh, but yes, you you're you're forced to use the things which are available because you cannot afford to just ignore the how fast the technology, especially in the AI field, uh, drives. 00:04:49.120 --> 00:04:55.439 So um, even for us, uh, it means to use uh US uh models, pharmacic open AI, and other things. 00:04:55.600 --> 00:05:02.959 Yes, there are uh alternatives from the Asia side, but Europe has too less to offer here currently. 00:05:03.120 --> 00:05:08.879 So if you are in the business and have to use AI, you are more or less forced to use even the uh US offerings. 00:05:09.120 --> 00:05:19.439 You have more and more now uh through the uh sovereignty um um discussions, uh more and more possibilities in Europe, but it's still too less. 00:05:19.600 --> 00:05:22.399 You cannot go without more or less in practice. 00:05:22.560 --> 00:05:29.600 Uh just take the example of uh the offering of uh clothed code and the clothes uh uh LM from Anthropic. 00:05:29.839 --> 00:05:31.360 It's uh really a frontier model. 00:05:31.439 --> 00:05:37.360 And if you want to do first-class development, really be on the frontier yourself, you have to use it, yes? 00:05:37.439 --> 00:05:45.680 Uh and even uh things like uh a lot of people talking, for instance, about the okay, let's do it here locally in Europe, or even locally in your own uh room. 00:05:45.759 --> 00:05:54.560 Uh it's uh nearly impossible currently because if you look on uh really locally I on own hardware, you don't get the necessary token per second. 00:05:54.800 --> 00:05:59.680 You don't get uh near this quality the frontier models from the US currently provides. 00:05:59.920 --> 00:06:10.000 So, in practice, uh it's uh it's a um a schizophrenic situation because we are we are talking about sovereignty on one side and would like to have more things locally in Europe. 00:06:10.079 --> 00:06:20.480 On the other side, we are forced to use the US and Asia offerings because of the quality and and the speed, uh, and especially in the cloud, and you cannot really do it uh on premise. 00:06:20.639 --> 00:06:35.519 So it's uh it's a hard uh situation currently where we would like to do more locally and on our own, but it's uh really not possible on the market, at least if you want to leverage the high quality and speed which is currently offered by others in the cloud. 00:06:35.839 --> 00:06:50.879 I mean, we have things like Aleph Alpha offering things in Europe, stack it is trying to do the hosting stuff, but yeah, it feels it feels more convenient to actually use the American and the Chinese offerings, but at the same time, convenience sometimes means lock-in as well. 00:06:50.959 --> 00:06:52.800 That's the that's the other problem with that. 00:06:52.959 --> 00:06:57.439 Uh but and I mean as you said, like on-prem is almost impossible because you can't buy hardware anymore. 00:06:57.519 -->

00:06:59.279 I mean, the RAM is basically out there. 00:06:59.360 --> 00:07:05.680 Like, I've I just got Nvidia just gave me a GTX Spark, and I I still have to set it up because I'm super excited about it. 00:07:05.839 --> 00:07:10.160 But I was like, this took this took like four months to deliver, and that was from NVIDIA directly. 00:07:10.240 --> 00:07:12.879 So it feels like the market is pretty empty on that. 00:07:13.120 --> 00:07:23.759 This NVIDIA technology definitely has a large potential to give us new possibilities, but yeah, only once it is really available and for an affordable price at the end. 00:07:23.920 --> 00:07:38.160 Uh, and um currently I see as a CTO, I see no possibility to really, at least for the next uh six or twelve months, to go without the uh cloud services uh of the frontier um um offerings in in the US. 00:07:38.399 --> 00:07:43.279 Mind you, we still got Microsoft and Google, which all have German server farms as well, if you wanted to use them. 00:07:43.360 --> 00:07:50.800 But I I I feel like both of them have kind of lost the pace compared to the others, to the smaller, faster moving ones. 00:07:51.040 --> 00:07:57.439 So it feels it feels kind of stuffy to actually start using Google Cloud or or Microsoft Cloud instead of OpenAI or Cloud. 00:07:57.680 --> 00:08:02.720 Yeah, OpenAI and Cloud idea from Anthropic is this, these are really the frontier ones. 00:08:02.800 --> 00:08:08.879 And if you as I said, if you want to be on the frontier from quality and speed, currently you have to use them. 00:08:09.040 --> 00:08:14.160 I dislike it myself as a European guy, but um currently I see no real alternative. 00:08:14.240 --> 00:08:23.920 But I hope it comes because to be fair, uh, the way it currently goes in the market, that there is no end, even when it comes to prices, it's fully sponsored. 00:08:24.000 --> 00:08:26.720 Even the high prices currently are still fully sponsored. 00:08:26.879 --> 00:08:31.759 They do not uh get the never necessary revenue what uh they have on the other side on costs. 00:08:31.920 --> 00:08:38.639 So even those uh companies like OpenAI and Entropic have large gaps between their revenue and uh their costs. 00:08:38.720 --> 00:08:40.240 So uh this is not the end. 00:08:40.320 --> 00:08:46.399 And if it becomes even more expensive and expensive over time, there will be uh an explosion, I think. 00:08:46.639 --> 00:08:48.879 This cannot uh work uh any longer. 00:08:48.960 --> 00:08:51.200 So I definitely hope that something changes. 00:08:51.440 --> 00:08:55.679 But either technologies, for instance, like Nvidia Spark, could help here. 00:08:55.759 --> 00:09:12.240 Uh, the other would be perhaps that uh new transformer architectures and other things on the conceptual side perhaps are invented, and this way um, as good um from the quality side, models can be established with a lot less uh resource uh uh consumption. 00:09:12.320 --> 00:09:16.480 So this would be another possibility, but currently it is for the for the next six or twelve months. 00:09:16.559 --> 00:09:19.840 I personally don't see it that uh this world changes here. 00:09:20.159 --> 00:09:28.480 Well, Jeff showed a good way of actually how you can actually pre-u uh save on tokens by having a different model that's pre-trained, that's pre-weighted. 00:09:28.639 --> 00:09:30.000 These kinds of things are possible. 00:09:30.080 --> 00:09:41.679 Also, you can put your own racks in front and basically make sure that you that in the company, when you have 11,000 people, that some knowledge that has been done has is cached locally with a with a vector database or something like that. 00:09:41.840 --> 00:09:44.399 That makes uh that makes the for lesser consumption. 00:09:44.559 --> 00:09:50.480 I'm very happy that that token maxing thing is off the table again, because that was the most ridiculous thing I heard in a while. 00:09:50.639 --> 00:09:59.679 Like it was like it's like when we got paid for lines of code, and you're like, great, I'm gonna write the most ineffective code that is really, really long and make myself a car for that. 00:10:03.200 --> 00:10:10.559 Seeing that you are a patchy person and open source person, it very much uh cares of my uh of my uh uh statue. 00:10:10.879 --> 00:10:16.720 We throw the world or the word open around a lot in the AI space, but then it never is open. 00:10:16.799 --> 00:10:19.360 It's only open weights, it's not open models. 00:10:19.519 --> 00:10:21.919 Do you think there is uh there is a chance for that? 00:10:22.080 --> 00:10:31.440 Because right now it seems like uh there's like all this up uh uh uh let's make it not up not public, so we can say it's actually an amazingly powerful thing. 00:10:31.600 --> 00:10:35.600 Like with open source, you cannot hide anything, but with closed source you can. 00:10:35.679 --> 00:10:39.600 And I think there's uh when it comes to getting investments, you need to hide things, right? 00:10:39.840 --> 00:10:41.840 Yeah, this is what currently happens. 00:10:42.000 --> 00:10:44.879 Uh as you mentioned, open weight is not open source here. 00:10:45.039 --> 00:10:49.519 It means the training data and other things is not available, it's just the resulting weights. 00:10:49.679 --> 00:10:59.919 Uh, a few of the uh LM vendors start uh more and more to really also show how is the training done, what is the training material, but it's it's very thin currently. 00:11:00.240 --> 00:11:10.000 Most of the stuff is is closed and uh, well, at maximum it's open weight at the end, but not really when it comes to uh do it yourself with the training data and all this stuff at once. 00:11:10.080 --> 00:11:11.840 So um it can become better. 00:11:12.000 --> 00:11:24.399 I would like as an open source guy that we have really open source models at the end, but uh I can imagine that um and I understand they protect this way their own investments by keeping or hiding things. 00:11:24.559 --> 00:11:33.759 Um

and the point is uh if you compare it with the open source software field, um, here it's easier because your investments are not such high. 00:11:33.919 --> 00:11:48.799 But once you have to spend millions to train things, I can also, as an open source guy, I understand that I have to do something uh more to protect it and not lose those things because this is really a lot more investment than we had on the bare software side. 00:11:48.960 --> 00:11:55.120 So, yes, even as an open source guy, I can understand it, but I dislike it that it's uh not uh really open source, what we see here. 00:11:55.440 --> 00:12:03.039 Now, the other the other tension that we feel right now is that everybody says, like, oh, agentic coding, don't do any code by hand anymore, that's only for old people. 00:12:03.200 --> 00:12:05.919 Uh, I still like writing code, it makes me happy. 00:12:06.000 --> 00:12:19.519 Yeah, and uh, but it feels like that there is a pace issue that like uh generated code in an agentic environment is so much faster to create and also much faster to discard as well as handwritten code. 00:12:19.679 --> 00:12:21.600 So, how do these two worlds come together? 00:12:21.759 --> 00:12:29.440 Because I guess in your space, uh it this is basically a tricky bit because the the generated code will never be uh will never be merged. 00:12:29.759 --> 00:12:36.399 Yeah, the point is uh my heart bleeds actually after four decades of intensive coding myself in this field. 00:12:36.639 --> 00:12:39.200 So, yes, I really it bleeds all the time. 00:12:39.360 --> 00:12:45.279 On the other hand, I have the job of a CTO in the industry, and uh uh here I have to do a different point of view. 00:12:45.360 --> 00:12:57.120 Um, I can just uh tell you with uh projects like agentic software engineering, this AZ.tools, what uh in the open source field, I try to make my peace with the situation by merging those things. 00:12:57.200 --> 00:13:05.200 So leverage from a Gentic eye on one side, even if yes, my heart bleeds, because bare coding, yes, it's dead, actually. 00:13:05.279 --> 00:13:22.720 Yeah, people like you and me, we still like it, and our heart bleeds, but on a bare fact uh uh level, we have to uh say uh manual coding is dead because the AI can do it faster, more precise, and actually cheaper at the end. 00:13:22.960 --> 00:13:37.679 So this we we have to accept uh it's a new uh era uh where we entered uh about in in the quarter one of uh 2025, um, and with also tools like Clot Code and the uh uh versions of Cloud uh LM at these times. 00:13:37.759 --> 00:13:44.720 So this is an agentic AI era where new possibilities exist, and it doesn't matter whether our uh heart bleeds, we have to cope with this. 00:13:44.799 --> 00:13:51.840 And uh well, I might try to make peace by fusing those things together and saying, okay, on one side we have agenti AI with new possibilities. 00:13:52.000 --> 00:13:57.039 On the other hand, uh, I like code, I like software, that's my

my my my life. 00:13:57.200 --> 00:14:05.200 And uh, how can we still produce quality results at the end, even if we are using uh uh AI to produce this stuff? 00:14:05.440 --> 00:14:16.159 And um, this is not easy because at the end, uh you can tell uh an uh in clot code, just enter a few prompts and tons of code immediately can get generated. 00:14:16.240 --> 00:14:20.720 And this is what you see all the time on Reddit when someone says, I have built something. 00:14:20.879 --> 00:14:23.759 Uh, I sometimes even get angry when I when I see this. 00:14:23.919 --> 00:14:25.600 I say you have nothing built. 00:14:25.759 --> 00:14:28.879 It was the AI, but it's okay, okay, on one side. 00:14:29.039 --> 00:14:46.320 Um, the point just is what those people most of the time create is convenience software, something a little utility or desktop app, uh, which where there is no risk or anything, and where it actually doesn't matter whether the code is really good or not, it's just essential that it works. 00:14:46.480 --> 00:14:48.639 And even if it breaks, you just restart it. 00:14:48.720 --> 00:14:49.840 There is no risk. 00:14:50.000 --> 00:15:07.759 Um, if you compare this with the regulated market where we are in the industry are working for banking, insurance is uh, etc., where software has to last for uh 15, 20, 25 years on one side, have to really be compliant to a lot of law uh issues, etc. 00:15:08.080 --> 00:15:12.799 Um, here you cannot just white coat something, so this doesn't work. 00:15:12.879 --> 00:15:16.399 Yeah, so we cannot directly transfer these the same idea, okay. 00:15:16.480 --> 00:15:22.960 Let's sit on a software engineer in front of uh code and let's enter a few prompts and get tons of code generated. 00:15:23.120 --> 00:15:31.279 Here we really have to find a balance between the new possibilities on one side with AI and the constraints on the other side. 00:15:31.440 --> 00:15:55.840 And this is more or less the difference for me between software programming or bear development and software engineering, because engineering, all the time, even in the past without AI, was about the processes and structuring and all those things around to really ensure that you get a transparent process where traceability exists, where you can at least ensure that certain qualities exist, etc. 00:15:56.080 --> 00:16:01.120 So this is more or less a risk mitigation process, if you want, uh software engineering. 00:16:01.200 --> 00:16:28.080 And yeah, in these days now, we have to bring those pieces together: the software engineering risk uh mitigation process on one side, and the possibilities of urgent AI coding tools like Look Code, where a lot of things can be generated in just a few minutes, uh, because they actually don't fit immediately together because there's uh an obvious clash, because it's different types of software you create, actually, with those different pieces. 00:16:28.720 --> 00:16:37.519 And you don't want to be a gatekeeper,

you want to actually allow people to be creative with generated code and generated apps. 00:16:37.600 --> 00:16:43.440 And I'm I mean, I'm seeing people in our sales team using using Claw to build little apps for themselves. 00:16:43.600 --> 00:16:56.799 That actually asking us in the engineering team would have done more, would have been like, yeah, we put it on the backlog and then we get to it when we get to it, rather than just empowering them to do the stuff that, as you said, is throwaway because it's only used internally. 00:16:57.039 --> 00:17:03.840 It's still an opportunity to leak your tokens and to leak your keys, uh, which happens a lot as well. 00:17:04.079 --> 00:17:09.440 But you don't want to be that that Starboard person that says, like, oh, only engineers can touch the code. 00:17:09.599 --> 00:17:10.960 It's actually fine to do that. 00:17:11.200 --> 00:17:22.400 At the same time, it reminds me of when what you see is what you get editors came out, and I was a front-end developer, and people like, oh, my cousin has Dreamweaver, he can make it for 200 euros rather than the 1200 that you want to have. 00:17:22.559 --> 00:17:24.240 And that's still the same thing. 00:17:24.319 --> 00:17:27.200 We or we sort of basic people in the past, like you know. 00:17:27.519 --> 00:17:31.440 Well, actually, this pattern goes over all the decals I was involved. 00:17:31.680 --> 00:17:37.920 Uh, in I can remember in the year 2000, we are talked about case, computer-aided software engineering. 00:17:38.079 --> 00:17:45.119 Later in the 2010, we talked about uh model-driven development, then we talked about low-code, no code. 00:17:45.279 --> 00:17:48.079 Currently, we are doing wipe coding and in a Gent D GI. 00:17:48.160 --> 00:17:50.480 It's always, I would say, the same pattern. 00:17:50.720 --> 00:17:53.599 People like to get rid of the nasty programming. 00:17:53.920 --> 00:17:59.039 Even I like programming, but a lot of people say, ah, it's nasty, it's a lot of effort I have to place. 00:17:59.119 --> 00:18:04.559 Uh, I want to draw just software or I sketch software, and it should come out and it immediately work. 00:18:04.880 --> 00:18:15.039 So there is a large track record in our society that uh in the IT society that we want to get rid of programming a little bit all the time with different approaches. 00:18:15.200 --> 00:18:22.799 And I would say the Atlantic AI area currently is the latest one, um, certainly the most powerful one, also, to get rid of this stuff. 00:18:22.880 --> 00:18:30.559 Uh, but it has its limits because it uh doesn't work for for, as we said, for convenience software, a little bit desktop app, utility. 00:18:30.720 --> 00:18:34.960 It it doesn't matter what the code is about, it doesn't matter what the it just has to work. 00:18:35.039 --> 00:18:38.559 And if it doesn't work, well, it's just this particular use case is broken. 00:18:38.720 --> 00:18:40.880 Nothing happens, there's no real risk. 00:18:41.119 --> 00:18:43.440 So uh I fully understand those approaches. 00:18:43.519 --> 00:18:51.839 Uh, even for low code, it's uh there are brilliant low-code and no-code platforms on the world, but also those have limits. 00:18:52.000 --> 00:19:04.640 And the limit more or less is obvious once you want to customize the solution uh towards more your shaping in the UI and particular functionalities, then though you have to work against those platforms all the time. 00:19:04.799 --> 00:19:07.759 And similar here now in the agentic AI era. 00:19:07.839 --> 00:19:19.039 Um, you can say, okay, create me something, but if this should have a certain quality, if it should be really this way and that way, you have to prompt a lot of things and it becomes more and more complicated. 00:19:19.200 --> 00:19:21.920 So it's, I would say, the same pattern all the time. 00:19:22.079 --> 00:19:29.359 People in the IT like to get rid of their nasty efforts to create software with different patterns over the time. 00:19:29.519 --> 00:19:31.119 Currently it's a Genti AI. 00:19:31.279 --> 00:19:33.759 And to a certain extent, it works great. 00:19:33.839 --> 00:19:37.039 Uh, but one has to understand where the limits are. 00:19:37.200 --> 00:19:48.799 And once you hit those limits, for us, it's for instance, we have well, a business information system with a few hundred lines of code where uh they are uh used in uh regulated markets like insurance and banking. 00:19:48.960 --> 00:19:52.640 There are different constraints than you have on a desktop application. 00:19:52.799 --> 00:20:05.839 And once you uh hit those limits, you have to find a way to fuse classical software engineering once again with a Gent DGI, because else you will not be able to cope with this situation of this limiting here. 00:20:07.279 --> 00:20:11.759 Yeah, I think the the uh it's an interesting world where we actually try to do that. 00:20:11.920 --> 00:20:20.319 I mean, I'm I'm actually happier to be gotten rid of the enterprise world where you had like UML, SOAP, XML, XSLT work back then. 00:20:20.400 --> 00:20:24.880 Like that was the thing where I'm like, okay, the web was so much faster, and now we got these things. 00:20:25.039 --> 00:20:30.720 One worry I have, I'm I mean, we just have a coding competition coming up, and I just said, like, do something creative. 00:20:30.880 --> 00:20:36.720 So I tried all the different uh different platforms out to give it an HTML document and say, like, do something creative with it. 00:20:36.960 --> 00:20:41.440 All of them gave me an HTML document back that I can put on GitHub and show people right now. 00:20:41.680 --> 00:20:46.720 In one provider, basically, like, oh, you have to use our IDE to actually access the thing. 00:20:46.960 --> 00:20:48.160 You do it for the first time. 00:20:48.240 --> 00:20:52.559 We need to verify you, you have to pay$20 to get access and create a key. 00:20:52.720 --> 00:21:02.720 And then it generated a uh a dynamic HTML

page, not the HTML, but dynamic HTML page, that basically had the key inside. 00:21:02.799 --> 00:21:06.079 And I'm like, okay, I can't share this on a public repo now. 00:21:06.319 --> 00:21:10.559 So a lot of this convenience is basically just a lock-in, isn't it? 00:21:10.880 --> 00:21:11.599 Yeah, sure. 00:21:11.839 --> 00:21:14.480 This is also the problem of those platforms. 00:21:14.559 --> 00:21:19.759 Uh, all platforms have more or less, uh, from the business side, this built-in lock-in effect. 00:21:20.000 --> 00:21:22.720 This is one of the reasons why vendors actually do this. 00:21:22.799 --> 00:21:24.799 Um, you have to accept it or not. 00:21:24.880 --> 00:21:29.359 Uh sometimes it's nasty and you don't want it because this login can be nasty. 00:21:29.519 --> 00:21:31.200 You cannot change the vendor easily. 00:21:31.279 --> 00:21:36.880 Uh, on the other hand, it is very convenient because you can uh immediately uh get your stuff running. 00:21:37.039 --> 00:21:40.160 It's uh I would say it's it's very similar to cloud offerings. 00:21:40.240 --> 00:21:40.960 Uh look here. 00:21:41.039 --> 00:21:47.200 Uh, if you go to AWS, they have uh an immense large amount of services they can offer you. 00:21:47.359 --> 00:21:54.960 But once you pick them, you get locked because you cannot, for instance, easily switch from AWS, Lambda to Azure Functions, for instance. 00:21:55.039 --> 00:21:59.759 Uhoretically, you could, but uh at the end, it's a lot of effort to transport the things. 00:22:00.240 --> 00:22:08.559 So, yes, uh it's uh clear that from a business perspective, AWS tries hard to establish this login on one side. 00:22:08.720 --> 00:22:14.880 So you have to accept uh is it uh in my case convenience I really want to have, then you have to accept the login. 00:22:15.279 --> 00:22:26.400 If you don't uh need this type of convenience and uh can put a more effort on it and want to avoid the login, yeah, you have to go another uh uh uh way and uh create more standalone solutions. 00:22:26.559 --> 00:22:33.039 So I would say it's always the same problem with platform with platforms, uh even with frameworks in programming. 00:22:33.279 --> 00:22:38.079 They give you limits, but within those limits, you have great convenience. 00:22:38.240 --> 00:22:42.160 But once you want to get out of those limits, you are out of the game. 00:22:43.519 --> 00:22:43.920 Cool. 00:22:44.079 --> 00:22:51.440 So let me talk, uh let's talk about your project right now that you're excited about, and that is ASE agentic software engineering. 00:22:51.839 --> 00:22:55.759 That sounds like it's actually two things that shouldn't fit together. 00:22:56.640 --> 00:22:57.039 Exactly. 00:22:57.359 --> 00:23:07.759 Agentic AI on one side with tools like Cload Code, and then came on the other side the software engineers, which are proud of their, as I said, uh traceable processes and all this other stuff. 00:23:07.920 --> 00:23:13.680 Um, here we first have to understand, and this is a misunderstanding for most people, software engineering is not software development. 00:23:13.759 --> 00:23:24.000 Okay, so software development on one side can be, if you want, broken down to ad hoc creating your user stories and other things and creating uh the software, so the the bare programming if you want. 00:23:24.079 --> 00:23:25.359 So for me, it's an onion model. 00:23:25.440 --> 00:23:28.559 You have programming, you have software development, you have software engineering. 00:23:28.640 --> 00:23:39.920 And software engineering is all those additional things for uh the processes together, traceable process, requirements engineering, software architecture, testing, project management, etc. 00:23:40.160 --> 00:23:40.400 Okay. 00:23:40.799 --> 00:23:46.400 And uh yes, it sounds uh those two things uh do not belong together, or how can they belong together? 00:23:46.559 --> 00:23:47.839 Well, for me, it was very simple. 00:23:48.000 --> 00:23:57.119 On one side, I want in the industry that uh even we are a software engineering company, have to do large-scale business information systems with hundred lines of code for enterprise customs. 00:23:57.200 --> 00:24:04.640 On one side, where software engineering is the core business, I want to leverage from a Gent DGI because uh I like Agen DGI. 00:24:04.720 --> 00:24:08.640 It's great what what possibilities exist, but yes, you immediately catch it. 00:24:08.720 --> 00:24:10.480 It doesn't directly belong together. 00:24:10.640 --> 00:24:18.559 So, what I've done is uh in this Gentic Software Engineering, or in short, AZ as I call it, in this AZ project, is to do a fusion here. 00:24:18.640 --> 00:24:25.680 Uh, technically, you can just imagine it's a plugin for a cloud code, uh GitHub Copilot, or uh OpenAI codex. 00:24:25.759 --> 00:24:32.000 So you have your Gentic AI tool, you plug in this ASI plugin, and you get additional functionalities. 00:24:32.160 --> 00:24:39.519 Those functionalities hook into this agent through so-called uh, yeah, here this is uh the setup of what you see on the screen. 00:24:39.599 --> 00:24:52.960 Um, you get additional skills, you get uh status lines, you get hooks, uh mods, uh whatever the particular underlying tool actually provides, and AZ adds additional functionalities. 00:24:53.119 --> 00:24:57.200 Those functionalities, what a software engineer in industrial practice needs all the time. 00:24:57.359 --> 00:25:25.519 So this starts from doing things like I can do my requirements engineering, um, in a fashion where not just entering a talk prompt, but I do either task-based, user story-based uh uh spectrum development or the other way of spectrum development, uh coherent binding document, but not in a way that I write a Word document or a confluence wiki or something like this as a document, but with the Gentic AI tool. 00:25:25.599 --> 00:25:34.480 So you can imagine, for instance, with AZ, if you do uh a specification, it means you

have on your uh JIT repo uh a few markdown files. 00:25:34.559 --> 00:25:53.680 But it's not just arbitrary markdown files, it's typed markdown files, which gets strictly linted in your build process and your gentic AI coding tool, like uh Cloud Code, through the help of AZ, gets known all the meter information, the meter model behind the specification, the schema, etc. 00:25:53.920 --> 00:26:01.359 So it understands your specification structurally wise, exactly the way it understands your TypeScript code, for instance. 00:26:01.519 --> 00:26:03.119 And this brings you onto a new level. 00:26:03.200 --> 00:26:06.880 So Cloed Code is great to do ad hoc things, it has a lot of knowledge. 00:26:07.039 --> 00:26:28.000 But when you want more strict software engineering through a requirements process here in this situation with a binding document, if you want to do task-based uh agentic development pipelines where engineers, though humans, and agents uh alternate in the in uh working on a particular user story, you need more. 00:26:28.079 --> 00:26:30.559 And this is what agentic software engineering provides. 00:26:30.640 --> 00:26:36.480 Um, agentic software engineering, for instance, has the possibility to create your user story in a task, move it to an engineer. 00:26:36.559 --> 00:26:42.400 The engineer takes this task, works on it through cloed code, and then it hands is it's picked up by an agent. 00:26:42.559 --> 00:26:48.799 This agent implements this code, and then the human once again comes, picks it up, uh, and does the review. 00:26:48.880 --> 00:26:55.279 And later another agent picks it up and merges, for instance, the feature from the feature branch to the master branch. 00:26:55.440 --> 00:27:04.640 So AZ is more or less putting the software engineering traditional processes onto agentic AI coding tools like Cloud Code. 00:27:04.960 --> 00:27:18.000 And yeah, it's a little bit strange fusion, perhaps, but it is actually what allows you at the end to drive software engineering on one side with or leverage from the possibilities of agentic AI tools like Cloud Code. 00:27:18.160 --> 00:27:24.880 So it's really a fusion of two things where on the first spot, yeah, do not belong together, but work like Wonder, actually. 00:27:25.920 --> 00:27:31.119 So, how does that differ from a harness environment that you put around your code that gets generated? 00:27:32.160 --> 00:27:38.720 Um, the harness itself is it's more or less pimped up with AZ on one side through hooks, mod skills, all the stuff. 00:27:38.799 --> 00:27:47.119 Uh so though, this this changes, but it still feels like your regular cloed code if you want, but you have additional features all the time which uh guide you. 00:27:47.279 --> 00:28:00.079 So, in technically, it's most of the time skills, but you have additional tools around this agent harness, like a Kanban board for visualizing uh your uh genti development pipeline, uh, all those additional things. 00:28:00.160 --> 00:28:03.279 So it's more or less an extension of Cloud Code, if you want. 00:28:03.440 --> 00:28:03.759 Yeah. 00:28:05.039 --> 00:28:07.839 Or any other environment as well, it works with the other ones as well. 00:28:07.920 --> 00:28:17.759 We could have copilots here line, open, yeah, because it just but the main focus is on Cloud Code because it's uh the most powerful uh Genti uh um agent harness currently on the market. 00:28:19.519 --> 00:28:32.400 Now uh uh it's interesting to see you have the uh you have the uh uh the testimonials, which is like the first testimonials from yourself, which is an interesting way to like how how people embraced it? 00:28:32.480 --> 00:28:34.880 Like, have you tried that with customers of MSG? 00:28:34.960 --> 00:28:39.200 Have you tried an MSG internally, or is this a side project that you're working on? 00:28:39.599 --> 00:28:41.119 Um all kinds of this. 00:28:41.200 --> 00:28:44.720 Uh on one side, I'm developing uh ASI with ASI itself. 00:28:44.799 --> 00:28:51.920 So all my open source uh projects in the last year are developed with ASI itself on on side projects. 00:28:52.079 --> 00:28:54.960 Then we are using it internally inside MSG. 00:28:55.279 --> 00:29:04.720 And uh we are also because we are doing not just uh creating products, we also do project uh with individualized uh custom software development for uh customers. 00:29:04.799 --> 00:29:06.319 We are using it also there. 00:29:06.720 --> 00:29:12.720 So it's used in multiple situations here because it's a toolkit, and as a toolkit, it's very flexible. 00:29:12.799 --> 00:29:15.759 Uh sometimes just part of ASI is used. 00:29:15.839 --> 00:29:23.759 Um, it can be even used just for requirement engineering, uh, and completely you can completely ignore all the code development stuff inside AZ. 00:29:24.000 --> 00:29:35.519 Uh, on the other hand, you can just do your open source project with it and just focus on bare metal coding, more or less, with the help of a Genty GI and AZ and ignore, for instance, requirements, engineering, testing, and other things. 00:29:35.680 --> 00:29:42.480 So, this is uh the the interesting point that it's a toolkit and you have a lot of tools inside there, and you can pick the particular tools. 00:29:42.640 --> 00:29:46.799 They are not tied to a particular process or workflow or something like this. 00:29:46.960 --> 00:29:53.440 And this is why on the website uh you can even read that you're in a driver's seat, so you decide what to do all the time. 00:29:53.599 --> 00:30:00.799 It's not a fixed framework which enforces a particular process or enforces a particular type of uh tool to use. 00:30:00.960 --> 00:30:13.119 So it's really a toolkit where you can choose on uh individual steps you do over your daily uh work hours, uh, what tool helps me and what uh can I ignore? 00:30:13.279 --> 00:30:16.799 So I'm using even AZ myself in completely different fashions. 00:30:16.880 --> 00:30:26.160 I'm sometimes even open

clothing code and do just brainstorming and other things through those skills because the ADC has even all those things a consultant needs. 00:30:26.240 --> 00:30:44.400 Uh just an example you want to find uh, for instance, software components on the internet, want to evaluate them with weighted decision matrices, want to uh create a decision for the customers, want to steal man this decision, or want an advocato diaboli to challenge your decision. 00:30:44.559 --> 00:30:46.400 So this has nothing to do with coding. 00:30:46.559 --> 00:30:49.599 This is also provided by AZ through a lot of skills. 00:30:49.759 --> 00:30:56.720 Um, because uh yeah, in the daily life of a software architect and software engineer, you have to do a lot more than just coding. 00:30:56.799 --> 00:31:05.920 So AZ supports also the coding part, but it provides a lot of additional functionalities for what software engineers need to do all the time. 00:31:06.079 --> 00:31:08.799 And this is uh a lot more than just coding. 00:31:09.119 --> 00:31:11.119 I love that it's really called Diaboli. 00:31:11.519 --> 00:31:12.400 This is amazing. 00:31:12.640 --> 00:31:23.839 Yeah, this is uh exactly this Diaboli is exactly the skill where you can give uh your um your statement or whatever you want and as an argument, and it tries to play the advocato diaboli. 00:31:24.160 --> 00:31:35.920 And uh yeah, it's interesting what happens there, uh, because uh you will be surprised uh what the AI actually finds as counter-arguments or antithesis for your arguments. 00:31:36.000 --> 00:31:37.279 I'm always surprised. 00:31:37.519 --> 00:31:45.440 My most uh um yeah enjoyed feature of AZ is actually named grilling, uh, just to get you an impression here. 00:31:45.519 --> 00:31:50.960 Uh, if you work with agentic AI coding tools like Lot Code, you can enter your prompt and it creates code. 00:31:51.039 --> 00:31:59.279 But at the end, you will be not fully satisfied with the result because uh you understand, okay, this I want a little bit different and that. 00:31:59.440 --> 00:32:05.359 It can be on the code level, it can be on the functionality side, it can be on the visible side on the UI. 00:32:05.519 --> 00:32:11.599 And what you do is uh you prompt against the result and let the uh Gen TI fix it. 00:32:11.680 --> 00:32:13.359 That's the usual thing people do. 00:32:13.519 --> 00:32:22.319 And what we found is that uh it helps a lot more if you create your task and before it is implemented, you get grilled on it. 00:32:22.480 --> 00:32:52.640 To get grilled on it means uh it uh the AI tries to relentlessly grill it your decisions, more or less, and find the open points because it's more efficient if the AI first tried to understand what later during implementation should be interpreted, and because it's open, it's most underspecified, so it's better to uh get questions on this and answer it before it is uh uh implemented than to post adjust the result all the time. 00:32:52.880 --> 00:33:05.599 And uh this is interesting because when I invented this feature over um multiple iterations, um people were rather skeptic and said, Okay, I should be grilled for I'm already uh telling in my prompt exactly what I want. 00:33:05.680 --> 00:33:21.119 And you would be surprised what uh then the bare uh world knowledge of an LM, a frontier LM, is able to do with your prompts and really figure out from the domain perspective, architecture perspective, implementation perspective, what the open points are. 00:33:21.359 --> 00:33:36.160 And uh in practice, it means you give a very precise, from your perspective, prompt, and the AI comes back with a table of about 10 questions, and I can tell you, more or less uh nine to ten questions are absolutely essential, and you are surprised. 00:33:36.400 --> 00:33:48.480 Oh, good that you actually uh uh raise those questions beforehand, because at the end, for all those decisions, which I usually would not have done, I had to do post adjustments. 00:33:48.559 --> 00:33:51.519 So, this is one more experience from practice. 00:33:51.680 --> 00:34:03.039 Um, agentic AI is great, but if you lap the agentic AI tool to really grill you for your decisions, the results become a lot better and you have to do less iterations all the time. 00:34:03.680 --> 00:34:15.360 That's what people keep saying about by coding as well, that instead of just prompting, you should actually ask the uh the LLM to interview you, to basically say, like, I want to build this, interview me about specs. 00:34:15.440 --> 00:34:22.639 And basically, much like a good uh software architect would do when we go to a customer, we just say, like, they're like, Oh, we want a login system. 00:34:22.719 --> 00:34:40.079 Okay, what kind of login system, what kind of issues, what kind of like you know, that grilling happens in in meetings normally, and it seems like we're getting away from that, like, just tell the machine to do things, but like let the machine ask me to find the unknown needs that we had before. 00:34:40.400 --> 00:34:40.800 Exactly. 00:34:40.880 --> 00:34:44.079 This is what some people call brief engineering currently. 00:34:44.320 --> 00:34:58.079 So you do a briefing of the machine, uh, and this way get a lot better uh results at the end because it's easier to work on particular domain-specific questions than you have to work on tons of generated code afterwards. 00:34:58.239 --> 00:35:00.320 Uh, so this is exactly what can be done. 00:35:00.400 --> 00:35:04.000 But you're uh here on the screen is uh, for instance, the Kanban board. 00:35:04.159 --> 00:35:09.280 Um, you as a um an old school uh IT veteran uh certainly will like that. 00:35:09.519 --> 00:35:10.960 This is grilling what you see here. 00:35:11.119 --> 00:35:17.920 This is exactly you get uh a question, get the the be grilled, and can click on answers, and below you see the Kanban board. 00:35:18.000 --> 00:35:22.800 And what you like is that in the left is the

web UI, and you have the same on the terminal. 00:35:22.880 --> 00:35:26.639 So even the Unix hackers like me uh have all the terminal available. 00:35:26.800 --> 00:35:30.320 This is also uh sometimes people hate me for this, uh, but I like it. 00:35:30.400 --> 00:35:38.559 I'm a Unix veteran, and uh AZA all skills have Unix uh long and short options and look like a Unix shell. 00:35:38.639 --> 00:35:43.679 And uh even uh such a Kanban board, what we see here on the screen is actually identically. 00:35:43.920 --> 00:35:50.480 If you just click on the uh uh screenshot on the uh right on the website, you see exactly the same on the terminal. 00:35:50.639 --> 00:35:52.880 So it's uh implemented twice actually. 00:35:52.960 --> 00:35:56.559 So the Unix terminal hackers can get uh the terminal exactly. 00:35:56.639 --> 00:36:02.159 This is the terminal version, so you see that it resembles uh even from the colors and other things, the web version. 00:36:02.239 --> 00:36:15.119 And this is the Kanban board where uh the lanes are more or less the agentic development pipeline where you put through your user stories, and uh you can imagine that the story is picked up by humans and agents all the time. 00:36:15.199 --> 00:36:18.880 Um, and this is about statistics in this uh in the pipeline, etc. 00:36:19.119 --> 00:36:31.440 So you see, it's a lot uh infrastructure in addition to what the regular agent GI coding tools like Cload Code provide in order to drive software engineering because soft engineering is a lot more than the bare coding part. 00:36:32.719 --> 00:36:36.320 But you're using you're using uh uh uh light mode. 00:36:36.559 --> 00:36:37.840 This is weird. 00:36:39.360 --> 00:36:41.360 Well, the website is in dark mode, yes. 00:36:41.440 --> 00:36:43.440 Uh uh, yes, it's it's real, it's true. 00:36:43.519 --> 00:36:51.280 Uh, this is in light mode here, and uh but well, with the help of a gente GI we can easily add also a dark mode in the future if it's light. 00:36:51.440 --> 00:36:56.480 Yes, it's it's uh the simple it's very simple because my terminal is in light mode, actually. 00:36:56.719 --> 00:36:59.599 And that's the reason why even the web version is in light mode. 00:37:00.000 --> 00:37:04.000 But light attracts bugs, that's the problem, isn't it? 00:37:04.480 --> 00:37:08.159 Oh well, it it it it it depends uh what we're used to. 00:37:08.239 --> 00:37:14.079 Uh, I'm working actually in in more or less in light mode all the time, uh, when it comes to the terminal since 40 years. 00:37:14.159 --> 00:37:17.679 So this is my preference, but it doesn't matter very much at the end. 00:37:17.840 --> 00:37:23.119 Most of the things uh inside AZ are uh, even for instance, uh things like uh the specification. 00:37:23.199 --> 00:37:26.159 Um, yeah, here you see it on the screen, just go uh up. 00:37:26.239 --> 00:37:32.239 Uh this is a light and dark mode uh stop here, no, a little bit down on spec, wherever the spec is here. 00:37:32.320

--> 00:37:35.679 Uh you see uh there's light and dark mode equally. 00:37:35.760 --> 00:37:41.760 So this is uh the generated specification out of AZ with uh uh sub-project is named Spec Book. 00:37:41.840 --> 00:37:48.400 And here you see it's light and dark mode for both people who like dark or the other one who like light. 00:37:48.800 --> 00:37:54.639 Now be honest, did you wipe code the website and the documentation or did you do it by hand? 00:37:54.960 --> 00:37:57.039 Uh no, it it depends uh what I'm doing. 00:37:57.280 --> 00:38:02.079 For instance, uh a website like AZ here is done with the Astro framework on one side. 00:38:02.239 --> 00:38:11.440 So what I've done manually is establish all uh this framework stuff, Astro, Astro configuration, uh Tailwind, CSS selection, all this stuff. 00:38:11.519 --> 00:38:14.239 But uh the particular content also came from me. 00:38:14.400 --> 00:38:21.280 But uh all the fiddling on the website with all uh the tailwind uh uh CSS classes is done by the AI at the end. 00:38:21.599 --> 00:38:41.440 So I I select uh the technology stack, I give the content, and the AI creates uh the particular layout here because uh, well, the last website I created myself was uh two or three years already ago, and I no longer have fun to do this stuff myself, even I'm as a web veteran can do it, obviously, but it's uh not necessary. 00:38:41.599 --> 00:38:51.119 On the other hand, uh AZ functionalities below are partly hard-coded, partly uh through the eye, but it's not really by coding, just uh to give you here an understanding from my side. 00:38:51.280 --> 00:39:00.239 Wipe coding means I tell with the prompt what I want, and I do not care about anything, how it actually looks, and I if it doesn't like uh what the result is I do another prompt. 00:39:00.320 --> 00:39:02.719 Uh but I do more as brief engineering through AZ. 00:39:02.800 --> 00:39:07.840 So I do a lot of initial thought uh together with the eye to brief the eye. 00:39:07.920 --> 00:39:13.039 And so at the end, it's not really an arbitrary code which comes out, it's exactly the code. 00:39:13.199 --> 00:39:17.360 It even is the code is exactly formatted uh as the other stuff. 00:39:17.440 --> 00:39:20.800 So it's a lot of restrictions what I apply with AZ. 00:39:21.119 --> 00:39:24.880 Uh so from I I do not call it wipe coding at the end, what I'm doing. 00:39:24.960 --> 00:39:28.480 I do a lot of brief engineering, give a lot of constraints to the AI. 00:39:28.880 --> 00:39:36.719 But yes, the AI creates a lot of code, obviously, because that's what I also want to leverage from uh agent myself. 00:39:37.280 --> 00:39:40.639 So AC AZ is on uh GitHub. 00:39:40.719 --> 00:39:45.679 Uh, do you have contributors from the outside already, or is it mostly people that you bugged in the office to help you out? 00:39:46.000 --> 00:39:52.880 Yeah, currently, well, it's uh as for most open source projects, they are around uh from the personal field of the primary

author. 00:39:52.960 --> 00:39:55.199 This was already at Apache Foundation and others. 00:39:55.280 --> 00:39:56.480 This is a classical thing. 00:39:56.559 --> 00:40:04.000 Yeah, the the contributors of AZ are currently around my CTO organization at the MSG, uh, because they are employees, uh friends of me. 00:40:04.159 --> 00:40:16.639 So this is well, as in all open source projects, there's one uh primary key figure that's this is me, and people in the personal or uh business uh uh yes circle or onion around this person. 00:40:16.719 --> 00:40:23.199 So very similar also on this project because it's a rather new project, it's just uh half a year uh old. 00:40:24.079 --> 00:40:27.440 But you do welcome pull requests and comments and yeah, sure. 00:40:27.599 --> 00:40:30.239 It's it's fully open source, Apache licensed. 00:40:30.320 --> 00:40:32.559 So, yes, obviously, it's fully open source. 00:40:32.960 --> 00:40:38.079 Everything is here available, it's not uh open weight or something like this, uh, as we talked before. 00:40:38.159 --> 00:40:40.960 It's a really a true open source project, yes. 00:40:41.360 --> 00:40:42.559 Well, very, very cool. 00:40:42.719 --> 00:40:43.440 Thank you so much. 00:40:43.519 --> 00:40:44.960 This was uh insightful. 00:40:45.039 --> 00:40:47.199 I think there's a lot of stuff that we can cover. 00:40:47.360 --> 00:40:52.800 Uh, might be inviting you back on the live show or something, so we can talk more about how it goes, how it went. 00:40:52.960 --> 00:40:59.280 Now, my last question when I go to the author page and I see you here, do you moonlight as a magician? 00:41:00.719 --> 00:41:10.000 As a magician, because it does look like a picture that you actually see it in in Las Vegas somewhere where it's like tonight, it's gonna be Ralph. 00:41:10.559 --> 00:41:16.559 Well, uh, I'm I'm actually not just an open source hacker, I'm also uh uh a guy standing on stage. 00:41:16.719 --> 00:41:19.119 So even the picture is something like this. 00:41:19.199 --> 00:41:22.719 Uh but and uh when it comes to software, well, perhaps it's uh magic. 00:41:22.880 --> 00:41:23.760 I don't know. 00:41:24.320 --> 00:41:28.960 You know, this has to be decided by others whether this is really artistic stuff I'm doing. 00:41:29.039 --> 00:41:37.119 Uh at least uh I hope I I've contributed a lot in the open source field for the last uh well over three decades now, and perhaps something is also magic. 00:41:37.280 --> 00:41:38.000 I don't know. 00:41:38.400 --> 00:41:45.760 Well, talking of which image magic just had a security flaw as well that needs fixing, so that's something we need to go for now as well. 00:41:46.079 --> 00:41:48.159 Well, this was this was delightful. 00:41:48.239 --> 00:41:56.800 It's great to have you on board, and it's great to have you uh at the are you gonna be at the event in Berlin again as well this year, or no, I was not participating, no. 00:41:57.280 --> 00:41:58.719 Will you come next year then?

00:41:59.280 --> 00:42:00.559 Perhaps. 00:42:00.880 --> 00:42:02.239 If I'm allowed it. 00:42:02.480 --> 00:42:15.599 The problem is in the industry, your time is immediate dramatically thin what is available to go to conferences and other things because you are such uh well drawn away with customer projects and and other things. 00:42:15.679 --> 00:42:19.280 So this is really I would like to enjoy to more get in contact with people. 00:42:19.360 --> 00:42:21.519 So I enjoyed even the invitation here. 00:42:21.679 --> 00:42:24.000 And if it time allows it, I can. 00:42:25.039 --> 00:42:25.920 Well, thanks very much. 00:42:26.079 --> 00:42:26.639 This was Dr. 00:42:26.719 --> 00:42:28.079 Ralph Engelschall, uh S. 00:42:28.239 --> 00:42:35.199 Engelshall, whatever that is, and uh that's my second first name, Sasha from MSG Group. 00:42:35.360 --> 00:42:43.599 And uh, we learned today about how you can actually merge the whole idea of uh uh agentic vibe coding and really, really boring old things. 00:42:43.760 --> 00:42:54.800 Oh, one last thing, uh OpenAI just announced that the new model that they're hiring is actually focused on uh on COBOL and Fortran to help with like fixing old banking environments. 00:42:54.960 --> 00:42:58.960 Is this something that you face where you have like really old languages that you have to support? 00:42:59.280 --> 00:43:04.159 We have definitely customers uh which have those old pieces in the banking and insurance field. 00:43:04.239 --> 00:43:05.119 Yes, definitely. 00:43:05.280 --> 00:43:16.559 So I can uh understand that OpenAI offers this particular model because uh there's certainly a large market uh to transform and migrate those old pieces of software into new eras. 00:43:16.639 --> 00:43:19.679 And it was rather hard in the past to do it without AI. 00:43:19.760 --> 00:43:24.400 So I even enjoyed that such specialized models came out, which can help here. 00:43:24.480 --> 00:43:25.039 Yeah. 00:43:25.360 --> 00:43:30.159 Yeah, because I mean freelancers that know COBOL can actually dictate their prices by now. 00:43:30.239 --> 00:43:35.840 Like, because if something that is broken, then it's something that is very vital for the bank or for an insurance company. 00:43:36.159 --> 00:43:43.760 Yeah, we had even people which were already retired and came back because they have this knowledge still available in their brain. 00:43:44.000 --> 00:43:44.480 Wow. 00:43:45.440 --> 00:43:48.079 But it was never fun to write COBOL, that was the issue with it. 00:43:48.159 --> 00:43:52.639 But like, and on the other hand, like I mean, what's your favorite language? 00:43:53.119 --> 00:44:01.840 My favorite language is actually TypeScript because it's uh a fantastic fusion of JavaScript and all the possibilities on the web, plus the static typing. 00:44:02.000 --> 00:44:08.960 And uh Anders Heilsberg, which we who invented this language, has really done a great job when it comes to the typing system. 00:44:09.119 --> 00:44:11.920 I really enjoy this typing system, the possibilities. 00:44:12.000 --> 00:44:20.400 So TypeScript is my personal preference, and that's the reason why my open source project for the last, I don't know, five or more years are only TypeScript in Node.js based. 00:44:21.199 --> 00:44:29.599 He does he did C sharp as well, and he was one of those people that in Microsoft I met in the cafe at Cafeteria from time to time, and I'm like, oh god, he's the guy that broke. 00:44:29.679 --> 00:44:31.920 And he's one of the nicest people in the planet. 00:44:32.000 --> 00:44:38.239 It's just like a really down-to-earth guy, and he's just a coder that wants to write things, and it's just wonderful to see that kind of stuff. 00:44:38.400 --> 00:44:39.280 I was choking the other. 00:44:41.280 --> 00:44:46.239 His results are wonderful, be it C sharp, is a great language, and the TypeScript also. 00:44:46.320 --> 00:44:53.039 It's uh a great fusion of the world of web with JavaScript and the classical statically typed languages. 00:44:53.119 --> 00:44:54.320 Yeah, really great. 00:44:54.639 --> 00:45:00.000 I was jokingly saying I'm going back to Perl, so nobody can actually read my code, neither agents nor humans. 00:45:00.239 --> 00:45:02.159 So uh we'll see how that works out. 00:45:02.320 --> 00:45:05.039 But yeah, the good old Pearl joke, exactly. 00:45:05.360 --> 00:45:07.679 You can what was it? 00:45:07.760 --> 00:45:11.679 Uh you can uh base a 64 it and not recognize on the other hand. 00:45:12.320 --> 00:45:13.280 Look like this. 00:45:14.639 --> 00:45:18.639 Same thing, like reverse, yeah, reverse engineering is gonna be tricky. 00:45:18.880 --> 00:45:19.840 Well, thanks very much. 00:45:20.000 --> 00:45:21.840 Uh this was Coffee with Developers. 00:45:22.000 --> 00:45:24.000 So uh we're gonna be on the platform soon. 00:45:24.079 --> 00:45:30.719 And uh, if you want to learn more about it, go to the GitHub, uh, try it out, try Ace and see if it actually works for your environment. 00:45:30.880 --> 00:45:34.159 I love that you also have a like why is it for you and why is it not for you? 00:45:34.320 --> 00:45:40.639 This is also a very good thing to get people to filter them out without uh so they're not uh actually disappointed by it. 00:45:40.800 --> 00:45:41.920 So that's a wonderful thing. 00:45:42.000 --> 00:45:44.400 So go and try it out, and that was it from us. 00:45:44.559 --> 00:45:45.599 Thanks very much. 00:45:46.000 --> 00:45:47.519 Thanks for the invitation.